v2.0
馃殌 Version 2.0 - complete overhaul
New Configuration Options
- Multiple directory Path support
- confidence_threshold: Determine required confidence. lowering may result in faster process times at more risk of false positives
- reprocess_all: Option to reprocess all audio tracks, not just undefined ones
Performance & Speed:
- Switched to faster-whisper (CTranslate2 backend) for 4-5x faster transcription speeds
- GPU acceleration support: Automatic CUDA detection with fallback to CPU
- Optimized compute types: Automatic selection of optimal compute types (float16 for GPU, int8 for CPU)
- CPU thread control: Configurable CPU thread allocation for better resource management
Enhanced Language Detection
- Percentage-based audio sampling: Intelligent sample extraction based on file duration percentages
- Improved VAD (Voice Activity Detection): Configurable minimum/maximum speech duration parameters
- Multi-stage detection: Primary attempt with VAD, fallback without VAD if audio is removed
- Confidence-based processing: Configurable confidence threshold (default: 0.9)
- Full-track analysis: For results below confidence_threshold, analyzes entire audio track for better accuracy
Finetuned Hallucination Detection
- Enhanced pattern recognition: Better detection of repetitive characters and common hallucination phrases
- Compression ratio analysis: Identifies highly repetitive (likely hallucinated) text
- Language-specific filtering: Improved detection of non-Latin script hallucinations
Improved Metadata Handling
- Language code normalization
- Track change tracking: Shows previous language when updating tracks
Full Changelog: 1.3...2.0