| { | |
| "paper_id": "O18-1001", | |
| "header": { | |
| "generated_with": "S2ORC 1.0.0", | |
| "date_generated": "2023-01-19T08:09:56.505811Z" | |
| }, | |
| "title": "Study and Implementation on Digit-related Speaker Verification", | |
| "authors": [ | |
| { | |
| "first": "Chung-Hung", | |
| "middle": [], | |
| "last": "\u5468\u5b97\u9d3b", | |
| "suffix": "", | |
| "affiliation": { | |
| "laboratory": "", | |
| "institution": "National Taiwan University", | |
| "location": {} | |
| }, | |
| "email": "" | |
| }, | |
| { | |
| "first": "", | |
| "middle": [], | |
| "last": "Chou", | |
| "suffix": "", | |
| "affiliation": { | |
| "laboratory": "", | |
| "institution": "National Taiwan University", | |
| "location": {} | |
| }, | |
| "email": "" | |
| }, | |
| { | |
| "first": "Jyh-Shing", | |
| "middle": [ | |
| "Roger" | |
| ], | |
| "last": "\u5f35\u667a\u661f", | |
| "suffix": "", | |
| "affiliation": { | |
| "laboratory": "", | |
| "institution": "Normal University", | |
| "location": { | |
| "country": "Taiwan" | |
| } | |
| }, | |
| "email": "" | |
| }, | |
| { | |
| "first": "", | |
| "middle": [], | |
| "last": "Jang", | |
| "suffix": "", | |
| "affiliation": { | |
| "laboratory": "", | |
| "institution": "Normal University", | |
| "location": { | |
| "country": "Taiwan" | |
| } | |
| }, | |
| "email": "[email protected]" | |
| }, | |
| { | |
| "first": "Shan-Wen", | |
| "middle": [], | |
| "last": "\u856d\u5584\u6587", | |
| "suffix": "", | |
| "affiliation": { | |
| "laboratory": "\u4e2d\u83ef\u96fb\u4fe1\u7814\u7a76\u9662 Chunghwa Telecom Laboratories, Taoyuan", | |
| "institution": "", | |
| "location": { | |
| "country": "Taiwan" | |
| } | |
| }, | |
| "email": "" | |
| }, | |
| { | |
| "first": "", | |
| "middle": [], | |
| "last": "Hsiao", | |
| "suffix": "", | |
| "affiliation": { | |
| "laboratory": "\u4e2d\u83ef\u96fb\u4fe1\u7814\u7a76\u9662 Chunghwa Telecom Laboratories, Taoyuan", | |
| "institution": "", | |
| "location": { | |
| "country": "Taiwan" | |
| } | |
| }, | |
| "email": "[email protected]" | |
| } | |
| ], | |
| "year": "", | |
| "venue": null, | |
| "identifiers": {}, | |
| "abstract": "Speaker recognition is an important biometric identification method. The biggest advantage of using such method is the simple requirement of its hardware, which only consists of a microphone. Therefore, it is widely implemented in mobile phones and call centers. The purpose of this thesis is to create a text-related speaker verification system, for which we", | |
| "pdf_parse": { | |
| "paper_id": "O18-1001", | |
| "_pdf_hash": "", | |
| "abstract": [ | |
| { | |
| "text": "Speaker recognition is an important biometric identification method. The biggest advantage of using such method is the simple requirement of its hardware, which only consists of a microphone. Therefore, it is widely implemented in mobile phones and call centers. The purpose of this thesis is to create a text-related speaker verification system, for which we", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Abstract", | |
| "sec_num": null | |
| } | |
| ], | |
| "body_text": [ | |
| { | |
| "text": "conduct three different approaches to analyze their result: dynamic time warping compares the differences between the MFCCs for digits at registration and digits at testing after applying forced alignment; sentence-level uses cosine similarity or PLDA to rate the two groups of ivector retrieved from the audios at registration and testing respectively; digit-level uses cosine similarity or PLDA to rate each i-vector of every digit in the audios after applying forced alignment.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "", | |
| "sec_num": null | |
| }, | |
| { | |
| "text": "Keywords: Speaker Verification, Forced Alignment, DTW, i-vector, PLDA. ", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "\u95dc\u9375\u8a5e\uff1a\u8a9e\u8005\u9a57\u8b49\uff0c\u5f37\u5236\u5c0d\u9f4a\uff0c\u52d5\u614b\u6642\u9593\u626d\u66f2\uff0ci-vector\uff0c\u6a5f\u7387\u7dda\u6027\u5224\u5225\u5206\u6790", | |
| "sec_num": null | |
| } | |
| ], | |
| "back_matter": [], | |
| "bib_entries": { | |
| "BIBREF0": { | |
| "ref_id": "b0", | |
| "title": "Fundamentals of speech recognition", | |
| "authors": [ | |
| { | |
| "first": "Lawrence", | |
| "middle": [ | |
| "R" | |
| ], | |
| "last": "Rabiner", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Biing-Hwang", | |
| "middle": [], | |
| "last": "Juang", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 1993, | |
| "venue": "", | |
| "volume": "14", | |
| "issue": "", | |
| "pages": "", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "Rabiner, Lawrence R., and Biing-Hwang Juang. Fundamentals of speech recognition. Vol. 14. Englewood Cliffs: PTR Prentice Hall, 1993..", | |
| "links": null | |
| }, | |
| "BIBREF1": { | |
| "ref_id": "b1", | |
| "title": "End-to-end text-dependent speaker verification", | |
| "authors": [ | |
| { | |
| "first": "Georg", | |
| "middle": [], | |
| "last": "Heigold", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 2016, | |
| "venue": "2016 IEEE International Conference on", | |
| "volume": "", | |
| "issue": "", | |
| "pages": "", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "Heigold, Georg, et al. \"End-to-end text-dependent speaker verification.\" Acoustics, Speech and Signal Processing (ICASSP), 2016 IEEE International Conference on. IEEE, 2016.", | |
| "links": null | |
| }, | |
| "BIBREF2": { | |
| "ref_id": "b2", | |
| "title": "Voice recognition algorithms using mel frequency cepstral coefficient (MFCC) and dynamic time warping (DTW) techniques", | |
| "authors": [ | |
| { | |
| "first": "Lindasalwa", | |
| "middle": [], | |
| "last": "Muda", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Mumtaj", | |
| "middle": [], | |
| "last": "Begam", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Irraivan", | |
| "middle": [], | |
| "last": "Elamvazuthi", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 2010, | |
