{"id":31498,"date":"2026-08-11T21:08:00","date_gmt":"2026-08-11T13:08:00","guid":{"rendered":"http:\/\/139.9.1.231\/?p=31498"},"modified":"2026-08-11T16:09:07","modified_gmt":"2026-08-11T08:09:07","slug":"is-text-all-you-need","status":"publish","type":"post","link":"http:\/\/139.9.1.231\/index.php\/2026\/08\/11\/is-text-all-you-need\/","title":{"rendered":"Is Text All You Need? \u6587\u672c\u4f5c\u4e3a\u8bed\u97f3\u5927\u6a21\u578b\u7684\u901a\u7528\u4fe1\u606f\u74f6\u9888"},"content":{"rendered":"\n

Is Text All You Need? Text as a Universal Information Bottleneck for Speech LLMs<\/strong><\/p>\n\n\n\n

\u6838\u5fc3\u4eae\u70b9: \u63d0\u51fa Convex Gate (C-Gate) \u8bed\u97f3-LLM \u6865\u63a5\u6a21\u5757<\/strong>\uff0c\u901a\u8fc7\u51f8\u5305\u7ea6\u675f\u5c06\u6bcf\u5e27\u8bed\u97f3\u8868\u793a\u4e3a LLM \u8bcd\u8868\u5d4c\u5165\u7684\u51f8\u7ec4\u5408\uff0c\u786e\u4fdd\u8868\u5f81\u4e25\u683c\u4f4d\u4e8e\u9884\u8bad\u7ec3 LLM \u7684\u8f93\u5165\u6d41\u5f62\u5185\uff0c\u6709\u6548\u907f\u514d\u8868\u5f81\u6f02\u79fb\u4e0e\u8bcd\u6c47\u9501\u5b9a\u3002<\/strong>\u8be5\u65b9\u6cd5\u5728 ASR \u4e0e\u60c5\u611f\u8bc6\u522b\u8054\u5408\u4efb\u52a1\u4e0a\u8868\u73b0\u4f18\u5f02\uff0cLibriSpeech WER \u76f8\u5bf9\u964d\u4f4e\u9ad8\u8fbe 48.7%\uff0c\u4e14\u60c5\u611f\u8bc6\u522b\u7cbe\u5ea6\u6301\u5e73\u6216\u8d85\u8d8a\u5355\u4efb\u52a1\u57fa\u7ebf\u3002\u7814\u7a76\u63ed\u793a\u8bed\u97f3\u4fe1\u606f\u7684\u6838\u5fc3\u8f7d\u4f53\u662f\u5d4c\u5165\u7a7a\u95f4\u4e2d\u7684\u65f6\u5e8f\u8f68\u8ff9\u800c\u975e\u79bb\u6563 Token \u8eab\u4efd\uff0c\u8bc1\u660e\u4e86\u51e0\u4f55\u5bf9\u9f50\u800c\u975e\u79bb\u6563\u5316\u624d\u662f\u8bed\u97f3-LLM \u63a5\u53e3\u8bbe\u8ba1\u7684\u6839\u672c\u56e0\u7d20\u3002<\/strong><\/p>\n\n\n\n

C-Gate\u4e0d\u662f\u628a\u8bed\u97f3\u751f\u786c\u5730\u7ffb\u8bd1\u6210\u6587\u5b57Token\uff0c\u800c\u662f\u628a\u8bed\u97f3\u8f6c\u6362\u6210LLM\u719f\u6089\u7684\u8fde\u7eed\u8bcd\u5411\u91cf\u8f68\u8ff9\uff0c\u8ba9\u6a21\u578b\u65e2\u80fd\u542c\u6e05\u201c\u8bf4\u4e86\u4ec0\u4e48\u201d\uff0c\u4e5f\u80fd\u611f\u77e5\u201c\u600e\u4e48\u8bf4\u7684\u201d\u3002<\/p><\/blockquote><\/figure>\n\n\n\n

\u8bba\u6587\u94fe\u63a5: https:\/\/arxiv.org\/pdf\/2606.09366v1<\/strong><\/a><\/p>\n\n\n\n

