{"id":28354,"date":"2025-08-19T10:14:00","date_gmt":"2025-08-19T02:14:00","guid":{"rendered":"http:\/\/139.9.1.231\/?p=28354"},"modified":"2025-08-19T15:15:05","modified_gmt":"2025-08-19T07:15:05","slug":"s2snd","status":"publish","type":"post","link":"http:\/\/139.9.1.231\/index.php\/2025\/08\/19\/s2snd\/","title":{"rendered":"S2SND  \u5728\u7ebf\/\u79bb\u7ebf\u7edf\u4e00\u63a8\u7406\u7684\u9ad8\u7cbe\u5ea6\u8bf4\u8bdd\u4eba\u65e5\u5fd7\u65b0\u6846\u67b6"},"content":{"rendered":"\n<p><strong>\u8bba\u6587\u9898\u76ee\uff1a<\/strong>Sequence-to-Sequence Neural Diarization with Automatic Speaker Detection and Representation<\/p>\n\n\n\n<p><strong>\u8bba\u6587\uff1a<\/strong><a href=\"https:\/\/arxiv.org\/abs\/2411.13849\">https:\/\/arxiv.org\/abs\/2411.13849<\/a><\/p>\n\n\n\n<p>\u6458\u81ea\uff1a<a href=\"https:\/\/mp.weixin.qq.com\/s\/s1JuYq5S2v2bfGnXPE0Qnw\">https:\/\/mp.weixin.qq.com\/s\/s1JuYq5S2v2bfGnXPE0Qnw<\/a><\/p>\n\n\n\n\n\n<h2><strong>\u7814\u7a76\u52a8\u673a<\/strong><\/h2>\n\n\n\n<p>\u201c\u8c01\u5728\u4ec0\u4e48\u65f6\u5019\u8bf4\u8bdd\u201d\u662f\u591a\u8bf4\u8bdd\u4eba\u8bed\u97f3\u7406\u89e3\u4e2d\u7684\u6838\u5fc3\u4efb\u52a1\u3002\u7136\u800c\uff0c\u4f20\u7edf\u7684\u8bf4\u8bdd\u4eba\u65e5\u5fd7\u7cfb\u7edf\u666e\u904d\u4f9d\u8d56\u4ece\u5b8c\u6574\u97f3\u9891\u63d0\u53d6\u8bf4\u8bdd\u4eba\u5d4c\u5165\u3001\u805a\u7c7b\u5339\u914d\u6216\u7aef\u5230\u7aef\u5efa\u6a21\u7b49\u65b9\u6cd5\uff0c\u5f80\u5f80\u53ea\u80fd\u79bb\u7ebf\u8fd0\u884c\uff0c<strong>\u96be\u4ee5\u9002\u5e94\u5b9e\u65f6\u5bf9\u8bdd\u3001\u4f1a\u8bae\u8f6c\u5199\u7b49\u5bf9\u7cfb\u7edf\u5ef6\u8fdf\u8981\u6c42\u6781\u9ad8\u7684\u5b9e\u9645\u573a\u666f\u3002<\/strong><\/p>\n\n\n\n<p>\u4e3a\u7a81\u7834\u8fd9\u4e9b\u9650\u5236\uff0c\u6b66\u6c49\u5927\u5b66\u4e0e\u6606\u5c71\u675c\u514b\u5927\u5b66\u7814\u7a76\u56e2\u961f\u63d0\u51fa\u4e86\u5168\u65b0\u6846\u67b6\u2014\u2014 S2SND\uff08Sequence-to-Sequence Neural Diarization\uff0c\u5e8f\u5217\u5230\u5e8f\u5217\u7684\u795e\u7ecf\u65e5\u5fd7\uff09\u3002\u8be5\u65b9\u6cd5\u57fa\u4e8e\u5e8f\u5217\u5230\u5e8f\u5217\u67b6\u6784\uff0c\u521b\u65b0\u6027\u5730\u878d\u5408\u4e86\u201c\u8bf4\u8bdd\u4eba\u81ea\u52a8\u68c0\u6d4b\u201d\u4e0e\u201c\u5d4c\u5165\u52a8\u6001\u63d0\u53d6\u201d\u673a\u5236\uff0c\u65e0\u9700\u805a\u7c7b\u4e0e\u5148\u9a8c\u7684\u8bf4\u8bdd\u4eba\u5d4c\u5165\uff0c\u5373\u53ef\u4ee5\u5728\u7ebf\u63a8\u7406\u7684\u5f62\u5f0f\u9010\u4e2a\u8bc6\u522b\u65b0\u52a0\u5165\u7684\u672a\u77e5\u8bf4\u8bdd\u4eba\uff0c\u5e76\u540c\u6b65\u5b8c\u6210\u8bed\u97f3\u6d3b\u52a8\u68c0\u6d4b\u4e0e\u8bf4\u8bdd\u4eba\u8868\u793a\u5efa\u6a21\u3002<\/p>\n\n\n\n<p>S2SND \u65e0\u9700\u805a\u7c7b\u6216\u6392\u5217\u4e0d\u53d8\u8bad\u7ec3\uff08PIT\uff09\uff0c\u5e76\u5177\u5907\u7edf\u4e00\u652f\u6301\u5728\u7ebf\u63a8\u7406\u4e0e\u79bb\u7ebf\u91cd\u89e3\u7801\u7b49\u7279\u6027\uff0c\u5728\u591a\u4e2a\u8bc4\u4f30\u6570\u636e\u96c6\u7684\u591a\u79cd\u8bc4\u6d4b\u6761\u4ef6\u4e0b\u5747\u53d6\u5f97\u4e86\u9886\u5148\u6027\u80fd\uff0c\u76f8\u5173\u6210\u679c\u5df2\u53d1\u8868\u4e8e\u8bed\u97f3\u9886\u57df\u6743\u5a01\u671f\u520a\u00a0IEEE Transactions on Audio, Speech and Language Processing\u3002<\/p>\n\n\n\n<h2><strong>\u65b9\u6cd5\u4ecb\u7ecd<\/strong><\/h2>\n\n\n\n<p>\u5982\u56fe1\u6240\u793a\uff0cS2SND \u6846\u67b6\u91c7\u7528\u5e8f\u5217\u5230\u5e8f\u5217\u67b6\u6784\uff0c\u4e3b\u8981\u5305\u542b\u4e09\u90e8\u5206\u6a21\u5757\uff1a\u57fa\u4e8eResNet-34\u7684\u7279\u5f81\u63d0\u53d6\u5668\uff08Extractor\uff09\uff0c\u57fa\u4e8eConformer\u7684\u7f16\u7801\u5668\uff08Encoder\uff09\uff0c\u548c\u521b\u65b0\u6027\u8bbe\u8ba1\u7684\u68c0\u6d4b\u89e3\u7801\u5668\uff08Detection Decoder\uff09\u548c\u8868\u5f81\u89e3\u7801\u5668\uff08Representation Decoder\uff09\u4e24\u4e2a\u90e8\u5206\uff0c\u5b9e\u73b0\u4e86\u8bf4\u8bdd\u4eba\u68c0\u6d4b\u4e0e\u5d4c\u5165\u63d0\u53d6\u7684\u53cc\u5411\u534f\u540c\u3002\u9996\u5148\uff0c\u8f93\u5165\u97f3\u9891\u7ecf\u8fc7&nbsp;ResNet \u7ed3\u6784\u7684\u5e27\u7ea7\u63d0\u53d6\u5668\u751f\u6210\u65f6\u5e8f\u8bf4\u8bdd\u4eba\u7279\u5f81\u5e8f\u5217\u3002\u968f\u540e\u9001\u5165\u57fa\u4e8e