{"id":29962,"date":"2026-01-10T18:27:18","date_gmt":"2026-01-10T10:27:18","guid":{"rendered":"http:\/\/139.9.1.231\/?p=29962"},"modified":"2026-01-10T18:27:20","modified_gmt":"2026-01-10T10:27:20","slug":"llm-based-asr-with-hotword-retrieval","status":"publish","type":"post","link":"http:\/\/139.9.1.231\/index.php\/2026\/01\/10\/llm-based-asr-with-hotword-retrieval\/","title":{"rendered":"\u57fa\u4e8e\u5927\u8bed\u8a00\u6a21\u578b\u7684\u8bed\u97f3\u8bc6\u522b\u4e0a\u4e0b\u6587\u504f\u7f6e\uff1a\u70ed\u8bcd\u68c0\u7d22\u4e0e\u5f3a\u5316\u5b66\u4e60\u65b9\u6cd5"},"content":{"rendered":"\n<ul class=\"has-light-gray-background-color has-background\"><li><strong><em>\u8bba\u6587\uff1aContextual Biasing for LLM-based ASR with Hotword Retrieval and Reinforcement Learning<\/em><\/strong><\/li><li><strong><em>\u4f5c\u8005\uff1a\u963f\u91cc\u5df4\u5df4\u901a\u4e49\u5b9e\u9a8c\u5ba4<\/em><\/strong><\/li><li><strong><em>\u8bba\u6587\u94fe\u63a5\uff1a<a href=\"https:\/\/arxiv.org\/abs\/2512.21828\">https:\/\/arxiv.org\/abs\/2512.21828<\/a><\/em><\/strong><\/li><li><strong><em>GLCLAP: A Novel Contrastive Learning Pre-trained Model for Contextual Biasing in ASR<\/em>: <em><a href=\"https:\/\/www.isca-archive.org\/interspeech_2025\/kong25_interspeech.pdf\">https:\/\/www.isca-archive.org\/interspeech_2025\/kong25_interspeech.pdf<\/a><\/em><\/strong><\/li><\/ul>\n\n\n\n\n\n<p>\u5728\u771f\u5b9e\u4e1a\u52a1\u91cc\uff0c\u5982\u679c\u4f60\u505a\u8fc7\u8bed\u97f3\u8bc6\u522b\u843d\u5730\uff0c\u5927\u6982\u7387\u4f1a\u9047\u5230\u7c7b\u4f3c\u7684\u5d29\u6e83\u77ac\u95f4\uff1a<\/p>\n\n\n\n<ul><li>\u5728\u533b\u7597\u573a\u666f\uff0c\u533b\u751f\u53e3\u8ff0\u4e00\u957f\u4e32\u836f\u54c1\u540d\u3001\u75c5\u7406\u540d\uff0c\u6a21\u578b\u80fd\u628a\u666e\u901a\u53e3\u8bed\u8bc6\u5f97\u5f88\u597d\uff0c\u4e00\u5230\u4e13\u4e1a\u540d\u8bcd\u5c31\u5f00\u59cb\u300c\u7f16\u300d\uff0c\u8fd8\u7ecf\u5e38\u628a\u8bcd\u8868\u91cc\u6ca1\u51fa\u73b0\u7684\u836f\u540d\u786c\u8bf4\u51fa\u6765\uff1b<\/li><li>\u5f71\u89c6\u5a92\u4f53\u3001\u77ed\u89c6\u9891\u9886\u57df\uff0c\u5267\u540d\u3001\u89d2\u8272\u540d\u3001\u827a\u4eba\u540d\u6bcf\u5929\u90fd\u5728\u66f4\u65b0\uff0c\u70ed\u8bcd\u8bcd\u8868\u8f7b\u8f7b\u677e\u677e\u51e0\u5341\u4e07\u6761\uff0c\u6a21\u578b\u4e00\u65e6\u201c\u8ba4\u4e0d\u4f4f\u3001\u8ba4\u4e0d\u51c6\u201d\uff0c\u7528\u6237\u641c\u4e0d\u51fa\u4e1c\u897f\u3002<\/li><\/ul>\n\n\n\n<p>\u5373\u4f7f\u5bf9\u4e8eLLM-ASR\u8fd9\u79cd\u5f3a\u5927\u7684\u8bed\u97f3\u8bc6\u522b\u6a21\u578b\uff0c\u5728\u843d\u5730\u8fc7\u7a0b\u4e2d\u4e5f\u7ed5\u4e0d\u5f00\u70ed\u8bcd\u8fd9\u4e2a\u8bdd\u9898\u3002<\/p>\n\n\n\n<p>\u901a\u4e49\u6700\u65b0\u5de5\u4f5c\u4e2d\u63d0\u51fa\u4e86\u4e00\u4e2a\u9762\u5411\u00a0LLM-ASR\u00a0\u7684\u53ef\u6269\u5c55\u4e0a\u4e0b\u6587\u504f\u7f6e\u6846\u67b6\uff0c\u628a<strong>\u201c\u70ed\u8bcd\u68c0\u7d22\u00a0+ LLM\u00a0\u81ea\u9002\u5e94\u00a0+\u00a0\u5f3a\u5316\u5b66\u4e60\u201d<\/strong>\u4e32\u6210\u4e00\u5957\u7cfb\u7edf\uff0c\u5728<strong>\u5927\u89c4\u6a21\u70ed\u8bcd\u573a\u666f\uff08\u8fd1\u00a010\u00a0\u4e07\u89c4\u6a21\u8bcd\u8868\uff09\u4e0b\u663e\u8457\u63d0\u5347\u4e86\u70ed\u8bcd\u8bc6\u522b\u80fd\u529b\uff0c\u540c\u65f6\u63d0\u5347\u4e86\u6574\u4f53\u8bc6\u522b\u6548\u679c\u3002<\/strong>\u5177\u4f53\u6765\u8bf4\uff0c\u9996\u5148\uff0c\u6269\u5c55\u4e86 Global\u2013Local Contrastive Language\u2013Audio Pre-trained\uff08GLCLAP\uff09\u6a21\u578b\uff0c\u901a\u8fc7\u5177\u5907\u9c81\u68d2\u6027\u7684\u6570\u636e\u589e\u5f3a\u4e0e\u6a21\u7cca\u5339\u914d\u673a\u5236\uff0c\u4ece\u5927\u89c4\u6a21\u8bcd\u8868\u4e2d\u68c0\u7d22\u51fa\u4e00\u4e2a\u7d27\u51d1\u7684 top-k \u70ed\u8bcd\u5019\u9009\u96c6\u5408\u3002\u5176\u6b21\uff0c\u5c06\u68c0\u7d22\u5230\u7684\u5019\u9009\u70ed\u8bcd\u4ee5\u6587\u672c\u63d0\u793a\u7684\u5f62\u5f0f\u6ce8\u5165\u5230 LLM-ASR \u6a21\u578b\u4e2d\uff0c\u5e76<strong>\u91c7\u7528GRPO\u8fdb\u884c\u5f3a\u5316\u5fae\u8c03<\/strong>\uff0c\u4f7f\u7528\u4efb\u52a1\u9a71\u52a8\u7684\u5956\u52b1\u51fd\u6570\u540c\u65f6<strong>\u4f18\u5316\u70ed\u8bcd\u8bc6\u522b\u6027\u80fd\u548c\u6574\u4f53\u8f6c\u5199\u51c6\u786e\u7387<\/strong>\u3002<\/p>\n\n\n\n<h3><strong>\u603b\u4f53\u6846\u67b6\uff1a\u68c0\u7d22\u00a0+\u00a0\u5f3a\u5316\u5b66\u4e60\uff0c\u4e24\u9636\u6bb5\u534f\u540c<\/strong><\/h3>\n\n\n\n<ol><li>\u70ed\u8bcd\u68c0\u7d22\uff08Hotword