| { | |
| "paper_id": "2019", | |
| "header": { | |
| "generated_with": "S2ORC 1.0.0", | |
| "date_generated": "2023-01-19T14:54:54.453061Z" | |
| }, | |
| "title": "Time Delay Neural Network-based Speaker Embedding Function Fine-tuned with Triplet Loss for Distance-based Speaker Recognition", | |
| "authors": [ | |
| { | |
| "first": "Chih-Ting", | |
| "middle": [], | |
| "last": "\u8449\u81f4\u5ef7", | |
| "suffix": "", | |
| "affiliation": { | |
| "laboratory": "", | |
| "institution": "National Sun Yat-sen University", | |
| "location": {} | |
| }, | |
| "email": "" | |
| }, | |
| { | |
| "first": "Po-Chin", | |
| "middle": [], | |
| "last": "Yehn", | |
| "suffix": "", | |
| "affiliation": { | |
| "laboratory": "", | |
| "institution": "National Sun Yat-sen University", | |
| "location": {} | |
| }, | |
| "email": "" | |
| }, | |
| { | |
| "first": "Su-Yu", | |
| "middle": [], | |
| "last": "Wang", | |
| "suffix": "", | |
| "affiliation": { | |
| "laboratory": "", | |
| "institution": "National Sun Yat-sen University", | |
| "location": {} | |
| }, | |
| "email": "" | |
| }, | |
| { | |
| "first": "Chia-Ping", | |
| "middle": [], | |
| "last": "Zhang", | |
| "suffix": "", | |
| "affiliation": { | |
| "laboratory": "", | |
| "institution": "National Sun Yat-sen University", | |
| "location": {} | |
| }, | |
| "email": "" | |
| }, | |
| { | |
| "first": "", | |
| "middle": [], | |
| "last": "Che", | |
| "suffix": "", | |
| "affiliation": { | |
| "laboratory": "", | |
| "institution": "National Sun Yat-sen University", | |
| "location": {} | |
| }, | |
| "email": "" | |
| }, | |
| { | |
| "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": "Bo-Cheng", | |
| "middle": [], | |
| "last": "Hsiao", | |
| "suffix": "", | |
| "affiliation": { | |
| "laboratory": "\u4e2d\u83ef\u96fb\u4fe1\u7814\u7a76\u9662 Chunghwa Telecom Laboratories, Taoyuan", | |
| "institution": "", | |
| "location": { | |
| "country": "Taiwan" | |
| } | |
| }, | |
| "email": "[email protected]" | |
| }, | |
| { | |
| "first": "Chung-Li", | |
| "middle": [], | |
| "last": "Chan", | |
| "suffix": "", | |
| "affiliation": { | |
| "laboratory": "\u4e2d\u83ef\u96fb\u4fe1\u7814\u7a76\u9662 Chunghwa Telecom Laboratories, Taoyuan", | |
| "institution": "", | |
| "location": { | |
| "country": "Taiwan" | |
| } | |
| }, | |
| "email": "[email protected]" | |
| }, | |
| { | |
| "first": "", | |
| "middle": [], | |
| "last": "Lu", | |
| "suffix": "", | |
| "affiliation": { | |
| "laboratory": "\u4e2d\u83ef\u96fb\u4fe1\u7814\u7a76\u9662 Chunghwa Telecom Laboratories, Taoyuan", | |
| "institution": "", | |
| "location": { | |
| "country": "Taiwan" | |
| } | |
| }, | |
| "email": "" | |
| } | |
| ], | |
| "year": "", | |
| "venue": null, | |
| "identifiers": {}, | |
| "abstract": "In this research work, we build a speaker recognition system based on the x-vector framework for speaker verification. During training, we propose to use the triplet loss to increase the distance between the embedding vectors from different speakers in high-dimensional space. During recognition, we use the European distance between test-utterance embedding vector and enrolled-speaker embedding vector for similarity measure, thus predicting the enrolled speaker with the minimum distance. The proposed system is evaluated with VoxCeleb speaker recognition dataset. The test set consists of utterances from 1,251 test speakers. The proposed", | |
| "pdf_parse": { | |
| "paper_id": "2019", | |
| "_pdf_hash": "", | |
| "abstract": [ | |
| { | |
| "text": "In this research work, we build a speaker recognition system based on the x-vector framework for speaker verification. During training, we propose to use the triplet loss to increase the distance between the embedding vectors from different speakers in high-dimensional space. During recognition, we use the European distance between test-utterance embedding vector and enrolled-speaker embedding vector for similarity measure, thus predicting the enrolled speaker with the minimum distance. The proposed system is evaluated with VoxCeleb speaker recognition dataset. The test set consists of utterances from 1,251 test speakers. The proposed", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Abstract", | |
| "sec_num": null | |
| } | |
| ], | |
| "body_text": [ | |
| { | |
| "text": "model achieves the top-1 recognition accuracy of 59.57% and the top-5 accuracy of 80.32%. ", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "", | |
| "sec_num": null | |
| }, | |
| { | |
| "text": "\u662f\u5e38\u898b\u7279\u5fb5\u6b63\u898f\u5316\u7684\u65b9\u6cd5\u4e4b\u4e00\uff0c\u85c9\u7531\u6574\u6bb5\u8a9e\u97f3\u7279\u5fb5\u7684\u5e73\u5747\u8207\u6a19\u6e96\u5dee\u5c0d\u7279\u5fb5\u9032\u884c\u6a19\u6e96\u5316\uff0c \u5047\u8a2d\u4e00\u97f3\u6a94\u7684\u97f3\u6846\u9577\u5ea6\u70ba \uff0c\u6bcf\u4e00\u97f3\u6846\u4e4b\u7279\u5fb5\u7dad\u5ea6\u70ba \uff0c1 \u2264 \u2264 \u4e141 \u2264 \u2264 \uff0c ) ( )\u8868 \u793a\u70ba\u7b2c \u500b\u97f3\u6846\u4e2d\u7b2c \u7dad\u7684\u7279\u5fb5\uff0c\u800c\u5c07 ) ( )\u6b63\u898f\u5316\u81f3 , ) ( )\u7684\u516c\u5f0f\u5982\u4e0b\uff1a , ) ( ) = ) ( ) \u2212 ( ) ( ) (1) \u5176\u4e2d ( ) = 1 1 ) ( ) 2 )34", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "\u95dc\u9375\u8a5e\uff1a\u6642\u5ef6\u795e\u7d93\u7db2\u8def\u3001\u8a9e\u8005\u8fa8\u8b58\u3001\u4e09\u5143\u7d44\u640d\u5931", | |
| "sec_num": null | |
| }, | |
| { | |
| "text": "(2)", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "\u95dc\u9375\u8a5e\uff1a\u6642\u5ef6\u795e\u7d93\u7db2\u8def\u3001\u8a9e\u8005\u8fa8\u8b58\u3001\u4e09\u5143\u7d44\u640d\u5931", | |
| "sec_num": null | |
| }, | |
| { | |
| "text": "( ) = 1 1( ) ( ) \u2212 ( )) 5 2 )34", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "\u95dc\u9375\u8a5e\uff1a\u6642\u5ef6\u795e\u7d93\u7db2\u8def\u3001\u8a9e\u8005\u8fa8\u8b58\u3001\u4e09\u5143\u7d44\u640d\u5931", | |
| "sec_num": null | |
| }, | |
| { | |
| "text": "(3) ", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "\u95dc\u9375\u8a5e\uff1a\u6642\u5ef6\u795e\u7d93\u7db2\u8def\u3001\u8a9e\u8005\u8fa8\u8b58\u3001\u4e09\u5143\u7d44\u640d\u5931", | |
| "sec_num": null | |
| } | |
| ], | |
| "back_matter": [], | |
| "bib_entries": { | |
| "BIBREF0": { | |
| "ref_id": "b0", | |
| "title": "X-vectors: Robust dnn embeddings for speaker recognition", | |
| "authors": [ | |
| { | |
| "first": "D", | |
| "middle": [], | |
| "last": "Snyder", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "D", | |
| "middle": [], | |
| "last": "Garcia-Romero", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "G", | |
| "middle": [], | |
| "last": "Sell", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "D", | |
| "middle": [], | |
| "last": "Povey", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "S", | |
| "middle": [], | |
| "last": "Khudanpur", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 2018, | |
| "venue": "Proc. ICASSP", | |
| "volume": "", | |
| "issue": "", | |
| "pages": "", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "D. Snyder, D. Garcia-Romero, G. Sell, D. Povey, and S. Khudanpur, \"X-vectors: Robust dnn embeddings for speaker recognition,\" in Proc. ICASSP, 2018.", | |
| "links": null | |
| }, | |
| "BIBREF1": { | |
| "ref_id": "b1", | |
