Stop mavwaked from hearing itself, and add barge-in (#287)

Playback was `go playAudio(reply)` — fire and forget, nobody holding the
process handle. Two audible consequences fell out of that.

She answered herself. The capture loop kept feeding the VAD while the
speaker was running, so her own reply came back in through the mic,
tripped the VAD, and was shipped to the daemon as a fresh command. There
is no acoustic echo canceller in this pipeline, so the fix is
half-duplex: while she is speaking, the capture side is muted. That part
is unconditional — it repairs a defect, it is not a new capability.

And talking over her did nothing, because there was no handle to cancel.
-barge-in now cuts playback when sustained energy clears a room-tuned
threshold (-barge-in-rms, default 0.12 normalised, over -barge-in-frames
consecutive frames, default 5). It is off by default: without an echo
canceller the only way to tell "he is talking over her" from "the mic is
hearing her" is that he is much louder, and how much louder depends on
where the mic sits.

The frame decision moved out of main.go into session.feed, behind a
player and an utteranceSender interface, so all of it is testable with
no mic, no speaker and no daemon. Nine tests cover the self-hearing
case, the off-by-default case, the consecutive-frame requirement,
speaker-leak-level audio not triggering, capturing the interrupting
utterance after a cut, and failed round-trips not starting playback.

The other seven items on #287 (partial STT, per-segment retry, mic
profiles, noise-floor calibration, short-response-while-speaking) are
untouched and stay on the task.

Makefile

@@ -16,7 +16,7 @@ PIPER_BIN     := $(shell pwd)/deps/piper/piper
 PIPER_MODEL   := $(shell pwd)/models/tts/ru_RU-irina-medium.onnx
 PIPER_ESPEAK  := $(shell pwd)/deps/piper/espeak-ng-data
 
-.PHONY: all build build-stt build-tts build-daemon build-client build-waked build-web build-poll build-caldav clean test fmt-check vet run-stt run-tts run-web download-embedder deps-go eval-router eval-recall eval-phrasing eval-models
+.PHONY: stt-fixtures test-stt-golden all build build-stt build-tts build-daemon build-client build-waked build-web build-poll build-caldav clean test fmt-check vet run-stt run-tts run-web download-embedder deps-go eval-router eval-recall eval-phrasing eval-models
 
 all: build
 
@@ -139,6 +139,17 @@ eval-models:
 	MAVEN_LLM_URL="$(MAVEN_LLM_URL)" $(GO) test -v -count=1 -timeout 60m \
 		-run TestLLMRouterBaseline ./internal/router/eval/
 
+# stt-fixtures — regenerate the golden STT audio in cmd/mavsttd/testdata from
+# the piper voices (#288). The committed WAVs are synthesised, never recorded,
+# so this is the only way they should ever change. TestGoldenAudioTranscription
+# then scores them against ggml-small; it self-skips when the model is absent.
+stt-fixtures:
+	./scripts/gen-stt-fixtures.sh
+
+test-stt-golden:
+	CGO_CFLAGS="$(CGO_CFLAGS)" CGO_LDFLAGS="$(CGO_LDFLAGS)" LD_LIBRARY_PATH="$(shell pwd)/deps/lib" \
+		$(GO) test -v -count=1 -run TestGolden ./cmd/mavsttd/
+
 run-stt: build-stt
 	LD_LIBRARY_PATH="$(shell pwd)/deps/lib" \
 		./mavsttd -socket /tmp/maven/stt.sock -model $(WHISPER_MODEL)

cmd/mavend/main.go

@@ -340,6 +340,10 @@ func run(args []string) error {
 		// retention loop. Off unless a capture block enables it, in which case
 		// all four capture methods answer ErrUnknownMethod.
 		wireCapture(srv, keeper, st, voiceW, phr, cfg)
+		// Voice identification (Vikunja #255). Enrolment plumbing only until a
+		// speaker-embedding model exists on disk; off entirely without a speaker
+		// block, so no wire path takes a voiceprint on a default box.
+		wireSpeaker(srv, st, cfg)
 	}
 
 	// WrapKeyFn — wraps the env key with a passkey credential public key and
@@ -485,6 +489,10 @@ func run(args []string) error {
 			wireModelSwap(srv, phr, cfg)
 			keeper := wireVision(ctx, srv, st, embedderOf(voiceW), cfg)
 			wireCapture(srv, keeper, st, voiceW, phr, cfg)
+			// Voice identification (Vikunja #255). Enrolment plumbing only until a
+			// speaker-embedding model exists on disk; off entirely without a speaker
+			// block, so no wire path takes a voiceprint on a default box.
+			wireSpeaker(srv, st, cfg)
 
 			// Start voice server.
 			if voiceW != nil {

cmd/mavend/speaker.go

@@ -0,0 +1,146 @@
+// mavend/speaker.go — core's half of voice identification (Vikunja #255,
+// docs/plans/10-speaker-recognition.md).
+//
+// # What is actually wired here, and what is not
+//
+// The enrolment plumbing is real: profiles are stored, listed and deleted, and
+// the wire methods exist as soon as a speaker block is configured. The
+// recognising half is NOT, and cannot be on this box, because there is no
+// speaker-embedding model on disk — no ECAPA, no x-vector, no titanet, no
+// wespeaker, nothing in /mnt/hdd1/llms but text ggufs. Until one is downloaded,
+// newSpeakerEmbedder returns nil, internal/speaker falls back to
+// speaker.Disabled, and every Identify answers ErrDisabled. The daemon logs
+// which half is off at startup rather than pretending.
+//
+// This is deliberately not papered over with a hand-rolled MFCC floor. A
+// biometric that is confidently wrong writes false claims about named people
+// into his memory, and that is worse than a capability that is honestly absent.
+//
+// # Off unless configured
+//
+// No speaker block, or one without enabled, ⇒ the three methods do not exist and
+// answer ErrUnknownMethod. On an unconfigured box there is no wire path that
+// takes a voiceprint at all.
+//
+// # The refused design step
+//
+// The plan asks for unknown speakers to be enrolled on first interaction. That
+// is refused in internal/speaker/enroll.go and there is no handler for it here:
+// no request shape in the protocol enrols whoever just spoke. Taking a biometric
+// of a guest who walked past the microphone is not something this daemon does.
+package main
+
+import (
+	"context"
+	"errors"
+	"log"
+
+	"github.com/kami/maven/internal/config"
+	"github.com/kami/maven/internal/ipc"
+	"github.com/kami/maven/internal/speaker"
+	"github.com/kami/maven/internal/store"
+)
+
+// speakerWiring holds the recognizer behind the three IPC handlers.
+type speakerWiring struct {
+	rec *speaker.Recognizer
+}
+
+// newSpeakerEmbedder loads the speaker-embedding model named by the config.
+//
+// It always returns nil today. The seam exists so that wiring a real model is a
+// change to this one function and nothing else: give it a loader, and Identify
+// starts working with no change to the store, the protocol, the auth table or
+// the handlers. See the plan document for what to download.
+func newSpeakerEmbedder(cfg *config.SpeakerConfig) speaker.Embedder {
+	if cfg == nil || cfg.ModelPath == "" {
+		return nil
+	}
+	log.Printf("speaker: model_path %q is configured but no embedding backend is built yet; "+
+		"enrolment and deletion work, recognition does not (Vikunja #255)", cfg.ModelPath)
+	return nil
+}
+
+// newSpeakerWiring builds the recognizer, or nil when the capability is off.
+func newSpeakerWiring(st *store.Store, cfg *config.Config) *speakerWiring {
+	if cfg == nil || cfg.Speaker == nil || !cfg.Speaker.Enabled {
+		return nil
+	}
+	if st == nil {
+		log.Print("speaker: enabled but there is no store to keep profiles in; staying off")
+		return nil
+	}
+	rec, err := speaker.New(newSpeakerEmbedder(cfg.Speaker), st.VectorMemory(), speaker.Config{
+		Threshold:  cfg.Speaker.Threshold,
+		MinSeconds: cfg.Speaker.MinSeconds,
+	})
+	if err != nil {
+		log.Printf("speaker: %v; staying off", err)
+		return nil
+	}
+	if rec.Enabled() {
+		log.Printf("speaker: recognition on, threshold %.2f", rec.Threshold())
+	} else {
+		log.Print("speaker: enrolment on, recognition BLOCKED — no speaker-embedding model " +
+			"on this box (see docs/plans/10-speaker-recognition.md)")
+	}
+	return &speakerWiring{rec: rec}
+}
+
+func (w *speakerWiring) enroll(ctx context.Context, req ipc.EnrollSpeakerReq) (ipc.EnrollSpeakerResp, error) {
+	p, err := w.rec.Enroll(ctx, req.ID, req.Name, req.Samples)
+	if err != nil {
+		return ipc.EnrollSpeakerResp{}, speakerErr(err)
+	}
+	return ipc.EnrollSpeakerResp{Speaker: toWireSpeaker(p)}, nil
+}
+
+func (w *speakerWiring) list(ctx context.Context) (ipc.ListSpeakersResp, error) {
+	ps, err := w.rec.List(ctx)
+	if err != nil {
+		return ipc.ListSpeakersResp{}, speakerErr(err)
+	}
+	out := make([]ipc.Speaker, 0, len(ps))
+	for _, p := range ps {
+		out = append(out, toWireSpeaker(p))
+	}
+	return ipc.ListSpeakersResp{Speakers: out, Enabled: w.rec.Enabled()}, nil
+}
+
+func (w *speakerWiring) forget(ctx context.Context, req ipc.ForgetSpeakerReq) error {
+	return speakerErr(w.rec.Forget(ctx, req.ID))
+}
+
+// toWireSpeaker drops the voiceprint. A listing says who is enrolled; it does
+// not hand the biometric back out over the socket.
+func toWireSpeaker(p speaker.Profile) ipc.Speaker {
+	return ipc.Speaker{ID: p.ID, Name: p.Name, Enrolled: p.Enrolled, Samples: p.Samples}
+}
+
+// speakerErr maps the package sentinels onto the wire vocabulary so a surface
+// can tell "you asked wrong" from "core broke".
+func speakerErr(err error) error {
+	switch {
+	case err == nil:
+		return nil
+	case errors.Is(err, speaker.ErrNotFound):
+		return ipc.ErrNoFact
+	case errors.Is(err, speaker.ErrBadID),
+		errors.Is(err, speaker.ErrBadFormat),
+		errors.Is(err, speaker.ErrTooShort):
+		return errors.Join(ipc.ErrBadParams, err)
+	default:
+		return err
+	}
+}
+
+// wireSpeaker attaches the three handlers when the capability is configured.
+func wireSpeaker(srv *ipc.Server, st *store.Store, cfg *config.Config) {
+	w := newSpeakerWiring(st, cfg)
+	if w == nil {
+		return
+	}
+	srv.EnrollSpeakerFn = w.enroll
+	srv.ListSpeakersFn = w.list
+	srv.ForgetSpeakerFn = w.forget
+}

cmd/mavsttd/golden_test.go

@@ -0,0 +1,323 @@
+package main
+
+// Golden-audio STT tests (Vikunja #288).
+//
+// These push real audio through the real whisper.cpp binding, so a bad model
+// path, a wrong language hint, a broken resample or a regressed silence gate
+// is caught by `make test` rather than by the owner talking to a daemon that
+// mishears him.
+//
+// The fixtures are piper-synthesised, not recorded — see
+// scripts/gen-stt-fixtures.sh. Nothing of the owner's voice is committed, and
+// any fixture can be rebuilt from the script plus a voice model.
+//
+// Matching is deliberately tolerant. Golden transcripts are model-dependent:
+// swapping ggml-small for a different whisper build moves punctuation, casing
+// and the odd word ending, and an exact-string assertion would turn every
+// model swap into a fixture rewrite. Each case therefore asserts two things —
+// the words that carry the intent are present, and the word error rate
+// against the reference stays under a per-case ceiling.
+
+import (
+	"context"
+	"encoding/json"
+	"os"
+	"path/filepath"
+	"strings"
+	"testing"
+	"unicode"
+
+	"github.com/kami/maven/internal/audio"
+	"github.com/kami/maven/internal/worker"
+)
+
+// goldenModelPath — the whisper model the golden tests run against. Same file
+// the Makefile's run-stt target uses. Overridable so a box that keeps its
+// models elsewhere can still run these.
+func goldenModelPath() string {
+	if p := os.Getenv("MAVEN_WHISPER_MODEL"); p != "" {
+		return p
+	}
+	return filepath.Join("..", "..", "models", "stt", "ggml-small.bin")
+}
+
+type goldenCase struct {
+	Name     string   `json:"name"`
+	WAV      string   `json:"wav"`
+	Lang     string   `json:"lang"`
+	Text     string   `json:"text"`
+	Keywords []string `json:"keywords"`
+	MaxWER   float64  `json:"max_wer"`
+}
+
+type goldenManifest struct {
+	Cases []goldenCase `json:"cases"`
+}
+
+func loadGoldenManifest(t *testing.T) goldenManifest {
+	t.Helper()
+	raw, err := os.ReadFile(filepath.Join("testdata", "golden_v1.json"))
+	if err != nil {
+		t.Fatalf("read golden manifest: %v", err)
+	}
+	var m goldenManifest
+	if err := json.Unmarshal(raw, &m); err != nil {
+		t.Fatalf("parse golden manifest: %v", err)
+	}
+	if len(m.Cases) == 0 {
+		t.Fatal("golden manifest has no cases")
+	}
+	return m
+}
+
+// normalizeTranscript lowercases, drops punctuation, folds the Russian ё onto
+// е (whisper is inconsistent about it and the router does not care), and
+// collapses whitespace. Everything the comparison does happens on this form.
+func normalizeTranscript(s string) []string {
+	var b strings.Builder
+	for _, r := range strings.ToLower(s) {
+		switch {
+		case r == 'ё':
+			b.WriteRune('е')
+		case unicode.IsLetter(r) || unicode.IsDigit(r):
+			b.WriteRune(r)
+		default:
+			b.WriteRune(' ')
+		}
+	}
+	return strings.Fields(b.String())
+}
+
+// wordErrorRate is the Levenshtein distance between two word sequences,
+// divided by the length of the reference. 0 means identical; it can exceed 1
+// when the hypothesis is much longer than the reference.
+func wordErrorRate(ref, hyp []string) float64 {
+	if len(ref) == 0 {
+		if len(hyp) == 0 {
+			return 0
+		}
+		return 1
+	}
+	prev := make([]int, len(hyp)+1)
+	cur := make([]int, len(hyp)+1)
+	for j := range prev {
+		prev[j] = j
+	}
+	for i := 1; i <= len(ref); i++ {
+		cur[0] = i
+		for j := 1; j <= len(hyp); j++ {
+			cost := 1
+			if ref[i-1] == hyp[j-1] {
+				cost = 0
+			}
+			cur[j] = min(prev[j]+1, min(cur[j-1]+1, prev[j-1]+cost))
+		}
+		prev, cur = cur, prev
+	}
+	return float64(prev[len(hyp)]) / float64(len(ref))
+}
+
+// missingKeywords returns the keywords absent from the hypothesis. A keyword
+// matches on prefix, so a different case ending ("воды" vs "воду") does not
+// fail the assertion — the router's stage-0 grammar is stem-shaped too.
+func missingKeywords(keywords []string, hyp []string) []string {
+	var missing []string
+	for _, kw := range keywords {
+		want := normalizeTranscript(kw)
+		if len(want) == 0 {
+			continue
+		}
+		if !containsSeq(hyp, want) {
+			missing = append(missing, kw)
+		}
+	}
+	return missing
+}
+
+func containsSeq(hyp, want []string) bool {
+	for i := 0; i+len(want) <= len(hyp); i++ {
+		ok := true
+		for j, w := range want {
+			// Prefix match, so inflection differences pass but
+			// distinct words do not.
+			if !looseWordMatch(hyp[i+j], w) {
+				ok = false
+				break
+			}
+		}
+		if ok {
+			return true
+		}
+	}
+	return false
+}
+
+func looseWordMatch(got, want string) bool {
+	if got == want {
+		return true
+	}
+	g, w := []rune(got), []rune(want)
+	n := len(w) - 1
+	if len(w) > 6 {
+		n = len(w) - 2
+	}
+	// Words of three runes or fewer have no room for a safe prefix: require
+	// an exact match rather than letting "час" pass for "часть".
+	if n < 3 || len(g) < n {
+		return false
+	}
+	return string(g[:n]) == string(w[:n])
+}
+
+// --- the model-backed test -------------------------------------------------
+
+func TestGoldenAudioTranscription(t *testing.T) {
+	m := loadGoldenManifest(t)
+
+	model := goldenModelPath()
+	if _, err := os.Stat(model); err != nil {
+		t.Skipf("whisper model %s absent (%v) — set MAVEN_WHISPER_MODEL or see AGENTS.md", model, err)
+	}
+
+	// Same gate thresholds as mavsttd's defaults, so a regression in the
+	// silence gate shows up here as an empty transcript.
+	h, err := newWhisperHandler(model, 300, 0.01)
+	if err != nil {
+		t.Fatalf("load whisper model %s: %v", model, err)
+	}
+	defer h.Close()
+
+	for _, c := range m.Cases {
+		t.Run(c.Name, func(t *testing.T) {
+			path := filepath.Join("testdata", c.WAV)
+			raw, err := os.ReadFile(path)
+			if err != nil {
+				t.Skipf("fixture %s absent (%v) — run scripts/gen-stt-fixtures.sh", path, err)
+			}
+			format, pcm, err := audio.PCMFromWAV(raw)
+			if err != nil {
+				t.Fatalf("%s is not canonical 16k mono PCM: %v", path, err)
+			}
+
+			resp, err := h.Transcribe(context.Background(), worker.TranscribeReq{
+				Audio: audio.Audio{Format: format, Bytes: pcm},
+				Lang:  c.Lang,
+			})
+			if err != nil {
+				t.Fatalf("transcribe %s: %v", c.WAV, err)
+			}
+			t.Logf("%s → %q (confidence %.3f)", c.WAV, resp.Text, resp.Confidence)
+
+			if strings.TrimSpace(resp.Text) == "" {
+				t.Fatalf("%s transcribed to empty text — the silence gate ate real speech", c.WAV)
+			}
+			if resp.Confidence <= 0 {
+				t.Errorf("%s: confidence %v, want > 0", c.WAV, resp.Confidence)
+			}
+
+			hyp := normalizeTranscript(resp.Text)
+			ref := normalizeTranscript(c.Text)
+
+			if missing := missingKeywords(c.Keywords, hyp); len(missing) > 0 {
+				t.Errorf("%s: missing keywords %v in %q", c.WAV, missing, resp.Text)
+			}
+			if wer := wordErrorRate(ref, hyp); wer > c.MaxWER {
+				t.Errorf("%s: WER %.2f > %.2f\n  want: %q\n  got:  %q", c.WAV, wer, c.MaxWER, c.Text, resp.Text)
+			}
+		})
+	}
+}
+
+// TestGoldenFixturesAreCanonical checks the committed audio without needing a
+// model, so a fixture regenerated at the wrong sample rate fails on every box.
+func TestGoldenFixturesAreCanonical(t *testing.T) {
+	m := loadGoldenManifest(t)
+	for _, c := range m.Cases {
+		path := filepath.Join("testdata", c.WAV)
+		raw, err := os.ReadFile(path)
+		if err != nil {
+			t.Errorf("fixture %s missing: %v", path, err)
+			continue
+		}
+		format, pcm, err := audio.PCMFromWAV(raw)
+		if err != nil {
+			t.Errorf("%s: %v", path, err)
+			continue
+		}
+		if !format.IsValid() {
+			t.Errorf("%s: format %+v is not canonical", path, format)
+		}
+		a := audio.Audio{Format: format, Bytes: pcm}
+		if d := a.Duration(); d < 0.5 || d > 10 {
+			t.Errorf("%s: duration %.2fs outside the sane 0.5–10s fixture range", path, d)
+		}
+		// The fixture must clear mavsttd's own silence gate, otherwise the
+		// model test below would be asserting on a gated empty string.
+		if reason := gateReason(pcmToF32(pcm), whisperSampleRate, 300, 0.01); reason != "" {
+			t.Errorf("%s: would be gated as %s", path, reason)
+		}
+		if len(c.Keywords) == 0 {
+			t.Errorf("%s: manifest case has no keywords", c.Name)
+		}
+		if c.MaxWER <= 0 || c.MaxWER > 1 {
+			t.Errorf("%s: max_wer %v outside (0,1]", c.Name, c.MaxWER)
+		}
+	}
+}
+
+func pcmToF32(b []byte) []float32 {
+	out := make([]float32, len(b)/2)
+	for i := range out {
+		s := int16(b[i*2]) | int16(b[i*2+1])<<8
+		out[i] = float32(s) / 32768.0
+	}
+	return out
+}
+
+// --- matcher unit tests (no model, no fixtures) ----------------------------
+
+func TestNormalizeTranscript(t *testing.T) {
+	got := normalizeTranscript(" Ещё,  Раз... ")
+	want := []string{"еще", "раз"}
+	if len(got) != len(want) || got[0] != want[0] || got[1] != want[1] {
+		t.Fatalf("normalizeTranscript = %v, want %v", got, want)
+	}
+}
+
+func TestWordErrorRate(t *testing.T) {
+	cases := []struct {
+		name     string
+		ref, hyp string
+		want     float64
+	}{
+		{"identical", "напомни мне через час", "Напомни мне через час.", 0},
+		{"one substitution", "напомни мне через час", "напомни мне через день", 0.25},
+		{"one deletion", "напомни мне через час", "напомни мне час", 0.25},
+		{"empty hypothesis", "напомни мне", "", 1},
+		{"both empty", "", "", 0},
+	}
+	for _, c := range cases {
+		t.Run(c.name, func(t *testing.T) {
+			got := wordErrorRate(normalizeTranscript(c.ref), normalizeTranscript(c.hyp))
+			if got != c.want {
+				t.Fatalf("WER = %v, want %v", got, c.want)
+			}
+		})
+	}
+}
+
+func TestMissingKeywords(t *testing.T) {
+	hyp := normalizeTranscript("Отметь, что я выпил воду.")
+	if got := missingKeywords([]string{"воды", "отметь"}, hyp); len(got) != 0 {
+		t.Fatalf("missingKeywords = %v, want none (inflection must not fail the match)", got)
+	}
+	if got := missingKeywords([]string{"календарю"}, hyp); len(got) != 1 {
+		t.Fatalf("missingKeywords = %v, want the absent keyword reported", got)
+	}
+	// A short word must match exactly — no 4-rune prefix shortcut that would
+	// let "час" pass for "часть".
+	hyp2 := normalizeTranscript("через час")
+	if got := missingKeywords([]string{"часть"}, hyp2); len(got) != 1 {
+		t.Fatalf("missingKeywords = %v, want %q reported missing", got, "часть")
+	}
+}

