| { | |
| "paper_id": "2020", | |
| "header": { | |
| "generated_with": "S2ORC 1.0.0", | |
| "date_generated": "2023-01-19T07:27:05.784257Z" | |
| }, | |
| "title": "NSYSU+CHT Speaker Verification System for Far-Field Speaker Verification Challenge 2020", | |
| "authors": [ | |
| { | |
| "first": "\u5f35\u80b2\u5609", | |
| "middle": [ | |
| "\uf02a" | |
| ], | |
| "last": "\u3001\u9673\u5609\u5e73", | |
| "suffix": "", | |
| "affiliation": {}, | |
| "email": "" | |
| }, | |
| { | |
| "first": "\uf02a", | |
| "middle": [], | |
| "last": "\u3001\u856d\u5584\u6587", | |
| "suffix": "", | |
| "affiliation": {}, | |
| "email": "" | |
| }, | |
| { | |
| "first": "\uf02b", | |
| "middle": [], | |
| "last": "\u3001\u8a79\u535a\u4e1e", | |
| "suffix": "", | |
| "affiliation": {}, | |
| "email": "" | |
| }, | |
| { | |
| "first": "\uf02b", | |
| "middle": [], | |
| "last": "\u3001\u5442\u4ef2\u7406 \uf02b", | |
| "suffix": "", | |
| "affiliation": {}, | |
| "email": "" | |
| }, | |
| { | |
| "first": "Yu-Jia", | |
| "middle": [], | |
| "last": "Zhang", | |
| "suffix": "", | |
| "affiliation": {}, | |
| "email": "" | |
| }, | |
| { | |
| "first": "Chia-Ping", | |
| "middle": [], | |
| "last": "Chen", | |
| "suffix": "", | |
| "affiliation": {}, | |
| "email": "cpchen@mail.cse.nsysu.edu.tw" | |
| }, | |
| { | |
| "first": "Shan-Wen", | |
| "middle": [], | |
| "last": "Hsiao", | |
| "suffix": "", | |
| "affiliation": {}, | |
| "email": "" | |
| }, | |
| { | |
| "first": "Bo-Cheng", | |
| "middle": [], | |
| "last": "Chan", | |
| "suffix": "", | |
| "affiliation": {}, | |
| "email": "" | |
| }, | |
| { | |
| "first": "Chung-Li", | |
| "middle": [], | |
| "last": "Lu", | |
| "suffix": "", | |
| "affiliation": {}, | |
| "email": "" | |
| } | |
| ], | |
| "year": "", | |
| "venue": null, | |
| "identifiers": {}, | |
| "abstract": "In this paper, we describe the system Team NSYSU+CHT has implemented for the 2020 Far-field Speaker Verification Challenge (FFSVC 2020). The single systems", | |
| "pdf_parse": { | |
| "paper_id": "2020", | |
| "_pdf_hash": "", | |
| "abstract": [ | |
| { | |
| "text": "In this paper, we describe the system Team NSYSU+CHT has implemented for the 2020 Far-field Speaker Verification Challenge (FFSVC 2020). The single systems", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Abstract", | |
| "sec_num": null | |
| } | |
| ], | |
| "body_text": [ | |
| { | |
| "text": "\u81ea\u52d5\u8a9e\u8005\u9a57\u8b49(Automatic Speaker Verification, ASV)\u7cfb\u7d71\u96a8\u8457\u6df1\u5ea6\u5b78\u7fd2\u6280\u8853\u7684\u767c\u5c55\uff0c\u6709\u8457 \u986f\u8457\u7684\u63d0\u6607\uff0c\u6bcf\u4e00\u5e74\u8209\u884c\u7684\u76f8\u95dc\u7af6\u8cfd\u66f4\u662f\u4e0d\u52dd\u679a\u8209\uff0c\u4e0d\u7ba1\u662f NIST Speaker Recognition Evalution (SRE) (NIST, 2019) \uff0c\u6291\u6216\u662f\u9632\u6b62\u6b3a\u9a19\u8a9e\u97f3\u653b\u64ca\u7684 ASVspoof (Todisco et al., 2019) \uff0c \u9019\u4e9b\u7af6\u8cfd\u90fd\u4fc3\u4f7f\u81ea\u52d5\u8a9e\u8005\u9a57\u8b49\u7cfb\u7d71\u65e5\u8da8\u6210\u719f\u3002\u76ee\u524d\u6700\u88ab\u5ee3\u6cdb\u63a1\u7528\u7684\u81ea\u52d5\u8a9e\u8005\u9a57\u8b49\u7cfb\u7d71\u662f \u57fa\u65bc\u5d4c\u5165(Embedding)\u7684\u67b6\u69cb\uff0c\u8a72\u67b6\u69cb\u7531\u524d\u7aef\u7684\u7279\u5fb5\u63d0\u53d6\u5668(Feature Extractor)\uff0c\u4ee5\u53ca\u5f8c\u7aef \u7684\u8a55\u5206\u5668(Scorer)\u7d44\u5408\u800c\u6210\u3002\u524d\u7aef\u5728\u97f3\u6846\u5c64(Frame level layer)\u5c07\u539f\u59cb\u8f38\u5165\u63d0\u53d6\u6210\u9ad8\u968e\u7684\u8868 \u5fb5\uff0c\u4e26\u7d93\u7531\u6c60\u5316\u5c64\u6574\u5408\u97f3\u6846\u5c64\u8cc7\u8a0a\u6210\u70ba\u97f3\u6bb5\u5c64(Segment level layer)\uff0c\u800c\u5f8c\u900f\u904e\u5168\u9023\u63a5\u5c64 \u63d0\u53d6\u5d4c\u5165\u4e26 softmax \u8a08\u7b97\u6a5f\u7387\u4ee5\u5206\u985e\u8a9e\u8005\u3002\u524d\u7aef\u67b6\u69cb\u5f9e\u50b3\u7d71\u7684\u6df1\u5ea6\u795e\u7d93\u7db2\u8def(Deep Neural Network, DNN) / i \u5411\u91cf(i-vector) (McLaren, Lei, & Ferrer, 2015 )\uff0c\u4e00\u76f4\u5230\u8fd1\u5e74\u5728\u8a9e\u8005\u9a57\u8b49 \u6bd4\u8cfd\u4e2d\uff0c\u5927\u653e\u7570\u5f69\u7684\u6642\u5ef6\u795e\u7d93\u7db2\u8def(Time Delay Neural Network, TDNN) / x \u5411\u91cf (x-vector) (Snyder, Garcia-Romero, Sell, Povey & Khudanpur, 2018 ) \u8207\u4ee5\u5176\u70ba\u5ef6\u4f38\u7684\u67b6\u69cb\uff1a\u64f4\u5c55\u6642\u5ef6 \u795e\u7d93\u7db2\u8def (Extend-TDNN) (Snyder et al., 2019) Li, S., Lu, X., Takashima, R., Shen, P., Kawahara, T., & Kawai, H. (2018) . Improving Retrieved from https://www.nist.gov/itl/iad/mig/nist-2019-speaker-recognition-evaluation Okabe, K., Koshinaka, T., & Shinoda, K. (2018) . Attentive statistics pooling for deep speaker embedding. In arXiv preprint arXiv:1803.10963.", | |
| "cite_spans": [ | |
| { | |
| "start": 164, | |
| "end": 186, | |
| "text": "(Todisco et al., 2019)", | |
| "ref_id": null | |
| }, | |
| { | |
| "start": 472, | |
| "end": 501, | |
| "text": "(McLaren, Lei, & Ferrer, 2015", | |
| "ref_id": null | |
| }, | |
| { | |
| "start": 582, | |
| "end": 635, | |
| "text": "(Snyder, Garcia-Romero, Sell, Povey & Khudanpur, 2018", | |
| "ref_id": "BIBREF4" | |
| }, | |
| { | |
| "start": 672, | |
| "end": 693, | |
| "text": "(Snyder et al., 2019)", | |
| "ref_id": "BIBREF5" | |
| }, | |
| { | |
| "start": 694, | |
| "end": 767, | |
| "text": "Li, S., Lu, X., Takashima, R., Shen, P., Kawahara, T., & Kawai, H. (2018)", | |
| "ref_id": null | |
| }, | |
| { | |
| "start": 880, | |
| "end": 915, | |
| "text": "Koshinaka, T., & Shinoda, K. (2018)", | |
| "ref_id": null | |
| } | |
| ], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "\u7dd2\u8ad6 (Introduction)", | |
| "sec_num": "1." | |
| }, | |
| { | |
| "text": "OpenSLR. (2020). Open Speech and Language Resources. Retrieved from https://openslr.org/", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "\u7dd2\u8ad6 (Introduction)", | |
| "sec_num": "1." | |
| } | |
| ], | |
| "back_matter": [], | |
| "bib_entries": { | |
| "BIBREF0": { | |
| "ref_id": "b0", | |
| "title": "The kaldi speech recognition toolkit", | |
| "authors": [ | |
| { | |
| "first": "D", | |
| "middle": [], | |
| "last": "Povey", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "A", | |
| "middle": [], | |
| "last": "Ghoshal", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "G", | |
| "middle": [], | |
| "last": "Boulianne", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "L", | |
| "middle": [], | |
| "last": "Burget", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "O", | |
| "middle": [], | |
| "last": "Glembek", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "N", | |
| "middle": [], | |
| "last": "Goel", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": ".", | |
| "middle": [ | |
| "." | |
| ], | |
| "last": "Vesely", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "K", | |
| "middle": [], | |
