IEEE SLT 2026 Challenge Aims to Extract Target Speakers from Real-World Chaos

The IEEE SLT 2026 REAL-TSE Challenge, announced in an arXiv paper on July 16, 2026, sets a new standard for target speaker extraction (TSE). Traditional TSE benchmarks rely on simulated read speech with clean conditions, but this challenge tests systems on genuine conversational recordings in Mandarin and English. The data includes natural speech overlap, reverberation, environmental noise, channel mismatches, and conversational dynamics — factors that destroy the performance of current models.

Participants must extract only the speech of a target speaker given a multi-speaker mixture and one or more enrollment utterances. The challenge defines two complementary tracks: an Online track for low-latency streaming extraction and an Offline track for full-context processing. Evaluation uses four metrics: Token Error Rate (TER), Speaker Similarity (SpkSim), DNSMOS (a perceptual quality metric), and target-speaker activity F1 score. The paper details the dataset creation, baseline systems, submitted model architectures, and condition-wise performance breakdowns.

Early results highlight that even state-of-the-art systems struggle with severe overlap and domain mismatch between enrollment and target utterances. The challenge also reveals that streaming models sacrifice accuracy for latency, while offline models show better overall quality but remain far from human-level extraction. The authors provide a set of lessons and recommendations for future real-world TSE benchmarks, including better simulation of natural conversation dynamics and smarter enrollment strategies.