omnius 1.0.417 → 1.0.419

package/dist/scripts/live-whisper.py

@@ -165,6 +165,24 @@ def main():
     window_samples = int(args.window_seconds * SAMPLE_RATE)
     last_text = ""
     running = True
+    noise_floor = None  # EMA of the quietest 0.5s block RMS — adapts to the capture chain
+
+    def block_energies(samples):
+        """RMS of consecutive 0.5s blocks — lets speech anywhere in the
+        window register even when most of the window is silence."""
+        block = SAMPLE_RATE // 2
+        n = max(1, len(samples) // block)
+        return [float(np.sqrt(np.mean(samples[i * block:(i + 1) * block] ** 2))) for i in range(n)]
+
+    def amplify(samples):
+        """Bring quiet-but-real speech toward full scale. Whisper performs far
+        better near normal levels; quiet capture chains (distant speakers,
+        low source gain, beamformed arrays) otherwise force users to shout
+        into the mic. Gain capped at 40x (~32dB), output clipped to [-1, 1]."""
+        peak = float(np.max(np.abs(samples))) if len(samples) else 0.0
+        if 0 < peak < 0.5:
+            samples = np.clip(samples * min(40.0, 0.9 / max(peak, 1e-4)), -1.0, 1.0)
+        return samples
 
     def read_stdin():
         """Read PCM16 from stdin in a background thread."""
@@ -197,6 +215,26 @@ def main():
                 # Take the last window_seconds of audio
                 window = audio_buf[-window_samples:].copy() if len(audio_buf) > window_samples else audio_buf.copy()
 
+            # Adaptive silence gate: whisper hallucinates repetitive garbage
+            # on silent input, but absolute thresholds reject real speech on
+            # quiet capture chains. Instead track the noise floor (quietest
+            # 0.5s block) and require the loudest block to rise above it.
+            energies = block_energies(window)
+            quiet = min(energies)
+            loud = max(energies)
+            if noise_floor is None:
+                noise_floor = quiet
+            elif quiet < noise_floor:
+                noise_floor = 0.7 * noise_floor + 0.3 * quiet  # drop fast
+            else:
+                noise_floor = 0.995 * noise_floor + 0.005 * quiet  # creep up slowly
+            gate = max(1.5e-4, noise_floor * 2.5)  # ≥ ~8dB above floor, ~-76dB min
+            if loud < gate:
+                continue
+
+            # Amplify toward full scale before transcription
+            window = amplify(window)
+
             # Transcribe
             try:
                 fp16 = (device == "cuda")
@@ -221,8 +259,11 @@ def main():
     with buf_lock:
         full_audio = audio_buf.copy()
 
-    if len(full_audio) >= SAMPLE_RATE:
+    final_loud = max(block_energies(full_audio)) if len(full_audio) >= SAMPLE_RATE else 0.0
+    final_gate = max(1.5e-4, (noise_floor or 0.0) * 2.5)
+    if len(full_audio) >= SAMPLE_RATE and final_loud >= final_gate:
         try:
+            full_audio = amplify(full_audio)
             fp16 = (device == "cuda")
             result = model.transcribe(
                 full_audio,

package/npm-shrinkwrap.json

@@ -1,12 +1,12 @@
 {
   "name": "omnius",
-  "version": "1.0.417",
+  "version": "1.0.419",
   "lockfileVersion": 3,
   "requires": true,
   "packages": {
     "": {
       "name": "omnius",
-      "version": "1.0.417",
+      "version": "1.0.419",
       "bundleDependencies": [
         "image-to-ascii"
       ],

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "omnius",
-  "version": "1.0.417",
+  "version": "1.0.419",
   "description": "AI coding agent powered by open-source models (Ollama/vLLM) — interactive TUI with agentic tool-calling loop",
   "type": "module",
   "main": "./dist/index.js",