local-transcript

# Local Transcript ## Overview Use this skill to turn a local media file into cleaned final transcript files in `.txt`, `.pdf`, or `.docx` format. Extract audio with `ffmpeg`, transcribe with `mlx-whisper` (Apple Silicon GPU) or `faster-whisper` (CPU fallback), then clean the transcript with deterministic replacements for known ASR bugs. LLM contextual proofreading is an opt-in second layer (`--llm-backend local|claude`) — see §LLM Proofreading Is Opt-In for why it is not the default. ## Workflow 1. Validate the input path. 2. Confirm the requested output format. 3. Check dependencies. 4. Resolve the ASR mode: `fast`, `balanced`, or `accurate`. 5. Reuse cached audio/raw transcript/clean transcript layers when available. 6. Extract or reuse 16 kHz mono WAV audio. 7. Transcribe with the selected ASR backend (language auto-detected or user-specified via `--language`). 8. Clean the transcript: simplified Chinese → deterministic replacements → (opt-in) LLM proofreading → post-LLM safety replacements. 9. Paragraphize and write the requested final file(s). ## Format Resolution Gate - If the user explicitly requests `txt`, `pdf`, or `word/docx`, use that format directly. - If the user requests multiple formats, generate all requested formats from the same cleaned transcript. - If the user asks to transcribe a file but does not specify an output format, ask a short follow-up question before execution: `Which output format do you want: txt, pdf, or docx?` - Do not guess the outpu...