{"id":31180,"date":"2026-06-02T17:03:02","date_gmt":"2026-06-02T09:03:02","guid":{"rendered":"http:\/\/139.9.1.231\/?p=31180"},"modified":"2026-06-02T17:03:04","modified_gmt":"2026-06-02T09:03:04","slug":"asr-roadmap","status":"publish","type":"post","link":"http:\/\/139.9.1.231\/index.php\/2026\/06\/02\/asr-roadmap\/","title":{"rendered":"ASR\u5927\u6a21\u578b\u53d1\u5c55\u8def\u7ebf"},"content":{"rendered":"\n<ul><li><strong>\u6587\u7ae0\u6765\u6e90\uff1a<a rel=\"noreferrer noopener\" href=\"https:\/\/mp.weixin.qq.com\/s\/Ph13am69CX8VxSSaIjJb0A\" target=\"_blank\">\u516c\u4f17\u53f7\u6bcf\u5468\u4e00\u4e2a\u5927\u6a21\u578b\u5e94\u7528<\/a><\/strong><\/li><li><a href=\"https:\/\/arxiv.org\/html\/2604.00610v1\" target=\"_blank\" rel=\"noreferrer noopener\">Speech LLMs are Contextual Reasoning Transcribers<\/a><\/li><li><a href=\"https:\/\/openreview.net\/pdf?id=JCujsFnDS7\" target=\"_blank\" rel=\"noreferrer noopener\">Whisfusion: Parallel ASR Decoding via a Diffusion Transformer<\/a><\/li><\/ul>\n\n\n\n<figure class=\"wp-block-image size-large\"><img loading=\"lazy\" width=\"1024\" height=\"683\" src=\"http:\/\/139.9.1.231\/wp-content\/uploads\/2026\/06\/11111-1024x683.webp\" alt=\"\" class=\"wp-image-31181\" srcset=\"http:\/\/139.9.1.231\/wp-content\/uploads\/2026\/06\/11111-1024x683.webp 1024w, http:\/\/139.9.1.231\/wp-content\/uploads\/2026\/06\/11111-300x200.webp 300w, http:\/\/139.9.1.231\/wp-content\/uploads\/2026\/06\/11111-768x512.webp 768w, http:\/\/139.9.1.231\/wp-content\/uploads\/2026\/06\/11111.webp 1080w\" sizes=\"(max-width: 1024px) 100vw, 1024px\" \/><figcaption>\u51c6\u786e\u7387\u4e0e\u901f\u5ea6\uff0cASR \u7684\u4e24\u6761\u8fdb\u5316\u8def\u7ebf<\/figcaption><\/figure>\n\n\n\n<p class=\"has-background\" style=\"background-color:#d4ebeb\">Whisper \u628a\u5f00\u6e90 ASR \u62c9\u5230\u4e86\u65b0\u9ad8\u5ea6\uff0c\u4f46 2025\u20132026 \u5e74\u7684\u9876\u4f1a\u8bba\u6587\u544a\u8bc9\u6211\u4eec\uff1a\u74f6\u9888\u5df2\u7ecf\u4e0d\u5728\u300c\u542c\u4e0d\u542c\u5f97\u6e05\u300d\uff0c\u800c\u5728\u300c\u600e\u4e48\u751f\u6210\u6587\u672c\u300d\u3002<strong>Microsoft \u7684 CoT-ASR \u8ba9\u5927\u6a21\u578b\u5148\u300c\u60f3\u300d\u518d\u300c\u5199\u300d\uff0cWhisfusion \u5219\u7528\u6269\u6563\u6a21\u578b\u5e76\u884c\u89e3\u7801\uff0c\u628a Whisper \u7684\u5ef6\u8fdf\u780d\u5230\u539f\u6765\u7684\u516b\u5206\u4e4b\u4e00\u3002<\/strong>\u672c\u6587\u6df1\u5ea6\u62c6\u89e3\u4e24\u7bc7\u4ee3\u8868\u8bba\u6587\uff0c\u5e2e\u4f60\u770b\u61c2 ASR \u8303\u5f0f\u8fc1\u79fb\u7684\u6765\u9f99\u53bb\u8109\u3002<\/p>\n\n\n\n<p class=\"has-background\" style=\"background-color:#e7d7ae\">\u7ed3\u8bba\uff1aLLM \u63a5\u5165 ASR \u540e\uff0c\u300c\u76f4\u63a5\u8f6c\u5199\u300d\u5e76\u6ca1\u6709\u5145\u5206\u91ca\u653e\u5927\u6a21\u578b\u80fd\u529b\u2014\u2014CoT-ASR \u7528\u94fe\u5f0f\u63a8\u7406\u628a WER \u964d 8.7%\u3001\u5b9e\u4f53\u9519\u8bef\u7387 EER \u964d 16.9%\uff1bWhisfusion \u7528\u975e\u81ea\u56de\u5f52\u6269\u6563\u89e3\u7801\uff0c\u76f8\u8fd1\u7cbe\u5ea6\u4e0b\u628a 20\u201330 \u79d2\u97f3\u9891\u7684\u89e3\u7801\u65f6\u95f4\u4ece 674.7ms \u538b\u5230 80.7ms\u3002<strong>\u4e00\u6761\u8def\u7ebf\u4f18\u5316\u300c\u51c6\u300d\uff0c\u4e00\u6761\u8def\u7ebf\u4f18\u5316\u300c\u5feb\u300d\uff0c\u5171\u540c\u6307\u5411\u65b0\u4e00\u4ee3 ASR \u67b6\u6784<\/strong>\u3002<\/p>\n\n\n\n<h2><strong>\u4e00\u3001\u524d\u8a00\uff1aASR \u4e3a\u4ec0\u4e48\u9700\u8981\u6362\u8303\u5f0f<\/strong><\/h2>\n\n\n\n<p>\u8fc7\u53bb\u5341\u5e74\uff0cASR \u7684\u4e3b\u7ebf\u6545\u4e8b\u662f\u300c\u66f4\u5927\u7684\u7f16\u7801\u5668 + \u66f4\u597d\u7684\u5bf9\u9f50\u300d\u3002Conformer\u3001Whisper\u3001SenseVoice\u2026\u2026\u51c6\u786e\u7387\u4e00\u8def\u6500\u5347\u3002\u4f46\u5f53 Speech LLM \u628a LLM \u63a5\u8fdb\u8bc6\u522b\u94fe\u8def\u540e\uff0c\u4e00\u4e2a\u5c34\u5c2c\u7684\u4e8b\u5b9e\u6d6e\u51fa\u6c34\u9762\uff1a<strong>\u5927\u6a21\u578b\u5728\u6587\u672c\u4fa7\u62e5\u6709\u7684\u63a8\u7406\u3001\u77e5\u8bc6\u3001\u4e0a\u4e0b\u6587\u7406\u89e3\u80fd\u529b\uff0c\u5728 ASR \u91cc\u51e0\u4e4e\u7528\u4e0d\u4e0a\u3002<\/strong><\/p>\n\n\n\n<p>\u539f\u56e0\u5f88\u7b80\u5355\u2014\u2014\u4f20\u7edf <strong>LLM-based ASR \u7684\u8bad\u7ec3\u76ee\u6807<\/strong>\u4ecd\u7136\u662f<strong>\u300c\u8bed\u97f3 \u2192 \u9010\u5b57\u8f6c\u5199\u300d<\/strong>\u3002\u8bed\u97f3\u548c\u6587\u672c\u627f\u8f7d\u7684\u4fe1\u606f\u9ad8\u5ea6\u91cd\u53e0\uff0c<strong>\u6a21\u578b\u88ab\u7ea6\u675f\u6210\u300c\u590d\u8bfb\u673a\u300d\uff0c\u800c\u4e0d\u662f\u300c\u7406\u89e3\u8005\u300d<\/strong>\u3002\u4e0e\u6b64\u540c\u65f6\uff0cWhisper \u5f0f<strong>\u81ea\u56de\u5f52\u89e3\u7801\u5668\u5fc5\u987b\u9010 token \u751f\u6210\uff0c<\/strong>\u6587\u672c\u8d8a\u957f\uff0c\u5ef6\u8fdf\u7ebf\u6027\u589e\u957f\uff0c\u5b9e\u65f6\u5b57\u5e55\u3001\u4f1a\u8bae\u8f6c\u5199\u3001\u7aef\u4fa7 ASR \u90fd\u6df1\u53d7\u5176\u82e6\u3002<\/p>\n\n\n\n<figure class=\"wp-block-image size-large\"><img loading=\"lazy\" width=\"1024\" height=\"683\" src=\"http:\/\/139.9.1.231\/wp-content\/uploads\/2026\/06\/22222-1024x683.webp\" alt=\"\" class=\"wp-image-31191\" srcset=\"http:\/\/139.9.1.231\/wp-content\/uploads\/2026\/06\/22222-1024x683.webp 1024w, http:\/\/139.9.1.231\/wp-content\/uploads\/2026\/06\/22222-300x200.webp 300w, http:\/\/139.9.1.231\/wp-content\/uploads\/2026\/06\/22222-768x512.webp 768w, http:\/\/139.9.1.231\/wp-content\/uploads\/2026\/06\/22222.webp 1080w\" sizes=\"(max-width: 1024px) 100vw, 1024px\" \/><\/figure>\n\n\n\n<p><strong>&nbsp;<\/strong><strong>2026 \u5e74\u7684\u4e24\u4e2a\u4fe1\u53f7<\/strong><\/p>\n\n\n\n<ul class=\"has-light-pink-background-color has-background\"><li>CoT-ASR\uff08Microsoft Core AI\uff09\uff1a\u628a Chain-of-Thought \u5f15\u5165 ASR\uff0cICLR\/arxiv 2026<\/li><li>Whisfusion\uff08ICLR 2026 \u6295\u7a3f\uff09\uff1aWhisper \u7f16\u7801\u5668 + \u6269\u6563\u5e76\u884c\u89e3\u7801<\/li><li>\u5171\u540c\u80cc\u666f\uff1aSpeech LLM \u89c4\u6a21\u5316\uff0c\u4f46 token \u5bc6\u5ea6\u5931\u8861\u4e0e AR \u5ef6\u8fdf\u6210\u4e3a\u4e24\u5927\u74f6\u9888<\/li><\/ul>\n\n\n\n<p><strong>\u300c<\/strong><em>\u8bba\u6587\u6570\u636e\u4ec5\u4f9b\u53c2\u8003\uff1bCoT-ASR \u57fa\u4e8e 3.8B Phi-4-mini + 38k \u5c0f\u65f6\u82f1\u6587\u6570\u636e\uff0cWhisfusion \u5728 LibriSpeech 960h \u4e0a\u5fae\u8c03\u3002\u843d\u5730\u65f6\u9700\u7ed3\u5408\u81ea\u5df1\u7684\u8bed\u79cd\u3001\u573a\u666f\u4e0e\u7b97\u529b\u91cd\u65b0\u8bc4\u4f30\u3002<\/em><strong>\u300d<\/strong><\/p>\n\n\n\n<h2><strong>\u4e8c\u3001CoT-ASR\uff1a\u8ba9\u5927\u6a21\u578b\u5148\u5206\u6790\uff0c\u518d\u8f6c\u5199<\/strong><\/h2>\n\n\n\n<p>\u8bba\u6587\u5168\u79f0 Speech LLMs are Contextual Reasoning Transcribers\uff0c\u4f5c\u8005\u6765\u81ea Microsoft Core AI\uff08Keqi Deng\u3001Jinyu Li \u7b49\uff09\u3002<\/p>\n\n\n\n<p>\u5b83\u8981\u56de\u7b54\u7684\u6838\u5fc3\u95ee\u9898\u662f\uff1a\u5982\u4f55\u628a LLM \u7684\u63a8\u7406\u80fd\u529b\u300c\u7ffb\u8bd1\u300d\u6210 ASR \u6536\u76ca\uff1f<\/p>\n\n\n\n<p><strong>\u258e&nbsp;<\/strong><strong>2.1 \u76f4\u63a5\u8f6c\u5199\u4e3a\u4f55\u6d6a\u8d39 LLM<\/strong><\/p>\n\n\n\n<p>\u73b0\u6709 Speech LLM \u901a\u5e38\u628a\u8bed\u97f3\u7f16\u7801\u5668\u8f93\u51fa\u62fc\u5728\u6587\u672c prompt \u524d\u9762\uff0c\u7136\u540e\u8ba9 LLM \u76f4\u63a5\u751f\u6210\u8f6c\u5199\u7ed3\u679c\u3002<strong>\u8bad\u7ec3 loss \u4e5f\u53ea\u76d1\u7763\u8f6c\u5199\u6587\u672c\u2014\u2014\u548c Conformer AED \u6ca1\u6709\u672c\u8d28\u533a\u522b\u3002<\/strong><\/p>\n\n\n\n<p>\u8bba\u6587\u6307\u51fa\uff0cASR \u5728\u4fe1\u606f\u8bba\u4e0a\u63a5\u8fd1\u300c\u5185\u5bb9\u4fdd\u6301\u6620\u5c04\u300d\uff1a\u8f93\u5165\u8bf4\u4ec0\u4e48\uff0c\u8f93\u51fa\u5c31\u5199\u4ec0\u4e48\uff0c\u8bed\u4e49\u53d8\u6362\u7a7a\u95f4\u6781\u5c0f\u3002LLM \u5728\u6d77\u91cf\u6587\u672c\u4e0a\u9884\u8bad\u7ec3\u83b7\u5f97\u7684\u5e38\u8bc6\u3001\u9886\u57df\u77e5\u8bc6\u3001\u6d88\u6b67\u80fd\u529b\uff0c\u5728\u300c\u53ea\u542c\u5c31\u5199\u300d\u7684\u6a21\u5f0f\u4e0b\u88ab\u4e25\u91cd\u538b\u5236\u3002<\/p>\n\n\n\n<p><strong>\u258e&nbsp;<\/strong><strong>2.2 \u94fe\u5f0f\u63a8\u7406\uff1aOne-Pass \u7684\u4e24\u6bb5\u5f0f\u8f93\u51fa<\/strong><\/p>\n\n\n\n<p>CoT-ASR \u7684\u5173\u952e\u8bbe\u8ba1\u662f\uff1a\u4e00\u6b21\u751f\u6210\uff08one-pass\uff09\uff0c\u4f46\u8f93\u51fa\u5206\u4e24\u6bb5\u3002\u6a21\u578b\u5148\u4ea7\u51fa\u300c\u8bed\u5883\u5206\u6790\u300d\uff08Contextual Analysis\uff09\uff0c\u518d\u4ea7\u51fa\u300c\u8f6c\u5199\u6587\u672c\u300d\u3002\u524d\u8005\u76f8\u5f53\u4e8e <strong>Chain-of-Thought<\/strong>\uff0c\u540e\u8005\u624d\u662f\u6700\u7ec8 ASR \u7ed3\u679c\u3002<\/p>\n\n\n\n<ul><li>\u8bed\u5883\u5206\u6790\uff1a\u63a8\u65ad\u8bf4\u8bdd\u573a\u666f\u3001\u4e3b\u9898\u3001\u53ef\u80fd\u7684\u4e13\u6709\u540d\u8bcd\u4e0e\u6b67\u4e49<\/li><li>\u8f6c\u5199\u6587\u672c\uff1a\u5728\u5206\u6790\u57fa\u7840\u4e0a\u751f\u6210\u66f4\u51c6\u786e\u7684\u8bc6\u522b\u7ed3\u679c<\/li><li>\u8bad\u7ec3\u6570\u636e\uff1a\u7528 Qwen2.5-14B \u4ece 3.8 \u4e07\u5c0f\u65f6\u8bed\u97f3\u81ea\u52a8\u6784\u9020\u300c\u5206\u6790 + \u8f6c\u5199\u300d\u5bf9<\/li><\/ul>\n\n\n\n<figure class=\"wp-block-image size-large\"><img loading=\"lazy\" width=\"1024\" height=\"683\" src=\"http:\/\/139.9.1.231\/wp-content\/uploads\/2026\/06\/3333-1024x683.webp\" alt=\"\" class=\"wp-image-31196\" srcset=\"http:\/\/139.9.1.231\/wp-content\/uploads\/2026\/06\/3333-1024x683.webp 1024w, http:\/\/139.9.1.231\/wp-content\/uploads\/2026\/06\/3333-300x200.webp 300w, http:\/\/139.9.1.231\/wp-content\/uploads\/2026\/06\/3333-768x512.webp 768w, http:\/\/139.9.1.231\/wp-content\/uploads\/2026\/06\/3333.webp 1080w\" sizes=\"(max-width: 1024px) 100vw, 1024px\" \/><\/figure>\n\n\n\n<p><strong>\u258e&nbsp;<\/strong><strong>2.3 CTC-guided Modality Adapter<\/strong><\/p>\n\n\n\n<p>\u8bed\u97f3\u5e27\u5e8f\u5217\u8fdc\u957f\u4e8e\u6587\u672c token\uff0c\u5982\u4f55\u628a Conformer \u7f16\u7801\u5668\u8f93\u51fa\u5bf9\u9f50\u5230 LLM \u9690\u7a7a\u95f4\uff0c\u662f Speech LLM \u7684\u7ecf\u5178\u96be\u9898\u3002<\/p>\n\n\n\n<p>CoT-ASR \u6ca1\u6709\u7b80\u5355\u7528\u4e24\u5c42 Linear \u6295\u5f71\uff0c\u800c\u662f\u63d0\u51fa CTC-guided Modality Adapter\u3002<\/p>\n\n\n\n<ul><li>\u6bcf\u5e27\u8ba1\u7b97 CTC blank \/ non-blank \u6982\u7387\u5206\u5e03<\/li><li>\u7528 non-blank \u5206\u5e03\u5bf9 LLM token embedding \u77e9\u9635\u505a\u52a0\u6743\u6c42\u548c\uff0c\u5f97\u5230\u5e27\u7ea7\u300c\u6587\u672c\u5316\u300d\u8868\u793a<\/li><li>\u4fdd\u7559\u5168\u90e8\u5e27\u4fe1\u606f\uff08\u542b