AIWisper v1.0.6

Improved

• Whisper Turbo Quality Optimization: Significant improvements to speech recognition quality
  • Added temperature=0.0 for deterministic output (reduces hallucinations)
  • Added condition_on_previous_text=False to prevent looping and error accumulation
  • Added hallucination_silence_threshold=2.0 to filter out phantom speech on silence
  • Added no_speech_threshold=0.5 for better silence detection
  • Optimized Silero VAD parameters for better speech detection:
    • threshold=0.5 for speech detection
    • min_speech_duration_ms=250 minimum speech duration
    • min_silence_duration_ms=2000 for segment separation
    • speech_pad_ms=400 padding around speech
  • Enabled word_timestamps=True for hallucination detection
• Go binding optimization: Updated whisper.cpp parameters
  • Temperature: 0.1 -> 0.0 (deterministic)
  • Temperature fallback: 0.3 -> 0.2 (less variability)
  • Added MaxContext=-1 to disable context (prevents looping)

Technical Details

• These changes apply to both faster-whisper (Python CLI) and native whisper.cpp (Go binding)
• Silero VAD in faster-whisper works alongside the existing Go-based VAD for chunk splitting
• No API changes - improvements are transparent to end users