local-transcript

Transcribing local media files into a polished transcript involves running a pipeline that extracts audio using ffmpeg, transcribes via mlx-whisper or faster-whisper, and applies LLM proofreading. It supports txt, pdf, and docx outputs with language auto-detection.

Yes, you can use faster-whisper for transcription on a CPU without Apple Silicon as it serves as a fallback backend. The pipeline automatically defaults to mlx-whisper for Apple Silicon GPU acceleration but switches to faster-whisper when GPU hardware is unavailable.

The best way to generate Chinese transcripts from local video is using a pipeline with language auto-detection and LLM-assisted proofreading. It ensures natural paragraph formatting and matches the spoken language, specifically producing Chinese output when the source audio is Chinese.

Yes, local-transcript supports exporting to docx and pdf formats, alongside txt. The pipeline generates these multi-format final deliverables directly from the transcribed media using python-docx and reportlab, providing cleaned transcripts for immediate downstream use.

ASR transcriptions contain errors in domain terms and proper nouns due to speech recognition limitations. The pipeline corrects these by applying deterministic fixes for common ASR errors, followed by LLM-based proofreading to ensure quality, consistency, and accurate terminology.

Caching accelerates repeated transcription runs on the same media by storing intermediate processing results. This prevents redundant audio extraction and ASR transcription, allowing the pipeline to quickly retrieve cached data and output the final polished transcripts for repeated analysis.