| { | |
| "paper_id": "2019", | |
| "header": { | |
| "generated_with": "S2ORC 1.0.0", | |
| "date_generated": "2023-01-19T14:54:53.030371Z" | |
| }, | |
| "title": "Experiments on In-House Far-Field Speech Recognition", | |
| "authors": [ | |
| { | |
| "first": "Hsuan-Sheng", | |
| "middle": [], | |
| "last": "\u90b1\u70ab\u76db", | |
| "suffix": "", | |
| "affiliation": { | |
| "laboratory": "\u4e2d\u83ef\u96fb\u4fe1\u7814\u7a76\u9662 Telecommunication Laboratories", | |
| "institution": "Chunghwa Telecom Co., Ltd", | |
| "location": { | |
| "country": "Taiwan" | |
| } | |
| }, | |
| "email": "" | |
| }, | |
| { | |
| "first": "Jyh-Her", | |
| "middle": [], | |
| "last": "Chiu", | |
| "suffix": "", | |
| "affiliation": { | |
| "laboratory": "\u4e2d\u83ef\u96fb\u4fe1\u7814\u7a76\u9662 Telecommunication Laboratories", | |
| "institution": "Chunghwa Telecom Co., Ltd", | |
| "location": { | |
| "country": "Taiwan" | |
| } | |
| }, | |
| "email": "" | |
| }, | |
| { | |
| "first": "", | |
| "middle": [], | |
| "last": "Yang", | |
| "suffix": "", | |
| "affiliation": { | |
| "laboratory": "\u4e2d\u83ef\u96fb\u4fe1\u7814\u7a76\u9662 Telecommunication Laboratories", | |
| "institution": "Chunghwa Telecom Co., Ltd", | |
| "location": { | |
| "country": "Taiwan" | |
| } | |
| }, | |
| "email": "" | |
| } | |
| ], | |
| "year": "", | |
| "venue": null, | |
| "identifiers": {}, | |
| "abstract": "In recent years, speech recognition applications have introduced a variety of remote operating systems, such as car voice assistants, smart speakers, etc. In these systems, far-field speech recognition plays a key role. This paper mainly presents our experiments and results on farfield speech recognition on smart speaker devices. We use data augmentation methods, simulated far-field speech and neural network-based acoustic models to reduce the character error rate (CER). In the experimental part, this paper recorded the parallel test corpus of three distances using the smart speaker. The 50cm situation corpus can be reduced from 13.31% to 8.41%, the relative improvement is 36.8%, and the 80cm situation corpus is reduced from 19.20% to 10.89% with relative improvement of 43.2%.", | |
| "pdf_parse": { | |
| "paper_id": "2019", | |
| "_pdf_hash": "", | |
| "abstract": [ | |
| { | |
| "text": "In recent years, speech recognition applications have introduced a variety of remote operating systems, such as car voice assistants, smart speakers, etc. In these systems, far-field speech recognition plays a key role. This paper mainly presents our experiments and results on farfield speech recognition on smart speaker devices. We use data augmentation methods, simulated far-field speech and neural network-based acoustic models to reduce the character error rate (CER). In the experimental part, this paper recorded the parallel test corpus of three distances using the smart speaker. The 50cm situation corpus can be reduced from 13.31% to 8.41%, the relative improvement is 36.8%, and the 80cm situation corpus is reduced from 19.20% to 10.89% with relative improvement of 43.2%.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Abstract", | |
| "sec_num": null | |
| } | |
| ], | |
| "body_text": [], | |
| "back_matter": [], | |
| "bib_entries": { | |
| "BIBREF0": { | |
| "ref_id": "b0", | |
| "title": "Understanding how deep belief networks perform acoustic modelling", | |
| "authors": [ | |
| { | |
| "first": "A", | |
| "middle": [], | |
| "last": "Mohamed", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "G", | |
| "middle": [], | |
| "last": "Hinton", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "G", | |
| "middle": [], | |
| "last": "Penn", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 2012, | |
| "venue": "Proc. ICASSP", | |
| "volume": "", | |
| "issue": "", | |
| "pages": "4273--4276", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "A. Mohamed, G. Hinton, and G. Penn, \"Understanding how deep belief networks perform acoustic modelling\", in Proc. ICASSP, 2012, pp. 4273-4276.", | |
| "links": null | |
| }, | |
| "BIBREF1": { | |
| "ref_id": "b1", | |
| "title": "Feature Learning in Deep Neural Networks -Studies on Speech Recognition Tasks", | |
| "authors": [ | |
| { | |
| "first": "D", | |
| "middle": [], | |
| "last": "Yu", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "M", | |
| "middle": [ | |
| "L" | |
| ], | |
| "last": "Seltzer", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "J", | |
| "middle": [], | |
| "last": "Li", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "J.-T", | |
| "middle": [], | |
| "last": "Huang", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "F", | |
| "middle": [], | |
| "last": "Seide", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 2013, | |
| "venue": "Proceedings of International Conference on Learning Representations", | |
| "volume": "", | |
| "issue": "", | |
| "pages": "", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "D. Yu., M. L. Seltzer, J. Li, J.-T. Huang , F. Seide, \"Feature Learning in Deep Neural Networks -Studies on Speech Recognition Tasks\", in Proceedings of International Conference on Learning Representations, May 2013.", | |
| "links": null | |
| }, | |
| "BIBREF2": { | |
| "ref_id": "b2", | |
| "title": "A time delay neural network architecture for efficient 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": null, | |
| "venue": "16th Annual Conference of the International Speech Communication Association (INTERSPEECH). ISCA", | |
| "volume": "", | |
| "issue": "", | |
| "pages": "", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "V. Peddinti, D. Povey, S. Khudanpur. \"A time delay neural network architecture for efficient modeling of long temporal contexts\", In: 16th Annual Conference of the International Speech Communication Association (INTERSPEECH). ISCA, Dresden, 205.", | |
| "links": null | |
| }, | |
| "BIBREF3": { | |
| "ref_id": "b3", | |
| "title": "A summary of the REVERB challenge: state-of-the-art and remaining challenges in reverberant speech processing research", | |
| "authors": [ | |
| { | |
| "first": "K", | |
| "middle": [], | |
| "last": "Kinoshita", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "M", | |
| "middle": [], | |
| "last": "Delcroix", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "S", | |
| "middle": [], | |
| "last": "Gannot", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "E", | |