| "venue": "", | |
| "volume": "", | |
| "issue": "", | |
| "pages": "", | |
| "other_ids": { | |
| "arXiv": [ | |
| "arXiv:1003.4083" | |
| ] | |
| }, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "Muda, Lindasalwa, Mumtaj Begam, and Irraivan Elamvazuthi. \"Voice recognition algorithms using mel frequency cepstral coefficient (MFCC) and dynamic time warping (DTW) techniques.\" arXiv preprint arXiv:1003.4083 (2010).", | |
| "links": null | |
| }, | |
| "BIBREF3": { | |
| "ref_id": "b3", | |
| "title": "Speaker identification and verification using Gaussian mixture speaker models", | |
| "authors": [ | |
| { | |
| "first": "Douglas", | |
| "middle": [ | |
| "A" | |
| ], | |
| "last": "Reynolds", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 1995, | |
| "venue": "Speech communication", | |
| "volume": "17", | |
| "issue": "", | |
| "pages": "91--108", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "Reynolds, Douglas A. \"Speaker identification and verification using Gaussian mixture speaker models.\" Speech communication 17.1-2 (1995): 91-108.", | |
| "links": null | |
| }, | |
| "BIBREF4": { | |
| "ref_id": "b4", | |
| "title": "Speaker verification using adapted Gaussian mixture models", | |
| "authors": [ | |
| { | |
| "first": "Douglas", | |
| "middle": [ | |
| "A" | |
| ], | |
| "last": "Reynolds", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "F", | |
| "middle": [], | |
| "last": "Thomas", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Robert", | |
| "middle": [ | |
| "B" | |
| ], | |
| "last": "Quatieri", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "", | |
| "middle": [], | |
| "last": "Dunn", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 2000, | |
| "venue": "Digital signal processing", | |
| "volume": "10", | |
| "issue": "", | |
| "pages": "19--41", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "Reynolds, Douglas A., Thomas F. Quatieri, and Robert B. Dunn. \"Speaker verification using adapted Gaussian mixture models.\" Digital signal processing 10.1-3 (2000): 19-41.", | |
| "links": null | |
| }, | |
| "BIBREF5": { | |
| "ref_id": "b5", | |
| "title": "Joint factor analysis versus eigenchannels in speaker recognition", | |
| "authors": [ | |
| { | |
| "first": "Patrick", | |
| "middle": [], | |
| "last": "Kenny", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 2007, | |
| "venue": "IEEE Transactions on Audio, Speech, and Language Processing", | |
| "volume": "15", | |
| "issue": "", | |
| "pages": "1435--1447", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "Kenny, Patrick, et al. \"Joint factor analysis versus eigenchannels in speaker recognition.\" IEEE Transactions on Audio, Speech, and Language Processing 15.4 (2007): 1435-1447.", | |
| "links": null | |
| }, | |
| "BIBREF6": { | |
| "ref_id": "b6", | |
| "title": "Front-end factor analysis for speaker verification", | |
| "authors": [ | |
| { | |
| "first": "Najim", | |
| "middle": [], | |
| "last": "Dehak", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 2011, | |
| "venue": "IEEE Transactions on Audio, Speech, and Language Processing", | |
| "volume": "19", | |
| "issue": "", | |
| "pages": "788--798", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "Dehak, Najim, et al. \"Front-end factor analysis for speaker verification.\" IEEE Transactions on Audio, Speech, and Language Processing 19.4 (2011): 788-798.", | |
| "links": null | |
| }, | |
| "BIBREF7": { | |
| "ref_id": "b7", | |
| "title": "Fisher discriminant analysis with kernels", | |
| "authors": [ | |
| { | |
| "first": "Sebastian", | |
| "middle": [], | |
| "last": "Mika", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 1999, | |
| "venue": "Proceedings of the 1999 IEEE signal processing society workshop", | |
| "volume": "", | |
| "issue": "", | |
| "pages": "", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "Mika, Sebastian, et al. \"Fisher discriminant analysis with kernels.\" Neural networks for signal processing IX, 1999. Proceedings of the 1999 IEEE signal processing society workshop..", | |
| "links": null | |
| }, | |
| "BIBREF9": { | |
| "ref_id": "b9", | |
| "title": "Probabilistic linear discriminant analysis for inferences about identity", | |
| "authors": [ | |
| { | |
| "first": "", | |
| "middle": [], | |
| "last": "Prince", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "J", | |
| "middle": [ | |
| "D" | |
| ], | |
| "last": "Simon", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "James", | |
| "middle": [ | |
| "H" | |
| ], | |
| "last": "Elder", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 2007, | |
| "venue": "IEEE 11th International Conference on", | |
| "volume": "", | |
| "issue": "", | |
| "pages": "", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "Prince, Simon JD, and James H. Elder. \"Probabilistic linear discriminant analysis for inferences about identity.\" Computer Vision, 2007. ICCV 2007. IEEE 11th International Conference on. IEEE, 2007.", | |
| "links": null | |
| }, | |
| "BIBREF10": { | |
| "ref_id": "b10", | |
| "title": "Analysis of i-vector length normalization in speaker recognition systems", | |
| "authors": [ | |
| { | |
| "first": "Daniel", | |
| "middle": [], | |
| "last": "Garcia-Romero", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Carol", | |
| "middle": [ | |
| "Y" | |
| ], | |
| "last": "Espy-Wilson", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 2011, | |
| "venue": "Twelfth Annual Conference of the International Speech Communication Association", | |
| "volume": "", | |
| "issue": "", | |
| "pages": "", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "Garcia-Romero, Daniel, and Carol Y. Espy-Wilson. \"Analysis of i-vector length normalization in speaker recognition systems.\" Twelfth Annual Conference of the International Speech Communication Association. 2011", | |
| "links": null | |
| }, | |
| "BIBREF11": { | |
| "ref_id": "b11", | |
| "title": "Spear: An open source toolbox for speaker recognition based on Bob", | |