cmd/mavsttd/testdata/golden_v1.json

@@ -0,0 +1,37 @@
+{
+  "note": "Golden STT fixtures. Audio is piper-synthesised, not recorded — see scripts/gen-stt-fixtures.sh. Regenerate with that script; do not hand-edit `wav`.",
+  "cases": [
+    {
+      "name": "ru_reminder",
+      "wav": "ru_reminder.wav",
+      "lang": "ru",
+      "text": "напомни мне через час позвонить маме",
+      "keywords": ["напомни", "час", "позвонить"],
+      "max_wer": 0.34
+    },
+    {
+      "name": "ru_fact",
+      "wav": "ru_fact.wav",
+      "lang": "ru",
+      "text": "отметь что я выпил воды",
+      "keywords": ["отметь", "воды"],
+      "max_wer": 0.34
+    },
+    {
+      "name": "ru_query",
+      "wav": "ru_query.wav",
+      "lang": "ru",
+      "text": "что у меня сегодня по календарю",
+      "keywords": ["сегодня", "календарю"],
+      "max_wer": 0.34
+    },
+    {
+      "name": "en_act",
+      "wav": "en_act.wav",
+      "lang": "en",
+      "text": "restart the web server and check the disk space",
+      "keywords": ["restart", "server", "disk"],
+      "max_wer": 0.34
+    }
+  ]
+}

cmd/mavwaked/main.go

@@ -12,8 +12,15 @@
 // (30ms frames, 16kHz PCM) matches silero-vad's input interface exactly, so
 // swapping energy-threshold for ONNX-inference is a local change in vad.go.
 //
+// While a reply is playing the capture side is muted (half-duplex): without
+// it, Maven's own voice comes back in through the mic and she answers
+// herself. -barge-in punches one hole in that gate — sustained energy well
+// above the speaker's leak level cuts playback so he can talk over her. It is
+// off by default because the threshold is room-specific; see playback.go.
+//
 //	usage:
 //	  mavwaked                          # default ALSA device, 127.0.0.1:9100
+//	  mavwaked -barge-in                # let him interrupt her mid-reply
 //	  mavwaked -device hw:1,0 -addr 10.42.0.1:9100
 //	  mavwaked -test file.wav           # read from file, no arecord
 package main
@@ -60,6 +67,9 @@ func run(args []string) error {
 	silenceMs := flag.Int("silence-ms", defaultSilenceMs, "silence ms to end utterance")
 	maxMs := flag.Int("max-ms", defaultMaxMs, "max utterance ms")
 	testFile := flag.String("test", "", "read PCM from file instead of arecord (testing only)")
+	bargeIn := flag.Bool("barge-in", false, "cut Maven off when he talks over her (needs a room-tuned -barge-in-rms)")
+	bargeRMS := flag.Int("barge-in-rms", defaultBargeRMS, "RMS x10000 a frame must clear to count as barge-in")
+	bargeFrames := flag.Int("barge-in-frames", defaultBargeFrames, "consecutive frames over -barge-in-rms before playback is cut")
 	flag.CommandLine.Parse(args)
 
 	ctx, stop := signal.NotifyContext(context.Background(), syscall.SIGINT, syscall.SIGTERM, syscall.SIGHUP)
@@ -117,14 +127,21 @@ func run(args []string) error {
 
 	defer src.Close()
 
-	return captureLoop(ctx, src, vad, vc, *lang)
+	var barge bargeInConfig
+	if *bargeIn {
+		barge = bargeInConfig{RMS: float64(*bargeRMS) / 10000.0, Frames: *bargeFrames}
+		log.Printf("mavwaked: barge-in on (rms %.4f x %d frames)", barge.RMS, barge.Frames)
+	}
+	sess := newSession(vad, newAplayPlayer(), &voiceSender{vc: vc}, *lang, barge)
+
+	return captureLoop(ctx, src, sess)
 }
 
-// captureLoop reads PCM from src, runs VAD, and sends complete utterances to
-// the voice server. Returns when ctx is done or src is exhausted.
-func captureLoop(ctx context.Context, src io.Reader, vad *VAD, vc *voice.Client, lang string) error {
+// captureLoop reads PCM from src and hands whole frames to the session.
+// Returns when ctx is done or src is exhausted.
+func captureLoop(ctx context.Context, src io.Reader, sess *session) error {
 	br := bufio.NewReaderSize(src, defaultReadSize)
-	frameBytes := vad.FrameSamples() * 2 // 480 samples × 2 bytes = 960 bytes per 30ms
+	frameBytes := sess.vad.FrameSamples() * 2 // 480 samples × 2 bytes = 960 bytes per 30ms
 
 	log.Printf("mavwaked: capture loop starting (frame=%d bytes, %dms)",
 		frameBytes, defaultFrameMs)
@@ -147,7 +164,7 @@ func captureLoop(ctx context.Context, src io.Reader, vad *VAD, vc *voice.Client,
 					// Flush partial frame.
 					partial = append(partial, buf[:n]...)
 					if len(partial) >= frameBytes {
-						if err := processFrame(partial[:frameBytes], vad, vc, lang); err != nil {
+						if err := sess.feed(ctx, partial[:frameBytes]); err != nil {
 							log.Printf("mavwaked: process frame: %v", err)
 						}
 						partial = partial[frameBytes:]
@@ -165,107 +182,31 @@ func captureLoop(ctx context.Context, src io.Reader, vad *VAD, vc *voice.Client,
 			partial = nil
 		}
 
-		if err := processFrame(full, vad, vc, lang); err != nil {
+		if err := sess.feed(ctx, full); err != nil {
 			log.Printf("mavwaked: process frame: %v", err)
 		}
 	}
 }
 
-// processFrame feeds one 30ms PCM frame to the VAD and sends any completed
-// utterance to the voice server.
-func processFrame(frame []byte, vad *VAD, vc *voice.Client, lang string) error {
-	samples := PCMToI16(frame)
-	utt, state := vad.Feed(samples)
+// voiceSender is the production utteranceSender: one PushToTalk round-trip
+// over the voice wire. SurfaceVoice (not the default SurfacePCClient that
+// c.PushToTalk uses) caps everything at L0, which is what makes an accidental
+// VAD trigger safe.
+type voiceSender struct{ vc *voice.Client }
 
-	if state == StateSpeech {
-		// Speech is in progress; nothing to send yet.
-		return nil
-	}
-
-	if utt.Bytes == nil {
-		// Still in silence, or short speech that didn't trigger.
-		return nil
-	}
-
-	// We have a complete utterance — send it to the voice server.
-	return sendUtterance(context.Background(), utt, vc, lang)
-}
-
-// sendUtterance sends audio to the voice server and plays the reply.
-func sendUtterance(ctx context.Context, utt audio.Audio, vc *voice.Client, lang string) error {
-	dur := utt.Duration()
-	log.Printf("mavwaked: utterance complete (%.2fs, %d bytes), sending...",
-		dur, len(utt.Bytes))
-
-	// Use SendRequest directly so we can set SurfaceVoice instead of the
-	// default SurfacePCClient that c.PushToTalk uses.
+func (s *voiceSender) Send(ctx context.Context, utt audio.Audio, lang string) (audio.Audio, error) {
 	var resp voice.PushToTalkResp
-	err := vc.SendRequest(ctx, voice.MethodPushToTalk, voice.PushToTalkReq{
+	err := s.vc.SendRequest(ctx, voice.MethodPushToTalk, voice.PushToTalkReq{
 		Audio:   utt,
 		Lang:    lang,
 		Surface: voice.SurfaceVoice,
 	}, &resp)
 	if err != nil {
-		return fmt.Errorf("push-to-talk: %w", err)
+		return audio.Audio{}, fmt.Errorf("push-to-talk: %w", err)
 	}
-
 	log.Printf("mavwaked: reply: %q (%.2fs audio)", resp.ReplyText, resp.ReplyAudio.Duration())
-
-	// Play the reply audio.
-	if len(resp.ReplyAudio.Bytes) > 0 {
-		go playAudio(resp.ReplyAudio)
-	} else {
-		log.Printf("mavwaked: empty reply audio (text only)")
-	}
-
 	if len(resp.RoutedChannels) > 0 {
 		log.Printf("mavwaked: also routed to: %v", resp.RoutedChannels)
 	}
-
-	return nil
-}
-
-// playAudio pipes PCM audio to aplay(1) for playback. Runs in a goroutine.
-func playAudio(a audio.Audio) {
-	// Build WAV header for aplay (or pipe raw PCM with the right format flags).
-	cmd := exec.Command("aplay",
-		"-f", "S16_LE",
-		"-r", fmt.Sprintf("%d", a.Format.SampleRate),
-		"-c", fmt.Sprintf("%d", a.Format.Channels),
-		"-t", "raw",
-	)
-
-	stdin, err := cmd.StdinPipe()
-	if err != nil {
-		log.Printf("mavwaked: aplay stdin pipe: %v", err)
-		return
-	}
-
-	if err := cmd.Start(); err != nil {
-		log.Printf("mavwaked: start aplay: %v", err)
-		return
-	}
-
-	// Write audio to aplay's stdin.
-	if _, err := stdin.Write(a.Bytes); err != nil {
-		log.Printf("mavwaked: write to aplay: %v", err)
-	}
-	_ = stdin.Close()
-
-	// Wait for playback to finish (with a timeout).
-	done := make(chan error, 1)
-	go func() {
-		done <- cmd.Wait()
-	}()
-
-	select {
-	case err := <-done:
-		if err != nil {
-			log.Printf("mavwaked: aplay: %v", err)
-		}
-	case <-time.After(30 * time.Second):
-		log.Printf("mavwaked: aplay timeout, killing")
-		_ = cmd.Process.Kill()
-		<-done
-	}
+	return resp.ReplyAudio, nil
 }

