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

2026-07-16

AI Summary

This work presents the first target speaker extraction (TSE) challenge specifically designed for realistic conversational scenarios, aiming to recover clean speech of a target speaker from complex audio recordings containing overlapping speech, reverberation, background noise, and dynamic multilingual (Mandarin–English) interactions, given only a short enrollment utterance of the target speaker. The system employs a speaker embedding–guided end-to-end separation architecture capable of both low-latency streaming inference and full-context modeling. Comprehensive evaluation using multiple metrics—including target estimation ratio (TER), speaker similarity (SpkSim), DNSMOS, and F1 score—not only establishes a new benchmark for TSE in real-world conditions but also offers key technical insights into robust speech extraction under challenging acoustic environments.

Abstract

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.

Problem

• target speaker extraction
• real-world conversational recordings
• multi-speaker mixture
• speaker separation
• speech extraction

Innovation

• Target Speaker Extraction
• Real-world Conversational Recordings
• Online/Offline TSE Tracks
• Multi-dimensional Evaluation Metrics
• Natural Speech Overlap