[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"doc-seo-341783-105":3,"detail-sidebar-cat-0-en-105":81,"doc-detail-341783-en":130},{"code":4,"msg":5,"data":6},0,"ok",{"site_id":7,"language":8,"slug":9,"title":10,"keywords":11,"description":12,"schema_data":13,"social_meta":74,"head_meta":76,"extra_data":78,"updated_unix":80},105,"en","chtnsysu-sound-event-detection-system-with-pretrained-embeddings-extracted-from-beats-model-for-dcase-2023-task-4-technical-report","CHT+NSYSU Sound Event Detection System with Pretrained Embeddings Extracted from BEATs Model for DCASE 2023 Task 4 - Technical Report","","This technical report details a submission for DCASE 2023 Task 4, focusing on sound event detection in domestic environments. The proposed system introduces FDY CRNN architectures that use BEATs embeddings, combining late-fusion features formed by concating Frequency Dynamic Convolution frame-level feature maps with BEATs frame embeddings. Training is performed with the mean teacher framework and Asymmetric Focal Loss to mitigate positive-negative sample imbalance. Two-stage mean teacher training and optional ensembling improve PSDS-scenario results versus the BEATs-embedding baseline.",{"@graph":14,"@context":73},[15,34,56],{"@type":16,"itemListElement":17},"BreadcrumbList",[18,23,27,31],{"item":19,"name":20,"@type":21,"position":22},"https://docshare.wps.com","Home","ListItem",1,{"item":24,"name":25,"@type":21,"position":26},"https://docshare.wps.com/document/","Document",2,{"item":28,"name":29,"@type":21,"position":30},"https://docshare.wps.com/document/research-report/","Research & Report",3,{"item":32,"name":10,"@type":21,"position":33},"https://docshare.wps.com/document/chtnsysu-sound-event-detection-system-with-pretrained-embeddings-extracted-from-beats-model-for-dcase-2023-task-4-technical-report/341783/",4,{"url":32,"name":10,"@type":35,"image":36,"author":41,"headline":10,"publisher":44,"fileFormat":47,"inLanguage":8,"description":12,"dateModified":48,"datePublished":49,"encodingFormat":47,"isAccessibleForFree":50,"interactionStatistic":51},"DigitalDocument",{"url":37,"@type":38,"width":39,"height":40},"https://docshare.wps.com/thumbnails/chtnsysu-sound-event-detection-system-with-pretrained-embeddings-extracted-from-beats-model-for-dcase-2023-task-4-technical-report/341783.png","ImageObject",300,407,{"name":42,"@type":43},"Kyle","Person",{"url":19,"name":45,"@type":46},"DocShare","Organization","application/pdf","2026-11-16","2026-09-22",true,{"@type":52,"interactionType":53,"userInteractionCount":55},"InteractionCounter",{"@type":54},"ViewAction",8,{"@type":57,"mainEntity":58},"FAQPage",[59,65,69],{"name":60,"@type":61,"acceptedAnswer":62},"What is the main focus of this technical report?","Question",{"text":63,"@type":64},"It presents a submission system for DCASE 2023 Task 4, targeting sound event detection in domestic environments using pretrained BEATs embeddings.","Answer",{"name":66,"@type":61,"acceptedAnswer":67},"How are BEATs embeddings integrated into the proposed system?",{"text":68,"@type":64},"Frame-level embeddings from BEATs are fused with CNN features, then passed through adaptive averaging pooling and fed into RNN and MLP classifiers.",{"name":70,"@type":61,"acceptedAnswer":71},"What training strategy and loss function are used to improve performance?",{"text":72,"@type":64},"The system uses the mean teacher framework with Asymmetric Focal Loss to reduce the imbalance between positive and negative samples, and applies two-stage mean teacher training.","https://schema.org",{"og:url":32,"og:type":75,"og:title":10,"og:site_name":45,"og:description":12},"article",{"robots":77,"canonical":32},"index,follow",{"doc_id":79,"site_id":7},341783,1791410330,{"code":4,"msg":82,"data":83},"success",[84,88,92,96,101,106,111,114,119,122,126],{"id":22,"doc_module":4,"doc_module_name":25,"category_name":85,"show_sort_weight":86,"slug":87},"Story & Novel",90,"story-novel",{"id":26,"doc_module":4,"doc_module_name":25,"category_name":89,"show_sort_weight":90,"slug":91},"Literature",80,"literature",{"id":33,"doc_module":4,"doc_module_name":25,"category_name":93,"show_sort_weight":94,"slug":95},"Exam",70,"exam",{"id":97,"doc_module":4,"doc_module_name":25,"category_name":98,"show_sort_weight":99,"slug":100},5,"Comic",60,"comic",{"id":102,"doc_module":4,"doc_module_name":25,"category_name":103,"show_sort_weight":104,"slug":105},6,"Technology",50,"technology",{"id":107,"doc_module":4,"doc_module_name":25,"category_name":108,"show_sort_weight":109,"slug":110},7,"Healthcare",40,"healthcare",{"id":55,"doc_module":4,"doc_module_name":25,"category_name":29,"show_sort_weight":112,"slug":113},30,"research-report",{"id":115,"doc_module":4,"doc_module_name":25,"category_name":116,"show_sort_weight":117,"slug":118},9,"Religion & Spirituality",20,"religion-spirituality",{"id":117,"doc_module":4,"doc_module_name":25,"category_name":120,"show_sort_weight":117,"slug":121},"World Cup","world-cup",{"id":123,"doc_module":4,"doc_module_name":25,"category_name":124,"show_sort_weight":123,"slug":125},10,"Lifestyle","lifestyle",{"id":127,"doc_module":4,"doc_module_name":25,"category_name":128,"show_sort_weight":97,"slug":129},19,"General","general",{"code":4,"msg":82,"data":131},{"doc_id":79,"user_id":132,"nickname":42,"user_avatar":133,"doc_module":4,"category_id":55,"category_name":29,"doc_title":10,"doc_description":12,"doc_content":134,"file_id":135,"file_url":136,"file_type":137,"file_size":138,"view_count":55,"is_deleted":4,"is_public":22,"is_downloadable":22,"audit_status":22,"page_count":33,"language":139,"language_code":8,"site_id":7,"html_lang":8,"table_of_contents":140,"faqs":141,"seo_title":142,"seo_description":12,"update_tm":143,"read_time":123,"preview_page_count":33,"preview_urls":144},3985741905716,"https://ap-avatar.wpscdn.com/davatar_994ba38a5ba835b3df7d355c54d3ed8d","CHT+NSYSU SOUND EVENT DETECTION SYSTEM WITH PRETRAINED EMBEDDINGS EXTRACTED FROM BEATS MODEL FOR DCASE 2023 TASK 4  \nTechnical Report  \nChia-Chuan Liu 1, Tzu-Hao Kuo 1, Chia-Ping Chen 1, Chung-Li Lu2, Bo-Cheng Chan2,  \nYu-Han Cheng2, Hsiang-Feng Chuang2  \n1 National Sun Yat-Sen University, Taiwan,  \n{m103040063,[m113040053](m113040053}@student.nsysu.edu.tw)[}](m113040053}@student.nsysu.edu.tw)[@student.nsysu.edu.tw](m113040053}@student.nsysu.edu.tw),  \n[cpchen@cse.nsysu.edu.tw](cpchen@cse.nsysu.edu.tw)  \n2 Chunghwa Telecom Laboratories, Taiwan,  \n{chungli,cbc,henacheng, [gotop](gotop}@cht.com.tw)[}](gotop}@cht.com.tw)[@cht.com.tw](gotop}@cht.com.tw)  \nABSTRACT  \nIn this technical report, we describe our submission system for DCASE 2023 Task4: sound event detection in domestic environments. We propose FDY CRNN systems using BEATs embeddings. The system adapted late-fusion to concate the feature maps from Frequency Dynamic Convolution and the frame-level embeddings from BEATs. After that, a classification layer produces the prediction from the late-fusion features. The system is trained by the mean teacher framework. We utilize Asymmetric Focal Loss as the supervised loss to alleviate the imbalance between positive and negative samples. Furthermore, we apply two-stage mean teacher training to utilize training data adequateately. Compared to PSDS-scenario 1 of 50% and PSDS-scenario 2 of 76.2% of the baseline system using BEATs embeddings. Our FDY CRNN system achieves 50.1% and  \n79.8%, respectively. The ensemble of the FDY CRNN system further improves the PSDS-scenario 1 to 52.5% and the PSDS-scenario 2 to 80 .4% .  \nIndex Terms— BEATs, FDY-CRNN, Selective kernel unit  \n1. INTRODUCTION  \nIn this technical report, we describe our submission systems for Detection and Classification of Acoustic Scenes and Events (DCASE) 2023 Task 4 A. In this year, a new baseline system [1] using pretrained embeddings extracted from Bidirectional Encoder representation from Audio Transformers (BEATs) [2] is proposed. BEATs is a transformer-based model trained on the large-scale Audioset [3] dataset. The model is the state-of-the-art system for audio classification tasks. As the output of BEATs, the frame-level embedding contains high-level classification information along temporal dimension, which is useful for SED systems. The new baseline system takes fusion features, which is the concatenation of framelevel embedding extracted from BEATs and feature maps produced by CNN, as input. Due to the temporal resolution of the frame-level embedding is finer, it needs to be downsampled to match the feature maps. There are three different downsampling methods, including average pooling, interpolation and RNN encoding. By fusing with the embeddings of BEATs, the newly baseline system has a huge progress in performance. Based on the baseline, we proposed our improvements to the architecture and training process. Firstly, we  \nused frequency dynamic convolution [4] and selective kernel convolution [5] to implement two different SED systems, called FDYCRNN and VGGSK respectively. The former improves feature extraction ability in the frequency domain, and the latter uses different kernel sizes with an attention mechanism to enhance discrimination between different event classes. Secondly, we applied noisy student [6], random consistency training [7] and asymmetric focal loss [8] during training. Thirdly, we employ an adaptive median filter to smooth strong label predictions. Our improved system achieves true PSDS-scenario 1 of 50.1% and true PSDS-scenario 2 of 79.8%, which outperform the baseline system of 50% and 76.4% . Furthermore, we ensemble multiple systems with different architecturesand training methods. The ensemble system further increases true PSDS-1 to 52 .5% and true PSDS-2 to 80 .4% .  \n2. METHODS  \n2.1. Model architectures  \nWe propose the VGGSK structure, consisting of selective kernel, residual structure [9] and Visual Geometry Group(VGG) convol","cbCaih9WNuyFI3SS","https://ap.wps.com/l/cbCaih9WNuyFI3SS","pdf",157262,"English","# Abstract\n# 1. Introduction\n# 2. Methods\n## 2.1. Model architectures\n## 2.2. Pretrained model BEATs\n## 2.3. Training process","[{\"question\":\"What is the main focus of this technical report?\",\"answer\":\"It presents a submission system for DCASE 2023 Task 4, targeting sound event detection in domestic environments using pretrained BEATs embeddings.\"},{\"question\":\"How are BEATs embeddings integrated into the proposed system?\",\"answer\":\"Frame-level embeddings from BEATs are fused with CNN features, then passed through adaptive averaging pooling and fed into RNN and MLP classifiers.\"},{\"question\":\"What training strategy and loss function are used to improve performance?\",\"answer\":\"The system uses the mean teacher framework with Asymmetric Focal Loss to reduce the imbalance between positive and negative samples, and applies two-stage mean teacher training.\"}]","CHT+NSYSU Sound Event Detection System with Pretrained Embeddings Extracted from BEATs Model for DCASE 2023 Task 4 - Technical Report | PDF",1790043783,[145,149,153,157],{"page_no":22,"url":146,"width":147,"height":148},"https://res-academy.cache.wpscdn.com/upload/kdocshare/2026-11-16/ff12c2642857d79b1321ec5bec06aeaa.webp",983,1272,{"page_no":26,"url":150,"width":151,"height":152},"https://res-academy.cache.wpscdn.com/upload/kdocshare/2026-11-16/17c70bf233da78db3cd89d6539c99d5a.webp",982,1270,{"page_no":30,"url":154,"width":155,"height":156},"https://res-academy.cache.wpscdn.com/upload/kdocshare/2026-11-16/4cfcc56de53285f8d73535319f8f5009.webp",981,1269,{"page_no":33,"url":158,"width":159,"height":160},"https://res-academy.cache.wpscdn.com/upload/kdocshare/2026-11-16/62ac4859ef3c761b9447b4fbb03a861a.webp",980,1268]