{"id":31417,"date":"2026-07-13T16:11:04","date_gmt":"2026-07-13T08:11:04","guid":{"rendered":"http:\/\/139.9.1.231\/?p=31417"},"modified":"2026-07-13T16:11:06","modified_gmt":"2026-07-13T08:11:06","slug":"hybrid-enrollment-neural-re-scoring","status":"publish","type":"post","link":"http:\/\/139.9.1.231\/index.php\/2026\/07\/13\/hybrid-enrollment-neural-re-scoring\/","title":{"rendered":"\u8bf4\u8bdd\u4eba\u9a8c\u8bc1\uff1aHybrid Enrollment + Neural Re-scoring \u8bba\u6587\u89e3\u8bfb"},"content":{"rendered":"\n

\u672c\u6587\u8ba8\u8bba\u7684\u662f\u77ed\u65f6\u957f\u8bf4\u8bdd\u4eba\u9a8c\u8bc1\uff08Short-duration Speaker Verification, SDSV\uff09\uff1a\u5728\u667a\u80fd\u97f3\u7bb1\u3001\u5bf9\u8bdd\u7ec8\u7aef\u6216\u7528\u6237\u81ea\u5b9a\u4e49\u5173\u952e\u8bcd\u5524\u9192\u573a\u666f<\/strong>\u4e2d\uff0c\u7cfb\u7edf\u5148\u68c0\u6d4b\u5230\u4e00\u53e5\u5f88\u77ed\u7684\u76ee\u6807\u77ed\u8bed\uff0c\u518d\u5224\u65ad\u8fd9\u53e5\u8bdd\u662f\u4e0d\u662f\u6ce8\u518c\u7528\u6237\u672c\u4eba\u8bf4\u7684\u3002<\/p>\n\n\n\n

\u8fd9\u7c7b\u6d4b\u8bd5\u8bed\u97f3\u901a\u5e38\u77ed\u4e8e 3 \u79d2\uff0c\u8eab\u4efd\u4fe1\u606f\u5f88\u5c11\uff0c\u4e14\u66f4\u5bb9\u6613\u53d7\u566a\u58f0\u3001\u97f3\u7d20\u8986\u76d6\u548c\u77ed\u8bed\u5185\u5bb9\u53d8\u5316\u5f71\u54cd\u3002\u8bba\u6587\u7684\u6838\u5fc3\u601d\u8def\u4e0d\u662f\u91cd\u65b0\u8bad\u7ec3\u4e00\u4e2a\u5927\u58f0\u7eb9\u6a21\u578b\uff0c\u800c\u662f\u51bb\u7ed3\u5df2\u6709\u8bf4\u8bdd\u4eba\u9aa8\u5e72\u6a21\u578b\uff0c\u5728\u5176\u4e0a\u8bad\u7ec3\u4e00\u4e2a\u8f7b\u91cf neural verifier\uff1a\u6ce8\u518c\u7aef\u540c\u65f6\u4f7f\u7528\u6587\u672c\u76f8\u5173\uff08TD\uff09\u77ed\u8bed\u548c\u6587\u672c\u65e0\u5173\uff08TI\uff09\u8f83\u957f\u8bed\u97f3\uff0c\u67e5\u8be2\u7aef\u4ecd\u662f TD \u77ed\u8bed\uff0c\u901a\u8fc7\u5168\u5c40\u4f59\u5f26\u76f8\u4f3c\u5ea6\u548c\u53cc\u5411\u5e27\u7ea7 cross-attention \u505a\u795e\u7ecf\u91cd\u6253\u5206\u3002<\/p>\n\n\n\n

1. \u4efb\u52a1\u80cc\u666f\uff1a\u77ed\u8bed\u97f3\u58f0\u7eb9\u9a8c\u8bc1\u4e3a\u4ec0\u4e48\u96be\uff1f<\/h2>\n\n\n\n

