{ "paper_id": "O06-1014", "header": { "generated_with": "S2ORC 1.0.0", "date_generated": "2023-01-19T08:07:49.906623Z" }, "title": "", "authors": [ { "first": "", "middle": [], "last": "\u5f35\u6587\u6770", "suffix": "", "affiliation": {}, "email": "" }, { "first": "", "middle": [], "last": "\uff0c\u9673\u9f0e\u5141", "suffix": "", "affiliation": {}, "email": "" }, { "first": "", "middle": [], "last": "\uff0c\u9673\u5b50\u548c", "suffix": "", "affiliation": {}, "email": "" }, { "first": "", "middle": [], "last": "\uff0c\u66fe\u5fd7\u4ec1", "suffix": "", "affiliation": {}, "email": "" }, { "first": "", "middle": [], "last": "\u3001\u5ed6\u5143\u752b", "suffix": "", "affiliation": {}, "email": "" }, { "first": "", "middle": [], "last": "\u570b\uf9f7\u53f0\uf963\u79d1\u6280\u5927\u5b78\u96fb\u5b50\u5de5\u7a0b\u5b78\u7cfb", "suffix": "", "affiliation": {}, "email": "" } ], "year": "", "venue": null, "identifiers": {}, "abstract": "", "pdf_parse": { "paper_id": "O06-1014", "_pdf_hash": "", "abstract": [], "body_text": [ { "text": "( , ) ( ) ( | ) ( ) ( | ) ( | ) K i j i j i i j k k i k P d w P d P w d P d P w z P z d = = = \u2211 (1) \u5176\u4e2d \u3001 \u3001 i d i w z \u6240\u4ee3\u8868\u7684\u5206\u5225\u662f document\u3001keyword\u3001latent", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "", "sec_num": null } ], "back_matter": [], "bib_entries": { "BIBREF0": { "ref_id": "b0", "title": "State-of-the-art in speaker recognition", "authors": [ { "first": "M", "middle": [], "last": "Faundez-Zanuy", "suffix": "" }, { "first": "E", "middle": [], "last": "Monte-Moreno", "suffix": "" } ], "year": 2005, "venue": "IEEE Aerospace and Electronic Systems Magazine", "volume": "20", "issue": "", "pages": "7--12", "other_ids": {}, "num": null, "urls": [], "raw_text": "M. Faundez-Zanuy, E. Monte-Moreno, \"State-of-the-art in speaker recognition\", IEEE Aerospace and Electronic Systems Magazine, Vol. 20, Issue 5, pp. 7-12, 2005.", "links": null }, "BIBREF1": { "ref_id": "b1", "title": "NIST -Speaker Recognition Evaluations", "authors": [], "year": null, "venue": "", "volume": "", "issue": "", "pages": "", "other_ids": {}, "num": null, "urls": [], "raw_text": "\"NIST -Speaker Recognition Evaluations\", http://www.nist.gov/speech/tests/spk/.", "links": null }, "BIBREF2": { "ref_id": "b2", "title": "The CSLP Speaker Recognition Evaluation (SRE)", "authors": [], "year": 2006, "venue": "", "volume": "", "issue": "", "pages": "", "other_ids": {}, "num": null, "urls": [], "raw_text": "The CSLP Speaker Recognition Evaluation (SRE) 2006, http://www.iscslp2006.org/.", "links": null }, "BIBREF3": { "ref_id": "b3", "title": "Chinese Corpus Consortium (CCC)", "authors": [], "year": null, "venue": "", "volume": "", "issue": "", "pages": "", "other_ids": {}, "num": null, "urls": [], "raw_text": "Chinese