{"id":31165,"date":"2026-07-11T13:46:20","date_gmt":"2026-07-11T05:46:20","guid":{"rendered":"http:\/\/139.9.1.231\/?p=31165"},"modified":"2026-07-11T13:46:22","modified_gmt":"2026-07-11T05:46:22","slug":"asr-moe","status":"publish","type":"post","link":"http:\/\/139.9.1.231\/index.php\/2026\/07\/11\/asr-moe\/","title":{"rendered":"ASR\u8bed\u97f3\u8bc6\u522b-MOE\u67b6\u6784\u8bba\u6587"},"content":{"rendered":"\n<div class=\"wp-block-group has-light-pink-background-color has-background\"><div class=\"wp-block-group__inner-container\">\n<p><strong>MOE- Conformer : <\/strong><\/p>\n\n\n\n<ul><li><strong>Mixture-of-Expert Conformer for Streaming Multilingual ASR:<a href=\"https:\/\/arxiv.org\/abs\/2305.15663\">https:\/\/arxiv.org\/abs\/2305.15663<\/a><\/strong><\/li><li><strong>Parameter-Efficient Conformers: <a href=\"https:\/\/arxiv.org\/pdf\/2209.08326\">https:\/\/arxiv.org\/pdf\/2209.08326<\/a><\/strong><\/li><\/ul>\n\n\n\n<p><strong>MoE Adapter<\/strong>:  <a href=\"https:\/\/arxiv.org\/pdf\/2601.02967\"><strong>https:\/\/arxiv.org\/pdf\/2601.02967<\/strong><\/a><\/p>\n<\/div><\/div>\n\n\n\n\n\n<h2><strong><a href=\"https:\/\/arxiv.org\/abs\/2305.15663\" target=\"_blank\" rel=\"noreferrer noopener\">\u8bba\u6587\uff1aMixture-of-Expert Conformer for Streaming Multilingual ASR<\/a><\/strong><\/h2>\n\n\n\n<figure class=\"wp-block-image size-full is-resized\"><img loading=\"lazy\" src=\"http:\/\/139.9.1.231\/wp-content\/uploads\/2026\/05\/image-11.png\" alt=\"\" class=\"wp-image-31171\" width=\"472\" height=\"511\" srcset=\"http:\/\/139.9.1.231\/wp-content\/uploads\/2026\/05\/image-11.png 649w, http:\/\/139.9.1.231\/wp-content\/uploads\/2026\/05\/image-11-277x300.png 277w\" sizes=\"(max-width: 472px) 100vw, 472px\" \/><\/figure>\n\n\n\n<p>\u8fd9\u7bc7\u8bba\u6587 <em>Mixture-of-Expert Conformer for Streaming Multilingual ASR<\/em> \u8ba8\u8bba\u7684\u662f\u4e00\u4e2a\u66f4\u504f\u5de5\u4e1a\u90e8\u7f72\u7684\u95ee\u9898\uff1a\u5982\u4f55\u8ba9\u4e00\u4e2a\u6d41\u5f0f\u7aef\u5230\u7aef ASR \u6a21\u578b\u540c\u65f6\u652f\u6301\u591a\u79cd\u8bed\u8a00\uff0c\u53c8\u4e0d\u628a\u63a8\u7406\u6210\u672c\u63a8\u5230\u7aef\u4fa7\u8bbe\u5907\u96be\u4ee5\u627f\u53d7\u7684\u7a0b\u5ea6\u3002\u4f5c\u8005\u9009\u62e9\u7684\u8def\u7ebf\u662f\u628a Mixture-of-Experts \u653e\u8fdb Conformer\uff0c\u628a\u6a21\u578b\u603b\u5bb9\u91cf\u505a\u5927\uff0c\u4f46\u6bcf\u6b21\u63a8\u7406\u53ea\u6fc0\u6d3b\u4e00\u5c0f\u90e8\u5206\u53c2\u6570\u3002<\/p>\n\n\n\n<h3>\u6458\u8981\uff1a\u5bb9\u91cf\u53d8\u5927\uff0c\u6fc0\u6d3b\u53c2\u6570\u4e0d\u7ebf\u6027\u53d8\u5927<\/h3>\n\n\n\n<p>\u8bba\u6587\u63d0\u51fa\u7684\u6a21\u578b\u662f\u5728\u6d41\u5f0f\u591a\u8bed\u79cd Conformer \u4e2d\u52a0\u5165 MoE \u5c42\u3002MoE \u5c42\u7531\u591a\u4e2a FFN \u4e13\u5bb6\u548c\u4e00\u4e2a softmax gate \u7ec4\u6210\uff0c\u6bcf\u4e2a\u8f93\u5165\u5e27\u53ea\u9009\u62e9\u6743\u91cd\u6700\u9ad8\u7684\u4e24\u4e2a\u4e13\u5bb6\u53c2\u4e0e\u8ba1\u7b97\u3002\u8fd9\u6837\uff0c\u4e13\u5bb6\u603b\u6570\u53ef\u4ee5\u589e\u52a0\uff0c\u6a21\u578b\u603b\u5bb9\u91cf\u4e5f\u53ef\u4ee5\u589e\u52a0\uff0c\u4f46\u63a8\u7406\u65f6\u6fc0\u6d3b\u7684\u4e13\u5bb6\u6570\u56fa\u5b9a\uff0c\u56e0\u6b64\u8ba1\u7b97\u548c\u6fc0\u6d3b\u53c2\u6570\u4e0d\u4f1a\u968f\u4e13\u5bb6\u6570\u91cf\u7ebf\u6027\u589e\u957f\u3002<\/p>\n\n\n<p><!-- formulas-added:2305-moe-routing --><\/p>\n\n\n<p>\u8bba\u6587\u4e2d\u7684 gate \u5148\u5bf9\u7b2c <code>l<\/code> \u5c42\u8f93\u5165 <code>x<\/code> \u505a\u7ebf\u6027\u6620\u5c04\uff0c\u518d\u901a\u8fc7 softmax \u5f97\u5230\u4e13\u5bb6\u6743\u91cd\uff1a<\/p>\n\n\n\n\\(\ng_l=\\mathrm{Softmax}(W_l\\cdot x)\n\\)\n\n\n\n<p>\u968f\u540e\u53ea\u53d6 top-2 expert\uff0c\u5e76\u628a\u4e24\u4e2a expert \u7684\u8f93\u51fa\u6309 gate \u6743\u91cd\u52a0\u6743\u6c42\u548c\uff1a<\/p>\n\n\n\n\\(\ny=\\sum_{i=1}^{2}g_{l,i}\\cdot e_{l,i}\n\\)\n\n\n\n<p>\u5b9e\u9a8c\u8986\u76d6 12 \u4e2a\u8bed\u8a00 locale\u3002\u76f8\u5bf9\u4e8e 180M \u53c2\u6570\u7684\u591a\u8bed\u79cd cascaded Conformer baseline\uff0cMoE-End \u6a21\u578b\u628a\u5e73\u5747 WER \u4ece 11.33 \u964d\u5230 9.98\uff0c\u7ea6 11.9% \u76f8\u5bf9\u6539\u5584\u3002\u4e0e\u540c\u7b49\u603b\u89c4\u6a21\u7684 dense baseline \u76f8\u6bd4\uff0cMoE \u8fbe\u5230\u7c7b\u4f3c WER\uff0c\u4f46\u63a8\u7406\u6fc0\u6d3b\u53c2\u6570\u7ea6\u4e3a 211M\uff0c\u5bf9\u6bd4 dense \u7684 400M \u66f4\u7701\u3002\u518d\u7ed3\u5408\u591a\u8bed\u79cd neural LM \u505a shallow fusion\uff0c\u5e73\u5747 WER \u8fd8\u80fd\u8fdb\u4e00\u6b65\u76f8\u5bf9\u964d\u4f4e\u7ea6 3%\u3002<\/p>\n\n\n\n<h3>\u5f15\u8a00\uff1a\u591a\u8bed\u79cd\u7edf\u4e00\u6a21\u578b\u7684\u5bb9\u91cf\u95ee\u9898<\/h3>\n\n\n\n<p>\u591a\u8bed\u79cd\u7aef\u5230\u7aef ASR \u7684\u5438\u5f15\u529b\u5f88\u76f4\u63a5\uff1a\u7528\u4e00\u4e2a\u6a21\u578b\u8bc6\u522b\u591a\u79cd\u8bed\u8a00\uff0c\u964d\u4f4e\u7ef4\u62a4\u548c\u90e8\u7f72\u590d\u6742\u5ea6\u3002\u8fc7\u53bb\u51e0\u5e74\uff0cCTC\u3001LSTM\u3001attention-based \u6a21\u578b\u4ee5\u53ca\u6d41\u5f0f RNN-T \u90fd\u5728\u591a\u8bed\u79cd ASR \u4e0a\u53d6\u5f97\u4e86\u8fdb\u5c55\u3002\u5c24\u5176\u662f\u7aef\u4fa7\u6d41\u5f0f\u573a\u666f\uff0c\u6a21\u578b\u65e2\u8981\u6709\u8bc6\u522b\u8d28\u91cf\uff0c\u53c8\u8981\u6ee1\u8db3\u4f4e\u5ef6\u8fdf\u548c\u4f4e\u8ba1\u7b97\u3002<\/p>\n\n\n\n<p>\u7ecf\u9a8c\u4e0a\uff0c\u6a21\u578b\u5bb9\u91cf\u8d8a\u5927\uff0c\u591a\u8bed\u79cd ASR \u8d8a\u5bb9\u6613\u53d7\u76ca\u3002Whisper\u3001USM \u7b49\u5927\u578b\u6a21\u578b\u4e5f\u8bf4\u660e\u4e86\u5927\u6570\u636e\u548c\u5927\u6a21\u578b\u5bf9\u8bed\u97f3\u8bc6\u522b\u8d28\u91cf\u7684\u63a8\u52a8\u4f5c\u7528\u3002\u4f46\u5927\u6a21\u578b\u7684\u4ee3\u4ef7\u662f\u8bad\u7ec3\u548c\u63a8\u7406\u6210\u672c\u3002\u5bf9\u7aef\u4fa7\u5e94\u7528\u6765\u8bf4\uff0c\u4e0d\u80fd\u7b80\u5355\u628a\u6a21\u578b\u6269\u5927\u5230\u6570\u5341\u4ebf\u53c2\u6570\u3002<\/p>\n\n\n\n<p>\u5df2\u6709\u4e00\u4e9b\u6548\u7387\u65b9\u6848\u4f9d\u8d56\u8bed\u8a00\u76f8\u5173\u7ec4\u4ef6\uff0c\u6bd4\u5982\u6309\u8bed\u8a00\u9009\u62e9 adapter \u6216\u4e8c\u9636\u6bb5\u6a21\u578b\u3002\u4f46\u6d41\u5f0f\u573a\u666f\u91cc\uff0c\u7a33\u5b9a\u9884\u6d4b\u8bed\u8a00\u4fe1\u606f\u672c\u8eab\u5c31\u4e0d\u5bb9\u6613\uff0c\u8fd8\u53ef\u80fd\u5f15\u5165\u9519\u8bef\u4f20\u64ad\u3002\u672c\u6587\u7684 MoE \u8def\u7ebf\u66f4\u76f4\u63a5\uff1a\u7531\u8f93\u5165\u8868\u793a\u52a8\u6001\u9009\u62e9\u4e13\u5bb6\uff0c\u4e0d\u9700\u8981\u663e\u5f0f\u8bed\u8a00\u6807\u7b7e\uff0c\u4e5f\u4e0d\u9700\u8981 ground-truth language information\u3002<\/p>\n\n\n\n<h3>\u76f8\u5173\u5de5\u4f5c\uff1a\u4e13\u5bb6\u6a21\u578b\u4e0e\u8bed\u8a00\u4fe1\u606f<\/h3>\n\n\n\n<p>\u8bba\u6587\u628a\u81ea\u5df1\u7684\u65b9\u6cd5\u653e\u5728\u51e0\u7c7b\u5de5\u4f5c\u4e4b\u95f4\u6bd4\u8f83\u3002\u7b2c\u4e00\u7c7b\u662f ASR \u4e2d\u5df2\u6709\u7684 MoE \u6a21\u578b\uff0c\u4f46\u8bb8\u591a\u5de5\u4f5c\u504f\u5355\u8bed\u79cd\uff0c\u6216\u8005\u9700\u8981\u989d\u5916\u7684\u5171\u4eab embedding \u7f51\u7edc\u6765\u505a\u4e13\u5bb6\u8def\u7531\u3002\u7b2c\u4e8c\u7c7b\u662f NLP \u548c\u89c6\u89c9\u4e2d\u7684 MoE\uff0c\u6bd4\u5982 Switch Transformer \u6216 DeepMoE\uff0c\u4e0d\u8fc7\u8fd9\u4e9b\u7ed3\u6784\u5728 ASR \u5c24\u5176\u662f\u6d41\u5f0f\u591a\u8bed\u79cd ASR \u4e2d\u7684\u76f4\u63a5\u6548\u679c\u5e76\u4e0d\u786e\u5b9a\u3002<\/p>\n\n\n\n<p>\u7b2c\u4e09\u7c7b\u662f informed-expert\uff1a\u6a21\u578b\u6839\u636e\u5df2\u77e5\u8bed\u8a00\u4fe1\u606f\u9009\u62e9\u67d0\u4e2a\u8bed\u8a00\u4e13\u5bb6\u3001adapter \u6216\u4e8c\u9636\u6bb5\u6a21\u5757\u3002\u8fd9\u79cd\u505a\u6cd5\u5728\u6709\u53ef\u9760\u8bed\u8a00\u6807\u7b7e\u65f6\u5f88\u81ea\u7136\uff0c\u4f46\u90e8\u7f72\u4e2d\u4f1a\u9047\u5230\u4e24\u4e2a\u9ebb\u70e6\uff1a\u8bed\u8a00\u4fe1\u606f\u8981\u4e48\u6765\u81ea\u5916\u90e8\uff0c\u8981\u4e48\u9700\u8981\u6a21\u578b\u5148\u9884\u6d4b\uff1b\u4e00\u65e6\u9884\u6d4b\u9519\u4e86\uff0c\u540e\u9762\u7684\u4e13\u5bb6\u9009\u62e9\u4e5f\u4f1a\u53d7\u5f71\u54cd\u3002\u672c\u6587\u7684 MoE \u4e0d\u663e\u5f0f\u4f7f\u7528\u8bed\u8a00\u4fe1\u606f\uff0crouting \u7531\u6a21\u578b\u4ece\u58f0\u5b66\u8868\u793a\u4e2d\u5b66\u51fa\u6765\u3002<\/p>\n\n\n\n<h3>MoE Conformer\uff1a\u628a\u4e13\u5bb6\u653e\u5728 FFN \u4f4d\u7f6e<\/h3>\n\n\n\n<p>\u57fa\u7840\u6a21\u5757\u662f Conformer\u3002\u4e00\u4e2a Conformer layer \u901a\u5e38\u5305\u542b\u4e24\u4e2a FFN\u3001\u4e2d\u95f4\u7684 self-attention \u548c convolution\u3002\u4f5c\u8005\u628a MoE \u4e3b\u8981\u7528\u4e8e\u66ff\u6362 Conformer \u91cc\u7684 FFN\uff0c\u5c24\u5176\u662f end FFN\u3002\u6bcf\u4e2a MoE \u5c42\u5305\u542b\u591a\u4e2a FFN \u4e13\u5bb6\u548c\u4e00\u4e2a router\u3002<\/p>\n\n\n\n<p>\u5bf9\u6bcf\u4e00\u5e27\u8868\u793a\uff0crouter \u901a\u8fc7 softmax \u8ba1\u7b97\u5404\u4e13\u5bb6\u6743\u91cd\uff0c\u7136\u540e\u9009\u51fa top-2 \u4e13\u5bb6\u3002\u4e24\u4e2a\u4e13\u5bb6\u7684\u8f93\u51fa\u6309\u8def\u7531\u6743\u91cd\u52a0\u6743\u6c42\u548c\uff0c\u5f97\u5230\u8be5 MoE \u5c42\u8f93\u51fa\u3002\u8bad\u7ec3\u548c\u63a8\u7406\u90fd\u4f7f\u7528 top-2\u3002\u4e3a\u4e86\u9632\u6b62\u4e13\u5bb6\u4f7f\u7528\u4e0d\u5747\u8861\uff0c\u8bba\u6587\u52a0\u5165\u8f85\u52a9\u8d1f\u8f7d\u5747\u8861\u635f\u5931\uff0c\u8ba9\u4e0d\u540c\u4e13\u5bb6\u90fd\u6709\u673a\u4f1a\u88ab\u8bad\u7ec3\u5230\u3002<\/p>\n\n\n\n<p>\u8fd9\u4e2a\u8bbe\u8ba1\u7684\u5173\u952e\u662f\u7a00\u758f\u6fc0\u6d3b\u3002\u6bd4\u5982\u603b\u5171\u6709 8 \u4e2a\u300116 \u4e2a\u6216 24 \u4e2a\u4e13\u5bb6\u65f6\uff0c\u6bcf\u5e27\u4ecd\u53ea\u8d70\u4e24\u4e2a\u4e13\u5bb6\u3002\u603b\u53c2\u6570\u4ee3\u8868\u6a21\u578b\u6f5c\u5728\u5bb9\u91cf\uff0c\u6fc0\u6d3b\u53c2\u6570\u4ee3\u8868\u63a8\u7406\u6210\u672c\uff1bMoE \u7684\u4f18\u52bf\u5c31\u5728\u4e8e\u8ba9\u8fd9\u4e24\u8005\u4e0d\u518d\u5b8c\u5168\u7ed1\u5b9a\u3002<\/p>\n\n\n\n<h3>\u5b9e\u9a8c\u8bbe\u7f6e<\/h3>\n\n\n<p><!