\u5728\u7528\u6237\u81ea\u5b9a\u4e49\u5173\u952e\u8bcd\uff08UDKWS\uff09\u7cfb\u7edf\u4e2d\uff0c\u5178\u578b\u94fe\u8def\u662f\uff1a\u5148\u901a\u8fc7 keyword spotting \u627e\u5230\u7528\u6237\u8bf4\u51fa\u7684\u76ee\u6807\u77ed\u8bed\uff0c\u518d\u5bf9\u8fd9\u4e2a\u77ed\u8bed\u7247\u6bb5\u505a\u8bf4\u8bdd\u4eba\u9a8c\u8bc1\u3002<\/strong>\u95ee\u9898\u5728\u4e8e\uff0c\u8fd9\u4e2a\u7247\u6bb5\u901a\u5e38\u53ea\u6709 0.8 \u5230 3 \u79d2\u3002\u76f8\u6bd4\u957f\u8bed\u97f3\u58f0\u7eb9\u9a8c\u8bc1\uff0c\u77ed\u8bed\u7ea7\u97f3\u9891\u4e2d\u7684\u8bf4\u8bdd\u4eba\u4fe1\u606f\u66f4\u5c11\uff0c\u56fa\u5b9a\u7ef4\u5ea6 embedding \u52a0\u4f59\u5f26\u76f8\u4f3c\u5ea6\u7684\u4f20\u7edf\u540e\u7aef\u66f4\u5bb9\u6613\u51fa\u73b0\u5206\u6570\u4e0d\u7a33\u5b9a\u3002<\/p>\n\n\n\n

\"\"
\u56fe 1\uff1a\u9488\u5bf9\u81ea\u5b9a\u4e49\u77ed\u8bed\u7684\u77ed\u65f6\u957f\u8bf4\u8bdd\u4eba\u9a8c\u8bc1\u4efb\u52a1\u793a\u610f\u56fe\u3002\u9996\u5148\uff0c\u8f93\u5165\u7684\u8bed\u97f3\u4f1a\u88ab\u4e00\u4e2a\u81ea\u5b9a\u4e49\u5173\u952e\u8bcd\u68c0\u6d4b\u6a21\u5757\u8fdb\u884c\u5904\u7406\uff0c\u7136\u540e\u4f1a\u88ab\u7528\u6765\u4e0e\u6587\u672c\u4f9d\u8d56\u578b\u6216\u6587\u672c\u72ec\u7acb\u578b\u7684\u6ce8\u518c\u8bed\u97f3\u8fdb\u884c\u9a8c\u8bc1\u3002<\/figcaption><\/figure>\n\n\n\n

\u8bba\u6587\u628a\u6ce8\u518c\u65b9\u5f0f\u5206\u4e3a\u4e24\u7c7b\uff1a<\/p>\n\n\n\n

  • TD enrollment<\/strong>\uff1a\u6ce8\u518c\u8bed\u97f3\u548c\u67e5\u8be2\u8bed\u97f3\u662f\u540c\u4e00\u7c7b\u77ed\u8bed\uff0c\u5185\u5bb9\u4e00\u81f4\uff0c\u97f3\u7d20\u66f4\u5bf9\u9f50\uff0c\u4f46\u6ce8\u518c\u65f6\u957f\u4e5f\u5f88\u77ed\uff0c\u8bf4\u8bdd\u4eba\u4fe1\u606f\u4e0d\u8db3\u3002<\/li>
  • TI enrollment<\/strong>\uff1a\u6ce8\u518c\u8bed\u97f3\u4e0d\u8981\u6c42\u548c\u67e5\u8be2\u77ed\u8bed\u5185\u5bb9\u4e00\u81f4\uff0c\u53ef\u4ee5\u66f4\u957f\uff0c\u8eab\u4efd\u4fe1\u606f\u66f4\u7a33\u5b9a\uff0c\u4f46\u5b58\u5728\u6587\u672c\u5185\u5bb9\u4e0d\u5339\u914d\u3002<\/li><\/ul>\n\n\n\n

    \u8bba\u6587\u7684\u5173\u952e\u89c2\u5bdf\u662f\uff1aTD \u7684\u6587\u672c\u4e00\u81f4\u6027\u6709\u4f18\u52bf\uff0c\u4f46\u53d7\u9650\u4e8e\u77ed\u65f6\u957f\uff1bTI \u6709\u5185\u5bb9 mismatch\uff0c\u4f46\u968f\u7740\u6ce8\u518c\u65f6\u957f\u589e\u52a0\uff0cspeaker representation \u4f1a\u8d8a\u6765\u8d8a\u7a33\u5b9a\u3002\u56e0\u6b64\uff0c\u771f\u5b9e\u7cfb\u7edf\u91cc\u4e0d\u5e94\u8be5\u53ea\u62bc TD \u6216 TI \u5355\u4e00\u8def\u7ebf\uff0c\u800c\u5e94\u8be5\u628a\u4e8c\u8005\u4e92\u8865\u8d77\u6765\u3002<\/p>\n\n\n\n

    2. \u65b9\u6cd5\u603b\u89c8\uff1a\u51bb\u7ed3\u9aa8\u5e72\uff0c\u53ea\u8bad\u7ec3\u8f7b\u91cf\u9a8c\u8bc1\u5668<\/h2>\n\n\n\n
    \"\"<\/figure>\n\n\n\n