Corpus Consortium (CCC), http://www.cccforum.org/.", "links": null }, "BIBREF4": { "ref_id": "b4", "title": "NIST 2001 Speaker Recognition Evaluation -Extended Data task", "authors": [], "year": null, "venue": "", "volume": "", "issue": "", "pages": "", "other_ids": {}, "num": null, "urls": [], "raw_text": "\"NIST 2001 Speaker Recognition Evaluation -Extended Data task\", http://www.nist.gov/speech/tests/spk/2001/extended-data/.", "links": null }, "BIBREF5": { "ref_id": "b5", "title": "The SuperSID project: exploiting high-level information for high-accuracy speaker recognition", "authors": [ { "first": "D", "middle": [ "A" ], "last": "Reynolds", "suffix": "" }, { "first": "W", "middle": [], "last": "Andrews", "suffix": "" }, { "first": "J", "middle": [], "last": "Campbell", "suffix": "" }, { "first": "J", "middle": [], "last": "Navratil", "suffix": "" }, { "first": "B", "middle": [], "last": "Peskin", "suffix": "" }, { "first": "A", "middle": [], "last": "Adami", "suffix": "" }, { "first": "Q", "middle": [], "last": "Jin", "suffix": "" }, { "first": "D", "middle": [], "last": "Klusacek", "suffix": "" }, { "first": "J", "middle": [], "last": "Abramson", "suffix": "" }, { "first": "R", "middle": [], "last": "Mihaescu", "suffix": "" }, { "first": "J", "middle": [], "last": "Godfrey", "suffix": "" }, { "first": "D", "middle": [], "last": "Jones", "suffix": "" }, { "first": "B", "middle": [], "last": "Xiang", "suffix": "" } ], "year": 2003, "venue": "Proc. ICASSP'03", "volume": "4", "issue": "", "pages": "784--787", "other_ids": {}, "num": null, "urls": [], "raw_text": "D.A. Reynolds, W. Andrews, J. Campbell, J. Navratil, B. Peskin, A. Adami, Q. Jin, D. Klusacek, J. Abramson, R. Mihaescu, J. Godfrey, D. Jones and B. Xiang, \"The SuperSID project: exploiting high-level information for high-accuracy speaker recognition,\" Proc. ICASSP'03, Vol. 4, pp. 784-787, 2003.", "links": null }, "BIBREF6": { "ref_id": "b6", "title": "Modeling prosodic dynamics for speaker recognition", "authors": [ { "first": "A", "middle": [ "G" ], "last": "Adami", "suffix": "" }, { "first": "R", "middle": [], "last": "Mihaescu", "suffix": "" }, { "first": "D", "middle": [ "A" ], "last": "Reynolds", "suffix": "" }, { "first": "J", "middle": [ "J" ], "last": "Godfrey", "suffix": "" } ], "year": 2003, "venue": "Proc. ICASSP'03", "volume": "4", "issue": "", "pages": "788--791", "other_ids": {}, "num": null, "urls": [], "raw_text": "A.G. Adami, R. Mihaescu, D.A. Reynolds and J.J. Godfrey, \"Modeling prosodic dynamics for speaker recognition,\" Proc. ICASSP'03, Vol. 4, pp. 788-791, 2003.", "links": null }, "BIBREF7": { "ref_id": "b7", "title": "Probabilistic Latent Prosody Analysis For Robust Speaker Verification", "authors": [ { "first": "Z", "middle": [ "H" ], "last": "Chen", "suffix": "" }, { "first": "Z", "middle": [ "R" ], "last": "Zeng", "suffix": "" }, { "first": "Y", "middle": [ "F" ], "last": "Liao", "suffix": "" }, { "first": "Y", "middle": [ "T" ], "last": "Juang", "suffix": "" } ], "year": 2006, "venue": "", "volume": "", "issue": "", "pages": "", "other_ids": {}, "num": null, "urls": [], "raw_text": "Z.H. Chen, Z.R. Zeng, Y.F. Liao, and Y.T. Juang, \"Probabilistic Latent Prosody Analysis For Robust Speaker Verification,\" ICASSP'06, 2006.", "links": null }, "BIBREF8": { "ref_id": "b8", "title": "MVA Processing of Speech Features", "authors": [ { "first": "C", "middle": [ "P" ], "last": "Chen", "suffix": "" }, { "first": "J", "middle": [], "last": "Bilmes", "suffix": "" } ], "year": null, "venue": "IEEE Trans. on Speech and Audio Processing", "volume": "", "issue": "", "pages": "", "other_ids": {}, "num": null, "urls": [], "raw_text": "C.P. Chen and J. Bilmes, \"MVA Processing of Speech Features\", to appear in IEEE Trans. on Speech and Audio Processing.", "links": null }, "BIBREF9": { "ref_id": "b9", "title": "Speaker Verification Using Adapted Gaussian Mixture Models", "authors": [ { "first": "D", "middle": [ "A" ], "last": "Reynolds", "suffix": "" }, { "first": "T", "middle": [], "last": "Quatieri", "suffix": "" }, { "first": "R", "middle": [], "last": "Dunn", "suffix": "" } ], "year": 2000, "venue": "Digital Signal Processing", "volume": "10", "issue": "", "pages": "19--41", "other_ids": {}, "num": null, "urls": [], "raw_text": "D.A. Reynolds, T. Quatieri and R. Dunn, \"Speaker Verification Using Adapted Gaussian Mixture Models,\" Digital Signal Processing, Vol. 10, pp. 19-41, January 2000.", "links": null }, "BIBREF10": { "ref_id": "b10", "title": "Speaker Adaptive Cohort Selection for Tnorm in Text-Independent Speaker Verification", "authors": [ { "first": "D", "middle": [], "last": "Sturim", "suffix": "" }, { "first": "D", "middle": [ "A" ], "last": "Reynolds", "suffix": "" } ], "year": 2005, "venue": "ICASSP'05", "volume": "", "issue": "", "pages": "", "other_ids": {}, "num": null, "urls": [], "raw_text": "D. Sturim and D.A. Reynolds, \"Speaker Adaptive Cohort Selection for Tnorm in Text-Independent Speaker Verification\", ICASSP'05,2005.", "links": null }, "BIBREF11": { "ref_id": "b11", "title": "LNKnet Pattern Classification Software", "authors": [], "year": null, "venue": "", "volume": "", "issue": "", "pages": "", "other_ids": {}, "num": null, "urls": [], "raw_text": "LNKnet Pattern Classification Software, http://www.ll.mit.edu/IST/lnknet/.", "links": null }, "BIBREF12": { "ref_id": "b12", "title": "Feature Warping for Robust Speaker Verification", "authors": [ { "first": "J", "middle": [], "last": "Pelecanos", "suffix": "" }, { "first": "S", "middle": [], "last": "Sridharan", "suffix": "" } ], "year": 