---
license: gpl-3.0
library_name: onnx
tags:
  - audio
  - source-separation
  - karaoke
  - onnx
  - webgpu
---

# Stemlight models

ONNX models used by [Stemlight](https://github.com/Thulrus/stemlight), which separates songs into stems in the browser.

## bs-roformer-karaoke.onnx

Splits a vocal track into **lead** and **backing** vocals.

- **Original model:** [anvuew/karaoke_bs_roformer](https://huggingface.co/anvuew/karaoke_bs_roformer) (BS-RoFormer, trained on becruily's lead/backing vocal data)
- **Changes:** exported to ONNX (fp32) with 8-second chunks; some operations were rewritten so it runs on WebGPU via onnxruntime-web. The weights were not changed.
- **Size:** ~210 MB
- **SHA-256:** `7bbf6a275f8468efa2e084f50437ee476faf61ad7de0bd21d8df9c23dea318a5`

## License

GPL-3.0, the same as the original model. All credit for training goes to anvuew.