{"id":5632,"date":"2026-10-23T13:55:25","date_gmt":"2026-10-23T05:55:25","guid":{"rendered":"https:\/\/www.15zhi.net\/blog\/?p=5632"},"modified":"2026-10-23T13:55:25","modified_gmt":"2026-10-23T05:55:25","slug":"202610-%e8%ae%ba%e6%96%87%e7%a0%94%e8%af%bb-use-a-unified-model-for-universal-sound-separation-and-extraction","status":"publish","type":"post","link":"https:\/\/www.15zhi.net\/blog\/202610-%e8%ae%ba%e6%96%87%e7%a0%94%e8%af%bb-use-a-unified-model-for-universal-sound-separation-and-extraction\/","title":{"rendered":"202610 \u8bba\u6587\u7814\u8bfb-USE: A Unified Model for Universal Sound Separation and Extraction"},"content":{"rendered":"\n

Author: <\/em><\/strong>Hongyu Wang, Chenda Li, Xin Zhou, Shuai Wang, Yanmin Qian<\/em>
Unit: <\/em><\/strong>Shanghai Jiao Tong University\uff1bNanjing University\uff1bShenzhen Loop Area Institute<\/em>
Journal: <\/em><\/strong>AAAI<\/em>
Year: <\/em><\/strong>2026<\/em><\/p>\n\n\n\n

\u4e00\u3001\u7814\u7a76\u80cc\u666f<\/h2>\n\n\n\n
\"\"<\/figure>\n\n\n\n

\u4e8c\u3001\u6838\u5fc3\u8d21\u732e<\/h2>\n\n\n\n
\"\"<\/figure>\n\n\n\n

\u4e09\u3001\u65b9\u6cd5<\/h2>\n\n\n\n

(1)\u5171\u4eab\u5206\u79bb\u4e3b\u5e72\uff1a \u6df7\u5408\u97f3\u9891\u7ecf\u8fc7 Encoder \u63d0\u53d6\u7279\u5f81\uff0cSeparator \u6839\u636e\u58f0\u6e90\u6761\u4ef6\u8868\u793a\u751f\u6210\u5206\u79bb\u63a9\u7801\uff0c\u518d\u7531 Decoder \u91cd\u5efa\u58f0\u6e90\u6ce2\u5f62\u3002<\/p>\n\n\n\n

(2)\u65e0\u5916\u90e8\u7ebf\u7d22\uff08USS\uff09\uff1a EDA Network \u6839\u636e\u6df7\u5408\u97f3\u9891\u7279\u5f81\u81ea\u52a8\u4f30\u8ba1\u58f0\u6e90\u6570\u91cf\uff0c\u5e76\u751f\u6210\u5bf9\u5e94\u7684 Attractor\uff0c\u5f15\u5bfc\u5171\u4eab\u5206\u79bb\u4e3b\u5e72\u5206\u79bb\u6240\u6709\u58f0\u6e90\u3002<\/p>\n\n\n\n

(3)\u6709\u5916\u90e8\u7ebf\u7d22\uff08TSE\uff09\uff1a Clue Network \u878d\u5408\u6587\u672c\u3001\u89c6\u9891\u548c\u58f0\u97f3\u6807\u7b7e\uff0c\u751f\u6210 Clue embedding\uff0c\u66ff\u4ee3 Attractor \u5f15\u5bfc\u540c\u4e00\u5206\u79bb\u4e3b\u5e72\u63d0\u53d6\u76ee\u6807\u58f0\u6e90\u3002<\/p>\n\n\n\n

\u6838\u5fc3\u673a\u5236\uff1a\u901a\u8fc7\u8bad\u7ec3\u9636\u6bb5\u7684\u8868\u793a\u5bf9\u9f50\uff0c\u4f7f Attractor \u4e0e Clue embedding \u80fd\u591f\u4f5c\u4e3a\u53ef\u66ff\u6362\u7684\u5206\u79bb\u6761\u4ef6\u3002<\/p>\n\n\n\n

\"\"<\/figure>\n\n\n\n

\u56db\u3001\u5b9e\u9a8c<\/h2>\n\n\n\n

1.\u5b9e\u9a8c\u51c6\u5907<\/p>\n\n\n\n

\"\"<\/figure>\n\n\n\n

2.\u5b9e\u9a8c\u7ed3\u679c\u2014\u2014Sound Separation & Source Counting<\/strong><\/strong><\/strong><\/strong><\/p>\n\n\n\n

