Problem
Research questions and friction points this paper is trying to address.
target speaker extraction
real-world TSE
acoustic mismatch
conversational overlap
reverberation and noise
The SonicAGI System for the REAL-TSE Challenge
🤖 AI Summary
This work addresses acoustic mismatch in real-world meeting scenarios caused by reverberation, noise, and irregular speaker overlaps by proposing a data-centric solution that combines simulated and real overlapping speech data, augmented with auxiliary supervision from a frozen offline enhancer. For online and offline enhancement tasks, the authors design two distinct architectures: SwiftNet-Lookahead, which achieves ultra-low latency of only 96 ms, and USEF-TFGridNet, which leverages frame-level registered cross-attention to jointly optimize perceptual quality and speaker fidelity during magnitude-domain fusion. Evaluated in the CHiME-7 DNS Challenge, the proposed systems demonstrate strong performance—SwiftNet-Lookahead secured second place in Track 1, while USEF-TFGridNet ranked fifth in Track 2—both significantly outperforming the respective baselines.
📝 Abstract
Real-world target speaker extraction (TSE) remains challenging because target speech, interference, and enrollment are recorded under mismatched acoustic conditions with reverberation, noise, and irregular conversational overlap. This paper describes the SonicAGI submission to the REAL-TSE Challenge (IEEE SLT 2026). We take a data-centric approach that combines fully simulated mixtures from clean speech with real meeting overlaps, and use a frozen offline enhancer to provide a denoised mirror of real targets for auxiliary supervision. For the online track, we introduce SwiftNet-Lookahead, which inserts a single bounded-lookahead module before a strictly causal iterative separator and keeps the total system latency at 96 ms. For the offline track, we use a frame-level enrollment cross-attention USEF-TFGridNet with a magnitude-domain fusion stage that trades off perceptual quality and speaker fidelity. In the official evaluation, SwiftNet-Lookahead ranks second in Track~1 and USEF-TFGridNet ranks fifth in Track~2, both exceeding the challenge baselines. These results suggest that real-data-oriented training and track-specific modeling are effective for conversational TSE.
Innovation
Methods, ideas, or system contributions that make the work stand out.
target speaker extraction
data-centric training
bounded-lookahead
cross-attention
magnitude-domain fusion
🔎 Similar Papers
No similar papers found.
💼 Related Jobs