| "last": "", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 2011, | |
| "venue": "Proceedings of IEEE 2011 Workshop on Automatic Speech Recognition and Understanding", | |
| "volume": "", | |
| "issue": "", | |
| "pages": "", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "Povey, D., Ghoshal, A., Boulianne, G., Burget, L., Glembek, O., Goel, N., . . . Vesely, K. (2011). The kaldi speech recognition toolkit. In Proceedings of IEEE 2011 Workshop on Automatic Speech Recognition and Understanding.", | |
| "links": null | |
| }, | |
| "BIBREF1": { | |
| "ref_id": "b1", | |
| "title": "Hi-mia: A far-field text-dependent speaker verification database and the baselines", | |
| "authors": [ | |
| { | |
| "first": "X", | |
| "middle": [], | |
| "last": "Qin", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "H", | |
| "middle": [], | |
| "last": "Bu", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "M", | |
| "middle": [], | |
| "last": "Li", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 2019, | |
| "venue": "", | |
| "volume": "", | |
| "issue": "", | |
| "pages": "", | |
| "other_ids": { | |
| "arXiv": [ | |
| "arXivpreprintarXiv:1912.01231" | |
| ] | |
| }, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "Qin, X., Bu, H., & Li, M. (2019). Hi-mia: A far-field text-dependent speaker verification database and the baselines. In arXiv preprint arXiv:1912.01231.", | |
| "links": null | |
| }, | |
| "BIBREF2": { | |
| "ref_id": "b2", | |
| "title": "Far-field end-to-end text-dependent speaker verification based on mixed training data with transfer learning and enrollment data augmentation", | |
| "authors": [ | |
| { | |
| "first": "X", | |
| "middle": [], | |
| "last": "Qin", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "D", | |
| "middle": [], | |
| "last": "Cai", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "M", | |
| "middle": [], | |
| "last": "Li", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 2019, | |
| "venue": "Proceedings of Interspeech 2019", | |
| "volume": "", | |
| "issue": "", | |
| "pages": "4045--4049", | |
| "other_ids": { | |
| "DOI": [ | |
| "10.21437/Interspeech.2019-1542" | |
| ] | |
| }, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "Qin, X., Cai, D., & Li, M. (2019). Far-field end-to-end text-dependent speaker verification based on mixed training data with transfer learning and enrollment data augmentation. In Proceedings of Interspeech 2019, 4045-4049. doi: 10.21437/Interspeech.2019-1542", | |
| "links": null | |
| }, | |
| "BIBREF3": { | |
| "ref_id": "b3", | |
| "title": "The ffsvc 2020 evaluation plan", | |
| "authors": [ | |
| { | |
| "first": "X", | |
| "middle": [], | |
| "last": "Qin", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "M", | |
| "middle": [], | |
| "last": "Li", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "H", | |
| "middle": [], | |
| "last": "Bu", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "R", | |
| "middle": [ | |
| "K" | |
| ], | |
| "last": "Das", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "W", | |
| "middle": [], | |
| "last": "Rao", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "S", | |
| "middle": [], | |
| "last": "Narayanan", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "H", | |
| "middle": [], | |
| "last": "Li", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 2020, | |
| "venue": "", | |
| "volume": "", | |
| "issue": "", | |
| "pages": "", | |
| "other_ids": { | |
| "arXiv": [ | |
| "arXivpreprintarXiv:2002.00387" | |
| ] | |
| }, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "Qin, X., Li, M., Bu, H., Das, R. K., Rao, W., Narayanan, S., & Li, H. (2020). The ffsvc 2020 evaluation plan. In arXiv preprint arXiv:2002.00387.", | |
| "links": null | |
| }, | |
| "BIBREF4": { | |
| "ref_id": "b4", | |
| "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": "Proceedings of IEEE International Conference on Acoustics, Speech and Signal Processing", | |
| "volume": "", | |
| "issue": "", | |
| "pages": "5329--5333", | |
| "other_ids": { | |
| "DOI": [ | |
| "10.1109/ICASSP.2018.8461375" | |
| ] | |
| }, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "Snyder, D., Garcia-Romero, D., Sell, G., Povey, D., & Khudanpur, S. (2018). X-vectors: Robust dnn embeddings for speaker recognition. In Proceedings of IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP 2018), 5329-5333. doi: 10.1109/ICASSP.2018.8461375", | |
| "links": null | |
| }, | |
| "BIBREF5": { | |
| "ref_id": "b5", | |
| "title": "The jhu speaker recognition system for the voices 2019 challenge", | |
| "authors": [ | |
| { | |
| "first": "D", | |
| "middle": [], | |
| "last": "Snyder", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "J", | |
| "middle": [], | |
| "last": "Villalba", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "N", | |
| "middle": [], | |
| "last": "Chen", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "D", | |
| "middle": [], | |
| "last": "Povey", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "G", | |
| "middle": [], | |
| "last": "Sell", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "N", | |
| "middle": [], | |
| "last": "Dehak", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "S", | |
| "middle": [], | |
| "last": "Khudanpur", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 2019, | |
| "venue": "Proeedings of Interspeech", | |
| "volume": "", | |
| "issue": "", | |
| "pages": "2468--2472", | |
| "other_ids": { | |
| "DOI": [ | |
| "10.21437/Interspeech.2019-2979" | |
| ] | |
| }, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "Snyder, D., Villalba, J., Chen, N., Povey, D., Sell, G., Dehak, N., & Khudanpur, S. (2019). The jhu speaker recognition system for the voices 2019 challenge. In Proeedings of Interspeech 2019, 2468-2472. doi: 10.21437/Interspeech.2019-2979", | |
| "links": null | |
| }, | |
| "BIBREF7": { | |
| "ref_id": "b7", | |
| "title": "Asvspoof2019: Future horizons in spoofed and fake audio detection", | |
| "authors": [ | |
| { | |
| "first": "K", | |
| "middle": [ | |
| "A" | |
| ], | |
| "last": "Lee", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 2019, | |
| "venue": "", | |
| "volume": "", | |
| "issue": "", | |
| "pages": "", | |
| "other_ids": { | |
| "arXiv": [ | |
| "arXiv:1904.05441" | |
| ] | |
| }, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "Lee, K. A. (2019). Asvspoof2019: Future horizons in spoofed and fake audio detection. In arXiv preprint arXiv:1904.05441.", | |
| "links": null | |
| }, | |
| "BIBREF8": { | |
| "ref_id": "b8", | |
| "title": "Additive margin softmax for face verification", | |
| "authors": [ | |
| { | |
| "first": "F", | |
| "middle": [], | |
| "last": "Wang", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "J", | |
| "middle": [], | |
| "last": "Cheng", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "W", | |