blank \u5e27\uff09\uff0c\u907f\u514d CTC \u538b\u7f29\u4e22\u4fe1\u606f<\/li><li>\u95e8\u63a7\u6b8b\u5dee\u5206\u652f\u8fdb\u4e00\u6b65\u878d\u5408\u539f\u59cb\u58f0\u5b66\u7279\u5f81<\/li><\/ul>\n\n\n\n<p>\u76f4\u89c9\u4e0a\uff1a\u6bcf\u4e00\u5e27\u7684 CTC \u5206\u5e03\u544a\u8bc9\u6211\u4eec\u300c\u8fd9\u4e00\u5e27\u6700\u50cf\u54ea\u4e2a\u5b57\u300d\uff0c<\/p>\n\n\n\n<p>\u518d\u6620\u5c04\u5230 LLM \u5df2\u7ecf\u719f\u6089\u7684 embedding \u7a7a\u95f4\u2014\u2014\u6bd4\u7eaf\u7ebf\u6027\u6295\u5f71\u66f4\u76f4\u63a5\u5730\u5229\u7528 LLM \u7684\u6587\u672c\u5148\u9a8c\u3002<\/p>\n\n\n\n<p><strong>2.4 \u7528\u6237\u5f15\u5bfc\u8f6c\u5199\uff1a\u6bd4\u70ed\u8bcd\u66f4\u300c\u8bed\u4e49\u5316\u300d<\/strong><\/p>\n\n\n\n<p>CoT-ASR \u8fd8\u652f\u6301 User Context \u6a21\u5f0f\uff1a\u7528\u6237\u63d0\u4f9b\u573a\u666f\u63cf\u8ff0\u6216\u5b9e\u4f53\u7ebf\u7d22\uff0c<strong>\u6a21\u578b\u8df3\u8fc7\u81ea\u751f\u6210\u63a8\u7406\uff0c\u76f4\u63a5\u8f6c\u5199<\/strong>\u3002\u8fd9\u7c7b\u4f3c\u300cPrompt ASR\u300d\uff0c\u4f46\u5229\u7528\u7684\u662f LLM \u7684 in-context learning\uff0c\u800c\u975e\u7b80\u5355\u70ed\u8bcd\u504f\u7f6e\u3002\u5b9e\u9a8c\u663e\u793a\uff0c\u52a0\u5165\u7528\u6237\u4e0a\u4e0b\u6587\u540e\uff0c\u5e73\u5747 EER \u4ece 9.17% \u8fdb\u4e00\u6b65\u964d\u5230 6.89%\uff0c\u76f8\u5bf9\u518d\u964d 24.9%\u3002Pharmacy \u9886\u57df EER \u4ece 5.97% \u964d\u5230 3.11%\uff0c\u533b\u7597\u573a\u666f\u6536\u76ca\u5c24\u5176\u660e\u663e\u3002<\/p>\n\n\n\n<figure class=\"wp-block-image size-large\"><img loading=\"lazy\" width=\"1024\" height=\"683\" src=\"http:\/\/139.9.1.231\/wp-content\/uploads\/2026\/06\/44444-1024x683.webp\" alt=\"\" class=\"wp-image-31200\" srcset=\"http:\/\/139.9.1.231\/wp-content\/uploads\/2026\/06\/44444-1024x683.webp 1024w, http:\/\/139.9.1.231\/wp-content\/uploads\/2026\/06\/44444-300x200.webp 300w, http:\/\/139.9.1.231\/wp-content\/uploads\/2026\/06\/44444-768x512.webp 768w, http:\/\/139.9.1.231\/wp-content\/uploads\/2026\/06\/44444.webp 1080w\" sizes=\"(max-width: 1024px) 100vw, 1024px\" \/><\/figure>\n\n\n\n<p><strong>&nbsp;<\/strong><strong>2.5 \u5b9e\u9a8c\u7ed3\u679c\uff1a\u5c0f\u6570\u636e\u8d85\u8d8a\u5927\u6a21\u578b<\/strong><\/p>\n\n\n\n<p>\u5728 LibriSpeech test-clean \u4e0a\uff0cCoT-ASR WER 2.20% vs Phi4MM \u57fa\u7ebf 2.41%\uff0c\u76f8\u5bf9\u964d 8.7%\u3002\u66f4\u503c\u5f97\u5173\u6ce8\u7684\u662f EER\uff08\u5b9e\u4f53\u9519\u8bef\u7387\uff09\uff1a8 \u4e2a\u884c\u4e1a\u6d4b\u8bd5\u96c6\u5e73\u5747 EER \u4ece 11.03% \u964d\u5230 9.17%\uff0c\u76f8\u5bf9\u964d 16.9%\u3002\u5bf9\u6bd4\u5f00\u6e90\u5927\u6a21\u578b\uff1aCoT-ASR \u4ec5\u7528 38k \u5c0f\u65f6\u6570\u636e\uff0c\u5e73\u5747 EER 9.17% \u5df2\u7565\u4f18\u4e8e Qwen3-Omni-30B\uff089.19%\uff09\u548c Whisper-large-v3\uff089.53%\uff09\u3002<\/p>\n\n\n\n<p>\u8bba\u6587\u8ba4\u4e3a\uff1a\u5bf9 ASR \u800c\u8a00\uff0c<strong>LLM \u53c2\u6570\u89c4\u6a21\u5e76\u975e\u4e07\u80fd\u94a5\u5319\uff0c\u300c\u4f1a\u4e0d\u4f1a\u7528 LLM \u7684\u63a8\u7406\u80fd\u529b\u300d\u624d\u662f\u5173\u952e<\/strong>\u3002<\/p>\n\n\n\n<p><strong>\u300c<\/strong><em>CoT-ASR \u7684\u542f\u793a\uff1aASR \u6b63\u5728\u4ece\u300c\u58f0\u5b66\u5206\u7c7b\u95ee\u9898\u300d\u8f6c\u5411\u300c\u8bed\u8a00\u7406\u89e3\u95ee\u9898\u300d\u3002\u4e13\u6709\u540d\u8bcd\u3001\u533b\u7597\u672f\u8bed\u3001\u6e38\u620f\u9ed1\u8bdd\u7b49\u573a\u666f\uff0cEER \u6307\u6807\u6bd4 WER \u66f4\u8d34\u8fd1\u771f\u5b9e\u4f53\u9a8c\u3002<\/em><strong>\u300d<\/strong><\/p>\n\n\n\n<h2><strong>\u4e09\u3001Whisfusion\uff1aWhisper \u7684\u5e76\u884c\u89e3\u7801\u9769\u547d<\/strong><\/h2>\n\n\n\n<p>Whisfusion\uff08Parallel ASR Decoding via a Diffusion Transformer\uff09\u662f ICLR 2026 \u6295\u7a3f\u8bba\u6587\uff0c\u5b83\u7784\u51c6\u7684\u662f\u53e6\u4e00\u4e2a\u75db\u70b9\uff1aWhisper \u7f16\u7801\u5668 30 \u79d2\u97f3\u9891\u4e00\u6b21\u524d\u5411\uff0c\u4f46\u89e3\u7801\u5668\u5fc5\u987b\u9010 token \u81ea\u56de\u5f52\u2014\u2014\u6587\u672c\u8d8a\u957f\uff0c\u8d8a\u6162\u3002<\/p>\n\n\n\n<p><strong>&nbsp;<\/strong><strong>3.1 \u67b6\u6784\u9519\u914d\uff1a\u6709\u5168\u91cf\u4e0a\u4e0b\u6587\uff0c\u5374\u53ea\u80fd\u987a\u5e8f\u751f\u6210<\/strong><\/p>\n\n\n\n<p>\u8bba\u6587 Figure 1 \u6e05\u6670\u5c55\u793a\uff1aWhisper-small \u7684\u7f16\u7801\u5668\u8017\u65f6\u51e0\u4e4e\u6052\u5b9a\uff0c\u89e3\u7801\u5668\u8017\u65f6\u968f\u8f93\u51fa\u8bcd\u6570\u7ebf\u6027\u589e\u957f\u300220\u201330 \u79d2\u97f3\u9891\u6bb5\u4e0a\uff0c\u89e3\u7801\u5360\u7aef\u5230\u7aef\u5ef6\u8fdf\u7684\u5927\u5934\u3002<\/p>\n\n\n\n<p>Whisper-Large-v3-turbo \u7b49\u84b8\u998f\u6a21\u578b\u7f13\u89e3\u4e86\u90e8\u5206\u95ee\u9898\uff0c\u4f46 AR \u672c\u8d28\u672a\u53d8\u3002<\/p>\n\n\n\n<p><strong>3.2 \u6838\u5fc3\u8bbe\u8ba1\uff1a\u51bb\u7ed3 Whisper + \u6269\u6563\u6587\u672c\u89e3\u7801\u5668<\/strong><\/p>\n\n\n\n<p>Whisfusion \u7684 hybrid \u67b6\u6784\uff1aWhisper \u7f16\u7801\u5668\u51bb\u7ed3\u4e0d\u52a8\uff0c<\/p>\n\n\n\n<p>\u53ea\u8bad\u7ec3\u8f7b\u91cf Cross-Attention Adapter \u548c Masked Diffusion Decoder\u3002<\/p>\n\n\n\n<ul><li>\u7f16\u7801\u5668\uff1a\u590d\u7528 Whisper \u9884\u8bad\u7ec3\u58f0\u5b66\u8868\u5f81\uff0c6.5k \u5c0f\u65f6\u6570\u636e\u5373\u53ef\u5fae\u8c03<\/li><li><strong>\u89e3\u7801\u5668\uff1a\u57fa\u4e8e Masked Diffusion Model\uff08MDM\uff09\uff0c\u6bcf\u6b65\u5e76\u884c\u66f4\u65b0\u5168\u90e8 token<\/strong><\/li><li><strong>\u63a8\u7406\uff1aParallel Diffusion Decoding\uff08PDD\uff09\uff0c\u591a\u5019\u9009\u5e76\u884c + \u7f6e\u4fe1\u5ea6\u7b5b\u9009<\/strong><\/li><\/ul>\n\n\n\n<figure class=\"wp-block-image size-large\"><img loading=\"lazy\" width=\"1024\" height=\"683\" src=\"http:\/\/139.9.1.231\/wp-content\/uploads\/2026\/06\/5555-1024x683.webp\" alt=\"\" class=\"wp-image-31209\" srcset=\"http:\/\/139.9.1.231\/wp-content\/uploads\/2026\/06\/5555-1024x683.webp 1024w, http:\/\/139.9.1.231\/wp-content\/uploads\/2026\/06\/5555-300x200.webp 300w, http:\/\/139.9.1.231\/wp-content\/uploads\/2026\/06\/5555-768x512.webp 768w, http:\/\/139.9.1.231\/wp-content\/uploads\/2026\/06\/5555.webp 1080w\" sizes=\"(max-width: 1024px) 100vw, 1024px\" \/><\/figure>\n\n\n\n<p><strong>3.3 \u6269\u6563\u89e3\u7801\u5982\u4f55\u5de5\u4f5c<\/strong><\/p>\n\n\n\n<p><strong>Masked Diffusion \u5728\u524d\u5411\u8fc7\u7a0b\u4e2d\u968f\u673a mask \u6587\u672c token\uff0c\u6a21\u578b\u5b66\u4e60\u4ece\u88ab mask \u7684\u5e8f\u5217\u4e2d\u6062\u590d\u539f\u6587\u3002\u63a8\u7406\u65f6\u4ece\u5168 mask \u5e8f\u5217\u51fa\u53d1\uff0c\u8fed\u4ee3\u53bb\u566a\u82e5\u5e72\u6b65\uff0c\u6bcf\u6b65\u6240\u6709\u4f4d\u7f6e\u540c\u65f6\u9884\u6d4b\u3002<\/strong><\/p>\n\n\n\n<p>\u4e0e AR \u7684\u5173\u952e\u5dee\u5f02\uff1aAR \u7b2c t \u4e2a token \u4f9d\u8d56\u524d t-1 \u4e2a\uff1b\u6269\u6563\u89e3\u7801\u6bcf\u6b65\u90fd\u80fd\u300c\u770b\u5230\u300d\u5b8c\u6574\u58f0\u5b66\u4e0a\u4e0b\u6587\u5e76\u53cc\u5411\u5efa\u6a21\u5168\u90e8 