-- formulas-added:2305-aux-loss --><\/p>\n\n\n<p>\u6a21\u578b\u8bad\u7ec3\u4f7f\u7528 RNN-T loss\uff0c\u5e76\u989d\u5916\u52a0\u5165\u4e13\u5bb6\u8d1f\u8f7d\u5747\u8861\u9879\u3002\u8bba\u6587\u4e2d\u7684 auxiliary loss \u5199\u6210\uff1a<\/p>\n\n\n\n\\(\nl_{\\mathrm{aux}}=\\frac{1}{N}\\sum_{i=1}^{N}c_i\\cdot m_i\n\\)\n\n\n\n<p>\u5176\u4e2d <code>m_i<\/code> \u662f\u7b2c <code>i<\/code> \u4e2a expert \u7684\u5e73\u5747 gate\uff0c<code>c_i<\/code> \u662f top-2 \u8def\u7531\u4e2d\u8be5 expert \u88ab\u9009\u62e9\u7684\u8ba1\u6570\u3002<\/p>\n\n\n\n<h4>\u6570\u636e<\/h4>\n\n\n\n<p>\u5b9e\u9a8c\u4f7f\u7528 12 \u4e2a\u8bed\u8a00 locale\uff1a\u7f8e\u5f0f\u82f1\u8bed\u3001\u4e2d\u6587\u3001\u6cd5\u8bed\u3001\u5fb7\u8bed\u3001\u65e5\u8bed\u3001\u7f8e\u5f0f\u897f\u73ed\u7259\u8bed\u3001\u897f\u73ed\u7259\u897f\u73ed\u7259\u8bed\u3001\u963f\u62c9\u4f2f\u8bed\u3001\u610f\u5927\u5229\u8bed\u3001\u5370\u5730\u8bed\u3001\u8461\u8404\u7259\u8bed\u548c\u4fc4\u8bed\u3002\u8bad\u7ec3\u6570\u636e\u6765\u81ea Voice Search\u3001YouTube \u7b49\u591a\u4e2a\u57df\uff0c\u603b\u8ba1\u7ea6 139.4M \u6761\u4eba\u5de5\u8f6c\u5199\u533f\u540d\u8bed\u97f3\u3002\u4e0d\u540c\u8bed\u8a00\u6570\u636e\u91cf\u5dee\u5f02\u5f88\u5927\uff0c\u4ece 0.5M \u5230 25.2M utterances \u4e0d\u7b49\u3002<\/p>\n\n\n\n<p>\u6d4b\u8bd5\u96c6\u6765\u81ea Voice Search \u6d41\u91cf\uff0c\u6bcf\u4e2a\u8bed\u8a00\u5927\u7ea6 1.4K \u5230 10K \u6761 utterances\uff0c\u4e0e\u8bad\u7ec3\u96c6\u4e0d\u91cd\u53e0\u3002\u8bc4\u4ef7\u6307\u6807\u662f WER\uff1b\u5bf9\u4e2d\u6587\u7b49\u8bed\u8a00\uff0c\u8bba\u6587\u6309\u5b57\u7b26\u8ba1\u7b97\u9519\u8bef\u7387\u3002<\/p>\n\n\n\n<h4>\u6a21\u578b\u7ec6\u8282<\/h4>\n\n\n\n<p>baseline \u662f\u4e00\u4e2a\u8bed\u8a00\u65e0\u5173\u7684\u591a\u8bed\u79cd transducer \u6a21\u578b\uff0c\u5305\u542b 7 \u5c42 causal Conformer encoder \u548c 10 \u5c42 non-causal cascaded encoder\u3002causal \u90e8\u5206\u4fdd\u8bc1\u6d41\u5f0f\uff0cnon-causal cascaded \u90e8\u5206\u63d0\u4f9b\u7ea6 0.9 \u79d2\u53f3\u4e0a\u4e0b\u6587\u3002\u6a21\u578b\u4f7f\u7528 separate decoders \u5206\u522b\u670d\u52a1 causal \u548c non-causal encoder\uff0c\u4ee5\u83b7\u5f97\u66f4\u597d\u8d28\u91cf\u3002baseline \u603b\u53c2\u6570\u7ea6 180M\u3002<\/p>\n\n\n\n<p>MoE \u6539\u9020\u4e3b\u8981\u53d1\u751f\u5728 cascaded encoder\u3002\u4f5c\u8005\u5c1d\u8bd5\u66ff\u6362 start FFN\u3001end FFN \u6216\u4e24\u8005\u90fd\u66ff\u6362\u3002\u6700\u591a\u4f7f\u7528 24 \u4e2a\u4e13\u5bb6\uff0c\u4f46\u6bcf\u6b21\u8bad\u7ec3\u548c\u63a8\u7406\u53ea\u9009 top-2\u3002\u8f93\u5165\u7279\u5f81\u4e3a 128 \u7ef4 log-Mel filterbank\uff0c\u7ecf\u8fde\u7eed\u5e27\u5806\u53e0\u5f62\u6210 512 \u7ef4\u8f93\u5165\uff0c\u5e76\u4e0b\u91c7\u6837\u5230 30ms \u5e27\u7387\uff1b\u8bad\u7ec3\u4e2d\u4f7f\u7528 SpecAug \u589e\u5f3a\u9c81\u68d2\u6027\u3002<\/p>\n\n\n\n<h3>\u7ed3\u679c\u4e0e\u6bd4\u8f83<\/h3>\n\n\n\n<h4>\u6d88\u878d\u5b9e\u9a8c<\/h4>\n\n\n\n<p>\u9996\u5148\u770b MoE \u653e\u5728\u54ea\u91cc\u3002baseline \u5e73\u5747 WER \u4e3a 11.33\u3002\u628a MoE \u653e\u5728 start FFN\uff0c\u5e73\u5747 WER \u4e3a 10.10\uff1b\u653e\u5728 end FFN\uff0c\u5e73\u5747 WER \u4e3a 9.98\uff1b\u4e24\u5904\u90fd\u653e\uff0c\u5e73\u5747 WER \u6700\u597d\uff0c\u4e3a 9.54\u3002\u4e0d\u8fc7\u4e24\u5904\u90fd\u653e\u4f1a\u589e\u52a0\u63a8\u7406\u6fc0\u6d3b\u53c2\u6570\u3002\u4f5c\u8005\u6700\u7ec8\u66f4\u591a\u91c7\u7528 MoE-End\uff0c\u56e0\u4e3a\u5b83\u5728\u8d28\u91cf\u548c\u6548\u7387\u4e4b\u95f4\u66f4\u5747\u8861\u3002<\/p>\n\n\n\n<figure class=\"wp-block-image size-full\"><img loading=\"lazy\" width=\"565\" height=\"505\" src=\"http:\/\/139.9.1.231\/wp-content\/uploads\/2026\/07\/image-16.png\" alt=\"\" class=\"wp-image-31351\" srcset=\"http:\/\/139.9.1.231\/wp-content\/uploads\/2026\/07\/image-16.png 565w, http:\/\/139.9.1.231\/wp-content\/uploads\/2026\/07\/image-16-300x268.png 300w\" sizes=\"(max-width: 565px) 100vw, 565px\" \/><\/figure>\n\n\n\n<p>\u4e13\u5bb6\u6570\u91cf\u65b9\u9762\uff0c8 experts \u7684 MoE-End \u5e73\u5747 WER \u4e3a 9.98\uff1b\u51cf\u5c11\u5230 4 experts \u540e\u4e3a 10.40\uff1b\u51cf\u5c11\u5230 2 experts \u540e\u4e3a 10.58\u3002\u7531\u4e8e\u63a8\u7406\u59cb\u7ec8\u6fc0\u6d3b top-2\uff0c\u4e13\u5bb6\u603b\u6570\u51cf\u5c11\u4e3b\u8981\u5f71\u54cd\u603b\u5bb9\u91cf\u800c\u4e0d\u662f\u6fc0\u6d3b\u53c2\u6570\u3002\u7ed3\u679c\u8bf4\u660e\uff0c\u989d\u5916\u4e13\u5bb6\u786e\u5b9e\u88ab\u6a21\u578b\u5229\u7528\u4e86\u3002<\/p>\n\n\n\n<figure class=\"wp-block-image size-full\"><img loading=\"lazy\" width=\"514\" height=\"535\" src=\"http:\/\/139.9.1.231\/wp-content\/uploads\/2026\/07\/image-17.png\" alt=\"\" class=\"wp-image-31352\" srcset=\"http:\/\/139.9.1.231\/wp-content\/uploads\/2026\/07\/image-17.png 514w, http:\/\/139.9.1.231\/wp-content\/uploads\/2026\/07\/image-17-288x300.png 288w\" sizes=\"(max-width: 514px) 100vw, 514px\" \/><\/figure>\n\n\n\n<p>MoE \u5c42\u6570\u4e5f\u5f88\u91cd\u8981\u3002\u53ea\u5728\u9694\u5c42\u4f7f\u7528 MoE\uff0c\u5e73\u5747 WER \u9000\u5230 10.50\uff1b\u53ea\u5728\u7b2c\u4e00\u4e2a Conformer \u5c42\u4f7f\u7528 MoE\uff0c\u4e3a 10.88\u3002\u5373\u4fbf\u53ea\u52a0\u4e00\u4e2a MoE \u5c42\u4e5f\u6bd4 baseline \u597d\uff0c\u4f46\u5b8c\u6574\u5730\u5728 end FFN \u4f4d\u7f6e\u52a0\u5165 MoE \u624d\u80fd\u53d1\u6325\u4e3b\u8981\u6548\u679c\u3002<\/p>\n\n\n\n<h4>\u4e0e dense baseline \u548c adapter \u6bd4\u8f83<\/h4>\n\n\n\n<p>\u4e0e 180M baseline \u76f8\u6bd4\uff0cMoE-End \u6a21\u578b\u603b\u53c2\u6570\u7ea6 400M\uff0c\u63a8\u7406\u6fc0\u6d3b\u7ea6 211M\uff0c\u5e73\u5747 WER \u4ece 11.33 \u964d\u5230 9.98\u3002\u4e3a\u4e86\u6392\u9664\u201c\u53ea\u662f\u6a21\u578b\u53d8\u5927\u201d\u7684\u56e0\u7d20\uff0c\u4f5c\u8005\u6784\u9020\u4e86\u4e00\u4e2a\u540c\u4e3a 400M \u7684\u5927 dense baseline\u3002\u8fd9\u4e2a dense \u6a21\u578b\u5e73\u5747 WER \u4e5f\u662f 9.98\uff0c\u4f46\u63a8\u7406\u9700\u8981\u6fc0\u6d3b 400M \u53c2\u6570\uff1bMoE \u53ea\u6fc0\u6d3b 211M\uff0c\u7ea6\u4e3a dense \u7684 53%\u3002<\/p>\n\n\n\n<figure class=\"wp-block-image size-full\"><img loading=\"lazy\" width=\"504\" height=\"472\" src=\"http:\/\/139.9.1.231\/wp-content\/uploads\/2026\/07\/image-18.png\" alt=\"\" class=\"wp-image-31353\" srcset=\"http:\/\/139.9.1.231\/wp-content\/uploads\/2026\/07\/image-18.png 504w, http:\/\/139.9.1.231\/wp-content\/uploads\/2026\/07\/image-18-300x281.png 300w\" sizes=\"(max-width: 504px) 100vw, 504px\" \/><\/figure>\n\n\n\n<p>\u4e0e\u57fa\u4e8e ground-truth language information \u7684 adapter \u6a21\u578b\u76f8\u6bd4\uff0cMoE \u7684\u610f\u4e49\u66f4\u660e\u663e\u3002Adapter \u6a21\u578b\u4f9d\u8d56\u771f\u5b9e\u8bed\u8a00\u4fe1\u606f\u9009\u62e9\u5bf9\u5e94\u6a21\u5757\uff1bMoE \u4e0d\u9700\u8981\u8bed\u8a00\u6807\u7b7e\uff0c\u53ea\u6839\u636e\u8f93\u5165\u52a8\u6001\u8def\u7531\u3002\u628a FFN multiplier \u8c03\u5c0f\u5e76\u589e\u52a0\u4e13\u5bb6\u6570\u540e\uff0c16 \u6216 24 experts \u7684 MoE \u5728\u5e73\u5747 WER \u4e0a\u63a5\u8fd1 adapter\uff0c\u4f46\u90e8\u7f72\u4e0a\u5c11\u4e86\u8bed\u8a00\u4fe1\u606f\u4f9d\u8d56\u3002<\/p>\n\n\n\n<h4>Shallow Fusion \u8fdb\u4e00\u6b65\u63d0\u5347<\/h4>\n\n\n\n<p>\u4f5c\u8005\u8fd8\u8bad\u7ec3\u4e86\u4e00\u4e2a 128M \u5de6\u53f3\u7684\u591a\u8bed\u79cd neural LM\uff0c\u5e76\u5728\u89e3\u7801\u65f6\u505a shallow fusion\u3002\u6587\u672c\u6570\u636e\u6765\u81ea 12 \u79cd\u8bed\u8a00\u7684\u76d1\u7763\u8bad\u7ec3\u6587\u672c\u548c\u989d\u5916 text-only \u6570\u636e\u3002\u52a0\u5165 LM \u540e\uff0cMoE \u6a21\u578b\u5e73\u5747 WER \u4ece 9.98 \u8fdb\u4e00\u6b65\u964d\u5230 9.68\uff0c\u7ea6 3% \u76f8\u5bf9\u6539\u5584\u3002<\/p>\n\n\n\n<figure class=\"wp-block-image size-large\"><img loading=\"lazy\" width=\"1024\" height=\"237\" src=\"http:\/\/139.9.1.231\/wp-content\/uploads\/2026\/07\/image-19-1024x237.png\" alt=\"\" class=\"wp-image-31354\" srcset=\"http:\/\/139.9.1.231\/wp-content\/uploads\/2026\/07\/image-19-1024x237.png 1024w, http:\/\/139.9.1.231\/wp-content\/uploads\/2026\/07\/image-19-300x70.png 300w, http:\/\/139.9.1.231\/wp-content\/uploads\/2026\/07\/image-19-768x178.png 