(1)\u5f15\u5165 EDA \u540e\uff0cUSE \u80fd\u591f\u5728\u4e0d\u9884\u5148\u6307\u5b9a\u58f0\u6e90\u6570\u91cf\u7684\u60c5\u51b5\u4e0b\u5b8c\u6210 2Mix \u548c 3Mix \u58f0\u97f3\u5206\u79bb\u3002<\/p>\n\n\n\n

(2)\u58f0\u6e90\u6570\u91cf\u4f30\u8ba1\u5728 2Mix \u548c 3Mix \u4e2d\u7684\u51c6\u786e\u7387\u5206\u522b\u8d85\u8fc7 80% \u548c 70%\uff0c\u4f46\u590d\u6742\u6df7\u5408\u573a\u666f\u4e0b\u4ecd\u5b58\u5728\u8ba1\u6570\u9519\u8bef\u3002<\/p>\n\n\n\n

(3)\u7b2c\u4e8c\u9636\u6bb5\u8054\u5408\u8bad\u7ec3\u540e\uff0c\u5206\u79bb\u6027\u80fd\u603b\u4f53\u4fdd\u6301\u6216\u5c0f\u5e45\u63d0\u5347\uff0c\u8bf4\u660e\u7edf\u4e00\u8bad\u7ec3\u672a\u660e\u663e\u635f\u5bb3\u81ea\u4e3b\u5206\u79bb\u80fd\u529b\u3002<\/p>\n\n\n\n

\"\"<\/figure>\n\n\n\n

3.\u5b9e\u9a8c\u7ed3\u679c\u2014\u2014Multi-modal Target Sound Extraction<\/strong><\/strong><\/strong><\/p>\n\n\n\n

(1)USE \u80fd\u591f\u5229\u7528\u6587\u672c\u3001\u89c6\u9891\u548c\u58f0\u97f3\u6807\u7b7e\u7b49\u4e0d\u540c\u7ebf\u7d22\u7ec4\u5408\u5b8c\u6210\u76ee\u6807\u58f0\u97f3\u63d0\u53d6\u3002<\/p>\n\n\n\n

(2)\u4e09\u79cd\u6a21\u6001\u8054\u5408\u4f7f\u7528\u65f6\u6574\u4f53\u8868\u73b0\u8f83\u597d\uff0c\u4ec5\u63d0\u4f9b\u5355\u4e00\u7ebf\u7d22\u65f6\u4e5f\u80fd\u5b8c\u6210\u63d0\u53d6\uff0c\u4f46\u4e0d\u540c\u6a21\u6001\u7684\u6548\u679c\u5b58\u5728\u5dee\u5f02\u3002<\/p>\n\n\n\n

(3)\u76f8\u6bd4 DCCRN\uff0cUSE \u5728\u8868\u4e2d\u5404\u7ebf\u7d22\u7ec4\u5408\u4e0b\u5747\u53d6\u5f97\u66f4\u9ad8\u7684 SNRi\uff0c\u8868\u660e\u591a\u6a21\u6001\u6761\u4ef6\u63d0\u53d6\u5177\u6709\u4e00\u5b9a\u4f18\u52bf\u3002<\/p>\n\n\n\n

\"\"<\/figure>\n\n\n\n

4.\u5b9e\u9a8c\u7ed3\u679c\u2014\u2014Unified Training Performance<\/strong><\/strong><\/strong><\/strong><\/p>\n\n\n\n

(1)\u8054\u5408\u8bad\u7ec3\u540e\u7684 USE \u5728\u76ee\u6807\u58f0\u97f3\u63d0\u53d6\u4efb\u52a1\u4e0a\uff0c\u4e0e\u5355\u72ec\u8bad\u7ec3\u7684 SepFormer \u548c BSRNN \u76f8\u6bd4\uff0c\u6027\u80fd\u603b\u4f53\u76f8\u5f53\u6216\u7565\u6709\u63d0\u5347\u3002<\/p>\n\n\n\n

(2)\u5728 2Mix \u548c 3Mix\u3001Seen \u548c Unseen \u573a\u666f\u4e2d\u5747\u5448\u73b0\u7c7b\u4f3c\u8d8b\u52bf\uff0c\u8bf4\u660e\u7edf\u4e00\u8bad\u7ec3\u80fd\u591f\u517c\u987e\u76ee\u6807\u63d0\u53d6\u4efb\u52a1\u3002<\/p>\n\n\n\n

