Audio to Word Document

Problem Solved

Audio files are often just raw speech and are hard to use directly for minutes, archives, or follow-up review. Manual transcription is slow, while generic online tools may send sensitive meeting content to external services. This skill targets Chinese audio and converts local mp3, wav, or m4a files into a structured .docx draft, reducing the manual step between listening and usable text.

How It Works

The skill runs faster-whisper locally and defaults to CPU with int8 inference, which is suitable for machines without a GPU; if a CUDA environment is already available, it can be used for acceleration. The transcript is packaged as a Word document with three parts: meeting header, transcript body, and timestamped line-by-line transcript. The body is merged into natural paragraphs by pauses, while the appendix keeps [HH:MM:SS] timestamps for context checks. The script also handles temporary ASCII path copies, traditional-to-simplified conversion, python-docx Chinese font settings, and hf-mirror.com model downloads, reducing friction on local desktop environments and domestic networks.

Scope and Notes

This is best suited for Chinese meetings, interviews, or voice notes, not real-time captions. The medium model on a typical CPU may run at roughly 1.2–2× realtime, so long recordings can take a long time and should be run in the background, with completion checked by DONE in the log and the presence of the final file. The transcript may still be inaccurate for specialized terms, homophones, or overlapping speakers, so users should verify key facts, names, and numbers before delivery.