768w, http:\/\/139.9.1.231\/wp-content\/uploads\/2026\/07\/image-19.png 1092w\" sizes=\"(max-width: 1024px) 100vw, 1024px\" \/><\/figure>\n\n\n\n<p>\u4e0d\u8fc7\u6539\u5584\u5e76\u975e\u6240\u6709\u8bed\u8a00\u90fd\u4e00\u81f4\u3002\u6cd5\u8bed\u6536\u76ca\u6700\u5927\uff0c\u4e2d\u6587\u548c\u5370\u5730\u8bed\u51fa\u73b0\u9000\u5316\u3002\u4f5c\u8005\u63a8\u6d4b\uff0c\u4e2d\u6587\u9000\u5316\u53ef\u80fd\u4e0e text-only \u6570\u636e\u91cc\u6df7\u5165\u7ca4\u8bed\u8f6c\u5199\u6709\u5173\uff1b\u5370\u5730\u8bed\u5219\u53ef\u80fd\u56e0\u4e3a text-only \u6570\u636e\u89c4\u6a21\u5f88\u5927\u4f46\u4e0e Search \u57df\u4e0d\u5b8c\u5168\u5339\u914d\uff0c\u9700\u8981\u66f4\u597d\u7684\u8fc7\u6ee4\u7b56\u7565\u3002<\/p>\n\n\n\n<h3>\u7ed3\u8bba\uff1aMoE \u7684\u90e8\u7f72\u4ef7\u503c\u5728\u4e8e\u201c\u4e0d\u9700\u8981\u8bed\u8a00\u6807\u7b7e\u201d<\/h3>\n\n\n\n<p>\u8fd9\u7bc7\u8bba\u6587\u5c55\u793a\u4e86 MoE \u5728\u6d41\u5f0f\u591a\u8bed\u79cd ASR \u4e2d\u7684\u4e00\u4e2a\u6e05\u6670\u7528\u9014\uff1a\u7528\u66f4\u5927\u7684\u603b\u5bb9\u91cf\u63d0\u5347\u591a\u8bed\u79cd\u8bc6\u522b\u8d28\u91cf\uff0c\u540c\u65f6\u901a\u8fc7 top-2 \u7a00\u758f\u6fc0\u6d3b\u63a7\u5236\u63a8\u7406\u6210\u672c\u3002\u6700\u91cd\u8981\u7684\u662f\uff0c\u6a21\u578b\u4e0d\u4f9d\u8d56\u8bed\u8a00\u6807\u7b7e\u5b8c\u6210\u4e13\u5bb6\u9009\u62e9\uff0c\u8fd9\u6bd4 adapter \u6216 per-language expert \u5728\u771f\u5b9e\u90e8\u7f72\u4e2d\u66f4\u7701\u5fc3\u3002<\/p>\n\n\n\n<p>\u4ece\u7ed3\u679c\u770b\uff0cMoE-End \u76f8\u5bf9\u4e8e baseline \u6709 11.9% \u5e73\u5747\u76f8\u5bf9 WER \u6539\u5584\uff1b\u4e0e\u540c\u89c4\u6a21 dense \u6a21\u578b\u76f8\u6bd4\uff0c\u8fbe\u5230\u7c7b\u4f3c\u8d28\u91cf\u4f46\u53ea\u6fc0\u6d3b\u7ea6 53% \u53c2\u6570\uff1b\u4e0e\u8bed\u8a00\u6807\u7b7e adapter \u76f8\u6bd4\uff0c\u8d28\u91cf\u63a5\u8fd1\u4f46\u8def\u7531\u66f4\u81ea\u52a8\u3002\u5bf9\u7aef\u4fa7\u3001\u6d41\u5f0f\u3001\u591a\u8bed\u79cd\u8fd9\u4e09\u4e2a\u7ea6\u675f\u540c\u65f6\u5b58\u5728\u7684\u573a\u666f\uff0c\u8fd9\u79cd\u201c\u52a8\u6001\u5bb9\u91cf\u201d\u601d\u8def\u5f88\u503c\u5f97\u7ee7\u7eed\u8ddf\u8fdb\u3002<\/p>\n\n\n\n<p><\/p>\n\n\n\n<h2><strong>\u5feb\u624b\u8bba\u6587\uff1a<a href=\"https:\/\/arxiv.org\/pdf\/2209.08326\" target=\"_blank\" rel=\"noreferrer noopener\">Parameter-Efficient Conformers<\/a>\uff0c\u5229\u7528MOE\u8fdb\u884c\u6a21\u578b\u88c1\u526a<\/strong><\/h2>\n\n\n\n<figure class=\"wp-block-image size-large\"><img loading=\"lazy\" width=\"1024\" height=\"619\" src=\"http:\/\/139.9.1.231\/wp-content\/uploads\/2026\/05\/image-12-1024x619.png\" alt=\"\" class=\"wp-image-31172\" srcset=\"http:\/\/139.9.1.231\/wp-content\/uploads\/2026\/05\/image-12-1024x619.png 1024w, http:\/\/139.9.1.231\/wp-content\/uploads\/2026\/05\/image-12-300x181.png 300w, http:\/\/139.9.1.231\/wp-content\/uploads\/2026\/05\/image-12-768x464.png 768w, http:\/\/139.9.1.231\/wp-content\/uploads\/2026\/05\/image-12.png 1039w\" sizes=\"(max-width: 1024px) 100vw, 1024px\" \/><\/figure>\n\n\n\n<p>\u8fd9\u7bc7\u8bba\u6587 <em>Parameter-Efficient Conformers via Sharing Sparsely-Gated Experts for End-to-End Speech Recognition<\/em> \u5173\u6ce8\u4e00\u4e2a\u975e\u5e38\u5b9e\u9645\u7684\u95ee\u9898\uff1aConformer \u5728\u7aef\u5230\u7aef\u8bed\u97f3\u8bc6\u522b\u91cc\u6548\u679c\u5f88\u597d\uff0c\u4f46\u6a21\u578b\u5c42\u6570\u548c\u53c2\u6570\u91cf\u4e0a\u6765\u4e4b\u540e\uff0c\u8bad\u7ec3\u3001\u90e8\u7f72\u3001\u7aef\u4fa7\u8fd0\u884c\u90fd\u4f1a\u53d8\u91cd\u3002\u4f5c\u8005\u7684\u601d\u8def\u4e0d\u662f\u7b80\u5355\u780d\u5c42\uff0c\u4e5f\u4e0d\u662f\u53ea\u505a\u666e\u901a\u7684\u53c2\u6570\u5171\u4eab\uff0c\u800c\u662f\u628a\u201c\u5171\u4eab Conformer \u5757\u201d\u548c\u201c\u7a00\u758f\u95e8\u63a7\u4e13\u5bb6\u201d\u7ed3\u5408\u8d77\u6765\uff0c\u8ba9\u5c11\u91cf\u53c2\u6570\u88ab\u91cd\u590d\u4f7f\u7528\uff0c\u540c\u65f6\u7528 MoE \u4fdd\u4f4f\u8868\u793a\u5bb9\u91cf\u3002<\/p>\n\n\n\n<h3>\u6458\u8981\uff1a\u5c11\u53c2\u6570\uff0c\u4e0d\u60f3\u5c11\u80fd\u529b<\/h3>\n\n\n\n<p>\u8bba\u6587\u7684\u6838\u5fc3\u76ee\u6807\u662f\u6784\u9020\u4e00\u4e2a\u53c2\u6570\u9ad8\u6548\u7684 Conformer \u7f16\u7801\u5668\u3002\u4f20\u7edf\u8de8\u5c42\u6743\u91cd\u5171\u4eab\u53ef\u4ee5\u51cf\u5c11\u53c2\u6570\uff0c\u4f46\u4e5f\u4f1a\u538b\u7f29\u6a21\u578b\u5bb9\u91cf\uff0c\u5bfc\u81f4\u8bc6\u522b\u6027\u80fd\u4e0b\u964d\u3002\u4f5c\u8005\u63d0\u51fa\u7684\u65b9\u6848\u662f\u5728\u5171\u4eab\u7684 Conformer \u5757\u4e2d\u52a0\u5165 sparsely-gated MoE\uff1a\u7b2c\u4e8c\u4e2a\u524d\u9988\u7f51\u7edc\u4e0d\u518d\u662f\u5355\u4e00\u8def\u5f84\uff0c\u800c\u662f\u4e00\u7ec4\u4e13\u5bb6\uff0c\u7531\u8def\u7531\u5668\u9009\u62e9\u5176\u4e2d\u4e00\u4e2a\u4e13\u5bb6\u53c2\u4e0e\u8ba1\u7b97\u3002\u8fd9\u6837\u603b\u53c2\u6570\u589e\u52a0\u4e86\u4e00\u4e9b\uff0c\u4f46\u6bcf\u6b21\u524d\u5411\u53ea\u6fc0\u6d3b\u4e00\u4e2a\u4e13\u5bb6\uff0c\u8ba1\u7b97\u91cf\u57fa\u672c\u4fdd\u6301\u5728\u975e MoE \u6a21\u578b\u7684\u6c34\u5e73\u3002<\/p>\n\n\n\n<p>\u4e3a\u4e86\u8ba9\u5171\u4eab\u5757\u5728\u4e0d\u540c\u6df1\u5ea6\u4f4d\u7f6e\u4ecd\u80fd\u9002\u914d\u4e0d\u540c\u5c42\u7ea7\u7684\u8868\u793a\uff0c\u8bba\u6587\u8fd8\u8ba9\u8def\u7531\u5668\u548c\u5f52\u4e00\u5316\u5c42\u4fdd\u6301\u72ec\u7acb\uff0c\u800c\u4e0d\u662f\u6240\u6709\u5185\u5bb9\u90fd\u5171\u4eab\u3002\u6700\u540e\uff0c\u4f5c\u8005\u7528\u5168\u53c2\u6570\u6a21\u578b\u4f5c\u4e3a teacher\uff0c\u901a\u8fc7\u9690\u85cf\u5c42\u8868\u793a\u7684\u77e5\u8bc6\u84b8\u998f\u8fdb\u4e00\u6b65\u5f25\u8865\u5171\u4eab\u6a21\u578b\u7684\u80fd\u529b\u635f\u5931\u3002\u5b9e\u9a8c\u663e\u793a\uff0c\u5728 AISHELL-1 \u4e0a\uff0c\u6700\u7ec8\u6a21\u578b\u7528\u7ea6\u4e09\u5206\u4e4b\u4e00\u7684\u7f16\u7801\u5668\u53c2\u6570\u53d6\u5f97\u4e86\u63a5\u8fd1\u5168\u53c2\u6570\u6a21\u578b\u7684 CER\u3002<\/p>\n\n\n\n<h3>\u5f15\u8a00\uff1aConformer \u5f88\u5f3a\uff0c\u4f46\u90e8\u7f72\u4e0d\u8f7b<\/h3>\n\n\n\n<p>\u7aef\u5230\u7aef ASR \u4e2d\uff0cTransformer \u548c Conformer \u5df2\u7ecf\u662f\u5f88\u5e38\u89c1\u7684\u7f16\u7801\u5668\u9009\u62e9\u3002Conformer \u5728 Transformer \u7684\u5168\u5c40\u5efa\u6a21\u57fa\u7840\u4e0a\u52a0\u5165\u5377\u79ef\u6a21\u5757\uff0c\u66f4\u9002\u5408\u8bed\u97f3\u8fd9\u79cd\u65e2\u6709\u957f\u7a0b\u4f9d\u8d56\u3001\u53c8\u6709\u5c40\u90e8\u7ed3\u6784\u7684\u5e8f\u5217\u3002\u76f8\u5bf9\u4f4d\u7f6e\u7f16\u7801\u3001Macaron \u98ce\u683c FFN\u3001\u5377\u79ef\u589e\u5f3a\u7b49\u8bbe\u8ba1\uff0c\u90fd\u8ba9\u5b83\u5728\u8bed\u97f3\u8bc6\u522b\u4e2d\u8868\u73b0\u7a33\u5b9a\u3002<\/p>\n\n\n\n<p>\u95ee\u9898\u5728\u4e8e\uff0c\u8fd9\u7c7b\u6a21\u578b\u5f80\u5f80\u53c2\u6570\u5197\u4f59\u3002\u76f4\u63a5\u5806\u5f88\u591a\u5c42\u53ef\u4ee5\u6362\u6765\u66f4\u5f3a\u8868\u8fbe\uff0c\u4f46\u4e5f\u5e26\u6765\u663e\u5b58\u3001\u5b58\u50a8\u548c\u63a8\u7406\u6210\u672c\u3002\u5df2\u6709\u5de5\u4f5c\u4f1a\u901a\u8fc7\u8de8\u5c42\u5171\u4eab\u53c2\u6570\u964d\u4f4e\u6a21\u578b\u89c4\u6a21\uff0c\u7c7b\u4f3c\u8ba9\u540c\u4e00\u4e2a block \u88ab\u91cd\u590d\u8c03\u7528\u591a\u6b21\u3002\u8fd9\u4e2a\u529e\u6cd5\u7701\u53c2\u6570\uff0c\u4f46\u526f\u4f5c\u7528\u4e5f\u660e\u663e\uff1a\u81ea\u7531\u53c2\u6570\u5c11\u4e86\uff0c\u6a21\u578b\u5bb9\u91cf\u4e0b\u964d\uff0c\u6027\u80fd\u5bb9\u6613\u6389\u3002<\/p>\n\n\n\n<p>\u4f5c\u8005\u7684\u5207\u5165\u70b9\u662f\uff1a\u65e2\u7136\u5171\u4eab\u4f1a\u635f\u5931\u5bb9\u91cf\uff0c\u90a3\u5c31\u5728\u5171\u4eab\u5757\u5185\u90e8\u5f15\u5165 MoE \u6765\u8865\u5bb9\u91cf\uff1b\u65e2\u7136 MoE \u53ef\u4ee5\u7a00\u758f\u6fc0\u6d3b\uff0c\u90a3\u5c31\u53ea\u8ba9\u5c11\u6570\u4e13\u5bb6\u53c2\u4e0e\u4e00\u6b21\u524d\u5411\uff0c\u907f\u514d\u8ba1\u7b97\u91cf\u8ddf\u7740\u603b\u53c2\u6570\u7ebf\u6027\u589e\u957f\u3002\u8fd9\u4e2a\u7ec4\u5408\u7279\u522b\u9002\u5408\u201c\u53c2\u6570\u5c11\u3001\u8ba1\u7b97\u4e0d\u80fd\u592a\u8d35\u201d\u7684\u573a\u666f\u3002<\/p>\n\n\n\n<h3>\u80cc\u666f\uff1aConformer Seq2Seq ASR<\/h3>\n\n\n\n<p>\u8bba\u6587\u4f7f\u7528\u7684\u662f attention-based encoder-decoder \u6846\u67b6\u3002\u7f16\u7801\u5668\u628a\u58f0\u5b66\u7279\u5f81\u5e8f\u5217\u53d8\u6210\u9ad8\u5c42\u8868\u793a\uff0c\u89e3\u7801\u5668\u6309 token \u9010\u6b65\u751f\u6210\u6587\u672c\u5e8f\u5217\uff0c\u8bad\u7ec3\u65f6\u4f18\u5316\u8d1f\u5bf9\u6570\u4f3c\u7136\uff0c\u63a8\u7406\u65f6\u7528 beam search \u627e\u66f4\u53ef\u80fd\u7684\u8f93\u51fa\u3002<\/p>\n\n\n<p><!