| "middle": [], | |
| "last": "Liu", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "H", | |
| "middle": [], | |
| "last": "Liu", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 2018, | |
| "venue": "IEEE Signal Processing Letters", | |
| "volume": "25", | |
| "issue": "7", | |
| "pages": "926--930", | |
| "other_ids": { | |
| "DOI": [ | |
| "10.1109/LSP.2018.2822810" | |
| ] | |
| }, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "Wang, F., Cheng, J., Liu, W., & Liu, H. (2018). Additive margin softmax for face verification. IEEE Signal Processing Letters, 25(7), 926-930. doi: 10.1109/LSP.2018.2822810", | |
| "links": null | |
| }, | |
| "BIBREF9": { | |
| "ref_id": "b9", | |
| "title": "Utterance-level aggregation for speaker recognition in the wild", | |
| "authors": [ | |
| { | |
| "first": "W", | |
| "middle": [], | |
| "last": "Xie", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "A", | |
| "middle": [], | |
| "last": "Nagrani", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "J", | |
| "middle": [ | |
| "S" | |
| ], | |
| "last": "Chung", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "A", | |
| "middle": [], | |
| "last": "Zisserman", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 2019, | |
| "venue": "Proceedings of 2019 IEEE International Conference on Acoustics, Speech and Signal Processing", | |
| "volume": "", | |
| "issue": "", | |
| "pages": "5791--5795", | |
| "other_ids": { | |
| "DOI": [ | |
| "10.1109/ICASSP.2019.8683120" | |
| ] | |
| }, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "Xie, W., Nagrani, A., Chung, J. S., & Zisserman, A. (2019). Utterance-level aggregation for speaker recognition in the wild. In Proceedings of 2019 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP 2019), 5791-5795. doi: 10.1109/ICASSP.2019.8683120", | |
| "links": null | |
| }, | |
| "BIBREF10": { | |
| "ref_id": "b10", | |
| "title": "Ghostvlad for set-based face recognition", | |
| "authors": [ | |
| { | |
| "first": "Y", | |
| "middle": [], | |
| "last": "Zhong", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "R", | |
| "middle": [], | |
| "last": "Arandjelovi\u0107", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "A", | |
| "middle": [], | |
| "last": "Zisserman", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 2018, | |
| "venue": "Proceedings of Asian Conference on Computer Vision", | |
| "volume": "", | |
| "issue": "", | |
| "pages": "35--50", | |
| "other_ids": { | |
| "DOI": [ | |
| "10.1007/978-3-030-20890-5_3" | |
| ] | |
| }, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "Zhong, Y., Arandjelovi\u0107, R., & Zisserman, A. (2018). Ghostvlad for set-based face recognition. In Proceedings of Asian Conference on Computer Vision 2018, 35-50. doi: 10.1007/978-3-030-20890-5_3", | |
| "links": null | |
| }, | |
| "BIBREF11": { | |
| "ref_id": "b11", | |
| "title": "Self-attentive speaker embeddings for text-independent speaker verification", | |
| "authors": [ | |
| { | |
| "first": "Y", | |
| "middle": [], | |
| "last": "Zhu", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "T", | |
| "middle": [], | |
| "last": "Ko", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "D", | |
| "middle": [], | |
| "last": "Snyder", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "B", | |
| "middle": [], | |
| "last": "Mak", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "D", | |
| "middle": [], | |
| "last": "Povey", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 2018, | |
| "venue": "Proceedings of", | |
| "volume": "", | |
| "issue": "", | |
| "pages": "3573--3577", | |
| "other_ids": { | |
| "DOI": [ | |
| "10.21437/Interspeech.2018-1158" | |
| ] | |
| }, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "Zhu, Y., Ko, T., Snyder, D., Mak, B., & Povey, D. (2018). Self-attentive speaker embeddings for text-independent speaker verification. In Proceedings of Interspeech 2018, 3573-3577. doi: 10.21437/Interspeech.2018-1158", | |
| "links": null | |
| } | |
| }, | |
| "ref_entries": { | |
| "FIGREF0": { | |
| "num": null, | |
| "text": "very deep time-delay neural network with vertical-attention for effectively training ctc-based asr systems. In Proceedings of 2018 IEEE Spoken Language Technology Workshop (SLT), 77-83. doi: 10.1109/SLT.2018.8639675 Liu, W., Wen, Y., Yu, Z., Li, M., Raj, B., & Song, L. (2017). Sphereface: Deep hypersphere embedding for face recognition. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 212-220. doi: 10.1109/CVPR.2017.713 Liu, W., Wen, Y., Yu, Z., & Yang, M. (2016). Large-margin softmax loss for convolutional neural networks. In Proceedings of ICML 2016, 48, 507-516. Liu, Y., He, L., & Liu, J. (2019). Large margin softmax loss for speaker verification. In arXiv preprint arXiv:1904.03479. Maaten, L. v. d., & Hinton, G. (2008). Visualizing data using t-sne. Journal of machine learning research, 9, 2579-2605. McLaren, M., Lei, Y., & Ferrer, L. (2015). Advances in deep neural network approaches to speaker recognition. In Proceedings of IEEE International Conference on Acoustics, ., Chung, J. S., Xie, W., & Zisserman, A. (2020). Voxceleb: Large-scale speaker verification in the wild. Computer Speech & Language, 60, 101027. doi: 10.1016/j.csl.2019", | |
| "type_str": "figure", | |
| "uris": null | |
| }, | |
| "TABREF0": { | |
| "text": "\uff1b\u800c\u539f\u5148\u57fa\u65bc\u5f71\u50cf\u8fa8\u8b58\u6240\u5efa\u7acb\u7684\u5377\u7a4d\u795e\u7d93 \u7db2\u8def\uff0c\u88ab\u8a8d\u70ba\u5728\u8a9e\u8005\u8fa8\u8b58\u4efb\u52d9\u4e2d\u4e5f\u80fd\u53d6\u5f97\u4e0d\u932f\u7684\u8868\u73fe\uff0c\u50cf\u662f\u6b98\u5dee\u795e\u7d93\u7db2\u8def(Residual Neural Network, ResNet) (Nagrani, Chung, Xie & Zisserman, 2020; Xie, Nagrani, Chung & Zisserman, 2019; Qin, Bu & Li, 2019)\uff1b\u53e6\u4e00\u65b9\u9762\uff0c\u4e5f\u6709\u8a31\u591a\u7814\u7a76\u662f\u91dd\u5c0d\u6c60\u5316\u5c64\u8207\u640d\u5931\u51fd\u6578 \u4f86\u4f5c\u6539\u9032\uff0c\u6c60\u5316\u5c64\u9664\u4e86\u5728\u6642\u5ef6\u795e\u7d93\u7db2\u8def\u4e2d\u6700\u5e38\u4f7f\u7528\u7684\u7d71\u8a08\u6c60\u5316\u5c64(Statistic Pooling)\u4e4b\u5916\uff0c \u81ea\u6ce8\u610f\u6c60\u5316\u5c64(Self-attentive Pooling) (Okabe, Koshinaka & Shinoda, 2018; Zhu, Ko, Snyder, Mak & Povey, 2018)\uff0cNetVLAD (Chen et al., 2018)\u90fd\u80fd\u4f7f\u7cfb\u7d71\u5728\u6574\u5408\u97f3\u6846\u5c64\u8cc7\u8a0a\u7684\u6548\u679c\u66f4", | |
| "num": null, | |
| "type_str": "table", | |
| "html": null, | |