2001, "venue": "Proc. ISCA Workshop on Speaker Recognition -2001: A Speaker Oddyssey", "volume": "", "issue": "", "pages": "", "other_ids": {}, "num": null, "urls": [], "raw_text": "J. Pelecanos and S. Sridharan, \"Feature Warping for Robust Speaker Verification,\" Proc. ISCA Workshop on Speaker Recognition -2001: A Speaker Oddyssey, June 2001.", "links": null }, "BIBREF13": { "ref_id": "b13", "title": "Histogram Equalization of Speech Representation for Robust Speech Recognition", "authors": [ { "first": "A", "middle": [], "last": "De La Torre", "suffix": "" }, { "first": "A", "middle": [ "M" ], "last": "Peinado", "suffix": "" }, { "first": "J", "middle": [ "C" ], "last": "Segura", "suffix": "" }, { "first": "J", "middle": [ "L" ], "last": "P-C", "suffix": "" }, { "first": "M", "middle": [ "C" ], "last": "Benitez", "suffix": "" }, { "first": "A", "middle": [ "J" ], "last": "Rubio", "suffix": "" } ], "year": 2005, "venue": "IEEE Trans. on Speech and Audio Processing", "volume": "13", "issue": "", "pages": "355--366", "other_ids": {}, "num": null, "urls": [], "raw_text": "A. de la Torre, A. M. Peinado, J. C. Segura, J. L. P-C, M. C. Benitez, A. J. Rubio, \"Histogram Equalization of Speech Representation for Robust Speech Recognition,\" IEEE Trans. on Speech and Audio Processing. Vol. 13, pp. 355-366, May 2005.", "links": null }, "BIBREF14": { "ref_id": "b14", "title": "Probabilistic Latent Semantic Analysis", "authors": [ { "first": "T", "middle": [], "last": "Hofmann", "suffix": "" } ], "year": 1999, "venue": "Proc. Uncertainty in Artificial Intelligence", "volume": "", "issue": "", "pages": "", "other_ids": {}, "num": null, "urls": [], "raw_text": "T. Hofmann, \"Probabilistic Latent Semantic Analysis,\" Proc. Uncertainty in Artificial Intelligence 1999, 1999.", "links": null }, "BIBREF15": { "ref_id": "b15", "title": "Unsupervised Learning by Probabilistic Latent Semantic Analysis", "authors": [ { "first": "T", "middle": [], "last": "Hofmann", "suffix": "" } ], "year": 2001, "venue": "Machine Learning", "volume": "", "issue": "", "pages": "177--196", "other_ids": {}, "num": null, "urls": [], "raw_text": "T. Hofmann, \"Unsupervised Learning by Probabilistic Latent Semantic Analysis,\" Machine Learning, pp. 177-196, 2001.", "links": null }, "BIBREF16": { "ref_id": "b16", "title": "Snack sound toolkit", "authors": [ { "first": "K", "middle": [], "last": "Sj\u00f6lander", "suffix": "" } ], "year": null, "venue": "", "volume": "", "issue": "", "pages": "", "other_ids": {}, "num": null, "urls": [], "raw_text": "K. Sj\u00f6lander, \"Snack sound toolkit\", http://www.speech.kth.se/snack/.", "links": null } }, "ref_entries": { "TABREF0": { "type_str": "table", "html": null, "num": null, "text": "Email: yfliao@ntut.edu.tw \u6458\u8981 \u5728\u672c\uf941\u6587\u4e2d\uff0c\u6211\u5011\u63a2\u8a0e\u5f37\u5065\u5f0f\u6f22\u8a9e\u6587\u5b57\u7279\u5b9a(text-dependent, TD)\u8207\u6587\u5b57\uf967\u7279\u5b9a(text-independent, TI) \u8a9e\u8005\u9a57\u8b49\u7cfb\u7d71\uff0c\u4e3b\u8981\u662f\u91dd\u5c0d\u6f22\u8a9e\u7684\u8072\u8abf\u8a9e\u8a00\u7279\u6027\uff0c\u63d0\u51fa\u6f5b\u5728\u97fb\uf9d8\u5206\u6790(latent prosody analysis, LPA) \u53ca\u9ad8\u65af\u6df7\u5408\u6a21\u578b(Gaussian mixture model, GMM)\uf978\u7a2e\u65b9\u5f0f\uff0c\u5206\u5225\u7528\uf92d\u5efa\u7f6e\u6bcf\u4f4d\u8a9e\u8005\u7684\u97fb\uf9d8\ufa08\u70ba\u6a21 \u578b\u53ca\u80fd\uf97e\u8207\u97f3\u9ad8\u8ecc\u8de1(pitch contour)\u7684\u52d5\u614b\u8b8a\u5316\u6a21\u578b\u3002\u5be6\u9a57\u7d50\u679c\u986f\u793a\u5728\u4f7f\u7528 ISCSLP-SRE \u8a9e\uf9be\u4e4b\u6f22 \u8a9e\u6587\u5b57\u7279\u5b9a\u8207\u6587\u5b57\uf967\u7279\u5b9a\u8a9e\u8005\u9a57\u8b49\u5be6\u9a57\u60c5\u6cc1\u4e0b\uff0c\u4f7f\u7528\u97fb\uf9d8\u8a0a\u606f(prosodic information)\uf92d\u8f14\u52a9\u50b3\u7d71 \u4f7f\u7528\u983b\u8b5c\u7279\u5fb5(spectral features)\u4e4b\u8a9e\u8005\u9a57\u8b49\u7cfb\u7d71\uff0c\u53ef\u6709\u6548\u63d0\u5347\u7cfb\u7d71\u6548\u80fd\u3002", "content": "
\u4e2d\uff0c\u5c0d\u65bc\u77ed\u7a0b\u97fb\uf9d8\u65b9\u9762\u901a\u5e38\u6703\u4f7f\u7528\u9ad8\u65af\u6df7\u5408\u6a21\u578b\uf92d\u7d71\u8a08\u97fb\uf9d8\u8a0a\u606f\uff0c\u80fd\u6355\u6349\u5230\u5982\u97f3\u9ad8\u8207\u80fd\uf97e\u7684\u5206 2. \u6587\u5b57\u7279\u5b9a\u8207\u6587\u5b57\uf967\u7279\u5b9a\u4e4b\u8a9e\u8005\u9a57\u8b49\u7cfb\u7d71\u67b6\u69cb \u8a9e\u8a00\u6bd4\u8f03\u4e4b\u4e0b\u53ef\u77e5\u6f22\u8a9e\u5c6c\u65bc\u4e00\u7a2e\u8072\u8abf\u8a9e\u8a00\uff0c\u96b1\u85cf\u65bc\u5167\u7684\u8072\u8abf\uf901\u662f\u500b\u95dc\u9375\u7684\u56e0\u7d20\uff0c\u5c07\u6703\u5927\u5927\u5730\u5f71\u97ff
\u4f48\u3001\u97f3\u9ad8\u8207\u80fd\uf97e\u7684\u659c\uf961\u4ee5\u53ca\u97f3\u9ad8\u8207\u80fd\uf97e\u7684\u6301\u7e8c\u6642\u9593\u7b49\u97fb\uf9d8\u7279\u5fb5\uff0c\u800c\u9577\u7a0b\u97fb\uf9d8\u6a21\u578b\u5247\u901a\u5e38\u6709 N-gram \u5716\u4e00\u53ca\u5716\u4e8c\u6240\u793a\u5206\u5225\u70ba\u6587\u5b57\uf967\u7279\u5b9a\u8207\u6587\u5b57\u7279\u5b9a\u9a57\u8b49\u4e4b\u6574\u5408\u67b6\u69cb\uff0c\u5c0d\u65bc\u6587\u5b57\uf967\u7279\u5b9a\uf92d\uf96f\uff0c\u5c07\u6709\u97f3\u6846 \u97fb\uf9d8\u8ecc\u8de1\u7684\u8b8a\u5316\u3002
\u53ca discrete hidden Markov model(DHMM) [6] \uf978\u7a2e\u65b9\u6cd5\uff0c\u53ef\u4ee5\u8868\u73fe\u51fa\u97fb\uf9d8\u8a0a\u606f\u96a8\u6642\u9593\u7684\u9577\u7a0b\u8b8a (frame)\u548c\u8a9e\u8005\uf978\u7a2e\u5c64\u7d1a\u4e00\u8d77\u4f7f\u7528\uff0c\u4e3b\u8981\u662f\u56e0\u70ba\u8003\uf97e\u5230\u8a3b\u518a\u8207\u6e2c\u8a66\u8a9e\uf9be\uf969\uf97e\u7684\u95dc\u4fc2\uff0c\u5728\u8a9e\u8005\u5c64\u7d1a\u6240 \u4e00\u822c\uf92d\uf96f\uff0cprosody state N-gram \u8a9e\u8005\u6a21\u578b [6-8] \u5df2\u7d93\u662f\u4ee5\u97fb\uf9d8\u7279\u5fb5\u70ba\u4e3b\u4e4b\u9a57\u8b49\u7cfb\u7d71\u6240\u63a1\u7d0d\u7684