-- formulas-added:2209-aed-nll --><\/p>\n\n\n<p>\u8bba\u6587\u4e2d\u5148\u628a AED \u7684\u9010 token \u9884\u6d4b\u6982\u7387\u5199\u6210\u4e0b\u9762\u8fd9\u4e2a\u5f62\u5f0f\uff0c\u5176\u4e2d <code><em>y<sub>&lt;s<\/sub><\/em><\/code> \u8868\u793a\u5f53\u524d\u4f4d\u7f6e\u4e4b\u524d\u7684 token \u524d\u7f00\uff1a<\/p>\n\n\n\n\\(\nP(y_s \\mid y_{&lt;s}, x)=\\mathrm{Trfm}(y_{&lt;s},x)\n\\)\n\n\n\n<p>\u5bf9\u5e94\u7684\u6700\u5927\u4f3c\u7136\u8bad\u7ec3\u76ee\u6807\uff0c\u4e5f\u5c31\u662f\u8d1f\u5bf9\u6570\u4f3c\u7136\u635f\u5931\u4e3a\uff1a<\/p>\n\n\n\n\\(\nL_{\\mathrm{nll}}(\\theta)=-\\frac{1}{S}\\sum_{s=1}^{S}\\log P(y_s\\mid y_{&lt;s},x)\n\\)\n\n\n\n<p>Conformer \u5757\u7531\u4e24\u4e2a FFN\u3001\u4e00\u4e2a\u591a\u5934\u81ea\u6ce8\u610f\u529b\u6a21\u5757\u548c\u4e00\u4e2a\u5377\u79ef\u6a21\u5757\u7ec4\u6210\u3002\u4e24\u4e2a FFN \u91c7\u7528\u534a\u6b65\u6b8b\u5dee\u98ce\u683c\uff0c\u6ce8\u610f\u529b\u8d1f\u8d23\u957f\u7a0b\u4f9d\u8d56\uff0c\u5377\u79ef\u8d1f\u8d23\u5c40\u90e8\u6a21\u5f0f\u3002\u672c\u6587\u7684 MoE \u6539\u9020\u53d1\u751f\u5728\u7b2c\u4e8c\u4e2a FFN\uff1a\u4f5c\u8005\u628a\u5b83\u66ff\u6362\u6210\u4e00\u4e2a\u7a00\u758f\u95e8\u63a7\u7684\u4e13\u5bb6\u96c6\u5408\uff0c\u4e5f\u5c31\u662f MoE-Conformer block\u3002<\/p>\n\n\n<p><!-- formulas-added:2209-conformer-block --><\/p>\n\n\n<p>\u8bba\u6587\u628a\u4e00\u4e2a MoE-Conformer block \u7684\u8ba1\u7b97\u5199\u6210\u56db\u6b65\u3002\u6700\u540e\u4e00\u6b65\u4e2d\uff0c\u7b2c\u4e8c\u4e2a FFN \u88ab\u66ff\u6362\u6210 MoE \u7248\u672c\uff1a<\/p>\n\n\n\n\\(\n\\begin{aligned}\nz_t^{(1)} &amp;= z_t + \\frac{1}{2}\\mathrm{FFN}(z_t),\\\\\nz_t^{(2)} &amp;= z_t^{(1)} + \\mathrm{MHSA}(z_t^{(1)}),\\\\\nz_t^{(3)} &amp;= z_t^{(2)} + \\mathrm{Conv}(z_t^{(2)}),\\\\\n\\hat{z}_t &amp;= \\mathrm{LayerNorm}\\left(z_t^{(3)}+\\frac{1}{2}\\mathrm{FFN}^{(\\mathrm{MoE})}(z_t^{(3)})\\right).\n\\end{aligned}\n\\)\n\n\n\n<h3>\u65b9\u6cd5\uff1a\u5171\u4eab\u7a00\u758f\u95e8\u63a7\u4e13\u5bb6<\/h3>\n\n\n\n<h4>Conformer \u53c2\u6570\u5171\u4eab<\/h4>\n\n\n\n<p>\u4f5c\u8005\u628a\u8fde\u7eed\u7684 C \u4e2a Conformer \u5757\u770b\u4f5c\u4e00\u7ec4\uff0c\u518d\u5806\u53e0 G \u7ec4\u3002\u4e0d\u540c\u7ec4\u4e2d\u76f8\u540c\u4f4d\u7f6e\u7684\u5757\u5171\u4eab\u53c2\u6570\uff0c\u76f8\u5f53\u4e8e\u4e00\u7ec4\u5757\u88ab\u9012\u5f52\u8c03\u7528 G \u6b21\u3002\u8fd9\u6837\u505a\u7684\u597d\u5904\u5f88\u76f4\u63a5\uff1a\u5982\u679c\u60f3\u8981 12 \u6b21\u53d8\u6362\uff0c\u4e0d\u4e00\u5b9a\u771f\u7684\u4fdd\u5b58 12 \u5957\u7f16\u7801\u5668\u53c2\u6570\uff0c\u53ef\u4ee5\u7528\u66f4\u5c11\u7684\u5757\u53cd\u590d\u8ba1\u7b97\u3002<\/p>\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\/2026\/07\/image-5.png\" alt=\"\" class=\"wp-image-31330\" width=\"154\" height=\"288\" srcset=\"http:\/\/139.9.1.231\/wp-content\/uploads\/2026\/07\/image-5.png 307w, http:\/\/139.9.1.231\/wp-content\/uploads\/2026\/07\/image-5-160x300.png 160w\" sizes=\"(max-width: 154px) 100vw, 154px\" \/><\/figure><\/div>\n\n\n\n<p>\u4f46\u662f\u5171\u4eab\u4e0d\u662f\u767d\u6765\u7684\u3002\u5171\u4eab\u5757\u5728\u6d45\u5c42\u548c\u6df1\u5c42\u9762\u5bf9\u7684\u8868\u793a\u5206\u5e03\u4e0d\u4e00\u6837\uff0c\u5982\u679c\u5b8c\u5168\u7528\u540c\u4e00\u5957\u53c2\u6570\u3001\u540c\u4e00\u5957\u8def\u7531\u3001\u540c\u4e00\u5957\u5f52\u4e00\u5316\u7edf\u8ba1\uff0c\u6a21\u578b\u4f1a\u5f88\u96be\u540c\u65f6\u9002\u914d\u4e0d\u540c\u6df1\u5ea6\u7684\u8868\u793a\u3002\u56e0\u6b64\u540e\u9762\u4e24\u4e2a\u8bbe\u8ba1\uff0c\u4e5f\u5c31\u662f\u72ec\u7acb\u8def\u7531\u5668\u548c\u72ec\u7acb\u5f52\u4e00\u5316\uff0c\u5c31\u53d8\u5f97\u5f88\u5173\u952e\u3002<\/p>\n\n\n\n<h4>MoE \u52a8\u6001\u8def\u7531<\/h4>\n\n\n\n<p>MoE \u6a21\u5757\u7531 E \u4e2a\u5e76\u884c FFN \u4e13\u5bb6\u548c\u4e00\u4e2a router \u7ec4\u6210\u3002\u5bf9\u6bcf\u4e2a\u65f6\u95f4\u6b65\u7684\u8868\u793a\uff0crouter \u8f93\u51fa\u5404\u4e13\u5bb6\u7684\u6982\u7387\uff0c\u8bba\u6587\u91c7\u7528 top-1 \u9009\u62e9\uff0c\u53ea\u6fc0\u6d3b\u5f97\u5206\u6700\u9ad8\u7684\u4e13\u5bb6\u3002\u4e5f\u5c31\u662f\u8bf4\uff0c\u867d\u7136\u6a21\u578b\u91cc\u5b58\u7740\u591a\u4e2a\u4e13\u5bb6\u53c2\u6570\uff0c\u4f46\u6bcf\u6b21\u8ba1\u7b97\u53ea\u8d70\u5176\u4e2d\u4e00\u4e2a FFN\u3002<\/p>\n\n\n<p><!-- formulas-added:2209-moe-routing --><\/p>\n\n\n<p>top-1 MoE \u7684\u8def\u7531\u8fc7\u7a0b\u5982\u4e0b\u3002router \u5148\u4ea7\u751f\u5404 expert \u7684 gate \u5206\u6570\uff0c\u518d\u9009\u62e9\u6700\u5927\u5206\u6570\u5bf9\u5e94\u7684 expert\uff1a<\/p>\n\n\n\n\\(\n\\begin{aligned}\ng &amp;= [g_0,\\cdots,g_{E-1}]=\\mathrm{softmax}(\\mathrm{router}(z_t^{(3)})),\\\\\ni^* &amp;= \\arg\\max_{0\\le i\\le E-1} g_i,\\\\\n\\mathrm{FFN}^{(\\mathrm{MoE})}(z_t^{(3)}) &amp;= g_{i^*}\\mathrm{FFN}_{i^*}(z_t^{(3)}).\n\\end{aligned}\n\\)\n\n\n\n<p>\u8fd9\u4e2a\u8bbe\u8ba1\u628a\u201c\u5bb9\u91cf\u201d\u548c\u201c\u8ba1\u7b97\u201d\u90e8\u5206\u89e3\u8026\uff1a\u603b\u53c2\u6570\u66f4\u591a\uff0c\u6f5c\u5728\u8868\u8fbe\u7a7a\u95f4\u66f4\u5927\uff1b\u4f46\u6fc0\u6d3b\u53c2\u6570\u4e0d\u589e\u52a0\u592a\u591a\uff0c\u63a8\u7406\u8ba1\u7b97\u4ecd\u63a5\u8fd1\u666e\u901a FFN\u3002\u4e3a\u4e86\u907f\u514d\u6240\u6709\u6837\u672c\u90fd\u6324\u5411\u540c\u4e00\u4e2a\u4e13\u5bb6\uff0c\u4f5c\u8005\u52a0\u5165 load balancing loss\uff0c\u540c\u65f6\u5728\u8bad\u7ec3\u65f6\u7ed9 router \u52a0\u9ad8\u65af\u566a\u58f0\uff0c\u8ba9\u4e13\u5bb6\u9009\u62e9\u66f4\u5206\u6563\u3002<\/p>\n\n\n<p><!-- formulas-added:2209-balance-loss --><\/p>\n\n\n<p>\u8d1f\u8f7d\u5747\u8861\u635f\u5931\u7528\u4e8e\u9f13\u52b1 expert \u88ab\u66f4\u5747\u5300\u5730\u4f7f\u7528\uff1a<\/p>\n\n\n\n\\(\nL_{\\mathrm{balance}}=E\\sum_{i=0}^{E-1}f_i\\bar{g}_i\n\\)\n\n\n\n<h4>\u72ec\u7acb\u8def\u7531\u5668\u4e0e\u5f52\u4e00\u5316<\/h4>\n\n\n\n<p>\u8bba\u6587\u6ca1\u6709\u628a\u6240\u6709 MoE router \u90fd\u4e00\u8d77\u5171\u4eab\uff0c\u800c\u662f\u8ba9\u6bcf\u4e2a MoE \u6a21\u5757\u62e5\u6709\u81ea\u5df1\u7684 router\u3002\u76f4\u89c9\u4e0a\uff0c\u540c\u4e00\u4e2a\u5171\u4eab\u5757\u5728\u7b2c 1 \u6b21\u3001\u7b2c 6 \u6b21\u3001\u7b2c 12 \u6b21\u9012\u5f52\u8c03\u7528\u65f6\uff0c\u8f93\u5165\u8868\u793a\u5df2\u7ecf\u5904\u5728\u4e0d\u540c\u5c42\u7ea7\uff1b\u5982\u679c\u8def\u7531\u8def\u5f84\u5b8c\u5168\u4e00\u81f4\uff0c\u5c31\u4f1a\u9650\u5236\u4e13\u5bb6\u9009\u62e9\u7684\u7075\u6d3b\u6027\u3002<\/p>\n\n\n\n<p>\u5f52\u4e00\u5316\u5c42\u4e5f\u7c7b\u4f3c\u3002LayerNorm\u3001BatchNorm \u7684\u7edf\u8ba1\u548c\u7f29\u653e\u504f\u79fb\u53c2\u6570\u5bf9\u8868\u793a\u5206\u5e03\u5f88\u654f\u611f\u3002\u4f5c\u8005\u8ba9\u5f52\u4e00\u5316\u6a21\u5757\u4fdd\u6301\u72ec\u7acb\uff0c\u4f7f\u4e0d\u540c\u5c42\u7ea7\u7684\u8868\u793a\u80fd\u591f\u7ef4\u6301\u5404\u81ea\u5408\u9002\u7684\u7edf\u8ba1\u72b6\u6001\u3002\u8bba\u6587\u8fd8\u628a\u5f52\u4e00\u5316\u4e2d\u7684 scale \u548c offset \u770b\u4f5c\u4e00\u79cd\u8f7b\u91cf adapter\uff0c\u7528\u5f88\u5c11\u53c2\u6570\u589e\u5f3a\u5171\u4eab\u5757\u7684\u9002\u914d\u80fd\u529b\u3002<\/p>\n\n\n\n<h4>\u9690\u85cf\u5c42\u77e5\u8bc6\u84b8\u998f<\/h4>\n\n\n\n<p>\u5171\u4eab\u6a21\u578b\u518d\u806a\u660e\uff0c\u6bd5\u7adf\u53c2\u6570\u5c11\u3002\u4f5c\u8005\u7528\u5168\u53c2\u6570 Conformer \u7f16\u7801\u5668\u4f5c\u4e3a teacher\uff0c\u8ba9\u5171\u4eab\u6a21\u578b\u7684\u7f16\u7801\u5668\u8f93\u51fa\u5c3d\u91cf\u63a5\u8fd1 teacher \u7684\u9690\u85cf\u8868\u793a\u3002\u8fd9\u91cc\u4e0d\u662f\u53ea\u84b8\u998f\u6700\u7ec8\u9884\u6d4b\u5206\u5e03\uff0c\u800c\u662f\u76f4\u63a5\u7ea6\u675f\u9690\u85cf embedding \u7684 L2 \u8ddd\u79bb\u3002\u8fd9\u6837\u505a\u7684\u76ee\u7684\uff0c\u662f\u8ba9\u5c0f\u6a21\u578b\u5b66\u4e60 full model \u7684\u4e2d\u95f4\u8868\u5f81\u8f68\u8ff9\u3002<\/p>\n\n\n<p><!-- formulas-added:2209-kd-loss --><\/p>\n\n\n<p>hidden embedding \u84b8\u998f\u635f\u5931\u76f4\u63a5\u7ea6\u675f student encoder \u8f93\u51fa <code>h_t<\/code> \u4e0e teacher encoder \u8f93\u51fa <code>h_t'<\/code> \u7684\u8ddd\u79bb\uff1a<\/p>\n\n\n\n\\(\nL_{\\mathrm{kd}}=\\frac{1}{T}\\sum_{t=0}^{T-1}\\lVert h_t-h_t&#8217;\\rVert_2\n\\)\n\n\n\n<h4>\u8bad\u7ec3\u76ee\u6807<\/h4>\n\n\n\n<p>\u6700\u7ec8\u635f\u5931\u7531\u4e09\u90e8\u5206\u7ec4\u6210\uff1a\u4e3b\u4efb\u52a1\u7684\u8d1f\u5bf9\u6570\u4f3c\u7136\u3001MoE \u7684\u8d1f\u8f7d\u5747\u8861\u635f\u5931\u3001\u9690\u85cf\u5c42\u77e5\u8bc6\u84b8\u998f\u635f\u5931\u3002\u8d1f\u8f7d\u5747\u8861\u9879\u8d1f\u8d23\u8ba9\u4e13\u5bb6\u4e0d\u584c\u7f29\uff0c\u84b8\u998f\u9879\u8d1f\u8d23\u8ba9\u5171\u4eab\u6a21\u578b\u8d34\u8fd1\u5168\u53c2\u6570 teacher\u3002\u8bba\u6587\u8fd8\u5728\u5b9e\u9a8c\u4e2d\u52a0\u5165 CTC loss \u6765\u8f85\u52a9\u5bf9\u9f50\u3002<\/p>\n\n\n<p><!