| name | transcript-generator |
| description | Transcribe a local audio/video file (lecture recording, meeting, podcast) into a text transcript plus SRT subtitles, fully offline using whisper.cpp on Apple Silicon. Use when asked to transcribe an mp4/mov/m4a/wav/mp3, generate a transcript or captions from a media file, or get the text of a recorded lecture. |
| argument-hint | <path-to-media-file> [model] |
Transcript Generator — Local Media → Text + Subtitles
You turn a local audio/video file into a clean .txt transcript and a timestamped .srt
subtitle file, fully offline, using whisper.cpp (Metal-accelerated on Apple Silicon).
Built for macOS + Homebrew. The file size of the video is irrelevant — only the audio
duration matters.
Arguments
- Media file path:
$0 (e.g. ~/Downloads/lecture.mp4). Required. If missing, ask the user.
- Model:
$1 (optional). One of small.en, medium.en (default), large-v3.
small.en (~470 MB) — fast draft, clean English audio.medium.en (~1.5 GB) — default, best balance for lectures in English.large-v3 (~3 GB) — best accuracy, heavy jargon, or non-English / mixed-language
audio (e.g. Portuguese). The .en models are English-only.
Step 0 — Verify input
Run ls -lh "<media>". If the file does not exist, stop and tell the user what you looked for.
Step 1 — Ensure dependencies
Check which ffmpeg whisper-cli. If either is missing:
brew install whisper-cpp ffmpeg
Run this in the background (it can take a few minutes) and wait for it to finish before
continuing. If brew itself is missing, stop and tell the user to install Homebrew.
Step 2 — Ensure the model (cached & reused)
Models live in ~/.cache/whisper/ as ggml-<model>.bin and are reused across runs.
If the chosen model file is missing, download it (in the background, 0.5–3 GB):
mkdir -p ~/.cache/whisper
curl -L -C - -o ~/.cache/whisper/ggml-<model>.bin \
https://huggingface.co/ggerganov/whisper.cpp/resolve/main/ggml-<model>.bin
ggml-medium.en.bin is normally already cached, so the default run skips this step.
Step 3 — Extract audio (16 kHz mono WAV)
whisper.cpp requires a 16 kHz mono PCM WAV:
ffmpeg -y -i "<media>" -ar 16000 -ac 1 -c:a pcm_s16le "<basename>.wav"
For anything longer than a few minutes, run in the background; confirm exit 0 and that
the .wav exists. Optionally report duration with
ffprobe -v error -show_entries format=duration -of default=noprint_wrappers=1:nokey=1 "<basename>.wav".
Step 4 — Transcribe
Run in the background (long for >20 min audio; roughly real-time or faster on M-series):
whisper-cli -m ~/.cache/whisper/ggml-<model>.bin -f "<basename>.wav" \
-otxt -osrt -of "<output-base>" -t 8 -pp
- Produces
<output-base>.txt (clean text) and <output-base>.srt (timestamped).
- Default
<output-base> = the media path without its extension, in the same folder.
Honor an output directory if the user specifies one.
- Monitor progress by reading the background task's output file and looking for
progress = NN% and the latest [hh:mm:ss --> ...] timestamp.
Step 5 — Clean up & report
- Delete the intermediate
.wav (large and disposable). Keep the cached model.
- Never delete the user's original media file unless they explicitly ask.
- Report the output paths. If useful, read the
.txt and give a 3–5 bullet summary of what
the recording covers.
Notes
- Apple GPU (Metal) is used automatically;
-t 8 sets CPU threads for the non-GPU parts.
- This skill was extracted from the MLN601 Week-1 lecture transcription workflow.
- If the user has many recordings to do, keep the model cached and only re-run Steps 3–5.