| "authors": [ | |
| { | |
| "first": "Elie", | |
| "middle": [], | |
| "last": "Khoury", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Laurent", | |
| "middle": [ | |
| "El" | |
| ], | |
| "last": "Shafey", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "S\u00e9bastien", | |
| "middle": [], | |
| "last": "Marcel", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 2014, | |
| "venue": "Acoustics, Speech and Signal Processing", | |
| "volume": "", | |
| "issue": "", | |
| "pages": "", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "Khoury, Elie, Laurent El Shafey, and S\u00e9bastien Marcel. \"Spear: An open source toolbox for speaker recognition based on Bob.\" Acoustics, Speech and Signal Processing (ICASSP), 2014 IEEE International Conference on. IEEE, 2014.", | |
| "links": null | |
| } | |
| }, | |
| "ref_entries": { | |
| "TABREF0": { | |
| "type_str": "table", | |
| "content": "<table/>", | |
| "html": null, | |
| "text": "The 2018 Conference on Computational Linguistics and Speech Processing ROCLING 2018, pp. 1-15 \u00a9The Association for Computational Linguistics and Chinese Language Processing", | |
| "num": null | |
| }, | |
| "TABREF1": { | |
| "type_str": "table", | |
| "content": "<table><tr><td>\u4e8c\u3001\u8a9e\u8005\u9a57\u8b49\u7c21\u4ecb \u9019\u8aaa\u660e\u4e86\u70ba\u4ec0\u9ebc\u4f7f\u7528\u5e73\u5747\u8ddd\u96e2\u932f\u8aa4\u7387\u6bd4\u8f03\u4e0d\u6703\u7a69\u5b9a\u4e0b\u964d\u7684\u539f\u56e0\u3002 \u5716\u4e09\u3001\u4e0d\u4f7f\u7528 feature scaling \u6642\u8a3b\u518a\u97f3\u6a94\u6578\u5c0d\u65bc\u52d5\u614b\u6642\u9593\u626d\u66f2\u4e4b\u6bd4\u8f03 \u8a00\uff0c\u6578\u5b57\u7d1a\u8a9e\u8005\u9a57\u8b49\u5728\u8a3b\u518a\u6642\u5373\u53ef\u5f97\u5230 10 \u7d44\u4ee3\u8868\u8a9e\u8005\u7684 i-vector(\u5206\u5225\u70ba 1\u30010\u30012\u30019\u3001 \u65bc\u8a9e\u53e5\u7d1a\u3002\u6700\u5f8c\u4ea6\u53ef\u4ee5\u767c\u73fe\u6574\u9ad4\u6700\u597d\u7684\u8868\u73fe\u5927\u6982\u843d\u5728 PLDA \u4f7f\u7528\u5e73\u5747\u6642\uff0c\u5176\u932f\u8aa4\u7387\u843d\u5728 \u97f3\u7279\u8cea\u4e0d\u540c\u9020\u6210\uff0c\u539f\u56e0\u662f\u6211\u5011\u5728\u4eba\u5de5\u6a19\u8a18\u97f3\u6a94\u6642\u767c\u73fe\u6709\u4e9b\u6578\u5b57\u807d\u8d77\u4f86\u975e\u5e38\u50cf\u662f\u96dc\u97f3\uff0c\u6b64 \u5b57\u5c0d\u65bc\u932f\u8aa4\u7387\u4e26\u7121\u660e\u986f\u52a9\u76ca\uff0c\u6b64\u5916\u96d6\u7136\u62ff\u6389\u9019 2 \u500b\u932f\u8aa4\u7387\u6700\u9ad8\u7684\u6578\u5b57\u5c0d\u65bc\u932f\u8aa4\u7387\u7121\u660e\u986f \u5b57\u7d1a\u8a9e\u8005\u9a57\u8b49\u7cfb\u7d71\uff0c\u6211\u5011\u767c\u73fe\u8a9e\u53e5\u7d1a\u8a9e\u8005\u9a57\u8b49\u7cfb\u7d71\u6548\u679c\u660e\u986f\u6bd4\u6578\u5b57\u7d1a\u8a9e\u8005\u9a57\u8b49\u7cfb\u7d71\u4f86\u5f97</td></tr><tr><td>\u8fd1\u5e74\u4f86\u7531\u65bc\u667a\u6167\u578b\u624b\u6a5f\u53ca\u901a\u8a0a\u7db2\u8def\u7684\u666e\u53ca\uff0c\u904e\u53bb\u8a31\u591a\u7121\u6cd5\u9054\u6210\u7684\u4e8b\u60c5\u5df2\u7d93\u9010\u6f38\u8b8a\u5f97\u53ef\u884c\uff0c \u4f8b\u5982\uff1a\u904e\u53bb\u4eba\u5011\u90fd\u662f\u900f\u904e\u73fe\u91d1\u9032\u884c\u4ea4\u6613\uff0c\u4f46\u73fe\u5728\u56e0\u70ba\u667a\u6167\u578b\u624b\u6a5f\u7684\u666e\u53ca\uff0c\u884c\u52d5\u652f\u4ed8\u53ca\u7db2 \u8def\u8cfc\u7269\u5df2\u7d93\u80fd\u9010\u6f38\u53d6\u4ee3\u904e\u53bb\u4eba\u5011\u50b3\u7d71\u7684\u4ea4\u6613\u884c\u70ba\u3002\u7136\u800c\u4e5f\u56e0\u70ba\u9019\u4e9b\u4e8b\u60c5\u5e36\u4f86\u8a31\u591a\u65b0\u7684\u554f \u984c\uff0c\u4ee5\u884c\u52d5\u652f\u4ed8\u53ca\u7db2\u8def\u8cfc\u7269\u800c\u8a00\uff0c\u554f\u984c\u5373\u662f\u8981\u5982\u4f55\u8fa8\u8a8d\u4f7f\u7528\u8005\u4e4b\u8eab\u4efd\uff0c\u9032\u800c\u78ba\u8a8d\u70ba\u672c\u4eba \u4f7f\u7528\u800c\u975e\u76dc\u7528\u4fbf\u6210\u70ba\u73fe\u5728\u975e\u5e38\u91cd\u8981\u7684\u4e00\u500b\u8ab2\u984c\u3002 \u4efb\u4f55\u7684\u7cfb\u7d71\u90fd\u9700\u8981\u8fa8\u8b58\u3001\u78ba\u8a8d\u4f7f\u7528\u8005\u4e4b\u8eab\u4efd\uff0c\u56e0\u70ba\u5982\u679c\u7121\u6cd5\u78ba\u8a8d\u4f7f\u7528\u8005\u5176\u8eab\u4efd\u5c0e\u81f4\u500b\u4eba \u8cc7\u6599\u5916\u6d29\u662f\u975e\u5e38\u56b4\u91cd\u7684\u4e8b\u60c5\uff0c\u4f8b\u5982\uff1a\u4fe1\u7528\u5361\u8cc7\u6599\u906d\u5230\u76dc\u7528\u3001\u5bb6\u88e1\u906d\u5230\u5c0f\u5077\u5165\u4fb5\u3001\u79c1\u5bc6\u7167 \u7247\u6d41\u51fa\u2026\u7b49\u3002\u50b3\u7d71\u6211\u5011\u5728\u8fa8\u8b58\u4f7f\u7528\u8005\u7684\u65b9\u6cd5\u591a\u70ba\uff1a\u9470\u5319\u3001\u9580\u7981\u5361\u3001\u5bc6\u78bc\u9396\u2026\u7b49\uff0c\u4f46\u9019\u4e9b \u65b9\u6cd5\u5f80\u5f80\u90fd\u6703\u6709\u907a\u5931\u4e26\u4e14\u906d\u5230\u76dc\u7528\u7684\u98a8\u96aa\u3002\u8fd1\u5e74\u4f86\u7531\u65bc\u96fb\u8166\u904b\u7b97\u80fd\u529b\u7684\u5927\u5e45\u9032\u6b65\uff0c\u9678\u7e8c \u958b\u59cb\u6709\u4e0d\u5c11\u4eba\u63d0\u51fa\u5229\u7528\u4eba\u985e\u751f\u7406\u7279\u5fb5\u9032\u884c\u8fa8\u8b58\u4f7f\u7528\u8005\u8eab\u4efd\uff0c\u4f7f\u7528\u4eba\u985e\u751f\u7406\u7279\u5fb5\u53ef\u4ee5\u907f\u514d \u6389\u907a\u5931\u4e26\u906d\u76dc\u7528\u7684\u98a8\u96aa\u3002\u76ee\u524d\u5e02\u9762\u4e0a\u5e38\u898b\u7684\u5229\u7528\u4eba\u985e\u751f\u7406\u7279\u5fb5\u9a57\u8b49\u65b9\u6cd5\u6709\uff1a\u6307\u7d0b\u9a57\u8b49\u3001 \u81c9\u90e8\u9a57\u8b49\u3001\u8996\u7db2\u819c\u9a57\u8b49\u3001\u8072\u7d0b\u9a57\u8b49(\u8a9e\u8005\u9a57\u8b49)\u2026\u7b49\u3002 \u672c\u7814\u7a76\u76ee\u6a19\u70ba\u5efa\u7acb\u4e00\u5957\u6578\u5b57\u6587\u672c\u76f8\u95dc\u7684\u8a9e\u8005\u9a57\u8b49\u7cfb\u7d71\u4e26\u7528\u65bc\u884c\u52d5\u652f\u4ed8\uff0c\u56e0\u70ba\u884c\u52d5\u652f\u4ed8\u591a \u61c9\u7528\u5728\u667a\u6167\u578b\u624b\u6a5f\u4e14\u6bcf\u53f0\u667a\u6167\u578b\u624b\u6a5f\u7686\u64c1\u6709\u9304\u97f3\u4e4b\u529f\u80fd\uff0c\u53c8\u9304\u88fd\u8072\u97f3\u70ba\u4e00\u4ef6\u5bb9\u6613\u9054\u6210\u4e4b \u4e8b\u60c5\uff0c\u56e0\u6b64\u6211\u5011\u5e0c\u671b\u80fd\u85c9\u7531\u8a9e\u8005\u9a57\u8b49\u4f86\u9054\u6210\u78ba\u8a8d\u4f7f\u7528\u8005\u4e4b\u8eab\u4efd\u3002 \u8a9e\u8005\u9a57\u8b49\u70ba\u4f9d\u64da\u8aaa\u8a71\u8005\u5ba3\u7a31\u4e4b\u8eab\u5206\u53ca\u5176\u8a9e\u97f3\u5167\u5bb9\u5224\u65b7\u662f\u5426\u5c6c\u5be6\uff0c\u6b64\u7a2e\u61c9\u7528\u60c5\u5883\u5e38\u7528\u65bc\uff1a \u884c\u52d5\u652f\u4ed8\u3001\u9580\u7981\u7cfb\u7d71...