Retrieval\uff09\uff1a<br>\u4ece\u5927\u8bcd\u8868\u4e2d\uff0c\u4e3a\u5f53\u524d\u8bed\u97f3\u68c0\u7d22\u51fa\u4e00\u5c0f\u64ae\u6700\u76f8\u5173\u7684&nbsp;top-k&nbsp;\u70ed\u8bcd\uff1b<\/li><li>\u70ed\u8bcd\u611f\u77e5&nbsp;ASR&nbsp;\u9002\u914d\uff08Hotword-aware ASR Adaptation\uff09\uff1a<br>\u628a\u68c0\u7d22\u51fa\u7684\u70ed\u8bcd\u4ee5&nbsp;prompt&nbsp;\u5f62\u5f0f\u5582\u7ed9&nbsp;LLM-ASR\uff0c\u5e76\u7528\u5f3a\u5316\u5b66\u4e60\u4f18\u5316\u5176\u4f7f\u7528\u7b56\u7565\u3002<\/li><\/ol>\n\n\n\n<p>\u6574\u4f53\u7ed3\u6784\u53ef\u4ee5\u7c7b\u6bd4\u4e3a\u8bed\u97f3\u7248\u7684&nbsp;RAG\uff08Retrieval-Augmented Generation\uff09\uff1a<\/p>\n\n\n\n<ul><li>\u68c0\u7d22\u4fa7\uff1a\u57fa\u4e8e\u6539\u8fdb\u7248\u00a0<strong>GLCLAP<\/strong>\uff08<strong>Global\u2013Local Contrastive Language\u2013Audio Pre-trained Model<\/strong>\uff09<strong>\u505a\u97f3\u9891\u00a0\u2194\u00a0\u70ed\u8bcd\u6587\u672c\u7684\u5339\u914d<\/strong>\uff1b<\/li><li>\u8bc6\u522b\u4fa7\uff1a\u628a\u68c0\u7d22\u5230\u7684\u70ed\u8bcd\u653e\u5230\u00a0LLM-ASR\u7684\u6587\u672c\u00a0prompt\u00a0\u4e2d\uff0c\u7528\u00a0GRPO\uff08Generative Rejection-based Policy Optimization\uff09\u505a\u00a0RL\u00a0\u5fae\u8c03\uff0c\u8ba9\u6a21\u578b\u5b66\u4f1a\uff1a<ul><li>\u5bf9\u771f\u6b63\u51fa\u73b0\u7684\u70ed\u8bcd\u8981\u201c\u8ba4\u5f97\u51c6\u201d\uff1b<\/li><\/ul><ul><li>\u5bf9\u6ca1\u51fa\u73b0\u7684\u70ed\u8bcd\u4e0d\u8981\u201c\u778e\u731c\u201d\uff1b<\/li><\/ul><ul><li>\u517c\u987e\u6574\u4f53\u8f6c\u5199\u7684\u00a0WER \/\u00a0\u53e5\u5b50\u51c6\u786e\u7387\u3002<\/li><\/ul><\/li><\/ul>\n\n\n\n<h4><strong>\u589e\u5f3a\u7248\u00a0GLCLAP\u00a0\u70ed\u8bcd\u68c0\u7d22<\/strong><\/h4>\n\n\n\n<p>GLCLAP \u68c0\u7d22\u5668\u4ee5\u97f3\u9891\u4fe1\u53f7 <em>x<\/em> \u53ca\u5019\u9009\u504f\u7f6e\u8bcd\u96c6\u5408 <em>G={g<sub>1<\/sub>,g<sub>2<\/sub>,\u2026,g<sub>N<\/sub>}<\/em>\u4f5c\u4e3a\u8f93\u5165\uff0c\u5176\u4e2d <em>N<\/em> \u8868\u793a\u5019\u9009\u8bcd\u8868\u7684\u89c4\u6a21\u3002\u8be5\u68c0\u7d22\u5668\u7531\u4e24\u4e2a\u7ec4\u4ef6\u6784\u6210\uff1a\u97f3\u9891\u7f16\u7801\u5668\uff08A-enc\uff09\u548c\u6587\u672c\u7f16\u7801\u5668\uff08T-enc\uff09\u3002\u5bf9\u4e8e\u8f93\u5165\u97f3\u9891\uff0cA-enc \u63d0\u53d6\u4e00\u4e2a\u56fa\u5b9a\u7ef4\u5ea6\u7684\u97f3\u9891\u5d4c\u5165\u8868\u793a\uff0c\u8bb0\u4e3a <em>h<sub>audi<\/sub><\/em><sub>o<\/sub>\u3002\u540c\u65f6\uff0c\u5019\u9009\u96c6\u5408\u4e2d\u7684\u6bcf\u4e00\u4e2a\u504f\u7f6e\u8bcd <strong><em>g<sub>i<\/sub>\u2208G<\/em><\/strong> \u901a\u8fc7 T-enc \u7f16\u7801\u4e3a\u5bf9\u5e94\u7684\u8bed\u4e49\u5411\u91cf <em>e<sub>i\u200b<\/sub><\/em>\uff0c\u4ece\u800c\u5f97\u5230\u6587\u672c\u5d4c\u5165\u96c6\u5408<em> E={e<sub>1<\/sub>,e<sub>2<\/sub>,\u2026,e<sub>N<\/sub>}<\/em>\u3002\u968f\u540e\uff0c\u6211\u4eec\u8ba1\u7b97 <em>h<\/em><sub><em>audio<\/em>\u200b<\/sub> \u4e0e\u96c6\u5408<em> E <\/em>\u4e2d\u6240\u6709\u6587\u672c\u5d4c\u5165\u4e4b\u95f4\u7684\u76f8\u4f3c\u5ea6\u5f97\u5206\uff0c\u5e76\u9009\u53d6\u5f97\u5206\u6700\u9ad8\u7684 top-k \u4e2a\u504f\u7f6e\u8bcd\uff0c\u6784\u6210\u5b50\u96c6<em> G\u2032<\/em>\u3002\u8fd9\u4e9b\u88ab\u9009\u4e2d\u7684\u504f\u7f6e\u8bcd\u968f\u540e\u88ab\u62fc\u63a5\u5230\u504f\u7f6e\u63d0\u793a\uff08bias prompt\uff09\u4e2d\uff0c\u4ee5\u5f15\u5bfc\u6a21\u578b\u8fdb\u884c\u4e0a\u4e0b\u6587\u611f\u77e5\u7684\u8f6c\u5199\u3002<\/p>\n\n\n\n<figure class=\"wp-block-image size-large\"><img loading=\"lazy\" width=\"1024\" height=\"657\" src=\"http:\/\/139.9.1.231\/wp-content\/uploads\/2026\/01\/image-21-1024x657.png\" alt=\"\" class=\"wp-image-30004\" srcset=\"http:\/\/139.9.1.231\/wp-content\/uploads\/2026\/01\/image-21-1024x657.png 1024w, http:\/\/139.9.1.231\/wp-content\/uploads\/2026\/01\/image-21-300x192.png 300w, http:\/\/139.9.1.231\/wp-content\/uploads\/2026\/01\/image-21-768x493.png 768w, http:\/\/139.9.1.231\/wp-content\/uploads\/2026\/01\/image-21.png 1157w\" sizes=\"(max-width: 1024px) 100vw, 1024px\" \/><\/figure>\n\n\n\n<figure class=\"wp-block-image size-full\"><img