omnius 1.0.417 → 1.0.418

package/dist/scripts/live-whisper.py

@@ -197,6 +197,19 @@ 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()
 
+            # Silence gate: whisper hallucinates repetitive garbage on
+            # near-silent input (quiet mics, diluted multichannel downmixes).
+            # Skip windows with no meaningful energy instead of transcribing.
+            rms = float(np.sqrt(np.mean(window ** 2))) if len(window) else 0.0
+            peak = float(np.max(np.abs(window))) if len(window) else 0.0
+            if rms < 3e-4 or peak < 2e-3:  # ~-70dB RMS: definitely silence
+                continue
+
+            # Normalize quiet-but-real audio: whisper performs much better
+            # near full scale. Cap gain at 20x so noise floors aren't blasted.
+            if peak < 0.3:
+                window = window * min(20.0, 0.9 / max(peak, 1e-4))
+
             # Transcribe
             try:
                 fp16 = (device == "cuda")
@@ -221,8 +234,12 @@ def main():
     with buf_lock:
         full_audio = audio_buf.copy()
 
-    if len(full_audio) >= SAMPLE_RATE:
+    final_rms = float(np.sqrt(np.mean(full_audio ** 2))) if len(full_audio) else 0.0
+    if len(full_audio) >= SAMPLE_RATE and final_rms >= 3e-4:
         try:
+            final_peak = float(np.max(np.abs(full_audio)))
+            if 0 < final_peak < 0.3:
+                full_audio = full_audio * min(20.0, 0.9 / max(final_peak, 1e-4))
             fp16 = (device == "cuda")
             result = model.transcribe(
                 full_audio,

package/npm-shrinkwrap.json

@@ -1,12 +1,12 @@
 {
   "name": "omnius",
-  "version": "1.0.417",
+  "version": "1.0.418",
   "lockfileVersion": 3,
   "requires": true,
   "packages": {
     "": {
       "name": "omnius",
-      "version": "1.0.417",
+      "version": "1.0.418",
       "bundleDependencies": [
         "image-to-ascii"
       ],

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "omnius",
-  "version": "1.0.417",
+  "version": "1.0.418",
   "description": "AI coding agent powered by open-source models (Ollama/vLLM) — interactive TUI with agentic tool-calling loop",
   "type": "module",
   "main": "./dist/index.js",