Speech-LLM \u7684\u4e00\u4e2a\u6838\u5fc3\u95ee\u9898\u662f\uff1a\u8bed\u97f3\u7f16\u7801\u5668\u8f93\u51fa\u7684\u8fde\u7eed\u5411\u91cf\uff0c\u5e94\u8be5\u4ee5\u4ec0\u4e48\u5f62\u5f0f\u8fdb\u5165\u4e00\u4e2a\u5df2\u7ecf\u51bb\u7ed3\u7684\u6587\u672c\u5927\u6a21\u578b\uff1f\u5982\u679c\u5f3a\u5236\u8bed\u97f3\u8868\u793a\u63a5\u8fd1\u79bb\u6563\u6587\u5b57 Token\uff0cASR \u4f1a\u6bd4\u8f83\u5bb9\u6613\uff0c\u4f46\u60c5\u7eea\u3001\u97f5\u5f8b\u7b49\u526f\u8bed\u8a00\u4fe1\u606f\u53ef\u80fd\u4e22\u5931\uff1b\u5982\u679c\u5141\u8bb8\u9002\u914d\u5668\u751f\u6210\u4efb\u610f\u8fde\u7eed\u5411\u91cf\uff0c\u8868\u793a\u80fd\u529b\u867d\u7136\u66f4\u5f3a\uff0c\u5374\u53ef\u80fd\u504f\u79bb LLM \u8bad\u7ec3\u65f6\u719f\u6089\u7684\u8f93\u5165\u7a7a\u95f4\uff0c\u5bfc\u81f4\u81ea\u56de\u5f52\u89e3\u7801\u4e0d\u7a33\u5b9a\u3002<\/strong><\/p>\n\n\n\n

\u8bba\u6587\u63d0\u51fa Convex Gate\uff0c\u7b80\u79f0 C-Gate\u3002\u5b83\u89c4\u5b9a\u6bcf\u4e00\u4e2a\u8bed\u97f3\u5e27\u8868\u793a\u90fd\u5fc5\u987b\u662f\u82e5\u5e72 LLM \u539f\u59cb\u8bcd\u5d4c\u5165\u7684\u51f8\u7ec4\u5408\u3002\u6a21\u578b\u4e0d\u5fc5\u628a\u8bed\u97f3\u786c\u6620\u5c04\u6210\u67d0\u4e2a\u6587\u5b57 Token\uff0c\u4f46\u751f\u6210\u7684\u5411\u91cf\u59cb\u7ec8\u4f4d\u4e8e LLM \u8bcd\u5d4c\u5165\u8868\u7684\u51f8\u5305\u5185\u90e8\uff0c\u4ece\u800c\u5728\u201c\u79bb\u6563\u7b26\u53f7\u7ea6\u675f\u201d\u548c\u201c\u81ea\u7531\u8fde\u7eed\u8868\u793a\u201d\u4e4b\u95f4\u53d6\u5f97\u5e73\u8861\u3002<\/strong><\/p>\n\n\n\n

\u5b9e\u9a8c\u6700\u503c\u5f97\u5173\u6ce8\u7684\u7ed3\u8bba\u4e0d\u662f Top-16 \u9009\u4e2d\u4e86\u54ea\u4e9b\u8bcd\uff0c\u800c\u662f\uff1a\u771f\u6b63\u627f\u8f7d\u8bed\u97f3\u4fe1\u606f\u7684\u662f\u8fd9\u4e9b\u8fde\u7eed\u6df7\u5408\u5411\u91cf\u5728\u65f6\u95f4\u8f74\u4e0a\u5f62\u6210\u7684\u8f68\u8ff9\u3002\u6253\u4e71\u8f68\u8ff9\u987a\u5e8f\u6216\u66ff\u6362 LLM \u7684\u8bcd\u5d4c\u5165\u57fa\u5e95\uff0cASR \u548c\u60c5\u7eea\u8bc6\u522b\u90fd\u4f1a\u5d29\u6e83\u3002<\/p>\n\n\n\n

\"\"
Figure 1\uff0cC-Gate \u603b\u4f53\u67b6\u6784\u4e0e\u8bcd\u5d4c\u5165\u51f8\u5305\u793a\u610f\u56fe\u3002<\/em>C-Gate \u5e76\u4e0d\u662f\u76f4\u63a5\u5b66\u4e60\u4e00\u4e2a\u65b0\u7684\u201c\u8bed\u97f3 embedding \u7a7a\u95f4\u201d\uff0c\u800c\u662f\u901a\u8fc7 Query\/Key \u6253\u5206 \u2192 \u4ece\u51bb\u7ed3\u7684 LLM embedding \u4e2d\u9009\u62e9 Top-16 \u2192 \u5bf9\u539f\u59cb embedding \u505a\u52a0\u6743\u6c42\u548c<\/strong>\uff0c\u4ece\u800c\u628a Whisper \u7684\u8bed\u97f3\u8868\u793a\u6620\u5c04\u5230 LLM \u539f\u6709 embedding \u7a7a\u95f4\u7684\u51f8\u5305\u5185<\/strong>\u3002\u8fd9\u4e5f\u662f\u5b83\u4e0d\u9700\u8981 Value Projection \u548c\u540e\u7f6e MLP \u7684\u5173\u952e\u539f\u56e0\u3002<\/figcaption><\/figure>\n\n\n\n

