Loading page
Boot the stack and run your first analysis in about ten minutes.
AudioLens runs as a set of Docker Compose services: a CPU-only FastAPI control plane, one or more Celery workers that actually load models, Redis for sessions and job metadata, and a React frontend. This page gets the full stack running locally and walks through one analysis job.
git clone https://github.com/ECHO-Lit/ECHO-LIT.git AudioLens
cd AudioLens
cp Backend/.env.example Backend/.env
cp Frontend/.env.example Frontend/.env
docker compose up --buildTo pre-warm model weights instead of downloading them lazily on first job:
docker compose run --rm worker-model-local python3 -c \
"from transformers import pipeline; pipeline('automatic-speech-recognition', model='openai/whisper-base')"To tear the stack down: docker compose down (add -v to also wipe volumes).
| Service | URL |
|---|---|
| Frontend | http://localhost:8080 |
| API health check | http://localhost:8000/health |
| Redis | localhost:6379 |
Visit http://localhost:8080. A new anonymous session cookie is set
automatically — there's no sign-up or login.
Start from a built-in dataset — fetch it first with the scripts described
on /docs/datasets — or upload your own .wav, .mp3, .m4a, or .flac file, up to 100 MB and
10 minutes long.
Pick whisper-base (speech-to-text) or wav2vec2 (emotion
classification) from the model dropdown.
Start with a prediction job to see the model's raw output, then try
saliency to see which parts of the waveform drove it. The UI polls the
job until it completes.
The result appears on the shared timeline. Open another panel — Attention or Embeddings, for example — and it stays in sync with the same datapoint.