-- formulas-added:2209-total-loss --><\/p>\n\n\n\\(\nL=L_{\\mathrm{nll}}+\\frac{\\alpha}{C}\\sum L_{\\mathrm{balance}}+\\beta L_{\\mathrm{kd}}\n\\)\n\n\n\n<p>\u8fd9\u91cc <code>C<\/code> \u662f MoE module \u7684\u6570\u91cf\uff0c<code>\u03b1<\/code> \u548c <code>\u03b2<\/code> \u5206\u522b\u63a7\u5236\u8d1f\u8f7d\u5747\u8861\u635f\u5931\u4e0e\u84b8\u998f\u635f\u5931\u7684\u6743\u91cd\u3002<\/p>\n\n\n\n<h3>\u4e0e\u5df2\u6709\u5de5\u4f5c\u7684\u5173\u7cfb<\/h3>\n\n\n\n<p>MoE \u5e38\u88ab\u7528\u6765\u6269\u5927\u6a21\u578b\u5bb9\u91cf\uff0c\u5c24\u5176\u662f\u5728 NLP \u5927\u6a21\u578b\u91cc\uff0c\u901a\u8fc7\u6761\u4ef6\u8ba1\u7b97\u6269\u5c55\u5230\u5f88\u5927\u7684\u53c2\u6570\u89c4\u6a21\u3002\u4f46\u8fd9\u7bc7\u8bba\u6587\u4e0d\u662f\u8ffd\u6c42\u8d85\u5927\u89c4\u6a21\uff0c\u800c\u662f\u628a MoE \u5f53\u4f5c\u53c2\u6570\u9ad8\u6548\u5de5\u5177\uff1a\u5171\u4eab\u4e13\u5bb6\u3001\u91cd\u590d\u4f7f\u7528\u4e13\u5bb6\uff0c\u8ba9\u5c11\u91cf\u6a21\u5757\u53d1\u6325\u66f4\u5927\u4f5c\u7528\u3002<\/p>\n\n\n\n<p>\u8de8\u5c42\u6743\u91cd\u5171\u4eab\u4e5f\u4e0d\u662f\u65b0\u60f3\u6cd5\uff0cALBERT\u3001Universal Transformer \u4ee5\u53ca\u82e5\u5e72 ASR \u5de5\u4f5c\u90fd\u7528\u8fc7\u7c7b\u4f3c\u673a\u5236\u3002\u672c\u6587\u7684\u4e0d\u540c\u70b9\u5728\u4e8e\uff0c\u5b83\u6ca1\u6709\u53ea\u505a\u6734\u7d20\u5171\u4eab\uff0c\u800c\u662f\u5728\u5171\u4eab\u7ed3\u6784\u91cc\u52a0\u5165\u7a00\u758f\u4e13\u5bb6\uff0c\u540c\u65f6\u8ba9 router \u548c normalization \u72ec\u7acb\uff0c\u4ece\u800c\u51cf\u5c11\u5171\u4eab\u5e26\u6765\u7684\u5bb9\u91cf\u548c\u5206\u5e03\u9002\u914d\u95ee\u9898\u3002<\/p>\n\n\n\n<h3>\u5b9e\u9a8c\uff1a<\/h3>\n\n\n\n<h4>\u5b9e\u9a8c\u8bbe\u7f6e<\/h4>\n\n\n\n<p>\u5b9e\u9a8c\u4f7f\u7528 AISHELL-1 \u666e\u901a\u8bdd\u8bed\u97f3\u8bc6\u522b\u6570\u636e\u96c6\uff1a\u7ea6 150 \u5c0f\u65f6\u8bad\u7ec3\u8bed\u97f3\u300118 \u5c0f\u65f6\u5f00\u53d1\u96c6\u300110 \u5c0f\u65f6\u6d4b\u8bd5\u96c6\u3002\u8f93\u5165\u7279\u5f81\u4e3a 80 \u7ef4 FBANK\uff0c\u7a97\u53e3 25ms\u3001\u6b65\u957f 10ms\uff0c\u5e76\u4f7f\u7528\u5168\u5c40 CMVN\u3001\u901f\u5ea6\u6270\u52a8\u3001SpecAugmentation \u548c time stretch \u7b49\u589e\u5f3a\u624b\u6bb5\u3002\u8bcd\u8868\u5305\u542b 4235 \u4e2a\u4e2d\u6587\u5b57\u7b26\u4ee5\u53ca\u8d77\u6b62\u7b26\u53f7\u3002<\/p>\n\n\n\n<p>\u6a21\u578b\u524d\u7aef\u662f\u4e24\u5c42 CNN subsampling\uff0c\u628a\u5e27\u7387\u964d\u5230 25Hz\u3002\u7f16\u7801\u5668\u7ef4\u5ea6\u4e3a 256\uff0cMHSA \u4f7f\u7528 4 \u4e2a\u5934\uff0c\u5377\u79ef\u6838\u5927\u5c0f 15\uff0cFFN \u4e2d\u95f4\u7ef4\u5ea6 1024\u3002MoE-Conformer \u7684\u7b2c\u4e8c\u4e2a FFN \u4f7f\u7528 4 \u4e2a\u4e13\u5bb6\uff0c\u89e3\u7801\u5668\u662f 4 \u5c42 Transformer\u3002\u8bad\u7ec3 80 \u4e2a epoch\uff0c\u4f7f\u7528 PyTorch \u548c FastMoE \u5b9e\u73b0\u3002<\/p>\n\n\n\n<h4>\u7ed3\u679c\u4e0e\u5206\u6790<\/h4>\n\n\n\n<p>\u4e3b\u8868\u91cc\uff0c\u5168\u53c2\u6570 C12 \u7f16\u7801\u5668\u53c2\u6570\u91cf\u4e3a 21.58M\uff0c\u6d4b\u8bd5\u96c6 CER \u4e3a 4.93\u3002\u6700\u7ec8\u7684 C2-MoE4-G6-KD \u53ea\u6709 6.95M \u7f16\u7801\u5668\u53c2\u6570\uff0c\u6d4b\u8bd5\u96c6 CER \u4e3a 5.03\u3002\u6362\u53e5\u8bdd\u8bf4\uff0c\u5b83\u7528\u5927\u7ea6\u4e09\u5206\u4e4b\u4e00\u7684\u7f16\u7801\u5668\u53c2\u6570\uff0c\u505a\u5230\u4e86\u975e\u5e38\u63a5\u8fd1 full-parameter \u6a21\u578b\u7684\u7ed3\u679c\u3002<\/p>\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\/2026\/07\/image-6.png\" alt=\"\" class=\"wp-image-31332\" width=\"407\" height=\"237\" srcset=\"http:\/\/139.9.1.231\/wp-content\/uploads\/2026\/07\/image-6.png 709w, http:\/\/139.9.1.231\/wp-content\/uploads\/2026\/07\/image-6-300x175.png 300w\" sizes=\"(max-width: 407px) 100vw, 407px\" \/><\/figure><\/div>\n\n\n\n<p>\u6d88\u878d\u5b9e\u9a8c\u663e\u793a\uff0c\u5355\u72ec\u51cf\u5c11\u5757\u6570\u4f1a\u660e\u663e\u635f\u4f24\u6548\u679c\uff0c\u4f8b\u5982 C2 \u7684\u6d4b\u8bd5 CER \u4e3a 6.50\uff1b\u52a0\u5165 MoE \u540e\uff0cC2-MoE4 \u964d\u5230 6.22\uff0c\u8bf4\u660e\u4e13\u5bb6\u673a\u5236\u786e\u5b9e\u8865\u4e86\u4e00\u90e8\u5206\u5bb9\u91cf\u3002\u518d\u52a0\u5165\u8de8\u5c42\u5171\u4eab\u9012\u5f52\u8ba1\u7b97\u540e\uff0cC2-G6 \u4e3a 5.62\uff0c\u800c C2-MoE4-G6 \u8fbe\u5230 5.08\uff0c\u8bf4\u660e\u201c\u5171\u4eab + MoE\u201d\u7684\u7ec4\u5408\u6bd4\u4efb\u4e00\u5355\u72ec\u673a\u5236\u66f4\u6709\u4ef7\u503c\u3002<\/p>\n\n\n\n<figure class=\"wp-block-image size-large\"><img loading=\"lazy\" width=\"1024\" height=\"433\" src=\"http:\/\/139.9.1.231\/wp-content\/uploads\/2026\/07\/image-7-1024x433.png\" alt=\"\" class=\"wp-image-31334\" srcset=\"http:\/\/139.9.1.231\/wp-content\/uploads\/2026\/07\/image-7-1024x433.png 1024w, http:\/\/139.9.1.231\/wp-content\/uploads\/2026\/07\/image-7-300x127.png 300w, http:\/\/139.9.1.231\/wp-content\/uploads\/2026\/07\/image-7-768x325.png 768w, http:\/\/139.9.1.231\/wp-content\/uploads\/2026\/07\/image-7.png 1258w\" sizes=\"(max-width: 1024px) 100vw, 1024px\" \/><\/figure>\n\n\n\n<p>\u72ec\u7acb\u8def\u7531\u5668\u548c\u5f52\u4e00\u5316\u7684\u4f5c\u7528\u4e5f\u5f88\u660e\u663e\u3002C2-MoE4-G6 \u5982\u679c\u5168\u90e8\u5171\u4eab\uff0c\u6d4b\u8bd5 CER \u4e3a 6.00\uff1b\u53ea\u8ba9\u5f52\u4e00\u5316\u72ec\u7acb\uff0c\u964d\u5230 5.21\uff1b\u5f52\u4e00\u5316\u548c router \u90fd\u72ec\u7acb\u540e\uff0c\u8fdb\u4e00\u6b65\u5230 5.08\u3002\u8fd9\u8bf4\u660e\u5171\u4eab\u6a21\u578b\u6700\u6015\u7684\u4e0d\u662f\u53c2\u6570\u5c11\u672c\u8eab\uff0c\u800c\u662f\u4e0d\u540c\u6df1\u5ea6\u8868\u793a\u88ab\u8feb\u4f7f\u7528\u5b8c\u5168\u76f8\u540c\u7684\u9002\u914d\u8def\u5f84\u3002<\/p>\n\n\n\n<p>\u77e5\u8bc6\u84b8\u998f\u5e26\u6765\u7684\u63d0\u5347\u76f8\u5bf9\u6e29\u548c\uff0c\u4f46\u5728 C2-MoE4-G6 \u4e0a\u4ecd\u628a\u6d4b\u8bd5 CER \u4ece 5.08 \u63a8\u5230 5.03\u3002\u4f5c\u8005\u8fd8\u901a\u8fc7\u8f93\u5165\u8f93\u51fa L2 \u8ddd\u79bb\u89c2\u5bdf\u6a21\u578b\u5185\u90e8\u53d8\u5316\uff1a\u5e26\u72ec\u7acb router \u548c normalization \u7684\u5171\u4eab\u6a21\u578b\u66f4\u63a5\u8fd1\u5168\u53c2\u6570 C12 \u7684\u53d8\u5316\u66f2\u7ebf\uff0c\u800c\u5168\u5171\u4eab\u6a21\u578b\u66f2\u7ebf\u66f4\u4e0d\u7a33\u5b9a\u3002<\/p>\n\n\n\n<figure class=\"wp-block-image size-full\"><img loading=\"lazy\" width=\"601\" height=\"379\" src=\"http:\/\/139.9.1.231\/wp-content\/uploads\/2026\/07\/image-8.png\" alt=\"\" class=\"wp-image-31336\" srcset=\"http:\/\/139.9.1.231\/wp-content\/uploads\/2026\/07\/image-8.png 601w, http:\/\/139.9.1.231\/wp-content\/uploads\/2026\/07\/image-8-300x189.png 300w\" sizes=\"(max-width: 601px) 100vw, 601px\" \/><\/figure>\n\n\n\n<h3>\u7ed3\u8bba\u4e0e\u672a\u6765\u65b9\u5411<\/h3>\n\n\n\n<p>\u8fd9\u7bc7\u8bba\u6587\u7684\u4ef7\u503c\u5728\u4e8e\uff0c\u5b83\u7ed9\u51fa\u4e86\u4e00\u4e2a\u8f83\u5b8c\u6574\u7684\u53c2\u6570\u9ad8\u6548 Conformer \u65b9\u6848\uff1a\u7528\u8de8\u5c42\u5171\u4eab\u538b\u7f29\u53c2\u6570\uff0c\u7528\u7a00\u758f MoE \u6062\u590d\u5bb9\u91cf\uff0c\u7528\u72ec\u7acb\u8def\u7531\u5668\u548c\u5f52\u4e00\u5316\u9002\u914d\u4e0d\u540c\u6df1\u5ea6\u8868\u793a\uff0c\u518d\u7528\u9690\u85cf\u5c42\u84b8\u998f\u8865\u9f50\u5c0f\u6a21\u578b\u8868\u73b0\u3002\u6700\u7ec8\u6a21\u578b\u5728 AISHELL-1 \u4e0a\u4ee5\u7ea6\u4e09\u5206\u4e4b\u4e00\u7f16\u7801\u5668\u53c2\u6570\u63a5\u8fd1\u5168\u53c2\u6570\u6a21\u578b\u3002<\/p>\n\n\n\n<p>\u5b83\u4e5f\u7559\u4e0b\u4e86\u81ea\u7136\u7684\u540e\u7eed\u95ee\u9898\uff1a\u65b9\u6cd5\u662f\u5426\u80fd\u5728\u66f4\u5927\u89c4\u6a21\u3001\u591a\u8bed\u79cd\u6216\u66f4\u590d\u6742\u7684 ASR \u6570\u636e\u96c6\u4e0a\u4fdd\u6301\u4f18\u52bf\uff1f\u80fd\u5426\u8fc1\u79fb\u5230 RNN-T\u3001CTC \u6216\u5176\u4ed6\u7aef\u5230\u7aef ASR \u67b6\u6784\uff1f\u4ece\u5de5\u7a0b\u89d2\u5ea6\u770b\uff0c\u8fd9\u7c7b\u65b9\u6848\u7684\u5438\u5f15\u529b\u5f88\u5f3a\uff0c\u56e0\u4e3a\u5b83\u4e0d\u662f\u5355\u7eaf\u8ffd\u6c42\u5c0f\u6a21\u578b\uff0c\u800c\u662f\u5728\u201c\u53c2\u6570\u3001\u8ba1\u7b97\u3001\u8868\u8fbe\u5bb9\u91cf\u201d\u4e4b\u95f4\u505a\u66f4\u7ec6\u7684\u62c6\u5206\u3002<\/p>\n\n\n\n<h2><strong><a href=\"https:\/\/arxiv.org\/pdf\/2601.02967\" target=\"_blank\" rel=\"noreferrer noopener\">MoEAdapter for Large Audio Language Models: Sparsity, Disentanglement, and Gradient-Conflict-Free<\/a><\/strong><\/h2>\n\n\n\n<figure class=\"wp-block-image size-full\"><img loading=\"lazy\" width=\"1006\" height=\"751\" src=\"http:\/\/139.9.1.231\/wp-content\/uploads\/2026\/05\/image-13.png\" alt=\"\" class=\"wp-image-31173\" srcset=\"http:\/\/139.9.1.231\/wp-content\/uploads\/2026\/05\/image-13.png 1006w, http:\/\/139.9.1.231\/wp-content\/uploads\/2026\/05\/image-13-300x224.png 300w, http:\/\/139.9.1.231\/wp-content\/uploads\/2026\/05\/image-13-768x573.png 768w\" sizes=\"(max-width: 1006px) 100vw, 1006px\" \/><\/figure>\n\n\n\n<p>\u8fd9\u7bc7\u8bba\u6587 <em>MoE Adapter for Large Audio Language Models: Sparsity, Disentanglement, and Gradient-Conflict-Free<\/em> \u7684\u95ee\u9898\u610f\u8bc6\u5f88\u660e\u786e\uff1a\u5927\u8bed\u8a00\u6a21\u578b\u8981\u7406\u89e3\u771f\u5b9e\u4e16\u754c\uff0c\u4e0d\u80fd\u53ea\u770b\u6587\u672c\uff0c\u97f3\u9891\u662f\u5f88\u91cd\u8981\u7684\u8f93\u5165\u6a21\u6001\u3002\u4f46\u97f3\u9891\u5e76\u4e0d\u662f\u4e00\u79cd\u5747\u5300\u4fe1\u53f7\u3002\u8bed\u97f3\u3001\u97f3\u4e50\u3001\u73af\u5883\u58f0\u627f\u8f7d\u7684\u4fe1\u606f\u7ed3\u6784\u4e0d\u540c\uff0c\u5982\u679c\u7528\u4e00\u4e2a dense adapter \u628a\u6240\u6709\u97f3\u9891\u90fd\u538b\u8fdb\u540c\u4e00\u4e2a\u6587\u672c embedding \u7a7a\u95f4\uff0c\u5f88\u5bb9\u6613\u51fa\u73b0\u53c2\u6570\u66f4\u65b0\u65b9\u5411\u4e92\u76f8\u51b2\u7a81\u3002<\/p>\n\n\n\n<figure class=\"wp-block-image size-full\"><img loading=\"lazy\" width=\"1024\" height=\"506\" src=\"http:\/\/139.9.1.231\/wp-content\/uploads\/2026\/07\/image-9.png\" alt=\"\" class=\"wp-image-31339\" srcset=\"http:\/\/139.9.1.231\/wp-content\/uploads\/2026\/07\/image-9.png 