| "content": "<table><tr><td>62</td><td>NSYSU+CHT\u5718\u968a\u65bc2020\u9060\u5834\u8a9e\u8005\u9a57\u8b49\u6bd4\u8cfd\u4e4b\u8a9e\u8005\u9a57\u8b49\u7cfb\u7d71 NSYSU+CHT\u5718\u968a\u65bc2020\u9060\u5834\u8a9e\u8005\u9a57\u8b49\u6bd4\u8cfd\u4e4b\u8a9e\u8005\u9a57\u8b49\u7cfb\u7d71 NSYSU+CHT\u5718\u968a\u65bc2020\u9060\u5834\u8a9e\u8005\u9a57\u8b49\u6bd4\u8cfd\u4e4b\u8a9e\u8005\u9a57\u8b49\u7cfb\u7d71 NSYSU+CHT\u5718\u968a\u65bc2020\u9060\u5834\u8a9e\u8005\u9a57\u8b49\u6bd4\u8cfd\u4e4b\u8a9e\u8005\u9a57\u8b49\u7cfb\u7d71 NSYSU+CHT\u5718\u968a\u65bc2020\u9060\u5834\u8a9e\u8005\u9a57\u8b49\u6bd4\u8cfd\u4e4b\u8a9e\u8005\u9a57\u8b49\u7cfb\u7d71</td><td>57 \u5f35\u80b2\u5609 \u7b49 59 \u5f35\u80b2\u5609 \u7b49 61 \u5f35\u80b2\u5609 \u7b49 63 \u5f35\u80b2\u5609 \u7b49 65 \u5f35\u80b2\u5609 \u7b49</td></tr><tr><td colspan=\"3\">Discriminant Analysis, PLDA) (Kenny, 2010)\u6210\u70ba\u4e86\u6700\u5e38\u4f7f\u7528\u7684\u65b9\u6cd5\u4e4b\u4e00\u3002 \u8fd1\u5e74\u4f86\u96a8\u8457\u7269\u806f\u7db2\u8a2d\u5099\u8207\u667a\u6167\u5bb6\u5c45\u7522\u54c1\u7684\u666e\u53ca\uff0c\u77ed\u8a9e\u97f3\u6307\u4ee4\u7684\u8655\u7406\uff0c\u4ee5\u53ca\u5728\u9060\u5834\u566a 2.1.2 \u57fa\u65bc\u5377\u7a4d\u795e\u7d93\u7db2\u8def (CNN-based) \u6839\u64da (Nagrani et al., 2020; Xie et al., 2019; Qin, Bu & Li, 2019) \u7684\u7814\u7a76\uff0c\u90fd\u8868\u660e\u6b98\u5dee\u795e\u7d93 \u8def\u826f\u597d\u8403\u53d6\u7279\u5fb5\u7684\u80fd\u529b\u3002 \u8868 2. \u6642\u5ef6\u6b98\u5dee\u795e\u7d93\u7db2\u8def\u67b6\u69cb \u901a\u9053\u6578\u76f8\u540c\uff0c \uff0c \uff0c \u662f\u7531\u7db2\u8def\u8a13\u7df4\u5f97\u5230\u7684\u53c3\u6578\uff0c\u8a72\u516c\u5f0f\u7684\u524d\u534a\u90e8\u70ba softmax\uff0c\u8868 \u8f38\u5165 \u5c6c\u65bc\u7fa4 \u7684\u6a5f\u7387\uff0c\u5f8c\u534a\u90e8\u70ba\u8a08\u7b97 \u8207\u7fa4\u4e2d\u5fc3\u7684\u8ddd\u96e2\uff0c\u4e26\u4ee5\u524d\u534a\u90e8\u8a08\u7b97\u51fa\u4f86\u7684 \u7d50\u5408\u65b9\u5f0f\u5982\u5716 1\uff0c\u6211\u5011\u4f9d\u7167\u6a5f\u7387\u7dda\u6027\u5224\u5225\u5206\u6790\u8207\u6a5f\u7387\u7dda\u6027\u5224\u5225\u5206\u6790\u8abf\u9069\u8a55\u5206\u5668\uff0c\u5c07\u4e00\u500b \u6a21\u578b\u62c6\u5206\u6210\u5169\u500b\u5b50\u7cfb\u7d71\uff0c\u4e26\u4f7f\u7528 BOSARIS toolkit (Brummer & De Villiers, 2013) \u4f86\u6821\u6b63 3.2 \u8cc7\u6599\u589e\u5f37 (Data Augmentation) \u8a13\u7df4\u8cc7\u6599\u63a1\u7528\u8cc7\u6599\u589e\u5f37\uff0c\u4e00\u76f4\u4ee5\u4f86\u90fd\u662f\u88ab\u4f7f\u7528\u65bc\u589e\u5f37\u8a9e\u8005\u5d4c\u5165\u6a21\u578b\u7684\u5f37\u5065\u6027(Robustness)\uff0c \u795e\u7d93\u7db2\u8def-G \u548c\u878d\u5408\u6a21\u578b\u5728\u9a57\u8b49\u96c6\u4e0a\u6e2c\u8a66\u4e26\u4e0a\u50b3\u6210\u7e3e\uff0c\u4e26\u4ee5\u6709\u7121\u7d93\u904e\u5206\u6578\u878d\u5408\u4f86\u5340\u5206\u4e0a\u50b3 \u7684\u7cfb\u7d71\uff0c\u7d93\u904e\u5206\u6578\u878d\u5408\u7684\u878d\u5408\u6a21\u578b\u4f5c\u70ba\u6211\u5011\u7af6\u8cfd\u7684 Primary System 1\uff0c\u800c\u9019\u4e5f\u662f\u6211\u5011\u5728\u8a72 \u8868 6. \u5404\u96dc\u8a0a\u8868\u793a\u60c5\u5883\u8207\u5f71\u97ff\u7684\u7d50\u679c \u53c3\u8003\u6587\u737b (References) Single System 2 \u5247\u662f\u63a1\u53d6\u8207 Single System 1 \u76f8\u540c\u7684\u6a21\u578b\u67b6\u69cb\uff0c\u552f\u4e00\u7684\u4e0d\u540c\u5728\u65bc\u6c60\u5316 \u5c64\u5f9e\u539f\u4f86\u7684\u7d71\u8a08\u6c60\u5316\u5c64\u66ff\u63db\u6210 GhostVLAD\uff0c\u8a13\u7df4\u8cc7\u6599\u8207\u8d85\u53c3\u6578\u4e26\u7121\u505a\u4efb\u4f55\u7684\u66f4\u52d5\u3002\u6b64\u5916\uff0c [Table 6. Each noise condition and result] Br\u00fcmmer, N., & De Villiers, E. (2013). The bosaris toolkit: Theory, algorithms and code for</td></tr><tr><td colspan=\"3\">\u97f3\u7684\u771f\u5be6\u4f7f\u7528\u5834\u666f\u4e0b\uff0c\u6210\u70ba\u4e86\u81ea\u52d5\u8a9e\u8005\u9a57\u8b49\u7cfb\u7d71\u7684\u65b0\u6311\u6230\uff0c\u800c\u9304\u97f3\u8a2d\u5099\u7684\u4e0d\u5339\u914d\uff0c\u66f4\u518d \u7db2\u8def\u67b6\u69cb\uff0c\u5728\u6709\u566a\u97f3\u8207\u8ff4\u97ff (reverberate) \u7684\u9060\u5834\u74b0\u5883\u4e0b\uff0c\u5c0d\u65bc\u7279\u5fb5\u7684\u64f7\u53d6\u662f\u76f8\u7576\u51fa\u8272\u7684\uff0c [Table 2. Network architecture of TDResNet] softmax \u503c\u4f5c\u70ba\u8a72\u8ddd\u96e2\u7684\u6b0a\u91cd\uff0c\u518d\u5c07\u6240\u6709\u7d50\u679c\u76f8\u52a0\uff0c\u6700\u7d42\u628a\u6240\u6709\u7684\u7fa4\u4e32\u9023\u6210\u6700\u5f8c\u7684\u8f38\u51fa\u5411 \u6211\u5011\u7cfb\u7d71\u5206\u6578\u4e4b\u9593\u7684\u6b0a\u91cd\uff0c\u6821\u6b63\u8cc7\u6599\u96c6\u63a1\u7528 FFSVC 2020 \u958b\u767c\u96c6\u3002 \u800c\u8a72\u7bc7\u8ad6\u6587(Qin, Cai & Li, 2019)\u4e2d\u6709\u63d0\u53ca\uff0c\u5728\u9060\u5834\u7684\u74b0\u5883\u4e0b\uff0c\u8a13\u7df4\u8cc7\u6599\u8207\u6e2c\u8a66\u8cc7\u6599\u5b58\u5728 \u7af6\u8cfd\u7684\u6700\u4f73\u7cfb\u7d71\uff0c\u53e6\u5916\u6c92\u6709\u7d93\u904e\u5206\u6578\u878d\u5408\u7684\u6642\u5ef6\u6b98\u5dee\u795e\u7d93\u7db2\u8def\u8207\u6642\u5ef6\u6b98\u5dee\u795e\u7d93\u7db2\u8def-G \u5247 \u6211\u5011\u4e5f\u4f7f\u7528 t-distributed stochastic neighbor embedding (t-SNE) (Maaten & Hinton, 2008) surviving the new dcf. arXiv preprint arXiv:1304.2865.</td></tr><tr><td colspan=\"3\">\u52a0\u6df1\u4e86\u8b58\u5225\u7684\u96e3\u5ea6\uff0c\u70ba\u4e86\u63a8\u52d5\u8a72\u60c5\u666f\u4e0b\u7684\u81ea\u52d5\u8a9e\u8005\u9a57\u8b49\u7cfb\u7d71\u7814\u7a76\uff0cFFSVC 2020 (Qin et al., \u56e0\u6b64\u63a1\u7528\u6b98\u5dee\u795e\u7d93\u7db2\u8def\u4f5c\u70ba\u6211\u5011\u57fa\u65bc\u5377\u7a4d\u795e\u7d93\u7db2\u8def\u7684\u97f3\u6846\u5c64\u67b6\u69cb\uff0c\u4e26\u53c3\u8003 (Nagrani et al., \u91cf \uff0c\u800c GhostVLAD \u7684\u6539\u9032\u5728\u65bc\u5411\u5f8c\u50b3\u905e\u6642\uff0c\u6709\u4e9b\u7fa4\u4e26\u4e0d\u6703\u88ab\u5305\u542b\u5728\u6700\u5f8c\u7684\u8f38\u51fa\u7576\u4e2d\uff0c \u8457\u4e0d\u5339\u914d\u7684\u73fe\u8c61\uff0c\u56e0\u6b64\u70ba\u4e86\u8981\u6a21\u64ec\u9060\u5834\u7684\u74b0\u5883\uff0c\u6211\u5011\u91dd\u5c0d\u5e7e\u500b\u8f03\u5927\u7684\u8cc7\u6599\u96c6\uff0c\u4f7f\u7528 \u4f5c\u70ba Single System 1\u30012\uff0c \u5176\u4e2d\u53c8\u4ee5 Single System 2 \u8868\u73fe\u8f03\u4f73\u3002\u6240\u6709\u7d50\u679c\u5982\u8868 4 \u6240\u793a\u3002 \u4f86\u5206\u5225\u5c0d Single System 1\u30012 \u7684\u9ad8\u7dad\u5ea6\u5d4c\u5165\u8996\u89ba\u5316\uff0c\u4ee5\u6b64\u8a55\u4f30\u4e0d\u540c\u6c60\u5316\u5c64\u5c0d\u65bc\u6700\u7d42\u5d4c\u5165 Chen, J., Cai, W., Cai, D., Cai, Z., Zhong, H., & Li, M. (2018). End-to-end language</td></tr><tr><td colspan=\"3\">2020)\u56e0\u61c9\u800c\u751f\u3002 \u56e0\u6b64\uff0c\u672c\u8ad6\u6587\u65e8\u5728\u53c3\u52a0 FFSVC 2020\uff0c\u4e26\u91dd\u5c0d\u4efb\u52d9\u4e00\uff1a\u55ae\u4e00\u9ea5\u514b\u98a8\u9663\u5217\u7684\u9060\u5834\u6587\u672c \u76f8\u95dc\u8a9e\u8005\u9a57\u8b49(Far-Field Text-Dependent Speaker Verification from single microphone array) \u8207\u4efb\u52d9\u4e8c\uff1a\u55ae\u4e00\u9ea5\u514b\u98a8\u9663\u5217\u7684\u9060\u5834\u6587\u672c\u7121\u95dc\u8a9e\u8005\u9a57\u8b49(Far-Field Text-Independent Speaker Verification from single microphone array)\u63a1\u53d6\u4e0d\u540c\u7684\u89e3\u6c7a\u65b9\u6cd5\u3002\u6211\u5011\u4ee5\u57fa\u65bc\u6642\u5ef6\u795e\u7d93\u7db2\u8def \u7684\u64f4\u5c55\u6642\u5ef6\u795e\u7d93\u7db2\u8def\uff0c\u8207\u57fa\u65bc\u5377\u7a4d\u795e\u7d93\u7db2\u8def\u7684\u6b98\u5dee\u795e\u7d93\u7db2\u8def\uff0c\u5efa\u7acb\u4e86\u524d\u7aef\u7684\u7279\u5fb5\u63d0\u53d6\u5668\u3002 \u8072\u5b78\u7279\u5fb5\u63a1\u7528 FBank (Filter Bank)\u914d\u4e0a\u97f3\u8abf(Pitch)\u3002\u800c\u5f8c\u4e5f\u91dd\u5c0d\u6b98\u5dee\u795e\u7d93\u7db2\u8def\u67b6\u69cb\u9032\u884c\u4fee \u6539\uff0c\u5c07\u5176\u8207\u64f4\u5c55\u6642\u5ef6\u795e\u7d93\u7db2\u8def\u7d50\u5408\uff0c\u6210\u70ba\u4e00\u500b\u65b0\u7684\u7db2\u8def\u67b6\u69cb\u7a31\u70ba\u6642\u5ef6\u6b98\u5dee\u795e\u7d93\u7db2\u8def(Time Delay Residual Neural Network, TDResNet)\uff0c\u4e26\u4f7f\u7528\u6a5f\u7387\u7dda\u6027\u5224\u5225\u5206\u6790\u4f5c\u70ba\u5f8c\u7aef\u8a55\u5206\u5668\uff0c \u5206\u5225\u5be6\u9a57\u4e0a\u8ff0\u6a21\u578b\u67b6\u69cb\u5728\u5404\u4efb\u52d9\u4e0a\u7684\u8868\u73fe\u3002\u6b64\u5916\uff0c\u6211\u5011\u4e5f\u91dd\u5c0d\u6c60\u5316\u5c64\u505a\u6539\u8b8a\uff0c\u5c07\u539f\u5148\u7684 \u7d71\u8a08\u6c60\u5316\u5c64\u66ff\u63db\u6210 NetVLAD \u7684\u6539\u9032\uff1aGhostVLAD (Zhong, Arandjelovic & Zisserman, 2018)\u3002\u66f4\u591a\u7684\u5be6\u4f5c\u7d30\u7bc0\u5c07\u6703\u5728\u5f8c\u7e8c\u7684\u7ae0\u7bc0\u8a73\u7d30\u8aaa\u660e\u3002 2020) \u6240\u63d0\u5230\u7684 thin-ResNet \u67b6\u69cb\uff0c\u5be6\u4f5c\u4e86\u53c3\u6578\u91cf\u8f03\u5c11\u7684\u6b98\u5dee\u795e\u7d93\u7db2\u8def\uff0c\u63a5\u8457\u540c\u6a23\u5c07\u97f3 \u6846\u5c64\u7684\u8f38\u51fa\u7d93\u904e\u7d71\u8a08\u6c60\u5316\u5c64\u6574\u5408\uff0c\u800c\u5f8c\u7684\u97f3\u6bb5\u5c64\u8207\u64f4\u5c55\u6642\u5ef6\u795e\u7d93\u7db2\u8def\u4e0d\u540c\uff0c\u53ea\u63a1\u7528\u4e00\u5c64 \u5168\u9023\u63a5\u5c64\uff0c\u4e26\u5c07 softmax \u66ff\u63db\u6210 AM-softmax \u8a08\u7b97\u6a5f\u7387\u9032\u884c\u5206\u985e\u3002\u6bcf\u4e00\u5c64\u7686\u7d93\u904e\u6279\u91cf\u6a19 \u6e96\u5316\u8207 ReLU \u6fc0\u6d3b\u51fd\u6578\u3002\u6bcf\u4e00\u500b\u6b98\u5dee\u5340\u584a (Residual Block) \u7686\u4f7f\u7528\u6b98\u5dee\u9023\u7d50 (Residual connect) \u9023\u63a5\uff0c\u6700\u7d42\u67b6\u69cb\u5716\u5982\u8868 1\u3002 \u8868 1. \u6b98\u5dee\u795e\u7d93\u7db2\u8def\u67b6\u69cb \u5982\u6b64\u4e00\u4f86\uff0c\u518d\u8a13\u7df4\u7db2\u8def\u6642\uff0c\u80fd\u8b93\u7db2\u8def\u81ea\u4e3b\u5b78\u7fd2\u54ea\u4e9b\u7279\u5fb5\u4f5c\u7528\u8f03\u4f4e\uff0c\u61c9\u8a72\u88ab\u5206\u985e\u5230\u9700\u8981\u88ab \u6392\u9664\u7684\u7fa4\u4e2d\uff0c\u800c\u56e0\u70ba\u88ab\u6392\u9664\u7684\u7fa4\u4e0d\u6703\u53c3\u8207\u5230\u6574\u500b\u7db2\u8def\u6b0a\u91cd\u7684\u66f4\u65b0\uff0c\u56e0\u6b64\u5728\u8a13\u7df4\u4e2d\u4f3c\u6709\u975e \u6709\uff0c\u6240\u4ee5\u53c8\u88ab\u7a31\u70ba Ghost \u7fa4\uff0c\u9019\u4e5f\u662f\u9019\u500b\u65b9\u6cd5\u7684\u7531\u4f86\uff0c\u800c Ghost \u7fa4\u4e5f\u662f\u4e8b\u5148\u8a2d\u5b9a\u597d\u7684\u8d85 \u53c3\u6578\uff0c\u5b83\u5c07\u539f\u5148\u7684 \u500b\u7fa4\u984d\u5916\u589e\u52a0 \u500b Ghost \u7fa4\uff0c\u6700\u5f8c\u518d\u5c07 \u7684\u8f38\u51fa\uff0c \u53ea\u63a1\u7528 \uff0c\u5c07\u4ee3\u8868\u566a\u97f3\u7684 Ghost \u7fa4\u6392\u9664\u6389\u3002\u6211\u5011\u6309\u7167\u539f\u59cb\u8ad6\u6587\u4e2d\u7684\u8a2d\u5b9a\uff0c 8\uff0c 2\u3002 2.3 \u640d\u5931\u51fd\u6578 (Loss Function) \u8fd1\u5e74\u4f86\uff0c\u57fa\u65bc AM-Softmax \u640d\u5931\u51fd\u6578\u8a13\u7df4\u7684\u8a9e\u8005\u9a57\u8b49\u7cfb\u7d71\uff0c\u6bd4\u8d77\u50b3\u7d71\u7684 softmax \u6548\u679c \u6709\u8457\u5f88\u5927\u7684\u63d0\u6607(Y. Liu, He & Liu, 2019)\uff0c\u56e0\u6b64\u6bd4\u8d77\u539f\u5148\u7684 softmax\uff0c\u6211\u5011\u66f4\u504f\u5411\u63a1\u7528 AM-softmax\uff0c\u8a72\u640d\u5931\u51fd\u6578\u5c07\u89d2\u5ea6\u9593\u9694\u7684\u6982\u5ff5\u5f15\u5165 softmax \u3002AM-softmax \u640d\u5931\u51fd\u6578\u516c\u5f0f \u5982\u4e0b\uff1a \u5716 1. \u878d\u5408\u7b56\u7565\u7684\u793a\u610f\u5716 KALDI toolkit (Povey et al., 2011)\u4ee5\u8ff4\u97ff\u7684\u65b9\u5f0f\u589e\u5f37\u6211\u5011\u7684\u8a13\u7df4\u8cc7\u6599\uff0c\u6700\u7d42\u63a1\u7528\u7d93\u589e\u5f37\u904e \u5f8c\u7684\u8cc7\u6599\u4f86\u8a13\u7df4\u6a21\u578b\uff0c\u8868 3 \u70ba\u6211\u5011\u8a13\u7df4\u904e\u7a0b\u7684\u8cc7\u6599\u6578\u91cf\u8207\u4f7f\u7528\u65b9\u5f0f\u3002 \u8868 3. \u8a13\u7df4\u904e\u7a0b\u7684\u8cc7\u6599\u6578\u91cf\u8207\u4f7f\u7528\u65b9\u5f0f \u500b\u6b98\u5dee\u5340\u584a\u7d44\u5408\u800c\u6210\uff0c\u6c60\u5316\u5c64\u63a1\u7528\u7d71\u8a08\u6c60\u5316\u5c64\uff0c\u640d\u5931\u51fd\u6578\u662f AM-softmax\uff0c\u5d4c\u5165\u63d0\u53d6\u5668\u7684 \u5b78\u7fd2\u7684\u5f71\u97ff\uff0c\u7d50\u679c\u5c55\u793a\u65bc\u5716 2\uff0c\u6211\u5011\u53ef\u4ee5\u767c\u73fe\uff0c\u63a1\u7528 GhostVLAD \u6240\u64f7\u53d6\u51fa\u4f86\u7684\u5d4c\u5165\u7d93 t-SNE\uff0c\u5206\u7fa4\u8868\u73fe\u8f03\u7d71\u8a08\u6c60\u5316\u5c64\u4f73\uff0c\u5c24\u5176\u662f\u5728\u5716\u7247\u4e0a\u534a\u90e8\u9bae\u6709\u91cd\u758a\u8005\u3002\u800c\u5f9e minDCF \u7684\u8a55 \u4f30\u6a19\u6e96\u4f86\u770b\uff0c\u5206\u7fa4\u7d50\u679c\u8f03\u4f73\u7684 Single System 2 \u4e5f\u78ba\u5be6\u8868\u73fe\u8f03\u4f73\uff0c\u9019\u4e5f\u5c31\u8868\u793a\u5728 false alarm \u8207 miss \u76f8\u540c\u6b0a\u91cd\u7684\u689d\u4ef6\u4e0b\uff0cSingle System 2 \u7684\u9a57\u8b49\u6548\u679c\u6bd4 Single System 1 \u597d\u3002 \u524d\u7aef\u7684\u6a21\u578b\uff0c\u4e26\u4e14\u6bcf\u4e00\u500b\u6a21\u578b\u5206\u5225\u5c0d\u61c9\u5f8c\u7aef\u7684\u6a5f\u7387\u7dda\u6027\u5224\u5225\u5206\u6790\u548c\u6a5f\u7387\u7dda\u6027\u5224\u5225\u5206\u6790\u8abf \u9069\u8a55\u5206\u5668\uff0c\u56e0\u6b64\u6700\u7d42\u7684\u878d\u5408\u7d50\u679c\u7531 6 \u500b\u4e0d\u540c\u7684\u5b50\u7cfb\u7d71\u53c3\u8207\u878d\u5408\u5f8c\u7522\u751f\uff0c\u800c\u9019\u500b\u878d\u5408\u7cfb\u7d71 \u662f\u6211\u5011\u6240\u6709\u7cfb\u7d71\u4e2d\u6700\u4f73\u7684\uff0c\u65bc\u4efb\u52d9\u4e00\u4e0a EER 9.94%\uff0cminDCF 0.7703\uff0c\u5728 22 \u968a\u53c3\u8cfd\u968a\u4f0d \u4e2d\u6392\u540d 14 \u540d\uff1b\u65bc\u4efb\u52d9\u4e8c\u4e0a EER 10.31%\uff0cminDCF 0.8762\uff0c\u5728 19 \u968a\u53c3\u8cfd\u968a\u4f0d\u4e2d\u6392\u540d 11 \u540d\u3002 \u6700\u5f8c\u6e2c\u8a66\u8a9e\u901f\u7684\u5f71\u97ff\uff0c\u8868 7 \u986f\u793a\u5404\u8a9e\u901f\u5e73\u5747\u79d2\u901f\u8207\u5f71\u97ff\u7684\u7d50\u679c\uff0c\u7d50\u679c\u986f\u793a\u6162\u8a9e\u901f\u7684 \u6548\u679c\u662f\u6700\u597d\u7684\uff0c\u5176\u6b21\u662f\u5feb\u8a9e\u901f\uff0c\u6700\u5dee\u7684\u662f\u6b63\u5e38\u8a9e\u901f\uff0c\u6211\u5011\u4e5f\u767c\u73fe\uff0c\u7576\u8a3b\u518a\u8207\u6e2c\u8a66\u79d2\u901f\u5dee \u8ddd\u8d8a\u5927\uff0c\u6548\u679c\u4e5f\u8d8a\u5dee\u3002 \u8868 7. \u5404\u8a9e\u901f\u5e73\u5747\u79d2\u901f\u8207\u5f71\u97ff\u7684\u7d50\u679c identification using netfv and netvlad. In Proceedings of 11th International Symposium on Chinese Spoken Language Processing (ISCSLP 2018), 319-323. doi: 10.1109/ISCSLP.2018.8706687 Deng, J., Guo, J., Xue, N., & Zafeiriou, S. (2019). Arcface: Additive angular margin loss for deep face recognition. In Proceedings of the IEEE Conference on Computer Vision and \u8868 4. Single System 1 \u7684\u67b6\u69cb\u5982\u6642\u5ef6\u6b98\u5dee\u795e\u7d93\u7db2\u8def\u7ae0\u7bc0\u4e2d\u6240\u63cf\u8ff0\uff0c\u7531 5 \u5c64\u7684\u6642\u5ef6\u795e\u7d93\u7db2\u8def\u8207 6 Pattern Recognition, 4690-4699. doi: 10.1109/CVPR.2019.00482 4.2 \u4e3b\u7cfb\u7d71 (Primary System) Kenny, P. (2010). Bayesian speaker verification with heavy-tailed priors. In Proceedings of \u4f7f\u7528\u7d93 BOSARIS toolkit \u878d\u5408\u904e\u5f8c\u7684\u7cfb\u7d71\u4f5c\u70ba Primary System 1\uff0c\u9078\u7528 ID 2\u30013\u30014 \u4f5c\u70ba Odyssey 2010, 14.