token\u3002\u56e0\u6b64\u8f93\u51fa\u957f\u5ea6\u5bf9\u5ef6\u8fdf\u7684\u5f71\u54cd\u5927\u5e45\u51cf\u5f31\u2014\u2014\u8fd9\u6b63\u662f ASR \u9700\u8981\u7684\u7279\u6027\u3002<\/p>\n\n\n\n<p><strong>\u258e&nbsp;<\/strong><strong>3.4 Parallel Diffusion Decoding\uff08PDD\uff09<\/strong><\/p>\n\n\n\n<p>Whisfusion \u8fdb\u4e00\u6b65\u63d0\u51fa PDD \u7b56\u7565\uff1a\u6bcf\u6b65\u751f\u6210 k \u4e2a\u5e76\u884c\u5019\u9009\u5e8f\u5217\uff0c\u6309\u7f6e\u4fe1\u5ea6\u9009\u6700\u4f18\u3002<\/p>\n\n\n\n<p>\u589e\u52a0 k \u53ef\u63d0\u5347\u51c6\u786e\u7387\uff0c\u4f46\u5bf9 RTF \u5f71\u54cd\u6781\u5c0f\u2014\u2014\u56e0\u4e3a\u5e76\u884c\u5019\u9009\u5728\u540c\u4e00 GPU batch \u4e2d\u5b8c\u6210\u3002<\/p>\n\n\n\n<p><strong>\u25cf&nbsp;&nbsp;<\/strong>k=5\u219215\uff1aWER \u4ece 9.1% \u964d\u5230 8.3%\uff0cRTF \u51e0\u4e4e\u4e0d\u53d8<\/p>\n\n\n\n<p><strong>\u25cf&nbsp;&nbsp;<\/strong>Oracle WER 5.9%\uff0c\u6a21\u578b\u5b9e\u9645 8.3%\uff0c68.7% \u6837\u672c\u9009\u4e2d\u8fd1\u6700\u4f18\u5019\u9009<\/p>\n\n\n\n<p><strong>\u25cf&nbsp;&nbsp;<\/strong>\u4e24\u9636\u6bb5\u8bfe\u7a0b\u5b66\u4e60\uff1aStage1 \u5efa\u7acb\u57fa\u7840\uff0cStage2 \u5f15\u5165 PDD \u8fbe\u6700\u4f18<\/p>\n\n\n\n<figure class=\"wp-block-image size-large\"><img loading=\"lazy\" width=\"1024\" height=\"683\" src=\"http:\/\/139.9.1.231\/wp-content\/uploads\/2026\/06\/6666-1024x683.webp\" alt=\"\" class=\"wp-image-31211\" srcset=\"http:\/\/139.9.1.231\/wp-content\/uploads\/2026\/06\/6666-1024x683.webp 1024w, http:\/\/139.9.1.231\/wp-content\/uploads\/2026\/06\/6666-300x200.webp 300w, http:\/\/139.9.1.231\/wp-content\/uploads\/2026\/06\/6666-768x512.webp 768w, http:\/\/139.9.1.231\/wp-content\/uploads\/2026\/06\/6666.webp 1080w\" sizes=\"(max-width: 1024px) 100vw, 1024px\" \/><\/figure>\n\n\n\n<p><strong>\u258e&nbsp;<\/strong><strong>3.5 \u901f\u5ea6\u6570\u636e\uff1a8.4\u00d7 \u4e0d\u662f\u5671\u5934<\/strong><\/p>\n\n\n\n<p>LibriSpeech test-clean\uff1aWhisfusion WER 4.9%\uff0cWhisper-small 5.0%\uff0c\u7cbe\u5ea6\u6301\u5e73\u3002<\/p>\n\n\n\n<p>\u5728 20\u201330 \u79d2\u97f3\u9891\u6bb5\u4e0a\uff0c\u89e3\u7801\u65f6\u95f4 674.7ms \u2192 80.7ms\uff0c\u52a0\u901f 8.4\u00d7\u3002<\/p>\n\n\n\n<p>\u541e\u5410\u65b9\u9762\uff1aWhisfusion \u8d85 3100 tokens\/s\uff0cWhisper-small \u4ec5\u7ea6 103 tokens\/s\uff0c\u5dee\u8ddd 13 \u500d\u4ee5\u4e0a\u3002<\/p>\n\n\n\n<p>RTF 0.005 vs 0.031\uff0c\u610f\u5473\u7740 CPU\/GPU \u7b97\u529b\u9884\u7b97\u53ef\u4ee5\u5927\u5e45\u91ca\u653e\u3002<\/p>\n\n\n\n<p><strong>\u300c<\/strong><em><strong>Whisfusion \u7684\u5c40\u9650\uff1a\u957f\u97f3\u9891\uff0820\u201330s\uff09\u8bad\u7ec3\u6837\u672c\u7a00\u7f3a\uff0c\u8be5\u533a<\/strong>\u95f4 WER 15.9% \u504f\u9ad8\uff1b\u4e0e Oracle \u4ecd\u6709 2.4% \u5dee\u8ddd\uff0c\u5019\u9009\u9009\u62e9\u7b56\u7565\u8fd8\u6709\u4f18\u5316\u7a7a\u95f4\u3002\u4f46\u4f5c\u4e3a Whisper \u751f\u6001\u7684\u300c\u5e76\u884c\u89e3\u7801\u63d2\u4ef6\u300d\uff0c\u65b9\u5411\u975e\u5e38\u6e05\u6670\u3002<\/em><strong>\u300d<\/strong><\/p>\n\n\n\n<h2><strong>\u56db\u3001\u4e24\u6761\u8def\u7ebf\u5982\u4f55\u4e92\u8865<\/strong><\/h2>\n\n\n\n<p>CoT-ASR \u548c Whisfusion \u770b\u4f3c\u90fd\u5728\u300c\u6539\u9020 Whisper\/LLM ASR\u300d\uff0c<\/p>\n\n\n\n<p>\u4f46\u4f18\u5316\u76ee\u6807\u51e0\u4e4e\u6b63\u4ea4\uff1a\u4e00\u4e2a\u8ffd\u51c6\u786e\u7387\u5c24\u5176\u662f\u5b9e\u4f53\u8bc6\u522b\uff0c<\/p>\n\n\n\n<p>\u4e00\u4e2a\u8ffd\u89e3\u7801\u541e\u5410\u4e0e\u5ef6\u8fdf\u3002<\/p>\n\n\n\n<figure class=\"wp-block-image size-large\"><img