loading=\"lazy\" width=\"916\" height=\"342\" src=\"http:\/\/139.9.1.231\/wp-content\/uploads\/2026\/01\/image-23.png\" alt=\"\" class=\"wp-image-30008\" srcset=\"http:\/\/139.9.1.231\/wp-content\/uploads\/2026\/01\/image-23.png 916w, http:\/\/139.9.1.231\/wp-content\/uploads\/2026\/01\/image-23-300x112.png 300w, http:\/\/139.9.1.231\/wp-content\/uploads\/2026\/01\/image-23-768x287.png 768w\" sizes=\"(max-width: 916px) 100vw, 916px\" \/><\/figure>\n\n\n\n<p><strong>\u672c\u6587\u5bf9\u00a0GLCLAP\u00a0\u53c8\u505a\u4e86\u4e24\u65b9\u9762\u589e\u5f3a\u3002<\/strong><\/p>\n\n\n\n<p><strong>Robustness-Aware Data Augmentation\uff08RADA\uff09<\/strong>\u3002<\/p>\n\n\n\n<p>\u4e3a\u7f13\u89e3\u70ed\u8bcd\u89c4\u6a21\u6269\u5927\u4f1a\u5bfc\u81f4\u53ec\u56de\u7387\u4e0b\u964d\u53ca\u5e72\u6270\u9879\u589e\u591a\u7684\u95ee\u9898\uff0c\u6211\u4eec\u6784\u5efa\u4e86\u4e00\u5957\u9c81\u68d2\u6027\u611f\u77e5\u6570\u636e\u589e\u5f3a\uff08RADA\uff09\u6d41\u7a0b\uff0c\u7528\u4e8e\u7f29\u51cf\u70ed\u8bcd\u8bcd\u8868\u89c4\u6a21\u3002\u521d\u59cb\u70ed\u8bcd\u8bcd\u8868\u901a\u8fc7\u7f51\u7edc\u722c\u53d6\u9886\u57df\u76f8\u5173\u7684\u70ed\u8bcd\u6784\u5efa\u5f97\u5230\u3002\u5bf9\u4e8e\u5019\u9009\u96c6\u5408<em> G<\/em> \u4e2d\u7684\u6bcf\u4e00\u4e2a\u504f\u7f6e\u8bcd <em>g<sub>i<\/sub><\/em>\uff0c\u6211\u4eec\u9996\u5148\u5229\u7528TTS\u7cfb\u7edf\u5408\u6210\u5bf9\u5e94\u8bed\u97f3\uff08<strong>\u5728\u5fc5\u8981\u65f6\u7531\u5927\u8bed\u8a00\u6a21\u578b\u751f\u6210\u4e0a\u4e0b\u6587\u6587\u672c<\/strong>\uff09\uff0c\u968f\u540e\u4f7f\u7528\u73b0\u6709\u7684 ASR \u7cfb\u7edf\u5bf9\u5408\u6210\u8bed\u97f3\u8fdb\u884c\u89e3\u7801\uff0c\u4ee5\u68c0\u6d4b\u539f\u59cb LLM-ASR <strong>\u662f\u5426\u5df2\u5177\u5907\u5bf9\u8be5\u70ed\u8bcd\u7684\u7a33\u5b9a\u8bc6\u522b\u80fd\u529b<\/strong>\u3002\u82e5\u8be5\u70ed\u8bcd\u80fd\u591f\u88ab\u53ef\u9760\u8bc6\u522b\uff0c\u5219\u5c06\u5176\u4ece\u70ed\u8bcd\u8bcd\u8868\u4e2d\u79fb\u9664\uff1b\u53cd\u4e4b\uff0c\u5219\u4fdd\u7559\u8be5\u8bcd\u4f5c\u4e3a\u540e\u7eed\u504f\u7f6e\u5efa\u6a21\u7684\u76ee\u6807\u3002\u901a\u8fc7\u8be5\u6d41\u7a0b\uff0c<strong>\u6a21\u578b\u8bad\u7ec3\u4e0e\u63a8\u7406\u9636\u6bb5\u4ec5\u9700\u5173\u6ce8\u771f\u6b63\u5177\u6709\u8bc6\u522b\u96be\u5ea6\u7684\u70ed\u8bcd\uff0c\u4ece\u800c\u63d0\u5347\u6574\u4f53\u68c0\u7d22\u4e0e\u504f\u7f6e\u6548\u679c\u3002<\/strong>\u7ed3\u679c\uff1a\u8bcd\u8868\u89c4\u6a21\u4ece\u00a060\u00a0\u4e07\u7f29\u51cf\u5230\u7ea6\u00a09.8\u00a0\u4e07\u3002<\/p>\n\n\n\n<ul><li>\u68c0\u7d22\u96be\u5ea6\u663e\u8457\u964d\u4f4e\uff1b<\/li><li>\u51cf\u5c11\u5927\u91cf\u201c\u5e72\u6270\u201d\u7684\u8bcd\uff0c\u540e\u9762&nbsp;LLM-ASR\u4e5f\u4e0d\u4f1a\u88ab\u8fd9\u4e9b\u8bcd\u5e72\u6270\u3002<\/li><\/ul>\n\n\n\n<p><strong>\u6a21\u7cca\u5339\u914d\u7b56\u7565\uff08Fuzzy Matching Strategy\uff09<\/strong>\uff1a<\/p>\n\n\n\n<p>\u5728\u5b9e\u9645\u5e94\u7528\u573a\u666f\u4e2d\uff0c\u70ed\u8bcd\u5f80\u5f80\u96be\u4ee5\u901a\u8fc7\u4e25\u683c\u7684\u8bcd\u9762\u5339\u914d\u8fdb\u884c\u7ea6\u675f\uff0c\u5426\u5219\u5c06\u5bfc\u81f4\u70ed\u8bcd\u8bcd\u8868\u89c4\u6a21\u6025\u5267\u81a8\u80c0\uff0c\u8fdb\u800c\u964d\u4f4e\u7cfb\u7edf\u7684\u53ef\u6269\u5c55\u6027\u3002\u7136\u800c\uff0c\u5728 GLCLAP \u7684\u8bad\u7ec3\u8fc7\u7a0b\u4e2d\uff0c\u70ed\u8bcd\u901a\u5e38\u901a\u8fc7\u4e25\u683c\u7684\u8bcd\u6c47\u7ea7\u5339\u914d\u8fdb\u884c\u7ea6\u675f\uff0c\u8fd9\u4e0e\u771f\u5b9e\u90e8\u7f72\u73af\u5883\u5b58\u5728\u4e0d\u4e00\u81f4\u6027\u3002\u5728\u771f\u5b9e\u573a\u666f\u4e2d\uff0c<strong>\u7528\u6237\u53ef\u80fd\u4f1a\u4f7f\u7528\u76ee\u6807\u70ed\u8bcd\u7684\u4e0d\u540c\u5f62\u6001\uff08\u5982\u5c48\u6298\u53d8\u5316\uff09\u3001\u8bed\u4e49\u6539\u5199\uff08paraphrases\uff09\u6216\u4ec5\u90e8\u5206\u63d0\u53ca\u76ee\u6807\u672f\u8bed<\/strong>\u3002\u4e3a\u5f25\u5408\u8bad\u7ec3\u4e0e\u90e8\u7f72\u4e4b\u95f4\u7684\u5dee\u5f02\uff0c\u6211\u4eec\u5728 <strong>GLCLAP \u8bad\u7ec3\u9636\u6bb5\u5f15\u5165\u6a21\u7cca\u5339\u914d\u7b56\u7565<\/strong>\uff0c\u4f7f\u6a21\u578b\u80fd\u591f<strong>\u5b66\u4e60\u5230\u66f4\u5177\u8bed\u4e49\u4e0e\u53d1\u97f3\u9c81\u68d2\u6027\u7684\u70ed\u8bcd\u8868\u793a<\/strong>\uff0c\u4ece\u800c\u63d0\u5347\u5728\u590d\u6742\u771f\u5b9e\u573a\u666f\u4e0b\u7684\u68c0\u7d22\u4e0e\u4e0a\u4e0b\u6587\u504f\u7f6e\u80fd\u529b\u3002<\/p>\n\n\n\n<p>\u771f\u5b9e\u573a\u666f\u4e2d\uff0c\u7528\u6237\u8bf4\u7684\u70ed\u8bcd\u5e38\u5e38\u4e0d\u662f\u8bcd\u8868\u91cc\u7684\u6807\u51c6\u5f62\u5f0f\uff0c\u6bd4\u5982\uff1a<\/p>\n\n\n\n<ul><li>\u8bf4\u4e00\u4e2a\u836f\u54c1\u540d\uff0c\u53ef\u80fd\u6709\u80f6\u56ca\u3001\u9897\u7c92\u3001\u53e3\u670d\u6db2\u7b49\u3002<\/li><li>\u8bf4\u4e00\u4e2a\u5f71\u7247\u540d\uff0c\u53ef\u80fd\u6709\u7b2c\u4e8c\u90e8\u3001\u7eed\u3001\u65b0xxx\u7b49\u3002<\/li><\/ul>\n\n\n\n<p>\u5982\u679c\u8bad\u7ec3\u4e2d\u7684\u76d1\u7763\u53ea\u5141\u8bb8\u4e25\u683c\u7684\u5b57\u9762\u5339\u914d\uff0c\u68c0\u7d22\u6a21\u578b\u5c31\u4f1a\u5bf9\u8fd9\u4e9b\u53d8\u4f53\u7f3a\u4e4f\u9c81\u68d2\u6027\u3002\u56e0\u6b64\uff0c\u672c\u6587\u5f15\u5165\u4e86\u00a0<strong>fuzzy matching\uff08\u6a21\u7cca\u5339\u914d\u7b56\u7565\uff09<\/strong>\uff0c\u5728<strong>\u8bad\u7ec3\u9636\u6bb5\u5f15\u5165\u4e86\u7531\u751f\u6210\u5f0f\u4e0a\u4e0b\u6587\u53e5\u5b50\u5d4c\u5165\u4ee5\u53ca\u7ecf\u8fc7\u523b\u610f\u6270\u52a8\u7684\u504f\u7f6e\u8bcd\u53d8\u4f53\u6240\u6784\u6210\u7684\u6570\u636e\u589e\u5f3a<\/strong>\u3002<\/p>\n\n\n\n<ul><li>\u5728\u8bad\u7ec3\u6570\u636e\u4e2d\u589e\u52a0\u591a\u79cd\u53d8\u4f53\uff1a\u4eba\u4e3a\u6270\u52a8\u8bcd\u5f62\uff0c\u5982<strong>Tongyi\u00a0\u2192\u00a0Tongyi abc<\/strong>\u7b49\uff1b<\/li><\/ul>\n\n\n\n<p>\u8fd9\u6837\u505a\u7684\u6548\u679c\uff1a<\/p>\n\n\n\n<ul><li>\u66f4\u8d34\u5408\u771f\u5b9e\u4e1a\u52a1\u4e2d\u201c\u8bcd\u5f62\u53d8\u52a8\u3001\u8bf4\u6cd5\u591a\u6837\u201d\u7684\u573a\u666f\uff1b<\/li><li>\u68c0\u7d22\u6a21\u578b\u5bf9\u70ed\u8bcd\u7684\u8bed\u4e49\u53ca\u5f62\u6001\u53d8\u4f53\u66f4\u9c81\u68d2\u3002<\/li><\/ul>\n\n\n\n<h4>LLM-ASR<\/h4>\n\n\n\n<p>LLM-ASR \u7f51\u7edc\u7531\u97f3\u9891\u7f16\u7801\u5668\u3001\u9002\u914d\u5668\u4ee5\u53caLLM\u4e09\u90e8\u5206\u7ec4\u6210\u3002\u672c\u6587\u9009\u7528 Qwen2.5-7B \u4f5c\u4e3a LLM \u4e3b\u4f53\u3002\u5728\u97f3\u9891\u7f16\u7801\u5668\u65b9\u9762\uff0c\u5c06\u539f\u59cb\u7684 Conformer \u7f16\u7801\u5668\u6269\u5c55\u4e3a Conformer-MoE \u7f16\u7801\u5668\uff0c\u5177\u4f53\u505a\u6cd5\u662f\u5728\u6bcf\u4e00\u5c42 Conformer \u4e2d\uff0c\u5c06\u7b2c\u4e8c\u4e2a\u524d\u9988\u7f51\u7edc\uff08FFN\uff09\u6a21\u5757\u66ff\u6362\u4e3a\u6df7\u5408\u4e13\u5bb6\uff08MoE\uff09\u7ed3\u6784\u3002\u6211\u4eec\u5b9a\u4e49\u4e86 <em>K<sub>C<\/sub><\/em>\u4e2a\u5019\u9009\u4e13\u5bb6\uff0c\u5e76\u901a\u8fc7router\u4ece\u4e2d\u9009\u62e9<em> K<sub>S<\/sub><\/em>\u4e2a\u4e13\u5bb6\u8fdb\u884c\u52a0\u6743\u805a\u5408\uff0c\u540c\u65f6\u4fdd\u7559\u4e00\u4e2a\u4e13\u7528\u7684\u5171\u4eab\u4e13\u5bb6\u4ee5\u63d0\u4f9b\u901a\u7528\u5efa\u6a21\u80fd\u529b\u3002<\/p>\n\n\n\n<figure class=\"wp-block-image size-full\"><img loading=\"lazy\" width=\"1021\" height=\"502\" src=\"http:\/\/139.9.1.231\/wp-content\/uploads\/2026\/01\/image-19.png\" alt=\"\" class=\"wp-image-29971\" srcset=\"http:\/\/139.9.1.231\/wp-content\/uploads\/2026\/01\/image-19.png 1021w, http:\/\/139.9.1.231\/wp-content\/uploads\/2026\/01\/image-19-300x148.png 300w, http:\/\/139.9.1.231\/wp-content\/uploads\/2026\/01\/image-19-768x378.png 768w\" sizes=\"(max-width: 1021px) 100vw, 1021px\" \/><\/figure>\n\n\n\n<figure class=\"wp-block-image size-large\"><img loading=\"lazy\" width=\"1024\" height=\"721\" src=\"http:\/\/139.9.1.231\/wp-content\/uploads\/2026\/01\/image-20-1024x721.png\" alt=\"\" class=\"wp-image-29979\" srcset=\"http:\/\/139.9.1.231\/wp-content\/uploads\/2026\/01\/image-20-1024x721.png 1024w, http:\/\/139.9.1.231\/wp-content\/uploads\/2026\/01\/image-20-300x211.png 300w, http:\/\/139.9.1.231\/wp-content\/uploads\/2026\/01\/image-20-768x541.png 768w, http:\/\/139.9.1.231\/wp-content\/uploads\/2026\/01\/image-20.png 1055w\" sizes=\"(max-width: 1024px) 100vw, 1024px\" \/><\/figure>\n\n\n\n<p>LLM-ASR \u6a21\u578b\u603b\u53c2\u6570\u89c4\u6a21\u4e3a 10.5B\uff0c\u5176\u4e2d\u63a8\u7406\u9636\u6bb5\u7684\u6709\u6548\u6fc0\u6d3b\u53c2\u6570\u91cf\u4e3a 8.7B\u3002\u6574\u4f53\u67b6\u6784\u7531\u4e00\u4e2a 3.5B \u53c2\u6570\u89c4\u6a21\u7684\u97f3\u9891\u7f16\u7801\u5668\u548c\u4e00\u4e2a LLM \u89e3\u7801\u5668\u7ec4\u6210\u3002\u7f16\u7801\u5668\u91c7\u7528 