</td></tr><tr><td colspan=\"3\">\u800c\u5728 (Li et al., 2018) \u8ad6\u6587\u4e2d\uff0c\u4f5c\u8005\u63d0\u51fa\u4e86\u8207\u6211\u5011\u60f3\u6cd5\u76f8\u8fd1\u7684\u6642\u5ef6\u6b98\u5dee\u5340\u584a(Time Delay Residual Block, TDResBlock)\u67b6\u69cb\uff0c\u4f46\u4ed6\u5011\u9078\u64c7\u5c07\u6642\u5ef6\u795e\u7d93\u7db2\u8def\u6a21\u7d44\u52a0\u5165\u5230\u6b98\u5dee\u5340 \u584a\u4e2d\u3002\u9019\u6a23\u7684\u4e0d\u540c\u4e4b\u8655\u5728\u65bc\uff0c\u6211\u5011\u900f\u904e\u524d\u5e7e\u5c64\u7684\u64f4\u5f35\u5f97\u5230\u4e86\u56fa\u5b9a\u7684\u611f\u77e5\u7bc4\u570d\uff0c\u624d\u63a5\u8457\u4f7f \u7528\u6b98\u5dee\u5340\u584a\u4f86\u8403\u53d6\u8207\u6574\u5408\u7279\u5fb5\uff0c\u4f46\u4ed6\u5011\u7684\u4f5c\u6cd5\u5247\u662f\u8b93\u6bcf\u500b\u6b98\u5dee\u5340\u584a\u7684\u611f\u77e5\u7bc4\u570d\u7686\u4e0d\u76f8\u540c\uff0c \u56e0\u6b64\u6bcf\u500b\u6b98\u5dee\u5340\u584a\u6574\u5408\u8457\u4e0d\u540c\u611f\u77e5\u7bc4\u570d\u7684\u8cc7\u8a0a\uff0c\u540c\u6642\uff0c\u4ed6\u5011\u7684\u64f4\u5f35\u6578\u6700\u7d42\u53ef\u9054 11\uff0c\u4f46\u6211 \u5011\u8cc7\u6599\u7684\u97f3\u6846\u6578\u4e26\u4e0d\u8db3\u4ee5\u61c9\u4ed8\u9019\u9ebc\u5927\u7684\u64f4\u5f35\uff0c\u5f9e\u800c\u5c0e\u81f4\u6700\u7d42\u7d50\u679c\u53ef\u80fd\u6703\u53d7\u5230\u96f6\u586b\u5145\u7684\u5f71 \u97ff\u800c\u8b8a\u5dee\uff0c\u56e0\u6b64\u6211\u5011\u5728\u8a2d\u8a08\u6642\u5ef6\u6b98\u5dee\u795e\u7d93\u7db2\u8def\u67b6\u69cb\u6642\uff0c\u64f4\u5f35\u6578\u53ea\u5230 5\u3002 \uff2c \u2211 log \u2022 \u2022 \u2211 \u2022 , (2) \u4ee3\u8868\u7b2c \u500b\u8f38\u5165\u7684\u7279\u5fb5\u5411\u91cf\u8207\u6b0a\u91cd\u5411\u91cf\u7684\u89d2\u5ea6\u9918\u5f26\u503c\uff0c \u5247\u4ee3\u8868\u89d2\u5ea6\u908a\u754c\uff0c \u662f \u5c3a\u5ea6\u4fc2\u6578\uff0c\u7528\u65bc\u8abf\u6574\u89d2\u5ea6\u9918\u5f26\u503c\u7684\u5927\u5c0f\uff0c \u548c \u7686\u662f\u8d85\u53c3\u6578\uff0c\u9019\u500b\u640d\u5931\u51fd\u6578\u7684\u76ee\u6a19\u662f \u8981\u6700\u5927\u5316 \u4f86\u8b93\u7279\u5fb5\u5411\u91cf\u8207\u6b0a\u91cd\u5411\u91cf\u7684\u593e\u89d2\u6700\u5c0f\u3002\u6211\u5011\u53c3\u8003\u539f\u8ad6\u6587\u8a2d\u5b9a 30\uff0c 0.2\u3002 \u8a13\u7df4\u8cc7\u6599\u4f7f\u7528\u8a13\u7df4\u8cc7\u6599\u7ae0\u7bc0\u63d0\u5230\u7684 7 \u7a2e\u4e0d\u540c\u8cc7\u6599\u96c6\u3002\u800c\u56e0\u70ba\u4efb\u52d9\u4e00\u8207\u4efb\u52d9\u4e8c\u7684\u5dee\u5225\u5728\u65bc\uff0c 4.3 \u958b\u767c\u96c6\u5206\u6790 (Development Data Analysis) \u6211\u5011\u63a1\u7528\u5927\u91cf\u6587\u672c\u7121\u95dc\u8cc7\u6599\u4f86\u8a13\u7df4\u6211\u5011\u7684\u6a21\u578b\uff0c\u7528\u4ee5\u543b\u5408\u4efb\u52d9\u4e8c\u7684\u689d\u4ef6\uff0c\u4f46\u5982\u679c\u76f4\u63a5\u5957 \u7528\u5728\u4efb\u52d9\u4e00\u7684\u60c5\u5883\u4e0b\uff0c\u6240\u5f97\u5230\u7684\u6548\u679c\u6703\u5f88\u5dee\uff0c\u800c\u6700\u76f4\u63a5\u7684\u89e3\u6c7a\u65b9\u6cd5\u662f\u4f7f\u7528\u6587\u672c\u76f8\u95dc\u8cc7\u6599 \u91cd\u65b0\u8a13\u7df4\u5d4c\u5165\u63d0\u53d6\u5668\uff0c\u6216\u8005\u4ee5\u6587\u672c\u7121\u95dc\u8cc7\u6599\u8a13\u7df4\u597d\u7684\u6a21\u578b\u4f5c\u70ba\u9810\u8a13\u7df4(Pre-train)\u6a21\u578b\uff0c\u63a1 \u9077\u79fb\u5b78\u7fd2(Transfer Learning)\u7684\u65b9\u5f0f\uff0c\u4f7f\u7528\u6587\u672c\u76f8\u95dc\u8cc7\u6599\u8abf\u9069\u6a21\u578b\uff0c\u4f46\u9019\u5169\u7a2e\u65b9\u6cd5\u52e2\u5fc5\u5f97 \u82b1\u8cbb\u5927\u91cf\u7684\u6642\u9593\uff0c\u56e0\u6b64\u6211\u5011\u7684\u4f5c\u6cd5\u662f\u9078\u64c7\u5c07\u8a13\u7df4\u6a5f\u7387\u7dda\u6027\u5224\u5225\u5206\u6790\u8207\u6a5f\u7387\u7dda\u6027\u5224\u5225\u5206\u6790 \u8abf\u9069\u7684\u8cc7\u6599\u66f4\u63db\u6210\u6587\u672c\u76f8\u95dc\u8cc7\u6599\uff0c\u80fd\u5920\u7bc0\u7701\u6642\u9593\u4e14\u80fd\u9054\u5230\u4e00\u5b9a\u7684\u6548\u679c\u3002 \u6211\u5011\u7684\u8072\u5b78\u7279\u5fb5\uff0c\u63a1\u7528 KALDI 40 \u7dad\u7684 FBank \u914d 3 \u7dad\u7684\u97f3\u8abf\uff0c\u4e26\u4e14\u7d71\u4e00\u53d6\u6a23\u983b\u7387\u70ba 16kHz\uff0c\u97f3\u6846\u9577\u5ea6\u70ba 25-ms\uff0c\u97f3\u6846\u504f\u79fb\u70ba 10-ms\uff0c\u800c\u7279\u5fb5\u64f7\u53d6\u5b8c\u5f8c\uff0c\u4f7f\u7528\u57fa\u65bc\u80fd\u91cf\u7684\u8a9e\u97f3 \u6d3b\u6027\u5075\u6e2c(Energy-based Voice Activation Detection)\u4f86\u9664\u53bb\u6c92\u6709\u8072\u97f3\u7684\u8a9e\u97f3\u7247\u6bb5\uff0c\u8a31\u591a\u5be6\u9a57 \u8868\u660e\uff0c\u6709\u7121\u63a1\u7528\u57fa\u65bc\u80fd\u91cf\u7684\u8a9e\u97f3\u6d3b\u6027\u5075\u6e2c\u5c0d\u65bc\u7d50\u679c\u7684\u5f71\u97ff\u662f\u5f88\u5927\u7684\uff0c\u63a5\u8457\u91dd\u5c0d\u7279\u5fb5\u505a\u5012 \u4efb\u52d9\u4e00\u70ba\u6587\u672c\u76f8\u95dc\uff0c\u8a3b\u518a\u8207\u6e2c\u8a66\u97f3\u6a94\u7684\u5167\u5bb9\u7686\u70ba\"\u4f60\u597d\uff0c\u7c73\u96c5\"\u3002\u4efb\u52d9\u4e8c\u5247\u70ba\u6587\u672c\u7121\u95dc\uff0c \u91dd\u5c0d FFSVC 2020 \u958b\u767c\u96c6\uff0c\u6211\u5011\u4ee5\u4efb\u52d9\u4e00\u7684\u6642\u5ef6\u6b98\u5dee\u795e\u7d93\u7db2\u8def\u6e2c\u8a66\u4e86\u7a7a\u9593\u3001\u96dc\u8a0a\u8207\u8a9e\u901f \u5167\u5bb9\u8207\u667a\u6167\u5bb6\u5c45\u8a2d\u5099\u6307\u4ee4\u8207\u65e5\u5e38\u7528\u8a9e\uff0c\u56e0\u6b64\u6a5f\u7387\u7dda\u6027\u5224\u5225\u5206\u6790\u8207\u6a5f\u7387\u7dda\u6027\u5224\u5225\u5206\u6790\u8abf\u9069 \u5c0d\u65bc\u7d50\u679c\u7684\u5f71\u97ff\uff0c\u70ba\u4e86\u516c\u5e73\u6027\uff0c\u6211\u5011\u76f4\u63a5\u63a1\u7528\u7af6\u8cfd\u65b9\u63d0\u4f9b\u7684\u958b\u767c\u96c6 trials \u5171 53,996 \u7b46\u9032 \u4f9d\u7167\u4e0d\u540c\u4efb\u52d9\u4f7f\u7528\u4e0d\u540c\u8cc7\u6599\uff0c\u5728\u4efb\u52d9\u4e00\u6211\u5011\u4f7f\u7528 FFSVC 2020 \u8a13\u7df4\u96c6 (\u53ea\u4f7f\u7528\u7de8\u865f 1-30) \u884c\u6e2c\u8a66\uff0c\u4e26\u4f9d\u64da\u6211\u5011\u7684\u6e2c\u8a66\u60c5\u5f62\uff0c\u5f9e 53,996 \u7b46\u6e2c\u8a66\u4e2d\u6311\u9078\u6307\u5b9a\u7684\u914d\u5c0d\uff0c\u56e0\u6b64\u6bcf\u500b\u6e2c\u8a66\u60c5 \u52a0\u4e0a SLR-85\uff0c\u5176\u9304\u97f3\u5167\u5bb9\u7686\u662f\"\u4f60\u597d\uff0c\u7c73\u96c5\"\uff0c\u9019\u6a23\u505a\u7684\u76ee\u7684\u662f\u70ba\u4e86\u8207\u6587\u672c\u76f8\u95dc\u7684\u4efb \u5f62\u7684 trials \u6578\u91cf\u6703\u6709\u4e9b\u5fae\u7684\u5dee\u7570\u3002 \u52d9\u4e00\u6e2c\u8a66\u60c5\u5883\u76f8\u540c\uff0c\u800c\u6587\u672c\u7121\u95dc\u7684\u4efb\u52d9\u4e8c\u5247\u4f7f\u7528\u5168\u90e8\u7684 