mavsttd: silence gate — don't feed non-speech to whisper

Whisper hallucinates subtitle-credit boilerplate ("Редактор субтитров …") on
silence/room-noise, which then got stored as tap:voice facts. Gate before the
model: drop clips shorter than -min-ms (default 300) or below -silence-rms
(default 0.01 normalized RMS). Both are flags — the mic floor is hardware
specific. Returns empty transcript (same as whisper's no-segments path), so
nothing downstream changes.

gateReason is pure and unit-tested (silence/short/quiet → dropped, loud+long →
passes). ponytail: energy gate, not a real VAD; upgrade to WebRTC VAD or
whisper no_speech_prob if too blunt.

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

cmd/mavsttd/whisper_handler.go

@@ -11,17 +11,24 @@ import "C"
 import (
 	"context"
 	"fmt"
+	"log"
 	"math"
 	"unsafe"
 
 	"github.com/kami/maven/internal/worker"
 )
 
+// whisperSampleRate is the rate whisper.cpp requires; the pipeline resamples
+// to it before sending, so it's also the rate the silence gate assumes.
+const whisperSampleRate = 16000
+
 type whisperHandler struct {
-	ctx *C.struct_whisper_context
+	ctx    *C.struct_whisper_context
+	minMs  int     // clips shorter than this are dropped (hallucination bait)
+	minRMS float64 // clips quieter than this (normalized RMS) are treated as silence
 }
 
-func newWhisperHandler(modelPath string) (*whisperHandler, error) {
+func newWhisperHandler(modelPath string, minMs int, minRMS float64) (*whisperHandler, error) {
 	cparams := C.whisper_context_default_params()
 	cPath := C.CString(modelPath)
 	defer C.free(unsafe.Pointer(cPath))
@@ -30,7 +37,33 @@ func newWhisperHandler(modelPath string) (*whisperHandler, error) {
 	if ctx == nil {
 		return nil, fmt.Errorf("whisper: failed to init from %s", modelPath)
 	}
-	return &whisperHandler{ctx: ctx}, nil
+	return &whisperHandler{ctx: ctx, minMs: minMs, minRMS: minRMS}, nil
+}
+
+// gateReason returns a non-empty reason when audio must NOT reach whisper:
+// too short, or below the energy floor (silence / room noise). Whisper
+// hallucinates subtitle-credit boilerplate ("Редактор субтитров …",
+// "Субтитры сделал …") on non-speech input, so we drop it before the model
+// ever sees it. Pure (no CGo) so it's unit-tested directly.
+//
+// ponytail: plain RMS energy + min-duration, not a real VAD. The mic floor is
+// hardware-specific (both thresholds are flags on mavsttd) — upgrade to WebRTC
+// VAD / whisper's no_speech_prob if energy gating proves too blunt.
+func gateReason(samples []float32, rate, minMs int, minRMS float64) string {
+	if len(samples) == 0 {
+		return "empty"
+	}
+	if ms := len(samples) * 1000 / rate; ms < minMs {
+		return fmt.Sprintf("too short (%dms < %dms)", ms, minMs)
+	}
+	var sum float64
+	for _, s := range samples {
+		sum += float64(s) * float64(s)
+	}
+	if rms := math.Sqrt(sum / float64(len(samples))); rms < minRMS {
+		return fmt.Sprintf("silence (rms %.4f < floor %.4f)", rms, minRMS)
+	}
+	return ""
 }
 
 func (h *whisperHandler) Transcribe(ctx context.Context, req worker.TranscribeReq) (worker.TranscribeResp, error) {
@@ -49,6 +82,12 @@ func (h *whisperHandler) Transcribe(ctx context.Context, req worker.TranscribeRe
 		samples[i] = float32(s) / 32768.0
 	}
 
+	// Silence gate: drop non-speech before whisper hallucinates on it.
+	if reason := gateReason(samples, whisperSampleRate, h.minMs, h.minRMS); reason != "" {
+		log.Printf("mavsttd: gated audio (%s) — skipping whisper", reason)
+		return worker.TranscribeResp{Text: "", Confidence: 0}, nil
+	}
+
 	params := C.whisper_full_default_params(C.WHISPER_SAMPLING_GREEDY)
 	params.print_progress = false
 	params.print_realtime = false