loading=\"lazy\" width=\"1024\" height=\"683\" src=\"http:\/\/139.9.1.231\/wp-content\/uploads\/2026\/06\/7777-1024x683.webp\" alt=\"\" class=\"wp-image-31214\" srcset=\"http:\/\/139.9.1.231\/wp-content\/uploads\/2026\/06\/7777-1024x683.webp 1024w, http:\/\/139.9.1.231\/wp-content\/uploads\/2026\/06\/7777-300x200.webp 300w, http:\/\/139.9.1.231\/wp-content\/uploads\/2026\/06\/7777-768x512.webp 768w, http:\/\/139.9.1.231\/wp-content\/uploads\/2026\/06\/7777.webp 1080w\" sizes=\"(max-width: 1024px) 100vw, 1024px\" \/><\/figure>\n\n\n\n<p><strong>&nbsp;<\/strong><strong>4.1 \u8303\u5f0f\u5bf9\u6bd4<\/strong><\/p>\n\n\n\n<p><strong>\u2460&nbsp;&nbsp;<\/strong>CoT-ASR\uff1a\u6539\u300c\u751f\u6210\u5185\u5bb9\u300d\u2014\u2014 \u5148\u63a8\u7406\u518d\u8f6c\u5199\uff0c\u6fc0\u6d3b LLM \u77e5\u8bc6<\/p>\n\n\n\n<p><strong>\u2461&nbsp;&nbsp;<\/strong>Whisfusion\uff1a\u6539\u300c\u751f\u6210\u65b9\u5f0f\u300d\u2014\u2014 \u5e76\u884c\u6269\u6563\u66ff\u4ee3\u81ea\u56de\u5f52<\/p>\n\n\n\n<p><strong>\u2462&nbsp;&nbsp;<\/strong>CoT-ASR\uff1a\u9002\u5408\u533b\u7597\u3001\u91d1\u878d\u3001\u5ba2\u670d\u7b49\u5b9e\u4f53\u5bc6\u96c6\u573a\u666f<\/p>\n\n\n\n<p><strong>\u2463&nbsp;&nbsp;<\/strong>Whisfusion\uff1a\u9002\u5408\u5b9e\u65f6\u5b57\u5e55\u3001\u957f\u97f3\u9891\u6279\u8f6c\u3001\u7aef\u4fa7\u4f4e\u5ef6\u8fdf\u573a\u666f<\/p>\n\n\n\n<p><strong>\u258e&nbsp;<\/strong><strong>4.2 \u5bf9\u5de5\u7a0b\u843d\u5730\u7684\u542f\u793a<\/strong><\/p>\n\n\n\n<p><strong>\u25cf&nbsp;&nbsp;<\/strong>\u8bc4\u4f30\u6307\u6807\u8981\u5347\u7ea7\uff1aWER \u4e0d\u591f\uff0c\u5782\u76f4\u573a\u666f\u5e94\u8ddf\u8e2a EER \/ \u5b9e\u4f53\u53ec\u56de<\/p>\n\n\n\n<p><strong>\u25cf&nbsp;&nbsp;<\/strong>Speech LLM \u4e0d\u5fc5\u76f2\u8ffd\u53c2\u6570\u91cf\uff1a38k \u5c0f\u65f6 + \u63a8\u7406\u8303\u5f0f\u53ef\u51fb\u8d25 30B \u6a21\u578b<\/p>\n\n\n\n<p><strong>\u25cf&nbsp;&nbsp;<\/strong>\u89e3\u7801\u5668\u662f\u5ef6\u8fdf\u74f6\u9888\uff1a\u7f16\u7801\u5668\u91cf\u5316\u3001\u84b8\u998f\u4e4b\u5916\uff0cNAR \u6269\u6563\u662f\u4e0b\u4e00\u6218\u573a<\/p>\n\n\n\n<p><strong>\u25cf&nbsp;&nbsp;<\/strong>\u4e24\u8005\u53ef\u7ec4\u5408\uff1aWhisfusion \u5f0f\u5e76\u884c\u89e3\u7801 + CoT \u5f0f\u63a8\u7406 prompt\uff0c\u662f\u503c\u5f97\u63a2\u7d22\u7684\u65b9\u5411<\/p>\n\n\n\n<p><strong>\u258e&nbsp;<\/strong><strong>4.3 \u4e0e\u5f00\u6e90 ASR \u4f18\u5316\u7684\u5173\u7cfb<\/strong><\/p>\n\n\n\n<p>\u5982\u679c\u4f60\u5728\u7528 Whisper \/ SenseVoice \/ sherpa-onnx \u505a\u843d\u5730\uff0c<\/p>\n\n\n\n<p>\u8fd9\u4e24\u7bc7\u8bba\u6587\u63d0\u4f9b\u4e86\u300c\u4e0b\u4e00\u6b65\u8be5\u5f80\u54ea\u8d70\u300d\u7684\u8def\u7ebf\u56fe\uff1a\u51c6\u786e\u7387\u74f6\u9888 \u2192 \u8003\u8651\u5f15\u5165\u63a8\u7406\u5f0f\u8f6c\u5199\u6216 LLM \u540e\u5904\u7406\uff1b\u901f\u5ea6\u74f6\u9888 \u2192 \u5173\u6ce8 NAR\/Flow Matching\/\u6269\u6563\u89e3\u7801\uff0c<\/p>\n\n\n\n<p>\u800c\u975e\u4e00\u5473\u7f29\u5c0f beam\u3002<\/p>\n\n\n\n<p>Whisper-CD\uff08\u5bf9\u6bd4\u89e3\u7801\u6291\u5236\u5e7b\u89c9\uff09\u3001Distilling Conversations\uff08\u591a\u8f6e\u4e0a\u4e0b\u6587\u538b\u7f29\uff09\u7b49\u540c\u671f\u5de5\u4f5c\uff0c<\/p>\n\n\n\n<p>\u4e0e CoT-ASR \/ Whisfusion \u5171\u540c\u6784\u6210 2026 ASR \u8bba\u6587\u7c07\u2014\u2014\u6838\u5fc3\u4e3b\u9898\u90fd\u662f\uff1a\u8ba9 ASR \u66f4\u300c\u806a\u660e\u300d\u3001\u66f4\u300c\u5feb\u300d\u3002<\/p>\n\n\n\n<h2><strong>\u4e94\u3001\u8bba\u6587\u4fe1\u606f\u4e0e\u5ef6\u4f38\u9605\u8bfb<\/strong><\/h2>\n\n\n\n<p><strong>\u258e&nbsp;<\/strong><strong>CoT-ASR<\/strong><\/p>\n\n\n\n<p><strong>\u25cf&nbsp;&nbsp;<\/strong>\u8bba\u6587\uff1aSpeech