cmd/mavwaked/playback.go

@@ -0,0 +1,139 @@
+package main
+
+// Reply playback, and the half-duplex gate around it (Vikunja #287).
+//
+// Before this, playback was `go playAudio(reply)` — fire and forget, with no
+// handle on the running aplay. Two things fell out of that, and both are
+// audible:
+//
+//  1. Self-trigger. The capture loop keeps feeding the VAD while the speaker
+//     is playing, so Maven's own reply comes back in through the mic, trips
+//     the VAD, and is sent to the daemon as a fresh utterance. She answers
+//     herself. There is no acoustic echo canceller in this pipeline, so the
+//     only correct fix is half-duplex: while she is speaking, the capture
+//     side is muted.
+//
+//  2. No barge-in. Talking over her did nothing — there was nothing to
+//     cancel, because nobody held the process handle.
+//
+// The two are the same mechanism seen from opposite sides, so they live
+// together here. Echo suppression is unconditional (it fixes a bug). Barge-in
+// is off unless -barge-in is passed, because it needs a room-specific energy
+// threshold: with no echo canceller, the only way to tell "he is talking over
+// her" from "the mic is hearing her" is that he is louder, and how much
+// louder depends on where the mic sits relative to the speaker.
+
+import (
+	"log"
+	"os/exec"
+	"strconv"
+	"sync"
+	"time"
+
+	"github.com/kami/maven/internal/audio"
+)
+
+// player plays one reply at a time and can be cut off mid-utterance.
+type player interface {
+	// Play starts playback of a, replacing anything already playing, and
+	// returns immediately.
+	Play(a audio.Audio)
+	// Stop ends playback now. A no-op when nothing is playing.
+	Stop()
+	// Playing reports whether audio is currently going out of the speaker.
+	Playing() bool
+}
+
+// aplayPlayer pipes raw PCM to aplay(1). Stop kills the child, which is what
+// makes barge-in instant rather than "instant at the end of the sentence".
+type aplayPlayer struct {
+	mu      sync.Mutex
+	cmd     *exec.Cmd
+	playing bool
+	// gen rises on every Play/Stop so a finishing playback cannot clear the
+	// playing flag of the one that replaced it.
+	gen uint64
+}
+
+func newAplayPlayer() *aplayPlayer { return &aplayPlayer{} }
+
+func (p *aplayPlayer) Play(a audio.Audio) {
+	if len(a.Bytes) == 0 {
+		return
+	}
+	p.Stop()
+
+	cmd := exec.Command("aplay",
+		"-f", "S16_LE",
+		"-r", strconv.Itoa(a.Format.SampleRate),
+		"-c", strconv.Itoa(a.Format.Channels),
+		"-t", "raw",
+	)
+	stdin, err := cmd.StdinPipe()
+	if err != nil {
+		log.Printf("mavwaked: aplay stdin pipe: %v", err)
+		return
+	}
+	if err := cmd.Start(); err != nil {
+		log.Printf("mavwaked: start aplay: %v", err)
+		_ = stdin.Close()
+		return
+	}
+
+	p.mu.Lock()
+	p.gen++
+	gen := p.gen
+	p.cmd = cmd
+	p.playing = true
+	p.mu.Unlock()
+
+	go func() {
+		if _, err := stdin.Write(a.Bytes); err != nil {
+			// Broken pipe is the expected outcome of Stop().
+			log.Printf("mavwaked: write to aplay: %v", err)
+		}
+		_ = stdin.Close()
+
+		done := make(chan error, 1)
+		go func() { done <- cmd.Wait() }()
+		select {
+		case err := <-done:
+			if err != nil {
+				log.Printf("mavwaked: aplay: %v", err)
+			}
+		case <-time.After(30 * time.Second):
+			log.Printf("mavwaked: aplay timeout, killing")
+			if pr := cmd.Process; pr != nil {
+				_ = pr.Kill()
+			}
+			<-done
+		}
+
+		p.mu.Lock()
+		if p.gen == gen {
+			p.playing = false
+			p.cmd = nil
+		}
+		p.mu.Unlock()
+	}()
+}
+
+func (p *aplayPlayer) Stop() {
+	p.mu.Lock()
+	cmd := p.cmd
+	if cmd != nil {
+		p.gen++
+		p.playing = false
+		p.cmd = nil
+	}
+	p.mu.Unlock()
+	if cmd != nil && cmd.Process != nil {
+		_ = cmd.Process.Kill()
+	}
+}
+
+func (p *aplayPlayer) Playing() bool {
+	p.mu.Lock()
+	defer p.mu.Unlock()
+	return p.playing
+}

cmd/mavwaked/playback_test.go

@@ -0,0 +1,33 @@
+package main
+
+import (
+	"testing"
+
+	"github.com/kami/maven/internal/audio"
+)
+
+// The real player must be safe to poke when nothing is playing — the capture
+// loop calls Playing() on every 30ms frame, and Stop() lands on an idle
+// player whenever a barge-in races the end of a reply. Neither may need
+// aplay(1) to be installed.
+func TestAplayPlayerIdleIsSafe(t *testing.T) {
+	p := newAplayPlayer()
+	if p.Playing() {
+		t.Fatal("a fresh player reports playing")
+	}
+	p.Stop()
+	p.Stop()
+	if p.Playing() {
+		t.Fatal("playing after Stop on an idle player")
+	}
+	// Empty audio is a text-only turn: nothing to play, no process to spawn.
+	p.Play(audio.Audio{Format: audio.PCM16kMono})
+	if p.Playing() {
+		t.Fatal("empty audio started playback")
+	}
+}
+
+func TestAplayPlayerSatisfiesPlayer(t *testing.T) {
+	var _ player = newAplayPlayer()
+	var _ player = &fakePlayer{}
+}

cmd/mavwaked/session.go

@@ -0,0 +1,123 @@
+package main
+
+// The capture session: what happens to one 30ms frame, given whether Maven is
+// currently speaking. Split out of main.go's processFrame so the decision is
+// testable without a mic, a speaker, or a daemon (Vikunja #287).
+
+import (
+	"context"
+	"log"
+
+	"github.com/kami/maven/internal/audio"
+)
+
+// utteranceSender ships one complete utterance to the voice server and
+// returns the reply audio to play. The real one round-trips over the voice
+// wire; tests substitute a recorder.
+type utteranceSender interface {
+	Send(ctx context.Context, utt audio.Audio, lang string) (audio.Audio, error)
+}
+
+// bargeInConfig holds the two numbers barge-in needs. Zero Frames disables
+// barge-in entirely — the half-duplex gate still runs.
+type bargeInConfig struct {
+	// RMS is the normalised energy a frame must exceed to count as him
+	// talking over her rather than the mic hearing her. It is deliberately
+	// far above the VAD's own floor: the speaker leaks into the mic at
+	// roughly ambient level, a person talking at the mic does not.
+	RMS float64
+	// Frames is how many consecutive frames must clear RMS before playback
+	// is cut. One loud frame is a door closing; five in a row is a voice.
+	Frames int
+}
+
+// Enabled reports whether barge-in should be attempted at all.
+func (c bargeInConfig) Enabled() bool { return c.Frames > 0 && c.RMS > 0 }
+
+// session is the per-client capture state machine.
+type session struct {
+	vad    *VAD
+	player player
+	sender utteranceSender
+	lang   string
+	barge  bargeInConfig
+
+	// loudFrames counts consecutive over-threshold frames seen while she is
+	// speaking. Reset whenever a frame falls back under the threshold, and
+	// whenever playback ends.
+	loudFrames int
+
+	// counters, read by tests and logged on the way out.
+	suppressed int // frames dropped because she was speaking
+	bargeIns   int // times playback was cut because he spoke over her
+	sent       int // utterances shipped to the daemon
+}
+
+func newSession(vad *VAD, p player, s utteranceSender, lang string, barge bargeInConfig) *session {
+	return &session{vad: vad, player: p, sender: s, lang: lang, barge: barge}
+}
+
+// feed processes one 30ms PCM frame.
+//
+// While the player is running the capture side is muted: the VAD is not fed
+// and no utterance can be produced, so Maven's own reply cannot come back in
+// as a new command. The one thing that gets through is barge-in — sustained
+// energy well above the speaker's leak level cuts playback, and capture
+// resumes on the very next frame with a clean VAD.
+func (s *session) feed(ctx context.Context, frame []byte) error {
+	if s.player.Playing() {
+		s.suppressed++
+		if !s.barge.Enabled() {
+			return nil
+		}
+		if frameRMS(PCMToI16(frame)) < s.barge.RMS {
+			s.loudFrames = 0
+			return nil
+		}
+		s.loudFrames++
+		if s.loudFrames < s.barge.Frames {
+			return nil
+		}
+		// He is talking over her. Cut her off, drop the VAD state that
+		// accumulated from the echo, and start listening for real.
+		s.player.Stop()
+		s.bargeIns++
+		s.loudFrames = 0
+		s.vad.Reset()
+		log.Printf("mavwaked: barge-in — stopped playback")
+		return nil
+	}
+
+	// Not speaking. If we just stopped, make sure no echo-era state leaks
+	// into the next utterance.
+	if s.loudFrames != 0 {
+		s.loudFrames = 0
+		s.vad.Reset()
+	}
+
+	utt, state := s.vad.Feed(PCMToI16(frame))
+	if state == StateSpeech || utt.Bytes == nil {
+		return nil
+	}
+	return s.dispatch(ctx, utt)
+}
+
+// dispatch ships a complete utterance and plays whatever comes back.
+func (s *session) dispatch(ctx context.Context, utt audio.Audio) error {
+	log.Printf("mavwaked: utterance complete (%.2fs, %d bytes), sending...", utt.Duration(), len(utt.Bytes))
+	reply, err := s.sender.Send(ctx, utt, s.lang)
+	s.sent++
+	if err != nil {
+		return err
+	}
+	if len(reply.Bytes) == 0 {
+		log.Printf("mavwaked: empty reply audio (text only)")
+		return nil
+	}
+	// The VAD has been accumulating from the buffered mic stream while the
+	// round-trip blocked. None of it is a command — reset before the
+	// speaker opens, so the first post-reply frame starts clean.
+	s.vad.Reset()
+	s.player.Play(reply)
+	return nil
+}

cmd/mavwaked/session_test.go

@@ -0,0 +1,282 @@
+package main
+
+import (
+	"context"
+	"errors"
+	"math"
+	"testing"
+
+	"github.com/kami/maven/internal/audio"
+)
+
+// fakePlayer records Play/Stop instead of shelling out to aplay.
+type fakePlayer struct {
+	playing bool
+	plays   int
+	stops   int
+	last    audio.Audio
+}
+
+func (p *fakePlayer) Play(a audio.Audio) { p.playing = true; p.plays++; p.last = a }
+func (p *fakePlayer) Stop()              { p.playing = false; p.stops++ }
+func (p *fakePlayer) Playing() bool      { return p.playing }
+
+// fakeSender records what was shipped and hands back a canned reply.
+type fakeSender struct {
+	sent  []audio.Audio
+	reply audio.Audio
+	err   error
+}
+
+func (s *fakeSender) Send(_ context.Context, utt audio.Audio, _ string) (audio.Audio, error) {
+	s.sent = append(s.sent, utt)
+	return s.reply, s.err
+}
+
+func replyAudio() audio.Audio {
+	return audio.Audio{Format: audio.PCM16kMono, Bytes: make([]byte, 16000)}
+}
+
+// frameAt returns a 30ms frame whose RMS is approximately rms.
+func frameAt(rms float64) []byte {
+	amp := rms * math.Sqrt2 * 32768
+	f := make([]int16, frameSamples)
+	for i := range f {
+		f[i] = int16(amp * math.Sin(2*math.Pi*440*float64(i)/16000))
+	}
+	return pcmBytes(f)
+}
+
+func silentBytes() []byte { return make([]byte, frameSamples*2) }
+
+// newTestSession wires a session with fakes and a default VAD.
+func newTestSession(barge bargeInConfig) (*session, *fakePlayer, *fakeSender) {
+	p := &fakePlayer{}
+	s := &fakeSender{reply: replyAudio()}
+	return newSession(NewVAD(0, 0, 0, 0), p, s, "ru", barge), p, s
+}
+
+// speakThenPause drives a full utterance through the session: enough loud
+// frames to trigger, then enough silence to end it.
+func speakThenPause(t *testing.T, sess *session) {
+	t.Helper()
+	speechFrames := (defaultSpeechMs + defaultFrameMs - 1) / defaultFrameMs
+	silenceFrames := (defaultSilenceMs+defaultFrameMs-1)/defaultFrameMs + 2
+	loud := frameAt(0.35)
+	for i := 0; i < speechFrames+5; i++ {
+		if err := sess.feed(context.Background(), loud); err != nil {
+			t.Fatalf("feed loud frame %d: %v", i, err)
+		}
+	}
+	for i := 0; i < silenceFrames; i++ {
+		if err := sess.feed(context.Background(), silentBytes()); err != nil {
+			t.Fatalf("feed silent frame %d: %v", i, err)
+		}
+	}
+}
+
+func TestSessionSendsUtteranceAndPlaysReply(t *testing.T) {
+	sess, p, snd := newTestSession(bargeInConfig{})
+	speakThenPause(t, sess)
+
+	if len(snd.sent) != 1 {
+		t.Fatalf("sent %d utterances, want 1", len(snd.sent))
+	}
+	if snd.sent[0].Format != audio.PCM16kMono {
+		t.Errorf("utterance format = %+v, want canonical", snd.sent[0].Format)
+	}
+	if p.plays != 1 {
+		t.Errorf("plays = %d, want 1", p.plays)
+	}
+}
+
+// The bug this whole file exists for: while the speaker is running, the mic
+// hears Maven and the old code shipped that back as a fresh command.
+func TestSessionDoesNotHearItselfWhilePlaying(t *testing.T) {
+	sess, p, snd := newTestSession(bargeInConfig{})
+	speakThenPause(t, sess)
+	if !p.Playing() {
+		t.Fatal("expected playback to be running after the reply")
+	}
+
+	// Feed a long stretch of loud audio — Maven's own voice coming back in.
+	base := sess.suppressed
+	loud := frameAt(0.35)
+	for i := 0; i < 200; i++ {
+		if err := sess.feed(context.Background(), loud); err != nil {
+			t.Fatalf("feed echo frame %d: %v", i, err)
+		}
+	}
+
+	if len(snd.sent) != 1 {
+		t.Fatalf("sent %d utterances, want 1 — her own reply was captured as a command", len(snd.sent))
+	}
+	if got := sess.suppressed - base; got != 200 {
+		t.Errorf("suppressed %d of the 200 echo frames, want all of them", got)
+	}
+	if p.stops != 0 {
+		t.Errorf("stops = %d, want 0 — barge-in is off, nothing should cut her off", p.stops)
+	}
+}
+
+// With barge-in off, no amount of noise stops playback.
+func TestSessionBargeInDisabledByDefault(t *testing.T) {
+	sess, p, _ := newTestSession(bargeInConfig{})
+	if sess.barge.Enabled() {
+		t.Fatal("zero bargeInConfig must be disabled")
+	}
+	speakThenPause(t, sess)
+	veryLoud := frameAt(0.6)
+	for i := 0; i < 50; i++ {
+		_ = sess.feed(context.Background(), veryLoud)
+	}
+	if p.stops != 0 || sess.bargeIns != 0 {
+		t.Fatalf("stops = %d, bargeIns = %d, want 0 with barge-in off", p.stops, sess.bargeIns)
+	}
+}
+
+func TestSessionBargeInCutsPlayback(t *testing.T) {
+	barge := bargeInConfig{RMS: 0.12, Frames: 5}
+	sess, p, _ := newTestSession(barge)
+	speakThenPause(t, sess)
+	if !p.Playing() {
+		t.Fatal("expected playback after the reply")
+	}
+
+	// Four loud frames must not be enough — a door closing is not a voice.
+	veryLoud := frameAt(0.35)
+	for i := 0; i < 4; i++ {
+		_ = sess.feed(context.Background(), veryLoud)
+	}
+	if p.stops != 0 {
+		t.Fatalf("playback cut after 4 frames, want it to hold until %d", barge.Frames)
+	}
+
+	// The fifth cuts her off.
+	_ = sess.feed(context.Background(), veryLoud)
+	if p.stops != 1 || sess.bargeIns != 1 {
+		t.Fatalf("stops = %d, bargeIns = %d, want 1 and 1", p.stops, sess.bargeIns)
+	}
+	if p.Playing() {
+		t.Fatal("still playing after barge-in")
+	}
+}
+
+// A burst that falls back under the threshold resets the counter, so noise
+// spread over a whole reply never accumulates into a false barge-in.
+func TestSessionBargeInNeedsConsecutiveFrames(t *testing.T) {
+	sess, p, _ := newTestSession(bargeInConfig{RMS: 0.12, Frames: 5})
+	speakThenPause(t, sess)
+
+	veryLoud := frameAt(0.35)
+	quiet := frameAt(0.02)
+	for i := 0; i < 20; i++ {
+		_ = sess.feed(context.Background(), veryLoud)
+		_ = sess.feed(context.Background(), veryLoud)
+		_ = sess.feed(context.Background(), quiet)
+	}
+	if p.stops != 0 || sess.bargeIns != 0 {
+		t.Fatalf("stops = %d, bargeIns = %d, want 0 — two-frame bursts must not accumulate", p.stops, sess.bargeIns)
+	}
+}
+
+// Speaker leak sits near the room floor; it must never reach the barge-in bar.
+func TestSessionEchoLevelAudioNeverBargesIn(t *testing.T) {
+	sess, p, _ := newTestSession(bargeInConfig{RMS: 0.12, Frames: 5})
+	speakThenPause(t, sess)
+
+	base := sess.suppressed
+	leak := frameAt(0.05) // loud enough for the VAD, far under the barge bar
+	for i := 0; i < 300; i++ {
+		_ = sess.feed(context.Background(), leak)
+	}
+	if p.stops != 0 {
+		t.Fatalf("stops = %d, want 0 — speaker leak must not read as barge-in", p.stops)
+	}
+	if got := sess.suppressed - base; got != 300 {
+		t.Errorf("suppressed %d of the 300 leak frames, want all of them", got)
+	}
+}
+
+// After barge-in the VAD must start clean, so the interrupting speech is
+// captured as a whole utterance rather than joined onto echo state.
+func TestSessionCapturesTheInterruptingUtterance(t *testing.T) {
+	sess, p, snd := newTestSession(bargeInConfig{RMS: 0.12, Frames: 5})
+	speakThenPause(t, sess)
+
+	veryLoud := frameAt(0.35)
+	for i := 0; i < 5; i++ {
+		_ = sess.feed(context.Background(), veryLoud)
+	}
+	if p.stops != 1 {
+		t.Fatalf("expected barge-in, stops = %d", p.stops)
+	}
+
+	// He keeps talking; that is a new command.
+	speakThenPause(t, sess)
+	if len(snd.sent) != 2 {
+		t.Fatalf("sent %d utterances, want 2 — the interruption itself must be heard", len(snd.sent))
+	}
+	if p.plays != 2 {
+		t.Errorf("plays = %d, want 2", p.plays)
+	}
+}
+
+// A failed round-trip must surface as an error and must not start playback.
+func TestSessionSendErrorDoesNotPlay(t *testing.T) {
+	p := &fakePlayer{}
+	snd := &fakeSender{err: errors.New("boom")}
+	sess := newSession(NewVAD(0, 0, 0, 0), p, snd, "ru", bargeInConfig{})
+
+	speechFrames := (defaultSpeechMs + defaultFrameMs - 1) / defaultFrameMs
+	silenceFrames := (defaultSilenceMs+defaultFrameMs-1)/defaultFrameMs + 2
+	loud := frameAt(0.35)
+	var lastErr error
+	for i := 0; i < speechFrames+5; i++ {
+		_ = sess.feed(context.Background(), loud)
+	}
+	for i := 0; i < silenceFrames; i++ {
+		if err := sess.feed(context.Background(), silentBytes()); err != nil {
+			lastErr = err
+		}
+	}
+	if lastErr == nil {
+		t.Fatal("send error was swallowed")
+	}
+	if p.plays != 0 || p.Playing() {
+		t.Fatalf("plays = %d, playing = %v, want no playback on a failed round-trip", p.plays, p.Playing())
+	}
+}
+
+// An empty reply (text-only turn) must leave the capture side open.
+func TestSessionEmptyReplyLeavesCaptureOpen(t *testing.T) {
+	p := &fakePlayer{}
+	snd := &fakeSender{reply: audio.Audio{Format: audio.PCM16kMono}}
+	sess := newSession(NewVAD(0, 0, 0, 0), p, snd, "ru", bargeInConfig{})
+
+	speakThenPause(t, sess)
+	if p.plays != 0 {
+		t.Fatalf("plays = %d, want 0 for an empty reply", p.plays)
+	}
+	speakThenPause(t, sess)
+	if len(snd.sent) != 2 {
+		t.Fatalf("sent %d, want 2 — capture must stay open when there is no audio reply", len(snd.sent))
+	}
+}
+
+func TestBargeInConfigEnabled(t *testing.T) {
+	cases := []struct {
+		c    bargeInConfig
+		want bool
+	}{
+		{bargeInConfig{}, false},
+		{bargeInConfig{RMS: 0.12}, false},
+		{bargeInConfig{Frames: 5}, false},
+		{bargeInConfig{RMS: 0.12, Frames: 5}, true},
+	}
+	for _, tc := range cases {
+		if got := tc.c.Enabled(); got != tc.want {
+			t.Errorf("%+v.Enabled() = %v, want %v", tc.c, got, tc.want)
+		}
+	}
+}