Conformer \u67b6\u6784\u7684\u7f16\u7801\u5668\u5efa\u6a21\u957f\u65f6\u4f9d\u8d56\u5173\u7cfb\uff0c\u8f93\u51fa\u4f5c\u4e3a\u4e0b\u6e38\u89e3\u7801\u5668\u7684\u4e3b\u7279\u5f81\u8f93\u5165\u3002\u5728\u89e3\u7801\u9636\u6bb5\uff0c\u6a21\u578b\u5e76\u884c\u4f7f\u7528\u4e24\u4e2a\u529f\u80fd\u4e92\u8865\u7684\u89e3\u7801\u5668\uff1a\u68c0\u6d4b\u89e3\u7801\u5668\u5229\u7528\u7f16\u7801\u5668\u7684\u8f93\u51fa\u7279\u5f81\u4e0e\u76ee\u6807\u8bf4\u8bdd\u4eba\u58f0\u7eb9\u5d4c\u5165\u4f5c\u4e3a\u53c2\u8003\u4fe1\u606f\uff0c\u9884\u6d4b\u591a\u4e2a\u8bf4\u8bdd\u4eba\u7684\u8bed\u97f3\u6d3b\u52a8\u3002\u53cd\u4e4b\uff0c\u8868\u5f81\u89e3\u7801\u5668\u5219\u5229\u7528\u63d0\u53d6\u5668\u7684\u8f93\u51fa\u7279\u5f81\u4e0e\u76ee\u6807\u8bf4\u8bdd\u4eba\u8bed\u97f3\u6d3b\u52a8\u4f5c\u4e3a\u53c2\u8003\u4fe1\u606f\uff0c\u63d0\u53d6\u591a\u4e2a\u8bf4\u8bdd\u4eba\u7684\u58f0\u7eb9\u5d4c\u5165\u3002<\/p>\n\n\n\n<p>\u4e3a\u4e86\u6253\u901a\u4e24\u6761\u8def\u5f84\u7684\u534f\u540c\u4f18\u5316\uff0cS2SND \u5f15\u5165\u4e00\u4e2a\u53ef\u5b66\u4e60\u7684\u5168\u4f53\u8bf4\u8bdd\u4eba\u5d4c\u5165\u77e9\u9635\uff08Embedding Matrix\uff09\uff0c\u6bcf\u4e00\u4e2a\u884c\u5411\u91cf\u5bf9\u5e94\u8bad\u7ec3\u96c6\u4e2d\u4e00\u4e2a\u8bf4\u8bdd\u4eba\u7684\u58f0\u7eb9\u5d4c\u5165\u3002\u8bad\u7ec3\u65f6\uff0c\u8bed\u97f3\u6d3b\u52a8\u6807\u7b7e\u7531\u6807\u6ce8\u6570\u636e\u63d0\u4f9b\uff0c\u800c\u76ee\u6807\u8bf4\u8bdd\u4eba\u5d4c\u5165\u901a\u8fc7\u8be5\u77e9\u9635\u67e5\u8868\u83b7\u5f97\uff08one-hot \u6807\u7b7e\u4e0e\u5d4c\u5165\u77e9\u9635\u76f8\u4e58\uff09\uff0c\u4f5c\u4e3a\u68c0\u6d4b\u89e3\u7801\u5668\u7684\u8f85\u52a9\u67e5\u8be2\uff1b\u540c\u65f6\uff0c\u8be5\u77e9\u9635\u4e5f\u4f5c\u4e3a\u8868\u5f81\u89e3\u7801\u5668\u7684\u76d1\u7763\u76ee\u6807\uff0c\u7528\u4e8e\u5bf9\u5176\u9884\u6d4b\u51fa\u7684\u5d4c\u5165\u6267\u884c ArcFace&nbsp;Loss\u76d1\u7763\u3002\u68c0\u6d4b\u89e3\u7801\u5668\u8f93\u5165\u7684\u58f0\u7eb9\u5d4c\u5165\u548c\u8868\u5f81\u89e3\u7801\u5668\u8f93\u51fa\u7684\u58f0\u7eb9\u5d4c\u5165\u4e3a\u540c\u4e2a\u7279\u5f81\u7a7a\u95f4\u5185\u8054\u5408\u4f18\u5316\u6240\u5f97\u3002<\/p>\n\n\n\n<p>\u4e3a\u589e\u5f3a\u6a21\u578b\u5bf9\u65b0\u8bf4\u8bdd\u4eba\u7684\u68c0\u6d4b\u80fd\u529b\uff0c\u8bad\u7ec3\u4e2d\u91c7\u7528Masked Speaker Prediction\u673a\u5236\uff1a\u6bcf\u4e2a\u8bad\u7ec3\u6837\u672c\u4e2d\u4f1a\u968f\u673a\u5c4f\u853d\u4e00\u4e2a\u5df2\u77e5\u8bf4\u8bdd\u4eba\u5d4c\u5165\uff0c\u4f7f\u7528\u4e00\u4e2a\u53ef\u5b66\u4e60\u7684\u201c\u4f2a\u8bf4\u8bdd\u4eba\u5d4c\u5165\u201d\u4ee3\u66ff\uff0c\u5e76\u5c06\u5176\u8bed\u97f3\u6d3b\u52a8\u6807\u7b7e\u5bf9\u5e94\u8f93\u51fa\u4f4d\u7f6e\uff0c\u4fc3\u4f7f\u6a21\u578b\u5b66\u4f1a\u53d1\u73b0\u201c\u672a\u6ce8\u518c\u8bf4\u8bdd\u4eba\u201d\u7684\u8bed\u97f3\u3002\u53e6\u4e00\u65b9\u9762\uff0cS2SND\u8bbe\u8ba1\u4e86Target-Voice Embedding Extraction\u673a\u5236\uff1a\u5229\u7528\u8868\u5f81\u89e3\u7801\u5668\u4ece\u4ece\u591a\u8bf4\u8bdd\u4eba\u6df7\u5408\u97f3\u9891\u4e2d\u53cd\u5411\u63d0\u53d6\u5d4c\u5165\uff0c\u5b9e\u73b0\u68c0\u6d4b\u4e0e\u8868\u5f81\u7684\u95ed\u73af\u4f18\u5316\u3002\u6574\u4e2a\u6846\u67b6\u7aef\u5230\u7aef\u8bad\u7ec3\uff0c\u8054\u5408\u4f18\u5316\u8bed\u97f3\u6d3b\u52a8\u9884\u6d4b\u7684\u00a0BCE \u635f\u5931\u4e0e\u5d4c\u5165\u7a7a\u95f4\u7684 ArcFace \u635f\u5931\u3002\u00a0<\/p>\n\n\n\n<figure class=\"wp-block-image size-large\"><img loading=\"lazy\" width=\"1024\" height=\"418\" src=\"http:\/\/139.9.1.231\/wp-content\/uploads\/2025\/08\/image-28-1024x418.png\" alt=\"\" class=\"wp-image-28376\" srcset=\"http:\/\/139.9.1.231\/wp-content\/uploads\/2025\/08\/image-28-1024x418.png 1024w, http:\/\/139.9.1.231\/wp-content\/uploads\/2025\/08\/image-28-300x122.png 300w, http:\/\/139.9.1.231\/wp-content\/uploads\/2025\/08\/image-28-768x313.png 768w, http:\/\/139.9.1.231\/wp-content\/uploads\/2025\/08\/image-28.png 1047w\" sizes=\"(max-width: 1024px) 100vw, 1024px\" \/><\/figure>\n\n\n\n<p><strong><em>\u56fe1.