[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"doc-detail-83821-en":3,"doc-seo-83821-105":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},83821,2336464648322,"Aria","https://ap-avatar.wpscdn.com/avatar/2200025388227c56fec?_k=1778556882303663488",8,"Research & Report","Training-Free Model Selection and Domain-Aware Score Calibration for First-Shot Anomalous Sound Detection","First-shot anomalous sound detection must flag anomalies from machine types never seen in training using a single fixed threshold, while the test clips may come from either a data-rich source domain or a data-scarce target domain. Two open DCASE Task 2 issues are addressed: negative source/target AUC correlation and non-transfer from dev to evaluation. DACO applies a training-free post-hoc layer over frozen embeddings, combining domain-balanced, label-free model selection with per-domain quantile calibration controlled by a prior strength. On DCASE 2025, selection predicts evaluation over a grid and raises official metric Ω, with bounded replication conclusions across 2023–2024 and a forward DCASE 2026 test.","Training-Free Model Selection and Domain-Aware Score Calibration for First-Shot Anomalous Sound  \nDetection  \nGrach Mkrtchian  \narXiv :2607 .04526v 1 [ cs . SD] 5 Jul 2026  \nAbstract—First-shot anomalous sound detection in DCASE Challenge Task 2 must flag anomalies of unseen machine types with a single threshold, without knowing whether a test clip comes from the data-rich source domain (990 normal training clips) or the data-scarce target domain (10). Two organizerreported problems remain open: source-and target-domain AUCare negatively correlated across systems, and development-set performance does not predict evaluation-set performance. We address both with a training-free post-hoc layer over frozen audio embeddings: (i) per-domain quantile calibration shrunk toward a pooled map by a prior strength m, tracing a controllable source/target balance frontier, and (ii) a label-free cross-validated domain-balance criterion that ranks candidate configurations from training normals only, paired with a coarse developmentlabeled viability veto. On DCASE 2025, the criterion rankpredicts the official evaluation score across a 45-configuration grid (Spearman ρs =+0 .91; family-block bootstrap 95% CI [+0 .83 , +0 .95]) while development score is uninformative (+0 .06, CI [−0 .39 , +0 .31]). Criterion-based selection raises evaluation Ω from 55.83 to 59.34 (jackknife CI [2 .2 , 4.8]) and, on an extended grid, to 61.05—retrospectively fourth of 35 teams. Replication on DCASE 2023 and 2024 bounds the claim: development score is uninformative in all three years and well-balanced degenerate configurations recur (vetoed in every case), but under familyclustered uncertainty the criterion’s predictive evidence survives only in 2025; a fixed full-equalization default matches or beats criterion selection in both replication years, and the selection gain over development-based selection is significant only in 2025 (+5 .2, jackknife CI [1 .3 , 9.2]). A DCASE 2026 forward test is frozen before the per-clip evaluation ground truth becomes available; all headline numbers are reproduced by the official evaluator.  \nIndex Terms—Anomalous sound detection, machine condition monitoring, domain generalization, model selection, calibration, group-conditional quantile calibration, DCASE.  \nI. INTRODUCTION  \nMACHINE condition monitoring by sound must operate  \non machine types the developer has never recorded, be deployed with almost no data from the operating environment, and run at a fixed sensitivity chosen before any anomaly has been observed. DCASE Challenge Task 2 crystallizes this setting as first-shot unsupervised ASD under domain generalization [1], [2]: for each machine type, training data comprise 990 normal clips from a source domain and only  \n10 from a target domain; at test time the domain label is withheld, so anomalies from both domains must be separated from normal sounds with a single threshold, and the evaluation  \nG. Mkrtchian is an independent researcher (e-mail: [g.mkrtchyan.m@gmail.com](g.mkrtchyan.m@gmail.com)).  \nmachine types are disjoint from the development (henceforth“dev”) ones, forbidding per-machine tuning.  \nThe 2025 organizers document two failure modes that remain open [1] . First, source/target imbalance: across the top twenty teams, source- and target-domain AUC are negatively correlated, and only four teams beat the official baselinesin both domains simultaneously. Second, dev→evaluation non-transfer: “achieving high AUC values in the development dataset does not indicate high AUC in the evaluation dataset”—model selection on the dev set is unreliable.  \nThis article argues that both failure modes are, at their core, calibration and model-selection problems, and that they can be attacked post hoc, on top of any frozen embedding extractor and any training-free anomaly-score backend, at negligible computational cost. The raw anomaly scores of source- and target-domain normal clips live on different scales, so no single thres","cbCaii2om2R9psYa","https://ap.wps.com/l/cbCaii2om2R9psYa","pdf",574129,1,12,"English","en",105,"# Abstract\n# I. Introduction\n## First-shot unsupervised ASD under domain generalization\n## Open problems: source/target imbalance and dev→evaluation non-transfer\n## Proposed method (DACO)\n### Label-free domain-balance selection criterion\n### Per-domain quantile calibration with controllable strength\n# Main findings","[{\"question\":\"What makes first-shot anomalous sound detection challenging in DCASE Task 2?\",\"answer\":\"The method must separate anomalies from normal sounds using one threshold for machine types unseen during training, while the test clips may originate from either a data-rich source domain or a data-scarce target domain.\"},{\"question\":\"Why do dev set results fail to predict evaluation performance?\",\"answer\":\"Development-set model selection is unreliable because the calibration mismatch between source and target score scales is not resolved correctly by dev performance, leading to weak or inconsistent transfer to evaluation.\"},{\"question\":\"How does DACO perform training-free model selection and calibration?\",\"answer\":\"Daco selects candidate configurations using a label-free criterion based on calibrated score distribution balance between domains, then applies per-domain quantile calibration with a controllable prior strength m that smoothly interpolates between raw scoring and full domain equalization.\"}]",1784190641,30,{"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},"training-free-model-selection-and-domain-aware-score-calibration-for-first-shot-anomalous-sound-detection","",{"@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/training-free-model-selection-and-domain-aware-score-calibration-for-first-shot-anomalous-sound-detection/83821/",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 makes first-shot anomalous sound detection challenging in DCASE Task 2?","Question",{"text":75,"@type":76},"The method must separate anomalies from normal sounds using one threshold for machine types unseen during training, while the test clips may originate from either a data-rich source domain or a data-scarce target domain.","Answer",{"name":78,"@type":73,"acceptedAnswer":79},"Why do dev set results fail to predict evaluation performance?",{"text":80,"@type":76},"Development-set model selection is unreliable because the calibration mismatch between source and target score scales is not resolved correctly by dev performance, leading to weak or inconsistent transfer to evaluation.",{"name":82,"@type":73,"acceptedAnswer":83},"How does DACO perform training-free model selection and calibration?",{"text":84,"@type":76},"Daco selects candidate configurations using a label-free criterion based on calibrated score distribution balance between domains, then applies per-domain quantile calibration with a controllable prior strength m that smoothly interpolates between raw scoring and full domain equalization.","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},[93,97,101,105,110,115,120,122,127,130,134],{"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":106,"doc_module":4,"doc_module_name":45,"category_name":107,"show_sort_weight":108,"slug":109},5,"Comic",60,"comic",{"id":111,"doc_module":4,"doc_module_name":45,"category_name":112,"show_sort_weight":113,"slug":114},6,"Technology",50,"technology",{"id":116,"doc_module":4,"doc_module_name":45,"category_name":117,"show_sort_weight":118,"slug":119},7,"Healthcare",40,"healthcare",{"id":11,"doc_module":4,"doc_module_name":45,"category_name":12,"show_sort_weight":28,"slug":121},"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":106,"slug":137},19,"General","general"]