transcribe-summarizelisted

Transcribe audio locally on macOS, Windows or Linux, and optionally write it up as meeting notes. Use whenever the user has an audio or video file and wants what was said - "transcribe this", "what was said on the call", "turn this recording into notes", "write up this meeting", "subtitle this", "get me an .srt", a voice memo, a dictation, an interview, a lecture, a podcast. Runs on-device by default with Whisper (mlx-whisper on Apple Silicon, faster-whisper everywhere else, Parakeet when asked for); Groq, OpenAI and ElevenLabs Scribe are available but only when named explicitly, and never without first stating what will be sent and what it will cost. Normalises and silence-trims the audio with ffmpeg before decoding, which measurably beats reaching for a bigger model, and filters the fabricated segments Whisper invents over silence. Also produces a factual meeting-notes document and a PDF. Not for writing prose or summarising text you already have.
kevin-burns/transcribe-summarize · ★ 0 · AI & Automation · score 75

Install: claude install-skill kevin-burns/transcribe-summarize

# transcribe-summarize Audio in, a transcript out, and optionally a set of meeting notes. The default path runs entirely on the user's machine. **Two things make this different from calling Whisper yourself**, and both came out of a day of measurement rather than design: 1. **Whisper invents speech over silence, and a bigger model does not help.** On a real call with 40 s of joining silence, `large-v3` produced an 18-second segment of one word repeated 55 times. This skill filters those out and keeps them, with the metrics that rejected them, in a sidecar so the decision is auditable. 2. **Cleaning the audio beats choosing a model.** Normalising and silence-trimming took invented segments from 2 to 0 and fixed a word both `large-v3` and `turbo` had wrong. It is on by default. ## Setup `ffmpeg` and `ffprobe` must be on PATH — macOS `brew install ffmpeg`, Windows `winget install Gyan.FFmpeg`, Linux `apt install ffmpeg`. Everything else is stdlib. The **backend** brings its own library, and only the one selected, so pick it with `uv run --with`: | backend | platform | `--with` spec | |---|---|---| | `mlx-whisper` — default on Apple Silicon | macOS arm64 only | `'mlx-whisper>=0.4.2'` | | `faster-whisper` — default elsewhere | mac / Windows / Linux | `'faster-whisper>=1.2'` | | `parakeet` — opt-in | Apple Silicon verified | `'parakeet-mlx'` | | `groq`, `openai`, `elevenlabs`, `gemini` — opt-in, **network** | any | nothing to install | If the library is mi