\u00a0\u5e8f\u5217\u5230\u5e8f\u5217\u795e\u7ecf\u8bf4\u8bdd\u4eba\u65e5\u5fd7\uff08S2SND\uff09\u6846\u67b6\u793a\u610f\u56fe\u3002\u5176\u4e2d\uff0cDet. \u548c Rep. \u5206\u522b\u8868\u793a\u68c0\u6d4b\uff08Detection\uff09\u4e0e\u8868\u5f81\uff08Representation\uff09\u7684\u7f29\u5199\u3002<\/em><\/strong><\/p>\n\n\n\n<p>\u5982\u56fe2\u6240\u793a\uff0cS2SND \u652f\u6301\u4f4e\u5ef6\u8fdf\u7684\u5728\u7ebf\u63a8\u7406\uff0c\u91c7\u7528\u5757\u5f0f\uff08Blockwise\uff09\u8f93\u5165\u65b9\u5f0f\u5904\u7406\u957f\u97f3\u9891\u3002\u6bcf\u6b21\u8f93\u5165\u4e00\u4e2a\u97f3\u9891\u5757\uff0c\u5305\u542b\u4e09\u90e8\u5206\uff1a\u5de6\u4e0a\u4e0b\u6587\u3001\u5f53\u524d\u5757\u4e0e\u53f3\u4e0a\u4e0b\u6587\uff0c\u5206\u522b\u7528\u4e8e\u5efa\u6a21\u5386\u53f2\u4f9d\u8d56\u3001\u8f93\u51fa\u5f53\u524d\u7ed3\u679c\u4e0e\u63d0\u4f9b\u672a\u6765\u53c2\u8003\u3002\u6574\u4e2a\u97f3\u9891\u4ee5\u6ed1\u52a8\u7a97\u53e3\u65b9\u5f0f\u9010\u5757\u63a8\u8fdb\uff0c\u5f53\u524d\u5757\u7684\u7ed3\u679c\u5b9e\u65f6\u8f93\u51fa\uff0c\u4fdd\u8bc1\u56e0\u679c\u6027\u3002\u63a8\u7406\u8fc7\u7a0b\u4e2d\uff0c\u7cfb\u7edf\u7ef4\u62a4\u4e00\u4e2a\u8bf4\u8bdd\u4eba\u5d4c\u5165\u7f13\u51b2\u533a\uff08Speaker-Embedding Buffer\uff09\uff0c\u7528\u4e8e\u5b58\u50a8\u5f53\u524d\u5df2\u77e5\u8bf4\u8bdd\u4eba\u7684\u5d4c\u5165\u8868\u793a\u3002\u6bcf\u4e00\u8f6e\u63a8\u7406\u7684\u8f93\u5165\u8bf4\u8bdd\u4eba\u5d4c\u5165\u7531\u4e09\u90e8\u5206\u62fc\u63a5\u7ec4\u6210\uff1a<\/p>\n\n\n\n<ol><li>\u56fa\u5b9a\u5728\u9996\u4f4d\u7684\u8bad\u7ec3\u8fc7\u7a0b\u4e2d\u4f7f\u7528\u7684\u4f2a\u8bf4\u8bdd\u4eba\u5d4c\u5165\uff0c\u7528\u4e8e\u68c0\u6d4b\u662f\u5426\u6709\u65b0\u8bf4\u8bdd\u4eba\u51fa\u73b0\uff1b<\/li><li>\u4ece\u7f13\u51b2\u533a\u6309\u987a\u5e8f\u8bfb\u53d6\u7684\u5df2\u77e5\u8bf4\u8bdd\u4eba\u5d4c\u5165\uff0c\u7528\u4f5c\u7279\u5b9a\u4eba\u8eab\u4efd\u7684\u53c2\u8003\u4fe1\u606f\uff1b<\/li><li>\u7528\u8bad\u7ec3\u8fc7\u7a0b\u4e2d\u4f7f\u7528\u7684\u975e\u8bed\u97f3\u5d4c\u5165\u4f5c\u4e3a\u586b\u5145\uff0c\u786e\u4fdd\u4e00\u4e2a\u6570\u636e\u6279\u6b21\uff08Batch\uff09\u7684\u8bf4\u8bdd\u4eba\u6570\u7ef4\u5ea6\u4e00\u81f4\u3002<\/li><\/ol>\n\n\n\n<p>\u68c0\u6d4b\u89e3\u7801\u5668\u4ee5\u5f53\u524d\u7f16\u7801\u7279\u5f81\u4e0e\u4e0a\u8ff0\u5d4c\u5165\u5e8f\u5217\u4e3a\u8f93\u5165\uff0c\u9884\u6d4b\u6bcf\u4e2a\u8bf4\u8bdd\u4eba\u7684\u8bed\u97f3\u6d3b\u52a8\uff08\u5305\u62ec\u4f2a\u8bf4\u8bdd\u4eba\uff09\uff1b\u968f\u540e\uff0c\u8868\u5f81\u89e3\u7801\u5668\u518d\u4ee5\u63d0\u53d6\u5668\u7279\u5f81\u4e0e\u9884\u6d4b\u7684\u8bed\u97f3\u6d3b\u52a8\u4f5c\u4e3a\u8f93\u5165\uff0c\u53cd\u5411\u63d0\u53d6\u6240\u6709\u76ee\u6807\u8bf4\u8bdd\u4eba\u7684\u5d4c\u5165\u8868\u793a\u3002\u6a21\u578b\u901a\u8fc7\u8bc4\u4f30\u6bcf\u4e2a\u9884\u6d4b\u51fa\u7684\u8bed\u97f3\u6d3b\u52a8\u7684\u975e\u91cd\u53e0\u8bed\u97f3\u65f6\u957f\u4f5c\u4e3a\u5d4c\u5165\u8d28\u91cf\u6743\u91cd\uff0c\u5bf9\u5e94\u751f\u6210\u7684\u5d4c\u5165\u53ea\u6709\u5728\u8d85\u8fc7\u8bbe\u5b9a\u9608\u503c\u65f6\u624d\u4f1a\u5199\u5165\u7f13\u51b2\u533a\uff1a<\/p>\n\n\n\n<ol><li>\u82e5\u4f2a\u8bf4\u8bdd\u4eba\u7684\u8bed\u97f3\u6d3b\u52a8\u8d28\u91cf\u6ee1\u8db3\u9608\u503c\uff0c\u8bf4\u660e\u5b58\u5728\u65b0\u7684\u672a\u6ce8\u518c\u8bf4\u8bdd\u4eba\uff0c\u5176\u5d4c\u5165\u5c06\u4ee5\u65b0\u8eab\u4efd\u7f16\u53f7\u5199\u5165\u7f13\u51b2\u533a\uff1b<\/li><li>\u82e5\u5df2\u6709\u8bf4\u8bdd\u4eba\u7684\u8bed\u97f3\u6d3b\u52a8\u8d28\u91cf\u6ee1\u8db3\u9608\u503c\uff0c\u5219\u5bf9\u5e94\u5d4c\u5165\u5c06\u8ffd\u52a0\u5230\u8be5\u8bf4\u8bdd\u4eba\u7684\u7f13\u51b2\u5217\u8868\uff1b<\/li><li>\u4f4e\u8d28\u91cf\u5d4c\u5165\u5219\u88ab\u820d\u5f03\uff0c\u4ee5\u907f\u514d\u6c61\u67d3\u5df2\u7f13\u5b58\u7684\u5d4c\u5165\u3002<\/li><\/ol>\n\n\n\n<p>\u6bcf\u6b21\u63a8\u7406\u4ec5\u91c7\u7528\u5f53\u524d\u5757\u5bf9\u5e94\u7684\u65f6\u95f4\u8303\u56f4\u5185\u7684\u9884\u6d4b\u7ed3\u679c\u62fc\u63a5\u5230\u5386\u53f2\u7ed3\u679c\uff0c\u786e\u4fdd\u8f93\u51fa\u7684\u65f6\u95f4\u56e0\u679c\u6027\u3002\u6b64\u5916\uff0c\u5f53\u6574\u4e2a\u97f3\u9891\u63a8\u7406\u5b8c\u6210\u540e\uff0c\u7cfb\u7edf\u53ef\u5229\u7528\u7f13\u51b2\u533a\u4e2d\u805a\u96c6\u7684\u9ad8\u8d28\u91cf\u8bf4\u8bdd\u4eba\u5d4c\u5165\uff0c\u5bf9\u5168\u6bb5\u97f3\u9891\u8fdb\u884c\u4e00\u6b21\u79bb\u7ebf\u91cd\u89e3\u7801\uff08Re-Decoding\uff09\uff0c\u8fdb\u4e00\u6b65\u63d0\u5347\u6574\u4f53\u8bc6\u522b\u7cbe\u5ea6\u3002\u7531\u4e8e\u6a21\u578b\u7ed3\u6784\u5728\u5728\u7ebf\u4e0e\u79bb\u7ebf\u6a21\u5f0f\u4e0b\u5b8c\u5168\u4e00\u81f4\uff0c\u65e0\u9700\u989d\u5916\u8f6c\u6362\u6216\u91cd\u65b0\u5efa\u6a21\u3002<\/p>\n\n\n\n<figure