translation-audio

Translation — audio layer

Audio / video → Whisper transcript → Buthaina's text translation layer.

Built for an RTX 5090 on Vast.ai (large-v3, float16). A 4 GB laptop GPU auto-falls back to small or CPU.

Vast.ai (RTX 5090)

Rent a 5090 image with CUDA 12.8+. Then:

git clone https://huggingface.co/mohameddalii/translation-audio
cd translation-audio
chmod +x scripts/run_vast.sh
./scripts/run_vast.sh

Smoke test after the server is up:

python -m audio sample.wav --transcribe-only
# or
curl -s http://127.0.0.1:8000/health
curl -F "[email protected]" -F "translate=false" http://127.0.0.1:8000/translate

Load test concurrent users (CLI):

python -m audio.loadtest --file sample.wav --users 8 --requests 40 --no-translate

Locust UI (keep uvicorn running in another terminal):

pip install locust
locust -f locustfile.py --host http://127.0.0.1:8000 --web-host 0.0.0.0 --web-port 8089

On Vast Jupyter, try https://<your-jupyter-host>/proxy/8089/. Start with 4 users, spawn rate 1. The GPU serializes ASR, so extra users wait or get 429.

Env vars (already set in scripts/run_vast.sh):

Whisper vs Cohere (restart the server between runs — one backend per process):

python -m audio sample.wav --transcribe-only --backend whisper > whisper.json
python -m audio sample.wav --transcribe-only --backend cohere > cohere.json
python -m audio.compare whisper.json cohere.json

Layout

• audio/ — ASR, ffmpeg video→wav, FastAPI, GPU queue, load test
• shared/contracts.py — payload sent to the text layer
• text/stub.py — stand-in for Buthaina's translator