(3)\u7ed3\u5408\u524d\u8ff0\u5206\u79bb\u5b9e\u9a8c\uff0cUSE \u80fd\u591f\u5728\u540c\u4e00\u7f51\u7edc\u4e2d\u4fdd\u7559 USS \u548c TSE \u4e24\u79cd\u80fd\u529b\u3002<\/p>\n\n\n\n

\"\"<\/figure>\n\n\n\n

5.\u5b9e\u9a8c\u7ed3\u679c\u2014\u2014Attractor\u2013Clue Representation Alignment<\/strong><\/strong><\/strong><\/strong><\/strong><\/p>\n\n\n\n

(1)\u58f0\u6e90\u7279\u5f81\u5206\u5e03\uff1a \u4e0d\u540c\u7c7b\u578b\u58f0\u6e90\u5728\u7279\u5f81\u7a7a\u95f4\u4e2d\u5448\u73b0\u4e00\u5b9a\u805a\u96c6\u8d8b\u52bf\uff0c\u4f46\u4ecd\u5b58\u5728\u90e8\u5206\u4ea4\u53e0\u3002<\/p>\n\n\n\n

(2)\u8de8\u8868\u793a\u76f8\u4f3c\u6027\uff1a \u76f8\u540c\u58f0\u97f3\u7c7b\u578b\u7684 Attractor \u4e0e Clue embedding \u5728\u90e8\u5206\u533a\u57df\u5206\u5e03\u76f8\u8fd1\uff0c\u4f46\u5e76\u672a\u5b8c\u5168\u91cd\u5408\u3002<\/p>\n\n\n\n

(3)\u672a\u89c1\u7c7b\u522b\u8868\u73b0\uff1a Seen \u548c Unseen \u573a\u666f\u5747\u5448\u73b0\u4e00\u5b9a\u7684\u5206\u5e03\u89c4\u5f8b\uff0c\u4f46\u90e8\u5206\u8868\u793a\u4ecd\u5b58\u5728\u504f\u79fb\u3002<\/p>\n\n\n\n

\"\"<\/figure>\n\n\n\n

6.\u5b9e\u9a8c\u7ed3\u679c\u2014\u2014Variable Source Number Generalization<\/strong><\/strong><\/strong><\/strong><\/strong><\/strong><\/p>\n\n\n\n

(1)USE \u5728\u591a\u4e2a\u58f0\u97f3\u6570\u636e\u96c6\u53ca 2\uff5e6 \u4e2a\u58f0\u6e90\u7684\u6df7\u5408\u573a\u666f\u4e2d\u5747\u80fd\u591f\u8fdb\u884c\u5206\u79bb\uff0c\u4f46\u4e0d\u540c\u6570\u636e\u96c6\u548c\u58f0\u6e90\u6570\u91cf\u4e0b\u6027\u80fd\u5b58\u5728\u5dee\u5f02\u3002<\/p>\n\n\n\n

(2)\u4f7f\u7528 EDA \u9884\u6d4b\u58f0\u6e90\u6570\u91cf\u4e0e\u9884\u5148\u6307\u5b9a\u771f\u5b9e\u6570\u91cf\u76f8\u6bd4\uff0c\u5206\u79bb\u6027\u80fd\u603b\u4f53\u63a5\u8fd1\uff0c\u90e8\u5206\u573a\u666f\u4e2d\u751a\u81f3\u7565\u6709\u63d0\u5347\u3002<\/p>\n\n\n\n

(3)\u968f\u7740\u6df7\u5408\u58f0\u6e90\u6570\u91cf\u589e\u52a0\uff0c\u5206\u79bb\u96be\u5ea6\u901a\u5e38\u589e\u5927\uff0c\u8bf4\u660e\u58f0\u6e90\u6570\u91cf\u9002\u5e94\u80fd\u529b\u4ecd\u53d7\u573a\u666f\u590d\u6742\u5ea6\u5f71\u54cd\u3002<\/p>\n\n\n\n