| "middle": [], | |
| "last": "Habets", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "R", | |
| "middle": [], | |
| "last": "Haeb-Umbach", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "W", | |
| "middle": [], | |
| "last": "Kellermann", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "V", | |
| "middle": [], | |
| "last": "Leutnant", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "R", | |
| "middle": [], | |
| "last": "Maas", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "T", | |
| "middle": [], | |
| "last": "Nakatani", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "B", | |
| "middle": [], | |
| "last": "Raj", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 2016, | |
| "venue": "EURASIP Journal on Advances in Signal Processing", | |
| "volume": "", | |
| "issue": "", | |
| "pages": "", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "K. Kinoshita, M. Delcroix, S. Gannot, E. Habets, R. Haeb-Umbach, W. Kellermann, V. Leutnant, R. Maas, T. Nakatani, B. Raj, et al. \"A summary of the REVERB challenge: state-of-the-art and remaining challenges in reverberant speech processing research\", EURASIP Journal on Advances in Signal Processing, 2016.", | |
| "links": null | |
| }, | |
| "BIBREF4": { | |
| "ref_id": "b4", | |
| "title": "Blind separation and dereverberation of speech mixtures by joint optimization", | |
| "authors": [ | |
| { | |
| "first": "T", | |
| "middle": [], | |
| "last": "Yoshioka", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "T", | |
| "middle": [], | |
| "last": "Nakatani", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "M", | |
| "middle": [], | |
| "last": "Miyoshi", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "H", | |
| "middle": [ | |
| "G" | |
| ], | |
| "last": "Okuno", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 2011, | |
| "venue": "IEEE Transactions on Audio, Speech, and Language Processing", | |
| "volume": "19", | |
| "issue": "1", | |
| "pages": "69--84", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "T. Yoshioka, T. Nakatani, M. Miyoshi, and H.G. Okuno. \"Blind separation and dereverberation of speech mixtures by joint optimization\". IEEE Transactions on Audio, Speech, and Language Processing, 19(1):69-84, 2011.", | |
| "links": null | |
| }, | |
| "BIBREF5": { | |
| "ref_id": "b5", | |
| "title": "A study on data augmentation of reverberant speech for robust speech recognition", | |
| "authors": [ | |
| { | |
| "first": "T", | |
| "middle": [], | |
| "last": "Ko", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "", | |
| "middle": [], | |
| "last": "Peddinti", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "", | |
| "middle": [], | |
| "last": "Povey", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "", | |
| "middle": [], | |
| "last": "Seltzer", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "", | |
| "middle": [], | |
| "last": "Khudanpur", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 2017, | |
| "venue": "Proc. ICASSP", | |
| "volume": "", | |
| "issue": "", | |
| "pages": "", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "T Ko, V Peddinti, D Povey, ML Seltzer, S Khudanpur, \"A study on data augmentation of reverberant speech for robust speech recognition\", in Proc. ICASSP, 2017.", | |
| "links": null | |
| }, | |
| "BIBREF6": { | |
| "ref_id": "b6", | |
| "title": "Jhu aspire system: Robust lvcsr with tdnns, ivector adaptation and rnn-lms", | |
| "authors": [ | |
| { | |
| "first": "V", | |
| "middle": [], | |
| "last": "Peddinti", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "G", | |
| "middle": [], | |
| "last": "Chen", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "V", | |
| "middle": [], | |
| "last": "Manohar", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "T", | |
| "middle": [], | |
| "last": "Ko", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "D", | |
| "middle": [], | |
| "last": "Povey", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "S", | |
| "middle": [], | |
| "last": "Khudanpur", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 2015, | |
| "venue": "Proceedings of 2015 IEEE Workshop on Automatic Speech Recognition and Understanding (ASRU)", | |
| "volume": "", | |
| "issue": "", | |
| "pages": "539--546", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "V. Peddinti, G. Chen, V. Manohar, T. Ko, D. Povey, and S. Khudanpur, \"Jhu aspire system: Robust lvcsr with tdnns, ivector adaptation and rnn-lms,\" in Proceedings of 2015 IEEE Workshop on Automatic Speech Recognition and Understanding (ASRU). IEEE, 2015, pp. 539-546.", | |
| "links": null | |
| }, | |
| "BIBREF8": { | |
| "ref_id": "b8", | |
| "title": "Robust speech recognition in unknown reverberant and noisy conditions", | |
| "authors": [ | |
| { | |
| "first": "H", | |
| "middle": [], | |
| "last": "Mallidi", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "", | |
| "middle": [], | |
| "last": "Hermansk", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 2015, | |
| "venue": "Proceedings of 2015 IEEE Workshop on Automatic Speech Recognition and Understanding (ASRU)", | |
| "volume": "", | |
| "issue": "", | |
| "pages": "533--538", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "Mallidi, H. Hermansk et al., \"Robust speech recognition in unknown reverberant and noisy conditions,\" in Proceedings of 2015 IEEE Workshop on Automatic Speech Recognition and Understanding (ASRU). IEEE, 2015, pp. 533-538.", | |
| "links": null | |
| }, | |
| "BIBREF9": { | |
| "ref_id": "b9", | |
| "title": "Training Data Augmentation and Data Selection", | |
| "authors": [ | |
| { | |
| "first": "", | |
| "middle": [], | |
| "last": "Karafi\u00e1t Martin", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "", | |
| "middle": [], | |
| "last": "Karel", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "", | |
| "middle": [], | |
| "last": "Kate\u0159ina", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "", | |
| "middle": [], | |
| "last": "Marc", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "", | |
| "middle": [], | |
| "last": "Watanabe Shinji", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "\u010cernock\u00fd", | |
| "middle": [], | |
| "last": "Burget Luk\u00e1\u0161", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "", | |
| "middle": [], | |
| "last": "Jan", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "", | |
