diff --git a/Makefile b/Makefile
index 28f34fd..ec8aaf3 100644
--- a/Makefile
+++ b/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)
diff --git a/cmd/mavsttd/golden_test.go b/cmd/mavsttd/golden_test.go
new file mode 100644
index 0000000..0783a83
--- /dev/null
+++ b/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")
+	}
+}
diff --git a/cmd/mavsttd/testdata/en_act.wav b/cmd/mavsttd/testdata/en_act.wav
new file mode 100644
index 0000000..5a904ae
Binary files /dev/null and b/cmd/mavsttd/testdata/en_act.wav differ
diff --git a/cmd/mavsttd/testdata/golden_v1.json b/cmd/mavsttd/testdata/golden_v1.json
new file mode 100644
index 0000000..870a4d5
--- /dev/null
+++ b/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
+    }
+  ]
+}
diff --git a/cmd/mavsttd/testdata/ru_fact.wav b/cmd/mavsttd/testdata/ru_fact.wav
new file mode 100644
index 0000000..fe6a3ac
Binary files /dev/null and b/cmd/mavsttd/testdata/ru_fact.wav differ
diff --git a/cmd/mavsttd/testdata/ru_query.wav b/cmd/mavsttd/testdata/ru_query.wav
new file mode 100644
index 0000000..3fabbda
Binary files /dev/null and b/cmd/mavsttd/testdata/ru_query.wav differ
diff --git a/cmd/mavsttd/testdata/ru_reminder.wav b/cmd/mavsttd/testdata/ru_reminder.wav
new file mode 100644
index 0000000..b979e7b
Binary files /dev/null and b/cmd/mavsttd/testdata/ru_reminder.wav differ
diff --git a/scripts/gen-stt-fixtures.sh b/scripts/gen-stt-fixtures.sh
new file mode 100755
index 0000000..98d6dae
--- /dev/null
+++ b/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"