cmd/mavwaked/vad.go

@@ -23,6 +23,15 @@ const (
 	defaultSilenceMs = 800   // silence hold before declaring end-of-utterance
 	defaultMaxMs     = 10000 // cap single utterance at 10s
 	defaultMinRMS    = 0.01  // RMS floor (same as mavsttd)
+
+	// Barge-in thresholds. Only used when -barge-in is passed. The RMS is
+	// x10000 like -min-rms, and sits an order of magnitude above the VAD's
+	// own floor on purpose: with no acoustic echo canceller, a frame only
+	// counts as "he is talking over her" if it is far louder than what the
+	// speaker leaks back into the mic. 5 frames is 150ms — long enough that
+	// a door or a cough does not cut her off mid-sentence.
+	defaultBargeRMS    = 1200 // 0.12 normalised RMS
+	defaultBargeFrames = 5
 )
 
 // frameSamples — samples per 30ms frame at 16kHz.

docs/plans/10-speaker-recognition.md

@@ -1,27 +1,125 @@
 # Plan: Speaker Recognition
 
-**Goal:** Maven can distinguish between different speakers on the voice channel — recognize known voices (the user, family members) and tag facts/notes/transcripts with a speaker identity.
+**Goal:** Maven can tell who is speaking on the voice channel, and tag what she writes with
+who said it.
 
-**Done when:**
-- Speaker embedding extractor (e.g., ECAPA-TDNN or a simple MFCC + GMM) runs on incoming voice PCM before STT
-- Embedding is compared against enrolled speaker profiles (stored as vectors in the `memory_vectors` table alongside semantic memory)
-- Unknown speakers are enrolled on first interaction (prompt: "кто это?")
-- All voice fact/note writes are tagged with `speaker:<id>` in the value/source metadata
-- Speaker identity is available as context to the router, phraser, and replier ("ok, <name>")
+**Status (2026-08-01, Vikunja #255):** the enrolment half is shipped. The recognising half is
+**BLOCKED on a model download** — there is no speaker-embedding model on this box, and one
+was not invented to fill the gap. See "Blocked, and on what" below.
 
-**Scope:**
-- New `internal/speaker/` package — enrollment, recognition, embedding extraction
-- Reuses `internal/store.MemoryStore` for speaker vector storage (same `memory_vectors` table, different `source` prefix)
-- Reuses `internal/audio` for PCM preprocessing
-- Integration point: `cmd/mavend/voice.go:HandlePushToTalk` — speaker ID extracted before STT, passed through context
+## What shipped
 
-**Steps:**
-1. Research speaker embedding approaches — simplest floor: MFCC + cosine similarity via `github.com/mjibson/go-dsp` or a pre-trained ONNX model (SpeechBrain ECAPA)
-2. Create `internal/speaker/recognizer.go` — `Recognizer` interface: `Identify(pcm []float32) (SpeakerID, confidence)`, `Enroll(id, pcm)`
-3. Create `internal/speaker/store.go` — speaker profile CRUD via `store.MemoryStore`: `Insert("speaker:<id>", embedding, meta)`, `Search(embedding, k)`
-4. Create `internal/speaker/enroll.go` — enrollment flow: capture N seconds of audio, extract embedding, prompt for name via TTS + STT round-trip
-5. Wire into `cmd/mavend/voice.go:HandlePushToTalk` — run speaker ID on the PCM before STT; pass speaker ID through `context.Context` to `applyAction`
-6. Tag all voice-written facts/notes with speaker ID — `Source` becomes `tap:voice:speaker:<id>` or metadata field
-7. Add IPC methods `MethodEnrollSpeaker`, `MethodListSpeakers`, `MethodRemoveSpeaker`
-8. Add speaker config block to `voice` in `config.Config` — `{speaker_recognition: true, model_path}`
-9. Test with 2+ recorded voice samples — verify correct identification and rejection of unknown speakers
+| Piece | Where | State |
+|---|---|---|
+| `Recognizer` — identify, list, get, forget | `internal/speaker/recognizer.go` | done; `Identify` answers `ErrDisabled` until a model exists |
+| Enrolment — several samples, averaged, re-normalised | `internal/speaker/enroll.go` | done |
+| Profile shape, id validation, cosine similarity | `internal/speaker/speaker.go` | done |
+| Profile storage as `speaker:<id>` vectors | `internal/memory` `Catalog` + `internal/store/memory.go` | done, no schema migration |
+| Config block, off by default | `internal/config` `SpeakerConfig` | done |
+| `enroll_speaker` / `list_speakers` / `forget_speaker` | `internal/ipc` | done, absent unless configured |
+| Authority rows | `internal/auth/policy.go` | done — enrol step-up, forget write, list read |
+| Daemon wiring + honest startup log | `cmd/mavend/speaker.go` | done |
+| Embedding backend | `newSpeakerEmbedder` | **BLOCKED** — returns nil, seam only |
+| Tagging voice writes with the speaker | `cmd/mavend/voice.go` | not wired; nothing to tag with yet |
+
+## Blocked, and on what
+
+A voiceprint needs a speaker-embedding model. The box was searched: `/mnt/hdd1/llms` holds
+sixteen ggufs across seven families and every one of them is a text model. There is no ECAPA,
+no x-vector, no titanet, no wespeaker, and no `.onnx` under `/mnt/hdd1` at all. There are also
+no enrolment samples, because nothing has ever recorded any.
+
+To unblock, two things are needed and neither can be done from inside the repo:
+
+1. **A model.** SpeechBrain ECAPA-TDNN exported to ONNX (`speechbrain/spkrec-ecapa-voxceleb`,
+   192-dim) is the usual choice and runs on CPU in well under a second for a few seconds of
+   audio. Download it per the recipe in `AGENTS.md`, put it beside the other models so the
+   bind mount picks it up, and point `speaker.model_path` at it.
+2. **An implementation of one function.** `newSpeakerEmbedder` in `cmd/mavend/speaker.go` is
+   the entire seam: give it an ONNX session that turns `audio.Audio` into a `[]float32` and
+   `Identify` starts working. Nothing else changes — not the store, not the protocol, not the
+   authority table, not the handlers. `internal/onnx` already loads the e5 embedder, so the
+   runtime wiring exists to copy.
+3. **Enrolment samples**, three or more per person, recorded deliberately.
+
+### Why there is no fallback
+
+The original plan offered "a simple MFCC + GMM" as the floor. That is refused. MFCC cosine
+distance is a channel and loudness detector as much as a voice detector: it will happily match
+two different people who sit at the same distance from the same microphone, and it drifts when
+the room changes. A general classifier that is sometimes wrong is a nuisance; a **biometric**
+that is confidently wrong writes false claims about named people into his memory, and then
+those claims get recalled as fact. For this capability a bad floor is worse than none, so the
+shipped state is honest absence: `speaker.Disabled`, `ErrDisabled`, and a startup line saying
+so.
+
+## The refusals, and why
+
+- **Unknown speakers are NOT enrolled on first interaction.** The plan's fourth "done when"
+  bullet asked for exactly that, with a TTS "кто это?" prompt. It is refused in
+  `enroll.go`'s doc comment and there is no request shape in the protocol that could express
+  it. Enrolling a voice is taking a biometric of a person; doing it automatically to whoever
+  walks past the microphone does it to guests who are not party to the exchange, and a
+  synthesised question into a room is not consent from whoever happens to answer. Enrolment is
+  an explicit act: an id, a name, and samples recorded for the purpose.
+- **One sample is not enough.** Three separate utterances and nine seconds minimum. A profile
+  built from one sentence encodes that sentence as much as the person, and the threshold then
+  behaves unpredictably against everything else.
+- **An unknown voice stays unknown.** Below threshold, `Identify` returns `ErrUnknown` naming
+  the closest profile in the error text for diagnosis, never as an answer. Guessing who is in
+  the room is how false memories about people get written.
+- **Deletion is one authority rung below enrolment.** Everywhere else in `policy.go` the
+  destructive direction is gated at least as hard as the constructive one. Here that would be
+  backwards: getting rid of a biometric must never be the harder half.
+- **The voiceprint never crosses the socket.** `ListSpeakersResp` carries ids, names, dates
+  and sample counts. The vector stays in core.
+- **Off unless configured.** No `speaker` block ⇒ the three methods answer
+  `ErrUnknownMethod`. There is no wire path on a default box that takes a voiceprint.
+
+## Storage
+
+Profiles live in the existing `memory_vectors` table under the `speaker:` id prefix, as the
+plan intended, so there is no migration. What that needed was a wider interface than
+`memory.Store`: `memory.Catalog` adds `ByPrefix` and `Delete`. `Delete` is the load-bearing
+one — a voiceprint someone asked to be rid of has to actually go, and a search-only store
+cannot do that. `InMemoryStore.Insert` also became an upsert by id, matching what the
+persistent store already did, so re-enrolling replaces a profile instead of stacking a second
+one behind the first.
+
+Profiles do not collide with note or fact vectors: they are only ever read through
+`ByPrefix("speaker:")`, and a note search never returns one because the prefix is not in its
+query path.
+
+## Config
+
+```json
+"speaker": {
+  "enabled": true,
+  "model_path": "/opt/maven/models/spk/ecapa-voxceleb.onnx",
+  "lib_path": "/opt/maven/lib",
+  "threshold": 0.7,
+  "min_seconds": 2.0
+}
+```
+
+`Recognizes()` requires both `enabled` and a `model_path`, so a half-filled block reads as off
+rather than as a capability that fails every turn. With `enabled` and no model the daemon still
+attaches the three methods — profiles can be created, listed and deleted — and logs that
+recognition is blocked.
+
+## Still open
+
+- The embedding backend (above). Everything below waits on it.
+- **Tagging voice writes.** `Profile.Source("tap:voice")` already produces
+  `tap:voice:speaker:kami`, which is the shape step 6 asked for, but nothing calls it yet:
+  with no recogniser there is no id to tag with. When the model lands, the hook is in the
+  voice path before STT.
+- **Speaker as router/phraser context.** Same dependency. Note the persona constraint when it
+  arrives: Maven addresses the owner informally and speaks to him, so "ok, <name>" needs care
+  for anyone who is not him.
+- **An enrolment surface.** The three IPC methods exist; no page drives them. Enrolment is
+  step-up, so it belongs on `/dash` behind a passkey, with a per-profile forget button next to
+  each row — that button is the reason `list_speakers` exists.
+- **A speaker column on the meeting recorder** (#253). Attributing lines in a transcript is
+  the obvious pairing, and it is the place where getting attribution wrong is most damaging,
+  so it waits for a real model too.

internal/auth/auth_test.go

@@ -442,3 +442,33 @@ func TestRequirement_Capture(t *testing.T) {
 		t.Errorf("voice starting a capture = %v; want allowed", err)
 	}
 }
+
+// TestRequirement_Speaker — a voiceprint is a biometric of a named person, so
+// taking one is step-up: a deliberate act from a surface that can carry a
+// passkey gesture, never something the voice path does mid-conversation.
+//
+// Deletion is one rung lower, and that asymmetry is the point. Everywhere else
+// in the table the destructive direction is gated at least as hard as the
+// constructive one; for a biometric that would be backwards, because getting
+// rid of it must never be the harder half.
+func TestRequirement_Speaker(t *testing.T) {
+	if got := Requirement(ipc.MethodEnrollSpeaker); got != AuthStepUp {
+		t.Errorf("EnrollSpeaker authority = %v; want AuthStepUp", got)
+	}
+	if got := Requirement(ipc.MethodForgetSpeaker); got != AuthWrite {
+		t.Errorf("ForgetSpeaker authority = %v; want AuthWrite", got)
+	}
+	if got := Requirement(ipc.MethodListSpeakers); got != AuthRead {
+		t.Errorf("ListSpeakers authority = %v; want AuthRead", got)
+	}
+	// Voice cannot enrol anybody, however the utterance is phrased.
+	voice := Scope{Surface: SurfaceVoice, Module: "voice", SourceScope: []string{"*"}}
+	if err := Can(ipc.MethodEnrollSpeaker, voice, nil); err == nil {
+		t.Error("voice enrolling a speaker was allowed; want refused")
+	}
+	// But it can read the roster, which is what answering "кого ты знаешь?"
+	// needs.
+	if err := Can(ipc.MethodListSpeakers, voice, nil); err != nil {
+		t.Errorf("voice listing speakers = %v; want allowed", err)
+	}
+}

internal/auth/policy.go

@@ -75,6 +75,22 @@ func Requirement(m ipc.Method) Authority {
 		// exist at all unless the operator enabled a capture block, and no
 		// recording can begin without someone saying so.
 		return AuthWrite
+	case ipc.MethodEnrollSpeaker:
+		// Taking a voiceprint (Vikunja #255). AuthStepUp, and unlike recording a
+		// meeting there is no reason to soften it: enrolment is not a thing anyone
+		// does by voice mid-conversation. It is a deliberate sit-down with a
+		// surface that can carry a passkey gesture, and it writes a biometric of a
+		// named person. If the gesture is inconvenient, that is the correct amount
+		// of friction for this particular write.
+		return AuthStepUp
+	case ipc.MethodForgetSpeaker:
+		// Deleting a voiceprint. One rung BELOW enrolment on purpose. Everywhere
+		// else in this table the destructive direction is gated at least as hard
+		// as the constructive one, and here that would be wrong: getting rid of a
+		// biometric must never be the harder half. The worst a caller at this rung
+		// can do is make Maven stop recognising someone, which is the state the
+		// box ships in anyway.
+		return AuthWrite
 	case ipc.MethodWriteFact:
 		return AuthWrite
 	case ipc.MethodAssertStepUp:
@@ -113,6 +129,10 @@ func Requirement(m ipc.Method) Authority {
 		// "что ты записываешь?" — the read side of the recorder. It reports a
 		// label, a start time and a byte count, begins nothing and keeps nothing.
 		ipc.MethodCaptureStatus,
+		// Who is enrolled. Returns ids, names and enrolment dates — never the
+		// voiceprints themselves, which stay in core. Listing the people Maven can
+		// recognise is exactly the read a surface needs to offer a "forget" button.
+		ipc.MethodListSpeakers,
 		// The read side of the model swap: which model is resident, which ones are
 		// allowlisted. It loads nothing and changes nothing.
 		ipc.MethodModelStatus:

internal/config/config.go

@@ -212,6 +212,11 @@ type Config struct {
 	// default. See CaptureConfig.
 	Capture *CaptureConfig `json:"capture,omitempty"`
 
