README.md

---
title: Ankui Align
emoji: 🎤
colorFrom: indigo
colorTo: blue
sdk: docker
app_port: 7860
pinned: false
short_description: Kazakh forced-alignment backend for the Ankui app
---

# Ankui Align

Forced-alignment microservice for the **Ankui** karaoke-creation app.

- Isolates vocals with torchaudio **HDEMUCS**, then runs **MMS-fl102** (Kazakh
  adapter) CTC forced alignment via `torchaudio.functional.forced_align`.
- Falls back to an energy-based line spread if alignment is degenerate, so it
  never returns fake all-zero timings.

## API (async)

- `GET /` or `/health` → `{"status":"ok","ready":bool}`
- `POST /align` → `202 {"jobId": "..."}` — body `{"lyrics","audio_b64","ext"}`
- `GET /jobs/{id}` → `{"status":"running|done|error","progress",​"lines":[...]}`

Set `ANKUI_ALIGN_KEY` as a Space secret to require an `X-Ankui-Key` header.
Uploaded audio is processed in a temp file and deleted immediately.