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/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.

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"