+	// Speaker — voice identification (Vikunja #255). nil / absent ⇒ no
+	// voiceprint is ever computed and nobody can be enrolled. Enabling it needs
+	// a speaker-embedding model, which is not on this box. See SpeakerConfig.
+	Speaker *SpeakerConfig `json:"speaker,omitempty"`
+
 	// MCP — Model Context Protocol servers Maven connects OUT to (Vikunja
 	// #251). nil / absent / no enabled server ⇒ no connection is made and no
 	// tool is discovered, like every other capability that reaches outside the
@@ -619,6 +624,45 @@ func (c *CaptureConfig) MaxDuration() time.Duration {
 	return time.Duration(c.MaxMinutes) * time.Minute
 }
 
+// SpeakerConfig — voice identification (internal/speaker,
+// docs/plans/10-speaker-recognition.md).
+//
+// Absent, or enabled=false, ⇒ no voiceprint is computed for any turn, the
+// enrolment methods do not exist, and nobody can be enrolled. A voiceprint is
+// biometric data about a person, so this one is off until someone typed a model
+// path on purpose.
+//
+// It cannot currently be turned on: there is no speaker-embedding model on this
+// box. See the plan document for what to download.
+type SpeakerConfig struct {
+	// Enabled — may she work out who is speaking. Default false.
+	Enabled bool `json:"enabled,omitempty"`
+
+	// ModelPath — an ECAPA-TDNN (or equivalent) speaker-embedding ONNX model.
+	// Required; without it the recognizer runs disabled and says so once.
+	ModelPath string `json:"model_path,omitempty"`
+
+	// LibPath — onnxruntime shared library, as for the text embedder. Empty ⇒
+	// the same default the embedder block uses.
+	LibPath string `json:"lib_path,omitempty"`
+
+	// Threshold — cosine similarity a match must beat. 0 ⇒
+	// speaker.DefaultThreshold (0.7). Lower it and she starts calling guests by
+	// his name, which is the expensive direction of this error.
+	Threshold float64 `json:"threshold,omitempty"`
+
+	// MinSeconds — least speech an identification will look at. 0 ⇒
+	// speaker.DefaultMinSeconds (2s).
+	MinSeconds float64 `json:"min_seconds,omitempty"`
+}
+
+// Recognizes reports whether voice identification should be wired. Safe on a
+// nil receiver, and false without a model path — enabled with nothing to embed
+// with is a misconfiguration, not a capability.
+func (s *SpeakerConfig) Recognizes() bool {
+	return s != nil && s.Enabled && strings.TrimSpace(s.ModelPath) != ""
+}
+
 // WeatherConfig configures the weather provider for voice queries.
 type WeatherConfig struct {
 	Provider        string `json:"provider,omitempty"`         // "open-meteo" or "" → stub

internal/config/senses_test.go

@@ -22,6 +22,9 @@ func TestSensesOffByDefault(t *testing.T) {
 	if cfg.Capture.MaxDuration() != 0 {
 		t.Error("a nil capture block invented a duration")
 	}
+	if cfg.Speaker.Recognizes() {
+		t.Error("speaker recognition is on with no speaker block")
+	}
 }
 
 // The recorder is the capability that most needs its default to be off, so it
@@ -158,3 +161,63 @@ func TestMediaWithoutVisionIsValid(t *testing.T) {
 		t.Error("vision came on by itself")
 	}
 }
+
+// A voiceprint is a biometric of a named person. Nothing about it turns on by
+// itself: no speaker block means no recognition, and no enrolment either.
+func TestSpeakerIsOffUntilExplicitlyEnabled(t *testing.T) {
+	var cfg Config
+	if err := json.Unmarshal([]byte(`{}`), &cfg); err != nil {
+		t.Fatal(err)
+	}
+	if cfg.Speaker.Recognizes() {
+		t.Error("speaker recognition came on with no config at all")
+	}
+	var empty Config
+	if err := json.Unmarshal([]byte(`{"speaker":{}}`), &empty); err != nil {
+		t.Fatal(err)
+	}
+	if empty.Speaker.Recognizes() {
+		t.Error("an empty speaker block enabled recognition")
+	}
+}
+
+// Enabled alone is not enough: recognition needs a model, and on this box there
+// is none. Recognizes() must stay false so the daemon reports the honest state
+// instead of claiming a capability it cannot perform.
+func TestSpeakerNeedsBothEnabledAndAModel(t *testing.T) {
+	var cfg Config
+	if err := json.Unmarshal([]byte(`{"speaker":{"enabled":true}}`), &cfg); err != nil {
+		t.Fatal(err)
+	}
+	if cfg.Speaker.Recognizes() {
+		t.Error("enabled with no model_path claimed to recognise")
+	}
+	var only Config
+	if err := json.Unmarshal([]byte(`{"speaker":{"model_path":"/opt/x.onnx"}}`), &only); err != nil {
+		t.Fatal(err)
+	}
+	if only.Speaker.Recognizes() {
+		t.Error("a model_path alone enabled recognition")
+	}
+}
+
+func TestSpeakerBlockParsesFromJSON(t *testing.T) {
+	const raw = `{"speaker":{"enabled":true,"model_path":"/opt/maven/models/spk/ecapa.onnx",` +
+		`"lib_path":"/opt/maven/lib","threshold":0.62,"min_seconds":1.5}}`
+	var cfg Config
+	if err := json.Unmarshal([]byte(raw), &cfg); err != nil {
+		t.Fatalf("unmarshal: %v", err)
+	}
+	if !cfg.Speaker.Recognizes() {
+		t.Fatal("speaker did not parse as enabled")
+	}
+	if cfg.Speaker.ModelPath != "/opt/maven/models/spk/ecapa.onnx" {
+		t.Errorf("model_path = %q", cfg.Speaker.ModelPath)
+	}
+	if cfg.Speaker.LibPath != "/opt/maven/lib" {
+		t.Errorf("lib_path = %q", cfg.Speaker.LibPath)
+	}
+	if cfg.Speaker.Threshold != 0.62 || cfg.Speaker.MinSeconds != 1.5 {
+		t.Errorf("thresholds = %+v", cfg.Speaker)
+	}
+}

internal/ipc/api.go

@@ -301,6 +301,51 @@ type CaptureStatusResp struct {
 	Bytes   int       `json:"bytes,omitempty"`
 }
 
+// EnrollSpeakerReq — register a voice (Vikunja #255).
+//
+// Samples are separate utterances recorded deliberately for this purpose, not
+// audio harvested from ordinary turns. internal/speaker requires several of
+// them totalling enough seconds, and refuses one long clip: a profile built
+// from a single sentence encodes that sentence as much as the person.
+//
+// There is no "enrol whoever just spoke" request shape, and that omission is
+// the point. Taking a biometric of a guest because they walked past the
+// microphone is not something a wire protocol should make easy.
+type EnrollSpeakerReq struct {
+	ID      string        `json:"id"`
+	Name    string        `json:"name,omitempty"`
+	Samples []audio.Audio `json:"samples"`
+}
+
+// Speaker — one enrolled voice as a surface sees it. The voiceprint itself is
+// never sent: a listing says who is enrolled, it does not hand out the
+// biometric.
+type Speaker struct {
+	ID       string    `json:"id"`
+	Name     string    `json:"name"`
+	Enrolled time.Time `json:"enrolled"`
+	Samples  int       `json:"samples"`
+}
+
+// EnrollSpeakerResp — the profile that was written.
+type EnrollSpeakerResp struct {
+	Speaker Speaker `json:"speaker"`
+}
+
+// ListSpeakersResp — who is enrolled, sorted by id. Enabled is false when no
+// embedding model is wired, which is this box's state: the profiles can be
+// listed and deleted, nothing can be recognised.
+type ListSpeakersResp struct {
+	Speakers []Speaker `json:"speakers"`
+	Enabled  bool      `json:"enabled"`
+}
+
+// ForgetSpeakerReq — delete one voiceprint. This is the request that must
+// always work; a biometric someone asked to be rid of has to actually go.
+type ForgetSpeakerReq struct {
+	ID string `json:"id"`
+}
+
 // SwapModelReq — load another resident model without restarting the daemon
 // (Vikunja #250). ModelPath must be one of the paths in phraser.swap_models;
 // anything else is ErrForbidden, and an unconfigured allowlist makes the whole

internal/ipc/client.go

@@ -515,6 +515,34 @@ func (c *Client) CaptureStatus(ctx context.Context) (CaptureStatusResp, error) {
 	return r, nil
 }
 
+// EnrollSpeaker registers a voice from several deliberately recorded samples
+// (Vikunja #255). ErrUnknownMethod means no speaker block is configured, which
+// is the default: on an unconfigured box there is no way to take a voiceprint.
+func (c *Client) EnrollSpeaker(ctx context.Context, req EnrollSpeakerReq) (EnrollSpeakerResp, error) {
+	var r EnrollSpeakerResp
+	if err := c.call(ctx, MethodEnrollSpeaker, req, &r); err != nil {
+		return EnrollSpeakerResp{}, err
+	}
+	return r, nil
+}
+
+// ListSpeakers reports who is enrolled. The voiceprints themselves stay in
+// core. Enabled is false when profiles exist but no embedding model is wired,
+// so a surface can say "enrolled, not recognising" rather than implying Maven
+// knows who is talking.
+func (c *Client) ListSpeakers(ctx context.Context) (ListSpeakersResp, error) {
+	var r ListSpeakersResp
+	if err := c.call(ctx, MethodListSpeakers, nil, &r); err != nil {
+		return ListSpeakersResp{}, err
+	}
+	return r, nil
+}
+
+// ForgetSpeaker deletes one voiceprint.
+func (c *Client) ForgetSpeaker(ctx context.Context, id string) error {
+	return c.call(ctx, MethodForgetSpeaker, ForgetSpeakerReq{ID: id}, nil)
+}
+
 // SwapModel asks core to load another resident model (Vikunja #250).
 // ErrUnknownMethod means core has no phraser.swap_models allowlist configured;
 // ErrForbidden means the path is not on it, or step-up was not asserted. A

internal/ipc/server.go

@@ -473,6 +473,13 @@ type Server struct {
 	CaptureStopFn   CaptureStopFunc
 	CaptureStatusFn CaptureStatusFunc
 
+	// Speaker* — voice identification (Vikunja #255). Set by the daemon only
+	// when a speaker block is configured; nil ⇒ all three methods answer
+	// ErrUnknownMethod, so on an unconfigured box no wire path enrols a voice.
+	EnrollSpeakerFn EnrollSpeakerFunc
+	ListSpeakersFn  ListSpeakersFunc
+	ForgetSpeakerFn ForgetSpeakerFunc
+
 	// UnlockFn — unwraps the store encryption key from the wrapped blob using
 	// the passkey credential public key, opens the encrypted store, and wires
 	// the rest of the daemon (voice, loop, delivery). Set by the daemon when
@@ -510,6 +517,12 @@ type CaptureAppendFunc func(ctx context.Context, req CaptureAppendReq) (CaptureA
 type CaptureStopFunc func(ctx context.Context, req CaptureStopReq) (CaptureStopResp, error)
 type CaptureStatusFunc func(ctx context.Context) (CaptureStatusResp, error)
 
+// EnrollSpeakerFunc / ListSpeakersFunc / ForgetSpeakerFunc — the core-side
+// halves of voice enrolment.
+type EnrollSpeakerFunc func(ctx context.Context, req EnrollSpeakerReq) (EnrollSpeakerResp, error)
+type ListSpeakersFunc func(ctx context.Context) (ListSpeakersResp, error)
+type ForgetSpeakerFunc func(ctx context.Context, req ForgetSpeakerReq) error
+
 // CheckFunc — the auth hook signature. Wired by the daemon (auth.Gate.Check
 // satisfies this); dispatch calls it once per request after param-unmarshal
 // independence (it gets the raw params, may unmarshal what it needs — ipc
@@ -1024,6 +1037,43 @@ func (s *Server) dispatch(ctx context.Context, req Request) (json.RawMessage, er
 		}
 		return nil, fmt.Errorf("%w: %s", ErrUnknownMethod, req.Method)
 
+	case MethodEnrollSpeaker:
+		if s.EnrollSpeakerFn != nil {
+			var p EnrollSpeakerReq
+			if err := unmarshalParams(req.Params, &p); err != nil {
+				return nil, err
+			}
+			resp, err := s.EnrollSpeakerFn(ctx, p)
+			if err != nil {
+				return nil, err
+			}
+			return marshalResult(resp), nil
+		}
+		return nil, fmt.Errorf("%w: %s", ErrUnknownMethod, req.Method)
+
+	case MethodListSpeakers:
+		if s.ListSpeakersFn != nil {
+			resp, err := s.ListSpeakersFn(ctx)
+			if err != nil {
+				return nil, err
+			}
+			return marshalResult(resp), nil
+		}
+		return nil, fmt.Errorf("%w: %s", ErrUnknownMethod, req.Method)
+
+	case MethodForgetSpeaker:
+		if s.ForgetSpeakerFn != nil {
+			var p ForgetSpeakerReq
+			if err := unmarshalParams(req.Params, &p); err != nil {
+				return nil, err
+			}
+			if err := s.ForgetSpeakerFn(ctx, p); err != nil {
+				return nil, err
+			}
+			return marshalResult(nil), nil
+		}
+		return nil, fmt.Errorf("%w: %s", ErrUnknownMethod, req.Method)
+
 	case MethodModelStatus:
 		if s.ModelStatusFn != nil {
 			resp, err := s.ModelStatusFn(ctx)

internal/ipc/speaker_test.go

@@ -0,0 +1,123 @@
+package ipc
+
+import (
+	"context"
+	"errors"
+	"testing"
+	"time"
+
+	"github.com/kami/maven/internal/audio"
+)
+
+// The default that matters most for a biometric: on a core that was never
+// configured with a speaker block, there is no wire path that takes a
+// voiceprint, and none that lists the ones that might exist.
+func TestSpeaker_OffUnlessConfigured(t *testing.T) {
+	_, _, cli, _ := newServerWithStore(t)
+	ctx := context.Background()
+
+	if _, err := cli.EnrollSpeaker(ctx, EnrollSpeakerReq{ID: "kami"}); !errors.Is(err, ErrUnknownMethod) {
+		t.Errorf("EnrollSpeaker error = %v, want ErrUnknownMethod", err)
+	}
+	if _, err := cli.ListSpeakers(ctx); !errors.Is(err, ErrUnknownMethod) {
+		t.Errorf("ListSpeakers error = %v, want ErrUnknownMethod", err)
+	}
+	if err := cli.ForgetSpeaker(ctx, "kami"); !errors.Is(err, ErrUnknownMethod) {
+		t.Errorf("ForgetSpeaker error = %v, want ErrUnknownMethod", err)
+	}
+}
+
+// Enrolment carries several samples across the boundary byte for byte — a
+// profile averaged over the wrong bytes is a profile of nobody.
+func TestSpeaker_EnrollCrossesTheWire(t *testing.T) {
+	_, srv, cli, _ := newServerWithStore(t)
+	ctx := context.Background()
+
+	enrolled := time.Now().UTC().Truncate(time.Second)
+	var gotID, gotName string
+	var gotSamples [][]byte
+
+	srv.EnrollSpeakerFn = func(_ context.Context, req EnrollSpeakerReq) (EnrollSpeakerResp, error) {
+		gotID, gotName = req.ID, req.Name
+		for _, s := range req.Samples {
+			gotSamples = append(gotSamples, s.Bytes)
+		}
+		return EnrollSpeakerResp{Speaker: Speaker{
+			ID: req.ID, Name: req.Name, Enrolled: enrolled, Samples: len(req.Samples),
+		}}, nil
+	}
+
+	mk := func(b byte, n int) audio.Audio {
+		buf := make([]byte, n)
+		for i := range buf {
+			buf[i] = b
+		}
+		return audio.Audio{Format: audio.PCM16kMono, Bytes: buf}
+	}
+	samples := []audio.Audio{mk(1, 64), mk(2, 96), mk(3, 128)}
+
+	resp, err := cli.EnrollSpeaker(ctx, EnrollSpeakerReq{ID: "kami", Name: "Ками", Samples: samples})
+	if err != nil {
+		t.Fatalf("EnrollSpeaker: %v", err)
+	}
+	if gotID != "kami" || gotName != "Ками" {
+		t.Errorf("server saw id=%q name=%q", gotID, gotName)
+	}
+	if len(gotSamples) != 3 {
+		t.Fatalf("server saw %d samples, want 3", len(gotSamples))
+	}
+	for i, want := range samples {
+		if string(gotSamples[i]) != string(want.Bytes) {
+			t.Errorf("sample %d altered in transit", i)
+		}
+	}
+	if resp.Speaker.Samples != 3 || !resp.Speaker.Enrolled.Equal(enrolled) {
+		t.Errorf("profile came back wrong: %+v", resp.Speaker)
+	}
+}
+
+// A listing says who is enrolled and whether recognition actually works. On
+// this box the honest answer is "enrolled, not recognising", and the response
+// has to be able to say so — otherwise a surface implies Maven knows who is
+// talking when nothing on disk can tell.
+func TestSpeaker_ListReportsDisabledRecognition(t *testing.T) {
+	_, srv, cli, _ := newServerWithStore(t)
+	ctx := context.Background()
+
+	srv.ListSpeakersFn = func(context.Context) (ListSpeakersResp, error) {
+		return ListSpeakersResp{
+			Speakers: []Speaker{{ID: "kami", Name: "Ками", Samples: 3}},
+			Enabled:  false,
+		}, nil
+	}
+
+	resp, err := cli.ListSpeakers(ctx)
+	if err != nil {
+		t.Fatalf("ListSpeakers: %v", err)
+	}
+	if len(resp.Speakers) != 1 || resp.Speakers[0].ID != "kami" {
+		t.Fatalf("speakers = %+v", resp.Speakers)
+	}
+	if resp.Enabled {
+		t.Error("Enabled = true; the seam must be able to report that nothing recognises")
+	}
+}
+
+// Deletion reaches core with the id intact and reports success. This is the
+// request that must always work.
+func TestSpeaker_ForgetReachesCore(t *testing.T) {
+	_, srv, cli, _ := newServerWithStore(t)
+	ctx := context.Background()
+
+	var forgot string
+	srv.ForgetSpeakerFn = func(_ context.Context, req ForgetSpeakerReq) error {
+		forgot = req.ID
+		return nil
+	}
+	if err := cli.ForgetSpeaker(ctx, "гость"); err != nil {
+		t.Fatalf("ForgetSpeaker: %v", err)
+	}
+	if forgot != "гость" {
+		t.Errorf("core forgot %q, want %q", forgot, "гость")
+	}
+}