| "title": "Representation learning: A review and new perspectives", | |
| "authors": [ | |
| { | |
| "first": "Y", | |
| "middle": [], | |
| "last": "Bengio", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "A", | |
| "middle": [], | |
| "last": "Courville", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "P", | |
| "middle": [], | |
| "last": "Vincent", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 2013, | |
| "venue": "IEEE transactions on pattern analysis and machine intelligence", | |
| "volume": "35", | |
| "issue": "", | |
| "pages": "1798--1828", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "Y. Bengio, A. Courville, and P. Vincent. \"Representation learning: A review and new perspectives,\" IEEE transactions on pattern analysis and machine intelligence, vol. 35, no. 8, pp. 1798-1828, 2013.", | |
| "links": null | |
| }, | |
| "BIBREF2": { | |
| "ref_id": "b2", | |
| "title": "A time delay neural network architecture forefficient modeling of long temporal contexts", | |
| "authors": [ | |
| { | |
| "first": "V", | |
| "middle": [], | |
| "last": "Peddinti", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "D", | |
| "middle": [], | |
| "last": "Povey", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "S", | |
| "middle": [], | |
| "last": "Khudanpur", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 2015, | |
| "venue": "Proc. Interspeech", | |
| "volume": "", | |
| "issue": "", | |
| "pages": "", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "V. Peddinti, D. Povey, and S. Khudanpur, \"A time delay neural network architecture forefficient modeling of long temporal contexts,\" in Proc. Interspeech, 2015.", | |
| "links": null | |
| }, | |
| "BIBREF3": { | |
| "ref_id": "b3", | |
| "title": "Robust text-independent speaker identification using gaussian mixture speaker models", | |
| "authors": [ | |
| { | |
| "first": "D", | |
| "middle": [ | |
| "A" | |
| ], | |
| "last": "Reynolds", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "R", | |
| "middle": [ | |
| "C" | |
| ], | |
| "last": "Rose", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 1995, | |
| "venue": "IEEE transactions on speech and audio processing", | |
| "volume": "3", | |
| "issue": "1", | |
| "pages": "72--83", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "D. A. Reynolds and R. C. Rose, \"Robust text-independent speaker identification using gaussian mixture speaker models,\" IEEE transactions on speech and audio processing, vol. 3, no. 1, pp. 72-83, 1995.", | |
| "links": null | |
| }, | |
| "BIBREF4": { | |
| "ref_id": "b4", | |
| "title": "Speaker verification using adapted gaussian mixture models", | |
| "authors": [ | |
| { | |
| "first": "D", | |
| "middle": [ | |
| "A" | |
| ], | |
| "last": "Reynolds", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "T", | |
| "middle": [ | |
| "F" | |
| ], | |
| "last": "Quatieri", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "R", | |
| "middle": [ | |
| "B" | |
| ], | |
| "last": "Dunn", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 2000, | |
| "venue": "Digital signal processing", | |
| "volume": "10", | |
| "issue": "1-3", | |
| "pages": "19--41", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "D. A. Reynolds, T. F. Quatieri, and R. B. Dunn, \"Speaker verification using adapted gaussian mixture models,\" Digital signal processing, vol. 10, no. 1-3, pp. 19-41, 2000.", | |
| "links": null | |
| }, | |
| "BIBREF5": { | |
| "ref_id": "b5", | |
| "title": "Joint factor analysis versus eigenchannels in speaker recognition", | |
| "authors": [ | |
| { | |
| "first": "P", | |
| "middle": [], | |
| "last": "Kenny", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "G", | |
| "middle": [], | |
| "last": "Boulianne", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "P", | |
| "middle": [], | |
| "last": "Ouellet", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "P", | |
| "middle": [], | |
| "last": "Dumouchel", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 2007, | |
| "venue": "IEEE Transactions on Audio, Speech, and Language Processing", | |
| "volume": "15", | |
| "issue": "4", | |
| "pages": "1435--1447", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "P. Kenny, G. Boulianne, P. Ouellet, and P. Dumouchel, \"Joint factor analysis versus eigenchannels in speaker recognition,\" IEEE Transactions on Audio, Speech, and Language Processing, vol. 15, no. 4, pp. 1435-1447, 2007.", | |
| "links": null | |
| }, | |
| "BIBREF6": { | |
| "ref_id": "b6", | |
| "title": "Discriminative and generative approaches for long-and short-term speaker characteristics modeling: application to speaker verification", | |
| "authors": [ | |
| { | |
| "first": "N", | |
| "middle": [], | |
| "last": "Dehak", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 2009, | |
| "venue": "", | |
| "volume": "", | |
| "issue": "", | |
| "pages": "", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "N. Dehak, Discriminative and generative approaches for long-and short-term speaker characteristics modeling: application to speaker verification. PhD thesis, \u00b4 Ecole de technologie sup\u00b4erieure, 2009.", | |
| "links": null | |
| }, | |
| "BIBREF7": { | |
| "ref_id": "b7", | |
| "title": "Front-end factor analysis for speaker verification", | |
| "authors": [ | |
| { | |
| "first": "N", | |
| "middle": [], | |
| "last": "Dehak", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "P", | |
| "middle": [ | |
| "J" | |
| ], | |
| "last": "Kenny", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "R", | |
| "middle": [], | |
| "last": "Dehak", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "P", | |
| "middle": [], | |
| "last": "Dumouchel", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "P", | |
| "middle": [], | |
| "last": "Ouellet", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 2011, | |
| "venue": "IEEE Transactions on Audio, Speech, and Language Processing", | |
| "volume": "19", | |
| "issue": "4", | |
| "pages": "788--798", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "N. Dehak, P. J. Kenny, R. Dehak, P. Dumouchel, and P. Ouellet, \"Front-end factor analysis for speaker verification,\" IEEE Transactions on Audio, Speech, and Language Processing, vol. 19, no. 4, pp. 788-798, 2011.", | |
| "links": null | |
| }, | |
| "BIBREF8": { | |
| "ref_id": "b8", | |
| "title": "Deep neural networks for small footprint text-dependent speaker verification", | |
| "authors": [ | |
| { | |
| "first": "E", | |
| "middle": [], | |
| "last": "Variani", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "X", | |
| "middle": [], | |
| "last": "Lei", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "E", | |
| "middle": [], | |
| "last": "Mcdermott", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "I", | |
| "middle": [ | |
| "L" | |
| ], | |
| "last": "Moreno", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "J", | |
| "middle": [], | |
| "last": "Gonzalez-Dominguez", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 2014, | |
| "venue": "2014 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP)", | |
| "volume": "", | |
| "issue": "", | |