\u7b49\u3002\u53e6\u5916\u8a9e\u8005\u9a57\u8b49\u53c8\u53ef\u4f9d\u7167\u8aaa\u8a71\u8005\u8a9e\u97f3\u4e4b\u5167\u5bb9\u5206\u70ba\u4e09\u5927\u985e\uff0c\u5206\u5225 \u70ba\uff1a\u672c\u6587\u7368\u7acb(text-independent) \u3001\u672c\u6587\u76f8\u95dc(text-relative) \u3001\u672c\u6587\u76f8\u4f9d(text-dependent)\u3002 \u672c\u6587\u7368\u7acb\u70ba\u8a9e\u97f3\u5167\u5bb9\u53ef\u4ee5\u70ba\u4efb\u610f\u7684\uff0c\u800c\u672c\u6587\u76f8\u95dc\u70ba\u8a9e\u97f3\u5167\u5bb9\u5fc5\u9808\u5728\u67d0\u4e9b\u7bc4\u570d\u5167\uff0c\u4f8b\u5982\uff1a \u50c5\u80fd\u662f\u6578\u5b57\u6216\u8005\u984f\u8272\u2026\u7b49\uff0c\u800c \u672c\u6587\u76f8\u4f9d\u5247\u70ba\u5b8c\u5168\u9650\u5b9a\u8a9e\u97f3\u4e4b\u5167\u5bb9\uff0c\u4f8b\u5982\u50c5\u80fd\u662f Hey, Siri \u3001 OK, Google \u7b49\u906d\u9650\u5236\u4e4b\u8a9e\u97f3\u5167\u5bb9\u3002 \u5728\u8a9e\u97f3\u8a0a\u865f\u8655\u7406\u4e2d\u7279\u5fb5\u901a\u5e38\u6211\u5011\u6703\u4f7f\u7528\u6885\u723e\u983b\u7387\u5012\u8b5c\u4fc2\u6578(Mel-scale Frequency Cepstral Coefficients, MFCC)[1]\uff0c\u800c\u8a9e\u8005\u9a57\u8b49\u4e2d\u9664 End-to-End[2]\u5916\u4e5f\u591a\u534a\u662f\u5148\u62bd\u53d6 MFCC \u5f8c\u518d \u9032\u884c\u5176\u5b83\u8655\u7406\uff0c\u5176\u76f8\u95dc\u65b9\u6cd5\u6709\uff1a\u52d5\u614b\u6642\u9593\u626d\u66f2(Dynamic Time Warping, DTW)[3]\u3001\u9ad8 \u65af\u6df7\u5408\u6a21\u578b(Gaussian Mixture Model, GMM)[4]\u3001\u901a\u7528\u80cc\u666f\u6a21\u578b(Universal Background Model, UBM)[5]\u3001\u806f\u5408\u56e0\u5b50\u5206\u6790(Joint factor Analysis, JFA)[6]\u3001i-vector[7]\u3002\u53e6\u5916\u503c\u5f97 \u4e00\u63d0\u7684\u662f\uff0ci-vector \u5728\u8a13\u7df4\u7e3d\u9ad4\u8b8a\u7570\u77e9\u9663\u6642\u6703\u53d7\u5230\u4e0d\u540c\u901a\u9053\u7684\u5e72\u64fe\uff0c\u56e0\u6b64\u4e00\u822c\u6703\u4f7f\u7528\u7dda \u6027\u5224\u5225\u5206\u6790(Linear Discriminant Analysis, LDA)[8]\u9032\u884c\u4fe1\u9053\u88dc\u511f\u4ee5\u53ca\u4f7f\u7528\u6a5f\u7387\u7dda\u6027\u5224 \u5225\u5206\u6790(Probability Linear Discriminant Analysis, PLDA)[9, 10]\u9032\u884c\u8a55\u5206\u3002 \u4e09\u3001\u8a9e\u6599\u914d\u7f6e\u53ca\u8a55\u4f30\u6a19\u6e96 (\u4e00)\u8a9e\u6599\u4ecb\u7d39 \u672c\u7814\u7a76\u4e4b\u4f86\u6e90\u8a9e\u6599\u5206\u70ba\u5169\u5927\u985e\uff0c\u7b2c\u4e00\u985e\u70ba\u6211\u5011\u5728\u53f0\u5927\u8cc7\u5de5\u7cfb\u5f35\u667a\u661f\u6559\u6388\u5728\u8ab2\u5802\u4e0a\u6240\u641c\u96c6 \u7684\u9304\u97f3\uff0c\u7b2c\u4e8c\u985e\u70ba\u4ea4\u5927\u96fb\u6a5f\u7cfb\u738b\u9038\u5982\u6559\u6388\u63d0\u4f9b\u7d66\u6211\u5011\u7684\u9304\u97f3\uff0c\u4ee5\u4e0b\u6211\u5011\u5c07\u5c0d\u9019\u5169\u985e\u8a9e\u6599 \u9032\u884c\u8aaa\u660e\u3002 1. \u53f0\u5927\u5f35\u667a\u661f\u6559\u6388\u4e4b\u8a9e\u6599 \u672c\u8a9e\u6599\u5171\u6709 191 \u53d6\u6a23\u983b\u7387\u70ba\u55ae\u8072\u9053 16KHz\uff0c\u97f3\u8cea\u70ba 16bit\u3002\u6b64\u5916\u8a72\u8a9e\u6599\u5305\u542b\u6a19\u8a18\u6a94\u5c0d\u61c9\u5230\u6bcf\u500b\u97f3\u6a94\u4e2d\u5404 \u500b\u6578\u5b57\u767c\u97f3\u958b\u59cb\u8207\u7d50\u675f\u7684\u6642\u9593\u3002 (\u4e8c)\u8cc7\u6599\u914d\u7f6e \u6211\u5011\u5c07\u7b2c\u4e09\u7ae0\u7b2c\u4e00\u7bc0\u4e4b\u4e8c\u7d44\u8a9e\u6599\u62c6\u6210\u8a13\u7df4\u8cc7\u6599\u8207\u6e2c\u8a66\u8cc7\u6599\u3002\u8a13\u7df4\u8cc7\u6599\u90e8\u5206\u6703\u4f9d\u7167\u539f\u6027\u5225 \u6bd4\u4f8b\u5f9e\u53f0\u5927\u8a9e\u6599\u4e2d\u53d6 41 \u4f4d\u8a9e\u8005(\u5176\u4e2d\u7537\u751f 34 \u4f4d\u3001\u5973\u751f 7 \u4f4d) \u3001\u4ea4\u5927\u8a9e\u6599\u4e2d\u4e4b\u6240\u6709\u8a9e\u8005 (\u5176\u4e2d\u7537\u751f 50 \u4f4d\u3001\u5973\u751f 50 \u4f4d) \uff0c\u6e2c\u8a66\u8cc7\u6599\u90e8\u5206\u53d6\u5269\u9918\u4e4b\u53f0\u5927\u8a9e\u6599(\u5176\u4e2d\u7537\u751f 126 \u4f4d\u3001 \u5973\u751f 24 \u4f4d) \u3002 \u8a13\u7df4\u8cc7\u6599\u6703\u5168\u90e8\u7528\u65bc\u8a13\u7df4\u6a21\u578b\uff0c\u800c\u6e2c\u8a66\u8cc7\u6599\u6703\u5206\u70ba\u8a9e\u8005\u6b63\u78ba (\u63a5\u53d7) \u53ca\u8a9e\u8005\u932f\u8aa4 (\u62d2\u7d55) \u9019\u4e8c\u7a2e\u60c5\u6cc1\u3002\u73fe\u5728\u6211\u5011\u5047\u8a2d\u8a3b\u518a\u97f3\u6a94\u4f7f\u7528\u4e8c\u7d44\u97f3\u6a94\uff0c\u5247\u63a5\u53d7\u4e4b\u7b46\u6578\u6703\u5171\u6709 150 \u4f4d\u8a9e\u8005 \u53ef\u7528\u65bc\u9a57\u8b49\uff0c\u6240\u4ee5\u62d2\u7d55\u7b46\u6578\u5171\u6709 178800 \u7b46\u3002\u5176\u6b63\u78ba\u932f\u8aa4\u6bd4\u70ba 1200:178800 = 1:149\u3002 (\u4e09)\u8a55\u4f30\u6a19\u6e96 \u8a9e\u8005\u9a57\u8b49\u7684\u932f\u8aa4\u53ef\u4ee5\u5206\u6210\u932f\u8aa4\u63a5\u53d7(False Acceptance)\u53ca\u932f\u8aa4\u62d2\u7d55(False Rejection)\u4e8c \u985e\u3002\u932f\u8aa4\u63a5\u53d7\u70ba\u8072\u97f3\u7247\u6bb5\u4e26\u975e\u70ba\u5176\u5ba3\u7a31\u4e4b\u8a9e\u8005\u4f46\u7cfb\u7d71\u8aa4\u5224\u70ba\u662f\u5176\u5ba3\u7a31\u4e4b\u8a9e\u8005\uff0c\u6211\u5011\u7a31\u4e4b \u70ba\u932f\u8aa4\u63a5\u53d7\uff0c\u800c\u932f\u8aa4\u62d2\u7d55\u70ba\u8072\u97f3\u7247\u6bb5\u70ba\u5176\u5ba3\u7a31\u4e4b\u8a9e\u8005\u4f46\u7cfb\u7d71\u8aa4\u5224\u70ba\u4e0d\u662f\u5176\u5ba3\u7a31\u4e4b\u8a9e\u8005\uff0c \u6211\u5011\u7a31\u4e4b\u70ba\u932f\u8aa4\u62d2\u7d55\u3002 \u6211\u5011\u5e0c\u671b\u80fd\u5920\u5f9e\u5168\u90e8\u7684\u6e2c\u8a66\u4e2d\u627e\u4e00\u500b\u9580\u6abb\u503c\uff0c\u4f7f\u5f97\u932f\u8aa4\u63a5\u53d7\u6bd4\u4f8b(False Acceptance Rate, FAR)\u7b49\u65bc\u932f\u8aa4\u62d2\u7d55\u6bd4\u4f8b(\u8a18\u4e4b\u7d50\u679c\u3002 (\u3127)\u52d5\u614b\u6642\u9593\u626d\u66f2\u8a9e\u8005\u9a57\u8b49 \u6211\u5011\u4f9d\u7167\u7b2c\u4e09\u7ae0\u7b2c\u4e8c\u7bc0\u7684\u8cc7\u6599\u914d\u7f6e\u5c0d\u65bc\u6240\u6709\u6e2c\u8a66\u5229\u7528\u52d5\u614b\u6642\u9593\u626d\u66f2\u6bd4\u8f03\u8072\u97f3\u4e4b\u9593\u7684\u8ddd \u96e2\u3002\u5176\u5c0d\u4e00\u7b46\u6e2c\u8a66\u4e4b\u8a73\u7d30\u4f5c\u6cd5\uff0c\u6211\u5011\u5148\u4f7f\u7528\u4eba\u5de5\u6a19\u8a18(\u5f37\u5236\u5c0d\u9f4a)\u7cbe\u6e96\u5730\u5207\u51fa\u5404\u500b\u6578\u5b57 \u4e4b\u7247\u6bb5\uff0c\u518d\u4f7f\u7528\u52d5\u614b\u6642\u9593\u626d\u66f2\u8a08\u7b97\u8a3b\u518a\u8207\u6e2c\u8a66\u5c0d\u61c9\u6578\u5b57\u9593\u7684\u8ddd\u96e2\uff0c\u800c\u9019\u4e9b\u6578\u5b57\u4e4b\u8ddd\u96e2\u52a0 \u7e3d\u5373\u70ba\u9019\u7b46\u6e2c\u8a66\u4e4b\u8ddd\u96e2\u3002\u7531\u65bc\u53ef\u80fd\u6709\u591a\u500b\u8a3b\u518a\u97f3\u6a94\u4e4b\u60c5\u6cc1\uff0c\u56e0\u6b64\u6211\u5011\u5206\u5225\u53ef\u4ee5\u4f7f\u7528\u5e73\u5747 \u8ddd\u96e2\u6216\u662f\u6700\u77ed\u8ddd\u96e2\u4f5c\u70ba\u6700\u7d42\u8ddd\u96e2\u3002\u53cd\u8986\u5c0d\u65bc\u6240\u6709\u6e2c\u8a66\u505a\u5b8c\u5f8c\uff0c\u6211\u5011\u53ef\u4ee5\u5f97\u5230\u975e\u5e38\u591a\u7684\u8ddd \u96e2\uff0c\u518d\u5229\u7528\u9019\u4e9b\u8ddd\u96e2\u6211\u5011\u5373\u53ef\u8a08\u7b97 EER\uff0c\u5716\u3127\u5373\u70ba\u672c\u7cfb\u7d71\u4e4b\u6d41\u7a0b\u5716\u3002 \u5716\u3127\u3001\u52d5\u614b\u6642\u9593\u626d\u66f2\u6d41\u7a0b\u5716 \u5716\u4e8c\u70ba\u6211\u5011\u4f7f\u7528 feature scaling \u6642\u8a3b\u518a\u97f3\u6a94\u6578\u5c0d\u65bc\u52d5\u614b\u6642\u9593\u626d\u66f2\u4e4b\u6bd4\u8f03\u76f4\u689d\u5716\uff0c\u6211\u5011\u53ef \u7528\u6700\u77ed\u8ddd\u96e2\u5247\u6709\u660e\u986f\u7684\u7a69\u5b9a\u4e0b\u964d\u4e4b\u8da8\u52e2\u3002\u7d93\u904e\u5206\u6790\u767c\u73fe\u9019\u7a2e\u73fe\u8c61\u662f\u56e0\u70ba\u300c\u5e73\u5747\u300d\u7528\u65bc\u8ddd \u96e2\u4e26\u4e0d\u516c\u5e73\uff0c\u8209\u4f8b\u4f86\u8aaa\uff1a\u5c0f\u660e\u5728\u65e9\u4e0a\u6642\u9304\u97f3\u8a3b\u518a\uff0c\u5728\u4e2d\u5348\u6642\u53c8\u9032\u884c\u9304\u97f3\u8a3b\u518a\uff0c\u665a\u4e0a\u9032\u884c \u5247\u8a72\u8072\u97f3\u6703\u8207\u5c0f\u660e\u7684\u4e09\u500b\u8a3b\u518a\u97f3\u6a94\u9032\u884c DTW \u7b97\u51fa\u8ddd\u96e2\uff0c\u5982\u679c\u8a72\u8072\u97f3\u662f\u5728\u65e9\u4e0a\u9304\u88fd\u5247\u80af \u5b9a\u6703\u8ddf\u5c0f\u660e\u5728\u65e9\u4e0a\u8a3b\u518a\u7684\u8072\u97f3\u6700\u50cf\uff0c\u4f46\u6703\u8ddf\u4e2d\u5348\u3001\u665a\u4e0a\u7684\u8072\u97f3\u5247\u8f03\u4e0d\u50cf\uff0c\u5982\u679c\u6211\u5011\u5728\u9019 \u908a\u4f7f\u7528\u5e73\u5747\u8ddd\u96e2\u5247\u6703\u88ab\u4e2d\u5348\u5403\u98ef\u53ca\u665a\u4e0a\u5237\u7259\u6240\u5f71\u97ff\uff0c\u4f46\u4f7f\u7528\u6700\u5c0f\u8ddd\u96e2\u5247\u4e0d\u6703\u53d7\u5230\u5f71\u97ff\u3002 0.00% 1 2 3 4 5 \u8a3b\u518a\u97f3\u6a94\u6578 \u91cd\u758a\u5927\u5c0f\u70ba 512\uff0c\u901a\u7528\u80cc\u666f\u6a21\u578b\u90e8\u5206\u6211\u5011\u7684\u9ad8\u65af\u6df7\u5408\u6a21\u578b\u4f7f\u7528 128 \u500b\u9ad8\u65af\u5206\u4f48\u6700\u591a\u8fed\u4ee3 \u6578\u5b57\u7d1a\u8a9e\u8005\u9a57\u8b49\u4e4b\u8a3b\u518a\u6d41\u7a0b\u53ca\u6e2c\u8a66\u6d41\u7a0b\u591a\u534a\u8207\u8a9e\u53e5\u7d1a\u8a9e\u8005\u9a57\u8b49\u7cfb\u7d71\u76f8\u540c\uff0c\u6700\u5927\u5dee\u7570\u5728\u65bc \u6578\u5b57\u7d1a\u8a9e\u8005\u9a57\u8b49\u7cfb\u7d71\u6703\u5c07\u4e00\u6bb5\u5e8f\u5217\u5982\uff1a1029384756 \u62c6\u6210 1\u30010\u30012\u30019\u30013\u30018\u30014\u30017\u30015\u3001 6 \u5f8c\uff0c\u518d\u5c07\u500b\u5225\u6578\u5b57\u7576\u6210\u300c\u8a9e\u53e5\u300d\u9001\u5165\u8a9e\u53e5\u7d1a\u8a3b\u518a(\u5176\u6d41\u7a0b\u5716\u5982\u5716\u56db) \u3002\u4ee5\u4e0a\u8ff0\u4f8b\u5b50\u800c \u5011\u53ef\u4ee5\u5f9e\u5716\u4e2d\u767c\u73fe\u4f7f\u7528\u5e73\u5747\u5206\u6578\u4ee5\u53ca\u52a0\u6b0a\u5206\u6578\u5c0d\u65bc\u932f\u8aa4\u7387\u4e26\u4e0d\u6703\u6709\u592a\u5927\u7684\u5f71\u97ff\u3002\u6b64\u5916\u53ef \u4ee5\u767c\u73fe PLDA \u96a8\u8457\u7dad\u5ea6\u4e0a\u5347\u800c\u6574\u9ad4\u932f\u8aa4\u7387\u4e5f\u6709\u660e\u986f\u4e0a\u5347\u7684\u8da8\u52e2\uff0c\u6211\u5011\u8a8d\u70ba\u5176\u539f\u56e0\u76f8\u540c \u9996\u5148\u6211\u5011\u6aa2\u67e5\u4f7f\u7528\u55ae\u7368\u6578\u5b57\u7684\u932f\u8aa4\u7387\uff0c\u5982\u5716\u4e5d\u6240\u793a\u3002\u6211\u5011\u53ef\u4ee5\u767c\u73fe\u589e\u52a0\u8a3b\u518a\u97f3\u6a94\u6578\u6642\u6bcf \u500b\u6578\u5b57\u7684\u932f\u8aa4\u7387\u90fd\u6709\u660e\u986f\u4e0b\u964d\uff0c\u4f46\u6bcf\u500b\u6578\u5b57\u7684\u932f\u8aa4\u7387\u5dee\u7570\u975e\u5e38\u660e\u986f\uff0c\u932f\u8aa4\u7387\u6700\u9ad8\u7684\u524d\u56db \u689d\u7dda\u4f9d\u5e8f\u70ba\uff1adigit5\u3001digit2\u3001digit1\u3001digit8\u3002\u6211\u5011\u5c0d\u65bc\u6709\u9019\u7a2e\u73fe\u8c61\u7684\u731c\u6e2c\u662f\u6bcf\u500b\u6578\u5b57\u7684\u767c \u6389\u932f\u8aa4\u7387\u6700\u9ad8\u7684\u524d n \u500b\u6578\u5b57\u4e4b\u76f4\u689d\u5716\uff0c\u5f9e\u5716\u4e2d\u53ef\u4ee5\u767c\u73fe\u7576\u6211\u5011\u62ff\u6389 1 \u500b\u6578\u5b57\u5f8c\u932f\u8aa4\u7565 \u78ba\u5be6\u6709\u4e9b\u5fae\u63d0\u5347\uff0c\u62ff\u6389 2 \u500b\u6578\u5b57\u6642\u932f\u8aa4\u7387\u53c8\u518d\u5ea6\u56de\u5230\u4e0d\u62ff\u6389\u4efb\u4f55\u6578\u5b57\u4e4b\u932f\u8aa4\u7387\uff0c\u62ff\u6389 3 \u500b\u4ee5\u4e0a\u6578\u5b57\u5f8c\u932f\u8aa4\u7387\u958b\u59cb\u9010\u6f38\u5347\u9ad8\u3002\u56e0\u6b64\u6211\u5011\u53ef\u4ee5\u5f9e\u5be6\u9a57\u4e2d\u63a8\u65b7\u9019 2 \u500b\u932f\u8aa4\u7387\u6700\u9ad8\u7684\u6578 \u6211\u5011\u5f9e\u5be6\u9a57\u4e2d\u5f97\u77e5 i-vector \u614b\u6642\u9593\u626d\u66f2\u65b9\u6cd5\u4e4b\u6548\u679c\u986f\u7136\u6bd4 i-vector \u65b9\u6cd5\u4f86\u5f97\u5dee\u3002\u53e6\u5916\u6bd4\u8f03\u8a9e\u53e5\u7d1a\u8a9e\u8005\u9a57\u8b49\u7cfb\u7d71\u8207\u6578 \u5716\u4e03\u70ba\u6578\u5b57\u7d1a\u4f7f\u7528\u4e0d\u540c\u8a55\u5206\u6a21\u578b\u3001\u4e0d\u540c i-vector \u7dad\u5ea6\u3001\u4e0d\u540c\u5206\u6578\u7d44\u5408\u4e4b\u6bd4\u8f03\u76f4\u689d\u5716\uff0c\u6211 \u6700\u5f8c\u7684\u9304\u97f3\u8a3b\u518a\uff0c\u56e0\u6b64\u6211\u5011\u6709\u5c0f\u660e\u7684\u4e09\u500b\u8a3b\u518a\u97f3\u6a94\u3002\u800c\u73fe\u5728\u6709\u4e00\u500b\u5c0f\u660e\u7684\u8072\u97f3\u8981\u9a57\u8b49\uff0c 2.00% 4.00% EER \u7528\u8a55\u5206\u6a21\u578b\u5373\u53ef\u8f38\u51fa\u5206\u6578\uff0c\u53cd\u8986\u5c0d\u65bc\u6240\u6709\u6e2c\u8a66\u505a\u5b8c\u5f8c\uff0c\u6211\u5011\u5373\u53ef\u8a08\u7b97 EER \u4f5c\u70ba\u8a55\u91cf\u6a19 \u6e96\u3002 \u5be6\u9a57\u8a2d\u5b9a\u90e8\u5206\uff0c\u6211\u5011\u7279\u5fb5\u53c3\u6578\u4f7f\u7528 39 \u7dad\u7684 MFCC\uff0c\u62bd\u53d6 MFCC \u4f7f\u7528\u7684\u7a97\u5927\u5c0f\u70ba 1024\u3001 \u5716\u516d\u3001\u8a9e\u53e5\u7d1a\u4f7f\u7528\u4e0d\u540c\u8a55\u5206\u6a21\u578b\u4ee5\u53ca\u4e0d\u540c i-vector \u7dad\u5ea6\u4e4b\u6bd4\u8f03 (\u4e09)\u6578\u5b57\u7d1a\u8a9e\u8005\u9a57\u8b49 \u5716\u4e03\u3001\u6578\u5b57\u7d1a\u4f7f\u7528\u4e0d\u540c\u8a55\u5206\u6a21\u578b\u3001\u4e0d\u540c i-vector \u7dad\u5ea6\u3001\u4e0d\u540c\u5206\u6578\u7d44\u5408\u4e4b\u6bd4\u8f03 i-vector \u7dad\u5ea6 \u8aa4\u7387\u3002\u7136\u800c\u6578\u5b57\u7d1a\u8a9e\u8005\u9a57\u8b49\u7cfb\u7d71\u6548\u679c\u8f03\u5dee\u4e26\u4e0d\u76f4\u89c0\uff0c\u56e0\u70ba\u6578\u5b57\u7d1a\u8a9e\u8005\u9a57\u8b49\u7cfb\u7d71\u63a1\u7528\u7684\u65b9 \u5f0f\u662f\u6587\u672c\u76f8\u4f9d\u7684\u65b9\u5f0f\uff0c\u76f4\u89c0\u4e0a\u6211\u5011\u6703\u8a8d\u70ba\u6b64\u7a2e\u65b9\u5f0f\u6548\u679c\u61c9\u8a72\u8f03\u4f73\uff0c\u4f46\u5be6\u969b\u4e0a\u537b\u4e26\u975e\u5982\u6b64\u3002 \u56e0\u6b64\u4e0b\u9762\u6211\u5011\u5c07\u63a2\u8a0e\u6578\u5b57\u7d1a\u8a9e\u8005\u9a57\u8b49\u7cfb\u7d71\u8f03\u5dee\u7684\u53ef\u80fd\u539f\u56e0\u3002 \u8a08\u7b97\u53cd\u800c\u5c0e\u81f4\u6574\u9ad4\u932f\u8aa4\u7387\u4e0a\u5347\u3002\u5716\u5341\u3127\u70ba\u6211\u5011\u4f7f\u7528 PLDA \u8a55\u5206\u6a21\u578b\u5728 50 \u7dad i-vector \u62ff \u78ba\u5be6\u5305\u542b\u8457\u8a9e\u8005\u7279\u5fb5\u3002 \u4e94\u3001\u7d50\u8ad6 \u70ba\u767c\u97f3\u7279\u8cea\u4ee5\u53ca\u9577\u5ea6\u9020\u6210\u7684\u5f71\u97ff\uff0c\u56e0\u6b64\u6211\u5011\u8a8d\u70ba\u5f88\u6709\u53ef\u80fd\u662f\u9019\u4e9b\u932f\u8aa4\u7387\u8f03\u9ad8\u7684\u6578\u5b57\u52a0\u5165 vector \u7dad\u5ea6\u4e0a\u8868\u73fe\u90fd\u6bd4\u5207\u6578\u5b57\u4f86\u5f97\u597d\u8a31\u591a\uff0c\u56e0\u6b64\u6211\u5011\u53ef\u4ee5\u63a8\u65b7\u9023\u7e8c\u6578\u5b57\u767c\u97f3\u4e2d\u9593\u7684\u7279\u5fb5 \u5716\u5341\u4e8c\u3001\u4e0d\u5207\u6578\u5b57\u8207\u5207\u6578\u5b57\u65bc\u8a9e\u53e5\u7d1a\u4e4b\u6bd4\u8f03\u76f4\u689d\u5716 \u5f9e\u4e0a\u8ff0\u5be6\u9a57\u4e2d\u53ef\u4ee5\u767c\u73fe\u67d0\u4e9b\u6578\u5b57\u932f\u8aa4\u7387\u78ba\u5be6\u6bd4\u8f03\u9ad8\uff0c\u4e5f\u767c\u73fe\u4e86\u9019\u4e9b\u932f\u8aa4\u7387\u8f03\u9ad8\u53ef\u80fd\u662f\u56e0 \u7387\u5e73\u5747\u843d\u5728 0.64%\u9644\u8fd1\u3001\u4e0d\u5207\u6578\u5b57\u932f\u8aa4\u7387\u5927\u81f4\u843d\u5728 0.27%\u9644\u8fd1\uff0c\u4e0d\u5207\u6578\u5b57\u5728\u4efb\u610f\u7684 i-\u4ee5\u767c\u73fe\u4f7f\u7528\u5e73\u5747\u8ddd\u96e2\u7684 EER \u4e26\u4e0d\u6703\u56e0\u70ba\u8a3b\u518a\u97f3\u6a94\u6578\u7684\u589e\u52a0\u800c\u6709\u660e\u986f\u7684\u7a69\u5b9a\u4e0b\u964d\uff0c\u4f46\u4f7f \u5716\u4e8c\u3001\u4f7f\u7528 feature scaling \u6642\u8a3b\u518a\u97f3\u6a94\u6578\u5c0d\u65bc\u52d5\u614b\u6642\u9593\u626d\u66f2\u4e4b\u6bd4\u8f03 \u6b64\u5916\u53ef\u4ee5\u89c0\u5bdf\u5230\u52d5\u614b\u6642\u9593\u626d\u66f2\u61c9\u7528\u65bc\u8a9e\u8005\u9a57\u8b49\u4e0a\u6548\u679c\u975e\u5e38\u7684\u4e0d\u7406\u60f3\uff0c\u9019\u8207\u6211\u5011\u904e\u53bb\u7684\u8a8d \u77e5\u5927\u4e0d\u76f8\u540c\uff0c\u7d93\u904e\u4e00\u4e9b\u5206\u6790\u53ca\u731c\u6e2c\u5f8c\u767c\u73fe\u53ef\u80fd\u662f\u9032\u884c feature scaling \u6240\u9020\u6210\u7684\uff0c\u9019\u6a23\u731c \u6e2c\u7684\u539f\u56e0\u5728\u65bc feature scaling \u6703\u5c0e\u81f4\u6bcf\u4e00\u7dad\u4e2d\u7684 MFCC \u8ddd\u96e2\u5f71\u97ff\u8b8a\u76f8\u540c\uff0c\u4f46\u8a9e\u8005\u53ca\u6578\u5b57 \u7279\u6027\u53ef\u80fd\u4e3b\u8981\u5305\u542b\u65bc MFCC \u4e2d\u7684\u67d0\u4e9b\u7279\u5b9a\u7dad\u5ea6\u9032\u800c\u5c0e\u81f4\u5340\u5206\u8a9e\u8005\u7684\u80fd\u529b\u4e0b\u964d\uff0c\u5716\u4e09\u5373 \u70ba\u6211\u5011\u70ba\u4e86\u9a57\u8b49\u731c\u6e2c\u6240\u9032\u884c\u5be6\u9a57\u4e4b\u7d50\u679c\uff0c\u53ef\u4ee5\u767c\u73fe EER \u78ba\u5be6\u6709\u5927\u5e45\u964d\u4f4e\u3002 0.00% 5.00% 10.00% 15.00% 20.00% 1 2 3 4 5 EER \u8a3b\u518a\u97f3\u6a94\u6578 (\u4e8c)\u8a9e\u53e5\u7d1a\u8a9e\u8005\u9a57\u8b49 \u5224\u5225\u5206\u6790\u7684\u8a9e\u8005\u7a7a\u9593\u4ee5\u53ca\u901a\u9053\u7a7a\u9593\u4e4b\u77e9\u9663\u7dad\u5ea6\u7686\u8a2d\u5b9a\u70ba 50 \u7dad\uff0c\u6700\u591a\u8fed\u4ee3 50 \u6b21 3\u30018\u30014\u30017\u30015\u30016 \u7684 i-vector) \uff0c\u6e2c\u8a66\u6642\u5c07\u8a3b\u518a\u6642\u5f97\u5230\u4e4b 10 \u7d44\u4ee3\u8868\u8a9e\u8005\u7684 i-vector \u8207\u5207\u597d 1.33%\u3002 \u5916\u4ea6\u6709\u4e9b\u6578\u5b57\u518d\u9023\u7e8c\u5538\u8d77\u4f86\u6642\u6703\u88ab\u7701\u7565\u90e8\u5206\u97f3\u6216\u8005\u5538\u904e\u5feb\u7684\u73fe\u8c61\uff0c\u7d93\u904e\u89c0\u5bdf\u767c\u73fe\u9019\u4e9b\u6578 \u63d0\u5347\uff0c\u4f46\u537b\u4e5f\u53ef\u4ee5\u6e1b\u5c11\u4f7f\u7528\u8005\u6240\u9700\u82b1\u8cbb\u7684\u6642\u9593\u3002 \u62ff\u6389\u932f\u8aa4\u7387\u6700\u9ad8\u7684\u524d n \u500b\u6578\u5b57 \u597d\u3002\u6b64\u5916\u7d93\u904e\u8fd1\u4e00\u6b65\u5206\u6790\u6578\u5b57\u7d1a\u8a9e\u8005\u9a57\u8b49\u7cfb\u7d71\u6211\u5011\u767c\u73fe\u4ee5\u4e0b\u73fe\u8c61\uff1a 7.0% \u8a3b\u518a\u97f3\u6a94\u6578\u5c0d\u65bc DTW \u4e4b\u5f71\u97ff (\u4f7f\u7528 feature scaling) \u8207\u8a3b\u518a\u97f3\u6a94\u4e4b\u5e73\u5747\u8ddd\u96e2 \u8207\u8a3b\u518a\u97f3\u6a94\u4e4b\u6700\u5c0f\u8ddd\u96e2 6.00% scaling) vector\uff0c\u62bd\u51fa\u6e2c\u8a66\u97f3\u6a94\u7684 MFCC \u4e26\u5c0d MFCC \u9032\u884c\u5012\u983b\u8b5c\u5e73\u5747\u503c\u8207\u8b8a\u7570\u6578\u6b63\u898f\u5316\uff0c\u4e26\u5c07\u9019 \u4e9b\u7d93\u904e\u5012\u983b\u8b5c\u5e73\u5747\u503c\u8207\u8b8a\u7570\u6578\u6b63\u898f\u5316\u5f8c\u7684 MFCC \u4e1f\u9032\u5168\u8b8a\u7570\u7a7a\u9593\u6a21\u578b\u5f8c\u6211\u5011\u5373\u53ef\u5f97\u5230 \u8a72\u6e2c\u8a66\u97f3\u6a94\u4e4b i-vector\u3002\u6700\u7d42\u6211\u5011\u5c07\u6e2c\u8a66\u97f3\u6a94\u4e4b i-vector \u4ee5\u53ca\u5ba3\u7a31\u5c6c\u65bc\u8a9e\u8005\u7684 i-vector \u5229 0.00% 50 100 150 200 250 300 350 400 50 100 150 200 250 300 350 400 \u6b64\u7576\u6211\u5011\u4f7f\u7528\u52d5\u614b\u6642\u9593\u626d\u66f2\u9032\u884c\u6bd4\u8f03\u8ddd\u96e2\u6642\u6703\u53d7\u5230\u9019\u4e9b\u7279\u5fb5\u6240\u5f71\u97ff\uff0c\u9032\u800c\u5f71\u97ff\u6574\u9ad4\u7684\u932f \u7a97\u7684\u500b\u6578\u4e4b\u76f8\u95dc\u6027\uff0c\u6b64\u9593\u63a5\u53ef\u4ee5\u8b49\u5be6\u767c\u97f3\u7684\u9577\u5ea6\u8207\u96e3\u5ea6\u8207\u932f\u8aa4\u7387\u6210\u6b63\u6bd4\u7684\u95dc\u4fc2\u3002 i-vector \u7dad\u5ea6 \u5716\u5341\u4e8c\u5373\u70ba\u4e0d\u5207\u6578\u5b57\u8207\u5207\u6578\u5b57\u65bc\u8a9e\u53e5\u7d1a\u4e4b\u6bd4\u8f03\u76f4\u689d\u5716\uff0c\u6211\u5011\u53ef\u4ee5\u767c\u73fe\u5207\u6578\u5b57\u5f8c\u5408\u4f75\u932f\u8aa4 i-vector \u7dad\u5ea6 0.00% 1.00% \u7d71\u6b21\u4f73\uff0c\u52d5\u614b\u6642\u9593\u626d\u66f2\u8a9e\u8005\u9a57\u8b49\u7cfb\u7d71\u6700\u5dee\u3002\u52d5\u614b\u6642\u9593\u626d\u66f2\u70ba\u6700\u5dee\u7684\u539f\u56e0\u6211\u5011\u53ef\u4ee5\u5f88\u76f4\u89c0 \u7684\u731c\u5230\u662f\u56e0\u70ba MFCC \u4e2d\u5305\u542b\u8457\u96dc\u8a0a\u3001\u97f3\u91cf...\u7b49\u7279\u5fb5\uff0c\u7136\u800c\u9019\u4e9b\u7279\u5fb5\u4e26\u975e\u8a9e\u8005\u7684\u7279\u5fb5\uff0c\u56e0 \u5011\u53ef\u4ee5\u660e\u986f\u767c\u73fe\u6709\u5448\u73fe\u53cd\u6bd4\u4e4b\u8da8\u52e2\u3002\u5716\u5341\u4e4b\u7d50\u679c\u7531\u500b\u5225\u6578\u5b57\u7684\u932f\u8aa4\u7387\u4f86\u89c0\u5bdf EER \u8207\u97f3 50 100 150 200 250 300 350 400 \u4f9d\u7167\u6578\u5b57\u767c\u97f3\u6642\u9593\u53ea\u53d6\u6578\u5b57\u767c\u97f3\u6642\u9593\u5167\u5c0d\u61c9\u7a97\u7684 MFCC\u3002 0.00% \u8b8a\u5316\u3002\u6b64\u5916\u6211\u5011\u4ea6\u5c0d\u8a3b\u518a\u97f3\u6a94\u6578\u8207\u932f\u8aa4\u7387\u7684\u95dc\u4fc2\u505a\u51fa\u5ea7\u6a19\u5716\u5982\u5716\u4e5d\u6240\u793a\uff0c\u5f9e\u5ea7\u6a19\u5716\u4e2d\u6211 \u4e86\u3002\u56e0\u6b64\u6211\u5011\u6c7a\u5b9a\u5c07\u8a9e\u53e5\u7d1a\u7684\u97f3\u6a94\u4e5f\u9032\u884c\u76f8\u540c\u4f5c\u6cd5\uff0c\u5176\u505a\u6cd5\u662f\u5148\u62bd\u53d6\u6574\u53e5\u8a71\u4e4b MFCC \u518d \u8a3b\u518a\u97f3\u6a94\u6578\u5c0d\u65bc DTW \u4e4b\u5f71\u97ff (\u4e0d\u4f7f\u7528 feature \u5716\u56db\u3001\u8a9e\u53e5\u7d1a\u8a3b\u518a\u6d41\u7a0b\u5716 \u5716\u56db\u70ba\u8a9e\u53e5\u7d1a\u8a3b\u518a\u6d41\u7a0b\u5716\uff0c\u8f38\u5165\u70ba\u4e00\u4f4d\u8a9e\u8005\u7684 n \u500b\u97f3\u6a94\uff0c\u62bd\u51fa\u97f3\u6a94\u7684 MFCC \u5f8c\u5c0d MFCC \u9032\u884c\u5012\u983b\u8b5c\u5e73\u5747\u503c\u8207\u8b8a\u7570\u6578\u6b63\u898f\u5316\uff0c\u6700\u5f8c\u5c07\u9019\u4e9b\u7d93\u904e\u5012\u983b\u8b5c\u5e73\u5747\u503c\u8207\u8b8a\u7570\u6578\u6b63\u898f\u5316\u5f8c\u7684 MFCC \u4e1f\u9032\u5168\u8b8a\u7570\u7a7a\u9593\u6a21\u578b\u5f8c\u6211\u5011\u5373\u53ef\u5f97\u5230\u8a72\u4ee3\u8868\u8a9e\u8005\u4e4b i-vector\u3002 \u5716\u4e94\u3001\u8a9e\u53e5\u7d1a\u6e2c\u8a66\u6d41\u7a0b\u5716 \u5716\u4e94\u70ba\u8a9e\u53e5\u7d1a\u6e2c\u8a66\u6d41\u7a0b\u5716\uff0c\u8f38\u5165\u70ba\u4e00\u500b\u6e2c\u8a66\u7684\u97f3\u6a94\u4ee5\u53ca\u8a72\u6e2c\u8a66\u97f3\u6a94\u5ba3\u7a31\u5c6c\u65bc\u8a9e\u8005\u7684 i-\u5716\u516d\u70ba\u6211\u5011\u4f9d\u7167\u4e0a\u8ff0\u5be6\u9a57\u8a2d\u5b9a\u6240\u8dd1\u51fa\u4f86\u7684\u5be6\u9a57\u7d50\u679c\uff0c\u6211\u5011\u53ef\u4ee5\u767c\u73fe\u5728\u6211\u5011\u7684\u8cc7\u6599\u4e0a\u8a55\u5206 \u65b9\u5f0f\u4f7f\u7528\u9918\u5f26\u76f8\u4f3c\u5ea6\u5728 i-vector \u7dad\u5ea6\u8f03\u4f4e\u7684\u6642\u5019\u8868\u73fe\u90fd\u6bd4\u4f7f\u7528\u6a5f\u7387\u7dda\u6027\u5224\u5225\u5206\u6790\u4f86\u5f97\u5dee\uff0c \u4f46\u662f\u96a8\u8457 i-vector \u7dad\u5ea6\u4e0a\u5347\uff0c\u9918\u5f26\u76f8\u4f3c\u5ea6\u4e4b\u8868\u73fe\u53cd\u800c\u6bd4\u6a5f\u7387\u7dda\u6027\u5224\u5225\u5206\u6790\u4f86\u5f97\u597d\u3002\u5c0d\u65bc \u9019\u500b\u73fe\u8c61\u4e4b\u770b\u6cd5\u662f\uff1a\u7531\u65bc\u6211\u5011\u7684\u8cc7\u6599\u50c5\u6709\u6578\u5b57\u4e14\u8a9e\u8005\u6578\u76ee\u70ba\u5e7e\u767e\u4eba\uff0c\u56e0\u6b64\u4f7f\u7528\u6a5f\u7387\u7dda\u6027 \u5224\u5225\u5206\u6790\u5728\u8f03\u9ad8\u7dad\u5ea6\u7684 i-vector \u6642\u53cd\u800c\u6709\u53ef\u80fd\u9020\u6210\u904e\u5ea6\u64ec\u5408\u7684\u60c5\u5f62\uff0c\u6700\u5f8c\u4ea6\u53ef\u4ee5\u767c\u73fe\u9918 \u5f26\u76f8\u4f3c\u5ea6\u53ca\u6a5f\u7387\u7dda\u6027\u5224\u5225\u5206\u6790\u90fd\u5728 i-vector \u7dad\u5ea6 300 \u6642\u6709\u6700\u4f73\u7684\u8868\u73fe\u3002 2.00% \u5716\u516b\u70ba\u932f\u8aa4\u6b0a\u8861\u5716\uff0c\u5f9e\u5716\u4e2d\u6211\u5011\u53ef\u4ee5\u767c\u73fe\u8a9e\u53e5\u7d1a\u8a9e\u8005\u9a57\u8b49\u7cfb\u7d71\u6700\u4f73\uff0c\u6578\u5b57\u7d1a\u8a9e\u8005\u9a57\u8b49\u7cfb \u7684\u8a0a\u865f\u5716\uff0c\u78ba\u5be6\u767c\u73fe\u5230\u932f\u8aa4\u8f03\u9ad8\u7684\u6578\u5b57\u767c\u97f3\u8f03\u70ba\u7c21\u55ae\uff0c\u800c\u932f\u8aa4\u7387\u8f03\u4f4e\u7684\u6578\u5b57\u767c\u97f3\u5247\u8f03\u6709 \u7fd2\u6163\u7684\u8154\u8abf\uff0c\u4f46\u56e0\u70ba\u6211\u5011\u8981\u628a\u6578\u5b57\u5f37\u5236\u62c6\u958b\u6240\u4ee5\u5c0e\u81f4\u593e\u96dc\u5728\u6578\u5b57\u4e2d\u9593\u7684\u9019\u4e9b\u8154\u8abf\u4e5f\u6d88\u5931 0.20% 0.10% 0.20% 0.30% 0.40% 0.50% 0.60% EER sentence-level i-vector \u7dad\u5ea6\u5be6\u9a57 cosine PLDA \u7684\u6578\u5b57\u4e00\u4e00\u5c0d\u61c9\u9001\u5165\u8a9e\u53e5\u7d1a\u6e2c\u8a66(\u5176\u6d41\u7a0b\u5716\u5982\u5716\u4e94)\u5373\u53ef\u5f97\u5230 10 \u7d44\u5206\u6578\uff0c\u6700\u5f8c\u518d\u5c07\u9019 (\u56db)\u932f\u8aa4\u5206\u6790 \u5b57\u591a\u534a\u53ea\u6709\u6bcd\u97f3\uff0c\u82e5\u540c\u6642\u5305\u542b\u6bcd\u97f3\u53ca\u5b50\u97f3\u5247\u6bd4\u8f03\u4e0d\u6703\u6709\u6b64\u7a2e\u73fe\u8c61\u3002 (1) \u79fb\u9664\u932f\u8aa4\u8f03\u9ad8\u7684\u6578\u5b57\u5c0d\u65bc\u932f\u8aa4\u7387\u5f71\u97ff\u4e0d\u5927\u3002 6.0% \u4e9b\u5206\u6578\u9032\u884c\u7d44\u5408\u5373\u53ef\u5f62\u6210\u6700\u7d42\u4e4b\u5206\u6578\uff0c\u4e0b\u9762\u70ba\u6211\u5011\u7684\u4e8c\u7a2e\u7d44\u5408\u65b9\u6cd5\u3002 1. \u4f7f\u7528\u5e73\u5747\u5206\u6578 = \u2211 \u2032 * 1 10 9 =0 2. \u4f7f\u7528\u52a0\u6b0a\u5206\u6578 = \u2211 9 =0 * \u2032 \u5716\u516b\u3001\u932f\u8aa4\u6b0a\u8861\u5716 \u70ba\u4e86\u9a57\u8b49\u6bcf\u500b\u6578\u5b57\u932f\u8aa4\u7387\u5dee\u7570\u5f88\u5927\u662f\u5426\u5982\u6211\u5011\u731c\u6e2c\uff0c\u6211\u5011\u89c0\u5bdf\u6578\u5b57 0 \u5230 9 \u5728 Audacity \u6578\u5b57\u767c\u97f3\u4e2d\u9593\u7684\u4e00\u4e9b\u7279\u5fb5\u662f\u5426\u6709\u53ef\u80fd\u88ab\u79fb\u9664\u4e86\uff0c\u8209\u4f8b\u4f86\u8aaa\uff1a\u591a\u6578\u4eba\u7684\u767c\u97f3\u6642\u5e38\u5e36\u8457\u4e00\u4e9b 0.40% EER 3.00% EER \u5716\u4e5d\u3001\u55ae\u7368\u6578\u5b57\u4f7f\u7528 50 \u7dad i-vector \u65bc\u4e0d\u540c\u8a3b\u518a\u97f3\u6a94\u6578\u4e4b\u932f\u8aa4\u7387\u8da8\u52e2\u5716 \u7531\u65bc\u62ff\u6389\u932f\u8aa4\u7387\u8f03\u9ad8\u7684\u6578\u5b57\u5f8c\u6211\u5011\u7684\u932f\u8aa4\u7387\u4ecd\u7136\u6bd4\u8a9e\u53e5\u7d1a\u9ad8\u51fa\u8a31\u591a\uff0c\u6211\u5011\u9032\u800c\u731c\u6e2c\u9023\u7e8c 0.60% 4.00% 3.0% 1 2 3 4 5 6 7 8 \u8a3b\u518a\u97f3\u6a94\u6578 \u5716\u5341\u3001\u500b\u5225\u6578\u5b57\u932f\u8aa4\u7387\u8207\u7a97\u7684\u500b\u6578\u4e4b\u95dc\u4fc2\u5ea7\u6a19\u5716 0.80% 5.00% 4.5% 6.0% frame \u500b\u6578 \u4e0d\u5207\u6578\u5b57 \u5207\u6578\u5b57 \u5343 6.00% 7.5% 9 9.3 9.6 9.9 10.2 10.5 10.8 11.1 11.4 11.7 12 12.3 12.6 sentence-level i-vector \u7dad\u5ea6\u5be6\u9a57 9.0% 10.5% 5% 7.00% digit-level fusion \u6bd4\u8f03 cosine\u5e73\u5747 PLDA\u5e73\u5747 cosine\u52a0\u6b0a PLDA\u52a0\u6b0a 12.0% 13.5% 15.0% 16.5% 18.0% 19.5% 21.0% 22.5% 25.5% 28.5% \u8a3b\u518a\u97f3\u6a94\u6578\u4e4b\u5f71\u97ff digit0 digit1 digit2 digit3 digit4 digit5 digit6 digit7 digit8 digit9 30% (2) \u767c\u97f3\u8f03\u9577\u7684\u6578\u5b57\u6bd4\u767c\u97f3\u8f03\u77ed\u7684\u6578\u5b57\u7528\u65bc\u8a9e\u8005\u9a57\u8b49\u6548\u679c\u8f03\u4f73\u3002 5.0% \u6578\u5b57 EER \u8207 frame \u500b\u6578\u4e4b\u5ea7\u6a19\u5716 0 1 2 3 4 5 6 7 8 9 (3) \u767c\u97f3\u8f03\u70ba\u8907\u96dc\u4e4b\u6578\u5b57\u7684\u8a9e\u8005\u9a57\u8b49\u6548\u679c\u4e5f\u6bd4\u767c\u97f3\u7c21\u55ae\u7684\u6578\u5b57\u4f86\u5f97\u597d\u3002 4.0% EER (4) \u6578\u5b57\u7d1a\u8a9e\u8005\u9a57\u8b49\u7cfb\u7d71\u6548\u679c\u8f03\u5dee\u7684\u539f\u56e0\u4e4b\u4e00\u662f\u7531\u65bc\u5f37\u5236\u5207\u958b\u5c0e\u81f4\u9023\u7e8c\u6578\u5b57\u767c\u97f3\u9593\u7684 3.0% 27.0% 5 25% \u8a9e\u8005\u7279\u5fb5\u6d88\u5931\u6240\u81f4\u3002 2.0% 24.0% EER 0 1 3 4 6 \u5716\u5341\u3127\u3001\u62ff\u6389\u932f\u8aa4\u7387\u6700\u9ad8\u7684\u524d n \u500b\u6578\u5b57\u4e4b\u76f4\u689d\u5716 7 10% n 9 0 1 2 3 4 5 6 7 8 9 15% 0.0% EER 8 20% 1.0% 2 (5)</td></tr></table>", | |
| "html": null, | |
| "text": "\u4f4d\u8a9e\u8005\uff0c\u5176\u4e2d\u7537\u751f\u5171\u6709 160 \u4f4d\u3001\u5973\u751f\u5171\u6709 31 \u4f4d\uff0c\u5ba4\u5167\u9304\u97f3\u5171\u6709 172 \u4f4d\u3001 \u5ba4\u5916\u9304\u97f3\u5171\u6709 19 \u4f4d\u3002\u6bcf 1 \u4f4d\u8a9e\u8005\u7686\u6709 10 \u7d44\u9304\u97f3\uff0c\u6bcf\u7d44\u9304\u97f3\u4e4b\u5167\u5bb9\u7686\u70ba 0 \u81f3 9 \u7684\u6392\u5217\u7d44 \u5408(\u9577\u5ea6\u70ba 10) \uff0c\u9304\u97f3\u88dd\u7f6e\u7686\u900f\u904e\u5be6\u9a57\u5ba4\u958b\u767c\u4e4b Android APP \u9032\u884c\u9304\u97f3\uff0c\u53d6\u6a23\u983b\u7387\u70ba\u55ae \u8072\u9053 44.1 KHz\uff0c\u97f3\u8cea\u70ba 16bit\u3002\u6b64\u5916\u6211\u5011\u6709\u91dd\u5c0d\u6240\u6709\u7684\u97f3\u6a94\u7d93\u7531\u4eba\u5de5\u6a19\u8a18\u51fa 0 \u81f3 9 \u5728\u8a72 \u97f3\u6a94\u767c\u97f3\u958b\u59cb\u7684\u6642\u9593\u4ee5\u53ca\u767c\u97f3\u7d50\u675f\u7684\u6642\u9593\uff0c\u5176\u6a19\u8a18\u6240\u63a1\u7528\u4e4b\u7a0b\u5f0f\u70ba Audacity\u3002 2. \u4ea4\u5927\u738b\u9038\u5982\u6559\u6388\u4e4b\u8a9e\u6599 \u672c\u8a9e\u6599\u5171\u6709 100 \u4f4d\u8a9e\u8005\uff0c\u5176\u4e2d\u7537\u751f\u5171\u6709 50 \u4f4d\u3001\u5973\u751f\u5171\u6709 50 \u4f4d\uff0c\u7686\u70ba\u5ba4\u5167\u9304\u97f3\u3002\u6bcf 1 \u4f4d \u8a9e\u8005\u7686\u6709 10 \u7d44\u9304\u97f3\uff0c\u6bcf\u7d44\u9304\u97f3\u4e4b\u5167\u5bb9\u7686\u70ba\u9577\u5ea6 4 -12 \u7684\u96a8\u6a5f\u6578\u5b57\uff0c\u9304\u97f3\u88dd\u7f6e\u70ba\u9ea5\u514b\u98a8\uff0c False Rejection rate, FRR)\uff0c \u8a72\u6bd4\u4f8b\u6211\u5011\u7a31\u4e4b\u70ba\u76f8\u540c\u932f\u8aa4\u6bd4\u4f8b (Equal Error Rate, EER)\uff0c\u800c EER \u5373\u70ba\u6211\u5011\u7528\u4f86\u8a55\u4f30\u7cfb\u7d71\u7684\u6a19\u6e96\u3002\u53e6\u5916\u4e00\u63d0\u7684\u662f\uff0c\u96d6\u7136 \u4e00\u822c\u4f86\u8aaa EER \u5373\u80fd\u5920\u53cd\u61c9\u7cfb\u7d71\u7684\u6548\u80fd\uff0c\u4f46\u7576\u7cfb\u7d71\u8981\u5be6\u969b\u4e0a\u7dda\u6642\u4ecd\u9700\u8981\u4f9d\u7167\u61c9\u7528\u60c5\u5883\u5c0d \u9580\u6abb\u503c\u9032\u884c\u62bd\u63db\u3002\u8209\u4f8b\u4f86\u8aaa\u7576\u7528\u65bc\u9ad8\u5b89\u5168\u6027\u7684\u7cfb\u7d71\u6642\uff0c\u6211\u5011\u66f4\u5728\u4e4e\u7684\u6703\u662f FAR \u800c\u975e EER\uff0c \u5f9e\u4e0a\u8ff0\u4e2d\u5373\u6703\u767c\u73fe\u6703\u56e0\u70ba\u60c5\u5883\u7684\u4e0d\u540c\u800c\u5c0d\u65bc FAR \u53ca FRR \u6709\u4e0d\u540c\u7684\u8981\u6c42\uff0c\u800c\u9019\u90e8\u5206\u5247\u53ef \u4ee5\u7d93\u904e\u932f\u8aa4\u6b0a\u8861\u5716\u89c0\u5bdf\u5f8c\u9078\u64c7\u9069\u5408\u7684\u9580\u6abb\u503c\u3002 \u56db\u3001\u5be6\u9a57\u8207\u5206\u6790 \u6240\u6709\u7684\u5be6\u9a57\u63a1\u7528\u4e4b\u4f5c\u696d\u7cfb\u7d71\u70ba Ubuntu 16.04\u3001\u8655\u7406\u5668\u70ba Intel i7 8700\u3001\u8a18\u61b6\u9ad4\u70ba DDR 2400 16GB 4 \u689d\u3001\u7a0b\u5f0f\u8a9e\u8a00\u70ba Python3.6\uff0c\u8a9e\u8005\u9a57\u8b49\u5de5\u5177\u7bb1\u63a1\u7528 SPEAR [11]\uff0c\u9810\u8a2d\u4f7f\u7528\u4eba\u5de5\u6a19 \u7684\u65b9\u6cd5\u4f7f\u7528\u4e8c\u500b\u97f3\u6a94\u8a3b\u518a\u6642 EER \u8a9e\u53e5\u7d1a\u53ef\u4ee5\u843d\u5728 0.2% \uff0c\u800c \u6578\u5b57\u7d1a EER \u843d\u5728 1.33%\uff0c\u4f46\u4f7f\u7528\u52d5\u614b\u6642\u9593\u626d\u66f2\u65b9\u6cd5\u5247 EER \u7121\u6cd5\u4f4e\u65bc 3%\uff0c\u7531\u6b64\u53ef\u898b\u52d5 \u6587\u672c\u5167\u5bb9\u8f03\u8907\u96dc\u6642 i-vector \u7dad\u5ea6\u61c9\u8a72\u589e\u52a0\uff0c\u6587\u672c\u5167\u5bb9\u8f03\u7c21\u6613\u6642 i-vector \u7dad\u5ea6\u61c9\u8a72\u6e1b", | |
| "num": null | |
| } | |
| } | |
| } | |
| } |