---
title: UniSE Speech Enhancement
emoji: 🔊
colorFrom: indigo
colorTo: blue
sdk: gradio
sdk_version: 6.15.1
app_file: app.py
short_description: Unified AR-LM-based speech enhancement & separation
python_version: "3.12"
startup_duration_timeout: 30m
---

# UniSE: Unified Speech Enhancement

This Space runs the **UniSE** model from the paper "UniSE: A Unified Framework for Decoder-only Autoregressive LM-Based Speech Enhancement" (arXiv:2510.20441).

## Supported Tasks

- **Speech Enhancement (SE)**: General-purpose denoising and clarity improvement (noise, reverb, packet loss).
- **Target Speaker Extraction (TSE)**: Extract a target speaker from a mixture using reference enrollment audio.
- **Speech Separation (SS)**: Separate mixed speakers into individual streams.

## How it works

1. Extract hidden states from WavLM layers and obtain a unified representation by averaging across layers.
2. Use the decoder-only autoregressive language model to predict speech tokens autoregressively.
3. Decode the tokens into audio using the BiCodec neural audio codec.

## References

- 📄 [Paper (arXiv:2510.20441)](https://arxiv.org/abs/2510.20441)
- 🤗 [Model on Hugging Face](https://huggingface.co/QuarkAudio/QuarkAudio-UniSE)
- 💻 [Source Code on GitHub](https://github.com/alibaba/unified-audio/tree/main/QuarkAudio-UniSE)