\"\"<\/figure>\n\n\n\n

\u4e94\u3001\u603b\u7ed3<\/h2>\n\n\n\n

(1)\u63d0\u51fa USE \u7edf\u4e00\u6846\u67b6\uff0c\u5c06\u901a\u7528\u58f0\u97f3\u5206\u79bb\uff08USS\uff09\u4e0e\u76ee\u6807\u58f0\u97f3\u63d0\u53d6\uff08TSE\uff09\u6574\u5408\u5230\u540c\u4e00\u4e2a\u7f51\u7edc\u4e2d\uff0c\u6839\u636e\u662f\u5426\u5b58\u5728\u5916\u90e8\u76ee\u6807\u7ebf\u7d22\u7075\u6d3b\u5207\u6362\u81ea\u4e3b\u5206\u79bb\u548c\u6761\u4ef6\u63d0\u53d6\u6a21\u5f0f\u3002<\/p>\n\n\n\n

(2)\u901a\u8fc7 EDA \u81ea\u52a8\u4f30\u8ba1\u58f0\u6e90\u6570\u91cf\u5e76\u751f\u6210\u58f0\u6e90 Attractor\uff0c \u4f7f\u58f0\u97f3\u5206\u79bb\u4e0d\u518d\u8981\u6c42\u7528\u6237\u63d0\u524d\u6307\u5b9a\u9700\u8981\u5206\u79bb\u7684\u58f0\u6e90\u6570\u91cf\uff0c\u63d0\u9ad8\u4e86\u6a21\u578b\u5bf9\u590d\u6742\u6df7\u5408\u97f3\u9891\u7684\u9002\u5e94\u80fd\u529b\u3002<\/p>\n\n\n\n

(3)\u901a\u8fc7\u591a\u6a21\u6001 Clue Network \u548c\u8de8\u4efb\u52a1\u8868\u793a\u5bf9\u9f50\uff0c \u5c06\u6587\u672c\u3001\u89c6\u9891\u53ca\u58f0\u97f3\u6807\u7b7e\u7ebf\u7d22\u6620\u5c04\u5230\u4e0eAttractor \u76f8\u517c\u5bb9\u7684\u8868\u793a\u7a7a\u95f4\uff0c\u4f7f\u6a21\u578b\u53ef\u4ee5\u4f7f\u7528\u4e0d\u540c\u6765\u6e90\u7684\u7ebf\u7d22\u63a7\u5236\u540c\u4e00\u4e2a\u5206\u79bb\u7f51\u7edc\u3002<\/p>\n\n\n\n

(4)\u8bba\u6587\u7684\u6838\u5fc3\u521b\u65b0\u4e0d\u662f\u5355\u72ec\u63d0\u9ad8\u5206\u79bb\u7f51\u7edc\u7684\u6027\u80fd\uff0c\u800c\u662f\u5efa\u7acb\u7edf\u4e00\u7684\u201c\u58f0\u6e90\u6761\u4ef6\u8868\u793a\u201d\u3002 \u8fd9\u4e00\u8868\u793a\u65e2\u53ef\u4ee5\u4ece\u6df7\u5408\u97f3\u9891\u4e2d\u81ea\u4e3b\u4ea7\u751f\uff0c\u4e5f\u53ef\u4ee5\u7531\u5916\u90e8\u7ebf\u7d22\u751f\u6210\uff0c\u4ece\u800c\u8fde\u63a5\u81ea\u52a8\u58f0\u6e90\u53d1\u73b0\u4e0e\u76ee\u6807\u58f0\u97f3\u63d0\u53d6\u3002<\/p>\n\n\n\n

\u516d\u3001\u5bf9\u9f50\u601d\u8003<\/h2>\n\n\n\n

1.\u65b9\u6cd5\u521b\u65b0\u70b9\u2014\u2014\u5168\u57df\u6570\u636e\u62d3\u6251\u5efa\u6a21: USE \u63d0\u4f9b\u4e86\u4e00\u79cd\u4ece\u6df7\u5408\u97f3\u9891\u4e2d\u81ea\u4e3b\u5efa\u7acb\u58f0\u6e90\u8868\u793a\u3001\u518d\u901a\u8fc7\u5916\u90e8\u7ebf\u7d22\u8fdb\u884c\u76ee\u6807\u5173\u8054\u7684\u65b9\u6cd5\u3002EDA \u53ef\u4ee5\u751f\u6210\u4e0d\u540c\u58f0\u6e90\u5bf9\u5e94\u7684 Attractor\uff0cClue Network \u5219\u5c06\u8bed\u4e49\u6216\u89c6\u89c9\u7ebf\u7d22\u6620\u5c04\u5230\u76f8\u540c\u8868\u793a\u7a7a\u95f4\uff0c\u4f7f\u5185\u90e8\u58f0\u6e90\u8868\u793a\u4e0e\u5916\u90e8\u6761\u4ef6\u5efa\u7acb\u8054\u7cfb\uff0c\u4e3a\u540e\u7eed\u58f0\u6e90\u7ea7\u8bed\u4e49\u7ed1\u5b9a\u4e0e\u7ed3\u6784\u5316\u58f0\u573a\u5efa\u6a21\u63d0\u4f9b\u4e86\u53c2\u8003\u3002<\/p>\n\n\n\n

