| { | |
| "paper_id": "O17-1012", | |
| "header": { | |
| "generated_with": "S2ORC 1.0.0", | |
| "date_generated": "2023-01-19T07:59:32.142909Z" | |
| }, | |
| "title": "\u57fa\u65bc i-vector \u8207 PLDA \u4e26\u4f7f\u7528 GMM-HMM \u5f37\u5236\u5c0d\u4f4d\u4e4b \u81ea\u52d5\u8a9e\u8005\u5206\u6bb5\u6a19\u8a18\u7cfb\u7d71 Speaker Diarization based on I-vector PLDA Scoring and using GMM-HMM Forced Alignment", | |
| "authors": [ | |
| { | |
| "first": "\u5f35\u4e58\u82e5", | |
| "middle": [], | |
| "last": "Cheng-Jo", | |
| "suffix": "", | |
| "affiliation": {}, | |
| "email": "" | |
| }, | |
| { | |
| "first": "Ray", | |
| "middle": [], | |
| "last": "Chang", | |
| "suffix": "", | |
| "affiliation": { | |
| "laboratory": "", | |
| "institution": "National Taiwan University", | |
| "location": {} | |
| }, | |
| "email": "" | |
| }, | |
| { | |
| "first": "Hung-Shin", | |
| "middle": [], | |
| "last": "\u674e\u9d3b\u6b23", | |
| "suffix": "", | |
| "affiliation": {}, | |
| "email": "" | |
| }, | |
| { | |
| "first": "", | |
| "middle": [], | |
| "last": "Lee", | |
| "suffix": "", | |
| "affiliation": {}, | |
| "email": "" | |
| }, | |
| { | |
| "first": "Hsin-Min", | |
| "middle": [], | |
| "last": "\u738b\u65b0\u6c11", | |
| "suffix": "", | |
| "affiliation": {}, | |
| "email": "" | |
| }, | |
| { | |
| "first": "", | |
| "middle": [], | |
| "last": "Wang", | |
| "suffix": "", | |
| "affiliation": {}, | |
| "email": "" | |
| }, | |
| { | |
| "first": "Jyh-Shing", | |
| "middle": [ | |
| "Roger" | |
| ], | |
| "last": "\u5f35\u667a\u661f", | |
| "suffix": "", | |
| "affiliation": {}, | |
| "email": "" | |
| }, | |
| { | |
| "first": "", | |
| "middle": [], | |
| "last": "Jang", | |
| "suffix": "", | |
| "affiliation": { | |
| "laboratory": "", | |
| "institution": "National Taiwan University", | |
| "location": {} | |
| }, | |
| "email": "" | |
| } | |
| ], | |
| "year": "", | |
| "venue": null, | |
| "identifiers": {}, | |
| "abstract": "", | |
| "pdf_parse": { | |
| "paper_id": "O17-1012", | |
| "_pdf_hash": "", | |
| "abstract": [], | |
| "body_text": [ | |
| { | |
| "text": "[2] Najim Dehak, Patrick Kenny, R\u00b4eda Dehak, Pierre Dumouchel,and Pierre Ouellet,", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "", | |
| "sec_num": null | |
| } | |
| ], | |
| "back_matter": [], | |
| "bib_entries": { | |
| "BIBREF0": { | |
| "ref_id": "b0", | |
| "title": "Factor Analysisfor Speaker Verification", | |
| "authors": [ | |
| { | |
| "first": "", | |
| "middle": [], | |
| "last": "Front-End", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 2010, | |
| "venue": "IEEE Transactions on Audio, Speech, and Language Processing", | |
| "volume": "", | |
| "issue": "", | |
| "pages": "", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "Front-End Factor Analysisfor Speaker Verification,\" IEEE Transactions on Audio, Speech, and Language Processing, 2010.", | |
| "links": null | |
| }, | |
| "BIBREF1": { | |
| "ref_id": "b1", | |
| "title": "Clustering-based i-vector formulation for speaker recognition", | |
| "authors": [ | |
| { | |
| "first": "H.-S", | |
| "middle": [], | |
| "last": "Lee", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 2014, | |
| "venue": "Proc. Interspeech", | |
| "volume": "", | |
| "issue": "", | |
| "pages": "", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "H.-S. Lee et al. , \"Clustering-based i-vector formulation for speaker recognition,\" in Proc. Interspeech, 2014.", | |
| "links": null | |
| }, | |
| "BIBREF2": { | |
| "ref_id": "b2", | |
| "title": "An overview of automatic speaker diarisation systems", | |
| "authors": [ | |
| { | |
| "first": "S", | |
| "middle": [], | |
| "last": "Tranter", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "D", | |
| "middle": [], | |
| "last": "Reynolds", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 2006, | |
| "venue": "IEEE Transactions on Audio, Speech, and Language Processing", | |
| "volume": "", | |
| "issue": "", | |
| "pages": "", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "S. Tranter and D. Reynolds, \"An overview of automatic speaker diarisation systems,\" IEEE Transactions on Audio, Speech, and Language Processing, Sept. 2006.", | |
| "links": null | |
| }, | |
| "BIBREF3": { | |
| "ref_id": "b3", | |
| "title": "Speaker diarization: A review of recent research", | |
| "authors": [ | |
| { | |
| "first": "X", | |
| "middle": [], | |
| "last": "Miro", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "S", | |
| "middle": [], | |
| "last": "Bozonnet", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "N", | |
| "middle": [], | |
| "last": "Evans", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "C", | |
| "middle": [], | |
| "last": "Fredouille", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "G", | |
| "middle": [], | |
| "last": "Friedland", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "O", | |
| "middle": [], | |
| "last": "Vinyals", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 2012, | |
| "venue": "IEEE Trans. Audio Speech Lang. Process", | |
| "volume": "20", | |
| "issue": "2", | |
| "pages": "356--370", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "X. Anguera Miro, S. Bozonnet, N. Evans, C. Fredouille, G. Friedland, and O. Vinyals, ''Speaker diarization: A review of recent research,'' IEEE Trans. Audio Speech Lang. Process., vol. 20, no. 2, pp. 356-370, Feb. 2012.", | |
| "links": null | |
| }, | |
| "BIBREF4": { | |
| "ref_id": "b4", | |
| "title": "Speaker Diarization with PLDA I-Vector Scoring and Unsupervised Calibration", | |