\u4e00\u3001\u4e24\u79cd\u4f20\u7edf Speech-LLM \u63a5\u53e3\u7684\u77db\u76fe<\/h2>\n\n\n\n
  • \u8fd1\u79bb\u6563 Token \u5bf9\u9f50\uff1a<\/strong>\u901a\u8fc7 CTC \u6216\u6587\u5b57\u5bf9\u9f50\u8ba9\u8bed\u97f3\u8868\u793a\u63a5\u8fd1 one-hot Token \u5206\u5e03\uff0c\u6709\u5229\u4e8e\u8f6c\u5199\uff0c\u4f46\u5bb9\u6613\u5f62\u6210 lexical lock-in\uff0c\u538b\u7f29\u60c5\u7eea\u3001\u97f5\u5f8b\u7b49\u4fe1\u606f\u3002<\/li>
  • \u65e0\u7ea6\u675f\u8fde\u7eed\u8868\u793a\uff1a<\/strong>Q-Former \u6216\u8fde\u63a5\u5668\u53ef\u4ee5\u8f93\u51fa\u4efb\u610f\u7a20\u5bc6\u5411\u91cf\uff0c\u8868\u8fbe\u80fd\u529b\u8f83\u5f3a\uff0c\u4f46\u5bb9\u6613\u53d1\u751f basis drift\uff0c\u5373\u8bed\u97f3\u5411\u91cf\u504f\u79bb\u51bb\u7ed3 LLM \u7684\u8f93\u5165\u5206\u5e03\u3002<\/li><\/ul>\n\n\n\n

    C-Gate \u7684\u89c2\u70b9\u662f\uff1a\u5173\u952e\u4e0d\u662f\u8bed\u97f3\u8868\u793a\u662f\u5426\u79bb\u6563\uff0c\u800c\u662f\u5b83\u7684\u51e0\u4f55\u4f4d\u7f6e\u662f\u5426\u4e0e LLM \u7684\u8f93\u5165\u7a7a\u95f4\u517c\u5bb9\u3002LLM \u552f\u4e00\u771f\u6b63\u8bad\u7ec3\u8fc7\u7684\u8fde\u7eed\u8f93\u5165\u7a7a\u95f4\u5c31\u662f\u81ea\u5df1\u7684\u8bcd\u5d4c\u5165\u8868\uff0c\u56e0\u6b64\u8bed\u97f3\u6865\u63a5\u5411\u91cf\u5e94\u8be5\u88ab\u9650\u5236\u5728\u8be5\u5d4c\u5165\u8868\u7684\u51f8\u5305\u4e2d\u3002<\/p>\n\n\n\n

    \u4e8c\u3001C-Gate \u6a21\u578b\u8bbe\u8ba1<\/h2>\n\n\n\n

    1. \u51f8\u5305\u7ea6\u675f<\/h3>\n\n\n\n

    \u8bbe\u51bb\u7ed3 LLM \u7684\u8bcd\u5d4c\u5165\u8868\u4e3a \\(E=\\{E_1,\\ldots,E_V\\}\\)\uff0c\u5176\u51f8\u5305\u5b9a\u4e49\u4e3a\uff1a<\/p>\n\n\n\n\\(\n\\operatorname{conv}(E)=\\left\\{\\sum_{v=1}^{V}\\alpha_vE_v:\\alpha_v\\ge 0,\n\\sum_{v=1}^{V}\\alpha_v=1\\right\\}\n\\)\n\n\n\n

    \u53ea\u8981\u8bed\u97f3\u5411\u91cf\u91c7\u7528\u975e\u8d1f\u3001\u548c\u4e3a 1 \u7684\u8bcd\u5d4c\u5165\u52a0\u6743\u7ec4\u5408\uff0c\u5b83\u5c31\u4e0d\u4f1a\u79bb\u5f00 LLM \u719f\u6089\u7684\u8f93\u5165\u51e0\u4f55\uff0c\u540c\u65f6\u4ecd\u7136\u53ef\u4ee5\u5728\u4e0d\u540c\u8bcd\u5411\u91cf\u4e4b\u95f4\u8fde\u7eed\u79fb\u52a8\uff0c\u4e0d\u5fc5\u9000\u5316\u6210\u67d0\u4e2a\u79bb\u6563 Token\u3002<\/p>\n\n\n\n

