cohr

Independent project. Not affiliated with or endorsed by Cohere.

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.

Status: v1.0.0. The pipeline, the app and the evals meet the v1 targets (Phase 6e). Recordings in 14 languages (Transcribe's), notes in Command R7B's 23 languages plus "Other (beta)" languages translated by Tiny Aya. See PLAN.md and NOTES.md.

Requirements

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

Setup

uv sync
uv run cohr models download     # the five pipeline models, 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

Pages: New run (upload or link a recording and slides, pick the languages), Notes (cited notes, with a Sources panel of the audio and slides beside or below them), Ask (questions about a run, answered with citations), Runs (status, size, prune, delete).

Command line

uv run cohr run meeting.m4a --slides deck.pdf --target de   # creates runs/<id>/ and processes 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

Settings: COHR_RUNS_DIR (default ./runs), COHR_HF_HOME (model cache location).

Deleting models

Remove one model 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

uv run cohr models lists the five repos. 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 (its dry run lists R7B, Tiny Aya, North Micro Vision and the judge). If you use it for other caches, run hf cache prune --dry-run first. cohr models download fetches anything removed.

Evals

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

• 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.
• 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

Two layers with separate licenses:

• This project's code: Apache-2.0 (LICENSE, NOTICE).
• Model weights: downloaded at runtime, never redistributed, 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.