| "authors": [ | |
| { | |
| "first": "G", | |
| "middle": [], | |
| "last": "Sell", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "D", | |
| "middle": [], | |
| "last": "Garcia-Romero", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 2014, | |
| "venue": "Proceedings of the IEEE Spoken Language Technology Workshop", | |
| "volume": "", | |
| "issue": "", | |
| "pages": "", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "G. Sell and D. Garcia-Romero, \"Speaker Diarization with PLDA I-Vector Scoring and Unsupervised Calibration,\" in Proceedings of the IEEE Spoken Language Technology Workshop, 2014.", | |
| "links": null | |
| }, | |
| "BIBREF5": { | |
| "ref_id": "b5", | |
| "title": "Analysis of I-vector Length Normalization in Speaker Recognition Systems", | |
| "authors": [ | |
| { | |
| "first": "Daniel", | |
| "middle": [], | |
| "last": "Garcia", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "-", | |
| "middle": [], | |
| "last": "Romero", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Carol", | |
| "middle": [ | |
| "Y" | |
| ], | |
| "last": "Espy-Wilson", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 2011, | |
| "venue": "Proceedings of Interspeech", | |
| "volume": "", | |
| "issue": "", | |
| "pages": "", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "Daniel Garcia-Romero and Carol Y. Espy-Wilson, \"Analysis of I-vector Length Normalization in Speaker Recognition Systems,\" in Proceedings of Interspeech, 2011.", | |
| "links": null | |
| }, | |
| "BIBREF6": { | |
| "ref_id": "b6", | |
| "title": "Exploiting Intra-Conversation Variability for Speaker Diarization", | |
| "authors": [ | |
| { | |
| "first": "Stephen", | |
| "middle": [], | |
| "last": "Shum", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Najim", | |
| "middle": [], | |
| "last": "Dehak", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Ekapol", | |
| "middle": [], | |
| "last": "Chuangsuwanich", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Douglas", | |
| "middle": [], | |
| "last": "Reynolds", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Jim", | |
| "middle": [], | |
| "last": "Glass", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 2011, | |
| "venue": "Proceedings of Interspeech", | |
| "volume": "", | |
| "issue": "", | |
| "pages": "", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "Stephen Shum, Najim Dehak, Ekapol Chuangsuwanich, Douglas Reynolds, and Jim Glass, \"Exploiting Intra-Conversation Variability for Speaker Diarization,\" in Proceedings of Interspeech, 2011.", | |
| "links": null | |
| } | |
| }, | |
| "ref_entries": { | |
| "FIGREF0": { | |
| "type_str": "figure", | |
| "uris": null, | |
| "text": "\u6458\u8981 \u8fd1\u5e74\u4f86\uff0ci-vector \u642d\u914d PLDA(Probability Linear Discriminant Analysis)\u7684\u7cfb\u7d71\u5df2\u7d93 \u5728\u81ea\u52d5\u8a9e\u8005\u5206\u6bb5\u6a19\u8a18(Speaker Diarization)\u7684\u7814\u7a76\u4e0a\u7372\u5f97\u4e86\u5f88\u597d\u7684\u7d50\u679c\u3002\u4e0d\u904e\uff0c\u7531\u65bc ivector \u9700\u8981\u7531\u8f03\u9577\u7684\u97f3\u8a0a\u7247\u6bb5\u62bd\u53d6\u51fa\u4f86\u624d\u5177\u6709\u8f03\u4f73\u7684\u8a9e\u8005\u7279\u6027\uff0c\u6240\u4ee5\u8f03\u7121\u6cd5\u6709\u6548\u5730\u8655\u7406 \u6642\u9593\u6975\u77ed\u7684\u8a9e\u53e5\u5340\u6bb5\u3002\u70ba\u6b64\uff0c\u672c\u8ad6\u6587\u63d0\u51fa\u4e00\u500b\u65b0\u7684\u81ea\u52d5\u8a9e\u8005\u5206\u6bb5\u6a19\u8a18\u6846\u67b6\uff1a\u5148\u7531 K \u5e73\u5747 (K-means)\u6f14\u7b97\u6cd5\u5f97\u5230\u521d\u6b65\u7684\u81ea\u52d5\u8a9e\u8005\u5206\u6bb5\u6a19\u8a18\u7d50\u679c\uff0c\u4e26\u64da\u6b64\u5efa\u7acb\u521d\u6b65\u8a9e\u8005\u6a21\u578b\uff0c\u518d \u914d\u5408\u5229\u7528 GMM-HMM(Gaussian Mixture Models-Hidden Markov Models)\u9032\u884c\u5f37\u5236\u5c0d\u4f4d (Forced Alignment)\u4ee5\u53ca\u8a9e\u8005\u5206\u7fa4(Speaker Clustering)\u4f86\u9032\u884c\u81ea\u52d5\u8a9e\u8005\u5206\u6bb5\u6a19\u8a18\u3002\u5f9e \u5be6\u9a57\u4e0a\u6211\u5011\u53ef\u4ee5\u767c\u73fe\uff0c\u96d6\u7136\u55ae\u7368\u5229\u7528 GMM-HMM \u8a9e\u8005\u5206\u7fa4\u4e26\u672a\u6bd4\u4f7f\u7528 GMM-HMM \u5f37 \u5236\u5c0d\u4f4d\u6240\u5f97\u5230\u7684\u53ec\u56de\u7387(Recall)\u4ee5\u53ca\u7cbe\u6e96\u7387(Precision)\u4f86\u5f97\u597d\uff0c\u4f46\u662f\u5229\u7528 GMM-HMM \u8a9e\u8005\u5206\u7fa4\u7684\u7d50\u679c\u518d\u91cd\u65b0\u9032\u884c GMM-HMM \u5f37\u5236\u5c0d\u4f4d\u537b\u53ef\u4ee5\u5f97\u5230\u8f03\u597d\u7684\u53ec\u56de\u7387\u4ee5\u53ca\u7cbe\u6e96 \u7387\uff0c\u6545\u7531 GMM-HMM \u8a9e\u8005\u5206\u7fa4\u4ee5\u5f97\u5230\u66f4\u7d30\u5c0f\u7684\u8a9e\u8005\u8aaa\u8a71\u5340\u6bb5\u5c0d\u81ea\u52d5\u8a9e\u8005\u5206\u6bb5\u6a19\u8a18\u7684\u554f \u984c\u662f\u6709\u5e6b\u52a9\u7684\u3002\u6b64\u5916\uff0c\u9019\u7bc7\u8ad6\u6587\u4e5f\u63a2\u8a0e\u91dd\u5c0d\u4e0d\u540c\u6642\u9593\u9577\u5ea6\u7684\u97f3\u8a0a\u7247\u6bb5\u5c0d\u81ea\u52d5\u8a9e\u8005\u5206\u6bb5\u6a19 \u8a18\u7684\u5f71\u97ff\u3002 \u95dc\u9375\u5b57\uff1a\u81ea\u52d5\u8a9e\u8005\u5206\u6bb5\u6a19\u8a18\uff0cI-vector\uff0cPLDA\uff0cGMM-HMM\uff0c\u5f37\u5236\u5c0d\u4f4d\uff0c\u8a9e\u8005\u5206\u7fa4 119", | |
| "num": null | |
| }, | |
| "TABREF0": { | |
| "num": null, | |
| "type_str": "table", | |
| "html": null, | |