    2. \u5168\u8bcd\u8868\u6253\u5206\u4e0e Top-16 \u6df7\u5408<\/h3>\n\n\n\n

    \u6a21\u578b\u4f7f\u7528\u51bb\u7ed3\u7684 Whisper-large-v3 \u7f16\u7801\u5668\u83b7\u5f97\u8bed\u97f3\u9690\u72b6\u6001\u3002\u7ecf\u8fc7\u6b65\u957f\u4e3a 4 \u7684\u5747\u503c\u6c60\u5316\u540e\uff0c\u6bcf\u4e2a\u8bed\u97f3\u72b6\u6001\u6620\u5c04\u6210\u67e5\u8be2\u5411\u91cf\uff1a<\/p>\n\n\n\n\\(\nq_t=\\operatorname{LN}(W_q\\tilde h_t),\\qquad K=W_kE\n\\)\n\n\n\n

    \u5176\u4e2d \\(K\\) \u662f\u6574\u4e2a\u8bcd\u8868\u7684\u6295\u5f71 Key\u3002\u6a21\u578b\u5bf9\u5b8c\u6574\u8bcd\u8868\u8ba1\u7b97\u76f8\u4f3c\u5ea6\uff1a<\/p>\n\n\n\n\\(\n\\pi_t=\\operatorname{softmax}\\left(\\frac{q_tK^{\\top}}{\\sqrt{d_p}\\,\\tau}\\right),\n\\qquad S_t=\\operatorname{TopK}_{16}(\\pi_t)\n\\)\n\n\n\n

    \u7136\u540e\u53ea\u4fdd\u7559\u6982\u7387\u6700\u9ad8\u7684 16 \u4e2a\u8bcd\u5d4c\u5165\uff0c\u5e76\u91cd\u65b0\u5f52\u4e00\u5316\uff1a<\/p>\n\n\n\n\\(\n\\alpha_{t,v}=\n\\begin{cases}\n\\dfrac{\\pi_{t,v}}{\\sum_{u\\in S_t}\\pi_{t,u}}, & v\\in S_t,\\\\\n0, & v\\notin S_t,\n\\end{cases}\n\\qquad\n\\tilde e_t=\\sum_{v\\in S_t}\\alpha_{t,v}E_v\\in\\operatorname{conv}(E)\n\\)\n\n\n\n

    \u8fd9\u91cc\u7684\u4e00\u4e2a\u5173\u952e\u8bbe\u8ba1\u662f\uff1a\u6295\u5f71\u53ea\u7528\u4e8e Query \u548c Key \u6253\u5206\uff0c<\/strong>Value \u59cb\u7ec8\u662f\u672a\u7ecf\u4fee\u6539\u7684\u539f\u59cb LLM \u8bcd\u5d4c\u5165\u3002Top-16 \u540e\u4e5f\u6ca1\u6709 Value Projection \u6216\u989d\u5916 MLP\uff0c\u56e0\u6b64\u51f8\u5305\u7ea6\u675f\u7531\u67b6\u6784\u672c\u8eab\u4fdd\u8bc1\uff0c\u800c\u4e0d\u662f\u4f9d\u8d56\u8bad\u7ec3\u635f\u5931\u8fd1\u4f3c\u5b9e\u73b0\u3002<\/strong><\/p>\n\n\n\n

    3. \u51bb\u7ed3 LLM \u7684\u7edf\u4e00\u89e3\u7801\u63a5\u53e3<\/h3>\n\n\n\n

    \u4efb\u52a1\u6307\u4ee4\u3001\u8bed\u97f3\u4f2a\u5d4c\u5165\u548c\u5386\u53f2\u8f93\u51fa Token \u88ab\u62fc\u63a5\u540e\u9001\u5165 Qwen2.5-7B-Instruct\uff1a<\/p>\n\n\n\n\\(\np_M(y_{1:N}\\mid x_{1:m},\\tilde e_{1:T})=\n\\prod_{i=1}^{N}p_M\\!\\left(y_i\\mid[E(x_{1:m});\\tilde e_{1:T};E(y_{<i})]\\right)\n\\)\n\n\n\n