CNN \u524d\u7aef\u5e76\u63a5\u5165 20 \u5c42 Conformer-MoE \u7ed3\u6784\u3002CNN \u524d\u7aef\u9996\u5148\u5bf9\u8f93\u5165\u7279\u5f81\u8fdb\u884c 4\u00d7 \u7684\u65f6\u95f4\u7ef4\u4e0b\u91c7\u6837\uff0c\u968f\u540e\u5c06\u5f97\u5230\u7684\u7279\u5f81\u9001\u5165 Conformer-MoE \u5806\u53e0\u6a21\u5757\u3002\u6bcf\u4e00\u5c42 MoE \u91c7\u7528 3-of-8 \u7684\u4e13\u5bb6\u8def\u7531\u7b56\u7565\uff0c\u9690\u85cf\u5c42\u7ef4\u5ea6\u4e3a 3584\u3002<\/p>\n\n\n\n<p>\u7f16\u7801\u5668\u7684\u8f93\u51fa\u8fdb\u4e00\u6b65\u7ecf\u8fc7\u4e00\u6b21 2\u00d7 \u7684\u5e27\u7ea7\u4e0b\u91c7\u6837\u4e0e\u7279\u5f81\u62fc\u63a5\u64cd\u4f5c\uff0c\u968f\u540e\u901a\u8fc7\u4e00\u4e2a\u4e24\u5c42\u7ebf\u6027\u9002\u914d\u5668\uff08adapter\uff09\uff0c\u6700\u7ec8\u4f5c\u4e3a\u8f93\u5165\u9001\u5165 LLM \u89e3\u7801\u5668\u3002\u5728\u57fa\u7840\u8bad\u7ec3\u9636\u6bb5\uff0c\u4ec5\u5bf9 LLM \u90e8\u5206\u65bd\u52a0 LoRA\u5fae\u8c03\uff0c\u5176\u4e2d LoRA \u7684\u79e9\uff08rank\uff09\u8bbe\u7f6e\u4e3a 64\uff0c\u7f29\u653e\u7cfb\u6570\uff08alpha\uff09\u4e3a 32\u3002\u5728\u4e0a\u4e0b\u6587\u504f\u7f6e\u8bad\u7ec3\u9636\u6bb5\uff0c\u5b66\u4e60\u7387\u8bbe\u4e3a 1\u00d710<sup>\u22125<\/sup>\uff0c\u5e76\u8054\u5408\u66f4\u65b0\u97f3\u9891\u7f16\u7801\u5668\u3001\u9002\u914d\u5668\u4ee5\u53ca LLM \u7684 LoRA \u53c2\u6570\u3002\u5728 GRPO \u8bad\u7ec3\u9636\u6bb5\uff0c\u7ee7\u7eed\u4f7f\u7528\u76f8\u540c\u7684\u5b66\u4e60\u7387\uff081\u00d710<sup>\u22125<\/sup>\uff09\uff0c\u4f46\u51bb\u7ed3\u7f16\u7801\u5668\u548c\u9002\u914d\u5668\uff0c\u4ec5\u66f4\u65b0 LLM \u7684 LoRA \u53c2\u6570\u3002\u6b64\u5916\uff0c\u6211\u4eec\u5c06 KL \u6563\u5ea6\u6b63\u5219\u9879\u7684\u6743\u91cd\u8bbe\u7f6e\u4e3a 0.04\uff0c\u5e76\u5728\u6bcf\u4e2a\u8bad\u7ec3\u6b65\u9aa4\u4e2d\u751f\u6210 6 \u4e2a\u5019\u9009\u54cd\u5e94\u7528\u4e8e\u7b56\u7565\u4f18\u5316\u3002<\/p>\n\n\n\n<h4>\u57fa\u4e8e\u5f3a\u5316\u5b66\u4e60\u5f15\u5bfc\u5224\u522b\u7684\u4e0a\u4e0b\u6587 ASR<\/h4>\n\n\n\n<p>\u4e3a\u6291\u5236\u7531\u4e0a\u4e0b\u6587\u504f\u7f6e\u5f15\u5165\u7684\u8bef\u68c0\uff08false positives\uff09\uff0c\u6211\u4eec\u5145\u5206\u5229\u7528 ASR \u6a21\u578b\u5728\u89e3\u7801\u9636\u6bb5\u5bf9\u504f\u7f6e\u8bcd\u8fdb\u884c\u5224\u522b\u7684\u80fd\u529b\u3002\u6a21\u578b\u91c7\u7528\u7ed3\u6784\u5316\u63d0\u793a\uff08structured prompt\uff09\u8fdb\u884c\u8bad\u7ec3\uff0c\u5f62\u5f0f\u5982\u4e0b\uff1a<\/p>\n\n\n\n<p>\u201c&lt;Audio&gt; \u8bf7\u5c06\u97f3\u9891\u8f6c\u5199\u4e3a\u6587\u672c\u3002\u53ef\u4f7f\u7528\u7684\u504f\u7f6e\u8bcd\u5305\u62ec\uff1a&lt;g\u2081&gt; &lt;g\u2082&gt; \u2026 &lt;g\u2096&gt;\u201d\u3002<\/p>\n\n\n\n<p>\u5176\u4e2d\uff0c<strong>\u63d0\u4f9b\u7684\u504f\u7f6e\u8bcd\u5217\u8868\u523b\u610f\u5305\u542b\u4e0e\u5f53\u524d\u8bed\u97f3\u65e0\u5173\u7684\u8bcd\u9879\u6216\u5e72\u6270\u8bcd\uff0c\u4ee5\u907f\u514d\u6a21\u578b\u8fc7\u5ea6\u4f9d\u8d56\u504f\u7f6e\u8bcd\u5e76\u63d0\u5347\u5176\u5224\u522b\u80fd\u529b\u3002<\/strong><\/p>\n\n\n\n<p>\u9664\u4e0a\u8ff0\u6570\u636e\u5c42\u9762\u7684\u589e\u5f3a\u7b56\u7565\u5916\uff0c\u6211\u4eec\u5728 LLM-ASR \u8bad\u7ec3\u8fc7\u7a0b\u4e2d\u8fdb\u4e00\u6b65\u5f15\u5165\u751f\u6210\u5f0f\u62d2\u7edd\u5f0f\u7b56\u7565\u4f18\u5316\uff08Generative Rejection-Based Policy Optimization\uff0cGRPO\uff09 \u8fd9\u4e00\u5f3a\u5316\u5b66\u4e60\u65b9\u6cd5\uff0c\u4ee5\u589e\u5f3a\u6a21\u578b\u5bf9\u504f\u7f6e\u8bcd\u7684\u533a\u5206\u80fd\u529b\u3002\u6240\u8bbe\u8ba1\u7684\u5956\u52b1\u51fd\u6570\u8054\u5408\u4f18\u5316\u591a\u4e2a\u76ee\u6807\uff0c\u5177\u4f53\u5305\u62ec\uff1a<\/p>\n\n\n\n<ul><li><strong>\u5339\u914d\u5956\u52b1\uff08match reward\uff09<\/strong>\uff1a\u82e5\u67d0\u5019\u9009\u504f\u7f6e\u8bcd\u540c\u65f6\u51fa\u73b0\u5728\u6a21\u578b\u8f93\u51fa\u4e0e\u53c2\u8003\u6807\u6ce8\u4e2d\uff0c\u6216\u540c\u65f6\u672a\u51fa\u73b0\u5728\u4e8c\u8005\u4e2d\uff0c\u5219\u5956\u52b1\u503c\u4e3a 1\uff1b\u5426\u5219\u5956\u52b1\u503c\u4e3a 0\uff1b<\/li><li><strong>\u57fa\u4e8e WER \u7684\u5956\u52b1\uff08WER-based reward\uff09<\/strong>\uff1a\u5956\u52b1\u5b9a\u4e49\u4e3a 1\u2212<em>WER<\/em>\uff0c\u4ee5\u4fdd\u8bc1\u6574\u4f53\u8f6c\u5199\u51c6\u786e\u7387\u3002<\/li><\/ul>\n\n\n\n<p>\u5728\u63a8\u7406\u9636\u6bb5\uff0c\u4e3a\u8fdb\u4e00\u6b65\u63d0\u5347\u6027\u80fd\uff0c\u6211\u4eec<strong>\u91c7\u7528\u8054\u5408\u675f\u641c\u7d22\uff08joint