class=\"wp-block-image size-large\"><img loading=\"lazy\" width=\"1024\" height=\"404\" src=\"http:\/\/139.9.1.231\/wp-content\/uploads\/2025\/08\/image-29-1024x404.png\" alt=\"\" class=\"wp-image-28377\" srcset=\"http:\/\/139.9.1.231\/wp-content\/uploads\/2025\/08\/image-29-1024x404.png 1024w, http:\/\/139.9.1.231\/wp-content\/uploads\/2025\/08\/image-29-300x118.png 300w, http:\/\/139.9.1.231\/wp-content\/uploads\/2025\/08\/image-29-768x303.png 768w, http:\/\/139.9.1.231\/wp-content\/uploads\/2025\/08\/image-29.png 1050w\" sizes=\"(max-width: 1024px) 100vw, 1024px\" \/><figcaption><strong><strong><em>\u56fe<\/em><\/strong><strong><em>2.\u00a0\u5e8f\u5217\u5230\u5e8f\u5217\u795e\u7ecf\u8bf4\u8bdd\u4eba\u65e5\u5fd7\uff08<\/em><\/strong><strong><em>S2SND\uff09\u63a8\u7406\u6d41\u7a0b\u56fe\u3002\u5176\u4e2d\uff0c<\/em><\/strong><br><strong><em>Det. \u548c Rep. \u5206\u522b\u8868\u793a\u68c0\u6d4b\uff08Detection\uff09\u4e0e\u8868\u5f81\uff08Representation\uff09\u7684\u7f29\u5199\u3002<\/em><\/strong><\/strong><\/figcaption><\/figure>\n\n\n\n<div class=\"wp-block-image\"><figure class=\"aligncenter size-full is-resized\"><img loading=\"lazy\" src=\"http:\/\/139.9.1.231\/wp-content\/uploads\/2025\/08\/image-30.png\" alt=\"\" class=\"wp-image-28378\" width=\"504\" height=\"373\" srcset=\"http:\/\/139.9.1.231\/wp-content\/uploads\/2025\/08\/image-30.png 504w, http:\/\/139.9.1.231\/wp-content\/uploads\/2025\/08\/image-30-300x222.png 300w\" sizes=\"(max-width: 504px) 100vw, 504px\" \/><\/figure><\/div>\n\n\n\n<h3>\u6d41\u5f0f\u63a8\u7406\u4f2a\u4ee3\u7801\uff1a<\/h3>\n\n\n\n<pre class=\"wp-block-code\"><code>Algorithm 1 Pseudocode of online inference in the Python-like style.\r\n\"\"\"\r\n- Extractor(), Encoder(), Det_Decoder(), Rep_Decoder(): neural network modules in S2SND models\r\n- W(): calculating embedding weight\r\nInputs\r\n- blocks: a sequence of input audio blocks\r\n- e_pse\/e_non: pseudo-speaker\/non-speech embedding\r\n- tau_1\/tau_2: threshold for pseudo-speaker\/enrolled-speaker embedding weight\r\n- lc\/lr: number of output VAD frames belonging to the current chunk \/ right context\r\n- N: speaker capacity\r\n- S: embedding dimension\r\nOutputs\r\n- dia_result: predicted target-speaker voice activities\r\n- emb_buffer: extracted speaker embeddings\r\n\"\"\"\r\ndia_result = {}             # \u521d\u59cb\u5316\u5206\u79bb\u7ed3\u679c\uff1a\u5b57\u5178\uff0ckey=spk_id\uff0cvalue=\u62fc\u63a5\u597d\u7684VAD\u5e27\u5e8f\u5217\r\nemb_buffer = {}             # \u521d\u59cb\u5316\u201c\u8bf4\u8bdd\u4eba\u5d4c\u5165\u7f13\u51b2\u533a\u201d\uff1akey=spk_id, value=&#091;(embedding, weight), ...]\r\nnum_frames = 0              # \u5df2\u7ecf\u8f93\u51fa\uff08\u62fc\u63a5\uff09\u7684VAD\u5e27\u8ba1\u6570\uff0c\u7528\u4e8e\u65b0\u51fa\u73b0\u7684\u8bf4\u8bdd\u4eba\u8865\u96f6\u5bf9\u9f50\r\nfor audio_block in blocks:  # \u5757\u5f0f\u6ed1\u7a97\u8bfb\u53d6\u4e0b\u4e00\u6bb5\u97f3\u9891\uff08\u542b\u5de6\/\u5f53\u524d\/\u53f3\u4e09\u90e8\u5206\uff09\r\n\r# \u6d41\u5f0f\uff1a\u9010\u5757\u5904\u7406\uff0c\u4fdd\u6301\u4f4e\u5ef6\u8fdf\u4e0e\u56e0\u679c\u6027\u3002\n    emb_list = &#091;e_pse]                    # \u8f93\u5165\u7ed9\u89e3\u7801\u5668\u7684\u201c\u8bf4\u8bdd\u4eba\u5d4c\u5165\u5e8f\u5217\u201d\u9996\u4f4d\u56fa\u5b9a\u653e\u4f2a\u8bf4\u8bdd\u4eba\u5d4c\u5165\r\n    spk_list = &#091;len(emb_buffer) + 1]      # \u5e76\u4e3a\u5b83\u5360\u4e00\u4e2a\u65b0ID\uff08\u201c\u6f5c\u5728\u65b0\u8bf4\u8bdd\u4eba\u201d\u7684\u5019\u9009ID\uff09\r\n\r# \u4f2a\u8bf4\u8bdd\u4eba\uff08pseudo-speaker\uff09\uff1a\u7528\u4e8e\u63a2\u6d4b\u662f\u5426\u6709\u65b0\u8bf4\u8bdd\u4eba\u3002\n# spk_list&#091;0] \u63d0\u524d\u9884\u7559\u4e00\u4e2a\u201c\u65b0ID\u201d\uff0c\u5982\u679c\u8fd9\u6b21\u5757\u91cc\u771f\u7684\u68c0\u6d4b\u5230\u65b0\u8bf4\u8bdd\u4eba\u4e14\u8d28\u91cf\u8fbe\u9608\u503c\uff0c\u5c31\u7528\u8fd9\u4e2a ID\u3002\n\n# \u5c06\u6bcf\u4e2a\u201c\u5df2\u77e5\u8bf4\u8bdd\u4eba\u201d\u7684\u6c47\u603b\u53c2\u8003\u5d4c\u5165\u62fc\u5230 emb_list\uff0c\u4e0e spk_list \u4e00\u4e00\u5bf9\u5e94\u3002\u8fd9\u6837\u89e3\u7801\u5668\u5c31\u80fd\u5bf9\u201c\u4f2a\u8bf4\u8bdd\u4eba + \u5df2\u77e5\u8bf4\u8bdd\u4eba\u4eec\u201d\u540c\u65f6\u8fdb\u884c\u8bed\u97f3\u6d3b\u52a8\u9884\u6d4b\u3002\n    for spk_id in emb_buffer.keys():      # \u904d\u5386\u5df2\u77e5\/\u5df2\u6ce8\u518c\u7684\u8bf4\u8bdd\u4eba\r\n        e_sum = torch.zeros(S)            # \u4e3a\u8be5\u8bf4\u8bdd\u4eba\u7d2f\u79ef\u6743\u91cd\u548c\r\n        w_sum = 0\r\n        for e_i, w_i in emb_buffer&#091;spk_id]:\r\n            e_sum += w_i * e_i            # \u8d28\u91cf\u52a0\u6743\u7684\u5d4c\u5165\u52a0\u548c\r\n            w_sum += w_i\r\n        emb_list.append(e_sum \/ w_sum)    # \u5f97\u5230\u8be5\u8bf4\u8bdd\u4eba\u7684\u201c\u53c2\u8003\u5d4c\u5165\u201d\uff08\u52a0\u6743\u5e73\u5747\uff09\r\n        spk_list.append(spk_id)           # \u8bb0\u5f55\u5176\u5bf9\u5e94\u7684\u8bf4\u8bdd\u4ebaID\uff08\u987a\u5e8f\u4e0eemb_list\u5bf9\u5e94\uff09\r\n\r\n# \u786e\u4fdd\u8fd9\u6b21\u524d\u5411\u4e2d\uff0c\u8bf4\u8bdd\u4eba\u7ef4\uff08batch\uff09\u957f\u5ea6\u56fa\u5b9a\u4e3a N\uff0c\u4fbf\u4e8e\u5f20\u91cf\u5316\u4e0e\u5e76\u884c\u3002\n    while len(emb_list) &lt; N:              # \u82e5\u8bf4\u8bdd\u4eba\u69fd\u4f4d\u4e0d\u591f\uff0c\u7528\u975e\u8bed\u97f3\u5d4c\u5165 e_non \u586b\u6ee1\uff08batch \u7ef4\u5ea6\u5bf9\u9f50\uff09\r\n        emb_list.append(e_non)\r\n\r    emb_tensor = torch.stack(emb_list)    # \u5f62\u72b6\uff1aN x S\r\n\r\n# \u63d0\u53d6\u5668\/\u7f16\u7801\u5668\u662f\u58f0\u5b66\u524d\u7aef\uff0c\u62bd\u53d6\u65f6\u5e8f\u8868\u793a\u3002\nX = Extractor(audio_block)            # \u63d0\u53d6\u5668\uff1a\u539f\u6ce2\u5f62 -> \u5e27\u7ea7\u7279\u5f81\uff0c\u5f62\u72b6 T x F\r\nX_hat = Encoder(X)                    # \u7f16\u7801\u5668\uff1a\u8fdb\u4e00\u6b65\u7f16\u7801\uff0c\u5f62\u72b6 T x D\r\n\r\n\n# Y_hat&#091;n]\uff1a\u7b2c n \u4e2a\u69fd\u4f4d\uff08\u4f2a\/\u5df2\u77e5\/\u586b\u5145\uff09\u7684\u8bed\u97f3\u6d3b\u52a8\uff08\u5e27\u7ea7\u6982\u7387\/0-1\uff09\u3002\n#E_hat&#091;n]\uff1a\u7b2c n \u4e2a\u69fd\u4f4d\u5bf9\u5e94\u7684\u8bf4\u8bdd\u4eba\u5d4c\u5165\u3002\n#\u4e24\u8005\u8054\u52a8\uff1a\u5148\u201c\u68c0\u6d4b\u54ea\u91cc\u5728\u8bf4\u8bdd\u201d\uff0c\u518d\u201c\u5728\u8fd9\u4e9b\u533a\u57df\u63d0\u4ee3\u8868\u5f81\u201d\u3002\nY_hat = Det_Decoder(X_hat, emb_tensor)  # \u68c0\u6d4b\u89e3\u7801\u5668\uff1a\u5bf9\u6bcf\u4e2a\u201c\u8bf4\u8bdd\u4eba\u69fd\u4f4d\u201d\u9884\u6d4bVAD\uff0c\u5f62\u72b6 N x T'\r\nE_hat = Rep_Decoder(X, Y_hat)           # \u8868\u5f81\u89e3\u7801\u5668\uff1a\u53cd\u5411\u63d0\u53d6\u6bcf\u4e2a\u69fd\u4f4d\u7684\u5d4c\u5165\uff0c\u5f62\u72b6 N x S\r\n\r\ny_pse = Y_hat&#091;0]                       # \u4f2a\u8bf4\u8bdd\u4eba\u7684 VAD\uff08T'\uff09\r\ne_pse = E_hat&#091;0]                       # \u4f2a\u8bf4\u8bdd\u4eba\u7684\u5d4c\u5165\uff08S\uff09\r\nw_pse = W(y_pse)                       # \u57fa\u4e8e\u975e\u91cd\u53e0\u8bed\u97f3\u65f6\u957f\u7b49\u6307\u6807\u8ba1\u7b97\u8d28\u91cf\u6743\u91cd\uff08\u6807\u91cf\uff09\r\n\r\n# \u5173\u952e\u70b9\uff1a\u53ea\u62fc\u5f53\u524d\u5757\u5bf9\u5e94\u7684 lc \u5e27\uff0c\u4fdd\u8bc1\u56e0\u679c\u8f93\u51fa\uff1b\u53f3\u4e0a\u4e0b\u6587\u53ea\u4f5c\u53c2\u8003\u4e0d\u8f93\u51fa\u3002\n# \u65b0\u8bf4\u8bdd\u4eba\u7b2c\u4e00\u6b21\u51fa\u73b0\uff0c\u9700\u8981\u5728\u5386\u53f2\u65f6\u95f4\u7ebf\u524d\u8fb9\u8865\u96f6\u5bf9\u9f50\n    if w_pse > tau_1:                      # \u82e5\u4f2a\u8bf4\u8bdd\u4eba\u8d28\u91cf\u8fc7\u9608\uff0c\u5224\u5b9a\uff1a\u51fa\u73b0\u65b0\u8bf4\u8bdd\u4eba\r\n        elapsed_y = torch.zeros(num_frames)      # \u4e3a\u65b0ID\u8865\u9f50\u5386\u53f2\u5e27\u76840\uff08\u8fc7\u53bb\u90fd\u672a\u8bf4\u8bdd\uff09\r\n        current_y = y_pse&#091;-(lc+lr) : -lr]        # \u53ea\u53d6\u5f53\u524d\u5757\u7684\u6709\u6548\u5e27\uff08\u53bb\u6389\u53f3\u4e0a\u4e0b\u6587\uff09\r\n        new_id = spk_list&#091;0]                     # \u53d6\u9884\u7559\u7684\u65b0\u8bf4\u8bdd\u4ebaID\r\n        dia_result&#091;new_id] = torch.cat(&#091;elapsed_y, current_y])  # \u62fc\u6210\u5168\u5c40\u65f6\u95f4\u7ebf\r\n        emb_buffer&#091;new_id] = &#091;(e_pse, w_pse)]    # \u628a\u8be5\u65b0\u4eba\u7684\u5d4c\u5165\u5199\u5165\u7f13\u51b2\u533a\r\n\r\n\n# \u66f4\u6b63\uff1a\u5faa\u73af\u4e0a\u754c\u4e0d\u5e94\u662f len(S)\uff08\u90a3\u662f\u5d4c\u5165\u7ef4\uff09\uff0c\u5e94\u8be5\u662f N \u6216 Y_hat.size(0)\u3002\n#\u5df2\u77e5\u8bf4\u8bdd\u4eba\u7684 dia_result&#091;spk_id] \u76f4\u63a5\u5728\u5df2\u6709\u65f6\u95f4\u7ebf\u4e0a\u62fc\u63a5\u672c\u5757\u7684 lc \u5e27\u3002\n#\u5f53\u8d28\u91cf\u597d\u65f6\uff0c\u5f80\u8be5\u8bf4\u8bdd\u4eba\u7684 emb_buffer \u8ffd\u52a0\u4e00\u6761 (embedding, weight)\uff0c\u540e\u7eed\u4f1a\u53c2\u4e0e\u52a0\u6743\u5e73\u5747\uff0c\u8d8a\u6eda\u8d8a\u7a33\u3002\n    for n in range(1, len(S)):             # \u3010\u8fd9\u91cc\u6709 Bug\u3011\u5e94\u4e3a range(1, N) \u6216 Y_hat.shape&#091;0]\r\n        y_n = Y_hat&#091;n]                     # \u7b2c n \u69fd\u4f4d\u7684 VAD\r\n        e_n = E_hat&#091;n]                     # \u7b2c n \u69fd\u4f4d\u7684\u5d4c\u5165\r\n        w_n = W(y_n)                       # \u8d28\u91cf\r\n        spk_id = spk_list&#091;n]               # \u69fd\u4f4d\u5bf9\u5e94\u7684\u8bf4\u8bdd\u4eba ID\uff08\u5df2\u77e5\uff09\r\n        dia_result&#091;spk_id] = torch.cat(&#091;dia_result&#091;spk_id], y_n&#091;-(lc+lr) : -lr]])  # \u62fc\u63a5\u672c\u5757\u6709\u6548\u5e27\r\n        if w_n > tau_2:                    # \u82e5\u8d28\u91cf\u597d\uff0c\u628a\u65b0\u5d4c\u5165\u8ffd\u52a0\u5230\u8be5\u8bf4\u8bdd\u4eba\u7684\u7f13\u51b2\u5217\u8868\r\n            emb_buffer&#091;spk_id].append((e_n, w_n))\r\n    num_frames += lc                       # \u5168\u5c40\u65f6\u95f4\u7ebf\u5411\u524d\u63a8\u8fdb lc \u5e27\r\n#\u6bcf\u5904\u7406\u5b8c\u4e00\u4e2a\u5757\uff0c\u5c31\u201c\u786e\u8ba4\u8f93\u51fa\u201d\u4e86 lc \u5e27\uff0c\u65f6\u95f4\u7ebf\u524d\u79fb\u3002<\/code><\/pre>\n\n\n\n<h2><strong>\u5b9e\u9a8c\u7ed3\u679c\u4e0e\u6027\u80fd\u5206\u6790<\/strong><\/h2>\n\n\n\n<p>\u672c\u7814\u7a76\u6784\u5efa\u4e86\u4e24\u7c7b\u6a21\u578b\u89c4\u6a21\u7684&nbsp;S2SND \u7cfb\u7edf\uff1aS2SND-Small \u548c S2SND-Medium\uff0c\u6700\u7ec8\u6a21\u578b\u53c2\u6570\u91cf\u5206\u522b\u7ea6\u4e3a16+M\u548c&nbsp;46M+\u3002\u4e3a\u652f\u6491\u5927\u89c4\u6a21\u8bad\u7ec3\uff0c\u6211\u4eec\u5206\u522b\u91c7\u7528\u4e86\u4e24\u4e2a\u4e0d\u540c\u89c4\u6a21\u7684\u8bf4\u8bdd\u4eba\u8bed\u6599\u4f5c\u4e3a\u4eff\u771f\u7d20\u6750\u3002\u5176\u4e2d\uff0cVoxCeleb2\u6570\u636e\u5e93\u7ea6&nbsp;100 \u4e07\u6761\u8bed\u97f3\uff0c\u8986\u76d6 6112 \u540d\u8bf4\u8bdd\u4eba\uff1bVoxBlink2\u6570\u636e\u5e93\u7ea6&nbsp;1000 \u4e07\u6761\u8bed\u97f3\uff0c\u8986\u76d6\u8d85\u8fc7 11 \u4e07\u540d\u8bf4\u8bdd\u4eba\u3002\u6b64\u5916\uff0c\u4e3a\u63d0\u5347\u4e2d\u578b\u6a21\u578b\u5bf9\u5927\u8bed\u6599\u7684\u5efa\u6a21\u80fd\u529b\uff0c\u672c\u7814\u7a76\u5f15\u5165\u4e86\u4ee5&nbsp;ResNet-152&nbsp;\u63d0\u53d6\u5668\u4f5c\u4e3a\u6559\u5e08\u6a21\u578b\u7684\u77e5\u8bc6\u84b8\u998f\u7b56\u7565\uff0c\u663e\u8457\u589e\u5f3a\u4e86\u8bf4\u8bdd\u4eba\u5d4c\u5165\u8d28\u91cf\u4e0e\u4e0b\u6e38\u63a8\u7406\u6027\u80fd\u3002\u6a21\u578b\u7ed3\u6784\u3001\u8bad\u7ec3\u7ec6\u8282\u4e0e\u5ef6\u8fdf\u8bbe\u7f6e\u7b49\u4fe1\u606f\u8be6\u89c1\u8bba\u6587\u539f\u6587\u3002\u6240\u6709\u5b9e\u9a8c\u5747\u57fa\u4e8e\u6807\u51c6\u7684&nbsp;DER\uff08Diarization Error Rate\uff09 \u6307\u6807\u8fdb\u884c\u8bc4\u4f30\uff0c\u4e0d\u4f7f\u7528&nbsp;Collar \uff0c\u4e25\u683c\u8861\u91cf\u7cfb\u7edf\u5728\u771f\u5b9e\u5e94\u7528\u6761\u4ef6\u4e0b\u7684\u5b9e\u9645\u8868\u73b0\u3002\u540c\u65f6\uff0c\u5206\u522b\u8bc4\u6d4b\u4f9d\u8d56\u548c\u4e0d\u4f9d\u8d56&nbsp;Oracle