    \u6240\u6709\u4efb\u52a1\u90fd\u4f7f\u7528\u6807\u51c6\u81ea\u56de\u5f52\u4ea4\u53c9\u71b5\uff0c\u4e0d\u9700\u8981 CTC\u3001\u5916\u90e8\u5206\u7c7b\u5934\u3001\u524d\u7f00\u5f3a\u5236\u6216\u4efb\u52a1\u4e13\u7528\u89e3\u7801\u5668\u3002<\/p>\n\n\n\n

    \u4e09\u3001\u53c2\u6570\u51bb\u7ed3\u4e0e\u591a\u4efb\u52a1\u8bad\u7ec3<\/h2>\n\n\n\n
    \u6a21\u5757<\/th>\u72b6\u6001<\/th><\/tr><\/thead>
    Whisper-large-v3 Encoder<\/td>\u51bb\u7ed3<\/td><\/tr>
    Qwen2.5-7B \u5d4c\u5165\u8868\u3001MLP\u3001LayerNorm\u3001LM Head<\/td>\u51bb\u7ed3<\/td><\/tr>
    C-Gate \u7684 Wq\u3001Wk\u3001LayerNorm\u3001\u6e29\u5ea6\u53c2\u6570<\/td>\u8bad\u7ec3\uff0c\u7ea62.49M<\/td><\/tr>
    Qwen\u7b2c0\u81f323\u5c42 Self-Attention \u7684 Q\/K\/V\/O \u6295\u5f71<\/td>\u8bad\u7ec3\uff0c\u7ea6704.75M<\/td><\/tr>
    \u603b\u53ef\u8bad\u7ec3\u53c2\u6570<\/td>707.25M<\/td><\/tr><\/tbody><\/table><\/figure>\n\n\n\n

    \u8bad\u7ec3\u4f7f\u7528\u4e00\u4e2a\u591a\u4efb\u52a1\u4ea4\u53c9\u71b5\u76ee\u6807\uff0c\u5e76\u901a\u8fc7\u52a8\u6001\u91cd\u52a0\u6743\u907f\u514d\u4e0d\u540c\u4efb\u52a1\u7684\u635f\u5931\u5c3a\u5ea6\u5931\u8861\uff1a<\/p>\n\n\n\n\\(\n\\mathcal{L}^{(t)}=\\sum_{i\\in\\mathcal{T}}w_i^{(t)}\\mathcal{L}_i^{(t)},\n\\qquad\nw_i^{(t)}\\propto\\left(\\frac{\\mathcal{L}_{i,\\mathrm{EMA}}^{(t)}}\n{\\mathcal{L}_{i,\\mathrm{init}}}\\right)^{\\alpha},\\quad \\alpha=1\n\\)\n\n\n\n

    EMA \u8870\u51cf\u7cfb\u6570\u4e3a 0.9\uff0c\u4efb\u52a1\u6743\u91cd\u622a\u65ad\u5728 0.2 \u81f3 5.0\u3002C-Gate-2T \u8054\u5408\u8bad\u7ec3 ASR \u548c\u60c5\u7eea\u4efb\u52a1\uff1bC-Gate-3T \u518d\u589e\u52a0\u8bed\u97f3\u63a8\u7406\u4efb\u52a1\uff0c\u7528\u6765\u6d4b\u8bd5\u5171\u4eab\u6865\u63a5\u8868\u793a\u7684\u80fd\u529b\u8fb9\u754c\u3002<\/p>\n\n\n\n

    \u56db\u3001\u8bad\u7ec3\u6570\u636e\u4e0e\u8bc4\u6d4b\u8bbe\u7f6e<\/h2>\n\n\n\n
    • ASR\uff1a<\/strong>960\u5c0f\u65f6 LibriSpeech\uff0c\u4f7f\u7528 test-clean \u62a5\u544a\u81ea\u56de\u5f52\u548c Teacher-Forced WER\u3002<\/li>
    • \u60c5\u7eea\uff1a<\/strong>\u7ea647\u5c0f\u65f6\u516c\u5f00\u60c5\u7eea\u8bed\u97f3\u6570\u636e\uff0cRAVDESS \u7528\u4f5c8\u5206\u7c7b\u7684\u5206\u5e03\u5916\u95ed\u96c6\u8bc4\u6d4b\u3002<\/li>
    • \u63a8\u7406\uff1a<\/strong>C-Gate-3T \u589e\u52a0\u8bed\u97f3\u63a8\u7406\u4efb\u52a1\uff0c\u8bc4\u6d4b\u5305\u62ec VoiceBench-BBH\u3001BBH-Heldout\u3001SpeechMMLU\u3001MMAU \u548c MMSU\u3002<\/li><\/ul>\n\n\n\n