1024w, http:\/\/139.9.1.231\/wp-content\/uploads\/2026\/07\/image-9-300x148.png 300w, http:\/\/139.9.1.231\/wp-content\/uploads\/2026\/07\/image-9-768x380.png 768w\" sizes=\"(max-width: 1024px) 100vw, 1024px\" \/><\/figure>\n\n\n\n<h3>\u6458\u8981\uff1a\u7528\u4e13\u5bb6\u5206\u5de5\u5904\u7406\u5f02\u8d28\u97f3\u9891<\/h3>\n\n\n\n<p>\u8bba\u6587\u63d0\u51fa MoE-Adapter\uff0c\u7528\u7a00\u758f Mixture-of-Experts \u66ff\u4ee3\u4f20\u7edf\u7684 dense audio adapter\u3002\u5b83\u4e0d\u662f\u8ba9\u6240\u6709\u97f3\u9891 token \u90fd\u901a\u8fc7\u540c\u4e00\u5957 FFN\uff0c\u800c\u662f\u7528\u52a8\u6001\u95e8\u63a7\u628a token \u8def\u7531\u5230\u82e5\u5e72\u4e13\u95e8\u4e13\u5bb6\uff0c\u540c\u65f6\u4fdd\u7559\u4e00\u5b9a\u5171\u4eab\u80fd\u529b\u6765\u6355\u6349\u5168\u5c40\u4e0a\u4e0b\u6587\u3002\u8fd9\u6837\uff0c\u8bed\u97f3\u3001\u97f3\u4e50\u3001\u73af\u5883\u58f0\u7b49\u4e0d\u540c\u5c5e\u6027\u53ef\u4ee5\u5728\u4e0d\u540c\u4e13\u5bb6\u5b50\u7a7a\u95f4\u4e2d\u88ab\u5efa\u6a21\uff0c\u4ece\u800c\u51cf\u8f7b\u68af\u5ea6\u51b2\u7a81\u3002<\/p>\n\n\n\n<p>\u5b9e\u9a8c\u57fa\u4e8e Qwen3-1.7B \u9aa8\u5e72\uff0c\u97f3\u9891\u524d\u7aef\u4f7f\u7528 Whisper-VQ tokenizer \u548c Whisper Encoder\u3002\u4f5c\u8005\u5728\u76f8\u540c\u53c2\u6570\u9884\u7b97\u4e0b\u6bd4\u8f83 dense adapter \u548c MoE-Adapter\uff1a\u4e24\u8005\u603b\u53c2\u6570\u7ea6 94.4M\uff0c\u4f46 MoE \u56e0\u7a00\u758f\u6fc0\u6d3b\uff0c\u63a8\u7406\u65f6\u53ea\u6fc0\u6d3b\u7ea6 70.8M \u53c2\u6570\u3002\u7ed3\u679c\u663e\u793a\uff0cMoE-Adapter \u5728 MMSU\u3001OBQA\u3001MMAU \u7b49\u97f3\u9891\u7406\u89e3\u548c\u63a8\u7406\u4efb\u52a1\u4e0a\u5747\u4f18\u4e8e dense baseline\uff0c\u5e76\u51cf\u5c11\u97f3\u9891\u8f93\u5165\u4e0e\u6587\u672c\u8f93\u5165\u4e4b\u95f4\u7684 modality gap\u3002<\/p>\n\n\n\n<h3>\u5f15\u8a00\uff1a\u97f3\u9891\u4e0d\u662f\u4e00\u79cd\u5355\u4e00\u5206\u5e03<\/h3>\n\n\n\n<p>\u5927\u8bed\u8a00\u6a21\u578b\u5728\u6587\u672c\u63a8\u7406\u4e0a\u5df2\u7ecf\u975e\u5e38\u5f3a\uff0c\u4f46\u53ea\u5904\u7406\u6587\u672c\u4f1a\u9650\u5236\u5b83\u4eec\u611f\u77e5\u73b0\u5b9e\u4e16\u754c\u7684\u80fd\u529b\u3002\u97f3\u9891\u5305\u542b\u4eba\u7c7b\u8bf4\u8bdd\u3001\u73af\u5883\u58f0\u97f3\u3001\u97f3\u4e50\u548c\u60c5\u7eea\u97f5\u5f8b\u7b49\u4fe1\u606f\uff0c\u662f\u591a\u6a21\u6001\u667a\u80fd\u7ed5\u4e0d\u5f00\u7684\u4e00\u73af\u3002\u5f53\u524d\u8bb8\u591a\u5927\u97f3\u9891\u8bed\u8a00\u6a21\u578b\u7684\u4e3b\u6d41\u505a\u6cd5\uff0c\u662f\u52a0\u4e00\u4e2a adapter\uff0c\u628a\u58f0\u5b66\u7279\u5f81\u6295\u5f71\u5230 LLM \u7684\u6587\u672c\u8bed\u4e49\u7a7a\u95f4\u91cc\u3002<\/p>\n\n\n\n<p>\u95ee\u9898\u5728\u4e8e\uff0c\u5f88\u591a adapter \u662f dense\u3001\u53c2\u6570\u5171\u4eab\u7684\uff1a\u6240\u6709\u97f3\u9891\u90fd\u7ecf\u8fc7\u540c\u4e00\u5957\u6295\u5f71\u5c42\u3002\u8fd9\u9690\u542b\u4e00\u4e2a\u5047\u8bbe\uff0c\u5373\u4e0d\u540c\u97f3\u9891\u7c7b\u578b\u53ef\u4ee5\u88ab\u540c\u4e00\u79cd\u6620\u5c04\u5747\u5300\u5904\u7406\u3002\u4f5c\u8005\u8ba4\u4e3a\u8fd9\u4e2a\u5047\u8bbe\u8fc7\u5f3a\u3002\u8bed\u97f3\u4e3b\u8981\u627f\u8f7d\u8bed\u4e49\u548c\u8bed\u8a00\u7ed3\u6784\uff0c\u97f3\u4e50\u66f4\u5173\u6ce8\u8282\u594f\u3001\u65cb\u5f8b\u548c\u60c5\u611f\uff0c\u73af\u5883\u58f0\u53c8\u6709\u81ea\u5df1\u7684\u58f0\u5b66\u6a21\u5f0f\u3002\u5b83\u4eec\u5728\u8868\u793a\u7a7a\u95f4\u4e2d\u53ef\u80fd\u4f4d\u4e8e\u4e0d\u540c\u6d41\u5f62\u3002<\/p>\n\n\n\n<p>\u5982\u679c\u4e00\u4e2a dense adapter \u540c\u65f6\u5b66\u4e60\u8fd9\u4e9b\u76f8\u4e92\u5dee\u5f02\u5f88\u5927\u7684\u76ee\u6807\uff0c\u4e0d\u540c\u6570\u636e\u7c7b\u578b\u7684\u68af\u5ea6\u53ef\u80fd\u671d\u76f8\u53cd\u65b9\u5411\u66f4\u65b0\u540c\u4e00\u7ec4\u53c2\u6570\u3002\u8fd9\u5c31\u662f\u8bba\u6587\u5f3a\u8c03\u7684 gradient conflict\u3002MoE-Adapter \u7684\u8d21\u732e\uff0c\u5c31\u662f\u7528\u52a8\u6001\u4e13\u5bb6\u8def\u7531\u628a\u8fd9\u4e9b\u51b2\u7a81\u62c6\u5f00\uff1a\u76f8\u4f3c\u5c5e\u6027\u5171\u4eab\u4e13\u5bb6\uff0c\u51b2\u7a81\u5c5e\u6027\u8fdb\u5165\u4e0d\u540c\u4e13\u5bb6\u3002<\/p>\n\n\n\n<h3>\u76f8\u5173\u5de5\u4f5c<\/h3>\n\n\n\n<h4>\u5927\u97f3\u9891\u8bed\u8a00\u6a21\u578b<\/h4>\n\n\n\n<p>\u65e9\u671f\u97f3\u9891\u95ee\u7b54\u6216\u8bed\u97f3\u4ea4\u4e92\u7cfb\u7edf\u5e38\u91c7\u7528\u7ea7\u8054\u7ba1\u7ebf\uff1a\u5148 ASR \u8f6c\u6587\u5b57\uff0c\u518d\u4ea4\u7ed9 LLM\u3002\u8fd9\u6837\u7684\u7cfb\u7edf\u5bb9\u6613\u53d7\u5230\u8bc6\u522b\u9519\u8bef\u4f20\u64ad\u5f71\u54cd\uff0c\u4e5f\u4f1a\u4e22\u5931\u8bed\u8c03\u3001\u60c5\u7eea\u3001\u97f3\u4e50\u548c\u73af\u5883\u58f0\u7b49\u975e\u6587\u5b57\u4fe1\u606f\u3002\u540e\u6765\u7684\u7aef\u5230\u7aef LALM \u901a\u8fc7\u53ef\u5b66\u4e60 adapter\uff0c\u628a\u58f0\u5b66\u7279\u5f81\u6620\u5c04\u5230\u6587\u672c\u7a7a\u95f4\uff0c\u8ba9 LLM \u76f4\u63a5\u6761\u4ef6\u5316\u5728\u97f3\u9891\u8868\u793a\u4e0a\u3002<\/p>\n\n\n\n<p>\u73b0\u6709 adapter \u5927\u81f4\u5206\u4e3a Q-Former \u7c7b\u548c linear projector \/ MLP projector \u7c7b\u3002\u540e\u8005\u7ed3\u6784\u7b80\u5355\u3001\u6548\u7387\u9ad8\uff0c\u56e0\u6b64\u88ab\u8bb8\u591a\u6700\u65b0\u6a21\u578b\u91c7\u7528\u3002\u4f46\u8fd9\u79cd\u5168\u5c40\u5171\u4eab\u6295\u5f71\u5c42\u96be\u4ee5\u9762\u5bf9\u97f3\u9891\u5185\u90e8\u7684\u5206\u5e03\u5dee\u5f02\u3002\u672c\u6587\u6b63\u662f\u9488\u5bf9\u8fd9\u4e2a\u74f6\u9888\uff0c\u628a\u7a00\u758f MoE \u5f15\u5165 audio-text alignment \u9636\u6bb5\u3002<\/p>\n\n\n\n<h4>MoE \u67b6\u6784<\/h4>\n\n\n\n<p>MoE \u7684\u57fa\u672c\u601d\u60f3\u662f\u8ba9\u4e0d\u540c\u4e13\u5bb6\u5904\u7406\u4e0d\u540c\u6837\u672c\u6216\u4e0d\u540c token\uff0c\u901a\u8fc7\u7a00\u758f\u95e8\u63a7\u5b9e\u73b0\u6761\u4ef6\u8ba1\u7b97\u3002\u5b83\u5df2\u7ecf\u5728\u8bed\u8a00\u6a21\u578b\u3001\u591a\u6a21\u6001\u6a21\u578b\u3001\u89c6\u89c9\u8bed\u8a00\u6a21\u578b\u7b49\u65b9\u5411\u8bc1\u660e\u4e86\u5bf9\u5f02\u8d28\u6570\u636e\u548c\u4efb\u52a1\u51b2\u7a81\u7684\u7f13\u89e3\u80fd\u529b\u3002\u97f3\u9891\u9886\u57df\u4e5f\u5f00\u59cb\u51fa\u73b0 MoE \u76f8\u5173\u5de5\u4f5c\uff0c\u4f8b\u5982\u751f\u6210\u3001\u533b\u7597\u97f3\u9891\u7279\u5f81\u9009\u62e9\u7b49\u3002<\/p>\n\n\n\n<p>\u4e0d\u8fc7\uff0c\u5728\u901a\u7528 audio-text alignment \u8fd9\u4e2a\u73af\u8282\uff0c\u4e3b\u6d41 LALM \u4ecd\u5927\u91cf\u4f9d\u8d56\u9759\u6001\u3001\u5171\u4eab\u53c2\u6570 adapter\u3002\u672c\u6587\u7684 MoE-Adapter \u4e0d\u53ea\u662f\u501f\u7528 MoE \u6269\u5bb9\u91cf\uff0c\u800c\u662f\u628a MoE \u4f5c\u4e3a\u4e00\u79cd\u201c\u89e3\u8026\u5de5\u5177\u201d\uff0c\u4e13\u95e8\u5904\u7406\u97f3\u9891\u5c5e\u6027\u4e4b\u95f4\u7684\u51b2\u7a81\u3002<\/p>\n\n\n\n<h3>\u65b9\u6cd5<\/h3>\n\n\n\n<h4>\u6574\u4f53\u6846\u67b6<\/h4>\n\n\n\n<p>\u6a21\u578b\u91c7\u7528\u7c7b\u4f3c Kimi-Audio \u7684 dual-stream \u97f3\u9891\u524d\u7aef\uff1a\u4e00\u6761\u8def\u5f84\u7528\u51bb\u7ed3 tokenizer \u63d0\u53d6\u79bb\u6563\u8bed\u4e49 token\uff0c\u53e6\u4e00\u6761\u8def\u5f84\u7528 speech encoder \u63d0\u53d6\u8fde\u7eed\u58f0\u5b66\u7279\u5f81\u3002\u4e24\u7c7b\u8868\u793a\u7ecf\u8fc7\u6295\u5f71\u548c\u878d\u5408\u540e\uff0c\u8fdb\u5165 adapter\u3002<\/p>\n\n\n\n<p>\u4f20\u7edf\u65b9\u6848\u4f1a\u7528 dense adapter \u628a\u878d\u5408\u97f3\u9891\u7279\u5f81\u6620\u5c04\u5230 LLM embedding \u7a7a\u95f4\u3002\u672c\u6587\u5219\u7528 MoE-Adapter \u5b8c\u6210\u8fd9\u4e00\u6b65\u3002\u6700\u7ec8\uff0cadapted audio embeddings \u4e0e\u6587\u672c token embeddings \u62fc\u63a5\uff0c\u4f5c\u4e3a LLM \u7684\u8f93\u5165\uff0c\u5e76\u7528\u6807\u51c6\u81ea\u56de\u5f52 next-token prediction \u8bad\u7ec3\u3002<\/p>\n\n\n\n<h4>\u7a00\u758f MoE Adapter<\/h4>\n\n\n\n<p>Dense adapter \u53ef\u4ee5\u770b\u4f5c\u4e00\u4e2a\u5355\u4f53 FFN\uff1a\u6240\u6709\u97f3\u9891 token \u90fd\u901a\u8fc7\u540c\u4e00\u7ec4\u6743\u91cd\u3002\u4f5c\u8005\u6307\u51fa\uff0c\u8fd9\u79cd\u8bbe\u8ba1\u5f3a\u5236\u540c\u4e00\u7ec4\u53c2\u6570\u540c\u65f6\u5bb9\u7eb3\u5f02\u8d28\u97f3\u9891\uff0c\u4f1a\u5f62\u6210\u4e0d\u5fc5\u8981\u7684\u4f18\u5316\u5e72\u6270\u3002<\/p>\n\n\n<p><!-- formulas-added:2601-dense-adapter --><\/p>\n\n\n<p>\u8bba\u6587\u5148\u628a dense adapter \u5199\u6210\u5355\u4f53 FFN \u6295\u5f71\u3002\u7ed9\u5b9a\u97f3\u9891 token <code>x<\/code>\uff0c\u8f93\u51fa embedding \u4e3a\uff1a<\/p>\n\n\n\n\\(\ny=\\mathcal{N}\\left(W_{d2}\\cdot\\sigma\\left(W_{d1}\\cdot\\mathcal{N}(x)\\right)\\right)\n\\)\n\n\n\n<p>MoE-Adapter \u628a\u5355\u4f53 FFN \u66ff\u6362\u6210\u4e13\u5bb6\u96c6\u5408\u3002\u6bcf\u4e2a\u4e13\u5bb6\u90fd\u662f\u8f7b\u91cf FFN\uff0crouter \u6839\u636e\u8f93\u5165 token \u8ba1\u7b97\u5404\u4e13\u5bb6\u5f97\u5206\uff0c\u5e76\u901a\u8fc7 Top-k \u9009\u62e9\u4fdd\u7559\u82e5\u5e72\u6d3b\u8dc3\u4e13\u5bb6\u3002\u88ab\u9009\u4e2d\u7684\u4e13\u5bb6\u8f93\u51fa\u6309\u95e8\u63a7\u6743\u91cd\u805a\u5408\uff0c\u5f62\u6210\u4e2d\u95f4\u8868\u793a\u3002\u968f\u540e\u518d\u7ecf\u8fc7\u8f93\u51fa\u6295\u5f71\u548c LayerNorm\uff0c\u5bf9\u9f50\u5230 LLM embedding \u7ef4\u5ea6\uff0c\u7528\u6765\u66ff\u6362\u8f93\u5165\u5e8f\u5217\u4e2d\u7684\u97f3\u9891\u5360\u4f4d token\u3002<\/p>\n\n\n<p><!