cohr

Local, offline meeting notes for Apple Silicon. A recording plus slides (or a screen-recording video) becomes notes in which every claim cites its timestamp [mm:ss] or [slide N]. All models run on MLX; nothing leaves your machine.

Independent project. Not affiliated with or endorsed by Cohere.

• Recordings in 14 languages (Cohere Transcribe's).
• Notes in Command R7B's 23 languages, plus "Other (beta)" languages translated by Tiny Aya.
• Ask questions about a meeting and search across all of them, with citations.
• Checked against the meeting: numbers, dates, owners and due dates in the notes are compared with the transcript and slides; uncited claims are flagged, never given a citation.

Current release: v1.11.0. What each release changed: GitHub releases. Design: PLAN.md; progress: NOTES.md.

Requirements

• Hardware: Apple Silicon Mac.
• Software: Python 3.13, uv, and ffmpeg on PATH (brew install ffmpeg).
• Memory: the default resident policy keeps each model loaded once it's used (peak ~17 GB, for 32 GB Macs). On a 16 GB Mac, choose sequential as the Memory policy on the New run page: it loads one model at a time (peak ~6 GB) and writes the same notes.
• Disk: ~14.4 GB for the models, plus each run's media until you prune it.
• Speed (M2 Max, 32 GB): about 6 minutes per hour of audio; slides add the Vision stage.

Setup

uv sync
uv run cohr models download   # the five pipeline models at pinned revisions (~14.4 GB)
uv run cohr models            # what's downloaded, with sizes

cohr models download checks free space first, and resumes an interrupted download when run again. Models go to the Hugging Face cache (~/.cache/huggingface/hub), or to $COHR_HF_HOME/hub when that's set. Only cohr models download uses the network; runs never do.

App

uv run streamlit run app.py

• New run: upload or link a recording and slides, and pick the languages.
• Notes: the cited notes, with a Sources panel of the audio and slides.
• Ask: questions about a run, answered with citations.
• Search: words, names and numbers across every run's transcript, slides and notes, and every meeting's action items by owner.
• Runs: status and size of each run; prune or delete it.

New run	Ask	Search

Screenshots show the eval set's generated meetings (evals/gen), not real recordings.

Command line

uv run cohr run meeting.m4a --slides deck.pdf --target de   # create runs/<id>/ and process it
uv run cohr run --video screen.mp4 --lang es                 # slides from the video's frames
uv run cohr runs                                             # list runs with size and status
uv run cohr resume <run>                                     # re-run from the first incomplete stage
uv run cohr prune <run>                                      # drop media, keep notes and JSON
uv run cohr rm <run>                                         # delete a run
uv run cohr search "budget" --kind note                      # search every run
uv run cohr index rebuild                                    # rebuild runs/index.db

--lang is the meeting's language and --target the notes' (both default to en).

The search index (runs/index.db) is rebuilt from the run files whenever it's missing or out of date, so deleting it is always safe.

Setting	Default	Meaning
COHR_RUNS_DIR	./runs	where runs are stored
COHR_HF_HOME	Hugging Face's cache	where models are stored ($COHR_HF_HOME/hub)

Deleting models

uv run cohr models lists the five model repos. Remove one with hf cache rm (it ships with huggingface_hub; add --cache-dir "$COHR_HF_HOME/hub" if you set COHR_HF_HOME). Run it with --dry-run first:

uv run hf cache rm model/mlx-community/c4ai-command-r7b-12-2024-4bit --dry-run
uv run hf cache rm model/mlx-community/c4ai-command-r7b-12-2024-4bit

cohr models download fetches anything removed.

Warning

Don't use hf cache prune: it removes revisions that no branch points to, and cohr pins exactly such revisions, so it deletes cohr's models (R7B, Tiny Aya, North Micro Vision and the judge). If you use it for other caches, run hf cache prune --dry-run first.

Evals

Quality is measured on meetings with known facts (evals/README.md, PLAN §18):

Set	What it covers	Time
quick	four 5-minute English meetings: clean, noisy, PDF deck, misspoken number	~10 min
lang	notes in German, Arabic, Hindi, Swahili and Bengali; Spanish and Japanese meetings	~10 min
long	one 2.5-hour meeting: two-pass notes and Ask at scale	~1 h
full	quick + lang + FLEURS clips in 14 languages + 5 real AMI meetings; the release check	~1 h

License

Three layers, with separate licenses:

• This project's code: Apache-2.0 (LICENSE, NOTICE).
• Model weights: downloaded at runtime, never redistributed, and governed by their authors' licenses. Cohere Transcribe and North Micro Vision are Apache-2.0, Silero VAD is MIT, and Tiny Aya Global and Command R7B are CC-BY-NC-4.0. With the default models, the app is for personal and research use only.
• Eval data (full only; fetched at eval time, never committed): AMI Meeting Corpus (University of Edinburgh et al.) and FLEURS (Google), both CC BY 4.0.