README.md

metadata

title: Vanta
emoji: 🎙
colorFrom: gray
colorTo: gray
sdk: docker
app_port: 7860
pinned: false
license: mit

Vanta — Target Speaker Extraction (backend)

FastAPI inference server for Vanta, a target speaker extraction (TSE) model. Upload a short reference clip of a target speaker and a noisy recording; get back the isolated voice of the target.

Endpoints

• GET /health — returns {status, checkpoint, device}
• POST /extract — multipart/form-data with fields mixture and enrollment. Pass ?include_residue=true to also get the residue (what was removed) as JSON.

Frontend lives separately (Vercel / Next.js) and talks to this API.