[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"doc-detail-86534-en":3,"doc-seo-86534-105":29,"detail-sidebar-cat-0-en-105":90},{"code":4,"msg":5,"data":6},0,"success",{"doc_id":7,"user_id":8,"nickname":9,"user_avatar":10,"doc_module":4,"category_id":11,"category_name":12,"doc_title":13,"doc_description":14,"doc_content":15,"file_id":16,"file_url":17,"file_type":18,"file_size":19,"view_count":20,"is_deleted":4,"is_public":20,"is_downloadable":20,"audit_status":20,"page_count":21,"language":22,"language_code":23,"site_id":24,"html_lang":23,"table_of_contents":25,"faqs":26,"seo_title":13,"seo_description":14,"update_tm":27,"read_time":28},86534,962075114765,"Quinn","https://ap-avatar.wpscdn.com/davatar_a8503ba1806abce46bf441b54a3ca4cd",6,"Technology","The SonicAGI System for the REAL-TSE Challenge","Real-world target speaker extraction (TSE) is difficult because target speech, interference, and enrollment are recorded under mismatched acoustic conditions with reverberation, noise, and irregular conversational overlap. The paper presents the SonicAGI submission to the REAL-TSE Challenge (IEEE SLT 2026). A data-centric pipeline combines fully simulated mixtures with real meeting overlaps and uses a frozen offline enhancer for auxiliary supervision. SwiftNet-Lookahead targets low-latency online processing at 96 ms, while USEF-TFGridNet with magnitude-domain fusion is used offline. Results rank second in Track 1 and fifth in Track 2, exceeding baselines, indicating real-data-oriented training and track-specific modeling effectiveness.","The SonicAGI System for the REAL-TSE Challenge  \nKai Li 1,2,†, Wendi Sang 1,†, Jintao Cheng 1 , Xiaolin Hu 1,2,3,∗  \n1Department of Computer Science and Technology, Institute for AI, BNRist, Tsinghua University, Beijing, China 2IDG/McGovern Institute for Brain Research, Tsinghua University, Beijing, China 3Chinese Institute for Brain Research (CIBR), Beijing, China †These authors contributed equally to this work. ∗ Corresponding author.  \narXiv :2607 . 1 1083v 1 [ cs . SD] 13 Jul 2026  \nAbstract—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-dataoriented training and track-specific modeling are effective for conversational TSE.  \nIndex Terms—target speaker extraction, speech separation, low latency, causal modeling, data simulation, conversational speech  \nI. INTRODUCTION  \nTarget speaker extraction (TSE) extracts a designated speaker from a mixture using an enrollment utterance [1] . It is a key front-end for hearing aids, meeting transcription, and speech interaction. Although recent speech and query-based separation methods perform well on curated or synthetic mixtures [2]–[4], real conversational recordings contain reverberation, background noise, irregular overlap, and crossscene enrollment. These factors expose a persistent mismatch between synthetic training and real evaluation conditions, making data realism a central bottleneck.  \nThe REAL-TSE Challenge [5] evaluates this setting on real Mandarin and English meeting and dinner-party recordings, scored by transcription error, target-speaker timing, speaker similarity, and perceptual quality. Track 1 requires online processing below 100 ms algorithmic latency and verifies causality through input-perturbation tests, whereas Track 2 allows full-context offline processing. Development data are held-out partitions of AISHELL-4 [6], AliMeeting [7], AMI [8], DipCo [9], and CHiME-6 [10]; their development and test splits are excluded from training.  \nOur SonicAGI submission is built around three components. First, we construct training data with a two-route sampler: Fully Simulated Mixing provides controlled speaker overlap from clean speech, while Real-recording Mixing samples overlapping utterances from real meetings; a frozen offline enhancer further produces a cleaned mirror of real targets for auxiliary supervision. Second, for Track 1, we propose SwiftNet-Lookahead, which inserts a single chunked bidirectional LSTM lookahead module before a strictly causal SwiftNet separator, so that future context is available within the latency budget but does not accumulate across iterative updates. Third, for Track 2, we use a frame-level enrollment cross-attention USEFTFGridNet and apply magnitude-domain fusion between the extractor output and its enhanced version, improving perceptual quality while  \npreserving target-speaker cues. The resulting systems rank second in Track 1 and fifth in Track 2, both outperforming the challenge baselines.  \nII. DA","cbCailwnwqIcso9i","https://ap.wps.com/l/cbCailwnwqIcso9i","pdf",235612,1,4,"English","en",105,"# I. Introduction\n# II. Data Construction\n## A. Source corpora\n## B. Fully Simulated Mixing route","[{\"question\":\"Why is real-world target speaker extraction (TSE) still challenging?\",\"answer\":\"Target and enrollment are recorded under mismatched acoustic conditions with reverberation and noise, and conversational overlap is irregular, creating a persistent mismatch between training and evaluation.