kuula (Estonian: "listen!") is local, diarized speech-to-text for Apple Silicon Macs, with first-class Estonian and English support.
Point it at an audio file, a video, or your Apple Voice Memos library and get a clean, speaker-labelled transcript in markdown. Everything runs on-device with MLX and Core ML; no audio, text or telemetry ever leaves your machine, and no accounts or API keys are needed.
Speaker 1 (00:00): Good morning everyone. Let's get started with the quarterly review.
Speaker 2 (00:10): Thanks for having me. I reviewed the numbers last night...
brew install lightheaded/tap/kuula
kuula setupOr with uv (PyPI release pending; install from GitHub):
brew install ffmpeg
uv tool install --python 3.13 git+https://github.com/lightheaded/kuula
kuula setupThen:
kuula memos --latest # transcribe your newest Voice Memo
kuula meeting.m4a # transcribe any audio/video filekuula setup walks you through everything interactive: model downloads, Voice Memos access, and the Finder right-click Quick Action. kuula doctor re-checks the installation any time.
- Private by construction. Recordings of meetings and 1-1s are sensitive. kuula never uploads anything, anywhere.
- Estonian is a first-class citizen. Generic multilingual models garble Estonian and mixed Estonian/English speech. kuula auto-detects the language and routes Estonian audio to TalTech's verbatim Estonian Whisper model, which handles code-switched English embedded in Estonian speech remarkably well.
- Speakers are labelled. Diarization is on by default via senko - no Hugging Face account, no gated models, works offline.
- Voice Memos without friction.
kuula memosreads your Voice Memos library directly (titles, dates, durations), so "transcribe yesterday's 1-1" is one command. No exporting, no dragging files. - Fast. Roughly 8-10x real time for diarized transcription on an M-series Pro - a 30-minute meeting takes about 3 minutes. The opt-in parakeet engine is faster still.
kuula memos # list recordings, newest first
kuula memos --latest # transcribe the newest
kuula memos "standup" # transcribe by title match
kuula memos "standup" --all # ...all matches
kuula transcribe interview.mp3 # any audio or video file
kuula interview.mp3 # shorthand for the above
kuula call.m4a -f srt # subtitle output
kuula call.m4a --stdout -f json # JSON to stdout for scripting
kuula call.m4a -l et # force Estonian
kuula call.m4a -d none # skip diarization (faster)Transcripts from files land next to the file; transcripts from Voice Memos land in ~/Downloads. Override with -o.
Output formats: md (default, with a YAML provenance header - disable with --no-frontmatter), txt, srt, vtt, json.
kuula setup installs a Finder Quick Action: right-click any audio/video file -> Quick Actions -> Transcribe with kuula. Notifications report progress; the transcript appears next to the file. See integrations/quick-action.
For transcribing straight from the Voice Memos app via the share sheet, see integrations/shortcuts.
The repo ships a skill that teaches Claude Code to drive kuula ("transcribe my latest voice memo and summarise the decisions"). Install:
ln -sfn "$(pwd)/skills/kuula" ~/.claude/skills/kuula| Audio | Engine | Model |
|---|---|---|
| Estonian (incl. mixed ET/EN) | mlx-whisper | TalTechNLP/whisper-large-v3-turbo-et-verbatim-2604 (MIT) |
| Everything else | mlx-whisper | mlx-community/whisper-large-v3-turbo (MIT) |
--engine parakeet (opt-in) |
parakeet-mlx | nvidia/parakeet-tdt-0.6b-v3 (CC-BY-4.0) |
Whisper output is lightly normalised (fillers dropped, tidy punctuation), which reads better for meeting notes. Parakeet is faster and fully verbatim - "uh"s and all - covering English and 24 other European languages; reach for it with --engine parakeet when speed or strict verbatimness matters.
With --lang auto (the default) kuula samples a few windows of the audio with a small Whisper model and votes on the language; Estonian wins ties because mixed Estonian/English meetings transcribe far better on the Estonian model than the other way round. Models are downloaded on first use (or ahead of time by kuula setup) and cached; the Estonian checkpoint is converted to MLX format once, locally.
-
senko (default): narcotic-sh/senko, Core ML-accelerated, MIT-licensed, no accounts or tokens. Language-agnostic acoustics.
-
pyannote (optional upgrade): the strongest open diarization, pyannote/speaker-diarization-community-1. The model is free but gated on Hugging Face, so it stays opt-in:
uv tool install --python 3.13 'kuula[pyannote]' # accept the terms at https://huggingface.co/pyannote/speaker-diarization-community-1 # create a read token at https://huggingface.co/settings/tokens export HF_TOKEN=... # your own token; kuula never stores it kuula meeting.m4a --diarizer pyannote --speakers 2
- Apple Silicon Mac (the ASR engines are MLX-based)
- macOS 13+;
kuula memostested against the current Voice Memos database format - Python 3.11-3.13 (
uvhandles this for you), ffmpeg kuula memosneeds your terminal to have Full Disk Access (System Settings -> Privacy & Security);kuula setupguides you there. Transcribing ordinary files needs no special permissions.
input -> ffmpeg (16 kHz mono wav) -> language probe (whisper-base)
-> ASR (parakeet-mlx | mlx-whisper) -> segments
-> diarizer (senko | pyannote) -> speaker turns
-> overlap merge -> md / txt / srt / vtt / json
The Voice Memos integration reads a temporary copy of the library's SQLite database (never the live one, never writing into Apple's folder). If Apple changes the schema, kuula falls back to plain file listing.
kuula itself is MIT. It stands on excellent open work, downloaded at runtime and never redistributed here:
- TalTechNLP Estonian Whisper models (MIT) - Tallinn University of Technology's Laboratory of Language Technology
- nvidia/parakeet-tdt-0.6b-v3 (CC-BY-4.0), via senstella/parakeet-mlx (Apache-2.0)
- senko (MIT), built on the 3D-Speaker pipeline
- mlx-whisper (MIT, Apple);
src/kuula/convert.pyadapts its HF-to-MLX conversion - pyannote (MIT code; the community-1 model weights are CC-BY-4.0, user-gated)
- OpenAI Whisper (MIT), the base of the Whisper family models