    \u8bba\u6587\u63d0\u51fa\u7684\u6846\u67b6\u7531\u4e24\u90e8\u5206\u7ec4\u6210\uff1a\u4e00\u4e2a\u51bb\u7ed3\u7684 pretrained speaker backbone\uff0c\u4ee5\u53ca\u4e00\u4e2a\u53ef\u8bad\u7ec3\u7684 neural verifier\u3002\u9aa8\u5e72\u6a21\u578b\u8d1f\u8d23\u63d0\u53d6 utterance-level \u548c frame-level speaker features\uff1bverifier \u8d1f\u8d23\u628a TD\u3001TI \u4e0e query \u4e4b\u95f4\u7684\u5168\u5c40\u548c\u5c40\u90e8\u8bc1\u636e\u878d\u5408\u6210\u6700\u7ec8\u9a8c\u8bc1\u5206\u6570\u3002<\/p>\n\n\n\n

    \u8bbe TI \u6ce8\u518c\u8bed\u97f3\u4e3a \\(X_{\\mathrm{ti}}^e\\)\uff0cTD \u6ce8\u518c\u8bed\u97f3\u4e3a \\(X_{\\mathrm{td}}^e\\)\uff0cTD \u67e5\u8be2\u8bed\u97f3\u4e3a \\(X_{\\mathrm{td}}^q\\)\u3002\u51bb\u7ed3\u9aa8\u5e72\u4f1a\u8f93\u51fa\u53e5\u7ea7\u8868\u793a\u548c\u5e27\u7ea7\u8868\u793a\uff1a<\/p>\n\n\n\n\\(\nX_{\\mathrm{ti}}^e \\rightarrow (E_{\\mathrm{ti},u}^e, E_{\\mathrm{ti},f}^e),\\quad\nX_{\\mathrm{td}}^e \\rightarrow (E_{\\mathrm{td},u}^e, E_{\\mathrm{td},f}^e),\\quad\nX_{\\mathrm{td}}^q \\rightarrow (E_{\\mathrm{td},u}^q, E_{\\mathrm{td},f}^q)\n\\)\n\n\n\n

    \u8fd9\u91cc \\(u\\) \u8868\u793a utterance-level embedding\uff0c\\(f\\) \u8868\u793a frame-level feature\u3002\u8bba\u6587\u4f7f\u7528\u7684\u9aa8\u5e72\u5305\u62ec ECAPA-TDNN\u3001CAM++ \u548c ERes2Net-L\uff0c\u5168\u90e8\u5728 Vox2 \u4e0a\u9884\u8bad\u7ec3\uff0c\u5e76\u5728\u672c\u6587\u8bad\u7ec3\u4e2d\u4fdd\u6301\u51bb\u7ed3\u3002\u8fd9\u4e00\u70b9\u5bf9\u5de5\u7a0b\u843d\u5730\u5f88\u91cd\u8981\uff1a\u4e0d\u9700\u8981\u63a8\u5012\u91cd\u8bad\u58f0\u7eb9\u6a21\u578b\uff0c\u53ea\u9700\u5728\u5df2\u6709\u58f0\u7eb9\u6a21\u578b\u4e0a\u52a0\u8f7b\u91cf\u91cd\u6253\u5206\u5934\u3002<\/p>\n\n\n\n

    3. \u5168\u5c40\u76f8\u4f3c\u5ea6\uff1a\u540c\u65f6\u4fdd\u7559 TI \u8eab\u4efd\u7a33\u5b9a\u6027\u548c TD \u5185\u5bb9\u4e00\u81f4\u6027<\/h2>\n\n\n\n

    verifier \u9996\u5148\u8ba1\u7b97\u4e24\u4e2a utterance-level \u4f59\u5f26\u76f8\u4f3c\u5ea6\uff1a<\/p>\n\n\n\n\\(\nS_{\\mathrm{ti}}=\\cos(E_{\\mathrm{ti},u}^e,E_{\\mathrm{td},u}^q),\\quad\nS_{\\mathrm{td}}=\\cos(E_{\\mathrm{td},u}^e,E_{\\mathrm{td},u}^q)\n\\)\n\n\n\n

