arXiv subjects

Zihan Qian

Publications and source records attributed to Zihan Qian.

3 recordsLinked to original sources

arXiv2026

SLT 2026 REAL-TSE Challenge: Real-world Target Speaker Extraction from Conversational Recordings

We introduce the REAL-TSE Challenge, an IEEE SLT 2026 satellite challenge on target speaker extraction~(TSE) from real conversational recordings. Given a multi-speaker mixture and one or more enrollment utterances from a target speaker, participating systems must recover only the target speech. Unlike simulated read-speech benchmarks, REAL-TSE evaluates Mandarin and English recordings that contain natural overlap, reverberation, noise, channel mismatch, and conversational dynamics. The challenge defines two complementary tracks: an Online track for low-latency streaming extraction and an Offline track for full-context processing. Systems are evaluated with Token Error Rate (TER), Speaker Similarity (SpkSim), DNSMOS, and target-speaker activity F1. This overview paper describes the task definition, datasets, baselines, evaluation protocol, submitted systems, condition-wise findings, and lessons for future real-world TSE benchmarks.

eess.AS↗

arXiv2026

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

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).

cs.SD↗

arXiv2024

A Novel Stratified Analysis Method for Testing and Estimating Overall Treatment Effects on Time-to-Event Outcomes Using Average Hazard with Survival Weight

Given the limitations of using the Cox hazard ratio to summarize the magnitude of the treatment effect, alternative measures that do not have these limitations are gaining attention. One of the recently proposed alternative methods uses the average hazard with survival weight (AH). This population quantity can be interpreted as the average intensity of the event occurrence in a given time window that does not involve study-specific censoring. Inference procedures for the ratio of AH and difference in AH have already been proposed in simple randomized controlled trial settings to compare two groups. However, methods with stratification factors have not been well discussed, although stratified analysis is often used in practice to adjust for confounding factors and increase the power to detect a between-group difference. The conventional stratified analysis or meta-analysis approach, which integrates stratum-specific treatment effects using an optimal weight, directly applies to the ratio of AH and difference in AH. However, this conventional approach has significant limitations similar to the Cochran-Mantel-Haenszel method for a binary outcome and the stratified Cox procedure for a time-to-event outcome. To address this, we propose a new stratified analysis method for AH using standardization. With the proposed method, one can summarize the between-group treatment effect in both absolute difference and relative terms, adjusting for stratification factors. This can be a valuable alternative to the traditional stratified Cox procedure to estimate and report the magnitude of the treatment effect on time-to-event outcomes using hazard.

stat.ME↗