-- formulas-added:2601-moe-adapter --><\/p>\n\n\n<p>\u6bcf\u4e2a expert \u672c\u8eab\u4e5f\u662f\u4e00\u4e2a\u8f7b\u91cf FFN\uff1a<\/p>\n\n\n\n\\(\nE_i(x)=W_{e2}^{(i)}\\cdot\\phi\\left(W_{e1}^{(i)}\\cdot\\mathcal{N}(x)\\right)\n\\)\n\n\n\n<p>router \u6839\u636e logits <code>s=xW_g<\/code> \u505a Top-k \u7a00\u758f\u9009\u62e9\uff0c\u518d softmax \u5f97\u5230\u95e8\u63a7\u6982\u7387\uff1a<\/p>\n\n\n\n\\(\nG(x)=\\mathrm{softmax}\\left(T_k(s)\\right),\\quad s=xW_g\n\\)\n\n\n\n<p>\u88ab\u9009\u4e2d\u7684 expert \u8f93\u51fa\u6309 gate \u6743\u91cd\u805a\u5408\uff0c\u5e76\u7ecf\u8fc7\u6700\u7ec8\u6295\u5f71\u5bf9\u9f50\u5230 LLM embedding \u7a7a\u95f4\uff1a<\/p>\n\n\n\n\\(\nh_{\\mathrm{MoE}}=\\sum_{i\\in I}G(x)_i\\cdot E_i(x)\n\\)\n\n\n\n\\(\ny_{\\mathrm{MoE}}=\\mathcal{N}\\left(W_P\\cdot h_{\\mathrm{MoE}}\\right)\n\\)\n\n\n\n<p>\u8fd9\u5957\u673a\u5236\u6709\u4e24\u4e2a\u6548\u679c\uff1a\u4e00\u662f\u7a00\u758f\u6fc0\u6d3b\u964d\u4f4e\u63a8\u7406\u6210\u672c\uff0c\u4e8c\u662f\u4e13\u5bb6\u5206\u5de5\u8ba9\u4e0d\u540c\u97f3\u9891\u5c5e\u6027\u8fdb\u5165\u4e0d\u540c\u5b50\u7a7a\u95f4\u3002\u5bf9\u4e8e\u8bed\u97f3\u3001\u97f3\u4e50\u3001\u73af\u5883\u58f0\u8fd9\u79cd\u5929\u7136\u5f02\u8d28\u8f93\u5165\uff0c\u7b2c\u4e8c\u70b9\u5c24\u5176\u91cd\u8981\u3002<\/p>\n\n\n\n<h4>\u8bad\u7ec3\u76ee\u6807<\/h4>\n\n\n\n<p>\u8bad\u7ec3\u76ee\u6807\u7531 next-token prediction loss \u548c auxiliary load-balancing loss \u7ec4\u6210\u3002\u524d\u8005\u8ba9\u6a21\u578b\u57fa\u4e8e\u97f3\u9891\u4e0a\u4e0b\u6587\u9884\u6d4b\u540e\u7eed\u6587\u672c token\uff0c\u662f\u4e3b\u4efb\u52a1\uff1b\u540e\u8005\u7528\u4e8e\u907f\u514d expert collapse\uff0c\u5373\u6240\u6709 token \u90fd\u6d8c\u5411\u5c11\u6570\u4e13\u5bb6\u3002<\/p>\n\n\n<p><!-- formulas-added:2601-training-loss --><\/p>\n\n\n<p>\u603b\u8bad\u7ec3\u76ee\u6807\u4e3a next-token prediction \u4e0e\u8d1f\u8f7d\u5747\u8861\u9879\u7684\u52a0\u6743\u548c\uff1a<\/p>\n\n\n\n\\(\nL=L_{\\mathrm{NTP}}+\\lambda L_{\\mathrm{aux}}\n\\)\n\n\n\n<p>\u5176\u4e2d\u4e3b\u4efb\u52a1 NTP loss \u5199\u6210\uff1a<\/p>\n\n\n\n\\(\nL_{\\mathrm{NTP}}=-\\sum_{t=1}^{T}\\log P(y_t\\mid y_{&lt;t},X;\\theta)\n\\)\n\n\n\n<p>\u8d1f\u8f7d\u5747\u8861\u635f\u5931\u4f1a\u540c\u65f6\u8003\u8651\u4e13\u5bb6\u7684\u91cd\u8981\u6027\u548c\u5b9e\u9645\u8d1f\u8f7d\uff0c\u8ba9\u4e0d\u540c\u4e13\u5bb6\u90fd\u88ab\u5145\u5206\u8bad\u7ec3\u3002\u8fd9\u91cc\u6709\u4e00\u4e2a\u5fae\u5999\u7684\u53d6\u820d\uff1a\u8fc7\u5f3a\u7684\u5747\u8861\u53ef\u80fd\u538b\u5236\u67d0\u4e9b\u81ea\u7136\u5f62\u6210\u7684\u4e13\u5bb6\u504f\u597d\uff0c\u4f46\u5b8c\u5168\u4e0d\u5747\u8861\u53c8\u4f1a\u635f\u5bb3\u9ad8\u5c42\u8bed\u4e49\u63a8\u7406\u7684\u6cdb\u5316\u3002\u8bba\u6587\u540e\u9762\u7684\u6d88\u878d\u548c\u5206\u6790\u4e13\u95e8\u8ba8\u8bba\u4e86\u8fd9\u4e2a\u77db\u76fe\u3002<\/p>\n\n\n<p><!-- formulas-added:2601-balance-loss --><\/p>\n\n\n<p>\u8bba\u6587\u5c06 expert importance \u4e0e expert load \u5206\u522b\u5b9a\u4e49\u4e3a\uff1a<\/p>\n\n\n\n\\(\n\\bar{P}_e=\\frac{1}{B}\\sum_{b=1}^{B}p_{b,e}\n\\)\n\n\n\n\\(\n\\bar{f}_e=\\frac{1}{B}\\sum_{b=1}^{B}r_{b,e}\n\\)\n\n\n\n<p>\u6700\u7ec8 auxiliary loss \u4e3a\uff1a<\/p>\n\n\n\n\\(\nL_{\\mathrm{aux}}=|\\mathcal{E}_R|\\sum_{e\\in\\mathcal{E}_R}\\bar{P}_e\\cdot\\bar{f}_e\n\\)\n\n\n\n<h3>\u5b9e\u9a8c<\/h3>\n\n\n\n<h4>\u5b9e\u9a8c\u8bbe\u7f6e<\/h4>\n\n\n\n<p>LLM \u9aa8\u5e72\u662f Qwen3-1.7B\uff0c\u97f3\u9891\u524d\u7aef\u4f7f\u7528 Whisper-VQ tokenizer \u548c Whisper Encoder\u3002\u8bad\u7ec3\u8bed\u6599\u89c4\u6a21\u4e3a 40B token\uff0c\u4f18\u5316\u5668\u4e3a AdamW\uff0c\u5b66\u4e60\u7387\u8c03\u5ea6\u91c7\u7528 Warmup-Stable-Decay\u3002\u4e3a\u4e86\u516c\u5e73\u6bd4\u8f83\uff0cdense adapter \u4e0e MoE-Adapter \u7684\u603b\u53c2\u6570\u9884\u7b97\u90fd\u9650\u5236\u5728\u7ea6 94.4M\u3002<\/p>\n\n\n\n<p>\u8bc4\u6d4b\u8986\u76d6\u51e0\u7c7b\u80fd\u529b\u3002MMAU \u7528\u4e8e\u97f3\u9891\u611f\u77e5\u548c\u526f\u8bed\u8a00\u7406\u89e3\uff0c\u8986\u76d6 speech\u3001sound\u3001music \u7b49\u573a\u666f\uff1bVoiceBench \u4e2d\u7684 MMSU \u548c OpenBookQA \u5b50\u96c6\u7528\u4e8e\u4e16\u754c\u77e5\u8bc6\u548c\u8bed\u4e49\u63a8\u7406\uff0c\u5b83\u4eec\u662f\u4ece\u6587\u672c\u63a8\u7406\u57fa\u51c6\u6539\u9020\u6765\u7684\u97f3\u9891\u7248\u672c\u3002\u6240\u6709\u8bc4\u6d4b\u91c7\u7528 greedy decoding\uff0c\u907f\u514d\u91c7\u6837\u968f\u673a\u6027\u5e72\u6270\u6bd4\u8f83\u3002<\/p>\n\n\n\n<h4>\u4e3b\u7ed3\u679c<\/h4>\n\n\n\n<p>\u5728\u77e5\u8bc6\u63a8\u7406\u4efb\u52a1\u4e0a\uff0cMoE-Adapter \u660e\u663e\u8d85\u8fc7 dense baseline\u3002MMSU \u7684 audio accuracy \u4ece 35.03 \u63d0\u5347\u5230 38.19\uff0cOBQA \u4ece 50.10 \u63d0\u5347\u5230 53.85\u3002\u5bf9\u6bd4\u6587\u672c\u8f93\u5165\u51c6\u786e\u7387\uff0c\u97f3\u9891\u8f93\u5165\u4ecd\u5b58\u5728\u660e\u663e gap\uff0c\u4f46 MoE \u628a\u8fd9\u4e2a\u5dee\u8ddd\u5206\u522b\u7f29\u5c0f\u4e86\u7ea6 3.16 \u548c 3.75 \u4e2a\u70b9\u3002<\/p>\n\n\n\n<figure class=\"wp-block-image size-full\"><img loading=\"lazy\" width=\"1014\" height=\"349\" src=\"http:\/\/139.9.1.231\/wp-content\/uploads\/2026\/07\/image-10.png\" alt=\"\" class=\"wp-image-31342\" srcset=\"http:\/\/139.9.1.231\/wp-content\/uploads\/2026\/07\/image-10.png 1014w, http:\/\/139.9.1.231\/wp-content\/uploads\/2026\/07\/image-10-300x103.png 300w, http:\/\/139.9.1.231\/wp-content\/uploads\/2026\/07\/image-10-768x264.png 768w\" sizes=\"(max-width: 1014px) 100vw, 1014px\" \/><\/figure>\n\n\n\n<p>\u5728 MMAU \u8fd9\u7c7b\u526f\u8bed\u8a00\u548c\u97f3\u9891\u611f\u77e5\u4efb\u52a1\u4e0a\uff0cMoE-Adapter \u4e5f\u4ece 59.79 \u63d0\u5347\u5230 61.50\u3002\u8fd9\u4e2a\u63d0\u5347\u8bf4\u660e\u4e13\u5bb6\u8def\u7531\u4e0d\u4ec5\u5bf9\u77e5\u8bc6\u63a8\u7406\u6709\u7528\uff0c\u4e5f\u80fd\u5e2e\u52a9\u6a21\u578b\u6355\u6349\u66f4\u590d\u6742\u7684\u58f0\u5b66\u7ebf\u7d22\u3002\u8bba\u6587\u5f3a\u8c03\uff0cMoE \u7684\u6536\u76ca\u4e0d\u662f\u5355\u7eaf\u53c2\u6570\u53d8\u591a\uff0c\u800c\u662f\u5728\u76f8\u8fd1\u603b\u53c2\u6570\u9884\u7b97\u4e0b\u66f4\u5408\u7406\u5730\u5206\u914d\u8868\u793a\u80fd\u529b\u3002<\/p>\n\n\n\n<h4>\u6d88\u878d\u5b9e\u9a8c<\/h4>\n\n\n\n<p>\u4e13\u5bb6\u914d\u7f6e\u65b9\u9762\uff0c\u9ed8\u8ba4\u7684 \u201c8 choose 4\u201d \u8868\u73b0\u6700\u5747\u8861\uff1aMMAU 61.50\u3001MMSU 38.19\u3001OBQA 53.85\u3002\u628a\u4e13\u5bb6\u6570\u6269\u5927\u5230 \u201c16 choose 4\u201d \u53cd\u800c\u53d8\u5dee\uff0c\u8bf4\u660e\u4e13\u5bb6\u603b\u6570\u4e0d\u662f\u8d8a\u591a\u8d8a\u597d\u3002\u628a\u8def\u7531\u53d8\u5f97\u8fc7\u7a00\u758f\uff0c\u4f8b\u5982 \u201c8 choose 1\u201d\uff0c\u4e5f\u4f1a\u663e\u8457\u4f24\u5bb3\u97f3\u9891\u63a8\u7406\u3002\u8bba\u6587\u7684\u7ed3\u8bba\u662f\uff0c\u4e13\u5bb6\u6570\u91cf\u3001\u6fc0\u6d3b\u6570\u91cf\u548c\u4e13\u5bb6\u5bb9\u91cf\u4e4b\u95f4\u9700\u8981\u5e73\u8861\uff0c\u800c\u4e0d\u662f\u76f2\u76ee\u6269\u67d0\u4e00\u4e2a\u7ef4\u5ea6\u3002<\/p>\n\n\n\n<p>\u8d1f\u8f7d\u5747\u8861\u635f\u5931\u7684\u6d88\u878d\u66f4\u6709\u610f\u601d\u3002\u53bb\u6389 EBL \u540e\uff0cMMAU \u4ece 61.50 \u5347\u5230 63.01\uff0c\u4f46 MMSU \u548c OBQA \u5206\u522b\u4e0b\u964d\u5230 37.37 \u548c 52.31\u3002\u4f5c\u8005\u89e3\u91ca\u8bf4\uff0cMMAU \u5f88\u5f02\u8d28\u4e14\u542b\u6709\u5927\u91cf\u4f4e\u5c42\u58f0\u5b66\u611f\u77e5\u6837\u672c\uff0c\u4e0d\u52a0\u5747\u8861\u65f6 router \u4f1a\u96c6\u4e2d\u4f7f\u7528\u5c11\u6570\u201c\u5f3a\u4e13\u5bb6\u201d\uff0c\u53cd\u800c\u6709\u5229\u4e8e\u8fd9\u7c7b\u611f\u77e5\u4efb\u52a1\uff1b\u4f46\u8fd9\u4f1a\u51cf\u5c11\u4e13\u5bb6\u591a\u6837\u6027\uff0c\u635f\u5bb3\u9700\u8981\u4e16\u754c\u77e5\u8bc6\u548c\u8bed\u4e49\u63a8\u7406\u7684\u4efb\u52a1<\/p>\n\n\n\n<figure class=\"wp-block-image size-full\"><img loading=\"lazy\" width=\"1014\" height=\"295\" src=\"http:\/\/139.9.1.231\/wp-content\/uploads\/2026\/07\/image-11.png\" alt=\"\" class=\"wp-image-31344\" srcset=\"http:\/\/139.9.1.231\/wp-content\/uploads\/2026\/07\/image-11.png 1014w, http:\/\/139.9.1.231\/wp-content\/uploads\/2026\/07\/image-11-300x87.png 300w, http:\/\/139.9.1.231\/wp-content\/uploads\/2026\/07\/image-11-768x223.png 768w\" sizes=\"(max-width: 1014px) 100vw, 1014px\" \/><\/figure>\n\n\n\n<figure class=\"wp-block-image size-large\"><img loading=\"lazy\" width=\"1024\" height=\"194\" src=\"http:\/\/139.9.1.231\/wp-content\/uploads\/2026\/07\/image-12-1024x194.png\" alt=\"\" class=\"wp-image-31345\" srcset=\"http:\/\/139.9.1.231\/wp-content\/uploads\/2026\/07\/image-12-1024x194.png 1024w, http:\/\/139.9.1.231\/wp-content\/uploads\/2026\/07\/image-12-300x57.png 300w, http:\/\/139.9.1.231\/wp-content\/uploads\/2026\/07\/image-12-768x146.png 768w, http:\/\/139.9.1.231\/wp-content\/uploads\/2026\/07\/image-12.png 1033w\" sizes=\"(max-width: 1024px) 100vw, 