      \u8bba\u6587\u6ca1\u6709\u5728\u4e3b\u4f53\u4e2d\u8be6\u7ec6\u5217\u51fa\u5168\u90e8 47 \u5c0f\u65f6\u60c5\u7eea\u8bad\u7ec3\u8bed\u6599\u548c\u63a8\u7406\u8bad\u7ec3\u6570\u636e\u7684\u9010\u9879\u6784\u6210\uff0c\u56e0\u6b64\u6570\u636e\u90e8\u5206\u7684\u53ef\u590d\u73b0\u4fe1\u606f\u5f31\u4e8e\u6a21\u578b\u7ed3\u6784\u90e8\u5206\u3002\u4f5c\u8005\u540c\u65f6\u5f3a\u8c03\uff0c\u591a\u9879\u9009\u62e9\u63a8\u7406\u6210\u7ee9\u53ef\u80fd\u53d7\u5230\u6587\u672c\u6377\u5f84\u3001\u7b54\u6848\u5148\u9a8c\u548c\u9009\u9879\u987a\u5e8f\u5f71\u54cd\uff0c\u56e0\u6b64\u628a\u5b83\u4eec\u89c6\u4e3a\u80fd\u529b\u8fb9\u754c\u6d4b\u8bd5\uff0c\u800c\u4e0d\u662f\u5b8c\u5168\u7531\u97f3\u9891\u9a71\u52a8\u7684\u63a8\u7406\u8bc1\u660e\u3002<\/p>\n\n\n\n

      \u4e94\u3001\u4e3b\u8981\u5b9e\u9a8c\u7ed3\u679c<\/h2>\n\n\n\n
      \u6a21\u578b<\/th>AR-WER<\/th>TF-WER<\/th>\u60c5\u7eea\u51c6\u786e\u7387<\/th>SpMMLU<\/th>MMAU<\/th>MMSU<\/th><\/tr><\/thead>
      C-Gate-ASR<\/td>7.76%<\/td>–<\/td>–<\/td>–<\/td>–<\/td>–<\/td><\/tr>
      C-Gate-Emotion<\/td>–<\/td>–<\/td>96.2%<\/td>–<\/td>–<\/td>–<\/td><\/tr>
      C-Gate-Reasoning<\/td>–<\/td>–<\/td>–<\/td>53.2%<\/td>44.0%<\/td>55.5%<\/td><\/tr>
      C-Gate-2T<\/td>4.78%<\/td>3.60%<\/td>97.1%<\/td>–<\/td>–<\/td>–<\/td><\/tr>
      C-Gate-3T<\/td>3.98%<\/td>3.89%<\/td>90.5%<\/td>61.4%<\/td>48.3%<\/td>60.6%<\/td><\/tr><\/tbody><\/table>
      \u6839\u636e Table 1\uff0cWER \u8d8a\u4f4e\u8d8a\u597d\uff0c\u5176\u4f59\u51c6\u786e\u7387\u8d8a\u9ad8\u8d8a\u597d\u3002<\/figcaption><\/figure>\n\n\n\n

      1. ASR \u4e0e\u60c5\u7eea\u4e4b\u95f4\u51fa\u73b0\u6b63\u8fc1\u79fb<\/h3>\n\n\n\n

      C-Gate-2T \u5c06\u81ea\u56de\u5f52 WER \u4ece\u5355\u4efb\u52a1\u6a21\u578b\u7684 7.76% \u964d\u81f3 4.78%\uff0c\u76f8\u5bf9\u964d\u4f4e 38.4%\uff1b\u4e0e\u6b64\u540c\u65f6\uff0c\u60c5\u7eea\u51c6\u786e\u7387\u4ece\u60c5\u7eea\u5355\u4efb\u52a1\u6a21\u578b\u7684 96.2% \u63d0\u5347\u5230 97.1%\u3002\u8fd9\u610f\u5473\u7740\u60c5\u7eea\u76d1\u7763\u6ca1\u6709\u5e72\u6270 ASR\uff0c\u53cd\u800c\u6291\u5236\u4e86 ASR-only \u6a21\u578b\u4e2d\u8f83\u4e25\u91cd\u7684\u63d2\u5165\u9519\u8bef\u3002<\/p>\n\n\n\n

