meizhong986/WhisperJAV

Implements a multi-stage noise-suppression architecture combining scene-based VAD segmentation, domain-specific linguistic normalization (handling Japanese onomatopoeia and dialect tokenization), and defensive decoding with log-probability thresholding to eliminate hallucinations in extended audio. Supports pluggable ASR backends (Whisper, Faster-Whisper, Qwen3-ASR, anime-whisper, HuggingFace Kotoba) through a decoupled ChronosJAV pipeline that separates text generation from forced-alignment timestamp inference. Offers seven processing modes with configurable scene detection, speech enhancement, and segmentation strategies, plus two-pass ensemble mode to merge outputs across different model architectures.

1,216 stars. Actively maintained with 148 commits in the last 30 days.