@vaibhavgupta0786/opencode-voice 0.3.7 → 0.3.8

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@vaibhavgupta0786/opencode-voice",
-  "version": "0.3.7",
+  "version": "0.3.8",
   "description": "Local-only push-to-talk dictation for OpenCode v2 \u2014 press F9, speak, press F9.",
   "type": "module",
   "license": "MIT",
@@ -53,7 +53,7 @@
     "@opentui/solid": "^0.5.0",
     "@types/node": "^26.0.0",
     "solid-js": "^1.9.0",
-    "typescript": "^5.4.0"
+    "typescript": "^7.0.2"
   },
   "repository": {
     "type": "git",

package/servers/mlx_server.py

@@ -28,6 +28,39 @@ import mlx_whisper
 MODEL = os.environ.get("MLX_MODEL", "mlx-community/whisper-small-mlx")
 PORT = int(os.environ.get("VOICE_STT_PORT", "8080"))
 
+# No-speech gate: refuse to transcribe clips that don't look spoken.
+# faster-whisper's default path has VAD + no_speech_threshold for this;
+# the mlx path inherits no such guard and hallucinates on room noise
+# (e.g. pure noise decoded into repeating nonsense Syllables).
+# Config: VOICE_NOSPEECH_MIN_MS / VOICE_NOSPEECH_ZCR, "0" disables.
+NOSPEECH_MIN_MS = float(os.environ.get("VOICE_NOSPEECH_MIN_MS", "200"))
+NOSPEECH_ZCR = float(os.environ.get("VOICE_NOSPEECH_ZCR", "0.4"))
+
+
+def speech_gate(samples: np.ndarray) -> tuple[bool, str]:
+    """Cheap numpy voice-likeness gate; (pass, reason).
+
+    Frames are 20ms hops at 16kHz. A frame counts as voiced when its RMS
+    exceeds 2% full scale. White noise has near-50% zero crossings vs
+    speech tones well under; noise that is loud enough to be voiced gets
+    refused on zero-crossing rate instead.
+    """
+    frame = int(16000 * 0.02)
+    n = len(samples) // frame
+    voiced_ms = 0.0
+    zcr = 0.0
+    if n:
+        rms = np.sqrt((samples[: n * frame].reshape(n, frame) ** 2).mean(axis=1))
+        voiced = int((rms > 0.02).sum())
+        voiced_ms = voiced * frame / 16000 * 1000
+        if len(samples) > 1:
+            zcr = (samples[1:] * samples[:-1] < 0).mean()
+    if NOSPEECH_MIN_MS and voiced_ms < NOSPEECH_MIN_MS:
+        return False, f"voiced={voiced_ms:.0f}ms < {NOSPEECH_MIN_MS:.0f}ms"
+    if NOSPEECH_ZCR and zcr > NOSPEECH_ZCR:
+        return False, f"zcr={zcr:.2f} > {NOSPEECH_ZCR} (noise-like)"
+    return True, f"voiced={voiced_ms:.0f}ms zcr={zcr:.2f}"
+
 print(f"loading {MODEL} ...", flush=True)
 # Warm up by loading weights once; per-request calls reuse them.
 _MODEL_REF = MODEL
@@ -80,6 +113,12 @@ class Handler(BaseHTTPRequestHandler):
                 self._json(400, {"error": "no file part"})
                 return
             samples = wav_bytes_to_float32(bytes(audio))
+            passed, gate = speech_gate(samples)
+            if not passed:
+                print(f"no-speech guard: {gate}", flush=True)
+                self._json(200, {"text": ""})
+                return
+            print(f"speech gate: {gate}", flush=True)
             try:
                 result = mlx_whisper.transcribe(samples, path_or_hf_repo=_MODEL_REF)
             finally:

package/skills/voice-check/SKILL.md

@@ -57,12 +57,27 @@ model loads from cache in roughly 10-25 s on a cold start:
 
     curl -s -m 3 http://127.0.0.1:8080/health
 
-## 5. If it still will not start
+## 5. Memory check
+
+Long-running STT servers ratchet memory silently — buffer allocators
+retain everything (measured: the mlx server held 7.7 GB after 25 days
+because each unique audio length adds GPU buffers that are never freed;
+a per-request `mx.clear_cache()` now caps it). Verify after boot:
+
+    PID=$(lsof -ti:8080 | head -1)
+    vmmap --summary $PID 2>/dev/null | grep "Physical footprint"
+
+Healthy: mlx server ≈ 0.7–1.5 GB · CPU (faster-whisper) ≈ 0.4–0.9 GB.
+If the footprint is already high (mlx > 2 GB, CPU > 1 GB), return to
+step 3, restart, and re-verify after one real take — the plateau is the
+health signal, not the size.
+
+## 6. If it still will not start
 
 Show the last 10-20 lines of the server log (`mlx-server.log` or
 `stt-server.log`) and report the error you see. Do not guess further.
 
-## 6. Report
+## 7. Report
 
-One line: server was up / has been restarted / failed (with the error).
-Remind the user to press Fn+F9 to dictate.
+One line: server was up / has been restarted / failed (with the error),
+plus the footprint reading. Remind the user to press Fn+F9 to dictate.