USEF-Laura-TSE
This repository provides the official model checkpoints for Discriminative–Generative Target Speaker Extraction with Decoder-Only Language Models.
The paper is currently under review. At this stage, we release the model checkpoints and inference code only. The training code will be made publicly available after the review process is complete.
Model Description
USEF-Laura-TSE is a two-stage target speaker extraction system that combines:
- a discriminative USEF-TFGridNet front-end for extracting the target speaker from a speech mixture; and
- a Laura-based generative back-end built with a decoder-only language model for improving perceptual quality and speech naturalness.
The generative back-end supports both autoregressive (AR) and non-autoregressive (NAR) inference. NAR inference additionally supports a configurable ratio between 0.0 and 1.0.
Model Files
| File | Description |
|---|---|
usef_laura_tse.pth | Complete USEF-Laura-TSE checkpoint, including the generative model and integrated front-end parameters. |
usef_front_end.pth | USEF-TFGridNet front-end checkpoint used to initialize the model structure. |
The FunCodec model and configuration required by the generative back-end are downloaded separately by the inference release.
Installation
Clone the official inference repository and install its dependencies:
git clone https://github.com/ZBang/USEF-Laura-TSE.git
cd USEF-Laura-TSE
python -m pip install -r requirements.txt
Download the released checkpoints:
python scripts/download_models.py
Inference
Autoregressive inference:
bash infer.sh --mode ar --output both
Non-autoregressive inference:
bash infer.sh --mode nar --ratio 1.0 --output both
The --output option supports front, generative, and both. See the GitHub repository for manifest formats, checkpoint paths, and complete usage instructions.
License
This model is released under the Creative Commons Attribution-NonCommercial 4.0 International License. Third-party components and models remain subject to their original licenses.