| "middle": [], | |
| "last": "Igor", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 2017, | |
| "venue": "New Era for Robust Speech Recognition: Exploiting Deep Learning", | |
| "volume": "", | |
| "issue": "", | |
| "pages": "245--260", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "KARAFI\u00c1T Martin, VESEL\u00dd Karel, \u017dMOL\u00cdKOV\u00c1 Kate\u0159ina, DELCROIX Marc, WATANABE Shinji, BURGET Luk\u00e1\u0161, \u010cERNOCK\u00dd Jan and SZ\u0150KE Igor. \"Training Data Augmentation and Data Selection,\" in New Era for Robust Speech Recognition: Exploiting Deep Learning. Heidelberg: Springer International Publishing, 2017, pp. 245- 260.", | |
| "links": null | |
| }, | |
| "BIBREF10": { | |
| "ref_id": "b10", | |
| "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 | |
| }, | |
| "BIBREF11": { | |
| "ref_id": "b11", | |
| "title": "Speaker embedding extraction with phonetic information", | |
| "authors": [ | |
| { | |
| "first": "Y", | |
| "middle": [], | |
| "last": "Liu", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "L", | |
| "middle": [], | |
| "last": "He", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "J", | |
| "middle": [], | |
| "last": "Liu", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "M", | |
| "middle": [ | |
| "T" | |
| ], | |
| "last": "Johnson", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 2018, | |
| "venue": "Proc. Interspeech", | |
| "volume": "", | |
| "issue": "", | |
| "pages": "2247--2251", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "Y. Liu, L. He, J. Liu, and M. T. Johnson, \"Speaker embedding extraction with phonetic information,\" in Proc. Interspeech 2018, pp. 2247-2251.", | |
| "links": null | |
| }, | |
| "BIBREF12": { | |
| "ref_id": "b12", | |
| "title": "iVector-based acoustic data selection", | |
| "authors": [ | |
| { | |
| "first": "Olivier", | |
| "middle": [], | |
| "last": "Siohan", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Michiel", | |
| "middle": [], | |
| "last": "Bacchiani", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 2013, | |
| "venue": "Proceedings of the 14th Annual Conference of the International Speech Communication Association (Interspeech)", | |
| "volume": "", | |
| "issue": "", | |
| "pages": "", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "Olivier Siohan and Michiel Bacchiani, \"iVector-based acoustic data selection,\" in Proceedings of the 14th Annual Conference of the International Speech Communication Association (Interspeech), Lyon, France, 2013.", | |
| "links": null | |
| }, | |
| "BIBREF14": { | |
| "ref_id": "b14", | |
| "title": "The kaldi speech recognition toolkit", | |
| "authors": [ | |
| { | |
| "first": "Y", | |
| "middle": [], | |
| "last": "Motlicek", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "P", | |
| "middle": [], | |
| "last": "Qian", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "", | |
| "middle": [], | |
| "last": "Schwarz", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 2011, | |
| "venue": "IEEE 2011 workshop on automatic speech recognition and understanding", | |
| "volume": "", | |
| "issue": "", | |
| "pages": "", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "Motlicek, Y. Qian, P. Schwarz et al., \"The kaldi speech recognition toolkit,\" in IEEE 2011 workshop on automatic speech recognition and understanding, no. EPFLCONF-192584. IEEE Signal Processing Society, 2011.", | |
| "links": null | |
| }, | |
| "BIBREF15": { | |
| "ref_id": "b15", | |
| "title": "Semi-Orthogonal Low-Rank Matrix Factorization for Deep Neural Networks", | |
| "authors": [ | |
| { | |
| "first": "D", | |
| "middle": [], | |
| "last": "Povey", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "G", | |
| "middle": [], | |
| "last": "Cheng", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Y", | |
| "middle": [], | |
| "last": "Wang", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "K", | |
| "middle": [], | |
| "last": "Li", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "H", | |
| "middle": [], | |
| "last": "Xu", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "M", | |
| "middle": [], | |
| "last": "Yarmohamadi", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "S", | |
| "middle": [], | |
| "last": "Khudanpur", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 2018, | |
| "venue": "Proc. Interspeech", | |
| "volume": "", | |
| "issue": "", | |
| "pages": "", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "D. Povey, G. Cheng, Y. Wang, K. Li, H. Xu, M. Yarmohamadi, and S. Khudanpur, \"Semi- Orthogonal Low-Rank Matrix Factorization for Deep Neural Networks,\" in Proc. Interspeech, 2018.", | |
| "links": null | |
| }, | |
| "BIBREF16": { | |
| "ref_id": "b16", | |
| "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": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "D. Snyder, G. Chen, and D. Povey, \"Musan: A music, speech, and noise corpus,\" arXiv, 2015.", | |
| "links": null | |
| }, | |
| "BIBREF17": { | |
| "ref_id": "b17", | |
| "title": "Investigation of transfer learning for ASR using LF-MMI trained neural networks", | |
| "authors": [ | |
| { | |
| "first": "P", | |
| "middle": [], | |
| "last": "Ghahremani", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "V", | |
| "middle": [], | |
| "last": "Manohar", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "H", | |
| "middle": [], | |
| "last": "Hadian", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "D", | |
| "middle": [], | |
| "last": "Povey", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "S", | |
| "middle": [], | |
| "last": "Khudanpur", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 2017, | |
| "venue": "Proc. ASRU", | |
| "volume": "", | |
| "issue": "", | |
| "pages": "", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "P. Ghahremani, V. Manohar, H. Hadian, D. Povey, and S. Khudanpur, \"Investigation of transfer learning for ASR using LF-MMI trained neural networks,\" in Proc. ASRU, 2017", | |
| "links": null | |
| } | |
| }, | |
| "ref_entries": { | |
| "TABREF0": { | |
| "type_str": "table", | |
| "num": null, | |