      C-Gate-3T \u52a0\u5165\u63a8\u7406\u4efb\u52a1\u540e\uff0cWER \u8fdb\u4e00\u6b65\u964d\u81f3 3.98%\uff0c\u76f8\u6bd4 C-Gate-ASR \u76f8\u5bf9\u964d\u4f4e 48.7%\uff0c\u4f46\u60c5\u7eea\u51c6\u786e\u7387\u4e0b\u964d\u5230 90.5%\u3002\u56e0\u6b64\u8bba\u6587\u5c06 C-Gate-2T \u89c6\u4e3a ASR \u4e0e\u526f\u8bed\u8a00\u4efb\u52a1\u4e4b\u95f4\u66f4\u5e72\u51c0\u7684\u5e73\u8861\u70b9\uff0c\u800c C-Gate-3T \u662f\u7528\u4e8e\u89c2\u5bdf\u591a\u4efb\u52a1\u5e72\u6270\u8fb9\u754c\u7684\u538b\u529b\u6d4b\u8bd5\u3002<\/p>\n\n\n\n

      2. \u63a8\u7406\u4efb\u52a1\u6709\u6240\u63d0\u5347\uff0c\u4f46\u4e0d\u662f\u7edd\u5bf9SOTA<\/h3>\n\n\n\n

      \u4e0e Reasoning-only \u6a21\u578b\u76f8\u6bd4\uff0cC-Gate-3T \u5728\u4e94\u9879\u63a8\u7406\u6d4b\u8bd5\u4e0a\u5747\u6709\u63d0\u5347\uff0c\u4f8b\u5982 BBH-Heldout \u4ece 23.6% \u63d0\u5347\u5230 40.0%\uff0cSpeechMMLU \u4ece 53.2% \u63d0\u5347\u5230 61.4%\uff0cMMSU \u4ece 55.5% \u63d0\u5347\u5230 60.6%\u3002<\/p>\n\n\n\n

      \u5728\u516c\u5f00\u6a21\u578b\u5c3a\u5ea6\u53c2\u8003\u4e2d\uff0cC-Gate-3T \u7684 LibriSpeech WER \u4e3a 3.98%\uff0c\u4e0d\u53ca Kimi-Audio \u7684 1.28% \u548c Qwen2-Audio \u7684 1.6%\uff1bMMAU 48.3% \u4e5f\u4f4e\u4e8e\u5927\u578b\u97f3\u9891\u6a21\u578b\u3002\u5b83\u7684\u4ef7\u503c\u4e3b\u8981\u662f\u4ee5\u76f8\u540c\u6570\u636e\u548c\u53ef\u8bad\u7ec3\u53c2\u6570\u9884\u7b97\u9a8c\u8bc1\u6865\u63a5\u51e0\u4f55\uff0c\u800c\u4e0d\u662f\u8ffd\u6c42\u7edd\u5bf9\u6392\u884c\u699c\u7b2c\u4e00\u3002<\/p>\n\n\n\n

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

      \u516d\u3001\u673a\u5236\u5206\u6790\uff1a\u4fe1\u606f\u4e0d\u662f\u79bb\u6563\u8bcd\uff0c\u800c\u662f\u65f6\u95f4\u8f68\u8ff9<\/h2>\n\n\n\n

      1. Top-16 \u5206\u5e03\u63a5\u8fd1\u5747\u5300<\/h3>\n\n\n\n

      Top-16 \u5f52\u4e00\u5316\u5206\u5e03\u7684\u71b5\u7ea6\u4e3a\u6700\u5927\u71b5\u7684 96.2% \u81f3 98.7%\uff0c\u8bf4\u660e\u6bcf\u4e00\u5e27\u5e76\u6ca1\u6709\u9009\u62e9\u4e00\u4e2a\u975e\u5e38\u660e\u786e\u7684\u6587\u5b57 Token\u3002Top-1 Token \u4e5f\u4e0d\u5bf9\u5e94\u771f\u5b9e\u8f6c\u5199\u8bcd\u3001\u60c5\u7eea\u6807\u7b7e\u6216\u53ef\u89e3\u91ca\u7c7b\u522b\uff0c\u56e0\u6b64 C-Gate \u4e0d\u662f\u4e00\u4e2a\u9690\u5f0f ASR Token \u68c0\u7d22\u5668\u3002<\/p>\n\n\n\n

