[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"doc-detail-82801-en":3,"doc-seo-82801-105":29,"detail-sidebar-cat-0-en-105":91},{"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},82801,4398048950312,"Violet","https://ap-avatar.wpscdn.com/avatar/400002538284de19e3c?_k=1778320343897328908",8,"Research & Report","MOSAIC: Interpretable Multi-Token Cross-Attention of Biophonetic and Self-Supervised Representations for Unified Voice Anti-Spoofing","Dominant voice anti-spoofing methods combine self-supervised learning (SSL) backbones with handcrafted features, but they offer limited transparency in how cues interact with transformer layers, and naive fusion restricts crossmodal learning. MOSAIC proposes an interpretable multi-token cross-attention framework that maps six semantic cue-group query tokens onto thirteen mean–std pooled WavLM-Large layer tokens. The 6×13 attention matrix and z-score activation analysis yield per-cue, per-layer attribution, improving unified LA/PA spoof detection performance across benchmarks.","MOSAIC: Interpretable Multi-Token Cross-Attention of Biophonetic and Self-Supervised Representations for Unified Voice Anti-Spoofing  \nYugwon Won  \narXiv :2607 .04314v1 [ ee ss .AS] 5 Jul 2026  \nAbstract—The dominant trend in voice anti-spoofing fuses selfsupervised (SSL) backbones (e.g., WavLM) with handcrafted features, yet such fusion typically lacks transparency in cueto-layer interactions, and simple concatenation limits crossmodal learning. We propose MOSAIC (Multi-token Oriented Speech Anti-spoofing via Integrated Cross-attention), an interpretable multi-token cross-attention framework that splits a 152-dimensional biophonetic feature vector into six semantic-group query tokens (Praat, phase, LFCC mean/std, sub-band mean/std) and attends them over thirteen mean–std pooled WavLM-Large transformer layers as keys/values. The resulting 6 × 13 attention matrix visualizes cue-to-layer alignment; a z-score analysis of the per-token activations shows that biophonetic/phase tokens activate more on bona fide speech while spectral/channel tokens activate more on spoofed speech — yielding per-cue, per-layer attribution that extends prior fusion approaches. Trained jointly with focal loss, a dual LA/PA domain-adversarial classifier, anda bona-fide-only VAE regularizer, MOSAIC attains EER 1.93 %/ 1.98 % on ASVspoof 2019 LA / PA — a single unified model that approaches the PA-specialized SOTA (LFCC-CMR, 1.34 %) while remaining competitive on LA — and 9.28 % / 6.21 % / 40.09 % on ASVspoof 2021 LA / DF / PA.  \nIndex Terms—Voice anti-spoofing, audio deepfake detection, cross-attention, self-supervised learning, WavLM, biophonetic features, interpretability, ASVspoof.  \nI. INTRODUCTION VOICE spoofing attacks—including text-to-speech (TTS)  \nsynthesis, voice conversion (VC), and replay — posea growing threat to automatic speaker verification (ASV) systems and to human listeners alike. The ASVspoof challenge series [1],[2] has established standardized benchmarks for two attack categories: Logical Access (LA, comprising TTS/VC) and Physical Access (PA, replay attacks) . Recent state of the art is dominated by self-supervised learning (SSL) backbones such as WavLM [3] and wav2vec 2.0 [4], paired with lightweight classifiers like AASIST [5] or attentive merging of hidden states [7] .  \nDespite this progress, several practical issues remain. Layerwise analyses of SSL models [7] identify which transformer layers contribute most to anti-spoofing, but the cue-to-layer mapping — which acoustic cue type aligns with which layer — is largely unaddressed, limiting transparency in safety-critical settings. Replay (PA) attacks are dominated by recording-channel distortions, which SSL features alone do not explicitly model; dedicated channel-aware methods such  \nY. Won is with the AI Security R&D Team, RaonSecure Co., Ltd., Seoul, Republic of Korea (e-mail: [linky1584@gmail.com](linky1584@gmail.com)).  \nas LFCC channel-magnitude-response (LFCC-CMR) [12] reach EER 1.34% on ASVspoof 2019 PA and 16.54% on ASVspoof 2021 PA with a linear-prediction spectrum estimator, but are typically evaluated as PA-only specialists with no LA/DF figures reported. The broader open challenge is the move from controlled 2019 PA conditions to the diverse, in-the-wild ASVspoof 2021 PA benchmark with unseen microphones and rooms [1], where single-model approaches that jointly handle LA and PA remain scarce. Recent SSL+handcrafted fusion—e.g., El Kheir et al.’s Two Views, One Truth [9], which pairs SSL features (as query) with spectral features MFCC/LFCC/CQCC (as key/value) for a 38 % relative EER reduction over SSL-only—retains a single uniform query side, so the attention map does not separate which acoustic cue type aligns with which SSL layer.  \nWe introduce MOSAIC with the following contributions:  \n• A 6-token multi-query cross-attention module that splits handcrafted features into six semantic-group queries (Praat glottal cues, STFT phase, LFCC mean/std, subband mean/std), ","cbCaihWnW0q06RjA","https://ap.wps.com/l/cbCaihWnW0q06RjA","pdf",488602,1,5,"English","en",105,"# Introduction\n## Voice spoofing challenges and ASVspoof benchmarks\n## Motivation for cue-to-layer interpretability\n# Proposed Method: MOSAIC\n## Biophonetic Features (152-dim)\n## MOSAIC architecture and training components","[{\"question\":\"What problem does MOSAIC address in existing voice anti-spoofing fusion methods?