VAD\u7684\u60c5\u51b5\uff0c\u4ee5\u6ee1\u8db3\u548c\u5176\u4ed6\u73b0\u6709\u65b9\u6cd5\u5728\u4e0d\u540c\u8bc4\u4f30\u6807\u51c6\u4e0b\u7684\u516c\u5e73\u5bf9\u6bd4\u3002<\/p>\n\n\n\n<p>\u5982\u88681\u6240\u793a\uff0c\u5728\u00a0DIHARD-II \u6d4b\u8bd5\u96c6\u4e2d\uff0cS2SND-Medium \u6a21\u578b\u7ed3\u5408 VoxBlink2 \u8bed\u6599\u4e0e\u77e5\u8bc6\u84b8\u998f\u8bad\u7ec3\uff0c\u5728\u5728\u7ebf\u63a8\u7406\u6a21\u5f0f\uff080.64s \u603b\u5ef6\u8fdf\uff09\u4e0b\u53d6\u5f97\u4e86 24.41% DER\uff0c\u5237\u65b0\u5f53\u524d\u6240\u6709\u5728\u7ebf\u7cfb\u7edf\u7684\u6700\u4f18\u7ed3\u679c\u3002\u5b8c\u6210\u4e00\u6b21\u79bb\u7ebf\u91cd\u89e3\u7801\u540e\uff0cDER \u8fdb\u4e00\u6b65\u964d\u4f4e\u81f3 21.95%\uff0c\u5373\u4fbf\u4e0d\u4f9d\u8d56 Oracle VAD\uff0c\u4e5f\u5168\u9762\u8d85\u8d8a\u73b0\u6709\u6700\u5f3a\u7684 EEND-GLA \u4e0e VBx \u7b49\u65b9\u6cd5\u3002\u76f8\u6bd4\u4e4b\u4e0b\uff0c\u4f20\u7edf EEND-EDA + STB \u65b9\u6cd5\u7684\u5728\u7ebfDER \u4ecd\u5728 30% \u4ee5\u4e0a\uff0cEEND-GLA-Large \u7cfb\u7edf\u4e5f\u4ec5\u5728\u79bb\u7ebf\u6761\u4ef6\u4e0b\u8fbe\u5230\u00a028.33%\u3002S2SND \u5b9e\u73b0\u4e86\u66f4\u4f4e\u5ef6\u8fdf\u3001\u66f4\u9ad8\u51c6\u786e\u7387\u7684\u53cc\u91cd\u7a81\u7834\u3002<\/p>\n\n\n\n<figure class=\"wp-block-image size-full\"><img loading=\"lazy\" width=\"508\" height=\"793\" src=\"http:\/\/139.9.1.231\/wp-content\/uploads\/2025\/08\/image-31.png\" alt=\"\" class=\"wp-image-28394\" srcset=\"http:\/\/139.9.1.231\/wp-content\/uploads\/2025\/08\/image-31.png 508w, http:\/\/139.9.1.231\/wp-content\/uploads\/2025\/08\/image-31-192x300.png 192w\" sizes=\"(max-width: 508px) 100vw, 508px\" \/><figcaption><strong><strong><em>\u8868<\/em><\/strong><strong><em>1.\u00a0S2SND \u6a21\u578b\u4e0e\u5176\u4ed6\u65b9\u6cd5\u5728 DIHARD-II \u6d4b\u8bd5\u96c6\u4e0a\u7684\u5bf9\u6bd4\u7ed3\u679c<\/em><\/strong><\/strong><\/figcaption><\/figure>\n\n\n\n<p>\u5982\u88682\u6240\u793a\uff0c\u5728\u66f4\u65b0\u4e00\u4ee3\u7684\u00a0DIHARD-III \u6570\u636e\u96c6\u4e0a\uff0cS2SND \u540c\u6837\u8868\u73b0\u51fa\u8272\u3002\u5728 0.80s \u5ef6\u8fdf\u8bbe\u7f6e\u4e0b\u7684\u5728\u7ebf\u6a21\u5f0f\uff0cS2SND-Medium \u8fbe\u5230 17.12% DER\uff0c\u660e\u663e\u4f18\u4e8e\u5404\u7c7b\u5728\u7ebf\u7cfb\u7edf\u3002\u5b8c\u6210\u79bb\u7ebf\u91cd\u89e3\u7801\u540e\uff0cDER \u8fdb\u4e00\u6b65\u964d\u81f3 15.13%\u3002\u503c\u5f97\u4e00\u63d0\u7684\u662f\uff0cS2SND \u7684\u5728\u7ebf DER \u5df2\u903c\u8fd1\u751a\u81f3\u4f18\u4e8e\u8bb8\u591a\u79bb\u7ebf\u7cfb\u7edf\uff08\u5982 EEND-M2F \u7684 16.07%\uff09\uff0c\u4f53\u73b0\u51fa\u5176\u5728\u65b0\u8bf4\u8bdd\u4eba\u68c0\u6d4b\u4e0e\u65f6\u5e8f\u5efa\u6a21\u65b9\u9762\u7684\u5f3a\u5927\u6cdb\u5316\u80fd\u529b\u3002<\/p>\n\n\n\n<p><\/p>\n\n\n\n<figure class=\"wp-block-image size-full\"><img loading=\"lazy\" width=\"507\" height=\"880\" src=\"http:\/\/139.9.1.231\/wp-content\/uploads\/2025\/08\/image-32.png\" alt=\"\" class=\"wp-image-28395\" srcset=\"http:\/\/139.9.1.231\/wp-content\/uploads\/2025\/08\/image-32.png 507w, http:\/\/139.9.1.231\/wp-content\/uploads\/2025\/08\/image-32-173x300.png 173w\" sizes=\"(max-width: 507px) 100vw, 507px\" \/><figcaption><strong><strong><em>\u8868<\/em><\/strong><strong><em>2.\u00a0S2SND \u6a21\u578b\u4e0e\u5176\u4ed6\u65b9\u6cd5\u5728 DIHARD-III \u6d4b\u8bd5\u96c6\u4e0a\u7684\u5bf9\u6bd4\u7ed3\u679c<\/em><\/strong><\/strong><\/figcaption><\/figure>\n\n\n\n<p>\u5982\u88682\u6240\u793a\uff0c\u5728\u66f4\u65b0\u4e00\u4ee3\u7684\u00a0DIHARD-III \u6570\u636e\u96c6\u4e0a\uff0cS2SND \u540c\u6837\u8868\u73b0\u51fa\u8272\u3002\u5728 0.80s \u5ef6\u8fdf\u8bbe\u7f6e\u4e0b\u7684\u5728\u7ebf\u6a21\u5f0f\uff0cS2SND-Medium \u8fbe\u5230 17.12% DER\uff0c\u660e\u663e\u4f18\u4e8e\u5404\u7c7b\u5728\u7ebf\u7cfb\u7edf\u3002\u5b8c\u6210\u79bb\u7ebf\u91cd\u89e3\u7801\u540e\uff0cDER \u8fdb\u4e00\u6b65\u964d\u81f3 15.13%\u3002\u503c\u5f97\u4e00\u63d0\u7684\u662f\uff0cS2SND \u7684\u5728\u7ebf DER \u5df2\u903c\u8fd1\u751a\u81f3\u4f18\u4e8e\u8bb8\u591a\u79bb\u7ebf\u7cfb\u7edf\uff08\u5982 