    \\(S_{\\mathrm{ti}}\\) \u66f4\u504f\u5411\u6355\u83b7\u7a33\u5b9a\u7684\u8bf4\u8bdd\u4eba\u8eab\u4efd\u4fe1\u606f\uff0c\\(S_{\\mathrm{td}}\\) \u66f4\u504f\u5411\u5229\u7528\u77ed\u8bed\u5185\u5bb9\u4e00\u81f4\u5e26\u6765\u7684\u5339\u914d\u4f18\u52bf\u3002\u5355\u770b\u8fd9\u4e24\u4e2a\u5206\u6570\u4ecd\u7136\u662f\u4f20\u7edf embedding \u540e\u7aef\u601d\u8def\uff0c\u6240\u4ee5\u8bba\u6587\u8fdb\u4e00\u6b65\u5f15\u5165\u5e27\u7ea7 cross-attention \u6765\u5904\u7406\u77ed\u8bed\u5185\u90e8\u7684\u5c40\u90e8\u5bf9\u9f50\u95ee\u9898\u3002<\/p>\n\n\n\n

    4. Parallel Cross-Attention\uff1a\u5728\u5e27\u7ea7\u522b\u91cd\u65b0\u5bf9\u9f50\u77ed\u8bed\u8bc1\u636e<\/h2>\n\n\n\n

    \u77ed\u65f6\u957f\u8bed\u97f3\u7684\u95ee\u9898\u4e0d\u662f\u53ea\u6709\u201c\u4fe1\u606f\u5c11\u201d\uff0c\u8fd8\u5305\u62ec\u5c40\u90e8\u97f3\u7d20\u548c\u65f6\u95f4\u4f4d\u7f6e\u4e0d\u7a33\u5b9a\u3002\u8bba\u6587\u4f7f\u7528\u5171\u4eab\u7684 parallel cross-attention \u6a21\u5757\uff0c\u5bf9 TD \u6ce8\u518c\u77ed\u8bed\u548c TD \u67e5\u8be2\u77ed\u8bed\u7684 frame-level features \u505a\u53cc\u5411\u6bd4\u8f83\u3002<\/p>\n\n\n\n

    \u6ce8\u518c\u5230\u67e5\u8be2\u65b9\u5411\uff1a<\/p>\n\n\n\n\\(\n\\tilde{Z}^{e}=\n\\mathrm{CrossAtt.}\n(Q=E_{\\mathrm{td},f}^{e},K=E_{\\mathrm{td},f}^{q},V=E_{\\mathrm{td},f}^{q})\n\\)\n\n\n\n

    \u67e5\u8be2\u5230\u6ce8\u518c\u65b9\u5411\uff1a<\/p>\n\n\n\n\\(\n\\tilde{Z}^{q}=\n\\mathrm{CrossAtt.}\n(Q=E_{\\mathrm{td},f}^{q},K=E_{\\mathrm{td},f}^{e},V=E_{\\mathrm{td},f}^{e})\n\\)\n\n\n\n

    \u968f\u540e\u5bf9\u4e24\u4e2a\u65b9\u5411\u7684\u8f93\u51fa\u505a\u65f6\u95f4\u7ef4 max pooling\uff0c\u5e76\u62fc\u63a5\u5f97\u5230\u5c40\u90e8\u5339\u914d\u7279\u5f81\uff1a<\/p>\n\n\n\n\\(\nh_f=[\\max(\\tilde{Z}^{e}) \\Vert \\max(\\tilde{Z}^{q})]\n\\)\n\n\n\n

    \u8fd9\u4e00\u6b65\u662f\u8bba\u6587\u65b9\u6cd5\u7684\u5173\u952e\uff1a\u5b83\u4e0d\u518d\u628a\u77ed\u8bed\u76f4\u63a5\u538b\u6210\u4e00\u4e2a\u5411\u91cf\u786c\u6bd4\uff0c\u800c\u662f\u8ba9\u6ce8\u518c\u77ed\u8bed\u548c\u67e5\u8be2\u77ed\u8bed\u5728\u5e27\u7ea7\u522b\u4e92\u76f8\u201c\u770b\u89c1\u201d\u5bf9\u65b9\uff0c\u4ece\u77ed\u8bed\u5185\u90e8\u627e\u5230\u66f4\u7ec6\u7c92\u5ea6\u7684\u5339\u914d\u8bc1\u636e\u3002\u4e2d\u6587\u89e3\u8bfb\u91cc\u5f3a\u8c03\u7684\u201c\u6ce8\u518c\u770b\u67e5\u8be2\u3001\u67e5\u8be2\u770b\u6ce8\u518c\uff0c\u628a\u77ed\u5e8f\u5217\u91cc\u5bf9\u5f97\u4e0a\u7684\u5c40\u90e8\u8bc1\u636e\u635e\u51fa\u6765\u201d\uff0c\u5bf9\u5e94\u7684\u5c31\u662f\u8fd9\u4e2a\u53cc\u5411 cross-attention \u6a21\u5757\u3002<\/p>\n\n\n\n

