RemixIT-TSE

Official model repository for:

RemixIT-TSE: Progressive Synthetic-to-Real Adaptation for Target Speech Extraction via Target-Aware Supervision and Remixing

📄 Paper
🎧 Audio Demo
💻 GitHub

Overview

RemixIT-TSE is a progressive synthetic-to-real adaptation framework for real-world target speech extraction (TSE).

Starting from a TSE model pretrained on synthetic mixtures, the framework progressively transfers the model to real-world conversational recordings through:

1. Synthetic-Data Pretraining (SDP)
2. Target-Aware Adaptation (TAA)
3. RemixIT-TSE Adaptation (RTA)

The adaptation process progressively shifts from:

Synthetic only → Synthetic + Real → Real only

To adapt RemixIT to TSE, RemixIT-TSE introduces two key modifications:

• Quality-aware pseudo-target filtering to retain reliable teacher pseudo-targets.
• Target-only supervision to avoid directly supervising the heterogeneous Non-target component.

Models

This repository provides the following checkpoints:

Model	Description
SDP	Source-domain baseline pretrained on synthetic data
RemixIT-TSE	Final model after progressive synthetic-to-real adaptation

The full training and adaptation pipeline will be released after paper acceptance.

Model Input

The model performs target speech extraction conditioned on:

• Mixture speech
• Target-speaker enrollment utterance

Audio sampling rate:

• 16 kHz

Model Output

The model outputs the estimated speech signal of the target speaker.

Evaluation

The models are evaluated on the SLT 2026 REAL-TSE Challenge evaluation sets.

Set	System	TER ↓	SIM ↑	OVRL ↑	P808 ↑	F1 ↑
EVAL-1	SDP	0.726	0.485	2.049	2.923	0.824
EVAL-1	RemixIT-TSE	0.680	0.533	2.173	3.088	0.837
EVAL-2	SDP	0.763	0.335	1.850	2.710	0.804
EVAL-2	RemixIT-TSE	0.713	0.408	2.040	2.978	0.837
Combined	SDP	0.748	0.395	1.929	2.790	0.812
Combined	RemixIT-TSE	0.700	0.458	2.093	3.022	0.837

On the unseen EVAL-2 set, RemixIT-TSE achieves a 6.53% relative TER reduction, together with relative improvements of 21.84% in speaker similarity, 9.89% in DNSMOS-P808, and 4.10% in target-activity F1 over the source-domain baseline.

Audio Demo

Audio examples are available on the project page:

RemixIT-TSE Audio Demo

The demo includes:

• Mixture
• Enrollment utterance
• SDP
• SAMoM (synthetic clean)
• SAMoM (real single-spk.)
• RemixIT-TSE

Usage

Inference code and detailed usage instructions will be provided with the full code release.

The checkpoints provided in this repository are intended for research and reproducibility purposes.

Citation

If you find this work useful, please consider citing:

@article{RemixIT-TSE2026,
  title={RemixIT-TSE: Progressive Synthetic-to-Real Adaptation for Target Speech Extraction via Target-Aware Supervision and Remixing},
  author={Yu Wang and Haixin Guan and Shuang Wei and Yanhua Long},
  year={2026},
  eprint={2609.35118},
  archivePrefix={arXiv},
  primaryClass={cs.SD},
  url={https://arxiv.org/abs/2609.35118}
}