| "content": "<table><tr><td>\u8a13\u7df4\u8a9e\u6599 \u97f3\u7bb11</td><td>\u8868\u4e00\u3001RIRND \u8cc7\u6599\u96c6 \u8868\u4e8c\u3001\u7a7a\u9593\u8108\u885d\u97ff\u61c9\u6a21\u64ec\u53c3\u6578\u8a2d\u5b9a \u8a9e\u8005\u6a5f\u7387\u5206\u4f48\u8f38\u51fa Softmax {0} GMM-UBM \u8a9e\u8005 GMM \u8a13\u7df4\u8a9e\u53e5n-RIR_small 1 7 =2 \u8868\u56db\u3001\u767c\u5c55\u8207\u6e2c\u8a66\u8a9e\u6599\u96c6 \u8868\u4e94\u3001\u7a7a\u9593\u6a21\u64ec\u6e2c\u8a66 CER \u7d50\u679c 1 \u8868\u4e03\u3001\u8a13\u7df4\u8cc7\u6599\u9078\u64c7 CER \u7d50\u679c 1 \u8868\u5341\u3001\u589e\u52a0\u8a9e\u6599\u6e2c\u8a66 CER \u7d50\u679c 1</td><td>512xN NN \u53c3\u6578</td></tr><tr><td colspan=\"3\">\u8a9e\u97f3\u8fa8\u8b58\u65b9\u9762\u5c31\u53d6\u5f97\u4e86\u91cd\u5927\u9032\u5c55[1,2,3]\uff0c\u5c0d\u65bc\u4efb\u4e00\u8a9e\u97f3\u8fa8\u8b58\u9818\u57df\uff0c\u7576\u63d0\u4f9b\u8db3\u5920\u4e14\u5177\u4ee3\u8868 \u6027\u7684\u8a13\u7df4\u8a9e\u6599\u6642\uff0cDNN \u5c31\u53ef\u4ee5\u5b78\u7fd2\u5176\u8072\u5b78\u672c\u8cea\u4e0a\u7684\u8b8a\u7570\u6027\uff0c\u5982\uff1a\u8a9e\u8005\u3001\u6027\u5225\u3001\u983b\u5bec\u3001 \u74b0\u5883\u7b49\u5dee\u7570\u3002\u7136\u800c\uff0c\u5728\u4e00\u500b\u771f\u5be6\u5ba4\u5167\u7684\u7a7a\u9593\u5167\uff0c\u53ef\u4ee5\u60f3\u50cf\u5230\u9060\u8ddd\u96e2\u8a9e\u97f3\u8fa8\u8b58\u80fd\u5920\u8b93\u6211\u5011 \u7684\u751f\u6d3b\u66f4\u4f73\u4fbf\u5229\uff0c\u4f46\u662f\u9060\u8ddd\u96e2\u8a9e\u97f3\u8fa8\u8b58\u4ecd\u7136\u662f\u4e00\u500b\u5177\u6709\u6311\u6230\u6027\u7684\u554f\u984c[4]\u3002 \u76ee\u524d\u5df2\u7d93\u63d0\u51fa\u8a31\u591a\u6280\u8853[4,5,6]\u4f86\u8655\u7406\u9060\u8ddd\u96e2\u8a9e\u97f3\u8fa8\u8b58\u554f\u984c\uff0c\u5176\u4e2d\u6700\u6709\u6548\u7684\u4f5c\u6cd5\u5c31\u662f\u8cc7\u6599 \u64f4\u5145(Data Augmentation)\uff0c\u5b83\u8b93 DNN \u6709\u6a5f\u6703\u53ef\u4ee5\u5b78\u7fd2\u5230\u771f\u5be6\u74b0\u5883\u4e2d\u53ef\u80fd\u906d\u9047\u5230\u7684\u60c5\u6cc1\uff0c \u53ea\u8981 DNN \u80fd\u5920\u5b78\u7fd2\u5230\u771f\u5be6\u74b0\u5883\u7684\u8072\u5b78\u7279\u5fb5\uff0c\u5728\u8a13\u7df4\u53ca\u6e2c\u8a66\u689d\u4ef6\u5339\u914d\u4e4b\u4e0b\u5c31\u6703\u6709\u4e0d\u932f\u7684 \u6548\u679c\u3002\u800c\u6536\u96c6\u5927\u91cf\u73fe\u5be6\u5404\u7a2e\u6a23\u5f0f\u7684\u74b0\u5883\u8a9e\u6599\u5f88\u8017\u8cbb\u5927\u91cf\u4eba\u529b\u53ca\u91d1\u9322\uff0c\u56e0\u6b64\uff0c\u5229\u7528\u6a21\u64ec\u65b9 \u6cd5\u7522\u751f\u8a13\u7df4\u8a9e\u6599\u662f\u4e00\u7a2e\u53ef\u884c\u7684\u9078\u64c7\uff0c\u85c9\u7531\u6a21\u64ec\u5404\u7a2e\u6a23\u5f0f\u74b0\u5883\u8a9e\u6599\u9032\u884c\u8cc7\u6599\u64f4\u5145\uff0c\u5c0d\u65bc\u5f37 \u5065\u8072\u5b78\u6a21\u578b\u4e0a\u80fd\u5f97\u5230\u975e\u5e38\u986f\u8457\u7684\u6548\u679c\uff0c\u53ef\u5f9e IARPA-ASPIRE \u9060\u8ddd\u96e2\u8fa8\u8b58\u7af6\u8cfd[7,8]\u4e2d\u770b \u5230\u4f7f\u7528\u8cc7\u6599\u64f4\u5145\u65b9\u6cd5\u5f97\u5230\u6700\u5927\u76f8\u5c0d 33%\u7684\u6539\u5584\u6548\u679c\u3002 \u672c\u6587\u4e3b\u8981\u5206\u4eab\u5229\u7528\u524d\u8ff0\u4f5c\u6cd5\u65bc\u5ba4\u5167\u667a\u6167\u97f3\u7bb1\u4e0a\u7684\u9060\u8ddd\u96e2\u8a9e\u97f3\u8fa8\u8b58\u5be6\u9a57\uff0c\u4f7f\u7528\u6a21\u64ec\u7684\u7a7a\u9593 \u8108\u885d\u97ff\u61c9(Room Impulse Response, RIR) \u6372\u7a4d\u65bc\u8a9e\u97f3\u4fe1\u865f\u4e2d\u4f86\u8868\u793a\u4e00\u500b\u7a7a\u9593\u7684\u6b98\u97ff (Reverberation)\uff0c\u85c9\u7531\u8abf\u6574\u4e0d\u540c\u53c3\u6578\u53ef\u4ee5\u6539\u5584\u667a\u6167\u97f3\u7bb1\u4e0a\u7684\u9060\u8ddd\u96e2\u8a9e\u97f3\u8fa8\u8b58\u554f\u984c\u3002 \u672c\u6587\u5c07\u5728\u63a5\u4e0b\u4f86\u7684\u7b2c\u4e8c\u7bc0\u6703\u4ecb\u7d39\u672c\u5be6\u9a57\u6240\u4f7f\u7528\u7684\u65b9\u6cd5\uff0c\u5728\u7b2c\u4e09\u7bc0\u5c07\u6703\u63cf\u8ff0\u672c\u6b21\u5be6\u9a57\u5728\u771f \u5be6\u97f3\u7bb1\u65bc\u4e0d\u540c\u8ddd\u96e2\u4e0b\u7684\u7d50\u679c\uff0c\u6700\u5f8c\u7d50\u8ad6\u6703\u653e\u5728\u7b2c\u56db\u7bc0\u4f5c\u8aaa\u660e\u3002 \u4e8c\u3001\u5be6\u9a57\u65b9\u6cd5 (\u4e00)\u6a21\u64ec\u7a7a\u9593\u8abf\u6574 \u5728 Ko[6]\u7684\u8ad6\u6587\u4e2d\uff0c\u4e3b\u8981\u900f\u904e\u93e1\u50cf\u6cd5\u7522\u751f\u4e86 3 \u7a2e RIR\uff0c\u5176\u7522\u751f\u65b9\u5f0f\u70ba\u5148\u4f9d\u7167\u4e3b\u8981\u8ddd\u96e2\u9650 \u5236\uff0c\u4f8b\u5982\u9577\u5bec 1~10 \u516c\u5c3a\u5167\uff0c\u9ad8 2~5 \u516c\u5c3a\u5167\uff0c\u4e82\u6578\u7522\u751f 200 \u7d44\u7a7a\u9593\u53c3\u6578\uff0c\u9019\u4e9b\u53c3\u6578\u5305\u542b \u4e86\u7a7a\u9593\u7684\u9577\u5bec\u9ad8\u3001\u63a5\u6536\u7aef\u4f4d\u7f6e\u8207\u7a7a\u9593\u7684\u5438\u6536\u4fc2\u6578\uff0c\u518d\u9650\u5236\u8072\u97f3\u4f86\u6e90\u7aef\u8207\u63a5\u6536\u7aef\u8ddd\u96e2\u70ba 0 \u8a73\u7d30\u8cc7\u6599\u5982\u8868\u4e00\u6240\u793a\u3002\u7136\u800c\uff0c\u6211\u5011\u4e3b\u8981\u7684\u61c9\u7528\u662f\u5728\u5bb6\u88e1\u7684\u5ba2\u5ef3\u5167\u4f7f\u7528\u667a\u6167\u97f3\u7bb1\uff0c\u4f7f\u7528 RIRND \u8cc7\u6599\u96c6\u672a\u5fc5\u5b8c\u5168\u7b26\u5408\u6211\u5011\u7684\u9700\u6c42\uff0c\u6240\u4ee5\u6211\u5011\u5617\u8a66\u8a2d\u8a08\u4ee5\u5ba2\u5ef3\u70ba\u7a7a\u9593\u5927\u5c0f\u7684\u53c3\u6578 \u4ee5\u7b26\u5408\u4f7f\u7528\u60c5\u5883\u3002\u6211\u5011\u5148\u4f9d\u7167\u5ba2\u5ef3\u7684\u53ef\u80fd\u7bc4\u570d\u4e82\u6578\u7522\u751f\u7a7a\u9593\u5927\u5c0f 200 \u7d44\uff0c\u518d\u4e82\u6578\u7522\u751f 100 \u7d44\u8072\u97f3\u4f86\u6e90\u7aef(\u5373\u8aaa\u8a71\u8005)\u53ca\u63a5\u6536\u7aef(\u5373\u97f3\u7bb1)\u4f4d\u7f6e\u8207\u6b98\u97ff\u6642\u9593\u7b49\u53c3\u6578\uff0c\u540c\u6a23\u7e3d\u5171 \u6709 20000 \u7d44\u7d50\u679c(\u4ee5\u4e0b\u7a31 RIR_exp) \u3002\u8aaa\u8a71\u8005\u8207\u97f3\u7bb1\u8ddd\u96e2\u7684\u7522\u751f\u65b9\u5f0f\u662f\u5148\u5c07\u7a7a\u9593\u5e73\u9762\u5283 \u5206\u6210 5 x 5 \u7684\u5340\u584a\uff0c\u7531\u5de6\u4e0a\u81f3\u53f3\u4e0b\u7de8\u865f\u70ba 0 \u81f3 24 \u865f\uff0c\u4e26\u5047\u8a2d\u97f3\u7bb1\u53ef\u80fd\u653e\u5728\u96fb\u8996\u65c1\u908a(\u7de8 \u865f 1\u30013) \u3001\u5ba2\u5ef3\u4e2d\u9593\u684c\u5b50(\u7de8\u865f 12)\u6216\u662f\u7a7a\u9593\u56db\u908a\u89d2\u843d(\u7de8\u865f 0\u30014\u300120\u300124) \uff0c\u5176\u4ed6\u7de8 \u865f\u70ba\u53ef\u80fd\u7684\u8aaa\u8a71\u8005\u4f4d\u7f6e\u3002\u81f3\u65bc\u5176\u4ed6\u53c3\u6578\u5982\u53cd\u5c04\u6b21\u6578\uff0c RIRND \u8a2d\u5b9a\u70ba 10\uff0c\u6211\u5011\u5247\u4e0d\u9650 \u6211\u5011\u521d\u6b65\u8cc7\u6599\u9078\u64c7\u5f8c\u50c5\u662f\u8b93\u4e0d\u540c RIR \u4e4b\u9593\u7684\u6df7\u5408\u6bd4\u4f8b\u4e0d\u540c\uff0c\u96d6\u7136\u9019\u6a23\u53ef\u4ee5\u78ba\u4fdd\u8207\u55ae\u4e00 \u5f59\u4e4b\u5916\uff0c\u6211\u5011\u4e5f\u89c0\u5bdf\u5230\u4e3b\u8981\u662f\u6211\u5011\u7684\u6a21\u578b\u8f03\u7121\u6cd5\u6b63\u78ba\u5730\u62d2\u7d55\u80cc\u666f\u4eba\u8072\u5e72\u64fe\uff0c\u5c0e\u81f4\u63d2\u5165\u932f \u6709\u6240\u6539\u5584\u3002 \u8207\u97f3\u7bb1\u7279\u6027\u6700\u76f8\u4f3c\u7684\u64f4\u5145\u8a9e\u53e5\u3002 1~2 \u516c\u5c3a\u53ca 3~4 \u516c\u5c3a\uff0c\u5982\u8868\u4e09\u6240\u793a\u3002 \u4e00\u53e5\u8a13\u7df4\u8a9e\u53e5\u4f7f\u7528\u4e0d\u540c\u7684\u64f4\u5145\u65b9\u5f0f\u90fd\u662f\u76f8\u540c\u8a9e\u8005\uff0c\u4e3b\u8981\u662f\u5e0c\u671b\u900f\u904e\u9019\u7a2e\u65b9\u5f0f\uff0c\u5617\u8a66\u9078\u51fa IBDEV \u8207 IB0515 \u7684\u7d50\u679c\uff0cGoogle \u8868\u73fe\u8f03\u597d\uff0c\u9664\u4e86\u5be6\u969b\u7528\u6236\u6bd4\u8f03\u5c11\u8aaa\u51fa\u8173\u672c\u4e2d\u7684\u7279\u6b8a\u8a5e \u4f8b\u4e5f\u5f9e 11.95%\u589e\u52a0\u5230 25.67%\uff0c\u4f46\u56e0\u70ba\u5be6\u9a57\u4e26\u672a\u52a0\u4e0a\u4fe1\u5fc3\u5ea6\u4f30\u6e2c\u7684\u6a5f\u5236\uff0c\u76f8\u4fe1\u52a0\u5165\u5f8c\u80fd \u4ee5 RIR_small \u8207 RIR_real \u64f4\u5145\u7684\u6548\u679c\u6700\u597d\u3002\u8cc7\u6599\u9078\u64c7\u7684\u6548\u679c\u4e0d\u5982\u9810\u671f\u7684\u539f\u56e0\u6216\u8a31\u662f\u56e0\u70ba \u66f4\u660e\u986f\uff0c\u4f8b\u5982 IBPH3-50cm \u53ef\u5f9e 13.31% \u964d\u81f3 8.41%\uff0c\u76f8\u5c0d\u6539\u5584 36.8%\uff0cIBPH3-80cm \u53ef \u751f\u51fa\u7684\u6a21\u64ec\u7d50\u679c\uff0c\u5176\u8aaa\u8a71\u8005\u8207\u97f3\u7bb1\u7684\u8ddd\u96e2\u5927\u90e8\u5206\u6703\u5206\u4f48\u5728 2~3 \u516c\u5c3a\u7684\u8ddd\u96e2\uff0c\u5176\u6b21\u5247\u662f \u6709\u5f88\u5927\u7684\u5dee\u7570\uff0c\u96d6\u7136\u4e3b\u8981\u90fd\u662f\u7528\u4f86\u5340\u5206\u8a9e\u8005\uff0c\u4f46\u662f\u4ee5\u6211\u5011\u7684\u8cc7\u6599\u9078\u64c7\u65b9\u5f0f\u4f86\u8aaa\uff0c\u5c0d\u65bc\u540c \u6b4c\u66f2\u540d\u7a31\u7b49\uff0c\u900f\u904e\u8a9e\u8a00\u6a21\u578b\u7684\u8abf\u6574\uff0c\u6e96\u78ba\u7387\u66f4\u9ad8\u3002\u4f46\u662f\u5728\u5be6\u969b\u74b0\u5883\u8207\u5be6\u969b\u7528\u6236\u4e0b\uff0c\u4f8b\u5982 Google \u7684\u932f\u8aa4\u5927\u90e8\u5206\u662f\u522a\u9664\u932f\u8aa4\uff0c\u800c\u5f9e\u8fa6\u516c\u5ba4\u74b0\u5883\u5230\u5be6\u969b\u74b0\u5883\uff0c\u6211\u5011\u7684\u6a21\u578b\u63d2\u5165\u932f\u8aa4\u6bd4 \u8207\u5be6\u969b\u74b0\u5883\u6700\u76f8\u4f3c\uff1b\u82e5\u4ee5 IBDEV \u767c\u5c55\u96c6\u8207 IB0515 \u6e2c\u8a66\u96c6\u7684\u55ae\u4e00 RIR \u7d50\u679c\u4f86\u770b\uff0c\u7684\u78ba 19.20%\uff1b\u800c\u5982\u679c\u6bd4\u8f03\u672a\u52a0\u5165\u64f4\u5145\u8cc7\u6599\u8207\u52a0\u5165\u5f8c\u7684\u6548\u679c\uff0c\u7686\u80fd\u6709\u6240\u63d0\u5347\uff0c\u4e14\u9060\u8ddd\u96e2\u7684\u6539\u5584 \u5011\u8a2d\u5b9a\u70ba 0.2~0.6\u3002\u8868\u4e8c\u662f\u6211\u5011\u7684\u53c3\u6578\u8a73\u7d30\u8a2d\u5b9a\uff0c\u6211\u5011\u4ea6\u9032\u4e00\u6b65\u7d71\u8a08\uff0c\u4ee5\u6211\u5011\u7684\u65b9\u6cd5\u7522 \u7406\u7684\u64f4\u5145\u7248\u672c\u3002i-vector \u8207 x-vector \u5169\u8005\u8655\u7406\u8a9e\u8005\u7684\u8072\u5b78\u7279\u5fb5\u65b9\u5f0f\u622a\u7136\u4e0d\u540c\uff0c\u53c3\u6578\u91cf\u4e5f IBD00 \u8207 IBQC5K \u7684\u7d50\u679c\uff0c\u4e3b\u8981\u662f\u56e0\u70ba\u8f03\u591a\u7279\u5225\u9818\u57df\u8a5e\u5f59\uff0c\u4f8b\u5982\u983b\u9053\u540d\u7a31\u3001\u7bc0\u76ee\u540d\u7a31\u6216 \u5217\u51fa CER \u932f\u8aa4\u5728\u8fa6\u516c\u5ba4\u8173\u672c\u9304\u97f3\u8207\u5be6\u969b\u74b0\u5883\u7684\u5206\u4f48\u60c5\u6cc1\uff0c\u5982\u8868\u5341\u4e09\u6240\u793a\uff0c\u53ef\u4ee5\u770b\u51fa \u61c9\u51fa\u8a9e\u97f3\u7684\u8a9e\u8005\u6216\u8072\u5b78\u7279\u6027\uff0c\u6839\u64da\u9019\u6a23\u7684\u5206\u4f48\u4f86\u770b\uff0c\u8868\u793a\u4ee5 RIR_exp \u64f4\u5145\u7684\u8a9e\u6599\u61c9\u8a72\u662f \u7576\u8ddd\u96e2\u8d8a\u9060\uff0c\u5176\u8fa8\u8b58\u6548\u679c\u660e\u986f\u964d\u4f4e\uff0c\u5982 IBPH3 \u672a\u52a0\u5165\u64f4\u5145\u8cc7\u6599\u7684 CER \u5f9e 5.69%\u589e\u52a0\u81f3 \u8aaa\uff0c\u5438\u6536\u4fc2\u6578\u8d8a\u9ad8\uff0c\u6b98\u97ff\u6642\u9593\u8d8a\u77ed\uff0c\u6211\u5011\u5247\u662f\u53c3\u8003 REVERB 2014 \u8a2d\u5b9a\u70ba 0.25~0.7\uff0c\u6211 \u5982\u5716\u4e8c\u6240\u793a\uff0c\u5047\u8a2d\u97f3\u7bb1\u70ba 7 \u500b\uff0c\u8a13\u7df4\u8a9e\u53e5 n-RIR_xxx \u70ba\u7b2c n \u53e5\u8a13\u7df4\u8a9e\u53e5\u7d93\u904e xxx \u65b9\u5f0f\u8655 \u8a66\u96c6\u90fd\u6709\u6539\u5584\uff0c\u8b93\u6a21\u578b\u66f4\u5f37\u5065\u3002\u6211\u5011\u7684\u6a21\u578b\u6548\u679c\u5728\u8173\u672c\u9304\u97f3\u7684\u8868\u73fe\u6bd4 Google \u597d\uff0c\u4f8b\u5982 \u8072\u9020\u6210\u63d2\u5165\u932f\u8aa4\uff0c\u4e26\u975e\u8a2d\u5099\u554f\u984c\uff0c\u9019\u4e5f\u662f\u6574\u9ad4 CER \u6bd4 Google \u7565\u5dee\u7684\u4e3b\u8981\u539f\u56e0\u3002\u6211\u5011\u4e5f \u4f8b\u70ba\u6700\u591a\uff0cx-vector \u5247\u662f\u4ee5 RIR_exp \u8207 RIR_small \u70ba\u6700\u591a\u3002\u5982\u679c i-vector \u8207 x-vector \u80fd\u53cd \u8a66\u7d50\u679c\u6703\u7528\u6b63\u898f\u5316\u5f8c\u7684\u932f\u8aa4\u7387\u8868\u793a\uff0c\u5373\u662f\u5c07\u932f\u8aa4\u7387\u9664\u4ee5\u57fa\u6e96\u932f\u8aa4\u7387\u3002\u9996\u5148\uff0c\u986f\u800c\u6613\u898b\u5730\uff0c \u5438\u6536\u4fc2\u6578\uff0c\u4e14\u5047\u8a2d\u7a7a\u9593\u5e73\u9762\u7686\u76f8\u540c\uff0c\u5176\u6b98\u97ff\u6642\u9593\u5247\u53ef\u900f\u904e Sabine \u516c\u5f0f\u63db\u7b97\u5f97\u51fa\uff0c\u7c21\u55ae\u5730 \u8a13\u7df4\u8a9e\u53e5\uff0c\u9078\u64c7\u8207\u767c\u5c55\u96c6\u8a9e\u8005\u6709\u6700\u591a\u4e14\u6700\u5927\u76f8\u4f3c\u5ea6\u7684\u64f4\u5145\u7248\u672c\u52a0\u5165\u3002\u8a13\u7df4\u8cc7\u6599\u9078\u64c7\u6d41\u7a0b \u96dc\u8a0a\u7684\u7d50\u679c\u70ba\u57fa\u6e96\uff0c\u7d2f\u7a4d\u52a0\u5165\u4e0d\u540c RIR \u7684\u8a13\u7df4\uff0c\u96d6\u7136\u8a13\u7df4\u6642\u9593\u589e\u52a0\u5f88\u591a\uff0c\u4f46\u5728\u6bcf\u4e00\u500b\u6e2c \u8ddf Google \u76f8\u6bd4\uff0c\u800c\u5ee0\u5546 C \u7684\u6548\u679c\u8f03\u5dee\uff0c\u89c0\u5bdf\u4e3b\u8981\u662f\u67d0\u4e9b\u4f7f\u7528\u8005\u7684\u8a2d\u5099\u74b0\u5883\u5e38\u51fa\u73fe\u80cc\u666f \u64c7\u5f8c\u7684 RIR \u5206\u4f48\uff0c\u5982\u8868\u4e5d\u6240\u793a\uff0c\u4e82\u6578\u662f\u5e73\u5747\u5206\u4f48\uff0ci-vector \u4ee5 RIR_exp \u8207 RIR_real \u7684\u6bd4 Error Rate, CER)\u70ba\u8a55\u4f30\u6a19\u6e96\uff0c\u5176\u7d50\u679c\u5982\u8868\u4e94\u3001\u8868\u516d\u6240\u793a\u3002\u7531\u65bc\u5546\u696d\u56e0\u7d20\uff0c\u5be6\u969b\u7528\u6236\u7684\u6e2c \u5236\uff1b\u6536\u97f3\u65b9\u5f0f\u6211\u5011\u8207 RIRND \u76f8\u540c\u5047\u8a2d\u70ba\u5168\u6307\u5411\u6027\uff1b\u6b98\u97ff\u6642\u9593\u53c3\u6578\u90e8\u5206\uff0cRIRND \u662f\u4f7f\u7528 \u53c3\u6578 \u8a2d\u5b9a\u503c RIR_small \u7a7a\u9593\u9577\u5bec 1~10 \u516c\u5c3a\uff0c\u9ad8\u5ea6 2~5 \u516c\u5c3a\uff0c\u5438\u6536\u4fc2\u6578 0.2~0.8\uff0c\u8072\u97f3\u4f86\u6e90\u8207 \u63a5\u6536\u4f4d\u7f6e\u8ddd\u96e2\u70ba 5 \u516c\u5c3a\u5167\uff0c\u4e82\u6578\u7522\u751f\u5171 20000 \u7d44 RIR_medium \u7a7a\u9593\u9577\u5bec 10~30 \u516c\u5c3a\uff0c\u9ad8\u5ea6 2~5 \u516c\u5c3a\uff0c\u5438\u6536\u4fc2\u6578 0.2~0.8\uff0c\u8072\u97f3\u4f86\u6e90 \u8207\u63a5\u6536\u4f4d\u7f6e\u8ddd\u96e2\u70ba 5 \u516c\u5c3a\u5167\uff0c\u4e82\u6578\u7522\u751f\u5171 20000 \u7d44 RIR_large \u7a7a\u9593\u9577\u5bec 30~50 \u516c\u5c3a\uff0c\u9ad8\u5ea6 2~5 \u516c\u5c3a\uff0c\u5438\u6536\u4fc2\u6578 0.2~0.8\uff0c\u8072\u97f3\u4f86\u6e90 \u8207\u63a5\u6536\u4f4d\u7f6e\u8ddd\u96e2\u70ba 5 \u516c\u5c3a\u5167\uff0c\u4e82\u6578\u7522\u751f\u5171 20000 \u7d44 RIR_real \u7531\u4e09\u5957\u8cc7\u6599\u96c6\u7d44\u6210\uff1a RWCP sound scene database: 1 \u500b\u5be6\u969b\u7a7a\u9593\uff0c\u9577\u5ea6\u70ba 6.66 \u516c\u5c3a\uff0c\u5bec\u5ea6 \u70ba 4.18 \u516c\u5c3a\uff0c\u8072\u97f3\u4f86\u6e90\u8207\u63a5\u6536\u8ddd\u96e2\u70ba 2 \u516c\u5c3a\uff0c\u97ff\u61c9\u6642\u9593\u70ba 0.3~1.3 \u79d2\uff0c\u5171 182 \u7d44 REVERB challenge database: 3 \u500b\u6a21\u64ec\u7a7a\u9593\uff0c\u97ff\u61c9\u6642\u9593\u5206\u5225\u70ba 0.25 \u79d2\u30010.5 \u79d2\u8207 0.7 \u79d2\uff0c2 \u7a2e\u8072\u97f3\u4f86\u6e90\u63a5\u6536\u4f4d\u7f6e\u8ddd\u96e2\u5206\u5225\u70ba 0.5 \u516c\u5c3a\u8207 2 \u516c\u5c3a\uff1b1 \u500b\u5be6\u969b\u7a7a\u9593\uff0c\u97ff\u61c9\u6642\u9593 0.7 \u79d2\uff0c2 \u7a2e\u8072\u97f3\u4f86\u6e90\u63a5\u6536\u4f4d\u7f6e\u8ddd\u96e2 \u5206\u5225\u70ba 1 \u516c\u5c3a\u8207 2.5 \u516c\u5c3a\uff0c\u5171 36 \u7d44 Aachen impulse response database: 4 \u500b\u5be6\u969b\u7a7a\u9593\uff0c\u7a7a\u9593\u5206\u5225\u70ba 3 x 1.8 x 0.5\u30015 x 6.4 x 2.9\u30018 x 5 x 3.1\u300110.8 x 10.9 x 3.15 \u516c\u5c3a\uff0c\u8072\u97f3\u4f86\u6e90\u63a5\u6536 \u8ddd\u96e2\u5206\u5225\u70ba 0.5, 1, 1.5 \u516c\u5c3a\u30011, 2, 3 \u516c\u5c3a\u30011.45, 1.7, 1.9, 2.25, 2.8 \u516c \u5c3a\u30014, 5.56, 7.1, 8.68, 10.2 \u516c\u5c3a\uff0c\u5171 107 \u7d44 (\u4e8c)\u8a13\u7df4\u8cc7\u6599\u9078\u64c7 \u6211\u5011\u5df2\u7d93\u900f\u904e\u6a21\u64ec RIR \u65b9\u5f0f\u4f86\u7522\u751f\u9060\u8ddd\u96e2\u7684\u8a0a\u865f\u6b98\u97ff\uff0c\u7136\u800c\u5c0d\u65bc\u6211\u5011\u76ee\u6a19\u7684\u5ba2\u5ef3\u74b0\u5883 \u97f3\u7bb1\u8fa8\u8b58\u4ecd\u4e0d\u4e00\u5b9a\u4e00\u81f4\uff0c\u5c24\u5176\u539f\u59cb\u8cc7\u6599\u5927\u591a\u4f86\u81ea\u65bc\u624b\u6a5f\u6216\u9ea5\u514b\u98a8\u3002\u5982\u6211\u5011\u6240\u77e5\uff0c\u5982\u679c\u9078 \u767c\u5c55\u96c6\u8207\u6240\u6709\u64f4\u5145\u8cc7\u6599\u62bd\u53d6 i-vector \u6216 x-vector \u7279\u5fb5\u5f8c\uff0c\u8a08\u7b97\u5176\u76f8\u4f3c\u5ea6\uff0c\u518d\u91dd\u5c0d\u6bcf\u4e00\u53e5 5~15dB \u53ca 13~20dB\uff0c\u540c\u6642\u6211\u5011\u4e5f\u5448\u73fe\u4e86 Google \u8fa8\u8b58\u7d50\u679c\u3002\u9996\u5148\uff0c\u6211\u5011\u4ee5\u7d93\u904e\u52a0\u6e1b\u901f\u8207 \u6240\u793a\uff0c\u5404\u7a2e\u5ee0\u5546\u7684\u97f3\u7bb1\u5728\u6a21\u578b\u52a0\u5165\u64f4\u5145\u8a9e\u6599\u5f8c\uff0c\u7686\u6709\u6539\u5584\uff1b\u5ee0\u5546 B \u8207 D \u7684\u6548\u679c\u5df2\u7d93\u80fd \u9078\u64c7\u5f8c\u53ea\u6709\u5728 IBPH3-50cm \u7684\u60c5\u6cc1\u4e0b\u6709\u6539\u5584\uff0c\u5176\u4ed6\u5247\u662f\u6c92\u6709\u6539\u5584\u3002\u6211\u5011\u5617\u8a66\u7d71\u8a08\u8cc7\u6599\u9078 \u6211\u5011\u9996\u5148\u6e2c\u8a66\u52a0\u5165\u4e0d\u540c RIR \u9032\u884c\u8a13\u7df4\u7684\u6548\u679c\uff0c\u5176\u4e2d\u5be6\u9a57\u7d50\u679c\u7686\u4ee5\u5b57\u932f\u8aa4\u7387(Character \u64da\u8a9e\u8a00\u6587\u5b57\u76f8\u95dc\u6027\uff0c\u6216\u662f\u4f9d\u64da\u8072\u5b78\u7279\u6027\u76f8\u95dc\u6027\u4f86\u6c7a\u5b9a[9]\u3002\u65bc\u672c\u8ad6\u6587\u4e2d\uff0c\u6211\u5011\u521d\u6b65\u5617\u8a66\u4f7f \u5f0f\uff0c\u5148\u4f7f\u7528\u539f\u59cb\u8cc7\u6599\u8a13\u7df4 i-vector \u53ca x-vector\uff0c\u518d\u6e96\u5099\u4e00\u90e8\u5206\u97f3\u7bb1\u8a9e\u6599\u7576\u4f5c\u767c\u5c55\u96c6\uff0c\u4e26\u5c0d \u96dc\u8a0a\u5247\u4f7f\u7528 MUSAN \u8cc7\u6599\u96c6[15]\uff0c\u5305\u542b\u4e00\u822c\u566a\u97f3\u3001\u97f3\u6a02\u53ca\u4eba\u8072\uff0cSNR \u5206\u5225\u70ba 0~15dB\u3001 \u7684\u60c5\u6cc1\uff0c\u522a\u9664\u7684\u932f\u8aa4\u8f03\u591a\u3002\u6211\u5011\u4e5f\u5448\u73fe\u4e0d\u540c\u5ee0\u5546\u97f3\u7bb1\u5728\u5be6\u969b\u7528\u6236\u74b0\u5883\u7684\u7d50\u679c\uff0c\u5982\u8868\u5341\u4e8c \u5247\u662f\u4e92\u6709\u9ad8\u4f4e\uff0c\u4f46\u5dee\u7570\u4e0d\u5927\u3002\u7136\u800c\uff0c\u5982\u679c\u8207\u8868\u4e94\u3001\u8868\u516d\u4f7f\u7528\u55ae\u4e00 RIR \u7684\u60c5\u6cc1\u76f8\u6bd4\uff0c\u8cc7\u6599 (\u4e00)\u6a21\u64ec\u7a7a\u9593\u8abf\u6574\u7d50\u679c \u5ea6\uff0c\u751a\u81f3\u53ef\u80fd\u5f97\u5230\u76f8\u540c\u6216\u66f4\u597d\u7684\u7d50\u679c\u3002\u8072\u97f3\u8cc7\u6599\u9078\u64c7\u7684\u65b9\u6cd5\u6709\u8a31\u591a\u7a2e\uff0c\u5927\u81f4\u4e0a\u53ef\u5206\u6210\u4f9d \u7528 i-vector \u6216\u662f x-vector\uff0c\u6548\u679c\u5927\u90e8\u5206\u90fd\u6703\u6bd4\u4e82\u6578\u9078\u64c7\u4f86\u5f97\u597d\uff1b\u800c\u4f7f\u7528 x-vector \u8207 i-vector \u6211\u5011\u4e5f\u52a0\u5165\u4e86\u52a0\u6e1b\u901f\u4ee5\u53ca\u52a0\u96dc\u8a0a\u5169\u7a2e\u8a9e\u6599\u64f4\u5145\u65b9\u6cd5\uff0c\u52a0\u901f\u8207\u6e1b\u901f\u5206\u5225\u70ba 1.1 \u500d\u8207 0.9 \u500d\uff0c \u8aa4\u8f03\u591a\u3002\u53e6\u5916\uff0c\u4e5f\u89c0\u5bdf\u5230 Google \u5728\u97f3\u7bb1\u9304\u97f3\u54c1\u8cea\u7a0d\u5dee\u7684\u60c5\u6cc1\u4e0b\uff0c\u5bb9\u6613\u51fa\u73fe\u6c92\u8fa8\u8b58\u7d50\u679c \u8ddd\u96e2\uff0c\u5982\u679c\u8ddd\u96e2\u7e2e\u5c0f\uff0c\u5247\u52a0\u5165\uff0c\u53cd\u4e4b\u5247\u4e0d\u52a0\u5165\u3002\u800c\u6211\u5011\u521d\u6b65\u7684\u4f5c\u6cd5\u5247\u662f\u4ee5\u5206\u985e\u8207\u6295\u7968\u65b9 \u64c7\u5f8c\u7684\u8a13\u7df4\u8cc7\u6599\u8207\u6e2c\u8a66\u74b0\u5883\u6216\u8a2d\u5099\u4f86\u6e90\u4e00\u81f4\uff0c\u8207\u4f7f\u7528\u5168\u90e8\u8cc7\u6599\u76f8\u6bd4\uff0c\u4e0d\u50c5\u80fd\u52a0\u5feb\u8a13\u7df4\u901f \u53c3\u6578 \u8a2d\u5b9a\u503c \u7a7a\u9593\u9577\u5ea6 (\u516c\u5c3a) [ 3.0, 3.5, 4.0, 4.5, 5.0, 5.5, 6.0, 6.5, 7.0 ] \u7a7a\u9593\u5bec\u5ea6 (\u516c\u5c3a) [ 3.0, 3.4, 3.8, 4.2, 4.6, 5.0 ] \u7a7a\u9593\u9ad8\u5ea6 (\u516c\u5c3a) [ 2.4, 2.6, 2.8, 3.0, 3.2, 3.4, 3.6, 3.8, 4.0, 4.2 ] \u8aaa\u8a71\u8005\u9ad8\u5ea6 (\u516c\u5c3a) [ 0.9, 1.1, 1.3, 1.5, 1.7 ] \u97f3\u7bb1\u9ad8\u5ea6 (\u516c\u5c3a) [ 0.4, 0.6, 0.8, 1.0, 1.2, 1.4 ] \u97f3\u901f (\u516c\u5c3a/\u79d2) 340 \u63a5\u6536\u7aef\u6307\u5411\u6027 \u5168\u6307\u5411 \u97ff\u61c9\u6642\u9593 (T60) [ 0.2, 0.3, 0.4, 0.5,0.6 ] \u53cd\u5c04\u6b21\u6578 \u7121\u9650\u5236 \u8aaa\u8a71\u8005\u4f4d\u7f6e\u7de8\u865f [ 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24 ] \u97f3\u7bb1\u4f4d\u7f6e\u7de8\u865f [ 0, 1, 3, 4, 12, 20, 24 ] \u8868\u4e09\u3001\u8aaa\u8a71\u8005\u8207\u97f3\u7bb1\u8ddd\u96e2\u5206\u4f48 \u8aaa\u8a71\u8005\u8207\u97f3\u7bb1\u8ddd\u96e2 \u7b46\u6578 \u5206\u4f48 (%) 0~1 \u516c\u5c3a 1871 9.355 1~2 \u516c\u5c3a 5869 29.345 2~3 \u516c\u5c3a 6774 33.870 3~4 \u516c\u5c3a 4347 21.735 4~5 \u516c\u5c3a 1139 5.695 \u4e09\u3001\u5be6\u9a57\u7d50\u679c \u63a5\u8457\u6211\u5011\u5448\u73fe\u8a13\u7df4\u8cc7\u6599\u9078\u64c7\u5f8c\u7684\u5be6\u9a57\u7d50\u679c\uff0c\u5982\u8868\u4e03\u3001\u8868\u516b\u6240\u793a\u3002\u53ef\u4ee5\u89c0\u5bdf\u5230\uff0c\u4e0d\u7ba1\u662f\u4f7f \u7d50\u679c\u4e2d\uff0c\u8a13\u7df4\u7684 epoch \u8a2d\u70ba 6\uff0c\u5176\u4ed6\u7684\u6a21\u578b\u8a2d\u5b9a\u8207\u65b9\u6cd5\u5be6\u9a57\u76f8\u540c\u3002\u9664\u4e86\u4f7f\u7528\u66f4\u591a\u8a9e\u6599\u5916\uff0c \u4e26\u900f\u904e\u4e00\u53e5\u6216\u591a\u53e5\u6279\u6b21\u7684\u65b9\u5f0f\uff0c\u5206\u5225\u8a08\u7b97\u6279\u6b21\u52a0\u5165\u524d\u8207\u52a0\u5165\u5f8c\u8207\u76ee\u6a19\u8cc7\u6599\u96c6\u5206\u4f48\u7684 KL \u5716\u4e00\u3001(a) i-vector (b) x-vector \u793a\u610f\u5716 TDNN5 {t} TDNN4 {t} TDNN3 {t-3, t, t +3} TDNN2 {t-2, t, t+2} TDNN1 [t-2, t+2] \u8a9e\u97f3\u7279\u5fb5 = 23\u7dad MFCC Pooling [0-T} TDNN6 {0} TDNN7 {0} (23x5)x512 (512x3)x512 (512x3)x512 512x512 512x1500 1500Tx3000 3000x512 512x512 (2048x60)x1 (2048x60)x1 (2048x60)x600 600x1 = + x \u8a9e\u8005\u8d85\u5411\u91cf M \u901a\u7528\u5e73\u5747\u8d85\u5411\u91cf m \u5168\u8b8a\u7570\u77e9\u9663 T i-vector w (a) (b) \u7dad\u5ea6 \u8a9e\u97f3\u7279\u5fb5 = 20\u7dad MFCC + \u4e00\u968e\u5dee\u91cf + \u4e8c\u968e\u5dee\u91cf = 60\u7dad \u8a13\u7df4 \u8a9e\u8005\u8abf\u9069 \u8a13\u7df4\u8a9e\u53e5n-RIR_medium 6 =1 \u7de8\u865f IBPH3 IBD00 IBQC5K IBDEV IB0515 K360 \u52a0\u5165\u4e0d\u540c RIR IBPH3-mic IBPH3-20cm IBPH3-50cm IBPH3-80cm K360 \u52a0\u5165\u4e0d\u540c RIR IBPH3-mic IBPH3-20cm IBPH3-50cm IBPH3-80cm STT2461 IBPH3-mic IBPH3-20cm IBPH3-50cm IBPH3-80cm \u4e32\u63a5\u8d85\u5411\u91cf (\u4e09)\u5be6\u9a57\u8a2d\u5b9a \u6211\u5011\u7684\u8a13\u7df4\u8a9e\u6599\u4f86\u81ea\u5546\u696d\u63a1\u8cfc\u3001\u5c08\u6848\u8a9e\u6599\u4ee5\u53ca\u7db2\u8def\u4e0a\u5404\u985e\u958b\u6e90\u8a9e\u6599\u7684\u96c6\u5408\uff0c\u5171\u6709 2461 \u5c0f\u6642(\u4ee5\u4e0b\u7a31 STT2461) \uff0c\u5176\u4e2d\u4f86\u6e90\u8a2d\u5099\u5305\u542b\u4e86\u9ea5\u514b\u98a8\u3001\u624b\u6a5f\u8207\u5ee3\u64ad\u7b49\u88dd\u7f6e\uff1b\u9304\u97f3\u6a23\u5f0f \u5305\u542b\u4e86\u8173\u672c\u9304\u97f3\u8207\u53e3\u8a9e\u5c0d\u8a71\uff1b\u60c5\u5883\u5305\u542b\u4e86\u65b0\u805e\u3001\u5404\u985e\u547d\u4ee4\u3001\u5404\u985e\u66f8\u7c4d\u8a9e\u53e5\u5ff5\u7a3f\u8207\u4e3b\u984c\u5f0f \u804a\u5929\u7b49\u65b9\u5f0f\uff1b\u8a9e\u8a00\u5247\u5305\u542b\u4e86\u4e2d\u6587\u3001\u82f1\u6587\u8207\u53f0\u8a9e\u4e09\u7a2e\u8a9e\u8a00\uff0c\u5176\u4e2d\u4ee5\u4e2d\u6587\u70ba\u4e3b\uff0c\u7d04\u4f54 76.8%\uff0c \u5176\u9918\u70ba 18.8%\u53ca 4.4%\u3002\u6211\u5011\u53e6\u5916\u5f9e\u8a13\u7df4\u8a9e\u6599\u4e2d\u53d6\u51fa\u7279\u5b9a\u8cc7\u6599\u96c6\uff0c\u5171 507 \u5c0f\u6642\u9032\u884c\u5be6\u9a57 (\u4ee5\u4e0b\u7a31 K360) \u3002\u6e2c\u8a66\u8a9e\u6599\u5247\u662f\u4e3b\u8981\u4f86\u81ea\u4e0d\u540c\u5ee0\u5546\u7684\u97f3\u7bb1\u8a2d\u5099\uff0c\u4e26\u5305\u542b\u5728\u8fa6\u516c\u5ba4\u74b0\u5883\u9304 \u88fd\u7684\u8173\u672c\u9304\u97f3\u4ee5\u53ca\u7528\u6236\u5be6\u969b\u74b0\u5883\u9ad4\u9a57\u6e2c\u8a66\u3002\u9304\u88fd\u8173\u672c\u7684\u61c9\u7528\u9818\u57df\u5305\u542b\u4e86\u5404\u7a2e\u8cc7\u8a0a\u67e5\u8a62\uff0c \u5982\u96fb\u8996\u983b\u9053\u3001\u7bc0\u76ee\u3001\u96fb\u5f71\u3001\u6b4c\u66f2\u3001\u5ee3\u64ad\u3001\u6709\u8072\u66f8\u3001\u80a1\u7968\u3001\u5e97\u5bb6\u3001\u5929\u6c23\u3001\u8def\u6cc1\u7b49\uff0c\u4e26\u642d\u914d \u97f3\u7bb12 \u97f3\u7bb17 \u2026 \u8a13\u7df4\u8a9e\u53e5n-RIR_large \u8a13\u7df4\u8a9e\u53e5n-RIR_real \u8a13\u7df4\u8a9e\u53e5n-RIR_exp 2 3 5 4 \u221a \u97f3\u7bb1\u8a9e\u8005\u8207\u64f4\u5145\u8a9e\u53e5\u8a08\u7b97 i-vector / x-vector \u76f8\u4f3c\u5ea6 \u97f3\u7bb1\u6295\u7968\u7d66 \u6700\u76f8\u4f3c\u64f4\u5145\u8a9e\u53e5 \u9078\u64c7\u8207\u6700\u591a\u97f3\u7bb1 \u76f8\u4f3c\u7684\u64f4\u5145\u8a9e\u53e5 \u985e\u578b \u6e2c\u8a66 \u6e2c\u8a66 \u6e2c\u8a66 \u767c\u5c55 \u6e2c\u8a66 NO_RIR 5.69 7.49 13.31 19.20 NO_RIR 5.69 7.49 13.31 19.20 Base + sp + noise 4.85 8.46 11.00 17.82 =0 =1 \u5c55\u96c6\u3002\u8a9e\u8a00\u6a21\u578b\u90e8\u5206\uff0c\u6211\u5011\u4f7f\u7528\u4ee5\u4e2d\u6587\u7db2\u9801\u8ddf\u65b0\u805e\u70ba\u4e3b\u7684 3-gram \u80cc\u666f\u6a21\u578b\uff0c\u518d\u5c07\u5404\u7a2e \u61c9\u7528\u60c5\u5883\u8a13\u7df4\u8a9e\u6599\u6574\u5408\u8a13\u7df4\u51fa\u4e00\u500b\u61c9\u7528\u5c0e\u5411\u7684 5-gram \u8a9e\u8a00\u6a21\u578b\uff0c\u6700\u5f8c\u900f\u904e\u5167\u63d2\u6cd5\u5c07\u5169 \u500b\u6a21\u578b\u5408\u4f75\uff0c\u521d\u6b65\u8a2d\u5b9a\u80cc\u666f\u6a21\u578b\u6bd4\u91cd\u70ba 0.3\u3002\u6b64\u5916\uff0c\u8a5e\u5178\u5927\u5c0f\u70ba 102 \u842c\u8a5e\uff0c\u8173\u672c\u6e2c\u8a66\u96c6 \u8a9e\u8a00\u8907\u96dc\u5ea6(Perplexity)\u7d04\u70ba 1087\uff0c\u5be6\u969b\u7528\u6236\u6e2c\u8a66\u96c6\u7d04\u70ba 1388\u3002\u7136\u800c\uff0c\u7531\u65bc\u8a5e\u5f59\u91cf\u5927\uff0c\u6e2c \u8a66\u53e5\u4e4b\u9593\u7684\u8907\u96dc\u5ea6\u5dee\u7570\u4e5f\u5f88\u5927\uff0c\u5f9e\u6578\u5341\u5230\u4e0a\u842c\u7686\u6709\u3002 2056 15546 5047 2960 7011 RIR_small 4.65 6.45 8.50 10.98 RIR_rand 5.25 6.15 8.67 11.35 +RIR_small + RIR_medium 4.60 6.79 8.85 13.45 \u53e5\u6578 \u5c0f\u6642\u6578 1.21 11.44 4.56 2.77 RIR_medium 4.78 6.24 8.83 11.79 RIR_ivector 4.90 6.12 8.41 11.32 +RIR_large + RIR_real 4.65 6.79 8.41 11.67 6.46 \u5e73\u5747\u79d2\u6578 2.13 2.65 3.25 3.38 RIR_large 4.99 6.29 9.64 12.27 RIR_xvector 4.99 6.10 7.95 11.28 +RIR_exp 4.62 6.82 8.11 11.51 3.32 \u6536\u97f3\u88dd\u7f6e \u9ea5\u514b\u98a8/\u97f3\u7bb1 \u97f3\u7bb1 \u97f3\u7bb1 \u97f3\u7bb1 \u97f3\u7bb1 \u74b0\u5883 \u8fa6\u516c\u5ba4 20/50/80 \u516c\u5206 \u8fa6\u516c\u5ba4 