    5. \u878d\u5408\u51b3\u7b56\u4e0e\u8bad\u7ec3\u76ee\u6807<\/h2>\n\n\n\n

    \u6700\u7ec8\uff0c\u6a21\u578b\u628a\u5c40\u90e8\u5e27\u7ea7\u7279\u5f81 \\(h_f\\)\u3001TI \u5168\u5c40\u76f8\u4f3c\u5ea6 \\(S_{\\mathrm{ti}}\\) \u548c TD \u5168\u5c40\u76f8\u4f3c\u5ea6 \\(S_{\\mathrm{td}}\\) \u8f93\u5165\u8f7b\u91cf MLP\uff0c\u8f93\u51fa\u6700\u7ec8\u9a8c\u8bc1\u5206\u6570\uff1a<\/p>\n\n\n\n\\(\nS=\\sigma(F(h_f,S_{\\mathrm{ti}},S_{\\mathrm{td}}))\n\\)\n\n\n\n

    \u5176\u4e2d \\(\\sigma(\\cdot)\\) \u662f sigmoid \u51fd\u6570\u3002\u8bad\u7ec3\u4f7f\u7528\u4e8c\u5206\u7c7b\u4ea4\u53c9\u71b5\uff1a<\/p>\n\n\n\n\\(\n\\mathcal{L}=-\n\\left[\ny\\log S+(1-y)\\log(1-S)\n\\right]\n\\)\n\n\n\n

    \\(y \\in \\{0,1\\}\\) \u8868\u793a enrollment \u548c query \u662f\u5426\u6765\u81ea\u540c\u4e00\u8bf4\u8bdd\u4eba\u3002\u5b9e\u9a8c\u4e2d verifier \u5305\u542b\u7ebf\u6027\u6295\u5f71\u5c42\u548c\u5bf9\u79f0 cross-attention \u6a21\u5757\uff0cattention \u4e3a 8 heads\uff0chidden dimension \u4e3a 128\uff1b\u8bad\u7ec3\u5728\u5355\u5f20 RTX 4090 \u4e0a\u8fdb\u884c\uff0cbatch size 256\uff0c\u8bad\u7ec3 25k steps\u3002<\/p>\n\n\n\n

    6. VoxPhrase \u6570\u636e\u96c6\uff1a\u4ece VoxCeleb \u81ea\u52a8\u5207\u51fa\u77ed\u8bed\u7ea7\u58f0\u7eb9\u9a8c\u8bc1\u6570\u636e<\/h2>\n\n\n\n

    \u8bba\u6587\u7684\u53e6\u4e00\u4e2a\u91cd\u8981\u8d21\u732e\u662f\u6784\u5efa VoxPhrase\uff0c\u7528\u6765\u6a21\u62df\u7528\u6237\u81ea\u5b9a\u4e49\u77ed\u8bed\u4e0b\u7684 SDSV\u3002\u6784\u5efa\u6d41\u7a0b\u662f\uff1a\u5148\u5bf9 VoxCeleb \u8bed\u97f3\u505a ASR \u83b7\u5f97 transcript\uff0c\u518d\u7528 forced alignment \u751f\u6210\u8bcd\u6216\u77ed\u8bed\u7ea7\u65f6\u95f4\u6233\uff0c\u968f\u540e\u901a\u8fc7 S2Phrase \u811a\u672c\u628a\u957f\u8bed\u97f3\u5207\u6210 0.8 \u5230 3 \u79d2\u7684\u77ed\u8bed\u7247\u6bb5\uff0c\u5e76\u8fc7\u6ee4\u4f4e\u8d28\u91cf\u5bf9\u9f50\u7ed3\u679c\u3002\u6bcf\u4e2a\u77ed\u8bed\u7247\u6bb5\u4fdd\u7559 speaker identity \u548c waveform\u3002<\/p>\n\n\n\n