| "content": "<table><tr><td>\u4e00\u3001\u7dd2\u8ad6 \u76ee\uff0c\u5c0d\u9019\u4e9b\u8a9e\u8005\u5efa\u7acb\u7c21\u55ae\u7684\u8a9e\u8005\u6a21\u578b\uff0c\u63a5\u8457\u6839\u64da\u5b83\u5011\u5f7c\u6b64\u4e4b\u9593\u7684\u7570\u540c\uff0c\u5617\u8a66\u53bb\u5408\u4f75\u5169 Extraction) \uff1b\u7b2c\u4e09\u90e8\u5206\u5247\u662f\u9032\u884c\u8a9e\u8005\u5206\u7fa4(Speaker Clustering) \uff1b\u7b2c\u56db\u90e8\u5206\u5247\u70ba\u7cfb\u7d71\u6548\u80fd \u5716\u4e8c\u3001\u5207\u5272\u7247\u6bb5\u7684\u6d41\u7a0b\u3002 \u6bb5\u70ba\u5206\u7fa4\u7684\u55ae\u4f4d\u800c\u5f97\u5230\u7684\u7d50\u679c\u59cb\u7d42\u9084\u662f\u592a\u904e\u9b06\u6563\u3002\u4e00\u822c\u800c\u8a00\uff0c\u6211\u5011\u5728\u8655\u7406\u8a9e\u97f3\u7684\u554f\u984c\u90fd \u6700\u5f8c\uff0c\u5716\u4e5d\u63cf\u8ff0\u4e86 K \u5e73\u5747\u6f14\u7b97\u6cd5\u3001\u5f37\u5236\u5c0d\u4f4d\u3001\u5229\u7528 GMM-HMM \u9032\u884c\u8a9e\u8005\u5206\u7fa4\u7684 \u6578\u70ba 2048\uff0c\u800c i-vector \u7684\u7dad\u5ea6\u5247\u70ba 600\u3002 \u5c0f\u7684\u8a9e\u8005\u7247\u6bb5\u518d\u9032\u884c\u884c\u5f37\u5236\u5c0d\u4f4d\u4ee5\u53ca\u8a9e\u8005\u5206\u7fa4\u662f\u5426\u6709\u52a9\u65bc\u81ea\u52d5\u8a9e\u8005\u5206\u6bb5\u6a19\u8a18\u7684\u6548\u80fd\uff0c \u5716\u5341\u4e94\u6211\u5011\u91dd\u5c0d\u7b2c\u4e8c\u6b21\u5f37\u5236\u5c0d\u4f4d\u4e0b\u53bb\u8a55\u4f30\u6bcf\u500b\u97f3\u8a0a\u7247\u6bb5\u7684\u9577\u5ea6\u8207\u91cd\u758a\u6642\u9593\u7684\u914d\u5c0d</td></tr><tr><td>\u96a8\u8457\u6642\u4ee3\u4e0d\u65b7\u7684\u6f14\u9032\uff0c\u4eba\u5011\u5728\u8655\u7406\u8a9e\u97f3\u7684\u6280\u8853\u4e5f\u6108\u4f86\u6108\u6210\u719f\u3002\u5c31\u62ff\u8a9e\u8005\u8fa8\u8b58 (Speaker Recognition)\u7684\u9818\u57df\u4f86\u8b1b\uff0c\u5f9e\u7576\u521d\u4f7f\u7528\u85c9\u8457\u9ad8\u65af\u6df7\u5408\u6a21\u578b(Gaussian Mixture Models, GMM)[1] \u4f86\u5efa\u7acb\u5ee3\u7fa9\u80cc\u666f\u6a21\u578b(Universal Background Models, UBM) \uff0c\u53ca\u81f3\u806f \u5408\u56e0\u7d20\u5206\u6790(Joint Factor Analysis, JFA) \uff0c\u5230\u76ee\u524d\u6700\u5ee3\u70ba\u6d41\u884c\u7684 i-vector [2][3]\uff0c\u5728\u5efa\u7acb\u7279 \u5169\u8a9e\u8005\u7684\u6a21\u578b\uff0c\u76f4\u5230\u627e\u5230\u6700\u4f73\u7684\u8a9e\u8005\u6578\u76ee\uff1b3)\u5c0d\u65bc\u6bcf\u4e00\u500b\u7fa4\u96c6\u90fd\u7d66\u4e88\u4e00\u500b\u8a9e\u8005\u8b58\u5225\uff0c \u4e26\u8a18\u9304\u8a9e\u8005\u5340\u9593\uff0c\u6700\u5f8c\u518d\u8207\u6b63\u78ba\u6a19\u8a18(Ground Truth)\u6bd4\u5c0d\u3002 \u56e0\u70ba\u6211\u5011\u6703\u5c07\u9304\u97f3\u8a18\u9304\u5207\u5272\u6210\u8a31\u591a\u53ea\u5305\u542b\u55ae\u4e00\u8a9e\u8005\u7684\u97f3\u8a0a\u7247\u6bb5\uff0c\u6240\u4ee5\u5728\u9032\u884c\u8a9e\u8005 \u5206\u7fa4\u7684\u6b65\u9a5f\u6642\uff0c\u53ef\u4ee5\u8996\u70ba\u5c0d\u6bcf\u500b\u97f3\u8a0a\u7247\u6bb5\u9032\u884c\u4e00\u9023\u4e32\u7684\u8a9e\u8005\u8fa8\u8b58\u3002\u5728\u9019\u500b\u60f3\u6cd5\u4e4b\u4e0b\uff0c \u7684\u8a55\u4f30\u65b9\u5f0f\u3002\u9019\u56db\u500b\u90e8\u5206\u5c07\u6703\u5728\u672c\u7ae0\u7684\u5404\u500b\u5c0f\u7bc0\u4e2d\u4e00\u4e00\u4ecb\u7d39\u3002\u5176\u4e2d\u7279\u5fb5\u53c3\u6578\u7684\u90e8\u5206\u6211\u5011 \u662f\u63a1\u7528\u4ee5\u4e0b\u5169\u500b\u7279\u5fb5\uff0c\u5206\u5225\u662f\u6885\u723e\u5012\u983b\u8b5c\u4fc2\u6578(Mel-Frequency Cepstrum Coefficients, MFCC) \uff0c\u4ee5\u53ca i-vector\u3002\u6700\u5f8c\u5728\u9032\u884c\u8a9e\u8005\u5206\u7fa4\u6642\uff0c\u6211\u5011\u63d0\u51fa\u4e86\u4e00\u500b\u7cfb\u7d71\u6548\u80fd\u7684\u8a55\u4f30\u6a5f\u5236\uff0c \u5e0c\u671b\u85c9\u7531\u6b64\u8a55\u4f30\u6a5f\u5236\u4f86\u5224\u65b7\u5206\u7fa4\u5f8c\u7684\u7d50\u679c\u8207\u6b63\u78ba\u6a19\u8a18\u7684\u5dee\u8ddd\u3002 (\u4e8c)\u7279\u5fb5\u62bd\u53d6 C1\u3001C2\uff0c\u7fa4\u5fc3\u7a31\u70ba Q1\u3001Q2\uff1b2)\u67e5\u8a62\u5269\u9918\u7684 i-vector \u5c0d Q1\u3001Q2 \u4e4b PLDA \u5206\u6578\uff0c\u4e26\u4e14\u6bd4\u8f03 \u5176\u5927\u5c0f\uff0c\u5982\u679c\u548c Q1 \u8005\u8f03\u9ad8\uff0c\u5247\u88ab\u5206\u914d\u5230 C1\uff0c\u53cd\u4e4b\u5247\u5206\u914d\u5230 C2\uff1b3)\u91cd\u65b0\u5b9a\u7fa9 C1 \u548c C2 \u7684 \u7fa4\u5fc3\uff0c\u76ee\u6a19\u70ba\u627e\u4e00\u500b\u548c\u7fa4\u5167\u6240\u6709 i-vector \u6700\u76f8\u4f3c\u7684\u4e00\u500b i-vector\uff0c\u8a08\u7b97\u7684\u65b9\u5f0f\u5982\u4e0b\u5f0f \u662f\u4ee5\u97f3\u6846(Frame)\u70ba\u55ae\u4f4d\uff0c\u901a\u5e38\u4e00\u500b\u97f3\u6846\u7684\u6642\u9593\u70ba 32 \u5fae\u79d2\uff0c\u76f8\u5c0d\u65bc\u6211\u5011\u7684\u97f3\u8a0a\u7247\u6bb5\u4ee5 \u79d2\u70ba\u55ae\u4f4d\u5be6\u5728\u5dee\u8ddd\u592a\u5927\u3002\u56e0\u6b64\uff0c\u6211\u5011\u5f9e\u7b2c\u4e00\u968e\u6bb5\u7684\u81ea\u52d5\u8a9e\u8005\u5206\u6bb5\u6a19\u8a18\u7d50\u679c\u4e2d\u5f97\u5230\u4e86\u5169\u4f4d \u8a9e\u8005\u6240\u6709\u7684\u8a9e\u8005\u7247\u6bb5\uff0c\u91dd\u5c0d\u9019\u4e9b\u8a9e\u8005\u7247\u6bb5\u62bd\u53d6\u5176 13 \u7dad\u7684\u6885\u723e\u5012\u983b\u8b5c\u4fc2\u6578\u4f86\u8a13\u7df4\u5169\u500b\u8a9e \u8005\u6a21\u578b\uff0c\u63a5\u8457\u5229\u7528 GMM-HMM \u53bb\u91cd\u65b0\u8abf\u6574\u6240\u6709\u8a9e\u8005\u5340\u6bb5\u7684\u7bc4\u570d\uff0c\u9019\u6a23\u7684\u52d5\u4f5c\u7a31\u505a\u5f37\u5236 \u8a9e\u8005\u5340\u6bb5\u97f3\u6846\u7e3d\u6578\u548c\u6b63\u78ba\u6a19\u8a18\u5167\u5ba2\u6236\u97f3\u6846\u7e3d\u6578\u7684\u6bd4\u4f8b\uff1b\u7cbe\u6e96\u7387\u5247\u8868\u793a\u843d\u5728\u6b63\u78ba\u6a19\u8a18\u5167 \u5c0d\u4f4d\u3002\u5716\u4e03\u662f\u5f37\u5236\u5c0d\u4f4d\u7684\u8aaa\u660e\u5716\uff0c\u91dd\u5c0d\u7b2c\u4e00\u968e\u6bb5\u81ea\u52d5\u8a9e\u8005\u6a19\u793a\u7684\u7d50\u679c\u6240\u8a13\u7df4\u7684\u8a9e\u8005\u6a21\u578b\uff0c \u7d50\u679c\u6bd4\u8f03\uff0c\u5176\u4e2d\u8f03\u660e\u986f\u89c0\u5bdf\u5230\u8b8a\u5316\u7684\u6211\u7528\u7d05\u8272\u65b9\u6846\u6a19\u8a18\uff0c\u4e26\u4e14\u653e\u5927\u986f\u793a\u65bc\u5716\u5341\u7576\u4e2d\u3002 \u7531\u5716\u5341\u7684(1)\u6211\u5011\u53ef\u4ee5\u767c\u73fe\uff0c\u5728\u6dfa\u85cd\u8272\u6a19\u8a18\u7684\u5ba2\u6236\u8aaa\u8a71\u5340\u6bb5\u4e2d\u5c1a\u5b58\u5728\u8457\u7d30\u5c0f\u9ec3\u8272\u5340 \u6bb5\uff0c\u4e5f\u5c31\u662f\u5ba2\u6236\u8aaa\u8a71\u5340\u6bb5\u3002\u56e0\u6b64\u5229\u7528 GMM-HMM \u9032\u884c\u8a9e\u8005\u5206\u7fa4\u53ef\u4ee5\u5c07\u96fb\u8a71\u9304\u97f3\u5167\u7d30 \u6211\u5011\u4f7f\u7528\u53ec\u56de\u7387(Recall)\u4ee5\u53ca\u7cbe\u6e96\u7387(Precision)\u4f86\u4f5c\u70ba\u8a55\u4f30\u81ea\u52d5\u8a9e\u8005\u5206\u6bb5\u6a19\u8a18 \u7684\u6a19\u6e96\uff0c\u5176\u4e2d\u6211\u5011\u7684\u53ec\u56de\u7387\u548c\u7cbe\u6e96\u7387\u662f\u5b9a\u7fa9\u5728\u5ba2\u6236\u7684\u8a9e\u8005\u7247\u6bb5\u3002\u7531\u65bc\u6211\u5011\u4f7f\u7528\u7684\u8cc7\u6599 \u5eab\u90fd\u6703\u7531\u5ba2\u670d\u5148\u958b\u59cb\u8aaa\u8a71\uff0c\u6240\u4ee5\u6211\u5011\u5728\u505a\u5b8c\u81ea\u52d5\u8a9e\u8005\u5206\u6bb5\u6a19\u8a18\u4e4b\u5f8c\u53d6\u7b2c\u4e8c\u4f4d\u8a9e\u8005\u505a\u70ba \u56e0\u6b64\u9084\u6703\u53e6\u5916\u6bd4\u8f03\u9019\u5169\u500b\u7d50\u679c\uff1a\u7b2c\u4e8c\u6b21 20 \u8f2a\u5f37\u5236\u5c0d\u4f4d\u7684\u7d50\u679c\uff0c\u548c\u7b2c\u4e8c\u6b21 GMM-HMM \u8a9e\u8005\u5206\u7fa4\u7684\u7d50\u679c\u3002 \u7531\u5716\u5341\u4e8c\u9084\u6709\u5716\u5341\u4e09\u53ef\u4ee5\u770b\u5f97\u51fa\u4f86 K \u5e73\u5747\u6f14\u7b97\u6cd5\u4e0d\u8ad6\u5728\u53ec\u56de\u7387\u4ee5\u53ca\u6e96\u78ba\u7387\u90fd\u6703\u5f97 \u5c0d\u81ea\u52d5\u8a9e\u8005\u5206\u6bb5\u6a19\u8a18\u7684\u5f71\u97ff\uff0c\u6211\u5011\u767c\u73fe\u5728\u9577\u5ea6\u70ba 5 \u7684\u914d\u5c0d\u5916\uff0c\u5176\u4ed6\u7684\u914d\u5c0d\u7d44\u5408\u5c0d\u81ea\u52d5\u8a9e\u8005\u5206\u6bb5\u6a19\u8a18\u7684\u5f71\u97ff\u4e26\u6c92\u6709\u5f88\u5927\u7684\u5dee\u5225\u3002\u5716\u5341\u516d\u63cf \u5c0f\u7684\u7247\u6bb5\u627e\u51fa\u4f86\u3002 \u5ba2\u6236\uff0c\u91dd\u5c0d\u5176\u8a9e\u8005\u5340\u6bb5\u4f86\u8a08\u7b97\u53ec\u56de\u7387\u4ee5\u53ca\u7cbe\u6e96\u7387\u3002\u53ec\u56de\u7387\u8868\u793a\u843d\u5728\u6b63\u78ba\u6a19\u8a18\u5167\u7684\u5ba2\u6236 \u5230\u6700\u5dee\u7684\u7d50\u679c\uff0c\u56e0\u70ba\u5b83\u518d\u9032\u884c\u8a9e\u8005\u5206\u7fa4\u7684\u55ae\u4f4d\u662f\u4ee5\u79d2\u70ba\u55ae\u4f4d\u9032\u884c\u7684\uff0c\u76f8\u5c0d\u65bc\u5176\u4ed6\u65b9\u5f0f \u8ff0\u4e86\u7b2c\u4e8c\u6b21\u5f37\u5236\u5c0d\u4f4d 20 \u8f2a\u7684\u53ec\u56de\u7387\u8207\u7cbe\u6e96\u7387\u7684\u95dc\u4fc2\u3002</td></tr><tr><td>\u5b9a\u7684\u8a9e\u8005\u6a21\u578b\u4e0a\uff0c\u5176\u6e96\u78ba\u6027\u5df2\u7d93\u6709\u76f8\u7576\u5e45\u5ea6\u7684\u63d0\u5347\u3002\u7136\u800c\uff0c\u6709\u6642\u5019\u6211\u5011\u4e0d\u9700\u8981\u77e5\u9053\u6bcf\u4e00 \u53e5\u5c0d\u8a71\u662f\u51fa\u81ea\u54ea\u4e00\u4f4d\u8a9e\u8005\uff0c\u56e0\u70ba\u5728\u67d0\u4e9b\u60c5\u5883\u4e2d\uff0c\u53ea\u6709\u67d0\u4e00\u4f4d\u8a9e\u8005\u662f\u6700\u91cd\u8981\u7684\uff0c\u800c\u5176\u4ed6\u4eba \u7684\u8072\u97f3\u76f8\u5c0d\u4e0a\u4e26\u6c92\u6709\u90a3\u9ebc\u95dc\u9375\u3002\u4f8b\u5982\uff0c\u5728\u8ffd\u8e64\u5acc\u7591\u72af\u7684\u72af\u7f6a\u9304\u97f3\u4e2d\uff0c\u6211\u5011\u53ea\u9700\u8981\u95dc\u6ce8\u5acc \u7591\u72af\u7684\u8072\u97f3\uff0c\u800c\u5176\u9918\u5728\u9304\u97f3\u4e2d\u51fa\u73fe\u7684\u4eba\u8072\u5c31\u6c92\u6709\u8fa8\u8b58\u5176\u8eab\u5206\u7684\u5fc5\u8981\uff0c\u53ea\u9700\u8981\u7d66\u4e88\u4ed6\u5011\u8a9e \u8005\u8b58\u5225(Speaker Identity)\u5373\u53ef\u3002\u4e00\u822c\u800c\u8a00\uff0c\u6211\u5011\u6703\u5c07\u9019\u985e\u53ea\u9700\u628a\u4e0d\u540c\u8a9e\u8005\u4ee5\u8a9e\u8005\u8b58\u5225 \u7684\u65b9\u5f0f\u6a19\u8a18\u4e0b\u4f86\u7684\u554f\u984c\u7d71\u7a31\u70ba\u81ea\u52d5\u8a9e\u8005\u5206\u6bb5\u6a19\u8a18(Automatic Speaker Diarization)\u7684\u554f \u984c\uff0c\u800c\u9019\u7a2e\u554f\u984c\u53c8\u53ef\u88ab\u7a31\u70ba\u300cWho Spoke When\u300d \uff0c\u4e5f\u5c31\u662f\u8981\u5c07\u4e00\u6bb5\u9304\u97f3\u8cc7\u6599\u4e2d\u7684\u8a9e\u8005\u5340 \u5206\u51fa\u4f86\uff0c\u4e26\u4e00\u4e00\u6a19\u793a\u4ed6\u5011\u7684\u8eab\u5206\u8b58\u5225\u4ee5\u53ca\u6642\u9593\u6233\u8a18(Time Stamp)[4]\u3002\u5728\u672c\u7bc7\u8ad6\u6587\u4e2d\uff0c \u6211\u5011\u53ea\u9700\u5224\u65b7\u5169\u500b\u97f3\u8a0a\u7247\u6bb5\u662f\u5426\u51fa\u81ea\u65bc\u540c\u4e00\u4f4d\u8a9e\u8005 [6]\u3002\u76ee\u524d\u5728\u8a9e\u8005\u8fa8\u8b58\u7684\u7814\u7a76\u4e2d\uff0ci-vector \u6280\u8853\u5df2\u7d93\u76f8\u7576\u6210\u719f\u4e14\u88ab\u5ee3\u6cdb\u4f7f\u7528\uff0c\u5b83\u80fd\u5c07\u4e0d\u540c\u9577\u5ea6\u7684\u8a9e\u97f3\u8f49\u63db\u6210\u4e00\u500b\u5177\u6709\u76f8\u540c \u7dad\u5ea6\u7684\u5411\u91cf\u4e14\u53ef\u4ee5\u4fdd\u7559\u5176\u4e2d\u7684\u8a9e\u8005\u8cc7\u8a0a\uff0c\u4e26\u5c07\u97f3\u8a0a\u7684\u901a\u9053\u566a\u97f3(Channel Noise)\u6ffe \u9664\u3002\u56e0\u6b64\uff0c\u5728\u81ea\u52d5\u8a9e\u8005\u5206\u6bb5\u6a19\u8a18\u4e2d\uff0c\u6211\u5011\u7684\u4e3b\u8981\u4efb\u52d9\u662f\u5c0d\u6240\u6709\u97f3\u8a0a\u7247\u6bb5\u62bd\u53d6\u5176 i-\u5716\u4e09\u3001\u7279\u5fb5\u62bd\u53d6\u7684\u6d41\u7a0b\u3002 \u0305 = argmax \u2211 ( , ) \uff0c\u2200 \u2208 \uff0c\u2200i \u2208 {1, 2} \u5229\u7528 GMM-HMM \u5c0d\u6574\u500b\u4e32\u63a5\u97f3\u6a94\u505a\u5f37\u5236\u5c0d\u4f4d\uff0c\u5176\u5f37\u5236\u5c0d\u4f4d\u7684\u7d50\u679c\u4e26\u4e0d\u6703\u6539\u8b8a\u8a9e\u8005\u5340\u6bb5 \u4ee5\u97f3\u6846\u505a\u70ba\u5206\u7fa4\u55ae\u4f4d\u592a\u904e\u7c97\u7565\uff0c\u4e0d\u904e\u4e5f\u662f\u97f3\u70ba\u5efa\u7acb\u65bc i-vector \u4ee5\u53ca PLDA \u7684\u6a5f\u5236\u4e0b\uff0c \u7684\u5ba2\u6236\u8a9e\u8005\u5340\u6bb5\u97f3\u6846\u7e3d\u6578\u548c\u5ba2\u6236\u8a9e\u8005\u5340\u6bb5\u97f3\u6846\u7e3d\u6578\u7684\u6bd4\u4f8b\u3002\u5716\u5341\u4e00\u7c21\u55ae\u7d66\u4e88\u4e00\u500b\u8a08\u7b97 \u2208 , \u2260 \u7684\u6578\u91cf\uff0c\u800c\u662f\u6539\u8b8a\u5b83\u5011\u7684\u76f8\u5c0d\u7bc4\u570d\u3002\u7b2c\u56db\u7bc0\u7684\u5be6\u9a57\u4e5f\u6703\u8aaa\u660e\uff0c\u5f37\u5236\u5c0d\u4f4d\u7684\u8a9e\u8005\u81ea\u52d5\u5206\u6bb5 \u5b83\u5c0d\u63a5\u4e0b\u4f86\u9032\u884c\u7684\u5f37\u5236\u5c0d\u4f4d\u6709\u4e00\u500b\u826f\u597d\u7684\u5206\u7fa4\u57fa\u790e\uff0c\u4f7f\u5f97\u8a9e\u8005\u5340\u6bb5\u7684\u7bc4\u570d\u9032\u884c\u5fae\u8abf\u4e4b \u53ec\u56de\u7387\u4ee5\u53ca\u7cbe\u6e96\u7387\u7684\u7bc4\u4f8b\uff0c\u5176\u4e2d C1\u3001C2 \u8868\u793a\u6b63\u78ba\u6a19\u8a18\u5167\u5ba2\u6236\u97f3\u6846\u7e3d\u6578\uff0cC'1\u3001C'2 \u8868\u793a \u63a5\u8457\uff0c\u6211\u5011\u5c07\u4e0a\u4e00\u5c0f\u7bc0\u5f97\u5230\u7684\u4e32\u63a5\u97f3\u6a94\u5207\u5272\u6210\u8a31\u591a\u76f8\u540c\u9577\u5ea6\u4e26\u4e14\u90e8\u5206\u91cd\u758a\u7684\u97f3\u8a0a\u7247 \u5176\u4e2d\uff0cPLDA(k, j)\u8868\u793a\u67e5\u8a62 k \u548c j \u9019\u5169\u500b i-vector \u7684 PLDA \u5206\u6578\uff1b4)\u91cd\u8907\u6b65\u9a5f 2)\u548c \u7d50\u679c\u5728\u9032\u884c 20 \u6b21\u4e4b\u524d\u5c31\u6703\u9054\u5230\u6536\u6582\u3002 \u5f8c\u53ef\u4ee5\u4f7f\u53ec\u56de\u7387\u4ee5\u53ca\u7cbe\u6e96\u7387\u6709\u5927\u5e45\u5ea6\u7684\u63d0\u5347\u3002\u6700\u5f8c\u5728\u9019\u4e94\u500b\u6bd4\u8f03\u65b9\u6cd5\u4e2d\uff0c\u6211\u5011\u7531\u7b2c\u4e8c \u5ba2\u6236\u8a9e\u8005\u5340\u6bb5\u97f3\u6846\u7e3d\u6578\uff0c\u800c E1\u3001E2 \u4ee3\u8868\u843d\u5728\u6b63\u78ba\u6a19\u8a18\u5167\u7684\u5ba2\u6236\u97f3\u6846\u7e3d\u6578\uff0c\u5247\u53ec\u56de\u7387\u4ee5 \u6bb5 [8]\uff0c\u4e26\u5c0d\u9019\u4e9b\u97f3\u8a0a\u7247\u6bb5\u62bd\u53d6 13 \u7dad\u6885\u723e\u5012\u983b\u8b5c\u4fc2\u6578\u4e4b\u5f8c\u518d\u5206\u5225\u6c42\u51fa\u5176 i-vector\u3002\u5728\u6b64 3)\u76f4\u5230 K \u5e73\u5747\u6f14\u7b97\u6cd5\u6536\u6582\u70ba\u6b62\u3002 \u53ca\u7cbe\u6e96\u7387\u7684\u8a08\u7b97\u5982\u4e0b\uff1a \u6b21\u5f37\u6b21\u5c0d\u4f4d\u5f97\u5230\u6700\u597d\u7684\u53ec\u56de\u7387(89.08 %)\u548c\u7cbe\u6e96\u7387(94.55 %) \u3002 \u7279\u5225\u5f37\u8abf\u7684\u662f\uff0c\u6b64\u8655\u7684\u97f3\u8a0a\u7247\u6bb5\u5167\u4e26\u975e\u53ea\u5305\u542b\u4e00\u4f4d\u8a9e\u8005\uff0c\u56e0\u70ba\u5b83\u7d93\u7531\u8a9e\u97f3\u7684\u4e32\u63a5\u800c\u4f86\uff0c \u5716\u4e00\u3001\u96d9\u8a9e\u8005\u4e4b\u81ea\u52d5\u8a9e\u8005\u5206\u6bb5\u6a19\u8a18\u7cfb\u7d71\u6d41\u7a0b\u3002 (\u4e00) \u5207\u5272\u7247\u6bb5 \u70ba\u4e86\u5f9e\u96fb\u8a71\u8a9e\u97f3\u4e2d\u5f97\u5230\u53ea\u5305\u542b\u4e00\u4f4d\u8a9e\u8005\u7684\u8aaa\u8a71\u7247\u6bb5\uff0c\u6211\u5011\u9810\u5148\u8a13\u7df4\u4e00\u7d44\u975c\u97f3\u6a21\u578b \u6240\u4ee5\u88e1\u9762\u53ef\u80fd\u4e0d\u53ea\u5305\u542b\u4e00\u4f4d\u8a9e\u8005\uff0c\u5206\u5225\u70ba\u4e0b\u9762\u4e09\u7a2e\u53ef\u80fd\uff1a1)\u97f3\u8a0a\u7247\u6bb5\u5167\u53ea\u6709\u5ba2\u670d\u7684\u8072 \u97f3\uff1b2)\u97f3\u8a0a\u7247\u6bb5\u5167\u53ea\u6709\u5ba2\u6236\u7684\u8072\u97f3\uff1b3)\u97f3\u8a0a\u7247\u6bb5\u5167\u540c\u6642\u542b\u6709\u5ba2\u6236\u8207\u5ba2\u670d\u7684\u8072\u97f3\uff0c\u4ee5\u4e0b \u7c21\u7a31\u70ba\u6df7\u5408\u3002\uff0c\u7531\u5716\u56db\u4e2d\u6211\u5011\u4efb\u610f\u53d6\u51fa\u4e09\u7a2e\u8a9e\u8005(\u5ba2\u6236\u3001\u5ba2\u670d\u3001\u6df7\u5408)\u4e4b\u5404\u5169\u6bb5\u97f3\u8a0a\u7247 \u6bb5\u6240\u62bd\u53d6\u51fa\u7684 i-vector\uff0c\u8a08\u7b97\u5176\u5f7c\u6b64\u4e4b\u9593\u7684 PLDA \u5206\u6578\uff0c\u53ef\u4ee5\u89c0\u5bdf\u51fa\u5ba2\u6236\u5c0d\u5ba2\u6236\u6216\u5ba2\u670d \u5716\u5341\u3001\u653e\u5927\u986f\u793a\u5f37\u5236\u5c0d\u4f4d\u8207 GMM-HMM \u8a9e\u8005\u5206\u7fa4\u7684\u5dee\u7570\u5716\u3002 20 \u8f2a\u4e4b\u524d\u53ec\u56de\u7387\u4ee5\u53ca\u7cbe\u6e96\u7387\u90fd\u6703\u9054\u5230\u98fd\u548c\uff0c\u4e26\u4e14\u7531\u5716\u5341\u4e8c\u548c\u5716\u5341\u4e09\u89c0\u5bdf\u5f97\u77e5\uff0c\u7b2c\u4e00\u6b21 (\u4e09)\u8a9e\u8005\u5206\u7fa4 \u503c\u5f97\u4e00\u63d0\u7684\u662f\uff0c\u6b64\u8655\u5c0d\u97f3\u8a0a\u7247\u6bb5\u9032\u884c\u8a9e\u8005\u5206\u7fa4\uff0c\u6703\u9047\u5230\u5982\u5716\u516d\u7684\u554f\u984c\uff1a\u5982\u679c\u524d\u4e00\u500b \u97f3\u8a0a\u7247\u6bb5\u88ab\u5206\u914d\u5230 C1\uff0c\u800c\u5f8c\u4e00\u500b\u97f3\u8a0a\u7247\u6bb5\u88ab\u5206\u914d\u5230 C2\uff0c\u90a3\u6211\u5011\u5982\u4f55\u53bb\u6c7a\u5b9a\u91cd\u758a\u90e8\u5206\u7684 \u985e\u5225\uff1f \u53ec\u56de\u7387(Recall) = 1 + 2 (3) \u5716\u5341\u56db\u3001\u97f3\u8a0a\u7247\u6bb5 4 \u79d2\uff0c\u91cd\u758a\u6642\u9593 0.5 \u79d2\u7684\u5f37\u5236\u5c0d\u4f4d\u548c\u7b2c\u4e8c\u6b21\u5f37\u5236\u5c0d\u4f4d\u7684\u53ec\u56de\u7387\u66f2 \u5716\u5341\u516d\u3001\u7b2c\u4e8c\u6b21\u5f37\u5236\u5c0d\u4f4d 20 \u8f2a\u7684\u5e73\u5747\u53ec\u56de\u7387\u8207\u5e73\u5747\u7cbe\u6e96\u7387\u7684\u5c0d\u61c9\u5716\u3002 1 + 2 \u7cbe\u6e96\u7387(Precision) = \u5716\u5341\u4e09\u3001\u5e73\u5747\u7cbe\u6e96\u7387\u7684\u6bd4\u8f03\u3002 \u7dda\u3002 1 + 2 \u2032 1 + \u2032 2 (4) \u6b64\u5916\u7531\u5716\u5341\u56db\u3001\u5716\u5341\u4e94\u53ef\u4ee5\u89c0\u5bdf\u5230\uff0c\u4e0d\u8ad6\u662f\u7b2c\u4e00\u6b21\u6216\u8005\u7b2c\u4e8c\u6b21\u5f37\u5236\u5c0d\u4f4d\u5728\u9032\u884c\u7b2c \u4e94\u3001\u7d50\u8ad6</td></tr><tr><td>\u6211\u5011\u6703\u5c07\u8eab\u5206\u8b58\u5225\u4ee5\u53ca\u6642\u9593\u6233\u8a18\u7d71\u7a31\u70ba\u8a9e\u8005\u5340\u6bb5(Speaker Region)\u3002 \u5ee3\u7fa9\u4f86\u8aaa\uff0c\u81ea\u52d5\u8a9e \u8005\u5206\u6bb5\u6a19\u8a18\u7684\u554f\u984c\u4e3b\u8981\u6703\u5206\u70ba\u5169\u7a2e\u985e\u578b\uff0c\u4e00\u7a2e\u662f\u6703\u8b70\u9304\u97f3(Conference Recordings) \uff0c\u53e6 \u4e00\u7a2e\u5247\u70ba\u5ee3\u64ad\u65b0\u805e(Broadcast News)[5]\u3002\u9019\u5169\u7a2e\u60c5\u5883\u6700\u5927\u7684\u5dee\u5225\u5728\u65bc\uff0c\u5ee3\u64ad\u65b0\u805e\u53ef\u4ee5 (Silence Model)\u4ee5\u53ca\u8a9e\u97f3\u6a21\u578b(Speech Model) \u3002\u9996\u5148\u5f9e 50 \u53e5\u96fb\u8a71\u8a9e\u97f3\u4e2d\u91dd\u5c0d\u975c\u97f3\u4ee5 \u5c0d\u5ba2\u670d\u7684 i-vector \u5f7c\u6b64\u9593\u7684 PLDA \u5206\u6578\u662f\u76f8\u5c0d\u8f03\u9ad8\u7684\uff0c\u4e0d\u904e\u6df7\u5408\u5c0d\u6df7\u5408\u7684 i-vector \u5f7c\u6b64 \u9593\u7684 PLDA \u5206\u6578\u537b\u6c92\u6709\u9019\u6a23\u7684\u95dc\u4fc2\u3002\u56e0\u6b64\u6839\u64da\u4e0d\u540c\u5ba2\u670d\u8207\u5ba2\u6236\u8072\u97f3\u7684\u6df7\u5408\u7a0b\u5ea6\uff0c\u5728 i-\u5716\u4e03\u3001\u4f7f\u7528 GMM-HMM \u9032\u884c\u5f37\u5236\u5c0d\u4f4d\u793a\u610f\u5716 \u6211\u5011\u7531\u5be6\u9a57\u5f97\u77e5\u5f37\u5236\u5c0d\u4f4d\u4e4b\u5f8c\u9032\u884c\u7684 GMM-HMM \u8a9e\u8005\u5206\u7fa4\u5f97\u5230\u8f03\u7d30\u5c0f\u7684\u8a9e\u8005\u5340 \u7528 GMM-HMM \u9032\u884c\u8a9e\u8005\u5206\u7fa4\u7684\u53ec\u56de\u7387\u4ee5\u53ca\u7cbe\u6e96\u7387\u90fd\u6bd4\u7b2c\u4e00\u6b21\u5f37\u5236\u5c0d\u4f4d\u9084\u8981\u597d\uff0c\u6240\u4ee5 \u9664\u4e86\u7b2c\u4e8c\u968e\u6bb5\u7684\u81ea\u52d5\u8a9e\u8005\u5206\u6bb5\u6a19\u793a\u7d50\u679c\u4e4b\u5916\uff0c\u70ba\u4e86\u60f3\u78ba\u8a8d\u627e\u51fa\u7d30\u5c0f\u8a9e\u8005\u5340\u6bb5\u7684\u7d50 \u6bb5\u6709\u52a9\u65bc\u8a13\u7df4\u51fa\u66f4\u597d\u7684\u8a9e\u8005\u6a21\u578b\uff0c\u4f7f\u7b2c\u4e8c\u6b21\u5f37\u5236\u5c0d\u4f4d\u80fd\u5f97\u5230\u66f4\u597d\u7684\u7d50\u679c\u3002\u503c\u5f97\u89c0\u5bdf\u7684\u662f\uff0c \u53ca\u8a9e\u97f3\u7684\u7247\u6bb5\u62bd\u53d6 13 \u7dad\u6885\u723e\u5012\u983b\u8b5c\u4fc2\u6578\uff0c\u7528\u4f86\u8a13\u7df4\u5177\u6709 32 \u500b\u6210\u5206(Component)\u7684 vector \u7684\u8868\u793a\u4e0a\u53ef\u8996\u70ba\u5169\u500b\u4e0d\u540c\u7684\u8a9e\u8005\u3002\u6b64\u5916\uff0c\u70ba\u4e86\u4e0d\u8b93\u8a9e\u8005\u7684\u8072\u97f3\u8b8a\u5316\u592a\u5927\uff0c\u4e26\u4e14\u589e \u5728\u78ba\u8a8d\u4e86\u6642\u9593\u8f03\u9577\u7684\u8a9e\u8005\u5340\u6bb5\u4e4b\u5f8c\uff0c\u6211\u5011\u5c07\u9762\u5c0d\u5728\u81ea\u52d5\u8a9e\u8005\u5206\u6bb5\u6a19\u8a18\u4e0a\u7684\u96e3\u984c\uff0c \u679c\u518d\u900f\u904e\u91cd\u65b0\u5c0d\u4f4d\u6703\u4e0d\u6703\u6709\u66f4\u597d\u7684\u6548\u679c\uff0c\u5728\u7b2c\u56db\u7ae0\u4e5f\u6703\u5c07\u5be6\u9a57\u7684\u7d50\u679c\u5c55\u793a\u51fa\u4f86\u3002 \u5229\u7528\u7b2c\u4e00\u6b21 GMM-HMM \u7684\u8a9e\u8005\u5206\u7fa4\u6240\u5f97\u5230\u7684\u8a9e\u8005\u5340\u6bb5\u53ef\u4ee5\u8a13\u7df4\u51fa\u66f4\u597d\u7684\u8a9e\u8005\u6a21\u578b\uff0c \u7b2c\u4e8c\u6b21\u4f7f\u7528 GMM-HMM \u9032\u884c\u8a9e\u8005\u5206\u7fa4\u7684\u7d50\u679c\u4e26\u6c92\u6709\u6bd4\u7b2c\u4e8c\u6b21\u5f37\u5236\u5c0d\u4f4d\u7684\u7d50\u679c\u9084\u8981\u4f86 GMM\u3002\u6709\u4e86\u9019\u5169\u500b\u6a21\u578b\u4e4b\u5f8c\uff0c\u5c31\u53ef\u4ee5\u91dd\u5c0d\u6bcf\u500b\u6e2c\u8a66\u97f3\u6a94\u900f\u904e\u6b64\u9ad8\u65af\u6df7\u5408\u6a21\u578b\u9032\u884c\u8a9e\u97f3 \u6c7a\u96d9\u8a9e\u8005\u4e4b\u81ea\u52d5\u8a9e\u8005\u5206\u6bb5\u6a19\u8a18\u554f\u984c\uff0c\u4e26\u4e14\u63a2\u8a0e\u5f97\u5230\u8a9e\u8005\u5206\u7fa4\u5f8c\u7684\u8a9e\u8005\u5340\u57df\u900f\u904e\u5f37\u5236\u5c0d \u5075\u6e2c(Voice Activity Detection, VAD)\u800c\u5f97\u5230\u8a31\u591a\u53ea\u5305\u542b\u4e00\u4f4d\u8a9e\u8005\u7684\u8aaa\u8a71\u7247\u6bb5\uff0c\u6700\u5f8c\u518d \u52a0\u7cfb\u7d71\u5728\u8655\u7406\u81ea\u52d5\u8a9e\u8005\u5206\u6bb5\u6a19\u8a18\u4e0a\u7684\u7cbe\u7d30\u5ea6\uff0c\u6211\u5011\u5617\u8a66\u4f7f\u7528\u53ef\u91cd\u758a\u7684\u97f3\u8a0a\u7247\u6bb5\uff0c\u800c\u4e14\u97f3 \u5c31\u662f\u5c07\u6975\u77ed\u6642\u9593\u7684\u8a9e\u8005\u5340\u6bb5\u6a19\u793a\u51fa\u4f86\u3002\u7d93\u7531\u5f37\u5236\u5c0d\u4f4d\uff0c\u6211\u5011\u53d6\u5f97\u4e86\u66f4\u70ba\u6e96\u78ba\u7684\u8a9e\u8005\u5340 \u5716\u4e5d\u3001K \u5e73\u5747\u6f14\u7b97\u6cd5\u3001\u5f37\u5236\u5c0d\u4f4d\u3001GMM-HMM \u8a9e\u8005\u5206\u7fa4\u7684\u7d50\u679c\u6bd4\u8f03\u3002 \u4e09\u3001\u8cc7\u6599\u5eab\u8207\u5be6\u9a57\u8a55\u4f30 \u5716\u5341\u4e00\u3001\u53ec\u56de\u7387\u3001\u7cbe\u6e96\u7387\u7684\u8a08\u7b97\u793a\u610f\u5716\u3002 \u9032\u800c\u63d0\u5347\u7b2c\u4e8c\u6b21\u5f37\u5236\u5c0d\u4f4d\u7684\u53ec\u56de\u7387\u4ee5\u53ca\u7cbe\u6e96\u7387\u3002\u4e0d\u904e\u7b2c\u4e8c\u6b21 GMM-HMM \u8a9e\u8005\u5206\u7fa4\u537b \u7684\u597d\uff0c\u6240\u4ee5\u6211\u5011\u8a8d\u70ba\u518d\u7e7c\u7e8c\u9032\u884c GMM-HMM \u8a9e\u8005\u5206\u7fa4\u8207\u548c\u5f37\u5236\u5c0d\u4f4d\u9019\u6a23\u7684\u5faa\u74b0\u5c0d\u81ea\u52d5</td></tr><tr><td>\u662f\u9810\u5148\u6f14\u7df4\u904e\u7684\uff0c\u6240\u4ee5\u5be6\u969b\u7684\u9304\u97f3\u60c5\u5883\u53ef\u80fd\u662f\u8a31\u591a\u8a9e\u8005\u4e00\u500b\u63a5\u8457\u4e00\u500b\u8a0e\u8ad6\u8b70\u984c\uff1b\u76f8\u5c0d\u5730\uff0c \u6703\u8b70\u9304\u97f3\u4e2d\u53c3\u8207\u8005\u7684\u767c\u8a00\u5177\u6709\u8f03\u9ad8\u81ea\u767c\u6027\uff0c\u6240\u4ee5\u8a9e\u8005\u8ddf\u8a9e\u8005\u7684\u5c0d\u8a71\u53ef\u80fd\u5728\u6642\u9593\u4e0a\u6703\u6709\u91cd \u758a\uff0c\u5728\u6703\u8b70\u9304\u97f3\u7576\u4e2d\u4e5f\u6709\u53ef\u80fd\u6703\u51fa\u73fe\u62cd\u624b\uff0c\u7b11\u8072\u7b49\u60c5\u6cc1\u51fa\u73fe\uff0c\u800c\u5728\u9019\u7bc7\u8ad6\u6587\u6240\u63a2\u8a0e\u7684\u60c5 \u5883\u662f\u4ecb\u65bc\u9019\u5169\u7a2e\u985e\u578b\u4e4b\u9593\u7684\u96fb\u8a71\u8a9e\u97f3\uff0c\u4e3b\u8981\u91dd\u5c0d\u5ba2\u670d\u8207\u5ba2\u6236\u7684\u96fb\u8a71\u9304\u97f3\uff0c\u56e0\u6b64\u5728\u4e00\u822c\u60c5 \u4f4d\uff0c\u4ee5\u53ca\u518d\u9032\u884c GMM-HMM \u8a9e\u8005\u5206\u7fa4\u662f\u5426\u6709\u52a9\u65bc\u63d0\u5347\u81ea\u52d5\u8a9e\u8005\u5206\u6bb5\u6a19\u8a18\u7684\u53ec\u56de\u7387\u4ee5 \u53ca\u7cbe\u6e96\u7387\u3002 \u5c07\u9019\u4e9b\u7247\u6bb5\u4e32\u63a5\u6210\u4e00\u500b\u65b0\u7684\u97f3\u6a94\uff0c\u76ee\u7684\u662f\u70ba\u4e86\u4e4b\u5f8c\u5728\u9032\u884c\u81ea\u52d5\u8a9e\u8005\u5206\u6bb5\u6a19\u8a18\u6642\u53ef\u4ee5\u5c07\u6642 \u9593\u6975\u77ed\u7684\u6bb5\u843d(\u4ecb\u65bc 0 \u81f3 1 \u79d2)\u6a19\u793a\u51fa\u4f86\u3002 \u8a0a\u7247\u6bb5\u7684\u9577\u5ea6\u8207\u91cd\u758a\u6642\u9593\u662f\u53ef\u8abf\u6574\u7684\uff0c\u5728\u7b2c\u56db\u7bc0\u7684\u5be6\u9a57\u6211\u6703\u63cf\u8ff0\u4e0d\u540c\u7684\u9577\u5ea6\u8207\u91cd\u758a\u6642\u9593 \u5c0d\u6211\u5011\u7684\u7cfb\u7d71\u6703\u9020\u6210\u751a\u9ebc\u6a23\u7684\u5f71\u97ff\u3002 \u5716\u516d\u3001\u91cd\u758a\u5340\u57df\u7684\u8a9e\u8005\u5206\u7fa4\u554f\u984c \u8a72\u5982\u4f55\u6c7a\u5b9a\u91cd\u758a\u90e8\u5206\u7684\u985e\u5225\u52e2\u5fc5\u6703\u5c0d\u5be6\u9a57\u7d50\u679c\u7522\u751f\u5f71\u97ff\uff0c\u56e0\u6b64\u6211\u5011\u7528\u4e00\u500b\u7c21\u55ae\u4e26\u4e14 \u6bb5\uff0c\u5229\u7528\u5b83\u5011\u4f86\u8a13\u7df4\u65b0\u7684\u8a9e\u8005\u6a21\u578b\uff0c\u4e4b\u5f8c\u518d\u7528 GMM-HMM \u9032\u884c\u8a9e\u8005\u5206\u7fa4\uff0c\u6982\u5ff5\u5982\u5716 \u516b\u6240\u793a\uff0c\u7531\u7b2c 20 \u8f2a\u7684\u5f37\u5236\u5c0d\u4f4d\u7d50\u679c\u8a13\u7df4\u51fa\u65b0\u7684\u8a9e\u8005\u6a21\u578b\uff0c\u63a5\u8457\u5c0d\u6574\u7684\u4e32\u63a5\u97f3\u6a94\u505a GMM-HMM \u8def\u5f91\u89e3\u78bc(Decode) \u3002\u548c\u5f37\u5236\u5c0d\u4f4d\u4e0d\u540c\u7684\u662f\uff0c\u5229\u7528 GMM-HMM \u9032\u884c\u8a9e\u8005 \u6bd4\u7b2c\u4e8c\u6b21\u5f37\u5236\u5c0d\u4f4d\u5f97\u5230\u7684\u7d50\u679c\u9084\u8981\u5dee\uff0c\u6211\u5011\u53ef\u4ee5\u5408\u7406\u63a8\u6e2c\u7d93\u904e\u7b2c\u4e8c\u6b21\u5f37\u5236\u5c0d\u4f4d\u5f8c\u5c31\u4e0d \u5716\u5341\u4e94\u3001\u7b2c\u4e8c\u6b21\u5f37\u5236\u5c0d\u4f4d 20 \u8f2a\u7684 F-Score\u3002 \u8a9e\u8005\u5206\u6bb5\u6a19\u8a18\u4e0d\u6703\u518d\u6709\u986f\u8457\u7684\u9032\u6b65\u3002 \u6700\u5f8c\u6211\u5011\u6703\u7531\u53ec\u56de\u7387\u4ee5\u53ca\u7cbe\u6e96\u7387\u6c42\u5f97 F-Score\uff0c\u4f86\u505a\u70ba\u6311\u9078\u6bd4\u8f03\u5404\u7a2e\u97f3\u8a0a\u7247\u6bb5\u7684\u9577\u5ea6\u4ee5 \u6211\u5011\u4f7f\u7528\u7684\u8cc7\u6599\u5eab\u662f\u7531\u4e2d\u570b\u4fe1\u8a17(China Trust)\u63d0\u4f9b\u4e4b 100 \u6bb5\u5ba2\u670d\u548c\u5ba2\u6236\u7684\u96fb\u8a71 \u53ca\u91cd\u758a\u6642\u9593\u7684\u8a55\u4f30\u65b9\u5f0f\uff0c\u5176\u8a08\u7b97\u7684\u516c\u5f0f\u5982(5)\u3002 \u9700\u8981\u518d\u9032\u884c\u7b2c\u4e8c\u6b21 GMM-HMM \u8a9e\u8005\u5206\u7fa4\u4f86\u5f97\u5230\u66f4\u7d30\u788e\u7684\u8a9e\u8005\u5340\u6bb5\u3002 \u6211\u5011\u4e5f\u767c\u73fe\u5728\u591a\u7d44\u97f3\u8a0a\u7247\u6bb5\u7684\u9577\u5ea6\u8207\u91cd\u758a\u6642\u9593\u7684\u7d44\u5408\u5c0d\u81ea\u52d5\u8a9e\u8005\u5206\u6bb5\u6a19\u8a18\u7684\u7d50\u679c \u8a9e\u97f3\uff0c\u6bcf\u6bb5\u96fb\u8a71\u8a9e\u97f3\u7684\u53d6\u6a23\u7387\u70ba 16,000 Hz\uff0c\u4e14\u90fd\u53ea\u5305\u542b\u5169\u4f4d\u8a9e\u8005\uff0c\u4e26\u7531\u5ba2\u670d\u5148\u958b\u59cb\u5c0d \u4ee5\u4e0b\u70ba\u672c\u8ad6\u6587\u7684\u7d50\u69cb\u8aaa\u660e\uff1a\u6211\u5011\u5c07\u5728\u7b2c\u4e8c\u7bc0\u4e2d\u4ecb\u7d39\u96d9\u8a9e\u8005\u4e4b\u81ea\u52d5\u8a9e\u8005\u5206\u6bb5\u6a19\u8a18\u7684 \u7cfb\u7d71\u67b6\u69cb\uff1b\u5728\u7b2c\u4e09\u7bc0\u4e2d\u4ecb\u7d39\u6211\u5011\u4f7f\u7528\u7684\u8cc7\u6599\u96c6\uff0c\u4ee5\u53ca\u8a55\u4f30\u81ea\u52d5\u8a9e\u8005\u5206\u6bb5\u6a19\u8a18\u7684\u6a19\u6e96\uff1b \u76f4\u89ba\u7684\u65b9\u5f0f\u4f86\u89e3\u6c7a\u9019\u500b\u554f\u984c\uff1a\u5047\u8a2d\u524d\u4e00\u500b i-vector \u53eb\u505a i\uff0c\u843d\u65bc C1 \u7fa4\u5167\uff1b\u5f8c\u4e00\u500b i-vector \u53eb\u505a j\uff0c\u843d\u65bc C2 \u7fa4\u5167\uff1b\u91cd\u758a\u90e8\u5206\u5beb\u4f5c Si\u2229j\uff0c\u60f3\u6cd5\u5c31\u662f\u627e\u51fa\u8f03\u9ad8\u7684 PLDA \u5206\u6578\uff0c\u8868\u793a\u8207\u54ea \u5206\u7fa4\u4e0d\u53ea\u6703\u6539\u8b8a\u8a9e\u8005\u5340\u6bb5\u7684\u7bc4\u570d\uff0c\u4e5f\u6703\u6539\u8b8a\u8a9e\u8005\u5340\u6bb5\u7684\u6578\u91cf\u3002 \u8a71\uff0c\u5e73\u5747\u9577\u5ea6\u70ba 4 \u5206 57 \u79d2\u3002\u5176\u4e2d\uff0c\u8cc7\u6599\u5eab\u4e26\u672a\u63d0\u4f9b\u6bcf\u6bb5\u8a9e\u97f3\u4e4b\u5ba2\u670d\u548c\u5ba2\u6236\u7684\u8a9e\u8005\u8cc7 \u6599\u3002 F-Score = 2 * \u4e26\u6c92\u6709\u592a\u5927\u7684\u5dee\u5225\uff0c\u4e0d\u904e\u7531\u65bc i-vector \u7684\u62bd\u53d6\u6642\u9593\u8207\u97f3\u8a0a\u7247\u6bb5\u7684\u6578\u91cf\u5448\u6b63\u76f8\u95dc\uff0c\u6240\u4ee5\u5efa * + (5) \u8b70\u53ef\u4ee5\u4f7f\u7528\u6642\u9593\u8f03\u9577\uff0c\u4e26\u4e14\u91cd\u758a\u6642\u9593\u8f03\u77ed\u7684\u97f3\u8a0a\u7247\u6bb5\u4f86\u9032\u884c\u81ea\u52d5\u8a9e\u8005\u5206\u6bb5\u6a19\u8a18\u3002</td></tr><tr><td>\u6cc1\u4e0b\u53ea\u6709\u5ba2\u670d\u8207\u5ba2\u6236\u5169\u4f4d\u8a9e\u8005\u3002 \u7b2c\u56db\u7bc0\u5c07\u8aaa\u660e\u5be6\u9a57\u7d50\u679c\u8207\u6578\u64da\u5206\u6790\uff1b\u5728\u7b2c\u4e94\u7bc0\u6211\u5011\u5c07\u70ba\u672c\u6587\u505a\u7d50\u8ad6\u3002 \u4e00\u7fa4\u5c31\u8d8a\u50cf\uff0c\u4f9d\u7167\u9019\u6a23\u7684\u60f3\u6cd5\u4f86\u6c7a\u5b9a\u91cd\u758a\u90e8\u5206\u7684\u5206\u7fa4\u3002 \u70ba\u4e86\u62bd\u53d6\u6885\u723e\u5012\u983b\u8b5c\u4fc2\u6578\u8207 i-vector\uff0c\u6bcf\u6bb5\u8a9e\u97f3\u6703\u5148\u964d\u4f4e\u53d6\u6a23\u7387\u70ba 8,000 Hz\u3002\u5176\u4e2d\uff0c \u5716\u5341\u4e8c\u3001\u5e73\u5747\u53ec\u56de\u7387\u7684\u6bd4\u8f03\u3002</td></tr><tr><td>\u4e00\u822c\u5728\u8655\u7406\u81ea\u52d5\u8a9e\u8005\u5206\u6bb5\u6a19\u8a18\u7684\u554f\u984c\u6703\u6d89\u53ca\u4e09\u500b\u6b65\u9a5f\uff1a1)\u5c07\u9304\u97f3\u8cc7\u6599\u5207\u5272\u6210\u8a31\u591a \u97f3\u8a0a\u7247\u6bb5\uff0c\u6211\u5011\u5e0c\u671b\u5728\u6bcf\u4e00\u500b\u97f3\u8a0a\u7247\u6bb5\u5167\u53ea\u5305\u542b\u4e00\u4f4d\u8a9e\u8005\u7684\u8072\u97f3\uff1b2)\u5c0d\u5207\u5272\u597d\u7684\u97f3\u8a0a \u7247\u6bb5\u9032\u884c\u8a9e\u8005\u5206\u7fa4\uff0c\u9019\u662f\u81ea\u52d5\u8a9e\u8005\u5206\u6bb5\u6a19\u8a18\u7684\u554f\u984c\u4e2d\u6700\u70ba\u95dc\u9375\u7684\u6b65\u9a5f\u3002\u50b3\u7d71\u7684\u81ea\u52d5\u8a9e \u8005\u5206\u6bb5\u6a19\u8a18\u662f\u8655\u7406\u672a\u77e5\u8a9e\u8005\u6578\u76ee\u7684\u8a9e\u97f3\u7d00\u9304\uff0c\u6240\u4ee5\u5728\u91dd\u5c0d\u8a9e\u8005\u5206\u7fa4\u7684\u554f\u984c\u4e2d\uff0c\u6700\u9996\u8981 P 1 = PLDA(i, Q 1 ) (1) \u6bcf\u4e00\u97f3\u6846\u7684\u9577\u5ea6\u70ba 32 ms\uff0c\u800c\u97f3\u6846\u4f4d\u79fb\u7684\u9577\u5ea6\u5247\u70ba 10 ms\u3002\u56e0\u70ba\u4e0a\u4e0b\u6587\u8cc7\u8a0a(Contextual \u56db\u3001\u5be6\u9a57\u7d50\u679c \u516d\u3001\u53c3\u8003\u6587\u737b \u4e8c\u3001\u96d9\u8a9e\u8005\u4e4b\u81ea\u52d5\u8a9e\u8005\u5206\u6bb5\u6a19\u8a18\u7cfb\u7d71\u67b6\u69cb \u96d9\u8a9e\u8005\u4e4b\u81ea\u52d5\u8a9e\u8005\u5206\u6bb5\u6a19\u8a18\u7684\u7cfb\u7d71\u6d41\u7a0b\u5982\u5716\u4e00\u6240\u793a\uff0c\u4e3b\u8981\u5206\u70ba\u56db\u500b\u90e8\u5206: \u7b2c\u4e00\u90e8\u5206 \u8ca0\u8cac\u5207\u5272\u7247\u6bb5(Segmentation) \uff0c\u5c07\u96fb\u8a71\u8a9e\u97f3\u5207\u5272\u6210\u8a31\u591a\u53ea\u5305\u542b\u4e00\u4f4d\u8a9e\u8005\u7684\u8aaa\u8a71\u7247\u6bb5\uff0c\u4e4b P 2 = PLDA(j, Q 2 ) (2) Information)\u8f03\u7121\u95dc\u65bc\u8a9e\u8005\u7279\u6027\uff0c\u6885\u723e\u5012\u983b\u8b5c\u4fc2\u6578\u53ea\u53bb\u975c\u614b\u7684 13 \u7dad\u90e8\u5206\uff0c\u800c\u4e0d\u8003\u616e\u52d5 2005\u30012006 \u4e26 Switchboard II-Phase 1-3 \u4ee5\u53ca Switchboard Cellular Part 1-2 \u4f86\u8a13\u7df4\u901a\u7528\u80cc \u5716\u627e\u51fa\u6700\u597d\u7684\u7d44\u5408\u3002\u9664\u4e86\u6bd4\u8f03 K \u5e73\u5747\u6f14\u7b97\u6cd5\u5f97\u5230\u7684\u7d50\u679c\u300120 \u8f2a\u5f37\u5236\u5c0d\u4f4d\u7684\u7d50\u679c\u3001 \u5716\u5341\u4e94\u3001\u97f3\u8a0a\u7247\u6bb5 4 \u79d2\uff0c\u91cd\u758a\u6642\u9593 0.5 \u79d2\u7684\u5f37\u5236\u5c0d\u4f4d\u548c\u7b2c\u4e8c\u6b21\u5f37\u5236\u5c0d\u4f4d\u7684\u7cbe\u6e96\u7387\u66f2 \u614b\u7684\u5dee\u7570\u3002\u800c\u5728\u62bd\u53d6 i-vector \u4ee5\u53ca\u8a08\u7b97 PLDA \u5206\u6578\u7684\u90e8\u5206\uff0c\u6211\u5011\u4ee5 NIST SRE 2004\u3001 \u6211\u5011\u91dd\u5c0d\u4e0d\u540c\u66f2\u97f3\u8a0a\u7247\u6bb5\u7684\u9577\u5ea6\u4ee5\u53ca\u91cd\u758a\u6642\u9593\u505a\u8a9e\u8005\u7684\u81ea\u52d5\u5206\u6bb5\u6a19\u8a18\u7684\u5be6\u9a57\uff0c\u4f01 [1] D.</td></tr><tr><td>\u7684\u554f\u984c\u662f\u300c\u7a76\u7adf\u6709\u5e7e\u4f4d\u8a9e\u8005\u300d\u3002 \u6700\u5ee3\u70ba\u4eba\u77e5\u7684\u65b9\u6cd5\u662f\uff0c\u6211\u5011\u5148\u5047\u8a2d\u4e00\u500b\u8db3\u5920\u591a\u7684\u8a9e\u8005\u6578 \u5f8c \u5c07 \u9019 \u4e9b \u8a9e \u8005 \u7247 \u6bb5 \u4e32 \u63a5 \u6210 \u4e00 \u500b \u4e32 \u63a5 \u97f3 \u6a94 \uff1b \u7b2c \u4e8c \u90e8 \u5206 \u70ba \u7279 \u5fb5 \u53c3 \u6578 \u7684 \u62bd \u53d6 ( Feature \u7531 K \u5e73\u5747\u6f14\u7b97\u6cd5\u6211\u5011\u5f97\u5230\u4e86\u7b2c\u4e00\u968e\u6bb5\u81ea\u52d5\u8a9e\u8005\u5206\u6bb5\u6a19\u8a18\u7684\u7d50\u679c\u3002\u7136\u800c\uff0c\u4ee5\u97f3\u8a0a\u7247 \u5716\u516b\u3001\u4f7f\u7528 GMM-HMM \u9032\u884c\u8a9e\u8005\u5206\u7fa4 \u666f\u6a21\u578b\u3001\u5168\u8b8a\u7570(Total Variability)\u6a21\u578b\u4ee5\u53ca PLDA \u6a21\u578b\u3002\u5176\u4e2d\uff0c\u901a\u7528\u80cc\u666f\u6a21\u578b\u7684\u6210\u5206 GMM-HMM \u8a9e\u8005\u5206\u7fa4\u7684\u7d50\u679c\u4e4b\u5916\uff0c\u6211\u9084\u60f3\u77e5\u9053\u5229\u7528 GMM-HMM \u9032\u884c\u8a9e\u8005\u5206\u7fa4\u5f97\u5230\u7d30 \u7dda\u3002</td></tr><tr><td>132</td></tr></table>", | |
| "text": "The 2017 Conference on Computational Linguistics and Speech Processing ROCLING 2017, pp. 119-135 \uf0d3 The Association for Computational Linguistics and Chinese Language Processing\u76f8\u4f3c\u77e9\u9663\u3002\u5728\u9019\u500b\u77e9\u9663\u4e2d\uff0c\u6211\u5011\u53ef\u4ee5\u67e5\u8a62\u5230\u6240\u6709\u5169\u5169\u4e0d\u540c\u7684 i-vector \u4e4b\u9593\u7684 PLDA \u5206\u6578\uff0c \u4e26\u5229\u7528\u6b64\u77e9\u9663\uff0c\u85c9\u8457 K \u5e73\u5747\u6f14\u7b97\u6cd5\u4f86\u5c0d\u8a9e\u8005\u5206\u7fa4\uff0c\u800c\u5f97\u5230\u7b2c\u4e00\u968e\u6bb5\u7684\u81ea\u52d5\u8a9e\u8005\u5206\u6bb5\u6a19 \u8a18\u7d50\u679c\uff0c\u5176\u6b65\u9a5f\u5982\u4e0b\uff1a1)\u5148\u96a8\u6a5f\u9078\u53d6\u5169\u500b i-vector \u7576\u4f5c\u5169\u7fa4\u7fa4\u5fc3\uff0c\u6211\u5011\u5c07\u9019\u5169\u7fa4\u7a31\u70ba \u79d2\uff0c\u91cd\u758a\u79d2\u6578\u70ba 0.5 \u79d2\u7684\u6642\u5019\u53ef\u4ee5 \u5f97\u5230\u6700\u597d\u7684 F-Score(91.54 %) \uff0c\u4e0d\u904e\u6574\u4f86\u4f86\u8aaa\uff0c\u9664\u4e86\u9577\u5ea6\u70ba 1 \u79d2\uff0c\u91cd\u758a\u79d2\u6578\u70ba 0.5 \u79d2 Reynolds, T. Quatieri, and R. Dunn, \"Speaker verification using adapted gaussian mixture models,\" Dig. Sig. Proc., 2000." | |
| } | |
| } | |
| } | |
| } |