EEND-M2F \u7684 16.07%\uff09\uff0c\u4f53\u73b0\u51fa\u5176\u5728\u65b0\u8bf4\u8bdd\u4eba\u68c0\u6d4b\u4e0e\u65f6\u5e8f\u5efa\u6a21\u65b9\u9762\u7684\u5f3a\u5927\u6cdb\u5316\u80fd\u529b\u3002<\/p>\n\n\n\n<p>\u9664\u4e86\u8bc6\u522b\u201c\u8c01\u5728\u4ec0\u4e48\u65f6\u5019\u8bf4\u8bdd\u201d\uff0c\u8bf4\u8bdd\u4eba\u65e5\u5fd7\u7cfb\u7edf\u8fd8\u5e94\u6709\u6548\u4f30\u8ba1\u97f3\u9891\u4e2d\u201c\u6709\u591a\u5c11\u4eba\u53c2\u4e0e\u5bf9\u8bdd\u201d\u3002\u56fe3\u5c55\u793a\u4e86\u4e0d\u540c\u7cfb\u7edf\u5728\u00a0DIHARD-III \u6d4b\u8bd5\u96c6\u4e0a\u7684\u8bf4\u8bdd\u4eba\u6570\u91cf\u9884\u6d4b\u6df7\u6dc6\u77e9\u9635\u3002\u76f8\u6bd4 EEND \u548c\u805a\u7c7b\u65b9\u6cd5\uff0cS2SND \u7684\u9884\u6d4b\u7ed3\u679c\u66f4\u96c6\u4e2d\u4e8e\u4e3b\u5bf9\u89d2\u7ebf\uff0c\u8868\u660e\u5176\u5728\u4e0d\u540c\u8bf4\u8bdd\u4eba\u6570\u573a\u666f\u4e0b\u5177\u6709\u66f4\u5f3a\u7684\u7a33\u5b9a\u6027\u548c\u6cdb\u5316\u80fd\u529b\u3002\u8fd9\u79cd\u80fd\u529b\u5f97\u76ca\u4e8e S2SND \u5728\u8bad\u7ec3\u9636\u6bb5\u5c06\u8bf4\u8bdd\u4eba\u68c0\u6d4b\u548c\u8868\u5f81\u8054\u5408\u4f18\u5316\u7684\u8bbe\u8ba1\uff0c\u4f7f\u6a21\u578b\u80fd\u591f\u5728\u65e0\u9700\u5148\u9a8c\u805a\u7c7b\u7684\u60c5\u51b5\u4e0b\uff0c\u81ea\u4e3b\u53d1\u73b0\u65b0\u8bf4\u8bdd\u4eba\u7684\u51fa\u73b0\u5e76\u6301\u7eed\u8ddf\u8e2a\uff0c\u4ece\u800c\u81ea\u7136\u5b8c\u6210\u8bf4\u8bdd\u4eba\u6570\u91cf\u7684\u5224\u65ad\u4e0e\u66f4\u65b0\u3002<\/p>\n\n\n\n<p><strong><em>\u56fe3.\u00a0DIHARD-III \u6d4b\u8bd5\u96c6\u4e0a\u7684\u8bf4\u8bdd\u4eba\u6570\u91cf\u9884\u6d4b\u6df7\u6dc6\u77e9\u9635\u3002Pyannote.audio v3.1\u3001VBx \u548c DiaPer \u7684\u7ed3\u679c\u7531\u5404\u81ea\u4f5c\u8005\u63d0\u4f9b\u3002\u8bc4\u4f30\u8fc7\u7a0b\u4e2d\u672a\u4f7f\u7528 Oracle VAD\u3002<\/em><\/strong><\/p>\n\n\n\n<figure class=\"wp-block-image size-large\"><img loading=\"lazy\" width=\"1024\" height=\"344\" src=\"http:\/\/139.9.1.231\/wp-content\/uploads\/2025\/08\/image-33-1024x344.png\" alt=\"\" class=\"wp-image-28397\" srcset=\"http:\/\/139.9.1.231\/wp-content\/uploads\/2025\/08\/image-33-1024x344.png 1024w, http:\/\/139.9.1.231\/wp-content\/uploads\/2025\/08\/image-33-300x101.png 300w, http:\/\/139.9.1.231\/wp-content\/uploads\/2025\/08\/image-33-768x258.png 768w, http:\/\/139.9.1.231\/wp-content\/uploads\/2025\/08\/image-33.png 1047w\" sizes=\"(max-width: 1024px) 100vw, 1024px\" \/><\/figure>\n","protected":false},"excerpt":{"rendered":"<p>\u8bba\u6587\u9898\u76ee\uff1aSequence-to-Sequence Neural Diarization with Autom &hellip; <a href=\"http:\/\/139.9.1.231\/index.php\/2025\/08\/19\/s2snd\/\" class=\"more-link\">\u7ee7\u7eed\u9605\u8bfb<span class=\"screen-reader-text\">S2SND  \u5728\u7ebf\/\u79bb\u7ebf\u7edf\u4e00\u63a8\u7406\u7684\u9ad8\u7cbe\u5ea6\u8bf4\u8bdd\u4eba\u65e5\u5fd7\u65b0\u6846\u67b6<\/span><\/a><\/p>\n","protected":false},"author":1,"featured_media":0,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":[],"categories":[4,9,34],"tags":[],"_links":{"self":[{"href":"http:\/\/139.9.1.231\/index.php\/wp-json\/wp\/v2\/posts\/28354"}],"collection":[{"href":"http:\/\/139.9.1.231\/index.php\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"http:\/\/139.9.1.231\/index.php\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"http:\/\/139.9.1.231\/index.php\/wp-json\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"http:\/\/139.9.1.231\/index.php\/wp-json\/wp\/v2\/comments?post=28354"}],"version-history":[{"count":28,"href":"http:\/\/139.9.1.231\/index.php\/wp-json\/wp\/v2\/posts\/28354\/revisions"}],"predecessor-version":[{"id":28398,"href":"http:\/\/139.9.1.231\/index.php\/wp-json\/wp\/v2\/posts\/28354\/revisions\/28398"}],"wp:attachment":[{"href":"http:\/\/139.9.1.231\/index.php\/wp-json\/wp\/v2\/media?parent=28354"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"http:\/\/139.9.1.231\/index.php\/wp-json\/wp\/v2\/categories?post=28354"},{"taxonomy":"post_tag","embeddable":true,"href":"http:\/\/139.9.1.231\/index.php\/wp-json\/wp\/v2\/tags?post=28354"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}