Docker container for Ukrainian Speech to Text inference Gradio app

Repository: https://github.com/ALERTua/stt_ukrainian_docker

GitHub Docker Registry: https://github.com/ALERTua/stt_ukrainian_docker/pkgs/container/stt_ukrainian_docker

Docker Hub: https://hub.docker.com/r/alertua/stt_ukrainian_docker

Description

Docker image for a gradio app that produces text from Ukrainian speech.

Used with https://github.com/ALERTua/stt-ukrainian-api to provide an OpenAI STT API endpoint to use it with Home Assistant.

I'll try to base this image on the most modern and effective SST model. This image is currently based on: Yehor/w2v-bert-uk-v2.1

Deployment

The best way is to use the docker-compose.yml

Or run directly with Docker:

docker run -d \
  -p 7860:7860 \
  -v ./docker_volumes/stt/data:/data \
  --name stt_ukrainian \
  ghcr.io/alertua/stt_ukrainian_docker:latest

Gradio Web UI

You can access the Gradio Web UI at http://{container_ip}:7860

/data volume structure

After the first run the data directory will look like this:

• .cache - contains models downloaded from HuggingFace Hub. ~2.4GB
• uv_cache - cache for installing prerequisites ~7.2gb
• venv - working environment ~7.4gb

Resources usage

• tag latest uses ~3 GiB of RAM

Things to do that I have no knowledge on (help appreciated)

• Make this use less RAM

Things to do that depend on the author's code

• dummy

Caveats

• The first start is slow as the models are downloaded and the prerequisites get installed.

• If you need a specific torch version, you can execute inside the running container:

  E.g. torch for my GTX1080ti

  cd /data
  source venv/bin/activate
  uv pip install torch torchaudio --index-url https://download.pytorch.org/whl/cu118 --force-reinstall

  Then restart the container.

  You can also execute this outside the container within the mounted virtual environment.