\",\"answer\":\"Existing fusion of SSL backbones and handcrafted features often lacks transparency about cue-to-layer interactions, and simple concatenation limits effective crossmodal learning. MOSAIC targets this by learning an interpretable cross-attention alignment between cues and SSL layers.\"},{\"question\":\"How does MOSAIC construct the multi-token cross-attention between biophonetic and SSL representations?\",\"answer\":\"It splits a 152-dimensional biophonetic feature vector into six semantic-group query tokens and attends over thirteen mean–std pooled WavLM-Large transformer layers as keys/values. The model produces a 6×13 attention matrix for cue-to-layer alignment.\"},{\"question\":\"What training components and losses does MOSAIC use to improve unified LA/PA performance?\",\"answer\":\"MOSAIC uses focal loss for spoof classification, a dual LA/PA domain-adversarial classifier with separate discriminators for LA and PA attack types, and a bona-fide-only VAE regularizer to reinforce stable representation learning.\"}]",1784183033,13,{"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":86,"head_meta":88,"extra_data":90,"updated_unix":27},"mosaic-interpretable-multi-token-cross-attention-of-biophonetic-and-self-supervised-representations-for-unified-voice-anti-spoofing","",{"@graph":35,"@context":85},[36,53,68],{"@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/research-report/",3,{"item":51,"name":13,"@type":42,"position":52},"https://docshare.wps.com/document/mosaic-interpretable-multi-token-cross-attention-of-biophonetic-and-self-supervised-representations-for-unified-voice-anti-spoofing/82801/",4,{"url":51,"name":13,"@type":54,"author":55,"headline":13,"publisher":57,"fileFormat":60,"inLanguage":23,"description":14,"dateModified":61,"datePublished":62,"encodingFormat":60,"isAccessibleForFree":63,"interactionStatistic":64},"DigitalDocument",{"name":9,"@type":56},"Person",{"url":40,"name":58,"@type":59},"DocShare","Organization","application/pdf","2026-07-17","2026-07-16",true,{"@type":65,"interactionType":66,"userInteractionCount":20},"InteractionCounter",{"@type":67},"ViewAction",{"@type":69,"mainEntity":70},"FAQPage",[71,77,81],{"name":72,"@type":73,"acceptedAnswer":74},"What problem does MOSAIC address in existing voice anti-spoofing fusion methods?","Question",{"text":75,"@type":76},"Existing fusion of SSL backbones and handcrafted features often lacks transparency about cue-to-layer interactions, and simple concatenation limits effective crossmodal learning. MOSAIC targets this by learning an interpretable cross-attention alignment between cues and SSL layers.","Answer",{"name":78,"@type":73,"acceptedAnswer":79},"How does MOSAIC construct the multi-token cross-attention between biophonetic and SSL representations?",{"text":80,"@type":76},"It splits a 152-dimensional biophonetic feature vector into six semantic-group query tokens and attends over thirteen mean–std pooled WavLM-Large transformer layers as keys/values. The model produces a 6×13 attention matrix for cue-to-layer alignment.",{"name":82,"@type":73,"acceptedAnswer":83},"What training components and losses does MOSAIC use to improve unified LA/PA performance?",{"text":84,"@type":76},"MOSAIC uses focal loss for spoof classification, a dual LA/PA domain-adversarial classifier with separate discriminators for LA and PA attack types, and a bona-fide-only VAE regularizer to reinforce stable representation learning.","https://schema.org",{"og:url":51,"og:type":87,"og:title":13,"og:site_name":58,"og:description":14},"article",{"robots":89,"canonical":51},"index,follow",{"doc_id":7,"site_id":24},{"code":4,"msg":5,"data":92},[93,97,101,105,109,114,119,122,127,130,134],{"id":20,"doc_module":4,"doc_module_name":45,"category_name":94,"show_sort_weight":95,"slug":96},"Story & Novel",90,"story-novel",{"id":46,"doc_module":4,"doc_module_name":45,"category_name":98,"show_sort_weight":99,"slug":100},"Literature",80,"literature",{"id":52,"doc_module":4,"doc_module_name":45,"category_name":102,"show_sort_weight":103,"slug":104},"Exam",70,"exam",{"id":21,"doc_module":4,"doc_module_name":45,"category_name":106,"show_sort_weight":107,"slug":108},"Comic",60,"comic",{"id":110,"doc_module":4,"doc_module_name":45,"category_name":111,"show_sort_weight":112,"slug":113},6,"Technology",50,"technology",{"id":115,"doc_module":4,"doc_module_name":45,"category_name":116,"show_sort_weight":117,"slug":118},7,"Healthcare",40,"healthcare",{"id":11,"doc_module":4,"doc_module_name":45,"category_name":12,"show_sort_weight":120,"slug":121},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":131,"doc_module":4,"doc_module_name":45,"category_name":132,"show_sort_weight":131,"slug":133},10,"Lifestyle","lifestyle",{"id":135,"doc_module":4,"doc_module_name":45,"category_name":136,"show_sort_weight":21,"slug":137},19,"General","general"]