---
title: Bytley Stems
emoji: 🎛️
colorFrom: purple
colorTo: pink
sdk: docker
app_port: 7860
pinned: false
---

# Bytley Demucs stem-separation backend

FastAPI service that splits a song into stems with **Demucs (htdemucs)** on the free
Hugging Face CPU tier. Same deployment pattern as the PDF→Word Space.

## Endpoints

- `GET /` — health: `{ok, service, model, maxMb}`
- `POST /separate?mode=stems` — multipart `file` → ZIP of `vocals.mp3, drums.mp3, bass.mp3, other.mp3`
- `POST /separate?mode=karaoke` — → ZIP of `instrumental.mp3, vocals.mp3`

## Deploy (Hugging Face Spaces)

1. huggingface.co → New Space → name `bytley-stems` → SDK: **Docker** → CPU basic (free).
2. Upload `app.py`, `requirements.txt`, `Dockerfile`, this `README.md`.
3. First build takes a while (torch + baked model ≈ 1–2 GB image). Watch Build logs.
4. Health check: `https://<user>-bytley-stems.hf.space/` → `{"ok":true,...}`.
5. Put that URL in `src/consts.ts` → `demucsApi: "https://<user>-bytley-stems.hf.space/separate"`.
6. Optional: Space secret `BYTLEY_TOKEN` + same value in `consts.ts` `demucsToken`.
7. Add the health URL to your cron pinger (cron-job.org / UptimeRobot, every 30 min) so the
   Space never hits the 48 h sleep.

## Notes

- A 3–4 min song takes roughly 4–10 min on the free 2-vCPU tier. One job at a time.
- Uploads are processed in a temp dir and deleted immediately after the response.
- Input capped at 30 MB / 10 minutes of audio.