80 \u516c\u5206 \u5be6\u969b\u74b0\u5883 \u5be6\u969b\u74b0\u5883 RIR_real 5.02 5.99 8.78 GOOGLE 12.62 15.95 19.55 24.24 11.67 RIR_exp 4.92 5.87 8.41 \u8868\u516b\u3001\u8a13\u7df4\u8cc7\u6599\u9078\u64c7 CER \u7d50\u679c 2 10.89 \u8868\u5341\u4e00\u3001\u589e\u52a0\u8a9e\u6599\u6e2c\u8a66 CER \u7d50\u679c 2 \u5be6\u969b\u74b0\u5883 \u985e\u578b \u8173\u672c\u9304\u97f3 \u8173\u672c\u9304\u97f3 \u8173\u672c\u9304\u97f3 \u5be6\u969b\u7528\u6236 K360 \u52a0\u5165\u4e0d\u540c RIR IBD00 IBQC5K IBDEV IB0515 \u8868\u516d\u3001\u7a7a\u9593\u6a21\u64ec\u6e2c\u8a66 CER \u7d50\u679c 2 NO_RIR 10.53 11.34 1.000 1.000 STT2461 IBD00 IBQC5K IBDEV IB0515 \u5be6\u969b\u7528\u6236 \u97f3\u7bb1\u5ee0\u5546 A B B: 1512 \u53e5 C: 3535 \u53e5 B B: 1623 \u53e5 C: 3226 \u53e5 D: 2162 \u53e5 \u5f9e 19.20%\u964d\u81f3 10.89%\uff0c\u76f8\u5c0d\u6539\u5584 43.2%\u3002\u800c\u5982\u679c\u6211\u5011\u89c0\u5bdf\u81ea\u884c\u6a21\u64ec\u7684 RIR_exp\uff0c\u5728\u5927 \u90e8\u5206\u7684\u6e2c\u8a66\u4e0a\u90fd\u80fd\u6709\u6539\u5584\uff0c\u4f46\u4ee4\u4eba\u610f\u5916\u7684\u662f\u5be6\u969b\u74b0\u5883\u52a0\u4e0a\u5be6\u969b\u7528\u6236\u7684\u767c\u5c55\u8207\u6e2c\u8a66\u96c6\uff0c\u96d6 \u7136\u6709\u6539\u5584\uff0c\u4f46\u662f\u6539\u5584\u5e45\u5ea6\u4e26\u4e0d\u5927\u3002\u7136\u800c\uff0c\u4f7f\u7528\u5176\u4ed6\u985e\u578b\u7684 RIR \u4e5f\u662f\u985e\u4f3c\u60c5\u6cc1\uff0c\u89c0\u5bdf\u7d50\u679c \u5176\u4e3b\u8981\u539f\u56e0\u662f\u5be6\u969b\u74b0\u5883\u9664\u4e86\u6709\u9060\u8ddd\u96e2\u6536\u97f3\u554f\u984c\u5916\uff0c\u9084\u6709\u80cc\u666f\u96dc\u8a0a\u5e72\u64fe\u5982\u96fb\u8996\u8072\u7684\u554f\u984c\u4ee5 \u53ca\u8a9e\u8a00\u6a21\u578b\u6db5\u84cb\u7387\u4e0d\u5920\uff0c\u5982\u65b0\u6b4c\u66f2\u3001\u65b0\u96fb\u5f71\u7b49 OOV(Out-of-Vocabulary)\u7684\u5f71\u97ff\u3002 K360 \u52a0\u5165\u4e0d\u540c RIR IBD00 IBQC5K IBDEV IB0515 NO_RIR 10.53 11.34 1.000 1.000 RIR_small 8.57 8.44 0.863 0.982 RIR_medium 8.53 8.43 0.903 0.974 RIR_large 8.91 8.89 0.909 0.976 RIR_real 8.65 9.05 0.919 0.961 RIR_exp 8.52 8.49 0.928 0.990 RIR \u7684\u8a9e\u6599\u6578\u91cf\u5927\u5c0f\u4e00\u81f4\uff0c\u4f46\u4e26\u6c92\u6709\u4f7f\u7528\u76f8\u95dc\u6027\u5206\u6578\u4f86\u9078\u64c7\u51fa\u6700\u76f8\u4f3c\u6216\u662f\u7be9\u9078\u6389\u4e0d\u76f8\u4f3c \u7684\u8cc7\u6599\uff0c\u53cd\u800c\u4e0d\u5982\u4f7f\u7528\u55ae\u4e00 RIR \u7684\u6548\u679c\uff0c\u672a\u4f86\u53ef\u5617\u8a66\u8abf\u6574\u6210\u5c07\u6240\u6709\u64f4\u5145\u8cc7\u6599\u4f9d\u5206\u6578\u6392 \u5e8f\u5f8c\u518d\u9078\u64c7\u51fa\u4e0d\u540c\u6578\u91cf\u9032\u884c\u5be6\u9a57\u3002 (\u4e09)\u589e\u52a0\u8a13\u7df4\u8a9e\u6599\u7d50\u679c RIR_rand 8.54 8.83 0.926 0.967 RIR_ivector 8.55 8.44 0.907 0.983 RIR_xvector 8.60 8.46 0.902 0.978 \u8868\u4e5d\u3001\u8a13\u7df4\u8cc7\u6599\u9078\u64c7\u6bd4\u4f8b\u5206\u4f48 \u6df7\u5408 RIR small medium large real exp random 20.01% 19.93% 20.07% 19.92% 20.07% i-vector 20.16% 12.10% 19.66% 22.43% 25.65% x-vector 29.87% 10.62% 2.98% 11.51% 45.02% Base + sp + noise 9.04 7.76 0.833 0.884 +RIR_small + RIR_medium 8.02 7.87 0.813 0.871 +RIR_large + RIR_real 7.92 7.65 0.777 0.847 +RIR_exp 7.85 7.26 0.762 0.855 GOOGLE 12.16 8.28 0.667 0.782 \u8868\u5341\u4e8c\u3001\u4e0d\u540c\u5ee0\u5546\u97f3\u7bb1\u6e2c\u8a66 CER \u7d50\u679c IB0515 \u5ee0\u5546 B \u5ee0\u5546 C \u5ee0\u5546 D Base + sp + noise 0.882 0.907 0.835 +RIR_small + RIR_medium 0.824 0.916 0.804 +RIR_large + RIR_real 0.800 0.887 0.801 +RIR_exp 0.832 0.898 0.775 GOOGLE 0.939 0.692 0.877 \u8868\u5341\u4e09\u3001\u6e2c\u8a66 CER \u932f\u8aa4\u5206\u4f48 \u6e2c\u8a66-\u6a21\u578b \u63d2\u5165 \u522a\u9664 \u53d6\u4ee3 IBD00-STT2461+RIR_exp 11.95% 22.08% 65.97% \u591a\u7a2e\u524d\u7db4\u8a9e\u6216\u5f8c\u7db4\u8a9e\uff0c\u5982\u300c\u6211\u8981\u770b \u5716\u4e8c\u3001\u8a13\u7df4\u8a9e\u6599\u9078\u64c7\u6d41\u7a0b\u793a\u610f\u5716 IBD00-GOOGLE 3.21% 65.01% 31.78% (\u4e8c)\u8a13\u7df4\u8cc7\u6599\u9078\u64c7\u7d50\u679c IB0515-STT2461+RIR_exp 25.67% 25.15% 49.18% \u6700\u5f8c\u6211\u5011\u5448\u73fe\u4f7f\u7528\u5168\u90e8\u8a9e\u6599 STT2461 \u9032\u884c\u8a13\u7df4\u7684\u7d50\u679c\uff0c\u5982\u8868\u5341\u3001\u8868\u5341\u4e00\u6240\u793a\u3002\u5728\u9019\u500b IB0515-GOOGLE 1.70% 86.14% 12.16%</td></tr></table>", | |
| "html": null, | |
| "text": "The 2019 Conference on Computational Linguistics and Speech Processing ROCLING 2019, pp. 35-46 \u00a9The Association for Computational Linguistics and Chinese Language Processing \u4e00\u3001\u7dd2\u8ad6 \u8fd1\u5e74\u4f86\uff0c\u5728\u5f15\u9032\u6df1\u5ea6\u985e\u795e\u7d93\u7db2\u7d61(Deep Neural Network, DNN)\u65bc\u8072\u5b78\u6a21\u578b\u8a13\u7df4\u4e4b\u5f8c\uff0c\u81ea\u52d5 OO \u53f0\u300d \u3001 \u300c\u6211\u60f3\u807d OO \u7684 OO\u300d \u3001 \u300cOO \u80a1\u50f9\u591a\u5c11\u300d \u3001 \u300c\u6700 \u8fd1\u7684 OO \u5728\u54ea\u88e1\u300d \u3001 \u300cOO \u6709\u6c92\u6709\u4e0b\u96e8\u300d\u7b49\uff1b\u6216\u662f\u8a9e\u97f3\u547d\u4ee4\uff0c\u5982\u8a2d\u5b9a\u9b27\u9418\u3001\u884c\u4e8b\u66c6\u3001\u8a02\u8eca \u7968\u3001\u8a2d\u5099\u63a7\u5236\uff0c\u5982\u300c\u8a2d\u5b9a O \u9ede O \u5206\u7684\u9b27\u9418\u300d \u3001 \u300c\u6253\u958b\u5ba2\u5ef3\u51b7\u6c23\u300d\u7b49\u3002\u5176\u4ed6\u8cc7\u8a0a\u5982\u8868\u56db\u6240 \u793a\uff0c\u5176\u4e2d\u9664\u4e86 IBPH3 \u662f\u5305\u542b\u9ea5\u514b\u98a8\u8207\u4e0d\u540c\u8ddd\u96e2\u97f3\u7bb1\u5c0d\u9f4a\u7684\u8a9e\u6599\u5916\uff0c\u5176\u4ed6\u5247\u90fd\u662f\u4e0d\u540c\u5ee0 \u5546\u7684\u97f3\u7bb1\u9304\u97f3\u8a9e\u6599\u3002 \u6211\u5011\u4f7f\u7528 Kaldi \u5de5\u5177[13]\u9032\u884c\u76f8\u95dc\u5be6\u9a57\uff0c\u63a1\u7528\u7684\u8072\u5b78\u6a21\u578b\u67b6\u69cb\u70ba TDNN-F[14]\uff0c\u7db2\u8def\u5c64\u6578 \u70ba 11 \u5c64\uff0c\u6bcf\u4e00\u5c64\u7dad\u5ea6\u70ba 1280 \u7dad\uff0cSVD \u5206\u89e3\u7dad\u5ea6\u70ba 256 \u7dad\uff0c\u6a21\u578b\u67b6\u69cb\u4e3b\u8981\u53c3\u8003 kaldi/egs/swbd/s5c/local/chain/tuning/run_tdnn_7n.sh\uff0c\u8a9e\u97f3\u7279\u5fb5\u4f7f\u7528 40 \u7dad MFCC\u30013 \u7dad PITCH \u53ca 100 \u7dad i-vector \u9032\u884c\u8a13\u7df4\u3002\u9032\u884c\u5be6\u9a57\u6642\uff0c\u539f\u59cb\u6a21\u578b\u7684 epoch \u8a2d\u70ba 4\uff0c\u52a0\u5165\u64f4\u5145 \u8cc7\u6599\u6642\u5247\u8a2d\u70ba 2\u3002\u9032\u884c\u8a13\u7df4\u8cc7\u6599\u9078\u64c7\u5be6\u9a57\u4e5f\u662f\u4f7f\u7528 Kaldi \u62bd\u53d6 i-vector \u8207 x-vector\uff0c\u6211\u5011" | |
| } | |
| } | |
| } | |
| } |