internal/ipc/wire.go

@@ -59,6 +59,9 @@ const (
 	MethodCaptureAppend          Method = "capture_append"
 	MethodCaptureStop            Method = "capture_stop"
 	MethodCaptureStatus          Method = "capture_status"
+	MethodEnrollSpeaker          Method = "enroll_speaker"
+	MethodListSpeakers           Method = "list_speakers"
+	MethodForgetSpeaker          Method = "forget_speaker"
 )
 
 // Request — one frame from module to core. Params is the JSON-encoded argument

internal/memory/store.go

@@ -3,6 +3,7 @@ package memory
 import (
 	"context"
 	"sort"
+	"strings"
 	"sync"
 )
 
@@ -19,6 +20,33 @@ type Store interface {
 	Search(ctx context.Context, vec []float32, topK int) ([]Result, error)
 }
 
+// Record is a stored vector read back whole — id, vector and metadata — as
+// opposed to Result, which is a search hit and carries a score instead of the
+// vector.
+type Record struct {
+	ID   string
+	Vec  []float32
+	Meta map[string]string
+}
+
+// Catalog is a Store that can also be enumerated by id prefix and deleted from.
+//
+// Search is not enough for every user of the vector table. Speaker profiles
+// (internal/speaker) need to list exactly their own rows without scoring
+// anything, because listing enrolled voices is not a similarity question, and
+// they need Delete because a voiceprint is data about a person and "forget this
+// voice" has to actually remove it. Note and fact recall use plain Store and are
+// unaffected.
+type Catalog interface {
+	Store
+	// ByPrefix returns every row whose id starts with prefix, in no particular
+	// order. An empty prefix returns everything.
+	ByPrefix(ctx context.Context, prefix string) ([]Record, error)
+	// Delete removes one row by id. Deleting a row that is not there is not an
+	// error: the caller asked for it to be gone and it is gone.
+	Delete(ctx context.Context, id string) error
+}
+
 // item is a single stored vector with metadata.
 type item struct {
 	id   string
@@ -32,14 +60,54 @@ type InMemoryStore struct {
 	items []item
 }
 
+// compile-time check: InMemoryStore satisfies Catalog.
+var _ Catalog = (*InMemoryStore)(nil)
+
 func NewInMemoryStore() *InMemoryStore {
 	return &InMemoryStore{}
 }
 
+// Insert upserts by id, matching the persistent store.MemoryStore: a repeated
+// id replaces the prior row rather than accumulating a second copy. Re-indexing
+// a note is an update, and re-enrolling a voice must replace the old voiceprint
+// rather than leave it searchable.
 func (s *InMemoryStore) Insert(_ context.Context, id string, vec []float32, meta map[string]string) error {
 	s.mu.Lock()
+	defer s.mu.Unlock()
+	for i := range s.items {
+		if s.items[i].id == id {
+			s.items[i] = item{id: id, vec: vec, meta: meta}
+			return nil
+		}
+	}
 	s.items = append(s.items, item{id: id, vec: vec, meta: meta})
-	s.mu.Unlock()
+	return nil
+}
+
+// ByPrefix implements Catalog.
+func (s *InMemoryStore) ByPrefix(_ context.Context, prefix string) ([]Record, error) {
+	s.mu.RLock()
+	defer s.mu.RUnlock()
+	var out []Record
+	for _, it := range s.items {
+		if !strings.HasPrefix(it.id, prefix) {
+			continue
+		}
+		out = append(out, Record{ID: it.id, Vec: append([]float32(nil), it.vec...), Meta: it.meta})
+	}
+	return out, nil
+}
+
+// Delete implements Catalog.
+func (s *InMemoryStore) Delete(_ context.Context, id string) error {
+	s.mu.Lock()
+	defer s.mu.Unlock()
+	for i := range s.items {
+		if s.items[i].id == id {
+			s.items = append(s.items[:i], s.items[i+1:]...)
+			return nil
+		}
+	}
 	return nil
 }
 

internal/memory/store_test.go

@@ -2,6 +2,7 @@ package memory
 
 import (
 	"context"
+	"fmt"
 	"math"
 	"testing"
 )
@@ -40,8 +41,9 @@ func TestTopKTruncation(t *testing.T) {
 	s := NewInMemoryStore()
 	ctx := context.Background()
 
+	// Distinct ids: Insert upserts by id, so ten rows need ten ids.
 	for i := 0; i < 10; i++ {
-		s.Insert(ctx, "", []float32{float32(i) / 10, 0, 0}, nil)
+		s.Insert(ctx, fmt.Sprintf("n%d", i), []float32{float32(i) / 10, 0, 0}, nil)
 	}
 
 	results, err := s.Search(ctx, []float32{1, 0, 0}, 3)

internal/speaker/enroll.go

@@ -0,0 +1,109 @@
+package speaker
+
+import (
+	"context"
+	"fmt"
+	"strconv"
+	"strings"
+	"time"
+
+	"github.com/kami/maven/internal/audio"
+)
+
+// Enroll registers a voice under an id and a spoken name.
+//
+// Several separate samples are required (MinEnrollSamples, MinEnrollSeconds
+// total): a profile built from one sentence encodes that sentence as much as the
+// person, and the resulting threshold behaviour is unpredictable. The samples
+// are embedded individually and the voiceprints averaged, then re-normalised.
+//
+// Re-enrolling an existing id REPLACES the profile. That is the intended way to
+// improve a weak one, and it is why the store upserts by id.
+//
+// # The refused step
+//
+// The plan document's fourth bullet reads "unknown speakers are enrolled on
+// first interaction (prompt: 'кто это?')". That is refused. Enrolling a voice is
+// taking a biometric of a person; doing it automatically the first time someone
+// walks past the microphone is doing it to guests, without them being part of
+// the exchange, and a TTS question into a room is not consent from whoever
+// happens to answer. Enrolment here is an explicit act: an id, a name, and
+// samples deliberately recorded for the purpose. An unknown voice stays unknown,
+// which the rest of the system is built to cope with.
+func (r *Recognizer) Enroll(ctx context.Context, id, name string, samples []audio.Audio) (Profile, error) {
+	id = NormalizeID(id)
+	if !ValidID(id) {
+		return Profile{}, fmt.Errorf("%w: %q", ErrBadID, id)
+	}
+	name = strings.TrimSpace(name)
+	if name == "" {
+		name = id
+	}
+	if len(samples) < MinEnrollSamples {
+		return Profile{}, fmt.Errorf("%w: %d sample(s), need %d separate ones",
+			ErrTooShort, len(samples), MinEnrollSamples)
+	}
+
+	var total float64
+	for i, s := range samples {
+		if !s.Format.IsValid() {
+			return Profile{}, fmt.Errorf("%w: sample %d: %+v", ErrBadFormat, i+1, s.Format)
+		}
+		total += seconds(s)
+	}
+	if total < MinEnrollSeconds {
+		return Profile{}, fmt.Errorf("%w: %.1fs total, need %.1fs",
+			ErrTooShort, total, MinEnrollSeconds)
+	}
+
+	// Embed first, store second. A model failure halfway through must not leave
+	// a half-built profile that would then be matched against.
+	var (
+		sum []float32
+		dim int
+	)
+	for i, s := range samples {
+		vec, err := r.embed(ctx, s)
+		if err != nil {
+			return Profile{}, fmt.Errorf("speaker: enroll %q sample %d: %w", id, i+1, err)
+		}
+		if sum == nil {
+			sum = make([]float32, len(vec))
+			dim = len(vec)
+		} else if len(vec) != dim {
+			// One model, one width. A mixed-width average would be nonsense.
+			return Profile{}, fmt.Errorf("%w: sample %d is %d wide, expected %d",
+				ErrBadVector, i+1, len(vec), dim)
+		}
+		for j, f := range vec {
+			sum[j] += f
+		}
+	}
+	mean, err := Normalize(sum)
+	if err != nil {
+		// Samples that cancel each other out to zero are not one voice.
+		return Profile{}, fmt.Errorf("speaker: enroll %q: %w", id, err)
+	}
+
+	p := Profile{
+		ID:       id,
+		Name:     name,
+		Enrolled: r.now().UTC(),
+		Samples:  len(samples),
+		Dim:      dim,
+		Vec:      mean,
+	}
+	meta := map[string]string{
+		"name":     p.Name,
+		"samples":  strconv.Itoa(p.Samples),
+		"enrolled": p.Enrolled.Format(time.RFC3339),
+		// kind marks the row for anything walking the vector table, so a future
+		// export or debug page can tell a voiceprint from a note embedding
+		// without parsing the id.
+		"kind": "speaker",
+	}
+	if err := r.cat.Insert(ctx, Prefix+id, mean, meta); err != nil {
+		return Profile{}, fmt.Errorf("speaker: enroll %q: %w", id, err)
+	}
+	return p, nil
+}

internal/speaker/recognizer.go

@@ -0,0 +1,212 @@
+package speaker
+
+import (
+	"context"
+	"fmt"
+	"sort"
+	"time"
+
+	"github.com/kami/maven/internal/audio"
+	"github.com/kami/maven/internal/memory"
+)
+
+// Recognizer holds the embedder and the enrolled profiles.
+//
+// The profiles live in the shared vector table under the "speaker:" id prefix,
+// which is what the plan asked for and what keeps them inside the encrypted
+// store rather than in a sidecar file. They are read through memory.Catalog
+// (ByPrefix / Delete) rather than Search, because "who is enrolled" is not a
+// similarity question and note recall must never rank a voiceprint.
+type Recognizer struct {
+	emb       Embedder
+	cat       memory.Catalog
+	threshold float64
+	minSec    float64
+	now       func() time.Time
+}
+
+// Config — the recognizer's knobs, built from config.SpeakerConfig.
+type Config struct {
+	// Threshold — cosine similarity a match must beat. 0 ⇒ DefaultThreshold.
+	Threshold float64
+	// MinSeconds — least speech an identification will look at. 0 ⇒
+	// DefaultMinSeconds.
+	MinSeconds float64
+}
+
+// New builds a Recognizer. emb nil ⇒ Disabled, which is this box's state and
+// makes every Identify answer ErrDisabled while enrolment and listing still
+// behave sensibly (they refuse for the same reason, with the same error).
+func New(emb Embedder, cat memory.Catalog, cfg Config) (*Recognizer, error) {
+	if cat == nil {
+		return nil, fmt.Errorf("speaker: no profile store")
+	}
+	if emb == nil {
+		emb = Disabled{}
+	}
+	th := cfg.Threshold
+	if th <= 0 {
+		th = DefaultThreshold
+	}
+	min := cfg.MinSeconds
+	if min <= 0 {
+		min = DefaultMinSeconds
+	}
+	return &Recognizer{emb: emb, cat: cat, threshold: th, minSec: min, now: time.Now}, nil
+}
+
+// Enabled reports whether an embedding model is actually wired. Surfaces use it
+// to say "recognition is off" once instead of failing every turn.
+func (r *Recognizer) Enabled() bool {
+	_, disabled := r.emb.(Disabled)
+	return !disabled
+}
+
+// Threshold is the configured match floor, for a status line.
+func (r *Recognizer) Threshold() float64 { return r.threshold }
+
+// Identify names the voice in a. ErrUnknown when nothing is close enough, which
+// is a normal answer and not a failure: a guest is a guest, and the caller
+// carries on with no speaker attached rather than guessing.
+//
+// Identification never decides whether Maven listens. It annotates the turn.
+func (r *Recognizer) Identify(ctx context.Context, a audio.Audio) (Match, error) {
+	if !a.Format.IsValid() {
+		return Match{}, fmt.Errorf("%w: %+v", ErrBadFormat, a.Format)
+	}
+	if seconds(a) < r.minSec {
+		return Match{}, fmt.Errorf("%w: %.1fs, need %.1fs", ErrTooShort, seconds(a), r.minSec)
+	}
+	vec, err := r.embed(ctx, a)
+	if err != nil {
+		return Match{}, err
+	}
+	profiles, err := r.List(ctx)
+	if err != nil {
+		return Match{}, err
+	}
+	if len(profiles) == 0 {
+		return Match{}, ErrNoProfiles
+	}
+
+	best := Match{Score: -2}
+	for _, p := range profiles {
+		if s := Similarity(vec, p.Vec); s > best.Score {
+			best = Match{Profile: p, Score: s}
+		}
+	}
+	if best.Score < r.threshold {
+		// The closest profile is reported in the error for a log line, because
+		// "не узнала, ближе всего Ками на 0.61" is what makes a threshold
+		// tunable. The caller must not use it as an identification.
+		return Match{}, fmt.Errorf("%w (closest %s at %.2f, need %.2f)",
+			ErrUnknown, best.Profile.ID, best.Score, r.threshold)
+	}
+	return best, nil
+}
+
+// List returns every enrolled profile, sorted by id so a listing is stable.
+func (r *Recognizer) List(ctx context.Context) ([]Profile, error) {
+	recs, err := r.cat.ByPrefix(ctx, Prefix)
+	if err != nil {
+		return nil, fmt.Errorf("speaker: list: %w", err)
+	}
+	out := make([]Profile, 0, len(recs))
+	for _, rec := range recs {
+		out = append(out, profileFromRecord(rec))
+	}
+	sort.Slice(out, func(i, j int) bool { return out[i].ID < out[j].ID })
+	return out, nil
+}
+
+// Get returns one profile by id.
+func (r *Recognizer) Get(ctx context.Context, id string) (Profile, error) {
+	id = NormalizeID(id)
+	if !ValidID(id) {
+		return Profile{}, fmt.Errorf("%w: %q", ErrBadID, id)
+	}
+	recs, err := r.cat.ByPrefix(ctx, Prefix+id)
+	if err != nil {
+		return Profile{}, fmt.Errorf("speaker: get: %w", err)
+	}
+	for _, rec := range recs {
+		if rec.ID == Prefix+id {
+			return profileFromRecord(rec), nil
+		}
+	}
+	return Profile{}, fmt.Errorf("%w: %q", ErrNotFound, id)
+}
+
+// Forget deletes a profile. This is the one operation that must always work:
+// a voiceprint is data about a person, and "перестань узнавать её" has to
+// actually remove it, not mark it inactive.
+func (r *Recognizer) Forget(ctx context.Context, id string) error {
+	id = NormalizeID(id)
+	if !ValidID(id) {
+		return fmt.Errorf("%w: %q", ErrBadID, id)
+	}
+	if _, err := r.Get(ctx, id); err != nil {
+		return err
+	}
+	if err := r.cat.Delete(ctx, Prefix+id); err != nil {
+		return fmt.Errorf("speaker: forget %q: %w", id, err)
+	}
+	return nil
+}
+
+// embed runs the model and normalises the result.
+func (r *Recognizer) embed(ctx context.Context, a audio.Audio) ([]float32, error) {
+	raw, err := r.emb.Embed(ctx, a)
+	if err != nil {
+		return nil, err
+	}
+	vec, err := Normalize(raw)
+	if err != nil {
+		return nil, err
+	}
+	return vec, nil
+}
+
+// profileFromRecord reads a stored row back into a Profile. A row with
+// unreadable metadata still yields a usable voiceprint — the vector is the part
+// that matters, and losing a name should not lose the enrolment.
+func profileFromRecord(rec memory.Record) Profile {
+	p := Profile{
+		ID:   trimPrefix(rec.ID),
+		Vec:  rec.Vec,
+		Dim:  len(rec.Vec),
+		Name: rec.Meta["name"],
+	}
+	if s := rec.Meta["samples"]; s != "" {
+		p.Samples = atoi(s)
+	}
+	if ts := rec.Meta["enrolled"]; ts != "" {
+		if t, err := time.Parse(time.RFC3339, ts); err == nil {
+			p.Enrolled = t
+		}
+	}
+	if p.Name == "" {
+		p.Name = p.ID
+	}
+	return p
+}
+
+func trimPrefix(id string) string {
+	if len(id) > len(Prefix) && id[:len(Prefix)] == Prefix {
+		return id[len(Prefix):]
+	}
+	return id
+}
+
+// atoi is a tolerant small-integer parse: metadata that is not a number reads
+// as 0 rather than failing the whole listing.
+func atoi(s string) int {
+	n := 0
+	for _, r := range s {
+		if r < '0' || r > '9' {
+			return 0
+		}
+		n = n*10 + int(r-'0')
+	}
+	return n
+}

