Local Audio Video Transcript Organizer

Problem

Podcasts, meeting videos, and raw transcripts often arrive with timestamps, repetitions, fragmented sentences, and ASR errors. This skill is aimed at a local workflow: it turns audio, video, links, subtitles, raw transcripts, or pasted notes into one shareable Chinese original-text document, not a summary or meeting minutes.

How It Works

By default it keeps stage artifacts under ./transcript-organizer-workspace/, including input/, audio_wav/, transcripts_raw/, transcripts_clean/, and final/. For media input, it extracts audio and transcodes it to 16kHz mono WAV; for links, it can download with yt-dlp; for subtitles or text drafts, it normalizes them directly. The ASR backend is selected by environment: Apple Silicon setups may use mlx-whisper, while other environments may use faster-whisper. The cleanup pass removes timestamps, speaker labels, filler, and repetition, and fixes common errors in names, product names, and technical terms while preserving meaning, numbers, examples, and argument chains. If the source is not Chinese, the final output is a faithful Chinese original-text version, not a simple translation or compressed takeaway list.

Boundaries

The skill is local-first: it does not create online documents or upload user content. It requires ffmpeg; link downloads need yt-dlp. If installation, memory, or login-gated links fail, check logs/ and fall back to local files, existing transcripts, or a smaller model. It is suitable for preserving shareable original expression, not for generating to-do lists, bilingual parallel text, or replacing human fact-checking.