      \"\"
      Top-16 \u8def\u7531\u71b5\u548c KL \u7edf\u8ba1\u3002<\/em><\/figcaption><\/figure>\n\n\n\n

      2. \u9009\u4e2d\u652f\u6301\u96c6\u5408\u7684\u65f6\u95f4\u53d8\u5316\u643a\u5e26\u4fe1\u606f<\/h3>\n\n\n\n

      \u76f8\u90bb\u8bed\u97f3\u5e27\u7684 16 \u4e2a\u652f\u6301 Token \u4e2d\uff0c\u5e73\u5747\u53ea\u6709\u7ea6 3 \u4e2a\u4f1a\u7ee7\u7eed\u4fdd\u7559\u3002\u5c06\u6574\u6bb5\u8bed\u97f3\u9009\u62e9\u8fc7\u7684\u652f\u6301 Token \u505a Bag-of-Supports \u63a2\u6d4b\uff0c\u5728 RAVDESS \u4e0a\u8fbe\u5230 77.7% \u51c6\u786e\u7387\uff0c\u9ad8\u4e8e C-Gate \u4e4b\u524d Whisper \u8868\u793a\u7684 67.8%\uff1b\u4f46\u5982\u679c\u628a\u6240\u6709\u5e27\u7684\u6865\u63a5\u5411\u91cf\u76f4\u63a5\u5e73\u5747\uff0c\u51c6\u786e\u7387\u53ea\u6709 16.8%\u3002<\/p>\n\n\n\n

      \u56e0\u6b64\uff0c\u4fe1\u606f\u5e76\u4e0d\u4e3b\u8981\u5b58\u5728\u4e8e\u67d0\u4e00\u5e27\u7684\u6982\u7387\u6743\u91cd\uff0c\u4e5f\u4e0d\u5728\u6574\u53e5\u5e73\u5747\u5411\u91cf\u4e2d\uff0c\u800c\u662f\u5728\u8bed\u97f3\u5e27\u6cbf\u7740\u8bcd\u5d4c\u5165\u51f8\u5305\u79fb\u52a8\u5f62\u6210\u7684\u6709\u5e8f\u8f68\u8ff9\u4e2d\u3002<\/p>\n\n\n\n

      3. \u540c\u4e00\u8f68\u8ff9\u53ef\u88ab\u4e0d\u540c\u4efb\u52a1\u4ee5\u4e0d\u540c\u65b9\u5f0f\u8bfb\u53d6<\/h3>\n\n\n\n

      \u5728\u60c5\u7eea Prompt \u4e0b\uff0cLLM \u7684\u6587\u672c\u5230\u8bed\u97f3\u6ce8\u610f\u529b\u8d8b\u5411\u4f4e\u79e9\uff0c\u7b49\u4ef7\u4e8e\u5bf9\u6574\u6bb5\u97f3\u9891\u8fdb\u884c\u5168\u5c40\u6c60\u5316\uff1b\u5728 ASR Prompt \u4e0b\uff0c\u6ce8\u610f\u529b\u79e9\u660e\u663e\u5347\u9ad8\uff0c\u4e0d\u540c\u8f93\u51fa\u4f4d\u7f6e\u4f1a\u9009\u62e9\u4e0d\u540c\u7684\u5c40\u90e8\u8bed\u97f3\u4f4d\u7f6e\u3002C-Gate-2T \u5728\u540c\u4e00\u4e2a\u6a21\u578b\u4e2d\u5373\u53ef\u6839\u636e Prompt \u4ece\u4f4e\u79e9\u5168\u5c40\u8bfb\u53d6\u5207\u6362\u5230\u4f4d\u7f6e\u654f\u611f\u8bfb\u53d6\u3002<\/p>\n\n\n\n

      \"\"
      Figure 2\uff0c\u60c5\u7eea\u4e0e ASR Prompt \u4e0b\u7684\u4efb\u52a1\u6761\u4ef6\u5316\u6ce8\u610f\u529b\u6709\u6548\u79e9\u3002<\/em><\/figcaption><\/figure>\n\n\n\n

      4. \u56e0\u679c\u5e72\u9884\u9a8c\u8bc1\u4e86\u51f8\u5305\u8f68\u8ff9\u7684\u91cd\u8981\u6027<\/h3>\n\n\n\n

      \u4f5c\u8005\u5728\u540c\u4e00\u4e2a C-Gate-3T \u68c0\u67e5\u70b9\u4e0a\u8fdb\u884c\u5e72\u9884\uff1a<\/p>\n\n\n\n