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

    VoxPhrase \u7684\u5173\u952e\u89c4\u6a21\u5982\u4e0b\uff1a<\/p>\n\n\n\n

    • \u8bad\u7ec3\u96c6\u6765\u81ea Vox2-dev\uff1a5,994 \u4e2a\u8bf4\u8bdd\u4eba\uff0c215,432 \u4e2a\u77ed\u8bed\u3002<\/li>
    • Eval-1 \u6765\u81ea Vox1\uff1a1,251 \u4e2a\u8bf4\u8bdd\u4eba\uff0c23,036 \u4e2a\u77ed\u8bed\u3002<\/li>
    • Eval-2 \u6765\u81ea Vox2-test\uff1a118 \u4e2a\u8bf4\u8bdd\u4eba\uff0c2,310 \u4e2a\u77ed\u8bed\u3002<\/li>
    • Eval-3 \/ Eval-4 \u6765\u81ea DeepMine\uff0c\u7528\u4e8e OOD \u6d4b\u8bd5\uff0c\u77ed\u8bed\u5206\u522b\u662f \u201cok google\u201d\uff08\u7ea6 2 \u79d2\uff09\u548c \u201cmy voice is my password\u201d\uff08\u7ea6 3 \u79d2\uff09\u3002<\/li><\/ul>\n\n\n\n

      \u4e3a\u4e86\u8ba9\u8bc4\u6d4b\u66f4\u63a5\u8fd1\u771f\u5b9e\u96be\u4f8b\uff0c\u8bba\u6587\u8fd8\u8bbe\u8ba1\u4e86 hard example mining\u3002\u5177\u4f53\u505a\u6cd5\u662f\u5148\u6309\u8bf4\u8bdd\u4eba\u805a\u5408\u77ed\u8bed\u6837\u672c\uff0c\u7528\u9884\u8bad\u7ec3 SV \u6a21\u578b\u6784\u9020 speaker prototype\uff0c\u518d\u8ba1\u7b97\u8bf4\u8bdd\u4eba\u4e4b\u95f4\u7684\u76f8\u4f3c\u5ea6\uff0c\u628a\u201c\u76f8\u4f3c\u4f46\u4e0d\u540c\u4eba\u201d\u7684\u914d\u5bf9\u9009\u4e3a hard negatives\u3002Eval-1 \u4e2d trials \u5305\u62ec top-1% 565,242\u3001top-5% 903,678\u3001top-10% 1,041,902 \u548c random 1,382,110\uff1bEval-2 \u4e2d\u5bf9\u5e94\u4e3a 26,904\u300152,702\u300165,086 \u548c 95,900\u3002<\/p>\n\n\n\n

      7. \u5b9e\u9a8c\u8bbe\u7f6e\uff1a\u4e09\u4e2a\u5f3a\u58f0\u7eb9\u9aa8\u5e72 + \u591a\u79cd\u6ce8\u518c\u65b9\u5f0f<\/h2>\n\n\n\n

      \u8bba\u6587\u4f7f\u7528\u4e09\u4e2a\u5f00\u6e90\u8bf4\u8bdd\u4eba\u6a21\u578b\u4f5c\u4e3a\u51bb\u7ed3\u9aa8\u5e72\uff1aECAPA-TDNN\uff0820.8M \u53c2\u6570\uff0cembedding \u7ef4\u5ea6 192\uff09\u3001CAM++\uff087.2M \u53c2\u6570\uff0cembedding \u7ef4\u5ea6 512\uff09\u548c ERes2Net-L\uff0820.5M \u53c2\u6570\uff0cembedding \u7ef4\u5ea6 192\uff09\u3002\u5b83\u4eec\u5728 VoxCeleb-O \u4e0a\u7684\u57fa\u7840 EER \u5206\u522b\u4e3a 0.86\u30010.65 \u548c 0.57\uff0c\u8bf4\u660e\u9aa8\u5e72\u672c\u8eab\u5df2\u7ecf\u662f\u5f3a\u57fa\u7ebf\u3002<\/p>\n\n\n\n

      \u5bf9\u6bd4\u7684 enrollment \u8bbe\u7f6e\u5305\u62ec\uff1a10 \u79d2 TI\u30013 \u79d2 TI\u3001TD phrase\uff080.8\u20133 \u79d2\uff09\uff0c\u4ee5\u53ca\u52a0\u5165 verifier \u540e\u7684\u6df7\u5408\u6ce8\u518c\u795e\u7ecf\u91cd\u6253\u5206\u3002\u6307\u6807\u4f7f\u7528 Equal Error Rate\uff08EER\uff0c\u8d8a\u4f4e\u8d8a\u597d\uff09\uff0c\u5e76\u62a5\u544a\u4e0d\u540c hard-negative \u96be\u5ea6\u4e0b\u7684\u7ed3\u679c\u3002<\/p>\n\n\n\n

      8. \u4e3b\u8981\u7ed3\u679c\uff1a\u6df7\u5408\u6ce8\u518c + \u795e\u7ecf\u91cd\u6253\u5206\u8de8\u9aa8\u5e72\u7a33\u5b9a\u63d0\u5347<\/h2>\n\n\n\n