beam search\uff09\u7b56\u7565\uff0c\u540c\u65f6\u89e3\u7801\u65e0\u4e0a\u4e0b\u6587\u7ea6\u675f\uff08context-free\uff09\u4e0e\u4e0a\u4e0b\u6587\u6761\u4ef6\u5316\uff08context-conditioned\uff09\u7684\u5019\u9009\u5047\u8bbe<\/strong>\u3002\u5728\u4fdd\u7559\u68c0\u7d22\u589e\u5f3a\u751f\u6210\uff08Retrieval-Augmented Generation\uff0cRAG\uff09\u4f18\u52bf\u7684\u540c\u65f6\uff0c\u8be5\u7b56\u7565<strong>\u6709\u6548\u964d\u4f4e\u4e86\u7531\u65e0\u5173\u504f\u7f6e\u8bcd\u5f15\u53d1\u7684\u5e7b\u89c9\u95ee\u9898<\/strong>\u3002<\/p>\n\n\n\n<h3>\u6570\u636e\u96c6\u548c\u6548\u679c<\/h3>\n\n\n\n<h4>Context-Biasing Training Datasets<\/h4>\n\n\n\n<p>LLM-ASR \u7cfb\u7edf\u5728\u603b\u8ba1\u6570\u767e\u4e07\u5c0f\u65f6\u7684\u8bed\u97f3\u6570\u636e\u4e0a\u8fdb\u884c\u8bad\u7ec3\u3002\u672c\u6587\u91cd\u70b9\u5173\u6ce8\u4e0a\u4e0b\u6587\u504f\u7f6e\u76f8\u5173\u7684\u6570\u636e\u8bbe\u7f6e\u3002\u5728\u5b8c\u6210\u57fa\u7840 LLM-ASR \u8bad\u7ec3\u4e4b\u540e\uff0c\u8fdb\u4e00\u6b65\u4f7f\u7528<strong>\u7ea6 200 \u4e07\u6761\u4e0e\u70ed\u8bcd\u548c\/\u6216\u4e0a\u4e0b\u6587\u5386\u53f2\u76f8\u5173\u7684\u8bed\u53e5\u5bf9\u6a21\u578b\u8fdb\u884c\u5fae\u8c03<\/strong>\uff0c\u8fd9\u4e9b\u8bed\u53e5\u4e3b\u8981\u901a\u8fc7 RADA \u6d41\u7a0b\u751f\u6210\u3002\u5728\u8bad\u7ec3\u6570\u636e\u6784\u6210\u4e0a\uff0c<strong>\u5305\u542b\u70ed\u8bcd\/\u4e0a\u4e0b\u6587\u7684\u8bed\u53e5\u4e0e\u4e0d\u5305\u542b\u70ed\u8bcd\/\u4e0a\u4e0b\u6587\u7684\u8bed\u53e5\u6309 1:8 \u7684\u6bd4\u4f8b\u8fdb\u884c\u6df7\u5408<\/strong>\uff0c\u5176\u4e2d<strong>\u975e\u504f\u7f6e\u6570\u636e\u5360\u4e3b\u5bfc\uff0c\u4ee5\u907f\u514d\u6a21\u578b\u5bf9\u4e0a\u4e0b\u6587\u504f\u7f6e\u7684\u8fc7\u5ea6\u4f9d\u8d56<\/strong>\u3002\u5bf9\u4e8e\u5305\u542b\u70ed\u8bcd\u7684\u8bed\u53e5\uff0c<strong>\u6bcf\u6761\u8bed\u53e5\u5305\u542b 1\u201310 \u4e2a\u70ed\u8bcd\uff0c\u5176\u4e2d\u7ea6\u4e00\u534a\u8bed\u53e5\u5305\u542b\u6b63\u786e\u7684\u76ee\u6807\u70ed\u8bcd\uff0c\u53e6\u4e00\u534a\u4e0d\u5305\u542b\u76ee\u6807\u70ed\u8bcd\uff0c\u4ece\u800c\u5728\u6b63\u3001\u8d1f\u70ed\u8bcd\u6837\u672c\u4e4b\u95f4\u5f62\u6210 1:1 \u7684\u6bd4\u4f8b\u5e73\u8861\u3002<\/strong><\/p>\n\n\n\n<h4>GLCLAP\u5fae\u8c03\u6570\u636e<\/h4>\n\n\n\n<p>\u5728\u5927\u89c4\u6a21\u901a\u7528 ASR \u6570\u636e\u4e0a\u5bf9\u6a21\u578b\u8fdb\u884c\u8bad\u7ec3\uff0c\u5176\u4e2d\u6587\u672c\u6807\u6ce8\u901a\u8fc7\u4ece\u5b8c\u6574\u8f6c\u5199\u4e2d\u968f\u673a\u88c1\u526a\u77ed\u8bed\u7684\u65b9\u5f0f\u83b7\u5f97\u3002\u968f\u540e\uff0c<strong>\u5728\u7b2c\u4e8c\u9636\u6bb5\u5bf9 GLCLAP \u68c0\u7d22\u5668\u8fdb\u884c\u5fae\u8c03<\/strong>\u65f6\uff0c\u6211\u4eec\u6784\u5efa\u4e86\u4e00\u4e2a<strong>\u9762\u5411\u7279\u5b9a\u9886\u57df\u7684\u97f3\u9891\u2013\u6587\u672c\u6570\u636e\u96c6\uff0c\u89c4\u6a21\u7ea6\u4e3a 25 \u4e07\u5bf9\uff0c\u4ee5\u66f4\u597d\u5730\u4f7f\u68c0\u7d22\u7ed3\u679c\u4e0e\u504f\u7f6e\u8bcd\u6240\u5c5e\u9886\u57df\u5bf9\u9f50<\/strong>\u3002<\/p>\n\n\n\n<h4>\u70ed\u8bcd\u8bcd\u8868\uff08Hotword Vocabulary\uff09<\/h4>\n\n\n\n<p>\u70ed\u8bcd\u8bcd\u8868\u901a\u8fc7\u7f51\u7edc\u6570\u636e\u6784\u5efa\uff0c\u4e3b\u8981\u8986\u76d6\u533b\u7597\u548c\u5a92\u4f53\uff08\u5f71\u89c6\uff09\u4e24\u4e2a\u9886\u57df\u3002\u5728\u5f97\u5230\u521d\u59cb\u8bcd\u8868\u540e\uff0c\u8fdb\u4e00\u6b65\u91c7\u7528 RADA \u7b56\u7565\u8fdb\u884c\u8fc7\u6ee4\u3002\u7ecf\u8fc7\u57fa\u4e8e RADA \u7684\u7b5b\u9009\uff0c\u70ed\u8bcd\u8bcd\u8868\u89c4\u6a21\u7531\u7ea6 60 \u4e07\u6761\u7f29\u51cf\u81f3 9.8 \u4e07\u6761\uff0c\u6709\u6548\u964d\u4f4e\u4e86\u8bcd\u8868\u89c4\u6a21\u5e76\u63d0\u5347\u4e86\u53ef\u7528\u6027\u3002<\/p>\n\n\n\n<h4>\u8bc4\u6d4b\u8bbe\u7f6e\uff08Evaluation\uff09<\/h4>\n\n\n\n<p>\u6784\u5efa\u4e86\u4e24\u4e2a\u9762\u5411\u7279\u5b9a\u9886\u57df\u7684\u6d4b\u8bd5\u96c6\uff1a<strong>Media<\/strong> \u548c <strong>Medical<\/strong>\uff0c\u6bcf\u4e2a\u6d4b\u8bd5\u96c6\u5305\u542b 240 \u6761\u8bed\u53e5\uff0c\u4e3b\u8981\u6765\u6e90\u4e8e\u5b9e\u9645\u7cfb\u7edf\u4e2d\u7684\u9519\u8bef\u6848\u4f8b\uff08bad