fix(voice): RMS-amplify STT input for weak voice

Condition the handset capture by RMS (not peak) before whisper: the
voice body is ~0.5-1.5% FS and a transient peak defeated peak-normalise,
leaving it inaudible. RMS-normalise to ~20% FS + clip. Recovers weak
captures whisper returned '' for (verified: -> 'C'est parti !').

tools/zacus-gateway/main.py

@@ -1941,17 +1941,18 @@ def _build_wav(pcm: bytes, sr: int, ch: int, sw: int) -> bytes:
     return out.getvalue()
 
 
-def _normalize_wav_for_stt(wav_bytes: bytes, target_peak: float = 0.7,
+def _normalize_wav_for_stt(wav_bytes: bytes, target_rms: float = 0.20,
-                           max_gain: float = 40.0, hp_cutoff_hz: float = 110.0) -> bytes:
+                           max_gain: float = 50.0, hp_cutoff_hz: float = 110.0) -> bytes:
-    """Condition a PLIP handset capture for Kyutai STT: high-pass then normalise.
+    """Condition a PLIP handset capture for the STT: high-pass then RMS-amplify.
 
-    Two problems with the SLIC handset path: (1) it injects low-frequency rumble
+    Two problems with the SLIC handset path: (1) low-frequency rumble/DC drift
-    / DC drift that swamps the speech and makes Kyutai output nothing — a
+    that swamps the speech — a high-pass (~110 Hz, box moving-average subtraction)
-    high-pass (~110 Hz, box moving-average subtraction) removes it; (2) the mic
+    removes it; (2) the voice body comes in VERY quiet (~0.5-1.5 % FS RMS). We
-    comes in very quiet (~2-5 % FS even at +24 dB PGA), so peak-normalise after
+    amplify by RMS (NOT peak — a transient peak defeats peak-normalisation and
-    filtering (gain capped so silence isn't blown up). WITHOUT the high-pass the
+    leaves the body inaudible) to ~20 % FS, then hard-clip the rare transient.
-    exact same capture transcribes empty; WITH it, clean French — verified on the
+    Verified: weak captures whisper returned '' for transcribe correctly once
-    bench."""
+    RMS-boosted (e.g. → "C'est parti !"). Silence gets blown up to noise but the
+    whisper hallucination filter drops the result."""
     if _np is None:
         return wav_bytes
     parsed = _wav_pcm(wav_bytes)
@@ -1967,11 +1968,11 @@ def _normalize_wav_for_stt(wav_bytes: bytes, target_peak: float = 0.7,
     if arr.size > k:
         lp = _np.convolve(arr, _np.ones(k, dtype=_np.float32) / k, mode="same")
         arr = arr - lp
-    peak = float(_np.max(_np.abs(arr))) or 1.0
+    rms = float(_np.sqrt(_np.mean(arr * arr))) or 1.0
-    gain = min((target_peak * 32767.0) / peak, max_gain)
+    gain = min((target_rms * 32767.0) / rms, max_gain)
     arr = _np.clip(arr * gain, -32768, 32767).astype(_np.int16)
-    logging.info("STT conditioning: high-pass %.0f Hz, peak %.1f%% FS → gain x%.1f",
+    logging.info("STT conditioning: high-pass %.0f Hz, RMS %.2f%% FS → gain x%.1f",
-                 hp_cutoff_hz, 100 * peak / 32768, gain)
+                 hp_cutoff_hz, 100 * rms / 32768, gain)
     return _build_wav(arr.tobytes(), sr, ch, sw)