      Table 2 \u7684\u6838\u5fc3\u7ed3\u8bba\u662f\uff1a\u5728 3 \u79d2\u6216 10 \u79d2 TI \u6ce8\u518c\u6761\u4ef6\u4e0b\uff0cTI \u901a\u5e38\u4f18\u4e8e\u7eaf TD\uff0c\u56e0\u4e3a\u66f4\u957f\u6ce8\u518c\u97f3\u9891\u63d0\u4f9b\u4e86\u66f4\u7a33\u5b9a\u7684\u8eab\u4efd\u4fe1\u606f\uff1b\u4f46\u5f53 TI \u6781\u77ed\u65f6\uff0cTD \u7684\u77ed\u8bed\u5185\u5bb9\u4e00\u81f4\u6027\u53c8\u53d8\u5f97\u91cd\u8981\u3002\u56e0\u6b64\u6700\u7a33\u7684\u65b9\u6848\u662f TD + TI \u6df7\u5408\u6ce8\u518c\uff0c\u518d\u901a\u8fc7 neural verifier \u91cd\u65b0\u6253\u5206\u3002<\/p>\n\n\n\n

      \u51e0\u4e2a\u4ee3\u8868\u6027\u6570\u5b57\u5982\u4e0b\uff1a<\/p>\n\n\n\n

      • ECAPA-TDNN\uff1aEval-1 \u5e73\u5747 EER \u4ece 10 \u79d2 TI \u7684 6.59 \u964d\u5230 5.75\uff1b3 \u79d2 TI \u4ece 8.23 \u964d\u5230 6.45\uff1bTD phrase \u4ece 10.06 \u964d\u5230 9.27\u3002<\/li>
      • CAM++\uff1aEval-1 \u5e73\u5747 EER \u4ece 10 \u79d2 TI \u7684 6.44 \u964d\u5230 5.35\uff1b3 \u79d2 TI \u4ece 8.15 \u964d\u5230 6.03\uff1bTD phrase \u4ece 9.15 \u964d\u5230 8.31\u3002<\/li>
      • ERes2Net-L\uff1aEval-1 \u5e73\u5747 EER \u4ece 10 \u79d2 TI \u7684 5.27 \u964d\u5230 4.54\uff1b3 \u79d2 TI \u4ece 6.51 \u964d\u5230 5.13\uff1bTD phrase \u4ece 7.96 \u964d\u5230 7.22\u3002<\/li><\/ul>\n\n\n\n

        \u5728\u6700\u96be\u7684 top-1% hard-negative \u573a\u666f\u4e2d\uff0c\u6539\u8fdb\u540c\u6837\u660e\u663e\u3002\u4f8b\u5982 CAM++ \u7684 10 \u79d2 TI top-1% EER \u4ece 11.33 \u964d\u5230 9.58\uff0c3 \u79d2 TI \u4ece 13.34 \u964d\u5230 10.47\uff1bERes2Net-L \u7684 10 \u79d2 TI top-1% EER \u4ece 9.32 \u964d\u5230 8.17\uff0c3 \u79d2 TI \u4ece 11.02 \u964d\u5230 8.99\u3002\u8bf4\u660e\u8be5\u65b9\u6cd5\u4e0d\u662f\u53ea\u5728\u5bb9\u6613\u6837\u672c\u4e0a\u8c03\u5206\uff0c\u800c\u662f\u5728\u76f8\u4f3c\u8bf4\u8bdd\u4eba\u6784\u6210\u7684 hard cases \u4e2d\u4e5f\u6709\u6548\u3002<\/p>\n\n\n\n

        9. TI \u65f6\u957f\u5206\u6790\uff1a\u4ec0\u4e48\u65f6\u5019 TI \u5f3a\uff0c\u4ec0\u4e48\u65f6\u5019 TD \u5f3a\uff1f<\/h2>\n\n\n\n
        \"\"<\/figure>\n\n\n\n

        Figure 4 \u5206\u6790\u4e86 TI enrollment duration \u5bf9 EER \u7684\u5f71\u54cd\u3002\u8bba\u6587\u62a5\u544a\uff0c\u5728 CAM++ \u7684 Eval-1 random \u8bbe\u7f6e\u4e0b\uff0c\u7eaf TD enrollment \u7684 EER \u4e3a 3.62%\uff0c\u52a0\u5165 verifier \u540e\u964d\u5230 3.09%\u3002\u800c 3 \u79d2 TI enrollment \u7684 EER \u4e3a 8.86%\uff0c\u8868\u73b0\u8f83\u5dee\uff1b\u968f\u7740 TI \u65f6\u957f\u4ece 1 \u79d2\u589e\u52a0\u5230 10 \u79d2\uff0cEER \u6301\u7eed\u4e0b\u964d\u3002\u5f53 TI \u65f6\u957f\u8d85\u8fc7 3 \u79d2\u65f6\uff0cTI \u5f00\u59cb\u4f18\u4e8e TD\uff1b\u5f53 TI \u5c0f\u4e8e 2 \u79d2\u65f6\uff0cTI \u4ecd\u5f31\u4e8e TD\u3002<\/p>\n\n\n\n