cases\uff09\u3002\u6bcf\u6761\u8bed\u53e5\u5747\u7531\u4eba\u5de5\u6807\u6ce8\u5176\u771f\u5b9e\u504f\u7f6e\u8bcd\uff0c\u5e76\u5c06\u8fd9\u4e9b\u504f\u7f6e\u8bcd\u52a0\u5165\u504f\u7f6e\u8bcd\u5217\u8868\u4e2d\u3002\u6b64\u5916\uff0c\u6211\u4eec\u8fd8\u6784\u5efa\u4e86\u4e00\u4e2a\u56de\u5f52\u6d4b\u8bd5\u96c6 <strong>General Task<\/strong>\uff0c\u5305\u542b\u7ea6 5,000 \u6761\u6807\u51c6 ASR \u8bed\u53e5\uff0c\u7528\u4e8e\u8bc4\u4f30\u901a\u7528\u8bc6\u522b\u6027\u80fd\u3002<\/p>\n\n\n\n<p>ASR \u8bc4\u6d4b\u4e2d\uff0c\u6211\u4eec\u91c7\u7528\u4e24\u9879\u6307\u6807\uff1a\uff08i\uff09\u53e5\u7ea7\u8bc6\u522b\u51c6\u786e\u7387\uff08Sentence-level Accuracy\uff0cSACC\uff09\uff0c\u4ee5\u53ca\uff08ii\uff09\u5173\u952e\u8bcd\u9519\u8bef\u7387\uff08Keyword Error Rate\uff0cKER\uff09\u3002<\/p>\n\n\n\n<h4>\u7ed3\u679c<\/h4>\n\n\n\n<p>\u5728 <strong>Media<\/strong> \u548c <strong>Medical<\/strong> \u6d4b\u8bd5\u96c6\u4e0a\u8bc4\u4f30\u57fa\u4e8e GLCLAP \u7684\u70ed\u8bcd\u68c0\u7d22\u6027\u80fd\u3002\u5982\u8868 1 \u548c\u8868 2 \u6240\u793a\uff0c\u9c81\u68d2\u6027\u611f\u77e5\u6570\u636e\u589e\u5f3a\uff08RADA\uff09\u4e0e\u6a21\u7cca\u5339\u914d\u7b56\u7565\u5747\u5bf9\u6574\u4f53\u6027\u80fd\u4ea7\u751f\u4e86\u6b63\u5411\u8d21\u732e\u3002\u5177\u4f53\u800c\u8a00\uff0c\u539f\u59cb\u70ed\u8bcd\u8bcd\u8868\u7ea6\u5305\u542b 60 \u4e07\u6761\u8bcd\u9879\uff0c\u5728\u5e94\u7528 RADA \u7b5b\u9009\u540e\u7f29\u51cf\u81f3 9.8 \u4e07\u6761\u3002\u6a21\u7cca\u5339\u914d\u4e0d\u4ec5\u66f4\u5951\u5408\u6211\u4eec\u7684\u8bc4\u6d4b\u6307\u6807\uff0c\u540c\u65f6\u4e5f\u66f4\u771f\u5b9e\u5730\u53cd\u6620\u4e86\u5b9e\u9645\u5e94\u7528\u573a\u666f\u4e2d\u7684\u504f\u7f6e\u8bcd\u4f7f\u7528\u60c5\u51b5\u3002\u6b64\u5916\uff0c\u7ed3\u679c\u663e\u793a\uff0c\u968f\u7740 top-kkk \u503c\u7684\u589e\u5927\uff0c\u53ec\u56de\u7387\u5448\u6301\u7eed\u4e0a\u5347\u8d8b\u52bf\u3002<\/p>\n\n\n\n<figure class=\"wp-block-image size-full\"><img loading=\"lazy\" width=\"575\" height=\"216\" src=\"http:\/\/139.9.1.231\/wp-content\/uploads\/2026\/01\/image-24.png\" alt=\"\" class=\"wp-image-30050\" srcset=\"http:\/\/139.9.1.231\/wp-content\/uploads\/2026\/01\/image-24.png 575w, http:\/\/139.9.1.231\/wp-content\/uploads\/2026\/01\/image-24-300x113.png 300w\" sizes=\"(max-width: 575px) 100vw, 575px\" \/><\/figure>\n\n\n\n<figure class=\"wp-block-image size-full\"><img loading=\"lazy\" width=\"551\" height=\"187\" src=\"http:\/\/139.9.1.231\/wp-content\/uploads\/2026\/01\/image-26.png\" alt=\"\" class=\"wp-image-30059\" srcset=\"http:\/\/139.9.1.231\/wp-content\/uploads\/2026\/01\/image-26.png 551w, http:\/\/139.9.1.231\/wp-content\/uploads\/2026\/01\/image-26-300x102.png 300w\" sizes=\"(max-width: 551px) 100vw, 551px\" \/><\/figure>\n\n\n\n<p>\u5c06 GLCLAP \u4e0e LLM-ASR \u6a21\u578b\u7ed3\u5408\uff0c\u5e76\u4ee5\u53e5\u7ea7\u8bc6\u522b\u51c6\u786e\u7387\uff08SACC\uff09\u548c\u5173\u952e\u8bcd\u9519\u8bef\u7387\uff08KER\uff09\u62a5\u544a\u6700\u7ec8\u8bc6\u522b\u6027\u80fd\u3002<strong>Base<\/strong> \u5217\u663e\u793a\u4e86\u5728\u4e0d\u4f7f\u7528\u4efb\u4f55\u504f\u7f6e\u63d0\u793a\uff08bias prompt\uff09\u7684\u60c5\u51b5\u4e0b\uff0c\u7ecf\u8fc7\u4e0a\u4e0b\u6587\u611f\u77e5\u5fae\u8c03\u7684 LLM-ASR \u6a21\u578b\u7684\u7ed3\u679c\u3002<\/p>\n\n\n\n<p><strong>\u53ec\u56de\u7387\u968f\u7740 top-k \u7684\u589e\u52a0\u6301\u7eed\u4e0a\u5347<\/strong>\uff0c\u4f46\u5728<strong>\u70ed\u8bcd\u6d4b\u8bd5\u96c6\u4e0a\u7684 KER \u548c SACC \u5e76\u672a\u5448\u5355\u8c03\u6539\u5584<\/strong>\u3002\u8fd9\u662f\u56e0\u4e3a<strong>\u8f83\u5927\u7684 top-k \u4f1a\u5411 LLM-ASR \u6a21\u578b\u5f15\u5165\u66f4\u591a\u5e72\u6270\u5019\u9009\u8bcd\uff0c\u4ece\u800c\u589e\u52a0\u8bc6\u522b\u5e72\u6270<\/strong>\u3002<\/p>\n\n\n\n<p>\u5728 <strong>General Task<\/strong> \u6d4b\u8bd5\u96c6\u4e0a\uff0c\u5927\u591a\u6570\u7ed3\u679c\u7565\u4f4e\u4e8e\u672a\u4f7f\u7528\u70ed\u8bcd\u7684\u57fa\u7ebf\u8868\u73b0\u3002\u7efc\u5408\u4e24\u4e2a\u70ed\u8bcd\u6d4b\u8bd5\u96c6\u7684\u7ed3\u679c\uff0c\u6211\u4eec\u8ba4\u4e3a <strong>top-2<\/strong> \u662f\u66f4\u4e3a\u5408\u9002\u7684\u9009\u62e9\u3002<\/p>\n\n\n\n<figure class=\"wp-block-image size-full\"><img loading=\"lazy\" width=\"616\" height=\"637\" src=\"http:\/\/139.9.1.231\/wp-content\/uploads\/2026\/01\/image-25.png\" alt=\"\" class=\"wp-image-30058\" srcset=\"http:\/\/139.9.1.231\/wp-content\/uploads\/2026\/01\/image-25.png 