Local, offline transcription for macOS 14+ on Apple Silicon. Turn audio/video or a YouTube URL into a readable transcript with optional speaker diarization.
uv tool install git+https://github.com/speech115/transcribe.gitRequires ffmpeg; install yt-dlp for YouTube sources.
Verify the installed executable and print the bundled agent skill with:
transcribe doctor
transcribe skilltranscribe call.m4a --speakers autoThe command prepares audio, runs vendored FluidAudio/Parakeet, optionally
diarizes, merges turns, and writes transcript.md, transcript.json, and
manifest.json (plus requested SRT/VTT subtitles).
Streaming diarization runs alongside ASR. Offline diarization runs after ASR.
Language and speaker count are automatic by default. Use --lang,
--speakers, --out, --out-root, and --formats when needed.
SRT/VTT cues use up to 80 text characters and 6 seconds; a single word keeps
its original timing. Existing transcript artifacts require --overwrite.
Artifacts are prepared in a temporary subdirectory of the output before publication, with
rollback on failure. The output directory and unrelated files are left in
place. A forced process kill during publication can leave a partial result;
previous artifacts remain in the output's .transcribe-*/previous directory.
Processing is local after the model cache is present. Media is not uploaded; network access is used only for the initial model cache and YouTube download.
v0.5.0 · single-maintainer project · Apache-2.0