        \u8fd9\u7ec4\u5b9e\u9a8c\u89e3\u91ca\u4e86\u8bba\u6587\u65b9\u6cd5\u4e3a\u4ec0\u4e48\u8981\u505a hybrid enrollment\uff1aTI \u4e0d\u662f\u5929\u7136\u66f4\u597d\uff0c\u5b83\u4f9d\u8d56\u8db3\u591f\u65f6\u957f\uff1bTD \u4e5f\u4e0d\u662f\u8fc7\u65f6\u65b9\u6848\uff0c\u5728\u6781\u77ed\u6ce8\u518c\u8bed\u97f3\u4e0b\uff0c\u77ed\u8bed\u5185\u5bb9\u4e00\u81f4\u6027\u4ecd\u7136\u6709\u4ef7\u503c\u3002\u7b80\u5355\u628a TI \u548c TD \u5206\u6570\u5e73\u5747\u5e76\u4e0d\u591f\uff0c\u8bba\u6587\u63d0\u5230 10 \u79d2\u65f6 TI+TD(mean) \u51e0\u4e4e\u6536\u655b\u5230 TI-only\uff082.03% vs. 1.98%\uff09\uff0c\u800c\u6df7\u5408\u6ce8\u518c + \u795e\u7ecf\u91cd\u6253\u5206\u80fd\u8fdb\u4e00\u6b65\u8fbe\u5230 1.6%\u3002\u771f\u6b63\u5e26\u6765\u5dee\u8ddd\u7684\u662f\u53ef\u5b66\u4e60\u7684 frame-level re-scoring\uff0c\u800c\u4e0d\u662f\u673a\u68b0\u5e73\u5747\u3002<\/p>\n\n\n\n

        10. OOD \u7ed3\u679c\uff1aDeepMine \u4e0a\u4e5f\u80fd\u964d EER<\/h2>\n\n\n\n

        \u8bba\u6587\u8fd8\u5728 DeepMine \u6784\u9020\u7684 Eval-3 \/ Eval-4 \u4e0a\u505a out-of-distribution \u8bc4\u4f30\u3002\u8fd9\u91cc\u7684\u77ed\u8bed\u5206\u522b\u662f \u201cok google\u201d \u548c \u201cmy voice is my password\u201d\u3002\u7ed3\u679c\u663e\u793a\uff0c\u5728 OOD \u573a\u666f\u4e2d TD enrollment \u901a\u5e38\u4f18\u4e8e\u77ed\u65f6 TI enrollment\uff0c\u56e0\u4e3a\u6587\u672c\u4e00\u81f4\u6027\u66f4\u91cd\u8981\uff1b\u4f46\u6df7\u5408\u6ce8\u518c + verifier \u4ecd\u7136\u53d6\u5f97\u6700\u597d\u7ed3\u679c\u3002<\/p>\n\n\n\n

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

        CAM++ \u4e0a\uff0cEval-3 \/ Eval-4 \u7684 EER \u4ece\u7eaf TD \u7684 8.17 \/ 6.19 \u964d\u5230\u6df7\u5408\u65b9\u6848\u7684 6.71 \/ 3.48\u3002ERes2Net-L \u4e0a\uff0c\u4ece\u7eaf TD \u7684 6.97 \/ 4.54 \u964d\u5230 4.88 \/ 2.38\u3002\u8fd9\u4e2a\u7ed3\u679c\u8bf4\u660e\uff0cverifier \u5b66\u5230\u7684\u4e0d\u662f\u53ea\u9002\u914d VoxPhrase \u57df\u5185\u6570\u636e\u7684\u6253\u5206\u504f\u7f6e\uff0c\u800c\u662f\u5bf9\u77ed\u8bed\u7ea7\u9a8c\u8bc1\u4e2d\u7684\u5c40\u90e8\u5339\u914d\u786e\u5b9e\u6709\u6cdb\u5316\u5e2e\u52a9\u3002<\/p>\n\n\n\n

        11. \u521b\u65b0\u70b9\u603b\u7ed3<\/h2>\n\n\n\n