1024px\" \/><\/figure>\n\n\n\n<h3>\u4e13\u5bb6\u5206\u5de5\u4e0e\u4f18\u5316\u52a8\u6001\u5206\u6790<\/h3>\n\n\n\n<h4>\u4e13\u5bb6\u5747\u8861\u5982\u4f55\u5f71\u54cd\u8def\u7531<\/h4>\n\n\n\n<p>\u4f5c\u8005\u5728 MMAU \u4e0a\u5206\u6790 speech\u3001sound\u3001music \u4e09\u7c7b\u6837\u672c\u7684\u4e13\u5bb6\u6fc0\u6d3b\u70ed\u529b\u56fe\u3002\u7ed3\u679c\u663e\u793a\uff0c\u6a21\u578b\u786e\u5b9e\u5b66\u51fa\u4e86\u6a21\u6001\u76f8\u5173\u7684\u4e13\u5bb6\u5206\u5de5\uff1a\u6709\u4e9b\u4e13\u5bb6\u4e3b\u8981\u670d\u52a1\u5355\u4e00\u7c7b\u522b\uff0c\u6709\u4e9b\u4e13\u5bb6\u5728 sound \u4e0e speech \u6216 sound \u4e0e music \u4e4b\u95f4\u5171\u4eab\u3002\u503c\u5f97\u6ce8\u610f\u7684\u662f\uff0c\u51e0\u4e4e\u6ca1\u6709\u4e13\u5bb6\u540c\u65f6\u4e13\u95e8\u670d\u52a1 speech \u548c music\u3002<\/p>\n\n\n\n<figure class=\"wp-block-image size-full\"><img loading=\"lazy\" width=\"1015\" height=\"334\" src=\"http:\/\/139.9.1.231\/wp-content\/uploads\/2026\/07\/image-13.png\" alt=\"\" class=\"wp-image-31346\" srcset=\"http:\/\/139.9.1.231\/wp-content\/uploads\/2026\/07\/image-13.png 1015w, http:\/\/139.9.1.231\/wp-content\/uploads\/2026\/07\/image-13-300x99.png 300w, http:\/\/139.9.1.231\/wp-content\/uploads\/2026\/07\/image-13-768x253.png 768w\" sizes=\"(max-width: 1015px) 100vw, 1015px\" \/><\/figure>\n\n\n\n<p>\u8fd9\u4e2a\u73b0\u8c61\u7b26\u5408\u76f4\u89c9\uff1a\u73af\u5883 sound \u548c speech\/music \u90fd\u53ef\u80fd\u5171\u4eab\u4e00\u4e9b\u4f4e\u5c42\u58f0\u5b66\u7279\u5f81\uff0c\u56e0\u6b64\u53ef\u4ee5\u4f5c\u4e3a\u201c\u6865\u201d\uff1b\u4f46 speech \u548c music \u5728\u65f6\u95f4\u7ed3\u6784\u3001\u8bed\u4e49\u7ec4\u7ec7\u4e0a\u5dee\u5f02\u66f4\u5927\uff0c\u4e0d\u9002\u5408\u5f3a\u884c\u585e\u8fdb\u540c\u4e00\u4e2a\u4e13\u5bb6\u3002EBL \u5e76\u4e0d\u4f1a\u6d88\u706d\u8fd9\u79cd\u5206\u5de5\uff0c\u800c\u662f\u9632\u6b62\u5c11\u6570\u4e13\u5bb6\u8fc7\u5ea6\u652f\u914d\uff0c\u4fdd\u7559\u4e00\u5b9a\u5747\u8861\u3002<\/p>\n\n\n\n<h4>\u68af\u5ea6\u51b2\u7a81\u4e0e\u7f13\u89e3\u673a\u5236<\/h4>\n\n\n\n<p>\u8bba\u6587\u7528\u4e24\u4e2a\u6307\u6807\u5206\u6790\u4f18\u5316\u8fc7\u7a0b\u3002\u7b2c\u4e00\u4e2a\u662f\u4e0d\u540c\u97f3\u9891\u7c7b\u522b\u68af\u5ea6\u4e4b\u95f4\u7684 cosine similarity\u3002dense adapter \u4e2d\uff0c\u4e0d\u540c\u7c7b\u522b\u7684\u68af\u5ea6\u7ecf\u5e38\u51fa\u73b0\u8d1f\u76f8\u4f3c\u5ea6\uff0c\u610f\u5473\u7740\u4e00\u4e2a\u7c7b\u522b\u7684\u66f4\u65b0\u65b9\u5411\u53ef\u80fd\u4f24\u5bb3\u53e6\u4e00\u4e2a\u7c7b\u522b\u3002MoE-Adapter \u5219\u628a\u8fd9\u4e9b\u76f8\u4f3c\u5ea6\u63a8\u5411\u66f4\u6b63\u7684\u65b9\u5411\uff0c\u8bf4\u660e\u4e13\u5bb6\u8def\u7531\u51cf\u5c11\u4e86\u7834\u574f\u6027\u5e72\u6270\u3002<\/p>\n\n\n\n<figure class=\"wp-block-image size-large\"><img loading=\"lazy\" width=\"1024\" height=\"385\" src=\"http:\/\/139.9.1.231\/wp-content\/uploads\/2026\/07\/image-14-1024x385.png\" alt=\"\" class=\"wp-image-31347\" srcset=\"http:\/\/139.9.1.231\/wp-content\/uploads\/2026\/07\/image-14-1024x385.png 1024w, http:\/\/139.9.1.231\/wp-content\/uploads\/2026\/07\/image-14-300x113.png 300w, http:\/\/139.9.1.231\/wp-content\/uploads\/2026\/07\/image-14-768x289.png 768w, http:\/\/139.9.1.231\/wp-content\/uploads\/2026\/07\/image-14.png 1054w\" sizes=\"(max-width: 1024px) 100vw, 1024px\" \/><\/figure>\n\n\n\n<p>\u7b2c\u4e8c\u4e2a\u662f gradient influence score\uff0c\u7528\u6765\u8861\u91cf\u57fa\u4e8e\u67d0\u4e00\u4efb\u52a1\u68af\u5ea6\u505a\u66f4\u65b0\u540e\uff0c\u5bf9\u53e6\u4e00\u4e2a\u4efb\u52a1\u635f\u5931\u662f\u5e2e\u52a9\u8fd8\u662f\u4f24\u5bb3\u3002dense adapter \u4e2d\uff0cspeech \u7684\u66f4\u65b0\u4f1a\u660e\u663e\u4f24\u5bb3 music \u548c sound\uff1bMoE-Adapter \u4e2d\uff0c\u5f71\u54cd\u5206\u6570\u66f4\u591a\u4e3a\u6b63\uff0c\u8bf4\u660e\u5b83\u4e0d\u662f\u7b80\u5355\u9694\u79bb\u4efb\u52a1\uff0c\u8fd8\u80fd\u901a\u8fc7\u5171\u4eab\u4e13\u5bb6\u4fdd\u7559\u6709\u76ca\u8fc1\u79fb\u3002\u4f8b\u5982 speech \u5bf9 sound \u7684\u66f4\u65b0\u53ef\u4ee5\u4ea7\u751f\u6b63\u5411\u5e2e\u52a9\uff0c\u800c music \u7684\u51b2\u7a81\u88ab\u66f4\u597d\u5730\u9694\u5f00\u3002<\/p>\n\n\n\n<figure class=\"wp-block-image size-large\"><img loading=\"lazy\" width=\"1024\" height=\"402\" src=\"http:\/\/139.9.1.231\/wp-content\/uploads\/2026\/07\/image-15-1024x402.png\" alt=\"\" class=\"wp-image-31348\" srcset=\"http:\/\/139.9.1.231\/wp-content\/uploads\/2026\/07\/image-15-1024x402.png 1024w, http:\/\/139.9.1.231\/wp-content\/uploads\/2026\/07\/image-15-300x118.png 300w, http:\/\/139.9.1.231\/wp-content\/uploads\/2026\/07\/image-15-768x302.png 768w, http:\/\/139.9.1.231\/wp-content\/uploads\/2026\/07\/image-15.png 1039w\" sizes=\"(max-width: 1024px) 100vw, 1024px\" \/><\/figure>\n\n\n\n<h3>\u7ed3\u8bba<\/h3>\n\n\n\n<p>\u8fd9\u7bc7\u8bba\u6587\u628a MoE-Adapter \u5b9a\u4f4d\u4e3a\u89e3\u51b3 LALM \u97f3\u9891\u5f02\u8d28\u6027\u7684\u7ed3\u6784\u5de5\u5177\u3002\u76f8\u6bd4 dense adapter\uff0c\u5b83\u7528\u52a8\u6001\u4e13\u5bb6\u5206\u5de5\u7f13\u89e3\u8bed\u97f3\u3001\u97f3\u4e50\u3001\u73af\u5883\u58f0\u4e4b\u95f4\u7684\u68af\u5ea6\u51b2\u7a81\uff0c\u5728\u76f8\u8fd1\u53c2\u6570\u9884\u7b97\u4e0b\u63d0\u5347\u97f3\u9891\u77e5\u8bc6\u63a8\u7406\u3001\u526f\u8bed\u8a00\u7406\u89e3\u548c\u8de8\u6a21\u6001\u5bf9\u9f50\u8868\u73b0\u3002\u66f4\u91cd\u8981\u7684\u662f\uff0c\u8bba\u6587\u4e0d\u4ec5\u7ed9\u51fa\u6307\u6807\u63d0\u5347\uff0c\u4e5f\u901a\u8fc7\u8def\u7531\u70ed\u529b\u56fe\u3001\u68af\u5ea6\u76f8\u4f3c\u5ea6\u548c\u5f71\u54cd\u5206\u6570\u89e3\u91ca\u4e86\u4e3a\u4ec0\u4e48 MoE \u6709\u6548\u3002<\/p>\n\n\n\n<h3>\u5c40\u9650\u6027<\/h3>\n\n\n\n<p>\u4f5c\u8005\u4e5f\u660e\u786e\u6307\u51fa\u4e86\u51e0\u4e2a\u9650\u5236\u3002\u7b2c\u4e00\uff0c\u5b9e\u9a8c\u76ee\u524d\u53ea\u5728 Qwen3-1.7B \u9aa8\u5e72\u4e0a\u9a8c\u8bc1\uff0c\u65b9\u6cd5\u662f\u5426\u9002\u7528\u4e8e\u5176\u4ed6 LLM \u5bb6\u65cf\u6216\u66f4\u5927\u89c4\u6a21\u6a21\u578b\uff0c\u4f8b\u5982 70B\uff0c\u8fd8\u9700\u8981\u5b9e\u9a8c\u3002\u7b2c\u4e8c\uff0c\u8bba\u6587\u6ca1\u6709\u7cfb\u7edf\u7814\u7a76\u7a00\u758f\u8def\u7531\u968f\u8bad\u7ec3\u6570\u636e\u89c4\u6a21\u589e\u957f\u7684 scaling law\u3002\u7b2c\u4e09\uff0c\u5f53\u524d\u4efb\u52a1\u96c6\u4e2d\u5728\u97f3\u9891\u7406\u89e3\u4e0e\u63a8\u7406\uff0c\u6ca1\u6709\u6269\u5c55\u5230\u751f\u6210\u5f0f\u97f3\u9891\u4efb\u52a1\u3002<\/p>\n\n\n\n<h3>\u9644\u5f55\uff1a\u8d85\u53c2\u6570\u548c\u5de5\u7a0b\u542b\u4e49<\/h3>\n\n\n\n<p>\u9644\u5f55\u5f3a\u8c03\uff0cdense baseline \u4e0e MoE-Adapter \u5728\u603b\u53c2\u6570\u9884\u7b97\u4e0a\u88ab\u4e25\u683c\u5bf9\u9f50\uff0c\u7ea6\u4e3a 94.4M\u3002MoE-Adapter \u7684\u6d3b\u8dc3\u53c2\u6570\u7ea6\u4e3a 70.8M\uff0c\u5927\u7ea6\u662f dense baseline \u7684 75%\u3002\u5171\u540c\u97f3\u9891\u524d\u7aef\u5305\u542b speech encoder\u3001audio hidden projection \u548c feature fusion\uff1bMoE \u7aef\u5219\u5305\u542b\u4e13\u5bb6\u96c6\u5408\u3001gate network \u548c aggregation block\u3002<\/p>\n\n\n\n<p>\u4ece\u5de5\u7a0b\u89d2\u5ea6\u770b\uff0c\u8fd9\u4e2a\u8bbe\u7f6e\u5f88\u5173\u952e\u3002\u5b83\u628a\u8bba\u6587\u7684\u7ed3\u8bba\u4ece\u201cMoE \u53c2\u6570\u66f4\u591a\u6240\u4ee5\u66f4\u5f3a\u201d\u62c9\u56de\u5230\u201c\u5728\u53ef\u6bd4\u9884\u7b97\u4e0b\uff0c\u7a00\u758f\u4e13\u5bb6\u5206\u5de5\u66f4\u9002\u5408\u5f02\u8d28\u97f3\u9891\u201d\u3002\u5982\u679c\u672a\u6765\u5927\u97f3\u9891\u8bed\u8a00\u6a21\u578b\u8981\u540c\u65f6\u5904\u7406\u8bed\u97f3\u95ee\u7b54\u3001\u73af\u5883\u58f0\u63a8\u7406\u3001\u97f3\u4e50\u7406\u89e3\u548c\u60c5\u7eea\u97f5\u5f8b\uff0cadapter \u5c42\u53ef\u80fd\u4e0d\u8be5\u518d\u662f\u4e00\u4e2a\u5355\u4f53\u6295\u5f71\u5668\uff0c\u800c\u5e94\u8be5\u5177\u5907\u66f4\u7ec6\u7c92\u5ea6\u7684\u8def\u7531\u548c\u5206\u5de5\u80fd\u529b\u3002<\/p>\n","protected":false},"excerpt":{"rendered":"<p>MOE- Conformer : Mixture-of-Expert Conformer for Stream &hellip; <a href=\"http:\/\/139.9.1.231\/index.php\/2026\/07\/11\/asr-moe\/\" class=\"more-link\">\u7ee7\u7eed\u9605\u8bfb<span class=\"screen-reader-text\">ASR\u8bed\u97f3\u8bc6\u522b-MOE\u67b6\u6784\u8bba\u6587<\/span><\/a><\/p>\n","protected":false},"author":1,"featured_media":0,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":[],"categories":[42,40,4,9,38,43,34],"tags":[],"_links":{"self":[{"href":"http:\/\/139.9.1.231\/index.php\/wp-json\/wp\/v2\/posts\/31165"}],"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=31165"}],"version-history":[{"count":41,"href":"http:\/\/139.9.1.231\/index.php\/wp-json\/wp\/v2\/posts\/31165\/revisions"}],"predecessor-version":[{"id":31355,"href":"http:\/\/139.9.1.231\/index.php\/wp-json\/wp\/v2\/posts\/31165\/revisions\/31355"}],"wp:attachment":[{"href":"http:\/\/139.9.1.231\/index.php\/wp-json\/wp\/v2\/media?parent=31165"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"http:\/\/139.9.1.231\/index.php\/wp-json\/wp\/v2\/categories?post=31165"},{"taxonomy":"post_tag","embeddable":true,"href":"http:\/\/139.9.1.231\/index.php\/wp-json\/wp\/v2\/tags?post=31165"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}