internal/speaker/speaker.go

@@ -0,0 +1,223 @@
+// Package speaker is voice identification (Vikunja #255,
+// docs/plans/10-speaker-recognition.md).
+//
+// The shape is the same seam internal/vision uses: an Embedder turns audio into
+// a voiceprint, a Recognizer compares one against the enrolled profiles, and a
+// Disabled floor refuses politely when nothing is wired. On this box nothing is
+// wired, and that is the honest state — see "Blocked" below.
+//
+// # A voiceprint is not like the other vectors
+//
+// Everything else in the vector table is something he wrote or said. A speaker
+// profile is biometric data about a person, quite possibly a person who never
+// asked for Maven to exist. The rules that follow from that are in the code:
+//
+//   - Enrolment is explicit and named. There is no "enrol the unknown voice
+//     automatically" path; see the refusal in enroll.go.
+//   - A profile is deletable, individually, and Forget really removes the row.
+//   - Below the threshold the answer is "I do not know", never the closest
+//     guess. A misattributed fact is worse than an unattributed one.
+//   - Nothing here gates whether Maven listens or answers. Identification
+//     annotates a turn; it never authorises one, and an unrecognised voice is
+//     not turned away.
+//   - Voiceprints never leave the box. They live in the encrypted store with
+//     everything else and are never search input to anything external.
+//
+// # Blocked
+//
+// There is no speaker-embedding model on this box: no ECAPA-TDNN, no x-vector,
+// no wespeaker or titanet ONNX anywhere under /mnt/hdd1 or models/ (checked
+// 2026-08-01; the only ONNX files are the e5 text embedder and the piper voice).
+// There are also no enrolment samples. So Recognizer runs against Disabled and
+// every Identify answers ErrDisabled until a model lands.
+//
+// The MFCC + GMM "simplest floor" in the plan document is refused rather than
+// deferred. A hand-rolled spectral distance would identify people confidently
+// and wrongly, and its output would be written into facts as "Ками said this".
+// For a biometric, a bad floor is worse than none: no answer is honest, and a
+// wrong answer is a false memory about a person.
+package speaker
+
+import (
+	"context"
+	"errors"
+	"math"
+	"strings"
+	"time"
+
+	"github.com/kami/maven/internal/audio"
+)
+
+// Prefix — the id prefix speaker profiles carry in the shared vector table.
+// It is what ByPrefix enumerates and what keeps voiceprints out of note recall.
+const Prefix = "speaker:"
+
+// DefaultThreshold — cosine similarity a match must beat to be a match.
+//
+// 0.7 is the usual operating point for ECAPA-style embeddings on clean speech
+// and it is deliberately on the strict side here. The two error directions are
+// not symmetric: refusing to name a voice costs a "не узнала", while naming the
+// wrong person writes his wife's remark into a fact attributed to him.
+const DefaultThreshold = 0.7
+
+// DefaultMinSeconds — how much speech an identification needs. Under about two
+// seconds a voiceprint is mostly noise and the similarity score is not worth
+// reading.
+const DefaultMinSeconds = 2.0
+
+// MinEnrollSamples / MinEnrollSeconds — what enrolment requires. Several
+// separate utterances, not one long one: a profile built from a single sentence
+// encodes that sentence's prosody as much as the voice.
+const (
+	MinEnrollSamples = 3
+	MinEnrollSeconds = 9.0
+)
+
+// Errors callers distinguish.
+var (
+	// ErrDisabled — no embedding model is wired. The state of this box.
+	ErrDisabled = errors.New("speaker: recognition is not configured")
+	// ErrTooShort — not enough speech to say anything about.
+	ErrTooShort = errors.New("speaker: not enough audio")
+	// ErrUnknown — audio embedded fine, but no enrolled profile is close
+	// enough. Not an error in the sense of something being broken: it is the
+	// correct answer for a guest, and the caller should carry on without a
+	// speaker rather than treat the turn as failed.
+	ErrUnknown = errors.New("speaker: voice not recognised")
+	// ErrNoProfiles — nobody is enrolled yet.
+	ErrNoProfiles = errors.New("speaker: nobody is enrolled")
+	// ErrNotFound — no profile with that id.
+	ErrNotFound = errors.New("speaker: no such profile")
+	// ErrBadID — an id that is empty or carries characters an id should not.
+	ErrBadID = errors.New("speaker: invalid profile id")
+	// ErrBadFormat — audio that is not the canonical 16 kHz mono PCM shape.
+	ErrBadFormat = errors.New("speaker: audio format not supported")
+	// ErrBadVector — an embedder returned something unusable (empty, or all
+	// zeroes, which normalises to nothing and would match everything equally).
+	ErrBadVector = errors.New("speaker: embedder returned an unusable vector")
+)
+
+// Embedder turns speech into a voiceprint. Implementations are expected to
+// return an L2-normalised vector, because the whole store compares by dot
+// product; Normalize is applied anyway rather than trusted.
+//
+// This is the seam a downloaded ECAPA-TDNN ONNX model plugs into. It is an
+// interface rather than a concrete ONNX type so the package is testable with no
+// model on disk, which is the only way it could be tested here at all.
+type Embedder interface {
+	Embed(ctx context.Context, a audio.Audio) ([]float32, error)
+	// Dim is the vector width, used to reject a profile recorded with a
+	// different model rather than silently scoring it as zero.
+	Dim() int
+}
+
+// Disabled is the floor: no model, no answers, no guesses.
+type Disabled struct{}
+
+// Embed always fails with ErrDisabled.
+func (Disabled) Embed(context.Context, audio.Audio) ([]float32, error) { return nil, ErrDisabled }
+
+// Dim is 0 for the disabled embedder.
+func (Disabled) Dim() int { return 0 }
+
+// Profile — one enrolled voice.
+//
+// Name is what she calls the person out loud ("Ками"). ID is the stable handle
+// used in sources and metadata. Samples records how many utterances the
+// voiceprint was averaged from, so a profile enrolled from the bare minimum is
+// visibly weaker than one built from ten.
+type Profile struct {
+	ID       string    `json:"id"`
+	Name     string    `json:"name"`
+	Enrolled time.Time `json:"enrolled"`
+	Samples  int       `json:"samples"`
+	Dim      int       `json:"dim"`
+
+	// Vec is the voiceprint. Not serialised to any surface: a listing tells him
+	// who is enrolled, it does not hand out the biometric itself.
+	Vec []float32 `json:"-"`
+}
+
+// Source is what a fact or note written during this speaker's turn is tagged
+// with, e.g. "tap:voice:speaker:kami". Attribution belongs in the source rather
+// than in the text, so it can be corrected or dropped later without rewriting
+// what was said.
+func (p Profile) Source(base string) string {
+	if p.ID == "" {
+		return base
+	}
+	return base + ":" + Prefix + p.ID
+}
+
+// Match — an identification result. Score is cosine similarity in [-1, 1].
+type Match struct {
+	Profile Profile
+	Score   float64
+}
+
+// ValidID reports whether an id is usable as a profile handle. Deliberately
+// narrow: lowercase letters, digits, dash and underscore. Ids end up in note
+// sources and in vector-table keys, so a permissive id would be a way to write
+// into a neighbouring key space.
+func ValidID(id string) bool {
+	if id == "" || len(id) > 64 {
+		return false
+	}
+	for _, r := range id {
+		switch {
+		case r >= 'a' && r <= 'z', r >= '0' && r <= '9', r == '-', r == '_':
+		default:
+			return false
+		}
+	}
+	return true
+}
+
+// NormalizeID lowercases and trims a proposed id before validating it, so
+// "Ками" typed as "Kami " does not fail for a reason nobody can see.
+func NormalizeID(id string) string {
+	return strings.ToLower(strings.TrimSpace(id))
+}
+
+// Normalize returns an L2-normalised copy of v, or ErrBadVector when there is
+// nothing to normalise. A zero vector is refused rather than passed on: it
+// scores 0 against everything, which reads as "no match" but for the wrong
+// reason and would hide a broken embedder.
+func Normalize(v []float32) ([]float32, error) {
+	if len(v) == 0 {
+		return nil, ErrBadVector
+	}
+	var sum float64
+	for _, f := range v {
+		if math.IsNaN(float64(f)) || math.IsInf(float64(f), 0) {
+			return nil, ErrBadVector
+		}
+		sum += float64(f) * float64(f)
+	}
+	norm := math.Sqrt(sum)
+	if norm == 0 {
+		return nil, ErrBadVector
+	}
+	out := make([]float32, len(v))
+	for i, f := range v {
+		out[i] = float32(float64(f) / norm)
+	}
+	return out, nil
+}
+
+// Similarity is the cosine similarity of two L2-normalised vectors. Different
+// widths score 0: a profile enrolled with another model must not accidentally
+// match, and 0 is below every sane threshold.
+func Similarity(a, b []float32) float64 {
+	if len(a) != len(b) || len(a) == 0 {
+		return 0
+	}
+	var sum float64
+	for i := range a {
+		sum += float64(a[i]) * float64(b[i])
+	}
+	return sum
+}
+
+// seconds is the playback length of a frame, for the minimum-audio checks.
+func seconds(a audio.Audio) float64 { return a.Duration() }