LLMs are Contextual Reasoning Transcribers<\/p>\n\n\n\n<p><strong>\u25cf&nbsp;&nbsp;<\/strong>\u673a\u6784\uff1aMicrosoft Core AI<\/p>\n\n\n\n<p><strong>\u25cf&nbsp;&nbsp;<\/strong>\u94fe\u63a5\uff1ahttps:\/\/arxiv.org\/html\/2604.00610v1<\/p>\n\n\n\n<p><strong>\u25cf&nbsp;&nbsp;<\/strong>\u9aa8\u5e72\uff1aPhi-4-mini 3.8B + Conformer \u7f16\u7801\u5668 + CTC Adapter<\/p>\n\n\n\n<p><strong>\u258e&nbsp;<\/strong><strong>Whisfusion<\/strong><\/p>\n\n\n\n<p><strong>\u25cf&nbsp;&nbsp;<\/strong>\u8bba\u6587\uff1aWhisfusion: Parallel ASR Decoding via a Diffusion Transformer<\/p>\n\n\n\n<p><strong>\u25cf&nbsp;&nbsp;<\/strong>\u4f1a\u8bae\uff1aICLR 2026\uff08under review\uff09<\/p>\n\n\n\n<p><strong>\u25cf&nbsp;&nbsp;<\/strong>\u94fe\u63a5\uff1ahttps:\/\/openreview.net\/pdf?id=JCujsFnDS7<\/p>\n\n\n\n<p><strong>\u25cf&nbsp;&nbsp;<\/strong>\u6570\u636e\uff1aLibriSpeech 960h \u5fae\u8c03\uff0c6.5k \u5c0f\u65f6\u6df7\u5408\u8bad\u7ec3<\/p>\n\n\n\n<p class=\"has-light-gray-background-color has-background\"><strong>\u603b\u7ed3 &nbsp; &nbsp; &nbsp; &nbsp; &nbsp;<br><\/strong>ASR \u6b63\u4ece\u300c\u76f4\u63a5\u8f6c\u5199\u300d\u8d70\u5411\u300c\u63a8\u7406\u5f0f\u8f6c\u5199\u300d\u4e0e\u300c\u5e76\u884c\u89e3\u7801\u300d\u4e24\u6761\u8def\u7ebf &nbsp; &nbsp; &nbsp; &nbsp; &nbsp;<br>\u2022 CoT-ASR\uff1aOne-Pass \u94fe\u5f0f\u63a8\u7406\uff0cWER -8.7%\uff0cEER -16.9%\uff0c38k \u5c0f\u65f6\u8d85\u8d8a 30B \u6a21\u578b &nbsp; &nbsp; &nbsp; &nbsp; &nbsp;<br>\u2022 Whisfusion\uff1aWhisper + \u6269\u6563 NAR \u89e3\u7801\uff0c20\u201330s \u97f3\u9891\u89e3\u7801\u52a0\u901f 8.4\u00d7 &nbsp; &nbsp; &nbsp; &nbsp; &nbsp;<br>\u2022 CTC Modality Adapter \u4e0e PDD \u5206\u522b\u662f\u4e24\u7bc7\u8bba\u6587\u7684\u5173\u952e\u5de5\u7a0b\u521b\u65b0 &nbsp; &nbsp; &nbsp; &nbsp; &nbsp;<br>\u2022 \u843d\u5730\u65f6\u6309\u573a\u666f\u9009\u8def\u7ebf\uff1a\u5b9e\u4f53\u51c6\u786e vs \u5b9e\u65f6\u5ef6\u8fdf\uff0c\u8bc4\u4f30\u6307\u6807\u4e5f\u8981\u76f8\u5e94\u5347\u7ea7<\/p>\n\n\n\n<p><\/p>\n","protected":false},"excerpt":{"rendered":"<p>\u6587\u7ae0\u6765\u6e90\uff1a\u516c\u4f17\u53f7\u6bcf\u5468\u4e00\u4e2a\u5927\u6a21\u578b\u5e94\u7528 Speech LLMs are Contextual Reasoning  &hellip; <a href=\"http:\/\/139.9.1.231\/index.php\/2026\/06\/02\/asr-roadmap\/\" class=\"more-link\">\u7ee7\u7eed\u9605\u8bfb<span class=\"screen-reader-text\">ASR\u5927\u6a21\u578b\u53d1\u5c55\u8def\u7ebf<\/span><\/a><\/p>\n","protected":false},"author":1,"featured_media":0,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":[],"categories":[29,4,9,38,34],"tags":[],"_links":{"self":[{"href":"http:\/\/139.9.1.231\/index.php\/wp-json\/wp\/v2\/posts\/31180"}],"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=31180"}],"version-history":[{"count":28,"href":"http:\/\/139.9.1.231\/index.php\/wp-json\/wp\/v2\/posts\/31180\/revisions"}],"predecessor-version":[{"id":31215,"href":"http:\/\/139.9.1.231\/index.php\/wp-json\/wp\/v2\/posts\/31180\/revisions\/31215"}],"wp:attachment":[{"href":"http:\/\/139.9.1.231\/index.php\/wp-json\/wp\/v2\/media?parent=31180"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"http:\/\/139.9.1.231\/index.php\/wp-json\/wp\/v2\/categories?post=31180"},{"taxonomy":"post_tag","embeddable":true,"href":"http:\/\/139.9.1.231\/index.php\/wp-json\/wp\/v2\/tags?post=31180"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}