Refine VAD parameters and transcription options in WhisperTranscriber for improved audio processing. Adjust max speech duration, min speech duration, and silence duration, and set chunk length to 12 seconds.

app.py

@@ -371,14 +371,16 @@ class WhisperTranscriber:
             beam_size=5,
             vad_filter=False,  # VAD is enabled by default for batched transcription
             vad_parameters=VadOptions(
-                max_speech_duration_s=whisper.feature_extractor.chunk_length,
-                min_speech_duration_ms=100,
-                speech_pad_ms=100,
+                max_speech_duration_s=12,
+                min_speech_duration_ms=150,   # ignore ultra-short blips
+                min_silence_duration_ms=150,  # split on short Mandarin pauses (if supported)                speech_pad_ms=100,
                 threshold=0.25,
                 neg_threshold=0.2,
+
             ),
             word_timestamps=True,
             initial_prompt=prompt,
+            condition_on_previous_text=False,  # avoid runaway context
             language_detection_segments=1,
             task="translate" if translate else "transcribe",
         )
@@ -387,6 +389,7 @@ class WhisperTranscriber:
         segments, transcript_info = batched_whisper.transcribe(
             audio_path, 
             batch_size=batch_size, 
+            chunk_length=12,   # ↓ from 30s default; try 10–15
             **options
         )
         segments = list(segments)