internal/speaker/speaker_test.go

@@ -0,0 +1,397 @@
+package speaker
+
+import (
+	"context"
+	"errors"
+	"math"
+	"strings"
+	"testing"
+	"time"
+
+	"github.com/kami/maven/internal/audio"
+	"github.com/kami/maven/internal/memory"
+)
+
+// fakeEmbedder returns a fixed vector per "voice", so a test can enrol one
+// person and present another without a model. Wobble adds a small perturbation
+// so repeated samples of one voice are close but not identical, which is what a
+// real embedder produces.
+type fakeEmbedder struct {
+	vec    []float32
+	err    error
+	calls  int
+	wobble float32
+}
+
+func (f *fakeEmbedder) Embed(_ context.Context, _ audio.Audio) ([]float32, error) {
+	f.calls++
+	if f.err != nil {
+		return nil, f.err
+	}
+	out := append([]float32(nil), f.vec...)
+	if f.wobble != 0 && len(out) > 1 {
+		out[0] += f.wobble * float32(f.calls)
+		out[1] -= f.wobble * float32(f.calls)
+	}
+	return out, nil
+}
+
+func (f *fakeEmbedder) Dim() int { return len(f.vec) }
+
+// speech builds n seconds of the canonical audio shape.
+func speech(sec float64) audio.Audio {
+	return audio.Audio{Format: audio.PCM16kMono, Bytes: make([]byte, int(sec*16000)*2)}
+}
+
+func newRec(t *testing.T, emb Embedder) (*Recognizer, memory.Catalog) {
+	t.Helper()
+	cat := memory.NewInMemoryStore()
+	r, err := New(emb, cat, Config{})
+	if err != nil {
+		t.Fatal(err)
+	}
+	return r, cat
+}
+
+func enrolSamples(n int, sec float64) []audio.Audio {
+	out := make([]audio.Audio, n)
+	for i := range out {
+		out[i] = speech(sec)
+	}
+	return out
+}
+
+// The state of this box: no model on disk. Every identification refuses rather
+// than guessing, and it says why.
+func TestDisabledRefusesEverything(t *testing.T) {
+	r, _ := newRec(t, nil)
+	if r.Enabled() {
+		t.Error("a recognizer with no model reports itself enabled")
+	}
+	if _, err := r.Identify(context.Background(), speech(5)); !errors.Is(err, ErrDisabled) {
+		t.Errorf("Identify = %v, want ErrDisabled", err)
+	}
+	if _, err := r.Enroll(context.Background(), "kami", "Ками", enrolSamples(3, 4)); !errors.Is(err, ErrDisabled) {
+		t.Errorf("Enroll = %v, want ErrDisabled", err)
+	}
+	// Listing still works: knowing that nobody is enrolled needs no model.
+	got, err := r.List(context.Background())
+	if err != nil || len(got) != 0 {
+		t.Errorf("List = %v, %v", got, err)
+	}
+}
+
+func TestNewRequiresAProfileStore(t *testing.T) {
+	if _, err := New(nil, nil, Config{}); err == nil {
+		t.Error("built a recognizer with nowhere to keep profiles")
+	}
+}
+
+func TestEnrollThenIdentify(t *testing.T) {
+	emb := &fakeEmbedder{vec: []float32{1, 0, 0, 0}, wobble: 0.01}
+	r, _ := newRec(t, emb)
+	ctx := context.Background()
+
+	p, err := r.Enroll(ctx, "Kami ", "Ками", enrolSamples(3, 4))
+	if err != nil {
+		t.Fatalf("enroll: %v", err)
+	}
+	if p.ID != "kami" {
+		t.Errorf("id = %q, want the normalised %q", p.ID, "kami")
+	}
+	if p.Name != "Ками" || p.Samples != 3 || p.Dim != 4 {
+		t.Errorf("profile = %+v", p)
+	}
+
+	m, err := r.Identify(ctx, speech(5))
+	if err != nil {
+		t.Fatalf("identify: %v", err)
+	}
+	if m.Profile.ID != "kami" || m.Profile.Name != "Ками" {
+		t.Errorf("match = %+v", m)
+	}
+	if m.Score < r.Threshold() {
+		t.Errorf("score %.3f is below the threshold it supposedly passed", m.Score)
+	}
+}
+
+// The error direction that matters. Naming the wrong person writes a false
+// memory about them, so a voice that is not close enough gets no name at all.
+func TestUnfamiliarVoiceIsNotGuessed(t *testing.T) {
+	emb := &fakeEmbedder{vec: []float32{1, 0, 0, 0}}
+	r, _ := newRec(t, emb)
+	ctx := context.Background()
+	if _, err := r.Enroll(ctx, "kami", "Ками", enrolSamples(3, 4)); err != nil {
+		t.Fatal(err)
+	}
+
+	// A different voice: orthogonal voiceprint, similarity 0.
+	emb.vec = []float32{0, 1, 0, 0}
+	m, err := r.Identify(ctx, speech(5))
+	if !errors.Is(err, ErrUnknown) {
+		t.Fatalf("Identify = %v, want ErrUnknown", err)
+	}
+	if m.Profile.ID != "" {
+		t.Errorf("a refused identification still handed back %q", m.Profile.ID)
+	}
+	// The log line needs the near miss to make the threshold tunable.
+	if !contains(err.Error(), "kami") {
+		t.Errorf("error does not name the closest profile: %v", err)
+	}
+}
+
+// Just under the threshold is still unknown. A boundary this important gets its
+// own test rather than being implied.
+func TestThresholdIsAFloorNotASuggestion(t *testing.T) {
+	cat := memory.NewInMemoryStore()
+	emb := &fakeEmbedder{vec: []float32{1, 0}}
+	r, err := New(emb, cat, Config{Threshold: 0.9})
+	if err != nil {
+		t.Fatal(err)
+	}
+	ctx := context.Background()
+	if _, err := r.Enroll(ctx, "kami", "Ками", enrolSamples(3, 4)); err != nil {
+		t.Fatal(err)
+	}
+	// cos ≈ 0.866, comfortably similar and still not similar enough.
+	emb.vec = []float32{0.866, 0.5}
+	if _, err := r.Identify(ctx, speech(5)); !errors.Is(err, ErrUnknown) {
+		t.Fatalf("0.866 against a 0.9 threshold = %v, want ErrUnknown", err)
+	}
+}
+
+func TestShortAudioIsRefusedBeforeTheModelRuns(t *testing.T) {
+	emb := &fakeEmbedder{vec: []float32{1, 0}}
+	r, _ := newRec(t, emb)
+	if _, err := r.Identify(context.Background(), speech(0.5)); !errors.Is(err, ErrTooShort) {
+		t.Fatalf("got %v, want ErrTooShort", err)
+	}
+	if emb.calls != 0 {
+		t.Error("a half-second of audio was sent to the model anyway")
+	}
+}
+
+func TestWrongAudioFormatIsRefused(t *testing.T) {
+	r, _ := newRec(t, &fakeEmbedder{vec: []float32{1, 0}})
+	bad := audio.Audio{
+		Format: audio.Format{SampleRate: 44100, Channels: 2, SampleBits: 16, Encoding: "pcm_s16le"},
+		Bytes:  make([]byte, 44100*4*5),
+	}
+	if _, err := r.Identify(context.Background(), bad); !errors.Is(err, ErrBadFormat) {
+		t.Fatalf("got %v, want ErrBadFormat", err)
+	}
+}
+
+func TestIdentifyWithNobodyEnrolled(t *testing.T) {
+	r, _ := newRec(t, &fakeEmbedder{vec: []float32{1, 0}})
+	if _, err := r.Identify(context.Background(), speech(5)); !errors.Is(err, ErrNoProfiles) {
+		t.Fatalf("got %v, want ErrNoProfiles", err)
+	}
+}
+
+// Enrolment is an explicit act with real samples behind it, not a byproduct of
+// someone speaking once.
+func TestEnrollmentRequiresSeveralRealSamples(t *testing.T) {
+	r, _ := newRec(t, &fakeEmbedder{vec: []float32{1, 0}})
+	ctx := context.Background()
+	cases := []struct {
+		name    string
+		samples []audio.Audio
+	}{
+		{"one long sample", enrolSamples(1, 30)},
+		{"two samples", enrolSamples(2, 10)},
+		{"three samples but seconds of audio", enrolSamples(3, 1)},
+		{"none at all", nil},
+	}
+	for _, c := range cases {
+		if _, err := r.Enroll(ctx, "kami", "Ками", c.samples); !errors.Is(err, ErrTooShort) {
+			t.Errorf("%s: %v, want ErrTooShort", c.name, err)
+		}
+	}
+}
+
+func TestEnrollRejectsBadIDs(t *testing.T) {
+	r, _ := newRec(t, &fakeEmbedder{vec: []float32{1, 0}})
+	for _, id := range []string{"", "  ", "../etc/passwd", "speaker:kami", "имя", "a/b", "x y"} {
+		if _, err := r.Enroll(context.Background(), id, "n", enrolSamples(3, 4)); !errors.Is(err, ErrBadID) {
+			t.Errorf("id %q accepted or wrong error: %v", id, err)
+		}
+	}
+}
+
+// Re-enrolling replaces the voiceprint. Leaving the old one searchable would
+// mean a person's rejected profile keeps matching them.
+func TestReEnrollReplaces(t *testing.T) {
+	emb := &fakeEmbedder{vec: []float32{1, 0, 0}}
+	r, _ := newRec(t, emb)
+	ctx := context.Background()
+	if _, err := r.Enroll(ctx, "kami", "Ками", enrolSamples(3, 4)); err != nil {
+		t.Fatal(err)
+	}
+	emb.vec = []float32{0, 1, 0}
+	if _, err := r.Enroll(ctx, "kami", "Ками", enrolSamples(4, 4)); err != nil {
+		t.Fatal(err)
+	}
+	list, err := r.List(ctx)
+	if err != nil {
+		t.Fatal(err)
+	}
+	if len(list) != 1 {
+		t.Fatalf("%d profiles after re-enrolling one person", len(list))
+	}
+	if list[0].Samples != 4 {
+		t.Errorf("sample count = %d, want the new 4", list[0].Samples)
+	}
+	// The new voiceprint is the one that matches.
+	if m, err := r.Identify(ctx, speech(5)); err != nil || m.Score < 0.99 {
+		t.Errorf("identify after re-enrol: %v (score %.3f)", err, m.Score)
+	}
+}
+
+// "Перестань узнавать её" has to actually delete the biometric.
+func TestForgetRemovesTheVoiceprint(t *testing.T) {
+	emb := &fakeEmbedder{vec: []float32{1, 0}}
+	r, cat := newRec(t, emb)
+	ctx := context.Background()
+	if _, err := r.Enroll(ctx, "guest", "Гостья", enrolSamples(3, 4)); err != nil {
+		t.Fatal(err)
+	}
+	if err := r.Forget(ctx, "Guest "); err != nil {
+		t.Fatalf("forget: %v", err)
+	}
+	recs, err := cat.ByPrefix(ctx, Prefix)
+	if err != nil {
+		t.Fatal(err)
+	}
+	if len(recs) != 0 {
+		t.Errorf("%d row(s) survived Forget", len(recs))
+	}
+	if _, err := r.Get(ctx, "guest"); !errors.Is(err, ErrNotFound) {
+		t.Errorf("Get after Forget = %v, want ErrNotFound", err)
+	}
+	if err := r.Forget(ctx, "guest"); !errors.Is(err, ErrNotFound) {
+		t.Errorf("second Forget = %v, want ErrNotFound", err)
+	}
+}
+
+// Voiceprints share the vector table with note and fact embeddings, so the
+// prefix has to actually partition it.
+func TestProfilesDoNotCollideWithNoteVectors(t *testing.T) {
+	emb := &fakeEmbedder{vec: []float32{1, 0}}
+	r, cat := newRec(t, emb)
+	ctx := context.Background()
+	if err := cat.Insert(ctx, "note:1", []float32{1, 0}, map[string]string{"text": "заметка"}); err != nil {
+		t.Fatal(err)
+	}
+	if _, err := r.Enroll(ctx, "kami", "Ками", enrolSamples(3, 4)); err != nil {
+		t.Fatal(err)
+	}
+	list, err := r.List(ctx)
+	if err != nil {
+		t.Fatal(err)
+	}
+	if len(list) != 1 || list[0].ID != "kami" {
+		t.Errorf("listing picked up a non-speaker row: %+v", list)
+	}
+	// And an identical note vector is never returned as a match.
+	m, err := r.Identify(ctx, speech(5))
+	if err != nil {
+		t.Fatal(err)
+	}
+	if m.Profile.ID != "kami" {
+		t.Errorf("matched %q", m.Profile.ID)
+	}
+}
+
+func TestEmbedderFailurePropagates(t *testing.T) {
+	emb := &fakeEmbedder{vec: []float32{1, 0}, err: errors.New("onnx fell over")}
+	r, _ := newRec(t, emb)
+	if _, err := r.Identify(context.Background(), speech(5)); err == nil {
+		t.Error("a model failure was reported as a successful identification")
+	}
+	if _, err := r.Enroll(context.Background(), "kami", "К", enrolSamples(3, 4)); err == nil {
+		t.Error("a model failure produced a profile")
+	}
+}
+
+// A zero vector scores 0 against everything, which reads as "no match" for the
+// wrong reason and would hide a broken model.
+func TestUnusableVectorsAreRefused(t *testing.T) {
+	r, _ := newRec(t, &fakeEmbedder{vec: []float32{0, 0, 0}})
+	if _, err := r.Enroll(context.Background(), "kami", "К", enrolSamples(3, 4)); !errors.Is(err, ErrBadVector) {
+		t.Errorf("zero vector: %v, want ErrBadVector", err)
+	}
+	if _, err := Normalize(nil); !errors.Is(err, ErrBadVector) {
+		t.Errorf("empty: %v", err)
+	}
+	if _, err := Normalize([]float32{float32(nan())}); !errors.Is(err, ErrBadVector) {
+		t.Errorf("NaN: %v", err)
+	}
+}
+
+func TestNormalizeProducesAUnitVector(t *testing.T) {
+	v, err := Normalize([]float32{3, 4})
+	if err != nil {
+		t.Fatal(err)
+	}
+	if got := Similarity(v, v); got < 0.999 || got > 1.001 {
+		t.Errorf("self-similarity = %f, want 1", got)
+	}
+}
+
+// A profile enrolled with another model must not accidentally match.
+func TestDifferentWidthsScoreZero(t *testing.T) {
+	if got := Similarity([]float32{1, 0}, []float32{1, 0, 0}); got != 0 {
+		t.Errorf("mismatched widths scored %f", got)
+	}
+}
+
+// Attribution belongs in the source, so it can be corrected without rewriting
+// what was said.
+func TestProfileSource(t *testing.T) {
+	p := Profile{ID: "kami"}
+	if got := p.Source("tap:voice"); got != "tap:voice:speaker:kami" {
+		t.Errorf("source = %q", got)
+	}
+	var anon Profile
+	if got := anon.Source("tap:voice"); got != "tap:voice" {
+		t.Errorf("unattributed source = %q, want the base unchanged", got)
+	}
+}
+
+func TestValidID(t *testing.T) {
+	for _, ok := range []string{"kami", "guest-2", "a_b", "x"} {
+		if !ValidID(ok) {
+			t.Errorf("%q rejected", ok)
+		}
+	}
+	for _, bad := range []string{"", "Kami", "имя", "a b", "a/b", "a:b", "..", strings.Repeat("a", 65)} {
+		if ValidID(bad) {
+			t.Errorf("%q accepted", bad)
+		}
+	}
+}
+
+func TestProfileMetadataSurvivesARoundTrip(t *testing.T) {
+	emb := &fakeEmbedder{vec: []float32{1, 0}}
+	r, _ := newRec(t, emb)
+	r.now = func() time.Time { return time.Date(2026, 8, 1, 12, 0, 0, 0, time.UTC) }
+	ctx := context.Background()
+	if _, err := r.Enroll(ctx, "kami", "Ками", enrolSamples(3, 4)); err != nil {
+		t.Fatal(err)
+	}
+	got, err := r.Get(ctx, "kami")
+	if err != nil {
+		t.Fatal(err)
+	}
+	if got.Name != "Ками" || got.Samples != 3 {
+		t.Errorf("profile = %+v", got)
+	}
+	if !got.Enrolled.Equal(time.Date(2026, 8, 1, 12, 0, 0, 0, time.UTC)) {
+		t.Errorf("enrolled = %v", got.Enrolled)
+	}
+}
+
+func contains(s, sub string) bool { return strings.Contains(s, sub) }
+
+func nan() float64 { return math.NaN() }

internal/store/memory.go

@@ -8,6 +8,7 @@ import (
 	"fmt"
 	"math"
 	"sort"
+	"strings"
 	"time"
 
 	"github.com/kami/maven/internal/memory"
@@ -37,8 +38,10 @@ func (s *Store) VectorMemory() *MemoryStore {
 	return &MemoryStore{db: s.db}
 }
 
-// compile-time check: MemoryStore satisfies the memory.Store interface.
+// compile-time check: MemoryStore satisfies the memory.Store interface, and the
+// wider Catalog that speaker profiles need (enumerate by prefix, delete by id).
 var _ memory.Store = (*MemoryStore)(nil)
+var _ memory.Catalog = (*MemoryStore)(nil)
 
 // Insert upserts a vector by id: a repeated id replaces the prior row rather
 // than accumulating duplicates (the note/fact ids are stable and unique, so a
@@ -95,6 +98,56 @@ func (m *MemoryStore) Search(ctx context.Context, vec []float32, topK int) ([]me
 	return out, nil
 }
 
+// ByPrefix returns every row whose id starts with prefix, vectors included.
+//
+// This is not a similarity query and deliberately does not score anything:
+// listing the enrolled voices is a question about which rows exist, and asking
+// it through Search would mean inventing a query vector to rank them by. The
+// prefix is matched with LIKE against an escaped pattern, so a profile id
+// containing % or _ cannot widen the match.
+func (m *MemoryStore) ByPrefix(ctx context.Context, prefix string) ([]memory.Record, error) {
+	pattern := escapeLike(prefix) + "%"
+	rows, err := m.db.QueryContext(ctx,
+		`SELECT id, vec, meta FROM memory_vectors WHERE id LIKE ? ESCAPE '\'`, pattern)
+	if err != nil {
+		return nil, fmt.Errorf("memory: by prefix %q: %w", prefix, err)
+	}
+	defer rows.Close()
+
+	var out []memory.Record
+	for rows.Next() {
+		var id, metaJSON string
+		var blob []byte
+		if err := rows.Scan(&id, &blob, &metaJSON); err != nil {
+			return nil, fmt.Errorf("memory: row: %w", err)
+		}
+		meta := map[string]string{}
+		if err := json.Unmarshal([]byte(metaJSON), &meta); err != nil {
+			return nil, fmt.Errorf("memory: unmarshal meta for %q: %w", id, err)
+		}
+		out = append(out, memory.Record{ID: id, Vec: decodeVec(blob), Meta: meta})
+	}
+	if err := rows.Err(); err != nil {
+		return nil, fmt.Errorf("memory: rows: %w", err)
+	}
+	return out, nil
+}
+
+// Delete removes one vector by id. A row that is not there is not an error —
+// "forget this voice" is satisfied either way.
+func (m *MemoryStore) Delete(ctx context.Context, id string) error {
+	if _, err := m.db.ExecContext(ctx, `DELETE FROM memory_vectors WHERE id = ?`, id); err != nil {
+		return fmt.Errorf("memory: delete %q: %w", id, err)
+	}
+	return nil
+}
+
+// escapeLike neutralises the LIKE wildcards in a literal prefix.
+func escapeLike(s string) string {
+	r := strings.NewReplacer(`\`, `\\`, `%`, `\%`, `_`, `\_`)
+	return r.Replace(s)
+}
+
 // encodeVec serializes a float32 slice as little-endian IEEE-754 bytes (4 bytes
 // per element) for the BLOB column.
 func encodeVec(v []float32) []byte {

scripts/gen-stt-fixtures.sh

@@ -0,0 +1,67 @@
+#!/usr/bin/env bash
+# gen-stt-fixtures.sh — regenerate the golden STT audio fixtures.
+#
+# The fixtures in cmd/mavsttd/testdata/*.wav are SYNTHESISED, not recorded.
+# They come out of the same piper voices maven speaks with, so nothing of the
+# owner's voice is committed and every fixture is reproducible from this
+# script plus the voice model. They are also small: 16 kHz mono s16le, a
+# couple of seconds each.
+#
+# Usage:
+#   scripts/gen-stt-fixtures.sh
+#
+# Voices are picked up from, in order, $PIPER_VOICE_RU / $PIPER_VOICE_EN, then
+# the repo's models/tts, then ~/esp-server/voices. The English voice is not
+# vendored; if it is missing the English fixture is skipped and the existing
+# one is left alone.
+set -euo pipefail
+
+root="$(cd "$(dirname "$0")/.." && pwd)"
+out="$root/cmd/mavsttd/testdata"
+piper="${PIPER_BIN:-$root/deps/piper/piper}"
+espeak="${PIPER_ESPEAK:-$root/deps/piper/espeak-ng-data}"
+
+pick_voice() {
+	for c in "$@"; do
+		[ -f "$c" ] && { echo "$c"; return 0; }
+	done
+	return 1
+}
+
+ru="$(pick_voice "${PIPER_VOICE_RU:-}" "$root/models/tts/ru_RU-irina-medium.onnx" "$HOME/esp-server/voices/ru_RU-irina-medium.onnx")" || {
+	echo "no russian piper voice found" >&2
+	exit 1
+}
+en="$(pick_voice "${PIPER_VOICE_EN:-}" "$root/models/tts/en_US-lessac-medium.onnx" "$HOME/esp-server/voices/en_US-lessac-medium.onnx")" || en=""
+
+# synth <voice> <out.wav> <text>
+# piper emits raw 22050 Hz s16le on stdout; ffmpeg resamples to the canonical
+# 16 kHz mono and writes a plain 44-byte-header WAV (-fflags bitexact keeps
+# ffmpeg's encoder LIST chunk out, so the bytes are stable across ffmpeg
+# builds and internal/audio.PCMFromWAV reads them without scanning).
+synth() {
+	local voice="$1" dest="$2" text="$3"
+	printf '%s' "$text" | LD_LIBRARY_PATH="$(dirname "$piper")" "$piper" \
+		--model "$voice" --config "$voice.json" \
+		--espeak_data "$espeak" --output_raw --quiet |
+		ffmpeg -hide_banner -loglevel error -y \
+			-f s16le -ar 22050 -ac 1 -i - \
+			-af "adelay=200,apad=pad_dur=0.2" \
+			-ar 16000 -ac 1 -c:a pcm_s16le -fflags bitexact "$dest"
+	echo "wrote $dest ($(stat -c%s "$dest") bytes)"
+}
+
+synth "$ru" "$out/ru_reminder.wav" "Напомни мне через час позвонить маме."
+synth "$ru" "$out/ru_fact.wav" "Отметь, что я выпил воды."
+synth "$ru" "$out/ru_query.wav" "Что у меня сегодня по календарю?"
+
+if [ -n "$en" ]; then
+	# Keep the English line free of words piper spells out letter by letter —
+	# "nginx" comes out of lessac as "engine X", which is a TTS artefact and
+	# would make the fixture assert on the wrong thing.
+	synth "$en" "$out/en_act.wav" "Restart the web server and check the disk space."
+else
+	echo "no english piper voice found — skipping en_act.wav" >&2
+fi
+
+echo "fixtures regenerated; expected transcripts live in $out/golden_v1.json"