Docker

The image includes everything: CUDA 12.8 and cuDNN 9, the server (Python 3.12), the OmniVoice voice engine in its own environment (Python 3.11 + torch), ffmpeg and Node.

Requirements

An NVIDIA GPU with an up-to-date driver.
Linux: Docker with the NVIDIA Container Toolkit.
Windows: Docker Desktop with WSL 2 (the GPU is passed to the container without installing anything else).

With Docker Compose

Clone the repository:

git clone https://github.com/alenj0x1/translay.git
cd translay

Set the owner password:
Terminal window
```
cp .env.example .env
```
and change TRANSLAY_OWNER_PASSWORD in .env.
Build and start:
Terminal window
```
docker compose up -d --build
```
Open http://localhost:8765 and sign in with that password.

The first time, it downloads the models (~6 GB) into the modelos volume; the container shows as healthy once the server responds. To follow progress:

docker compose logs -f

Why the password is needed

Outside Docker, anyone connecting from the same machine (127.0.0.1) is the owner without signing in. Inside the container, connections arrive from the Docker network, so nobody counts as local: without a password nobody could administer the app. TRANSLAY_OWNER_PASSWORD (or --owner-password) sets it at startup. More in Access.

Volumes

Path in the container	What it stores
`/app/data`	Your library, transcripts, voices, queue, preferences and keys (in the compose file, `./data`).
`/models`	Hugging Face models (`HF_HOME`). Keeping it avoids downloading them again.

You can mount the same data/ folder used by the local installation: the format is the same.

Without Compose

docker build -t translay .
docker run -d --gpus all -p 8765:8765 \
  -e TRANSLAY_OWNER_PASSWORD=tu-contraseña \
  -v "$PWD/data:/app/data" -v translay-modelos:/models \
  --name translay translay

The startup options go after the image name; they replace the defaults (--host 0.0.0.0 --no-browser), so include them:

docker run … translay --host 0.0.0.0 --no-browser --translator nllb-1.3b