| "pages": "4052--4056", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "E. Variani, X. Lei, E. McDermott, I. L. Moreno, and J. Gonzalez-Dominguez, \"Deep neural networks for small footprint text-dependent speaker verification,\" in 2014 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pp. 4052-4056, IEEE, 2014.", | |
| "links": null | |
| }, | |
| "BIBREF9": { | |
| "ref_id": "b9", | |
| "title": "Locally-connected and convolutional neural networks for small footprint speaker recognition", | |
| "authors": [ | |
| { | |
| "first": "Y", | |
| "middle": [], | |
| "last": "Chen", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "I", | |
| "middle": [], | |
| "last": "Lopez-Moreno", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "T", | |
| "middle": [ | |
| "N" | |
| ], | |
| "last": "Sainath", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "M", | |
| "middle": [], | |
| "last": "Visontai", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "R", | |
| "middle": [], | |
| "last": "Alvarez", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "C", | |
| "middle": [], | |
| "last": "Parada", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 2015, | |
| "venue": "Proc. Interspeech", | |
| "volume": "", | |
| "issue": "", | |
| "pages": "", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "Y.-h. Chen, I. Lopez-Moreno, T. N. Sainath, M. Visontai, R. Alvarez, and C. Parada , \"Locally-connected and convolutional neural networks for small footprint speaker recognition,\" in Proc. Interspeech, 2015.", | |
| "links": null | |
| }, | |
| "BIBREF10": { | |
| "ref_id": "b10", | |
| "title": "End-to-end text-dependent speaker verification", | |
| "authors": [ | |
| { | |
| "first": "G", | |
| "middle": [], | |
| "last": "Heigold", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "I", | |
| "middle": [], | |
| "last": "Moreno", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "S", | |
| "middle": [], | |
| "last": "Bengio", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "N", | |
| "middle": [], | |
| "last": "Shazeer", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 2016, | |
| "venue": "Proc. ICASSP", | |
| "volume": "", | |
| "issue": "", | |
| "pages": "", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "G. Heigold, I. Moreno, S. Bengio, and N. Shazeer, \"End-to-end text-dependent speaker verification,\" in Proc. ICASSP, 2016.", | |
| "links": null | |
| }, | |
| "BIBREF11": { | |
| "ref_id": "b11", | |
| "title": "Musan: A music, speech, and noise corpus", | |
| "authors": [ | |
| { | |
| "first": "D", | |
| "middle": [], | |
| "last": "Snyder", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "G", | |
| "middle": [], | |
| "last": "Chen", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "D", | |
| "middle": [], | |
| "last": "Povey", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 2015, | |
| "venue": "", | |
| "volume": "", | |
| "issue": "", | |
| "pages": "", | |
| "other_ids": { | |
| "arXiv": [ | |
| "arXiv:1510.08484" | |
| ] | |
| }, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "D. Snyder, G. Chen, and D. Povey, \"Musan: A music, speech, and noise corpus,\" arXiv preprint arXiv:1510.08484, 2015.", | |
| "links": null | |
| }, | |
| "BIBREF12": { | |
| "ref_id": "b12", | |
| "title": "Voxceleb: a large-scale speaker identification dataset", | |
| "authors": [ | |
| { | |
| "first": "A", | |
| "middle": [], | |
| "last": "Nagrani", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "J", | |
| "middle": [ | |
| "S" | |
| ], | |
| "last": "Chung", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "A", | |
| "middle": [], | |
| "last": "Zisserman", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 2017, | |
| "venue": "Proc. Interspeech", | |
| "volume": "", | |
| "issue": "", | |
| "pages": "", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "A. Nagrani, J. S. Chung, and A. Zisserman, \"Voxceleb: a large-scale speaker identification dataset,\" in Proc. Interspeech, 2017.", | |
| "links": null | |
| }, | |
| "BIBREF13": { | |
| "ref_id": "b13", | |
| "title": "Voxceleb2: Deep speaker recognition", | |
| "authors": [ | |
| { | |
| "first": "J", | |
| "middle": [ | |
| "S" | |
| ], | |
| "last": "Chung", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "A", | |
| "middle": [], | |
| "last": "Nagrani", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "A", | |
| "middle": [], | |
| "last": "Zisserman", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 2018, | |
| "venue": "Proc. Interspeech", | |
| "volume": "", | |
| "issue": "", | |
| "pages": "", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "J. S. Chung, A. Nagrani, and A. Zisserman, \"Voxceleb2: Deep speaker recognition,\" in Proc. Interspeech, 2018.", | |
| "links": null | |
| }, | |
| "BIBREF14": { | |
| "ref_id": "b14", | |
| "title": "Facenet: A unified embedding for face recognition and clustering", | |
| "authors": [ | |
| { | |
| "first": "F", | |
| "middle": [], | |
| "last": "Schroff", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "D", | |
| "middle": [], | |
| "last": "Kalenichenko", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "J", | |
| "middle": [], | |
| "last": "Philbin", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 2015, | |
| "venue": "Proceedings of the IEEE conference on computer vision and pattern recognition", | |
| "volume": "", | |
| "issue": "", | |
| "pages": "815--823", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "F. Schroff, D. Kalenichenko, and J. Philbin, \"Facenet: A unified embedding for face recognition and clustering,\" in Proceedings of the IEEE conference on computer vision and pattern recognition, pp. 815-823, 2015.", | |
| "links": null | |
| } | |
| }, | |
| "ref_entries": { | |
| "TABREF0": { | |
| "num": null, | |
| "text": "Keywords: TDNN, Speaker Recognition, Triplet Loss \u52a0\u5165\u566a\u97f3\u8207\u5229\u7528\u623f\u9593\u8108\u885d\u97ff\u61c9 (Room Impulse Response) \u52a0\u5165\u8ff4\u97ff\u4f86\u9032\u884c\u6578 \u64da\u589e\u5f37\u3002\u800c MUSAN \u8a9e\u6599\u5eab\u5167\u5bb9\u5305\u542b\u4e09\u500b\u90e8\u5206\uff0c\u5206\u5225\u70ba\u6f14\u8aaa (Speech) \u3001\u97f3\u6a02", | |
| "type_str": "table", | |
| "html": null, | |
| "content": "<table><tr><td>\u672c\u6587\u4e3b\u8981\u5206\u70ba\u4e94\u500b\u90e8\u4efd\uff1a\u7b2c\u4e00\u90e8\u4efd\u70ba\u7dd2\u8ad6\uff1b\u7b2c\u4e8c\u90e8\u4efd\u70ba\u76f8\u95dc\u7814\u7a76\u7684\u56de\u9867\u8207\u63a2\u8a0e\uff1b\u7b2c\u4e09\u90e8 \u6587\u7684\u8a9e\u8005\u8fa8\u8b58\u7cfb\u7d71\u4e5f\u4ee5 x-vector \u70ba\u57fa\u790e\u9032\u884c\u6539\u9032\u3002 \u97ff\uff0c\u4e0d\u50c5\u50c5\u589e\u52a0\u8cc7\u6599\u7684\u6578\u91cf\u8207\u591a\u6a23\u6027\uff0c\u4e5f\u66f4\u80fd\u4f7f\u7cfb\u7d71\u66f4\u52a0\u5f37\u5065\u3002\u6211\u5011\u4f7f\u7528 MUSAN \u8a9e</td></tr><tr><td>\u4efd\u70ba\u7814\u7a76\u65b9\u6cd5\u8207\u6d41\u7a0b\uff0c\u4ecb\u7d39\u4f7f\u7528\u8cc7\u6599\u96c6\u3001\u8cc7\u6599\u524d\u8655\u7406\u3001\u6a21\u578b\u67b6\u69cb\u3001\u8a13\u7df4\u6d41\u7a0b\u4ee5\u53ca\u8a3b\u518a\u8207 \u76f8\u4f3c\u5ea6\u6bd4\u5c0d\u7b49\u6d41\u7a0b\uff1b\u7b2c\u56db\u90e8\u4efd\u5247\u70ba\u5be6\u9a57\u7d50\u679c\u8207\u5206\u6790\uff0c\u8aaa\u660e\u5be6\u9a57\u74b0\u5883\u8207\u8a2d\u5b9a\uff0c\u4e26\u6839\u64da\u5be6\u9a57 \u7d50\u679c\u9032\u884c\u5206\u6790\uff1b\u7b2c\u4e94\u90e8\u4efd\u70ba\u7d50\u8ad6\u3002 \u4e09\u3001\u7814\u7a76\u65b9\u6cd5\u8207\u6d41\u7a0b (\u4e00) \u3001VoxCeleb \u8cc7\u6599\u96c6 \u6599\u5eab [12] (Music) \u8207\u566a\u97f3 (Noise) \uff0c\u6f14\u8aaa\u90e8\u5206\u7684\u5167\u5bb9\u70ba\u6717\u8b80\u66f8\u672c\u67d0\u7ae0\u7bc0\u5167\u5bb9\u7684\u8a9e\u97f3\u6216\u662f\u7f8e\u570b\u807d</td></tr><tr><td>VoxCeleb \u8cc7\u6599\u96c6\u53ef\u5206\u70ba VoxCeleb1 [13] \u8207 VoxCeleb2 [14]\uff0c\u5169\u8005\u7686\u70ba\u6587\u672c\u7121\u95dc (Text-</td></tr><tr><td>\u4e00\u3001\u7dd2\u8ad6 \u8b49\u6703\u6216\u8faf\u8ad6\u6703\u7684\u516c\u958b\u6f14\u8aaa\uff1b\u97f3\u6a02\u90e8\u5206\u5247\u6db5\u84cb\u53e4\u5178\u8207\u73fe\u4ee3\u6d41\u884c\u6a02\uff1b\u566a\u97f3\u90e8\u5206\u5305\u542b\u5404\u985e\u5e38</td></tr><tr><td>\u666f\u96dc\u97f3\u751a\u81f3\u662f\u5176\u4ed6\u4eba\u8aaa\u8a71\u7684\u8072\u97f3\uff0c\u8cc7\u6599\u96c6\u9664\u4e86\u63d0\u4f9b\u8a9e\u8005\u8eab\u5206\u4e4b\u5916\uff0c\u4e5f\u63d0\u4f9b\u8a72\u8a9e\u8005\u570b\u7c4d \u96a8\u8457\u5927\u6578\u64da\u7684\u6642\u4ee3\u5230\u4f86\uff0c\u6df1\u5ea6\u5b78\u7fd2\u6210\u70ba\u6642\u4e0b\u7684\u71b1\u9580\u8b70\u984c\u4e4b\u4e00\uff0c\u4e26\u6162\u6162\u8d70\u5165\u6211\u5011\u7684\u751f\u6d3b\u4e4b \u898b\u566a\u97f3\uff0c\u4f46\u4e0d\u5305\u62ec\u660e\u986f\u53ef\u8fa8\u8b58\u8aaa\u8a71\u5167\u5bb9\u7684\u4eba\u8072\u3002\u5728\u7522\u751f\u6578\u64da\u589e\u5f37\u5f8c\u7684\u97f3\u6a94\u5f8c\uff0c\u56e0\u70ba\u904b \u4e8c\u3001\u76f8\u95dc\u7814\u7a76 Independent) \u8a9e\u97f3\u8cc7\u6599\u96c6\uff0c\u5167\u5bb9\u6e90\u81ea\u65bc Youtube \u4e2d\u540d\u4eba\u7684\u5f71\u7247\uff0c\u56e0\u6b64\u5167\u5bb9\u53ef\u80fd\u6703\u6709\u80cc \u5716\u4e00\u3001\u6885\u723e\u5012\u983b\u8b5c\u4fc2\u6578\u8072\u5b78\u7279\u5fb5\u8655\u7406\u6d41\u3002</td></tr><tr><td>\u5728\u8a9e\u8005\u8fa8\u8b58\u8207\u9a57\u8b49\u6280\u8853\u767c\u5c55\u4e4b\u4e2d\uff0c\u9ad8\u65af\u6df7\u548c\u6a21\u578b (Gaussian Mixture Model) [4] \u626e\u6f14\u5341\u5206 \u7b97\u8cc7\u6e90\u7684\u8003\u91cf\uff0c\u6211\u5011\u4e26\u4e0d\u6703\u5168\u90e8\u4f7f\u7528\uff0c\u800c\u662f\u96a8\u6a5f\u53d6 1,000,000 \u500b\u6578\u64da\u589e\u5f37\u7684\u97f3\u6a94\u8207\u539f\u59cb \u4e2d\uff0c\u800c\u8eab\u5206\u9a57\u8b49\u8207\u8b58\u5225\u5c31\u662f\u4e00\u500b\u4e3b\u8981\u7684\u61c9\u7528\u7bc4\u7587\uff0c\u4ee5\u524d\u53ea\u51fa\u73fe\u5728\u96fb\u5f71\u88e1\u7684\u751f\u7269\u7279\u5fb5\u8fa8\u8b58 \u7cfb\u7d71\u4e5f\u9010\u6f38\u53ef\u4ee5\u5728\u6211\u5011\u7684\u65e5\u5e38\u751f\u6d3b\u4e2d\u767c\u73fe\u8e64\u8de1\uff0c\u5982\u8072\u7d0b\u9019\u985e\u751f\u7269\u7279\u5fb5\u4e0d\u4f3c\u50b3\u7d71\u7684\u9470\u5319\u6216 \u91cd \u8981 \u7684 \u89d2 \u8272 \uff0c \u800c \u9ad8 \u65af \u6df7 \u548c \u6a21 \u578b -\u901a \u7528 \u80cc \u666f \u6a21 \u578b (Gaussian Mixture Model-Universal \u4ee5\u53ca\u6027\u5225\uff0c\u5169\u8cc7\u6599\u96c6\u7684\u8cc7\u6599\u5206\u4f48\u72c0\u6cc1\u5982\u8868\u4e00\u3002VoxCeleb1 \u5b98\u65b9\u63d0\u4f9b\u5169\u7a2e\u8cc7\u6599\u96c6\u5206\u5272\u65b9 \u97f3\u6a94\u9032\u884c\u6a21\u578b\u7684\u8a13\u7df4\u3002 \u7531\u65bc\u5be6\u969b\u61c9\u7528\u6642\u6536\u97f3\u5bb9\u6613\u53d7\u74b0\u5883\u4ee5\u53ca\u9304\u97f3\u88dd\u7f6e\u7b49\u56e0\u7d20\u5e72\u64fe\uff0c\u6e2c\u8a66\u6642\u7684\u74b0\u5883\u53ef\u80fd\u8207\u8a13\u7df4\u97f3</td></tr><tr><td>\u662f\u9059\u63a7\u5668\uff0c\u53ef\u80fd\u6703\u6709\u907a\u5931\u7684\u98a8\u96aa\u5b58\u5728\uff0c\u57fa\u65bc\u8072\u7d0b\u7684\u8a9e\u8005\u8fa8\u8b58\u4e0d\u9700\u651c\u5e36\u984d\u5916\u7684\u7269\u54c1\uff0c\u50c5\u9700 Background Model, GMM-UBM) [5] \u66f4\u662f\u4e00\u500b\u91cd\u8981\u7684\u61c9\u7528\uff0c\u901a\u7528\u80cc\u666f\u6a21\u578b\u662f\u4e00\u7a2e\u5927\u578b\u7684 \u5f0f\uff0c\u8a9e\u8005\u9a57\u8b49\u5206\u5272 (Verification Split) \u8207\u8a9e\u8005\u8b58\u5225\u5206\u5272(Identification Split) \uff0c\u5169\u7a2e\u5206\u5272 \u6a94\u7684\u9304\u88fd\u74b0\u5883\u5927\u4e0d\u76f8\u540c\uff0c\u9032\u800c\u5f71\u97ff\u5be6\u969b\u4f7f\u7528\u6642\u7684\u6e96\u78ba\u7387\u3002\u9664\u4e86\u4f7f\u7528\u6578\u64da\u589e\u5f37\u76e1\u53ef\u80fd\u7684\u6a21</td></tr><tr><td>\u8005\u7684\u7279\u5fb5\u5206\u4f48\uff0c\u800c\u662f\u5148\u4f7f\u7528\u6240\u6709\u8a9e\u8005\u7684\u8cc7\u6599\u53bb\u8a13\u7df4\u4e00\u500b\u901a\u7528\u7684\u80cc\u666f\u6a21\u578b\uff0c\u8868\u793a\u51fa\u8a9e\u8005\u7121 \u8a9e\u8005\u9a57\u8b49\u5206\u5272\u4e4b\u9a57\u8b49\u96c6\u8207\u6e2c\u8a66\u96c6\u4e2d\u7684\u8a9e\u8005\u4e0d\u91cd\u8907\uff0c\u800c\u8a9e\u8005\u8b58\u5225\u5206\u5272\u4e4b\u8a13\u7df4\u96c6\u3001\u9a57\u8b49 (\u4e09) \u3001\u8072\u5b78\u7279\u5fb5\u8207\u6b63\u898f\u5316 \u8072\u97f3\u4fbf\u53ef\u4ee5\u9032\u884c\u8fa8\u8b58\uff0c\u9054\u5230\u66f4\u4fbf\u5229\u3001\u5b89\u5168\u7684\u6548\u679c\u3002 \u9ad8\u65af\u6df7\u548c\u6a21\u578b\uff0c\u4e26\u975e\u50cf\u50b3\u7d71\u9ad8\u65af\u6df7\u548c\u6a21\u578b\u91dd\u5c0d\u6bcf\u500b\u8a9e\u8005\u8a13\u7df4\u4e00\u500b\u9ad8\u65af\u6df7\u548c\u6a21\u578b\u4f86\u8868\u793a\u8a9e \u5305\u542b\u4e4b\u97f3\u6a94\u76f8\u540c\uff0c\u5dee\u5225\u50c5\u5728\u65bc\u4f9d\u4efb\u52d9\u76ee\u7684\u4e0d\u540c\u800c\u628a\u8cc7\u6599\u96c6\u9032\u884c\u4e0d\u540c\u7684\u5207\u5272\u5206\u6cd5\uff0c\u5176\u4e2d \u4eff\u5404\u7a2e\u4e0d\u540c\u7684\u74b0\u5883\u4e4b\u5916\uff0c\u6211\u5011\u4e5f\u52a0\u5165\u7279\u5fb5\u6b63\u898f\u5316\u7684\u65b9\u6cd5\u5e6b\u52a9\u6211\u5011\u589e\u52a0\u7cfb\u7d71\u7684\u5f37\u5065\u6027\uff0c\u5176</td></tr><tr><td>\u5176\u4e2d\uff0c\u8a9e\u8005\u9a57\u8b49\u662f\u7576\u524d\u71b1\u9580\u7684\u7814\u7a76\u9805\u76ee\uff0c\u8aaa\u8a71\u4eba\u9808\u5148\u5ba3\u7a31\u5176\u8eab\u4efd\uff0c\u518d\u7531\u8a9e\u8005\u9a57\u8b49\u7cfb\u7d71\u9032</td></tr><tr><td>\u884c\u6bd4\u5c0d\uff0c\u505a\u51fa\u662f\u5426\u901a\u904e\u9a57\u8b49\u7684\u6c7a\u65b7\u3002\u7136\u800c\u5728\u4e00\u4e9b\u5be6\u969b\u7684\u61c9\u7528\u4e0a\uff0c\u4eba\u5011\u958b\u59cb\u8ffd\u6c42\u6307\u4ee4\u7c21\u660e \u95dc (Speaker-independent) \u7684\u7279\u5fb5\u5206\u4f48\uff0c\u7136\u5f8c\u518d\u4f7f\u7528\u6307\u5b9a\u8a9e\u8005\u7684\u8cc7\u6599\u53bb\u8abf\u9069\u80cc\u666f\u6a21\u578b\u7522\u751f \u96c6\u3001\u6e2c\u8a66\u96c6\u4e2d\u7684\u8a9e\u8005\u76f8\u540c\u3002 \u6211\u5011\u6240\u4f7f\u7528\u7684\u8072\u5b78\u7279\u5fb5\u70ba\u6885\u723e\u5012\u983b\u8b5c\u4fc2\u6578 (Mel-Frequency Cepstral Coefficient, MFCC) \uff0c</td></tr><tr><td>\u8a9e\u8005\u6a21\u578b\u3002\u4e4b\u5f8c\uff0c\u70ba\u4e86\u89e3\u6c7a\u5728\u4e0d\u540c\u9304\u97f3\u88dd\u7f6e\u4e0a\uff0c\u540c\u8a9e\u8005\u7684\u9304\u97f3\u807d\u8d77\u4f86\u6703\u4e0d\u4e00\u6a23\u7684\u554f\u984c\uff0c \u662f\u4e00\u7a2e\u91dd\u5c0d\u4eba\u8033\u807d\u89ba\u800c\u8a2d\u8a08\u7684\u4e00\u7a2e\u8072\u5b78\u7279\u5fb5\uff0c\u7531\u65bc\u4eba\u8033\u5c0d\u4e0d\u540c\u983b\u7387\u7684\u8072\u97f3\u6709\u4e0d\u540c\u7684\u654f\u92b3</td></tr><tr><td>\u8207\u4fbf\u6377\u6027\uff0c\u82e5\u80fd\u53bb\u9664\u5ba3\u7a31\u8eab\u4efd\u7684\u6b65\u9a5f\uff0c\u4fbf\u53ef\u8b93\u4f7f\u7528\u8005\u7684\u9ad4\u9a57\u66f4\u4f73\u3002\u5982\u667a\u6167\u5bb6\u5ead\u7522\u54c1\u8a31\u591a</td></tr><tr><td>\u90fd\u63a1\u53d6\u8072\u63a7\u7684\u65b9\u5f0f\u4f86\u4e0b\u9054\u6307\u4ee4\uff0c\u9664\u4e86\u8fa8\u8b58\u4f7f\u7528\u8005\u6240\u4e0b\u9054\u7684\u6307\u4ee4\u4e4b\u5916\uff0c\u66f4\u5e0c\u671b\u80fd\u5c0d\u4e0b\u9054\u6307 \u806f\u5408\u56e0\u7d20\u5206\u6790 (Joint Factor Analysis, JFA) [6] \u63d0\u51fa\u5c07 GMM-UBM \u6240\u5f97\u51fa\u8d85\u7d1a\u5411\u91cf \u8868\u4e00\u3001VoxCeleb \u8cc7\u6599\u5206\u4f48\u8868 \u7a0b\u5ea6\uff0c\u6545\u6211\u5011\u5728\u983b\u7387\u5ea7\u6a19\u8ef8\u4f9d\u6885\u723e\u523b\u5ea6 (Mel Scale) \u914d\u7f6e\u5728\u4f4e\u983b\u8f03\u5bc6\u96c6\u3001\u9ad8\u983b\u8f03\u7a00\u758f\u7684</td></tr><tr><td>(Supervector) \u9032\u884c\u56e0\u7d20\u5206\u6790\uff0c\u53ef\u5206\u70ba\u901a\u9053\u5b50\u7a7a\u9593 (Channel Subspace) \u8207\u8a9e\u8005\u5b50\u7a7a\u9593 \u4e09\u89d2\u5e36\u901a\u6ffe\u6ce2\u5668 (Triangular Bandpass Filters)\uff0c\u8868\u793a\u4eba\u8033\u5c0d\u4f4e\u983b\u8072\u97f3\u611f\u53d7\u8f03\u70ba\u654f\u92b3\u4f46\u9762 \u4ee4\u7684\u4eba\u9032\u884c\u8fa8\u8b58\uff0c\u4f86\u9054\u5230\u4e00\u4e9b\u7c21\u55ae\u5ba2\u88fd\u5316\u56de\u61c9\u7684\u6548\u679c\uff0c\u4f8b\u5982\u6211\u5011\u82e5\u80fd\u5206\u8fa8\u4e0b\u9054\u6307\u4ee4\u7684\u662f (Speaker Subspace)\uff0cJFA \u76f8\u4fe1\u82e5\u80fd\u53bb\u9664\u901a\u9053\u56e0\u7d20\u7684\u5f71\u97ff\uff0c\u90a3\u6211\u5011\u5c31\u80fd\u53bb\u9664\u4e0d\u540c\u9304\u97f3\u689d\u4ef6 VoxCeleb1 VoxCeleb2 \u5c0d\u9ad8\u983b\u8072\u97f3\u4fbf\u76f8\u8f03\u70ba\u9072\u920d\u3002\u6885\u723e\u5012\u983b\u8b5c\u4fc2\u6578\u8072\u5b78\u7279\u5fb5\u8655\u7406\u6d41\u7a0b\u5982\u5716\u4e00\uff0c\u8a9e\u97f3\u8a0a\u865f\u7d93\u904e\u9810</td></tr><tr><td>\u5bb6\u4e2d\u54ea\u4f4d\u6210\u54e1\uff0c\u4fbf\u80fd\u63d0\u4f9b\u9069\u5408\u4e14\u5ba2\u88fd\u5316\u7684\u56de\u61c9\u7d66\u4f7f\u7528\u8005\uff0c\u5c31\u50cf\u662f\u4e00\u500b\u7c21\u55ae\u7684\u64ad\u97f3\u6a02\u6307\u4ee4\uff0c \u7684\u5f71\u97ff\uff0c\u4f7f\u7cfb\u7d71\u66f4\u5f37\u5065\u3002\u7136\u800c\u5728[7] \u4e2d\u537b\u767c\u73fe\uff0c\u901a\u9053\u90e8\u4efd\u4ecd\u7136\u5305\u542b\u8a9e\u8005\u8cc7\u8a0a\uff0c\u70ba\u4e86\u89e3\u6c7a \u9a57\u8b49\u5206\u5272 \u8fa8\u8b58\u5206\u5272 \u5f37\u8abf(Pre-emphasis)\uff0c\u4f86\u63d0\u5347\u9ad8\u983b\u7684\u90e8\u4efd\uff0c\u59cb\u4fe1\u865f\u7684\u983b\u8b5c\u8b8a\u5f97\u5e73\u5766\u3002\u4e4b\u5f8c\u5c07\u591a\u500b\u53d6\u6a23\u9ede\u96c6</td></tr><tr><td>\u5c0d\u5bb6\u4e2d\u9577\u8005\u53ef\u4ee5\u64ad\u653e\u53f0\u8a9e\u7d93\u5178\u6b4c\u66f2\uff0c\u800c\u5e74\u5c11\u8005\u5247\u53ef\u64ad\u653e\u6642\u4e0b\u5076\u50cf\u5718\u9ad4\u7684\u6b4c\u66f2\u3002</td></tr><tr><td>\u9019\u500b\u554f\u984c\uff0c\u63d0\u51fa\u4e86\u4e00\u7a2e\u5c07\u8a9e\u8005\u7a7a\u9593\u8207\u901a\u9053\u7a7a\u9593\u6574\u5408\u70ba\u55ae\u4e00\u7684\u5168\u5c40\u5dee\u7570\u7a7a\u9593 (Total dev test train dev test dev test \u5408\u6210\u4e00\u500b\u89c0\u6e2c\u55ae\u4f4d\uff0c\u7a31\u70ba\u97f3\u6846(Frame)\uff0c\u4e26\u5c0d\u6bcf\u4e00\u500b\u97f3\u6846\u4e58\u4e0a\u6f22\u540d\u7a97(Hamming window)</td></tr><tr><td>\u7136\u800c\uff0c\u5728\u5be6\u969b\u4f7f\u7528\u6642\u9700\u8981\u8fa8\u8b58\u7684\u8a9e\u8005\u6216\u662f\u8aaa\u8a3b\u518a\u8005\uff0c\u5927\u591a\u6642\u5019\u6703\u8207\u8a13\u7df4\u6642\u6240\u4f7f\u7528\u7684\u8a13\u7df4 Variability Space)\uff0c\u800c\u5c0d\u61c9\u7684\u5168\u5c40\u56e0\u5b50\u5247\u88ab\u7a31\u70ba i-vector (Identity Vector) [8]\u3002\u5728 2010 \u5e74 \u8a9e\u8005\u6578 1,211 40 1,251 1,251 1,251 5,994 118 \u4f86\u589e\u52a0\u97f3\u6846\u5de6\u7aef\u8207\u53f3\u7aef\u7684\u9023\u7e8c\u6027\u3002\u7531\u65bc\u8a0a\u865f\u5728\u6642\u57df\u4e0a\u7684\u8b8a\u5316\u5f88\u96e3\u770b\u51fa\u8a0a\u865f\u7684\u7279\u6027\uff0c\u56e0\u6b64</td></tr><tr><td>\u8cc7\u6599\u4e2d\u7684\u8a9e\u8005\u4e0d\u540c\uff0c\u6240\u4ee5\u7121\u6cd5\u7c21\u55ae\u5730\u4f7f\u7528\u57fa\u65bc Softmax \u7684\u5206\u985e\u5668\u4f86\u8655\u7406\uff0c\u70ba\u4e86\u89e3\u6c7a\u9019\u500b \u81f3 2016 \u5e74\u4e4b\u9593\u7684\u8a9e\u8005\u9a57\u8b49\u7cfb\u7d71\u5e7e\u4e4e\u90fd\u63a1\u7528 i-vector \u6216\u4ee5\u6b64\u70ba\u57fa\u790e\u9032\u884c\u6539\u9032\u3002 \u97f3\u6a94\u6578 148,642 4,874 138,361 6,904 8,251 \u6703\u5148\u7d93\u7531\u5feb\u901f\u5085\u7acb\u8449\u8f49\u63db(Fast Fourier Transfrom, FFT) \u5c07\u8a0a\u865f\u5f9e\u6642\u57df\u4fe1\u865f\u8f49\u63db\u5230\u983b\u57df\u4fe1 1,092,009 36,237 \u554f\u984c\uff0c\u5728\u672c\u8ad6\u6587\u4e2d\uff0c\u6211\u5011\u4ee5 x \u5411\u91cf [1] \u67b6\u69cb\u70ba\u57fa\u790e\uff0c\u4e26\u900f\u904e\u8868\u5fb5\u5b78\u7fd2 (Representation \u865f\u4e0a\uff0c\u4ee5\u80fd\u91cf\u5206\u4f48\u4f86\u89c0\u5bdf\u3002\u518d\u5c07\u5f97\u5230\u7684\u983b\u8b5c\u4e58\u4e0a\u591a\u7d44\u4e09\u89d2\u5e36\u901a\u6ffe\u6ce2\u5668\uff0c\u5f97\u5230\u6bcf\u4e00\u500b\u6ffe\u6ce2 \u53e6\u4e00\u65b9\u9762\uff0c\u7531\u65bc\u6df1\u5ea6\u5b78\u7fd2 (Deep Learning) \u5728\u5716\u50cf\u8fa8\u8b58\u7684\u6210\u529f\uff0c\u4eba\u5011\u4e5f\u5617\u8a66\u5c07\u6df1\u5ea6\u985e\u795e \u5728\u8cc7\u6599\u4f7f\u7528\u4e0a\uff0c\u6211\u5011\u4f7f\u7528 VoxCeleb2 \u9a57\u8b49\u96c6\u4f86\u8a13\u7df4\u6a21\u578b\u3001\u4f7f\u7528 VoxCeleb1 \u8b58\u5225\u5206\u5272\u7684 Learning) [2] \u7684\u65b9\u5f0f\uff0c\u5f9e\u6642\u5ef6\u795e\u7d93\u7db2\u8def (Time Delay Neural Network, TDNN) [3] \u4e2d\u53d6\u5f97 \u5d4c\u5165\u5411\u91cf (Embedding Vector) \uff0c\u4e26\u85c9\u7531\u5c0d\u8a3b\u518a\u8a9e\u97f3\u4e4b\u5d4c\u5165\u5411\u91cf\u9032\u884c\u8a3b\u518a\u7684\u52d5\u4f5c\uff0c\u70ba\u8a3b\u518a \u5668 \u8f38 \u51fa \u7684 \u5c0d \u6578 \u80fd \u91cf (Log energy) \u5f8c \uff0c \u5c07 \u5c0d \u6578 \u80fd \u91cf \u7d93 \u96e2 \u6563 \u9918 \u5f26 \u8f49 \u63db (Discrete Cosine \u7d93\u7db2\u8def\u61c9\u7528\u5728\u8a9e\u8005\u8fa8\u8b58\u7684\u4efb\u52d9\u4e0a\u3002\u5728 2014 \u5e74 d-vector [9] \u4f7f\u7528\u56db\u5c64 256 \u7dad\u96b1\u85cf\u5c64\u7684\u591a\u5c64 \u611f\u77e5\u5668\u9032\u884c\u8868\u5fb5\u5b78\u7fd2\uff0c\u4e26\u5f9e\u6700\u5f8c\u4e00\u5c64\u96b1\u85cf\u5c64\u64f7\u53d6\u51fa\u5d4c\u5165\u5411\u91cf\uff0c\u9019\u4e5f\u555f\u767c\u4e86\u5176\u4ed6\u4f7f\u7528\u5377\u7a4d \u6e2c\u8a66\u96c6\u4f86\u9032\u884c\u8a9e\u8005\u8fa8\u8b58\u7684\u6548\u80fd\u8a55\u4f30\u3001\u4f7f\u7528 VoxCeleb1 \u9a57\u8b49\u5206\u5272\u7684\u6e2c\u8a66\u96c6\u4f86\u9032\u884c\u8a9e\u8005\u8fa8 Transform, DCT)\u5f8c\u5373\u53ef\u5f97\u6885\u723e\u5012\u983b\u4fc2\u6578\u3002</td></tr><tr><td>\u8005\u5efa\u7acb\u8a9e\u8005\u6a21\u578b\uff0c\u5728\u6e2c\u8a66\u8a9e\u97f3\u9032\u5165\u7cfb\u7d71\u6642\uff0c\u6703\u5c0d\u6e2c\u8a66\u8a9e\u97f3\u4e4b\u5d4c\u5165\u5411\u91cf\u8207\u6240\u6709\u8a3b\u518a\u8005\u4e4b\u8a9e \u795e\u7d93\u7db2\u8def (Convolutional Neural Network, CNN) [10] \u6216\u662f\u905e\u8ff4\u795e\u7d93\u7db2\u8def (Recurrent \u8b58\uff0c\u66f4\u9032\u4e00\u6b65\u7684\u4f86\u7814\u7a76\u662f\u5426\u8a9e\u8005\u9a57\u8b49\u7684\u6e96\u78ba\u7387\u662f\u5426\u8207\u8a9e\u8005\u8fa8\u8b58\u6b63\u76f8\u95dc\u3002</td></tr><tr><td>\u8005\u6a21\u578b\u9032\u884c\u76f8\u4f3c\u5ea6\u6bd4\u5c0d\uff0c\u4e26\u9078\u51fa\u6700\u76f8\u4f3c\u8005\u505a\u70ba\u7cfb\u7d71\u5224\u5b9a\u6e2c\u8a66\u8a9e\u97f3\u6240\u5c6c\u4e4b\u8a9e\u8005\u3002\u70ba\u4e86\u4f7f\u6e96</td></tr><tr><td>Neural Network, RNN) [11] \u958b\u767c\u8a9e\u8005\u9a57\u8b49\u7cfb\u7d71\u7684\u60f3\u6cd5\u3002\u5230\u4e86 2016 \u5e74\uff0c\u4f7f\u7528\u6642\u5ef6\u795e\u7d93\u7db2</td></tr><tr><td>\u78ba\u7387\u63d0\u5347\uff0c\u6211\u5011\u4e0d\u50c5\u4f7f\u7528\u4ea4\u53c9\u71b5\u640d\u5931 (Cross Entropy Loss) \u4f86\u8a13\u7df4\u6a21\u578b\uff0c\u4e5f\u63a1\u7528\u4e09\u5143\u7d44</td></tr><tr><td>\u640d\u5931 (Triplet Loss) \u4f86\u5c0d\u6a21\u578b\u9032\u884c\u8abf\u9069\uff0c\u4f7f\u4e0d\u540c\u8a9e\u8005\u7684\u5d4c\u5165\u5411\u91cf\u5728\u9ad8\u7dad\u7a7a\u9593\u6709\u66f4\u597d\u7684\u5224 \u8def [3] \u7684 x-vector [1] \u88ab\u63d0\u51fa\uff0c\u5b83\u6700\u91cd\u8981\u7684\u7279\u8272\u5728\u65bc\u5c0d\u8a13\u7df4\u97f3\u6a94\u9032\u884c\u5982\uff1a\u52a0\u5165\u566a\u97f3\u3001\u8ff4 (\u4e8c) \u3001\u6578\u64da\u589e\u5f37</td></tr><tr><td>\u5225\u6027\u3002 \u97ff\u3001\u8b8a\u901f\u7b49\u6578\u64da\u589e\u5f37 (Data Augmentation) [12]\uff0c\u4f7f\u8a13\u7df4\u8cc7\u6599\u5448\u500d\u6578\u6210\u9577\u4e26\u7372\u5f97\u8d85\u8d8a\u7576\u6642 \u6211\u5011\u63a1\u7528\u6578\u64da\u589e\u5f37 (Data Augmentation) \u7684\u6280\u8853\uff0c\u5c0d\u8cc7\u6599\u52a0\u5165\u566a\u97f3\u3001\u6216\u662f\u5c0d\u97f3\u6a94\u52a0\u5165\u8ff4</td></tr><tr><td>\u5176\u4ed6\u7cfb\u7d71\u7684\u5f37\u5065\u6027\u3002\u5982\u4eca x-vector \u5df2\u7d93\u662f\u76ee\u524d\u6700\u4e3b\u6d41\u7684\u8a9e\u8005\u8b58\u5225\u8207\u9a57\u8b49\u7cfb\u7d71\u4e4b\u4e00\uff0c\u672c\u8ad6</td></tr></table>" | |
| }, | |
| "TABREF1": { | |
| "num": null, | |
| "text": "\u4f86\u5c07\u97f3\u6846\u5c64\u7d1a\u7684\u8cc7\u8a0a\u8f49\u70ba\u97f3\u6bb5\u5c64\u7d1a\u3002\u800c\u6a21\u578b\u6700\u5f8c\u4e00\u5c64\u8f38\u51fa\u5c64\u70ba 5,994 \u7dad softmax \uff0c\u53ef\u4ee5\u767c\u73fe\u50c5\u4f7f\u7528 softmax \u6642 EER \u70ba 9.64%\uff0c \u800c\u4f7f\u7528\u4e09\u5143\u7d44\u640d\u5931\u800c\u672a\u4f7f\u7528 softmax \u9810\u8a13\u7df4\u6a21\u578b\u6642\uff0cEER \u5247\u7565\u9ad8\u4f86\u5230 10.39%\uff0c\u4e0d\u904e\u4f7f \u6b64\u5916\uff0c\u5716\u516d\u70ba\u5404\u8a13\u7df4\u65b9\u6cd5\u5728\u8a9e\u8005\u9a57\u8b49\u4e0a\u5206\u6578\u7684\u5206\u4f48\u5716\uff0c\u6a6b\u8ef8\u70ba trials \u4e2d\u6bd4\u5c0d\u76ee\u6a19\u9593\u7684\u6b50 \u6c0f\u8ddd\u96e2\uff0c\u800c\u7e31\u8ef8\u4ee3\u8868\u9810\u6e2c\u7d50\u679c\u70ba\u8a72\u8ddd\u96e2\u6642\u8cc7\u6599\u7684\u6578\u76ee\uff0c\u7d05\u8272\u70ba target trials \u7684\u5206\u6578\u5206\u4f48\u60c5 \u6cc1\uff0c\u85cd\u8272\u70ba nontarget trials \u7684\u5206\u6578\u5206\u4f48\u60c5\u6cc1\u3002\u6211\u5011\u89c0\u5bdf\u7684\u91cd\u9ede\u6709\u4e8c\uff1a\u4e00\u662f target \u8207 nontarget", | |
| "type_str": "table", | |
| "html": null, | |
| "content": "<table><tr><td>\u7684\u516c\u5f0f\u5982\u4e0b\uff1a \u5728\u751f\u6210\u4e09\u5143\u7d44\u7684\u904e\u7a0b\u4e2d\uff0c\u6bcf\u7b46\u97f3\u6a94\u7686\u6703\u6210\u70ba\u9328\u9ede\uff0c\u8207\u6240\u6709\u53ef\u80fd\u7684\u6b63\u6a23\u672c\u53bb\u627e\u5c0b\u4e00\u500b\u7b26\u5408 \u6703\u8207\u5be6\u969b\u61c9\u7528\u6642\u6709\u6240\u4e0d\u540c\uff0c\u6240\u4ee5\u6211\u5011\u5fc5\u9808\u5c0d\u60f3\u8981\u8a3b\u518a\u7684\u4f7f\u7528\u8005\uff0c\u64f7\u53d6\u5176\u8a9e\u97f3\u7684\u5d4c\u5165\u5411\u91cf\uff0c \u7684\u9078\u64c7\u908a\u754c \u70ba 0.2\u3002 Triplet 23.68 % 45.58 %</td></tr><tr><td>\u6a5f\u7387\u3002 \u7b97\uff0c\u5230\u7b2c\u4e09\u5c64\u5247\u662f\u53d6\u76f8\u9130\u7684 7 \u500b\u97f3\u6846\uff0c\u7b2c\u56db\u5c64\u8207\u7b2c\u4e94\u5c64\u5247\u50c5\u53d6 1 \u500b\u3002\u6642\u5ef6\u795e\u7d93\u7db2\u8def\u4fbf\u662f <\uff1d4 || C \u2212 D || 5 5 < || C \u2212 F || < || C \u2212 D || 5 5 + (5) \u65bc\u95be\u503c\u7684\u8a3b\u518a\u8005\u70ba\u6700\u7d42\u7cfb\u7d71\u505a\u51fa\u7684\u8fa8\u8b58\u7d50\u679c\u3002 \u5206\u4f48\u91cd\u758a\u7684\u90e8\u4efd\uff0c\u4ee3\u8868\u6a21\u578b\u53ef\u80fd\u5206\u985e\u932f\u8aa4\u7684\u90e8\u4efd\uff0c\u91cd\u758a\u7684\u9762\u7a4d\u6108\u5c0f\u8868\u793a\u6108\u80fd\u5c07\u4e0d\u540c\u8a9e\u8005 (\u4e8c)\u5be6\u9a57\u7d50\u679c \u5728\u6211\u5011\u7684\u6a21\u578b\u4e2d\u97f3\u6846\u5c64\u7d1a\u70ba\u4e94\u5c64\u67b6\u69cb\uff0c\u7b2c\u4e00\u5c64\u8207\u7b2c\u4e8c\u5c64\u53d6\u76f8\u9130\u7684 5 \u500b\u97f3\u6846\u70ba\u8f38\u5165\u9032\u884c\u904b 9: = \u2212 1 < ( ( )) 9 \u689d\u4ef6\u7684\u8ca0\u6a23\u672c\u7d44\u6210\u4e09\u5143\u7d44\uff0c\u5176\u4e2d\u9328\u9ede\u8207\u6b63\u6a23\u672c\u7684\u7d44\u5408\u4e0d\u5f97\u91cd\u8907\uff0c\u4e5f\u5c31\u662f\u8aaa\uff0c\u4efb\u4e00\u9328\u9ede\u4e0d \u4ee5\u63d0\u4f9b\u8a3b\u518a\u7684\u529f\u80fd\uff0c\u800c\u6211\u5011\u63a1\u7528\u5c0d\u6240\u6709\u8a3b\u518a\u8a9e\u97f3\u5f97\u51fa\u7684\u5d4c\u5165\u5411\u91cf\u53d6\u5e73\u5747\u505a\u70ba\u8a3b\u518a\u8005\u7684\u8a9e (4) \u6703\u5728\u5b83\u7684\u6b63\u6a23\u672c\u88ab\u9078\u70ba\u9328\u9ede\u6642\u4f5c\u70ba\u6b63\u6a23\u672c\uff0c\u800c\u5c0d\u8ca0\u6a23\u672c\u7684\u9078\u64c7\u689d\u4ef6\u70ba\uff1a \u8005\u6a21\u578b\uff0c\u7576\u6e2c\u8a66\u8a9e\u97f3\u8f38\u5165\u6642\uff0c\u5247\u8207\u6240\u6709\u8a3b\u518a\u8a9e\u8005\u6a21\u578b\u6bd4\u8f03\u76f8\u4f3c\u5ea6\uff0c\u56de\u50b3\u76f8\u4f3c\u5ea6\u6700\u9ad8\u4e14\u9ad8 Softmax + Triplet 59.57 % 80.32 %</td></tr><tr><td>\u900f\u904e\u96b1\u85cf\u5c64\u7684\u5806\u758a\u4f86\u63d0\u7149\u9023\u7e8c\u97f3\u6846\u7684\u7279\u5fb5\uff0c\u4e14\u96a8\u8457\u96b1\u85cf\u5c64\u7684\u589e\u52a0\uff0c\u795e\u7d93\u7db2\u8def\u53ef\u4ee5\u6536\u96c6\u5230 \u66f4\u5927\u7bc4\u570d\u97f3\u6846\u7684\u8cc7\u8a0a\uff1b\u5728\u97f3\u6846\u5c64\u7d1a\u7d50\u675f\u6642\uff0c\u6703\u5728\u7d71\u8a08\u6c60\u5316\u5c64\u5c0d\u6240\u6709\u97f3\u6846\u8a08\u7b97\u5e73\u5747\u8207\u8b8a\u7570 \u6578\uff0c\u6574\u5408\u6210\u97f3\u6bb5\u5c64\u7d1a\u7684\u8cc7\u8a0a\u3002\u6b64\u5916\uff0c\u5728\u9019\u500b\u6a21\u578b\u4e2d\uff0c\u6bcf\u5c64\u96b1\u85cf\u5c64\u7686\u7d93\u904e\u6279\u91cf\u6a19\u6e96\u5316(Batch Normalization) \u8207 Rectified Linear Unit (ReLU)\u6fc0\u6d3b\u51fd\u6578\u3002\u7576\u6a21\u578b\u8a13\u7df4\u5b8c\u6210\u5f8c\uff0c\u6211\u5011\u5f9e Segment 1 \u8f38\u51fa\u53d6\u5f97 512 \u7dad\u7684\u5d4c\u5165\u5411\u91cf\u3002 \u70ba\u6700\u5f8c\u4e00\u5c64\u5206\u985e\u7684\u985e\u5225\u6578\uff1b < \u70ba\u8868\u793a\u8a72\u7b46\u8cc7\u6599\u7684\u771f\u5be6\u985e\u5225\uff0c\u50c5\u5728\u8a72\u7b46\u8cc7\u6599\u771f\u5be6\u985e\u5225\u5c6c \u65bc\u7b2c \u985e\u6642\u70ba 1\uff0c\u5176\u9918\u6642\u5019\u70ba 0\uff1b ( )\u70ba\u795e\u7d93\u7db2\u8def\u9810\u6e2c\u7b2c \u985e\u7684\u6a5f\u7387\u3002 \u5728\u76f8\u4f3c\u5ea6\u7684\u6bd4\u8f03\u4e0a\uff0c\u7531\u65bc\u4e09\u5143\u7d44\u640d\u5931\u51fd\u5f0f\u662f\u8a08\u7b97\u9328\u9ede\u8207\u6b63\u6a23\u672c\u3001\u9328\u9ede\u8207\u8ca0\u6a23\u672c\u6b50\u6c0f\u7a7a\u9593 \u6211\u5011\u5728\u672c\u7bc0\u6bd4\u8f03\u8a13\u7df4\u6d41\u7a0b\u4e2d\u662f\u5426\u4f7f\u7528\u4e09\u5143\u7d44\u640d\u5931\u4e4b\u9593\u7684\u5dee\u5225 (\u662f\u5426\u6709\u5716\u4e94\u4e2d\u7684\u4e09\u5143\u7d44 \u5206\u985e\u6e05\u695a\uff0c\u7cfb\u7d71\u7684\u6548\u80fd\u4e5f\u6108\u4f73\uff0c\u5f9e\u5716\u4e2d\u53ef\u4ee5\u767c\u73fe Softmax+Triplet \u9762\u7a4d\u6700\u5c0f\uff0c\u540c\u6642\u4e5f\u5728\u5be6 \u4e94\u3001\u7d50\u8ad6 C \u70ba\u9328\u9ede\u7684\u5d4c\u5165\u5411\u91cf\u3001 D \u70ba\u6b63\u6a23\u672c\u7684\u5d4c\u5165\u5411\u91cf\u3001 F \u70ba\u8ca0\u6a23\u672c\u7684\u5d4c\u5165\u5411\u91cf\uff0c \u70ba\u5b9a\u7fa9\u8ca0 \u6a23\u672c\u7684\u908a\u754c\u503c\u3002\u5728\u9078\u5b9a C \u8207 D \u7684\u60c5\u6cc1\u4e0b\uff0c\u8ca0\u6a23\u672c\u8207\u9328\u9ede\u7684\u8ddd\u96e2\u5fc5\u9808\u5c0f\u65bc\u6b63\u6a23\u672c\u8207\u9328\u9ede \u7684\u8ddd\u96e2\u52a0\u4e0a \uff0c\u4e14\u5927\u65bc\u6b63\u6a23\u672c\u8207\u9328\u9ede\u7684\u8ddd\u96e2\uff0c\u9019\u6a23\u9019\u7b46\u5176\u4ed6\u8a9e\u8005\u7684\u97f3\u6a94\u624d\u80fd\u88ab\u9078\u70ba\u8ca0\u6a23 \u672c\uff0c\u76ee\u7684\u5728\u65bc\u9078\u51fa\u534a\u96e3\u8ca0\u6a23\u672c (Semi-hard Negatives) \u4f86\u8a13\u7df4\uff0c\u5982\u5716\u56db\u6240\u793a\uff0c\u5982\u6b64\u4e00\u4f86\uff0c \u4e0b\u7684\u8ddd\u96e2\u5dee\uff0c\u6240\u4ee5\u76f8\u4f3c\u5ea6\u8a55\u4f30\u4f7f\u7528\u6b50\u6c0f\u8ddd\u96e2\u4f86\u8a08\u7b97\uff0c\u5f97\u51fa\u7684\u503c\u6108\u5c0f\u5247\u8868\u793a\u5169\u8005\u76f8\u4f3c\u5ea6\u6108 \u9a57\u4e2d\u6709\u6700\u597d\u7684\u6548\u679c\uff1b\u7b2c\u4e8c\u662f target \u8207 nontarget \u5206\u4f48\u4e2d\u5fc3\uff0c\u5728 Softmax \u6211\u5011\u770b\u5230\u5206\u5e03\u72c0\u6cc1 \u8a13\u7df4\u968e\u6bb5) \uff0c\u4e5f\u89c0\u5bdf\u662f\u5426\u4f7f\u7528 softmax \u9810\u8a13\u7df4\u6a21\u578b\u5c0d\u4e09\u5143\u7d44\u640d\u5931\u8a13\u7df4\u65b9\u6cd5 (\u662f\u5426\u6709\u5716\u4e94 \u4e2d\u7684 softmax \u8a13\u7df4\u968e\u6bb5) \u7684\u5f71\u97ff\uff0c\u63a2\u8a0e\u662f\u5426\u4e09\u5143\u7d44\u640d\u5931\u5728\u5c11\u4e86 softmax \u7684\u689d\u4ef6\u4e0b\uff0c\u662f\u5426 \u5728\u672c\u8ad6\u6587\u4e2d\uff0c\u6211\u5011\u4ee5 x \u5411\u91cf\u67b6\u69cb\u70ba\u57fa\u790e\uff0c\u958b\u767c\u8a9e\u8005\u8fa8\u8b58\u7cfb\u7d71\uff0c\u900f\u904e\u6539\u8b8a\u640d\u5931\u51fd\u5f0f\u7684\u65b9 \u8207\u5e38\u614b\u5206\u4f48\u76f8\u4f3c\uff0c\u4f46\u6709\u7d93\u4e09\u5143\u7d44\u8abf\u9069\u5f8c\uff0ctarget \u8207 nontarget \u7684\u5206\u4f48\u4e2d\u5fc3\u7686\u6709\u62c9\u958b\u5f7c\u6b64\u4e4b \u9ad8\uff0c\u6b50\u6c0f\u8ddd\u96e2\u7684\u516c\u5f0f\u5982\u4e0b\uff1a || \u2212 || 5 \u6cd5\uff0c\u5c07\u539f\u672c\u5f8c\u7aef\u7e41\u7463\u7684\u5206\u985e\u6d41\u7a0b\u7c21\u5316\u81f3\u8a08\u7b97\u6b50\u6c0f\u8ddd\u96e2\u4f86\u6bd4\u8f03\u6e2c\u8a66\u8a9e\u97f3\u8207\u8a3b\u518a\u8005\u8a9e\u97f3\u7684 \u9593\u8ddd\u96e2\u7684\u60c5\u5f62\u767c\u751f\uff0c\u5c55\u73fe\u4e09\u5143\u7d44\u62c9\u958b\u4e0d\u540c\u8a9e\u8005\u4e4b\u9593\u8ddd\u96e2\u7684\u6548\u679c\u3002 \u4f7f\u540c\u985e\u8cc7\u6599\u7684\u5d4c\u5165\u5411\u91cf\u5728\u9ad8\u7dad\u7a7a\u9593\u4e2d\u805a\u96c6\u3002\u5728\u4e09\u5143\u7d44\u640d\u5931\u7684\u640d\u5931\u51fd\u5f0f\u4e2d\u6211\u5011\u53ef\u4ee5\u770b\u51fa\uff0c F (7) \u8a13\u7df4\u7684\u6838\u5fc3\u6982\u5ff5\u5728\u65bc\u62c9\u958b\u4e0d\u540c\u985e\u5225\u4e4b\u9593\u7684\u8ddd\u96e2\uff0c\u4f46\u7f3a\u5c11\u4f7f\u540c\u985e\u5225\u5167\u7684\u8cc7\u6599\u5167\u805a\u7684\u80fd\u529b\uff0c \u76f8\u4f3c\u5ea6\u3002\u6b64\u5916\uff0c\u4e0d\u540c\u65bc\u5e38\u898b\u7684\u8a9e\u8005\u8b58\u5225\u8a13\u7df4\u96c6\u8207\u6e2c\u8a66\u96c6\u4e2d\u7684\u8a9e\u8005\u76f8\u540c\uff0c\u70ba\u4e86\u4f7f\u5982\u667a\u6167 5 = 1( ( ) \u2212 ( )) 5 \u53ef\u4ee5\u907f\u514d\u6a21\u578b\u6536\u6582\u5728\u5c40\u90e8\u6700\u5c0f\u503c (Local Minima)\u3002 <34 \u4f8b\u5982\u5728\u4e09\u5143\u7d44\u7684\u9078\u64c7\u4e2d\u4e26\u672a\u5c0d\u6b63\u6a23\u672c\u7684\u9078\u64c7\u505a\u51fa\u9650\u5236\uff0c\u7f3a\u5c11\u62c9\u8fd1\u6b63\u6a23\u672c\u8207\u9328\u9ede\u4e4b\u9593\u7684\u8ddd \u5bb6\u5ead\u7522\u54c1\u7b49\u61c9\u7528\u4e0a\u66f4\u52a0\u65b9\u4fbf\uff0c\u5728\u8a3b\u518a\u65b0\u589e\u6216\u522a\u9664\u7528\u6236\u4e0a\u4e0d\u53d7\u9650\u5236\uff0c\u6211\u5011\u5728\u8a13\u7df4\u96c6\u8a9e\u8005</td></tr><tr><td>\u8207 \u70ba\u8981\u8a08\u7b97\u76f8\u4f3c\u5ea6\u7684\u5169\u5411\u91cf\uff0c\u7686\u70ba \u7dad\u3002\u9019\u500b\u65b9\u6cd5\u8207\u6211\u5011\u5728\u8a08\u7b97\u4e09\u5143\u7d44\u640d\u5931\u6642\u8a08\u7b97\u5d4c \u96e2\u7684\u529f\u80fd\uff0c\u6240\u4ee5\u5728\u5c11\u4e86 softmax \u6642\uff0c\u4e5f\u7f3a\u5c11\u4e86\u4f7f\u540c\u985e\u8cc7\u6599\u5167\u805a\u7684\u80fd\u529b\uff0c\u50c5\u6191\u4e09\u5143\u7d44\u640d\u5931 \u8207\u6e2c\u8a66\u96c6\u8a9e\u8005\u4e0d\u540c\u7684\u689d\u4ef6\u4e0b\u9032\u884c\u8fa8\u8b58\uff0c\u5229\u7528\u5d4c\u5165\u5411\u91cf\u5c0d\u8a3b\u518a\u8005\u5efa\u7acb\u8a9e\u8005\u6a21\u578b\uff0c\u4e26\u5728\u6e2c</td></tr><tr><td>\u5165\u5411\u91cf\u4e4b\u9593\u8ddd\u96e2\u7684\u65b9\u6cd5\u76f8\u540c\uff0c\u4e5f\u5e0c\u671b\u80fd\u85c9\u6b64\u66f4\u7b26\u5408\u8a13\u7df4\u6642\u7684\u8a34\u6c42\uff0c\u9054\u5230\u66f4\u597d\u7684\u6548\u679c\u3002 \u5716\u4e94\u3001\u6a21\u578b\u8a13\u7df4\u6d41\u7a0b\u5716\u3002 \u628a\u4e0d\u540c\u985e\u5225\u7684\u8ddd\u96e2\u62c9\u958b\uff0c\u662f\u5426\u80fd\u5c0d\u8a13\u7df4\u8cc7\u6599\u672a\u66fe\u51fa\u73fe\u904e\u7684\u8a9e\u97f3\u5f97\u51fa\u5224\u5225\u6027\u4f73\u7684\u5d4c\u5165\u5411\u91cf \u8a66\u6642\u627e\u51fa\u8207\u6e2c\u8a66\u8a9e\u97f3\u6700\u76f8\u4f3c\u7684\u8a3b\u518a\u8005\u505a\u70ba\u7cfb\u7d71\u5224\u5b9a\u7684\u7d50\u679c\u3002\u5728\u5be6\u9a57\u4e2d\uff0c\u6211\u5011\u6bd4\u8f03\u662f\u5426</td></tr><tr><td>\u662f\u503c\u5f97\u63a2\u8a0e\u7684\u554f\u984c\u3002 (a) Softmax \u4f7f\u7528\u4e09\u5143\u7d44\u640d\u5931\u7684\u5dee\u7570\uff0c\u767c\u73fe\u7121\u8ad6\u5728\u8a9e\u8005\u9a57\u8b49\u4e0a\u6216\u662f\u8a9e\u8005\u8fa8\u8b58\u4e0a\u7686\u6709\u6240\u5e6b\u52a9\uff0c\u5728 (b) Triplet (c) Softmax+Triplet</td></tr><tr><td>\u56db\u3001\u5be6\u9a57\u7d50\u679c\u8207\u5206\u6790 \u5728\u6b64\u7cfb\u7d71\u4e2d\uff0c\u4f7f\u7528 30 \u7dad\u7684 MFCC \u7279\u5fb5\u70ba\u8f38\u5165\uff0c\u97f3\u6846 (Frame) \u9577\u5ea6\u70ba 25 \u6beb\u79d2\u3001\u6bcf\u6b21\u79fb \u9996\u5148\uff0c\u8868\u4e8c\u70ba\u5728 VoxCeleb1 \u9a57\u8b49\u5206\u5272\u7684\u6e2c\u8a66\u8cc7\u6599\u4e2d\u5229\u7528\u5b98\u65b9\u63d0\u4f9b\u7684 trials \u9032\u884c\u9a57\u8b49\u7684\u5be6 \u5716\u516d\u3001\u5728 VoxCeleb1 \u8a9e\u8005\u9a57\u8b49\u4e4b\u5206\u6578\u5206\u4f48\u5716 VoxCeleb1 \u8b58\u5225\u5206\u5272\u6e2c\u8a66\u96c6\u55ae\u4e00\u8f38\u51fa(top-1)\u7684\u8fa8\u8b58\u6b63\u78ba\u7387\u70ba 59.57 %\uff0c\u524d\u4e94\u500b\u8f38\u51fa</td></tr><tr><td>\u52d5 10 \u6beb\u79d2\u3002\u5305\u542b\u7279\u5fb5\u63d0\u53d6\u3001\u6578\u64da\u589e\u5f37\u7b49\u524d\u8655\u7406\u63a1\u7528 Kaldi \u4f5c\u70ba\u5be6\u4f5c\u5de5\u5177\uff0c\u6a21\u578b\u8a13\u7df4\u8207 \u9a57\u7d50\u679c\uff0c\u4e5f\u662f\u5728\u4f7f\u7528\u4e09\u5143\u7d44\u640d\u5931\u6642\u4f7f\u7528\u7684\u9a57\u8b49\u8cc7\u6599\u7684\u7d50\u679c\uff0c\u5305\u542b 40 \u4f4d\u8a9e\u8005\u5171 37,720 \u7b46 (top-5)\u7684\u8fa8\u8b58\u6b63\u78ba\u7387\u5247\u53ef\u4ee5\u9054\u5230 80.32%\uff0c\u4f46\u53e6\u4e00\u65b9\u9762\u6211\u5011\u4e5f\u5efa\u8b70\u5728\u4f7f\u7528\u4e09\u5143\u7d44\u640d\u5931</td></tr><tr><td>(\u4e00)\u5be6\u9a57\u8a2d\u5b9a \u76f8\u4f3c\u5ea6\u6bd4\u5c0d\u5247\u662f\u4f7f\u7528 TensorFlow \u4f5c\u70ba\u5be6\u4f5c\u5de5\u5177\u3002 trials\uff0ctarget : nontarget \u70ba 1:1\u3002\u6211\u5011\u9664\u4e86 EER \u4e4b\u5916\uff0c\u4e5f\u4f7f\u7528 minDCF (Minimum Decision \u5728\u8a9e\u8005\u8fa8\u8b58\u4e0a\uff0c\u6211\u5011\u4ee5 Top-1 \u6e96\u78ba\u7387\u8207 Top-5 \u6e96\u78ba\u7387\u70ba\u8a55\u4f30\u6a19\u6e96\uff0cTop-1 \u6e96\u78ba\u7387\u8868\u793a\u50c5 \u6642\uff0c\u4f7f\u7528 softmax \u9810\u8a13\u7df4\u6a21\u578b\uff0c\u53ef\u4ee5\u4f7f\u6a21\u578b\u66f4\u7a69\u5b9a\u4e14\u6709\u66f4\u597d\u7684\u6548\u679c\u3002</td></tr><tr><td>minDCF \u4e5f\u662f\u6700\u4f4e\uff0c\u70ba 0.6278\u3002 \u5be6\u9a57\u7d50\u679c\u5982\u8868\u4e09\uff0c\u6211\u5011\u767c\u73fe\u6709 softmax \u9810\u8a13\u7df4\u4e14\u7d93\u4e09\u5143\u7d44\u640d\u5931\u8abf\u9069\u4e4b\u5f8c\uff0c\u5728\u8a9e\u8005\u8fa8\u8b58\u4e0a \u5716\u4e09\u3001\u4e09\u5143\u7d44\u8a13\u7df4\u968e\u6bb5\u4e4b\u6642\u5ef6\u795e\u7d93\u7db2\u8def\u67b6\u69cb\u5716\uff1a\u53f3\u5074\u6578\u5b57\u5404\u5c64\u7684\u8f38\u5165\u8207\u8f38\u51fa\u7dad\u5ea6\uff1bL \u5728\u8a9e\u8005\u8fa8\u8b58\u7684\u5be6\u9a57\u4e2d\uff0c\u6211\u5011\u4f7f\u7528 VoxCeleb2 \u9a57\u8b49\u96c6\u5171 5,994 \u4f4d\u8a9e\u8005\u4f86\u8a13\u7df4\u6642\u5ef6\u795e\u7d93\u7db2\u8def\uff0c \u6211\u5011\u5f9e VoxCeleb1 \u8b58\u5225\u5206\u5272\u4e4b\u9a57\u8b49\u96c6\u6bcf\u4f4d\u8a9e\u8005\u53d6 5 \u53e5\u97f3\u6a94\u9032\u884c\u8a3b\u518a\uff0c\u518d\u7531\u6e2c\u8a66\u96c6\u4f86\u6e2c\u8a66 \u6a21\u578b\uff0c\u6e2c\u8a66\u96c6\u5171\u5305\u542b 1,251 \u4f4d\u8a9e\u8005\u8207 8,251 \u53e5\u97f3\u6a94\uff0c\u6a21\u578b\u7684\u8a13\u7df4\u6d41\u7a0b\u5982\u5716\u4e94\u3002 \u5728 softmax \u8a13\u7df4\u968e\u6bb5\u904e\u7a0b\u4e2d\uff0c\u6211\u5011\u53d6\u8a13\u7df4\u96c6\u4e2d\u6bcf\u4f4d\u8a9e\u8005 10 \u7b46\u97f3\u6a94\u505a\u70ba\u9a57\u8b49\u96c6\uff0c\u4e14\u4f7f\u7528 \u65e9\u505c\u6cd5 (Early Stopping) \u7684\u6a5f\u5236\uff0c\u6bcf\u8a13\u7df4\u5b8c\u4e00\u8f2a\u8a13\u7df4\u8cc7\u6599\u53bb\u8a08\u7b97\u4e00\u6b21\u5728\u9a57\u8b49\u96c6\u4e0a\u7684\u640d\u5931\uff0c \u82e5\u8a72\u640d\u5931\u9023\u7e8c\u4e0d\u4e0b\u964d 5 \u6b21\u5247\u505c\u6b62\u8a13\u7df4\u3002 \u4e94\u76f8\u4f3c\u8005\u4e2d\u6709\u6e2c\u8a66\u8a9e\u97f3\u6240\u5c6c\u8a9e\u8005\u5373\u7b97\u6b63\u78ba\u3002 Cost Function)\u4f86\u8a55\u4f30\u7cfb\u7d71\uff0c\u53c3\u6578\u8a2d\u5b9a\u6bd4\u7167 [14] \u7528\u4e09\u5143\u7d44\u640d\u5931\u4e14\u52a0\u4e0a softmax \u9810\u8a13\u7df4\u6a21\u578b\u505a\u70ba\u8d77\u59cb\u6b0a\u91cd\u7684\u8a71\uff0cEER \u6709\u6548\u964d\u4f4e\u81f3 6.84%\uff0c \u7cfb\u7d71\u5224\u65b7\u76f8\u4f3c\u5ea6\u6700\u9ad8\u8005\u70ba\u6e2c\u8a66\u8a9e\u97f3\u6240\u5c6c\u7684\u8a9e\u8005\u624d\u7b97\u6b63\u78ba\uff0cTop-5 \u6e96\u78ba\u7387\u5247\u662f\u7cfb\u7d71\u5224\u5b9a\u524d \u53c3\u8003\u6587\u737b</td></tr><tr><td>\u70ba\u97f3\u6a94\u97f3\u6846\u6578\u3002 \u5728\u4e09\u5143\u7d44\u8a13\u7df4\u968e\u6bb5\u6642\uff0c\u4f7f\u7528\u4e09\u5143\u7d44\u640d\u5931\u8a13\u7df4\u6a21\u578b\uff0c\u6211\u5011\u4e26\u4e0d\u5f9e\u6240\u6709\u7684\u8a13\u7df4\u8cc7\u6599\u4e2d\u7d44\u6210\u4e09 \u4e5f\u6709\u6240\u9032\u6b65\uff0c\u76f8\u8f03\u65bc\u50c5\u4f7f\u7528 softmax \u8a13\u7df4 Top-1 \u6e96\u78ba\u7387\u63d0\u5347\u4e86\u7d04 5%\uff0cTop-5 \u6e96\u78ba\u7387\u66f4\u4e0a</td></tr><tr><td>\u5716\u56db\u3001\u4e09\u5143\u7d44\u640d\u5931\u4e2d\u5404\u985e\u8ca0\u6a23\u672c\u7684\u5b9a\u7fa9\u5716\uff1a\u5716\u4e2d\u4e4b\u8ca0\u6a23\u672c\u70ba\u534a\u96e3\u8ca0\u6a23\u672c\u3002 \u5143\u7d44\uff0c\u800c\u662f\u6bcf\u6b21\u53d6 90 \u4f4d\u8a9e\u8005\uff0c\u6bcf\u4f4d\u8a9e\u8005\u53d6 20 \u53e5\u97f3\u6a94\uff0c\u5171 1,800 \u53e5\u97f3\u6a94\uff0c\u6211\u5011\u7531\u9019\u4e9b\u97f3 \u5347\u4e86\u7d04 6.3%\u3002\u4f46\u5f9e\u5be6\u9a57\u7d50\u679c\u4e2d\u6211\u5011\u4e5f\u767c\u73fe\uff0c\u6c92\u6709\u4f7f\u7528 softmax \u9810\u8a13\u7df4\u6a21\u578b\u7684\u8a71\u6548\u80fd\u6703\u5927</td></tr><tr><td>\u6a94\u69cb\u6210\u4e09\u5143\u7d44\uff0c\u4f86\u8a13\u7df4\u8207\u66f4\u65b0\u6a21\u578b\uff0c\u9019\u6a23\u5c0d\u8a9e\u8005\u9032\u884c\u63a1\u6a23\u7684\u65b9\u5f0f\u76f8\u5c0d\u65bc\u4e8b\u5148\u627e\u51fa\u6240\u6709\u8a9e \u8868\u4e8c\u3001\u5728 VoxCeleb1 \u8a9e\u8005\u9a57\u8b49\u4e4b\u5be6\u9a57\u7d50\u679c \u5927\u7684\u964d\u4f4e\uff0c\u9019\u9ede\u5728\u8a9e\u8005\u8fa8\u8b58\u6642\u5c24\u5176\u660e\u986f\u3002</td></tr><tr><td>\u5716\u4e8c\u3001Softmax \u8a13\u7df4\u968e\u6bb5\u4e4b\u6642\u5ef6\u795e\u7d93\u7db2\u8def\u67b6\u69cb\u5716\uff1a\u53f3\u5074\u6578\u5b57\u5404\u5c64\u7684\u8f38\u5165\u8207\u8f38\u51fa\u7dad\u5ea6\uff1bL \u5728\u8a13\u7df4\u5b8c\u6210\u5f8c\uff0c\u70ba\u4e86\u66f4\u6e05\u695a\u5f97\u5206\u8fa8\u51fa\u540c\u8a9e\u8005\u7684\u8a9e\u97f3\u8207\u5176\u4ed6\u8a9e\u8005\u7684\u8a9e\u97f3\u4e4b\u9593\u7684\u5dee\u7570\uff0c\u6211\u5011 \u5982\u6b64\u4e00\u4f86\uff0c\u4e09\u5143\u7d44\u640d\u5931\u7684\u640d\u5931\u51fd\u6578\u5b9a\u7fa9\u5982\u4e0b\uff1a \u8005\u7684\u4e09\u5143\u7d44\uff0c\u80fd\u66f4\u5feb\u901f\u5730\u53cd\u61c9\u6a21\u578b\u7684\u73fe\u6cc1\uff0c\u4e26\u627e\u51fa\u5c0d\u6539\u5584\u7576\u524d\u6a21\u578b\u6709\u5e6b\u52a9\u7684\u4e09\u5143\u7d44\u3002\u5728 \u8a13\u7df4\u65b9\u6cd5 EER minDCF</td></tr><tr><td>\u70ba\u97f3\u6a94\u97f3\u6846\u6578\u3002 \u6368\u68c4\u4e86\u7528\u4f86\u63d0\u53d6\u5d4c\u5165\u5411\u91cf\u7684\u90a3\u5c64\u4ee5\u5f8c\u7684\u6bcf\u4e00\u5c64\u7db2\u8def\uff0c\u5982\u5716\u4e09\u6240\u793a\uff0c\u4e26\u6539\u70ba\u4f7f\u7528\u4e09\u5143\u7d44\u640d \u5931 [15] \u4f86\u8abf\u6574\u6a21\u578b\uff0c\u4e09\u5143\u7d44\u7684\u5b9a\u7fa9\u5982\u4e0b\uff1a )I<DJK) = [|| C \u2212 D || 5 5 \u2212 || C \u2212 F || 5 5 + ] N \u4e09\u5143\u7d44\u8a13\u7df4\u968e\u6bb5\u6642\uff0c\u4f7f\u7528 VoxCeleb1 \u9a57\u8b49\u5206\u5272\u7684\u6e2c\u8a66\u8cc7\u6599\u70ba\u9a57\u8b49\u96c6\uff0c\u4e26\u9032\u884c\u8a9e\u8005\u9a57\u8b49\uff0c (6) \u4ee5\u8a9e\u8005\u9a57\u8b49\u7684 EER (Equal Error Rate) \u70ba\u5224\u65b7\u662f\u5426\u505c\u6b62\u8a13\u7df4\u7684\u6a19\u6e96\u3002\u6bcf\u6b21\u8a13\u7df4\u5b8c\u4e00\u6b21\u63a1 Softmax 9.64 % 0.7174 \u8868\u4e09\u3001\u5728 VoxCeleb1 \u8a9e\u8005\u8fa8\u8b58\u4e4b\u5be6\u9a57\u7d50\u679c</td></tr><tr><td>\u9328\u9ede (Anchor)\uff1a \u8a13\u7df4\u96c6\u4e2d\u4e00\u8a9e\u8005 A \u7684\u771f\u97f3\u6a94\u3002 \u6a23\u7684\u8cc7\u6599\u5f8c\uff0c\u4fbf\u6703\u9032\u884c\u4e00\u6b21\u8a9e\u8005\u9a57\u8b49\uff0c\u7531\u65bc\u6bcf\u6b21\u63a1\u6a23\u7684\u8cc7\u6599\u50c5 1,800 \u7b46\uff0c\u6240\u4ee5\u9023\u7e8c 10 \u6b21 Triplet 10.39 % 0.8919 \u8a13\u7df4\u65b9\u6cd5 Top-1 \u6e96\u78ba\u7387 Top-5 \u6e96\u78ba\u7387 (\u4e94)\u640d\u5931\u51fd\u6578 \u6b63\u6a23\u672c (Positive)\uff1a \u8a9e\u8005 A \u8207\u9328\u9ede\u4e0d\u540c\u7684\u53e6\u4e00\u97f3\u6a94\u3002 (\u516d)\u8a3b\u518a\u8207\u76f8\u4f3c\u5ea6\u6bd4\u8f03 \u7684\u53c3\u6578\u66f4\u65b0 EER \u7686\u6c92\u6709\u4e0b\u964d\u624d\u505c\u6b62\u8a13\u7df4\u3002 Softmax + Triplet 6.84 % 0.6278 Softmax 54.59 % 73.67 %</td></tr><tr><td>\u5728\u521d\u671f\u8a13\u7df4\u6642\u5ef6\u795e\u7d93\u7db2\u8def\u6642\uff0c\u6211\u5011\u4f7f\u7528\u4ea4\u53c9\u71b5\u640d\u5931\u70ba\u640d\u5931\u51fd\u5f0f\u4f86\u66f4\u65b0\u6a21\u578b\uff0c\u55ae\u7b46\u8cc7\u6599\u6642 \u8ca0\u6a23\u672c (Negative)\uff1a\u975e\u8a9e\u8005 A \u7684\u53e6\u4e00\u8a9e\u8005\u4e4b\u97f3\u6a94\u3002 \u4e0d\u540c\u65bc\u50b3\u7d71\u8a9e\u8005\u8b58\u5225\u4e2d\u8a13\u7df4\u96c6\u8a9e\u8005\u8207\u6e2c\u8a66\u8a9e\u8005\u91cd\u8907\uff0c\u6211\u5011\u9810\u8a2d\u5927\u591a\u6642\u5019\u8a13\u7df4\u96c6\u4e2d\u7684\u8a9e\u8005 \u5728\u8d85\u53c3\u6578\u65b9\u9762\uff0c\u5be6\u9a57\u4e2d\u5b78\u7fd2\u7387\u7686\u8a2d\u70ba 0.001\uff0c\u512a\u5316\u65b9\u6cd5\u4f7f\u7528 Adam \u6f14\u7b97\u6cd5\uff0c\u4e09\u5143\u7d44\u640d\u5931</td></tr></table>" | |
| } | |
| } | |
| } | |
| } |