omnius 1.0.424 → 1.0.425

package/dist/scripts/live-whisper.py

@@ -286,14 +286,35 @@ def _segment_filtered_text(result) -> str:
 
 
 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
+    """Bring quiet-but-real speech toward a healthy level for whisper.
+
+    Gain targets the SPEECH RMS (median of the loudest half of 0.2s blocks),
+    not the raw peak — a single click/pop no longer suppresses amplification
+    of otherwise-quiet speech. Gain is capped at 40x (~32dB) and bounded so
+    the peak never clips."""
+    if len(samples) == 0:
+        return samples
+    peak = float(np.max(np.abs(samples)))
+    if peak <= 0:
+        return samples
+    block = max(1, SAMPLE_RATE // 5)
+    energies = sorted(
+        float(np.sqrt(np.mean(samples[i:i + block] ** 2)))
+        for i in range(0, max(1, len(samples) - block + 1), block)
+    )
+    loud_half = energies[len(energies) // 2:] or energies
+    speech_rms = loud_half[len(loud_half) // 2]  # median of the loud half
+    target_rms = 0.08  # ≈ -22dBFS — comfortable whisper input level
+    if speech_rms >= target_rms:
+        return samples
+    # No-clip bound uses the 99th-percentile amplitude, not the absolute
+    # peak — an isolated click/pop may clip (harmless to ASR) instead of
+    # suppressing amplification of the actual speech.
+    p99 = float(np.percentile(np.abs(samples), 99.0))
+    gain = min(40.0, target_rms / max(speech_rms, 1e-5), 0.98 / max(p99, 1e-4))
+    if gain <= 1.05:
+        return samples
+    return np.clip(samples * gain, -1.0, 1.0)
 
 # ---------------------------------------------------------------------------
 # Main transcription loop — utterance FSM
@@ -302,7 +323,7 @@ def amplify(samples):
 def main():
     import argparse
     parser = argparse.ArgumentParser(description="Live Whisper transcription worker")
-    parser.add_argument("--model", default="base", help="Whisper model size (tiny/base/small/medium/large)")
+    parser.add_argument("--model", default="medium", help="Whisper model size (tiny/base/small/medium/large)")
     # Accepted for backward compatibility with older launchers (sliding-window era).
     parser.add_argument("--chunk-seconds", type=float, default=3, help=argparse.SUPPRESS)
     parser.add_argument("--window-seconds", type=float, default=10, help=argparse.SUPPRESS)

package/npm-shrinkwrap.json

@@ -1,12 +1,12 @@
 {
   "name": "omnius",
-  "version": "1.0.424",
+  "version": "1.0.425",
   "lockfileVersion": 3,
   "requires": true,
   "packages": {
     "": {
       "name": "omnius",
-      "version": "1.0.424",
+      "version": "1.0.425",
       "bundleDependencies": [
         "image-to-ascii"
       ],

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "omnius",
-  "version": "1.0.424",
+  "version": "1.0.425",
   "description": "AI coding agent powered by open-source models (Ollama/vLLM) — interactive TUI with agentic tool-calling loop",
   "type": "module",
   "main": "./dist/index.js",