616w, http:\/\/139.9.1.231\/wp-content\/uploads\/2026\/01\/image-25-290x300.png 290w\" sizes=\"(max-width: 616px) 100vw, 616px\" \/><\/figure>\n\n\n\n<p>GRPO \u8bad\u7ec3\u7684\u7ed3\u679c\u8868\u00a04\uff0c\u5f15\u5165 GRPO \u53ef\u4ee5\u5728\u5a92\u4f53\u548c\u533b\u7597\u8bbe\u5907\u4e0a\u7684 KER \u4e2d\u4ea7\u751f\u660e\u663e\u7684\u6027\u80fd\u63d0\u5347\u3002\u00a0\u6b64\u5916\uff0c\u5f97\u76ca\u4e8eGRPO\u4e2d\u4f7f\u7528\u7684\u57fa\u4e8e\u51c6\u786e\u6027\u7684\u5956\u52b1\uff0c\u901a\u7528\u4efb\u52a1\u7684\u53e5\u5b50\u51c6\u786e\u6027\u4e5f\u5f97\u5230\u4e86\u663e\u7740\u63d0\u9ad8\u3002<\/p>\n\n\n\n<h3><strong>\u5de5\u7a0b\u4e0e\u843d\u5730\u89c6\u89d2\u7684\u4e00\u4e9b\u542f\u53d1<\/strong><strong><\/strong><\/h3>\n\n\n\n<p>\u4ece\u5de5\u7a0b\u843d\u5730\u7684\u89d2\u5ea6\uff0c\u8fd9\u7bc7\u5de5\u4f5c\u6709\u51e0\u4e2a\u7279\u522b\u503c\u5f97\u5b9e\u8df5\u53c2\u8003\u7684\u70b9\uff1a<\/p>\n\n\n\n<ol><li>\u4e0d\u8981\u76f2\u4fe1\u201c\u5168\u91cf\u70ed\u8bcd\u8868\u201d\uff1a<ul><li>\u7528\u00a0RADA\u00a0\u5148\u7b5b\u4e00\u904d\u201c\u6a21\u578b\u5df2\u719f\u7ec3\u638c\u63e1\u201d\u7684\u8bcd\uff0c\u518d\u505a\u504f\u7f6e\uff0c\u80fd\u5927\u5e45\u964d\u4f4e\u590d\u6742\u5ea6\u4e0e\u5e72\u6270\uff1b<\/li><\/ul><\/li><li>\u68c0\u7d22\u4e0e\u89e3\u7801\u8981\u534f\u540c\u8bbe\u8ba1\uff1a<ul><li>\u4ec5\u6709\u9ad8\u53ec\u56de\u7684\u68c0\u7d22\u8fd8\u4e0d\u591f\uff0c\u8981\u4e0e\u00a0LLM\u00a0\u7684\u4f7f\u7528\u7b56\u7565\u5171\u8bbe\u8ba1\uff0c\u5426\u5219\u5bb9\u6613\u201c\u7ed9\u4e86\u6b66\u5668\u4f46\u4e0d\u4f1a\u7528\u201d\uff1b<\/li><\/ul><\/li><li>RL\u00a0\u5bf9\u00a0ASR\u00a0\u4e5f\u975e\u5e38\u6709\u7528\uff1a<ul><li>\u4f20\u7edf\u00a0ASR\u00a0\u591a\u7528\u00a0CE\/CTC\/Transducer\u00a0\u7b49\u635f\u5931\uff0c\u5f88\u96be\u76f4\u63a5\u5bf9\u63a5\u4efb\u52a1\u7ea7\u6307\u6807\uff1b<\/li><\/ul><ul><li>\u5f15\u5165\u7c7b\u4f3c\u00a0GRPO\u00a0\u8fd9\u6837\u7684\u00a0RL\u00a0\u65b9\u6cd5\uff0c\u53ef\u4ee5\u5728\u201c\u70ed\u8bcd\u8bc6\u522b\u00a0+\u00a0\u53e5\u5b50\u51c6\u786e\u201d\u8fd9\u6837\u7684\u7ec4\u5408\u76ee\u6807\u4e0a\u505a\u66f4\u76f4\u63a5\u7684\u4f18\u5316\uff1b<\/li><\/ul><\/li><li><strong>top-k\u00a0\u662f\u4e2a\u5173\u952e\u7684\u201c\u5de5\u7a0b\u8d85\u53c2\u201d\uff1a<\/strong><ul><li><strong>k\u00a0\u592a\u5c0f\uff0c\u53ec\u56de\u4e0d\u8db3\uff1b<\/strong><\/li><\/ul><ul><li><strong>k\u00a0\u592a\u5927\uff0c\u5e72\u6270\u8fc7\u591a\uff1b<\/strong><\/li><\/ul><ul><li>\u6700\u4f18\u70b9\u4f9d\u8d56\u4e1a\u52a1\u573a\u666f\u3001\u8bcd\u8868\u8d28\u91cf\u4e0e\u00a0LLM\u00a0\u5bb9\u91cf\uff0c\u9700\u8981\u901a\u8fc7\u7cfb\u7edf\u6027\u5b9e\u9a8c\u6765\u9009\u3002<\/li><\/ul><\/li><\/ol>\n","protected":false},"excerpt":{"rendered":"<p>\u8bba\u6587\uff1aContextual Biasing for LLM-based ASR with Hotword Re &hellip; <a href=\"http:\/\/139.9.1.231\/index.php\/2026\/01\/10\/llm-based-asr-with-hotword-retrieval\/\" class=\"more-link\">\u7ee7\u7eed\u9605\u8bfb<span class=\"screen-reader-text\">\u57fa\u4e8e\u5927\u8bed\u8a00\u6a21\u578b\u7684\u8bed\u97f3\u8bc6\u522b\u4e0a\u4e0b\u6587\u504f\u7f6e\uff1a\u70ed\u8bcd\u68c0\u7d22\u4e0e\u5f3a\u5316\u5b66\u4e60\u65b9\u6cd5<\/span><\/a><\/p>\n","protected":false},"author":1,"featured_media":0,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":[],"categories":[42,4,9,38,34],"tags":[],"_links":{"self":[{"href":"http:\/\/139.9.1.231\/index.php\/wp-json\/wp\/v2\/posts\/29962"}],"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=29962"}],"version-history":[{"count":85,"href":"http:\/\/139.9.1.231\/index.php\/wp-json\/wp\/v2\/posts\/29962\/revisions"}],"predecessor-version":[{"id":30083,"href":"http:\/\/139.9.1.231\/index.php\/wp-json\/wp\/v2\/posts\/29962\/revisions\/30083"}],"wp:attachment":[{"href":"http:\/\/139.9.1.231\/index.php\/wp-json\/wp\/v2\/media?parent=29962"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"http:\/\/139.9.1.231\/index.php\/wp-json\/wp\/v2\/categories?post=29962"},{"taxonomy":"post_tag","embeddable":true,"href":"http:\/\/139.9.1.231\/index.php\/wp-json\/wp\/v2\/tags?post=29962"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}