2.\u6280\u672f\u76ee\u6807\u70b9\u2014\u2014\u865a\u5b9e\u7a7a\u95f4\u5b6a\u751f\u63a8\u6f14: USE \u53ef\u4ee5\u5728\u672a\u77e5\u58f0\u6e90\u6570\u91cf\u7684\u60c5\u51b5\u4e0b\u81ea\u52a8\u5206\u79bb\u591a\u4e2a\u58f0\u97f3\uff0c\u4e5f\u53ef\u4ee5\u6839\u636e\u6587\u672c\u3001\u89c6\u9891\u548c\u7c7b\u522b\u6807\u7b7e\u63d0\u53d6\u6307\u5b9a\u76ee\u6807\uff0c\u4f46\u6a21\u578b\u5c1a\u672a\u663e\u5f0f\u5efa\u6a21\u58f0\u6e90\u65b9\u5411\u3001\u7a7a\u95f4\u4f4d\u7f6e\u3001\u957f\u671f\u8f68\u8ff9\u548c\u8de8\u65f6\u95f4\u8eab\u4efd\u4e00\u81f4\u6027\uff0c\u56e0\u6b64\u8fd8\u4e0d\u80fd\u76f4\u63a5\u6784\u6210\u5b8c\u6574\u7684\u52a8\u6001\u58f0\u573a\u8868\u793a\u3002<\/p>\n\n\n\n

3.\u4e2d\u8bd5\u4ea7\u54c1\u70b9\u2014\u2014\u57c3\u89c5\u6587\u65c5: USE \u53ef\u4ee5\u4ece\u666f\u533a\u3001\u8857\u9053\u7b49\u590d\u6742\u73af\u5883\u58f0\u97f3\u4e2d\u81ea\u4e3b\u5206\u79bb\u4e0d\u540c\u58f0\u6e90\uff0c\u4e5f\u53ef\u4ee5\u6839\u636e\u6307\u5b9a\u58f0\u97f3\u7c7b\u522b\u6216\u6587\u672c\u63cf\u8ff0\u63d0\u53d6\u7279\u5b9a\u73af\u5883\u58f0\u97f3\u3002<\/p>\n\n\n\n

<\/p>\n","protected":false},"excerpt":{"rendered":"

Author: Hongyu Wang, Chenda Li, Xin Zhou, Shuai Wang, Y […]<\/p>\n","protected":false},"author":74,"featured_media":0,"comment_status":"closed","ping_status":"","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[1],"tags":[],"class_list":["post-5632","post","type-post","status-publish","format-standard","hentry","category-events"],"_links":{"self":[{"href":"https:\/\/www.15zhi.net\/blog\/wp-json\/wp\/v2\/posts\/5632","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/www.15zhi.net\/blog\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/www.15zhi.net\/blog\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/www.15zhi.net\/blog\/wp-json\/wp\/v2\/users\/74"}],"replies":[{"embeddable":true,"href":"https:\/\/www.15zhi.net\/blog\/wp-json\/wp\/v2\/comments?post=5632"}],"version-history":[{"count":1,"href":"https:\/\/www.15zhi.net\/blog\/wp-json\/wp\/v2\/posts\/5632\/revisions"}],"predecessor-version":[{"id":5643,"href":"https:\/\/www.15zhi.net\/blog\/wp-json\/wp\/v2\/posts\/5632\/revisions\/5643"}],"wp:attachment":[{"href":"https:\/\/www.15zhi.net\/blog\/wp-json\/wp\/v2\/media?parent=5632"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/www.15zhi.net\/blog\/wp-json\/wp\/v2\/categories?post=5632"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/www.15zhi.net\/blog\/wp-json\/wp\/v2\/tags?post=5632"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}