Analysis Service

The analysis service is a micro-service for Echolalia. It launches long-running tasks, including:

1. The voice type classifier (VTC) [1][2] - note that different major versions are available, but at present (Dec 29, 2025) we use version 1
2. ChildProject's acoustics pipeline [3]

Setting Up

Note that we use Swarm as our orchestrator, but that Swarm is quite opinionated in how you use Docker and Docker Compose. The deployment workflow is therefore a little different from the Docker Compose (development) workflow.

Deployment

Please install the NVIDIA container toolkit

Initialise Docker Swarm (only need to do this once)

docker swarm init

Check your node name

docker node ls

And update

docker node update --label-add gpu=true <node-name>

Fill in the .env file and export into the shell

export $(cat .env | xargs)

Add the required images (TODO: add automated deployment to this repo)

docker build -t me/echoswarm-alice:latest ./alice-worker
docker build -t me/echoswarm-vtc:latest ./vtc-worker
docker build -t me/echoswarm-vtc-2:latest ./vtc-2-worker
docker build -t me/echoswarm-w2v2:latest ./w2v2-worker
docker build -t me/echoswarm-acoustics:latest ./acoustics-worker
docker build -t me/echoswarm-daemon:latest ./daemon

Then start with

docker stack deploy -c docker-compose.yml swarm-service

Useful Swarm Tips

# check status
docker service ls

# tasks (containers)
docker service ps echoswarm-alice

# logs
docker service logs -f echoswarm-alice

# shutdown
docker stack rm swarm-service

Development

For development, the .env.test file has already been filled in. Development may be hindered by not using a x86_64 architecture, or not being on Linux, and I suggest getting your hands on a VM if possible. This is a limitation of many of the models we run.

For development Docker Compose is the right tool. The main difference with Docker Swarm is that containers are not restarted automatically once they're brought down. Docker Swarm-compatible compose files are quite restrictive, and so we use docker-compose.dev.yml for development. Note that a model runner always shuts down after running a task to completion.

To run all the test containers

docker compose --env-file .env.test -f docker-compose.dev.yml up

References

[1] T. Charlot, T. Kunze, M. Poli, A. Cristia, E. Dupoux, and M. Lavechin, ‘BabyHuBERT: Multilingual Self-Supervised Learning for Segmenting Speakers in Child-Centered Long-Form Recordings’, arXiv [eess.AS]. 2025.

[2] M. Lavechin, R. Bousbib, H. Bredin, E. Dupoux, and A. Cristia, ‘An open-source voice type classifier for child-centered daylong recordings’, in Interspeech, 2020.

[3] L. Gautheron, N. Rochat, and A. Cristia, ‘Managing, storing, and sharing long-form recordings and their annotations’, Language Resources and Evaluation, Feb. 2022.