\"},{\"question\":\"What data-centric strategy does SonicAGI use for training?\",\"answer\":\"It combines fully simulated mixtures from clean speech with real-recording mixing that remixes segments from real meetings, and it uses a frozen offline enhancer to generate a denoised mirror of real targets for auxiliary supervision.\"},{\"question\":\"How are online and offline tracks modeled differently?\",\"answer\":\"For the online track, SwiftNet-Lookahead adds a bounded lookahead module before a strictly causal iterative separator to keep latency at 96 ms. For the offline track, USEF-TFGridNet with frame-level enrollment cross-attention applies magnitude-domain fusion between the extractor output and its enhanced version.\"}]",1784212457,10,{"code":4,"msg":30,"data":31},"ok",{"site_id":24,"language":23,"slug":32,"title":13,"keywords":33,"description":14,"schema_data":34,"social_meta":85,"head_meta":87,"extra_data":89,"updated_unix":27},"the-sonicagi-system-for-the-real-tse-challenge","",{"@graph":35,"@context":84},[36,52,67],{"@type":37,"itemListElement":38},"BreadcrumbList",[39,43,47,50],{"item":40,"name":41,"@type":42,"position":20},"https://docshare.wps.com","Home","ListItem",{"item":44,"name":45,"@type":42,"position":46},"https://docshare.wps.com/document/","Document",2,{"item":48,"name":12,"@type":42,"position":49},"https://docshare.wps.com/document/technology/",3,{"item":51,"name":13,"@type":42,"position":21},"https://docshare.wps.com/document/the-sonicagi-system-for-the-real-tse-challenge/86534/",{"url":51,"name":13,"@type":53,"author":54,"headline":13,"publisher":56,"fileFormat":59,"inLanguage":23,"description":14,"dateModified":60,"datePublished":61,"encodingFormat":59,"isAccessibleForFree":62,"interactionStatistic":63},"DigitalDocument",{"name":9,"@type":55},"Person",{"url":40,"name":57,"@type":58},"DocShare","Organization","application/pdf","2026-07-17","2026-07-16",true,{"@type":64,"interactionType":65,"userInteractionCount":20},"InteractionCounter",{"@type":66},"ViewAction",{"@type":68,"mainEntity":69},"FAQPage",[70,76,80],{"name":71,"@type":72,"acceptedAnswer":73},"Why is real-world target speaker extraction (TSE) still challenging?","Question",{"text":74,"@type":75},"Target and enrollment are recorded under mismatched acoustic conditions with reverberation and noise, and conversational overlap is irregular, creating a persistent mismatch between training and evaluation.","Answer",{"name":77,"@type":72,"acceptedAnswer":78},"What data-centric strategy does SonicAGI use for training?",{"text":79,"@type":75},"It combines fully simulated mixtures from clean speech with real-recording mixing that remixes segments from real meetings, and it uses a frozen offline enhancer to generate a denoised mirror of real targets for auxiliary supervision.",{"name":81,"@type":72,"acceptedAnswer":82},"How are online and offline tracks modeled differently?",{"text":83,"@type":75},"For the online track, SwiftNet-Lookahead adds a bounded lookahead module before a strictly causal iterative separator to keep latency at 96 ms. For the offline track, USEF-TFGridNet with frame-level enrollment cross-attention applies magnitude-domain fusion between the extractor output and its enhanced version.","https://schema.org",{"og:url":51,"og:type":86,"og:title":13,"og:site_name":57,"og:description":14},"article",{"robots":88,"canonical":51},"index,follow",{"doc_id":7,"site_id":24},{"code":4,"msg":5,"data":91},[92,96,100,104,109,112,117,122,127,130,133],{"id":20,"doc_module":4,"doc_module_name":45,"category_name":93,"show_sort_weight":94,"slug":95},"Story & Novel",90,"story-novel",{"id":46,"doc_module":4,"doc_module_name":45,"category_name":97,"show_sort_weight":98,"slug":99},"Literature",80,"literature",{"id":21,"doc_module":4,"doc_module_name":45,"category_name":101,"show_sort_weight":102,"slug":103},"Exam",70,"exam",{"id":105,"doc_module":4,"doc_module_name":45,"category_name":106,"show_sort_weight":107,"slug":108},5,"Comic",60,"comic",{"id":11,"doc_module":4,"doc_module_name":45,"category_name":12,"show_sort_weight":110,"slug":111},50,"technology",{"id":113,"doc_module":4,"doc_module_name":45,"category_name":114,"show_sort_weight":115,"slug":116},7,"Healthcare",40,"healthcare",{"id":118,"doc_module":4,"doc_module_name":45,"category_name":119,"show_sort_weight":120,"slug":121},8,"Research & Report",30,"research-report",{"id":123,"doc_module":4,"doc_module_name":45,"category_name":124,"show_sort_weight":125,"slug":126},9,"Religion & Spirituality",20,"religion-spirituality",{"id":125,"doc_module":4,"doc_module_name":45,"category_name":128,"show_sort_weight":125,"slug":129},"World Cup","world-cup",{"id":28,"doc_module":4,"doc_module_name":45,"category_name":131,"show_sort_weight":28,"slug":132},"Lifestyle","lifestyle",{"id":134,"doc_module":4,"doc_module_name":45,"category_name":135,"show_sort_weight":105,"slug":136},19,"General","general"]