Cactus — local meeting protocol (Track 8)

This is the team's local Track 8 app. It has been run on an Apple Silicon Mac with macOS and Python 3.12. Use an Apple Silicon Mac with enough memory and disk for the local models; Intel Macs and other platforms have not been tested. Install FFmpeg (including ffprobe) and have curl and tar available. The first setup needs internet access to install dependencies and download models. Processing uses local models; meeting audio and transcripts stay on the Mac.

One-time setup

From a clone of this repository:

cd ~/Downloads/track8
python3.12 -m venv .venv
.venv/bin/python -m pip install -r requirements.txt
mkdir -p models/whisper models/diarization models/ollama-runtime
.venv/bin/python - <<'PY'
from huggingface_hub import snapshot_download
snapshot_download(
    repo_id="mlx-community/whisper-large-v3-turbo-q4",
    revision="660c343bbf4e52ac257f0b7d952e5388e6f93bef",
    local_dir="models/whisper",
)
snapshot_download(
    repo_id="pyannote-community/speaker-diarization-community-1",
    revision="8a527374977391da736e0daaef26855d949d9685",
    local_dir="models/diarization",
)
PY
curl -fL https://github.com/ollama/ollama/releases/download/v0.34.3/ollama-darwin.tgz -o /private/tmp/track8-ollama.tgz
tar -xzf /private/tmp/track8-ollama.tgz -C models/ollama-runtime

Start Ollama in one terminal, from the repository root:

OLLAMA_HOST=127.0.0.1:11434 OLLAMA_NO_CLOUD=1 \
OLLAMA_MODELS="$PWD/models/ollama" models/ollama-runtime/ollama serve

In another terminal, download the extraction model once:

cd ~/Downloads/track8
OLLAMA_HOST=127.0.0.1:11434 OLLAMA_MODELS="$PWD/models/ollama" \
models/ollama-runtime/ollama pull qwen2.5:7b

Launch and export a protocol

Keep Ollama running. In the second terminal:

cd ~/Downloads/track8
export ASR_MODEL="$PWD/models/whisper"
export DIARIZATION_MODEL="$PWD/models/diarization"
export OLLAMA_MODEL=qwen2.5:7b
export OMP_NUM_THREADS=1
export VECLIB_MAXIMUM_THREADS=1
.venv/bin/python -m streamlit run app.py --server.address 127.0.0.1 --server.port 8501

Open http://127.0.0.1:8501. Upload a local MP3, WAV, or M4A recording. Click “1. Распознать запись / принять транскрипт”; correct speaker labels if needed. Click “2. Извлечь поручения локально”, review or edit the actions, then click “Скачать протокол DOCX”. The browser downloads the generated document; do not commit recordings, transcripts, or generated protocols. For batch processing and local output checks, see LOCAL_RUN.md.