Sandbox for speech experiments with a FastAPI backend and a Svelte frontend.

Recommended prerequisites:

• uv
• pnpm

cd backend
cp .env.example .env
uv run --all-extras fastapi dev main.py

# optional: skip --all-extras to use only some functionality

• API: http://localhost:8000
• Docs: http://localhost:8000/docs

The forced alignment tool requires a separate service. If you want it, run this in a separate process:

• Repo: https://github.com/stephenmac7/mfa-service
• Expected endpoint: http://localhost:8001

cd frontend
cp .env.example .env
pnpm install
pnpm run dev

This will start the app on http://localhost:5173.

This project uses lazy loading for various speech processing models and encoders. Dependencies are loaded on-demand when specific endpoints are first accessed, rather than at application startup. This improves startup time but means:

• First requests to each endpoint may be slower as models are loaded into memory
• Additional dependencies may be required depending on which features you use
• Some imports will fail until you install the necessary optional dependencies for the encoder/tokenizer you want to use

Note on Kanade: the kanade extra should be enough here. We rely on PyTorch's SDPA path in the pinned torch line for the attention optimizations we want.

(Instructions from Svelte -- have not tried this yet)

To create a production version of your app:

npm run build

You can preview the production build with npm run preview.

To deploy your app, you may need to install an adapter for your target environment.