# AlphaFlowTSE: One-Step Generative Target Speaker Extraction via Conditional AlphaFlow

Source: [AlphaFlowTSE: One-Step Generative Target Speaker Extraction via Conditional AlphaFlow](<https://www.sota2.com/research/paper/69b248b2d030084c22e2fac5>)

Collection index: [llms.txt](<https://www.sota2.com/research/llms.txt>)

Data updated: 2026-03-13T05:00:04.336Z

Authors: Duojia Li, Shuhan Zhang, Zihan Qian, Wenxuan Wu, Shuai Wang, Qingyang Hong, Lin Li, Haizhou Li

Published: 2026-03-11T12:14:25.000Z

## Abstract

In target speaker extraction (TSE), we aim to recover target speech from a multi-talker mixture using a short enrollment utterance as reference. Recent studies on diffusion and flow-matching generators have improved target-speech fidelity. However, multi-step sampling increases latency, and one-step solutions often rely on a mixture-dependent time coordinate that can be unreliable for real-world conversations. We present AlphaFlowTSE, a one-step conditional generative model trained with a Jacobian-vector product (JVP)-free AlphaFlow objective. AlphaFlowTSE learns mean-velocity transport along a mixture-to-target trajectory starting from the observed mixture, eliminating auxiliary mixing-ratio prediction, and stabilizes training by combining flow matching with an interval-consistency teacher-student target. Experiments on Libri2Mix and REAL-T confirm that AlphaFlowTSE improves target-speaker similarity and real-mixture generalization for downstream automatic speech recognition (ASR).

## Recorded benchmark results

A research result applies to its stated dataset, split, metric and evaluation configuration. Compare like-for-like settings, preserve the metric's unit and direction, and distinguish paper-reported results from independent reproduction. The best result recorded here is not a claim of exhaustive, current global coverage.

- [Target Speaker Extraction on Libri2Mix Clean (min 16 kHz)](<https://www.sota2.com/research/sota/target-speaker-extraction-on-libri2mix-clean-min-16-khz.md>): Target Speaker Extraction; Libri2Mix Clean min 16 kHz; PESQ: 3.27.

- [Target Speaker Extraction on Libri2Mix Noisy min 16 kHz](<https://www.sota2.com/research/sota/target-speaker-extraction-on-libri2mix-noisy-min-16-khz.md>): Target Speaker Extraction; Libri2Mix Noisy min 16 kHz; PESQ: 2.28.

- [Target Speaker Extraction on REAL-T DipCo English](<https://www.sota2.com/research/sota/target-speaker-extraction-on-real-t-dipco-english.md>): Target Speaker Extraction; REAL-T DipCo English; DNSMOS OVRL: 1.56.

- [Target Speaker Extraction on REAL-T AISHELL-4 Chinese](<https://www.sota2.com/research/sota/target-speaker-extraction-on-real-t-aishell-4-chinese.md>): Target Speaker Extraction; REAL-T AISHELL-4 Chinese; DNSMOS OVRL: 2.277.

- [Target Speaker Extraction on REAL-T AliMeeting Chinese](<https://www.sota2.com/research/sota/target-speaker-extraction-on-real-t-alimeeting-chinese.md>): Target Speaker Extraction; REAL-T AliMeeting Chinese; DNSMOS OVRL: 2.086.

- [Target Speaker Extraction on REAL-T AMI English](<https://www.sota2.com/research/sota/target-speaker-extraction-on-real-t-ami-english.md>): Target Speaker Extraction; REAL-T AMI English; DNSMOS OVRL: 2.169.

- [Target Speaker Extraction on REAL-T CHiME-6 English subset](<https://www.sota2.com/research/sota/target-speaker-extraction-on-real-t-chime-6-english-subset.md>): Target Speaker Extraction; REAL-T CHiME-6 English subset; DNSMOS OVRL: 1.858.

## Original sources and implementation

[Original paper](<https://arxiv.org/abs/2603.10701>)

[PDF](<https://arxiv.org/pdf/2603.10701.pdf>)