FFSVC 2020 \u8a13\u7df4\u96c6\u3002 \u9996\u5148\uff0c\u6211\u5011\u5be6\u9a57\u8a3b\u518a\u8207\u6e2c\u8a66\u5728\u4e0d\u540c\u7a7a\u9593\u8ddd\u96e2\u4e0b\u5c0d\u65bc\u7d50\u679c\u7684\u5f71\u97ff\uff0c\u5982\u8868 5 \u6240\u793a\uff0c0.25m\u3001 \u983b\u8b5c\u5e73\u5747\u503c\u8207\u8b8a\u7570\u6578\u6b63\u898f\u5316(Cepstral Mean And Variance Normalization, CMVN)\uff0c\u964d\u4f4e\u96e2 1m\u3001-1.5m\u30013m \u53ca 5m \u5206\u5225\u8868\u793a\u9304\u97f3\u88dd\u7f6e\u4e0d\u540c\u7684\u6536\u97f3\u8ddd\u96e2\uff0c0.25m \u70ba\u9304\u97f3\u88dd\u7f6e\u9762\u5c0d\u8aaa\u8a71 \u7fa4\u7279\u5fb5\u7684\u5f71\u97ff\uff0c\u4f7f\u6a21\u578b\u7684\u8a13\u7df4\u6548\u80fd\u63d0\u6607\u3002 \u4eba 0.25 \u516c\u5c3a\u9060\uff0c\u4ee5\u6b64\u985e\u63a8\uff0c\u800c\u8ca0\u865f\u5247\u8868\u793a\u6536\u97f3\u8ddd\u96e2\u76f8\u53cd\uff0c\u5373\u662f\u80cc\u5c0d\u8aaa\u8a71\u4eba\u4f86\u6536\u97f3\u3002\u6211\u5011</td></tr><tr><td colspan=\"3\">\u597d\uff1b\u640d\u5931\u51fd\u6578\u5247\u662f\u5f9e\u4eba\u81c9\u8fa8\u8b58\u9818\u57df\u501f\u9452\u800c\u4f86\uff0c\u5617\u8a66\u4e86\u8a31\u591a softmax \u4e0d\u540c\u7684\u8b8a\u9ad4\uff1aL-softmax (W. Liu, Wen, Yu & Yang, 2016)\u3001A-softmax (W. Liu et al., 2017)\u3001AM-softmax (Wang, Cheng, Liu & Liu, 2018)\u3001AAM-softmax (Deng, Guo, Xue & Zafeiriou, 2019)\u3002\u800c\u5f8c\u7aef\u7684\u8a55 \u5206 \u5668 \uff0c \u9664 \u4e86 \u9918 \u5f26 \u76f8 \u4f3c \u5ea6 (Cosine Similarity) \uff0c \u6a5f \u7387 \u7dda \u6027 \u5224 \u5225 \u5206 \u6790 (Probabilistic Linear \u6211\u5011\u53c3\u7167(Snyder et al., 2019)\u5efa\u7acb\u4e86\u64f4\u5c55\u6642\u5ef6\u795e\u7d93\u7db2\u8def\u4f5c\u70ba\u6211\u5011\u7684\u57fa\u6e96(Baseline)\uff0c\u5176\u67b6\u69cb \u4f7f\u7528\u4e86\u5341\u5c64\u4f86\u63d0\u53d6\u97f3\u6846\u5c64\u7684\u7279\u5fb5\uff0c\u4e26\u4e14\u5728\u7b2c 3\u30015\u30017 \u5c64\u4e2d\u4f7f\u7528\u5230\u4e86\u64f4\u5f35(dilation)\u7684\u6982\u5ff5\uff0c \u9019\u4e5f\u662f\u6642\u5ef6\u795e\u7d93\u7db2\u8def\u7684\u7cbe\u9ad3\u6240\u5728\uff0c\u4ee5\u64f4\u5f35\u4f86\u64f4\u5927\u97f3\u6846\u5c64\u8cc7\u8a0a\u7684\u611f\u77e5\u7bc4\u570d\uff0c\u64f4\u5f35\u6578\u5206\u5225\u70ba 2\u3001 3\u30014\uff0c\u7576\u64f4\u5f35\u6578\u70ba 2 \u6642\uff0c\u6211\u5011\u6703\u53d6\u76f8\u9130\u7684 5 \u500b\u97f3\u6846\u9032\u884c\u904b\u7b97\uff0c\u64f4\u5f35\u6578\u70ba 3 \u6642\uff0c\u5247\u53d6 7 \u500b \u97f3\u6846\uff0c\u4ee5\u6b64\u985e\u63a8\uff0c\u900f\u904e\u5c64\u5c64\u5806\u758a\uff0c\u56e0\u6b64\u6700\u7d42\u53ef\u4ee5\u611f\u77e5 23 \u500b\u97f3\u6846\u7684\u8cc7\u8a0a\u3002\u63a5\u8457\u5c07\u97f3\u6846\u5c64\u8f38 \u51fa\u7d93\u904e\u7d71\u8a08\u6c60\u5316\u5c64\uff0c\u5c07\u97f3\u6846\u5c64\u8cc7\u8a0a\u6574\u5408\u6210\u97f3\u6bb5\u5c64\u8cc7\u8a0a\uff0c\u800c\u5f8c\u7531\u5169\u5c64\u5168\u9023\u63a5\u5c64\u7d44\u6210\u97f3\u6bb5\u5c64\uff0c \u6700\u5f8c\u8f38\u51fa\u7d93 softmax \u8a08\u7b97\u6a5f\u7387\u9032\u884c\u5206\u985e\uff1b\u5728\u63a8\u8ad6\u968e\u6bb5\uff0c\u6211\u5011\u5f9e\u97f3\u6bb5\u5c64\u7684\u7b2c\u4e00\u5c64\u53d6\u51fa 512 \u7dad\u4ee3\u8868\u8a9e\u8005\u7684\u5d4c\u5165\u3002\u6bcf\u4e00\u5c64\u7686\u7d93\u904e\u6279\u91cf\u6a19\u6e96\u5316(Batch Normalization)\u8207 Rectified Linear Unit (ReLU)\u6fc0\u6d3b\u51fd\u6578\u3002 \u64f4\u5c55\u6642\u5ef6\u795e\u7d93\u7db2\u8def\u8207\u6b98\u5dee\u795e\u7d93\u7db2\u8def\u7684\u5dee\u5225\u5728\u65bc\uff0c\u64f4\u5c55\u6642\u5ef6\u795e\u7d93\u7db2\u8def\u7684\u524d\u5e7e\u5c64\u662f\u7531\u64c1\u6709\u64f4 \u5f35\u7684\u5377\u7a4d\u5c64\u6240\u7d44\u5408\u800c\u6210\uff0c\u900f\u904e\u64f4\u5f35\u4f86\u7372\u53d6\u8f03\u5927\u7bc4\u570d\u7684\u97f3\u6846\u5c64\u8cc7\u8a0a\uff0c\u96a8\u5f8c\u7dca\u63a5\u8457\u6578\u5c64\u7121\u64f4 \u5f35\u4e14\u76f8\u540c\u5377\u7a4d\u6838\u5927\u5c0f\u7684\u5377\u7a4d\u5c64\u4f86\u63d0\u53d6\u4e26\u6574\u5408\u97f3\u6846\u5c64\u8cc7\u8a0a\uff0c\u800c\u6b98\u5dee\u795e\u7d93\u7db2\u8def\u7684\u5c64\u6578\u96d6\u7136\u8f03 \u64f4\u5c55\u6642\u5ef6\u795e\u7d93\u7db2\u8def\u6df1\uff0c\u4f46\u5176\u6700\u7d42\u611f\u77e5\u7684\u97f3\u6846\u5c64\u7bc4\u570d\u537b\u4e0d\u5982\u64f4\u5c55\u6642\u5ef6\u795e\u7d93\u7db2\u8def\u4f86\u5f97\u5ee3\u95ca\uff0c \u6240\u4ee5\u6211\u5011\u8a8d\u70ba\u5169\u8005\u6709\u4e92\u88dc\u4e4b\u8655\uff0c\u64f4\u5c55\u6642\u5ef6\u795e\u7d93\u7db2\u8def\u5f8c\u5e7e\u5c64\u7121\u64f4\u5f35\u7684\u90e8\u4efd\uff0c\u5982\u679c\u63a1\u7528\u6b98\u5dee \u795e\u7d93\u7db2\u8def\u7684\u6a5f\u5236\uff0c\u5c31\u80fd\u8b93\u5f8c\u7e8c\u7684\u5c64\u6578\u8d8a\u6df1\u8d8a\u5927\uff0c\u4f7f\u5f97\u7279\u5fb5\u8403\u53d6\u80fd\u529b\u66f4\u597d\uff0c\u56e0\u6b64\u6309\u7167\u8a72\u60f3 \u6cd5\uff0c\u8a2d\u8a08\u4e86\u4e00\u500b\u6df7\u5408\u7684\u67b6\u69cb\uff0c\u67b6\u69cb\u5982\u8868 2\uff0c\u524d\u4e94\u5c64\u63a1\u7528\u4e86\u539f\u5148\u64f4\u5c55\u6642\u5ef6\u795e\u7d93\u7db2\u8def\u7684\u64f4\u5f35 \u8a2d\u8a08\uff0c\u64f4\u5f35\u6578\u5206\u5225\u70ba 2\u30013\u30014\u30015\uff0c\u5f8c\u5e7e\u5c64\u5247\u4f7f\u7528\u4e86\u4e0d\u540c\u901a\u9053\u5927\u5c0f\u7684\u6b98\u5dee\u5340\u584a\u5404 3 \u500b\uff0c\u8a72 \u67b6\u69cb\u4fdd\u7559\u4e86\u64f4\u5c55\u6642\u5ef6\u795e\u7d93\u7db2\u8def\u7372\u53d6\u8f03\u5927\u7bc4\u570d\u97f3\u6846\u5c64\u8cc7\u8a0a\u7684\u80fd\u529b\uff0c\u540c\u6642\u4e5f\u64c1\u6709\u6b98\u5dee\u795e\u7d93\u7db2 2.2 \u6c60\u5316\u5c64 (Pooling) \u9664\u4e86\u539f\u5148\u900f\u904e\u8a08\u7b97\u5e73\u5747\u503c\u8ddf\u6a19\u6e96\u5dee\u7684\u7d71\u8a08\u6c60\u5316\u5c64\u4e4b\u5916\uff0c\u70ba\u4e86\u89e3\u6c7a\u9060\u5834\u566a\u97f3\u5c0d\u65bc\u5206\u985e\u7684\u5f71 \u97ff\uff0c\u6211\u5011\u5617\u8a66\u4f7f\u7528 GhostVLAD (Xie et al., 2019)\u65b9\u6cd5\uff0c\u8a72\u65b9\u6cd5\u662f\u7531 NetVLAD \u70ba\u57fa\u790e\u6539 \u9032\u800c\u4f86\u7684\uff0cNetVLAD \u662f\u4e00\u7a2e\u53ef\u8a13\u7df4\u7684\u5206\u7fa4\u6cd5\uff0c\u4e3b\u8981\u505a\u6cd5\u662f\u5c07\u6bcf\u500b\u97f3\u6846\u5c64\u7684\u7279\u5fb5\u5206\u914d\u5230 \u4e0d\u540c\u7684\u7fa4\uff0c\u63a5\u8457\u8a08\u7b97\u8a72\u7279\u5fb5\u5230\u7fa4\u4e2d\u5fc3\u7684\u6b98\u5dee\u4e26\u7de8\u78bc\u6210\u6700\u5f8c\u7684\u8f38\u51fa\uff0c\u7522\u751f \u5927\u5c0f\u7684\u77e9 \u9663 \u3002\u4ee5\u4e0b\u70ba NetVLAD \u8a08\u7b97\u516c\u5f0f\uff1a , \u2211 \u2211 (1) \u5176\u4e2d \u8abf\u9069\u7684\u8a55\u5206\u5668\uff0c\u800c\u70ba\u4e86\u80fd\u5f97\u5230\u6700\u4f73\u7684\u7cfb\u7d71\u8868\u73fe\uff0c\u6211\u5011\u7d50\u5408\u4e86\u591a\u500b\u7cfb\u7d71\u6240\u8a08\u7b97\u51fa\u4f86\u7684\u5206\u6578\uff0c \u6a21\u578b\u4f7f\u7528 Nvidia GeForce GTX 1080 Ti GPU \u8a13\u7df4 6 \u500b epoch\u3002 GhostVLAD \u7684\u6642\u5ef6\u6b98\u5dee\u795e\u7d93\u7db2\u8def)\u548c\u878d\u5408\u6a21\u578b\uff0c\u4e26\u4e14\u50c5\u6709\u6642\u5ef6\u6b98\u5dee\u795e\u7d93\u7db2\u8def\u3001\u6642\u5ef6\u6b98\u5dee \u8868\u793a\u7fa4\u7e3d\u6578\uff0c\u662f\u4e00\u500b\u81ea\u8a02\u7684\u8d85\u53c3\u6578\uff0c \u8868\u793a\u6bcf\u4e00\u500b\u7fa4\u7684\u7dad\u5ea6\uff0c\u8207\u97f3\u6846\u5c64\u7684\u8f38\u51fa \u5f8c\u7aef\u8a55\u5206\u5668\u662f\u57fa\u65bc\u9ad8\u65af\u6a5f\u7387\u7dda\u6027\u5224\u5225\u5206\u6790\uff0c\u6211\u5011\u5148\u91dd\u5c0d\u64f7\u53d6\u51fa\u4f86\u7684\u8a9e\u8005\u5d4c\u5165\u4f5c\u5e73\u5747\u6b63\u898f \u5316\uff0c\u4f86\u964d\u4f4e\u8a9e\u8005\u5d4c\u5165\u6578\u503c\u7684\u8b8a\u7570\u6027\uff0c\u63a5\u8457\u7d93\u7531\u7dda\u6027\u5224\u5225\u5206\u6790 (Linear discriminant analysis, LDA) \u4f86\u5c07\u5d4c\u5165\u7684\u7dad\u5ea6\u964d\u7dad\u5230 250 \u7dad\uff0c\u4e26\u7528\u964d\u7dad\u904e\u5f8c\u7684\u5d4c\u5165\u8a13\u7df4\u6a5f\u7387\u7dda\u6027\u5224\u5225\u5206\u6790\uff0c\u4ee5 \u53ca\u7528\u65bc\u6a5f\u7387\u7dda\u6027\u5224\u5225\u5206\u6790\u8abf\u9069 (Adaptation) \u7684\u8abf\u6574\uff0c\u6700\u5f8c\u4ee5\u8a13\u7df4\u597d\u7684\u6a5f\u7387\u7dda\u6027\u5224\u5225\u5206\u6790 \u6a21\u578b\uff0c\u8a08\u7b97\u7d93\u8f49\u63db\u904e\u5f8c\u7684\u8a9e\u8005\u5d4c\u5165\u9593\u7684\u5206\u6578\u3002 2.4.2 \u5206\u6578\u878d\u5408 (Score Fusion) \u6bcf\u500b\u7cfb\u7d71\u90fd\u6709\u5176\u4e0d\u540c\u7684\u5d4c\u5165\u63d0\u53d6\u5668\u7684\u67b6\u69cb\uff0c\u4ee5\u53ca\u6a5f\u7387\u7dda\u6027\u5224\u5225\u5206\u6790\u548c\u6a5f\u7387\u7dda\u6027\u5224\u5225\u5206\u6790 \u7af6\u8cfd\u65b9\u63d0\u4f9b\u7684 FFSVC 2020 \u8a13\u7df4\u96c6\uff0c\u63a1\u7528\u9ea5\u514b\u98a8\u9663\u5217\u53ca\u624b\u6a5f\uff0c\u5728\u4e0d\u540c\u7a7a\u9593\u8ddd\u96e2\u3001\u4e0d\u540c\u96dc \u767c\u73fe\u5728 1m \u7684\u8ddd\u96e2\u4e0b\u6548\u679c\u6700\u597d\uff0c\u800c -1.5m \u7684\u8ddd\u96e2\u6548\u679c\u6700\u5dee\uff0c\u56e0\u70ba\u5176\u6536\u97f3\u65b9\u5411\u8207\u5176\u4ed6\u8ddd \u96e2\u76f8\u53cd\uff0c\u56e0\u6b64\u63a8\u6e2c\u6548\u679c\u5dee\u8207\u6536\u97f3\u65b9\u5411\u6709\u95dc\u3002 3.4 \u4f7f\u7528\u7af6\u8cfd\u65b9\u6240\u63d0\u4f9b\u7684 FFSVC 2020 \u958b\u767c\u96c6\u8207\u9a57\u8b49\u96c6\uff0c\u9304\u88fd\u65b9\u5f0f\u8207\u5167\u5bb9\u540c FFSVC 2020 \u8a13 \u7df4\u96c6\uff0c\u4f46\u5f7c\u6b64\u8a9e\u8005\u4e0d\u91cd\u758a\u3002\u4f9d\u7167\u7af6\u8cfd\u8981\u6c42\uff0c\u6e2c\u8a66\u65b9\u5f0f\u9700\u4ee5\u624b\u6a5f\u9304\u88fd\u7684\u97f3\u6a94\u4f5c\u70ba\u8a3b\u518a\uff0c\u9ea5 \u5728\u9019\u7bc7\u8ad6\u6587\u4e2d\uff0c\u6211\u5011\u53c3\u52a0\u4e86 FFSVC 2020\uff0c \u4e26\u57fa\u65bc\u6642\u5ef6\u795e\u7d93\u7db2\u8def\u8207\u5377\u7a4d\u795e\u7d93\u7db2\u8def\u5be6\u4f5c\u524d \u8868 5. \u4e0d\u540c\u7a7a\u9593\u8ddd\u96e2\u5f71\u97ff\u7684\u7d50\u679c [Table 5. Effects of different spatial distances] \u7aef\u7684\u7cfb\u7d71\uff0c\u540c\u6642\u4e5f\u5c07\u5169\u500b\u4e0d\u540c\u57fa\u790e\u7684\u7cfb\u7d71\u7d50\u5408\uff0c\u8a2d\u8a08\u51fa\u4e00\u500b\u65b0\u7684\u88ab\u7a31\u70ba\u6642\u5ef6\u6b98\u5dee\u795e\u7d93\u7db2 \u8a0a\u4ee5\u53ca\u4e0d\u540c\u8a9e\u901f\u4e0b\u9304\u88fd\u800c\u6210\uff0c\u9304\u97f3\u5167\u5bb9\u8207\u667a\u6167\u5bb6\u5c45\u7522\u54c1\u7684\u4f7f\u7528\u60c5\u5883\u6709\u95dc\uff0c\u800c\u9664\u4e86\u4f7f\u7528\u8a72 \u514b\u98a8\u9663\u5217\u9304\u88fd\u7684\u97f3\u6a94\u4f5c\u70ba\u6e2c\u8a66\u3002\u6240\u6709\u5be6\u9a57\u7686\u7d93\u7531\u958b\u767c\u96c6\u4f86\u6e2c\u8a66\u7d50\u679c\uff0c\u4e26\u4ee5\u5176\u7d50\u679c\u4f86\u9810\u4f30 \u8def\u7684\u7cfb\u7d71\uff0c\u5f8c\u7aef\u5be6\u4f5c\u6a5f\u7387\u7dda\u6027\u5224\u5225\u5206\u6790\u8207\u6a5f\u7387\u7dda\u6027\u5224\u5225\u5206\u6790\u8abf\u9069\u4e26\u7528\u4f86\u878d\u5408\u7cfb\u7d71\u3002\u5404\u7cfb \u8cc7\u6599\u4e4b\u5916\uff0c\u4f9d\u64da\u7af6\u8cfd\u8981\u6c42\uff0c\u6211\u5011\u4e5f\u5f9e OpenSLR (OpenSLR, 2020)\u4e0a\u6311\u9078\u958b\u653e\u8cc7\u6599\u96c6\uff0c\u5176 \u5728\u9a57\u8b49\u96c6\u4e0a\u7684\u8868\u73fe\uff0c\u56e0\u6b64\u958b\u767c\u96c6\u4e26\u7121\u53c3\u8207\u5230\u4efb\u4f55\u5f62\u5f0f\u7684\u8a13\u7df4\u4e2d\uff0c\u50c5\u7528\u4f86\u8a55\u4f30\u6a21\u578b\u8a13\u7df4\u7d50 \u7d71\u5206\u5225\u5728 FFSVC 2020 \u7684\u958b\u767c\u96c6\u8207\u9a57\u8b49\u96c6\u4e0a\u8a55\u4f30\uff0c\u5f9e\u7d50\u679c\u770b\u51fa\u6642\u5ef6\u6b98\u5dee\u795e\u7d93\u7db2\u8def\u52dd\u904e\u539f \u4e2d\u5305\u542b\u7af6\u8cfd\u8a55\u4f30\u8a08\u756b\u4e2d\u6709\u63d0\u53ca\u7684 SLR-85 (HI-MIA)\uff0c\u800c\u70ba\u4e86\u8981\u7b26\u5408\u672c\u7af6\u8cfd\u7684\u6e2c\u8a66\u60c5\u5883\uff0c \u679c\u7684\u597d\u58de\u3002 \u5148\u7684\u5169\u500b\u7cfb\u7d71\uff0c \u6b64\u5916\uff0c\u6211\u5011\u4e5f\u5be6\u9a57\u4e86 GhostVLAD \u4e26\u8207\u539f\u5148\u7684\u7d71\u8a08\u6c60\u5316\u5c64\u505a\u6bd4\u8f03\u3002\u6700 \u6211\u5011\u4e5f\u9078\u64c7\u4e86\u4e2d\u6587\u4e14\u9304\u97f3\u5167\u5bb9\u8207\u667a\u6167\u5bb6\u5c45\u6709\u95dc\u7684\u8cc7\u6599\u96c6\uff0c\u5206\u5225\u662f SLR-33 (AISHELL)\uff0c SLR-38 (FreeST Chinese)\uff0c\u4ee5\u53ca SLR-68 (MAGICDATA)\uff0c\u540c\u6642\u70ba\u4e86\u589e\u52a0\u8a13\u7df4\u8cc7\u6599\u7684\u8a9e\u8005 \u5011\u8a2d\u5b9a\u6279\u91cf\u5927\u5c0f(Batch Size)\u70ba 32\uff0c\u8d77\u59cb\u5b78\u7fd2\u7387\u70ba 0.001\uff0c\u4e26\u96a8\u8457\u8a13\u7df4\u8fed\u4ee3\u6578\u905e\u6e1b\u81f3 0.0001\uff0c \u64f4\u5c55\u6642\u5ef6\u795e\u7d93\u7db2\u8def\u3001\u6b98\u5dee\u795e\u7d93\u7db2\u8def\u3001\u6642\u5ef6\u6b98\u5dee\u795e\u7d93\u7db2\u8def\u3001\u6642\u5ef6\u6b98\u5dee\u795e\u7d93\u7db2\u8def-G(\u63a1\u7528 \u97f3\u7684\u97f3\u6a94\u4f86\u8a3b\u518a\u6642\uff0c\u6548\u679c\u6bd4\u8d77\u5176\u4ed6\u6709\u566a\u97f3\u7684\u4f86\u5f97\u597d\u3002 \u8cc7\u6599\u96c6\uff0c\u56e0\u6b64\u6700\u5f8c\u7e3d\u5171\u63a1\u7528\u4e86 7 \u500b\u4e0d\u540c\u7684\u8cc7\u6599\u96c6\u7528\u4ee5\u8a13\u7df4\u6a21\u578b\u3002\u81f3\u65bc\u8a13\u7df4\u7684\u8d85\u53c3\u6578\uff0c\u6211 \u6211\u5011\u7e3d\u5171\u5be6\u9a57\u4e86\u4e94\u7a2e\u4e0d\u540c\u7684\u6a21\u578b\u5728\u958b\u767c\u96c6\u4e0a\u7684\u8868\u73fe\uff0c\u9010\u4e00\u5c0d\u61c9\u8868 4 \u7684\u4e94\u500b\u7cfb\u7d71\uff0c\u5206\u5225\u662f \u8a3b\u518a\u8207\u6e2c\u8a66\u5728\u76f8\u540c\u7684\u566a\u97f3\u74b0\u5883\u4e0b\uff0c\u7d50\u679c\u90fd\u8868\u73fe\u7684\u8f03\u597d\uff0c\u5373\u5c0d\u89d2\u7dda\u7684\u90e8\u5206\uff0c\u800c\u7576\u63a1\u7528\u7121\u566a \u591a\u6a23\u6027\uff0c\u6211\u5011\u4e5f\u52a0\u5165\u4e86\u5728\u8a9e\u97f3\u4efb\u52d9\u4e0a\u7d93\u5e38\u4f7f\u7528 SLR-49 (VoxCeleb)\u8207 SLR-12 (LibriSpeech) \u7d42\uff0c\u6211\u5011\u7684\u6700\u4f73\u878d\u5408\u7cfb\u7d71\u80fd\u5728\u4efb\u52d9\u4e00\u4e0a\u9054\u5230 minDCF 0.7703\uff0cEER 9.94%\uff0c\u5728\u4efb\u52d9\u4e8c\u4e0a\u5247 4. \u7d50\u679c (Result) (a) TDResNet (b) TDResNet-G \u63a5\u8457\u6211\u5011\u5be6\u9a57\u96dc\u8a0a\u7684\u5f71\u97ff\uff0c\u8868 6 \u70ba\u5404\u96dc\u8a0a\u8868\u793a\u60c5\u5883\u8207\u5f71\u97ff\u7684\u7d50\u679c\uff0c\u5f9e\u7d50\u679c\u53ef\u4ee5\u770b\u51fa\uff0c \u662f minDCF 0.8762\uff0c EER 10.31%\u3002</td></tr></table>" | |
| } | |
| } | |
| } | |
| } |