\u5316\u3002\uf967\u904e\u9577\u7a0b\u97fb\uf9d8\u6a21\u578b\u901a\u5e38\u53d7\u9650\u65bc\u5927\uf97e\u8a9e\uf9be\u7684\u9700\u6c42\u554f\u984c\uff0c\u56e0\u70ba\u8981\u6709\u5145\u5206\u8a9e\uf9be\u624d\u80fd\u6709\u6548\u63cf\u8ff0\u97fb\uf9d8\u7684 \u9700\u7684\uf97e\u9060\u6bd4\u97f3\u6846\u5c64\u7d1a\u5927\u5f97\u591a\uff0c\u6cc1\u4e14\u73fe\u5be6\uf9fa\u6cc1\u4e2d\u7e3d\u662f\u53ea\u80fd\u7372\u5f97\u6709\u9650\u7684\u8a9e\uf9be\uf97e\uff1b\u53cd\u89c0\u6587\u5b57\u7279\u5b9a\u7684\u60c5\u6cc1 \u65b9\u6cd5\uff0c\u7136\u800c\u60f3\u8981\u80fd\u5920\u53ef\u9760\u5730\u4f30\u6e2c\u51fa\u6b64 N-gram \u8a9e\u8005\u6a21\u578b\uff0c\u5247\u64c1\u6709\u5927\uf97e\u7684\u8a13\uf996\u8207\u6e2c\u8a66\u8a9e\uf9be\u901a\u5e38\u662f\u5148
\u7279\u6027\uff0c\u6240\u4ee5\u91dd\u5c0d\u9019\u9ede\u7f3a\u5931\u6211\u5011\u5c07\u63d0\u51fa\u6f5b\u5728\u97fb\uf9d8\u5206\u6790\u65b9\u6cd5\uf92d\u5f97\u5230\u53ef\u9760\u7684\u97fb\uf9d8\u8a0a\u606f\u3002 \u5247\u53ea\u662f\u63a1\u53d6\u97f3\u6846\u5c64\u7d1a\u7684\u65b9\u5f0f\uff0c\u56e0\u70ba\u8a72\u8a9e\uf9be\u7684\u9577\ufa01\ufa26\u975e\u5e38\u7684\u7c21\u77ed\u3002 \u6c7a\u689d\u4ef6\uff0c\u8b6c\u5982\uf96f\u77e5\u540d\u7684 NIST2001-SRE Extended Data Task \u4e2d\uff0c\u5206\u5225\u4f7f\u7528 8 \uf906\u548c 2 \uf906\u7d04\uf978\u5206\u9418\u7684
\u672c\u6587\u7ae0\u4e2d\uff0c\u6211\u5011\u6703\u5728\u7cfb\u7d71\u524d\u7aef\u7684\u983b\u8b5c\u7279\u5fb5\u4f7f\u7528 mean subtraction, variance normalization, and \u6587\u5b57\uf967\u7279\u5b9a\u4efb\u52d9\u662f\u7531\u4e09\u7a2e\uf967\u540c\u6a21\u7d44\uf92d\u69cb\u6210\uff0c\u5982\u5716\u4e00\u6240\u793a\uff0c\u9996\u5148\u662f\u9ad8\u65af\u6df7\u5408\u6a21\u578b\u5c07\u97f3\u9ad8\u8207\u80fd\uf97e \u5c0d\u8a71\uf906\u5b50\u4f5c\u70ba\u8a13\uf996\u8207\u6e2c\u8a66\u3002\u7136\u800c\u5728\u6211\u5011\u6240\u63d0\u51fa\u7684\u6f5b\u5728\u97fb\uf9d8\u5206\u6790\u65b9\u6cd5\uf9e8\uff0c\u5927\uf97e\u8a9e\uf9be\u5c07\uf967\u6703\u662f\u5fc5\u9700\u689d
ARMA filtering (MVA) [9] \u53bb\u9664\u90e8\u4efd\u901a\u9053\uf967\u5339\u914d\u7684\u554f\u984c\uff0c\u63a5\u8457\u8a9e\u8005\u9a57\u8b49\u7cfb\u7d71\u5c07\u904b\u7528\uf967\u540c\u6a21\u7d44\uf92d\u6574 \u8ecc\u8de1\u7684\u52d5\u614b\u8b8a\u5316\u8207\u6240\u63d0\u51fa\u4e4b\u6f5b\u5728\u97fb\uf9d8\u5206\u6790\u65b9\u6cd5\u505a\u4e00\u5408\u4f75\uff0c\u5b8c\u6574\u7372\u53d6\u6bcf\u4f4d\u8a9e\u8005\u7684\u97fb\uf9d8\ufa08\u70ba\uff0c\u6700\u5f8c\u5247 \u4ef6\uff0c\u56e0\u70ba\u76f8\u540c\u7684\u8a9e\uf9be\u5eab\u4e0b\u5df2\u80fd\u6210\u529f\u5730\u904b\u7528\u5728\u6587\u5b57\uf967\u7279\u5b9a\u4e4b\u8a9e\u8005\u9a57\u8b49 [8]\uff0c\u4e14\u5e73\u5747\uf92d\uf96f\u50c5\u50c5\u53ea\u9700\u5171
\u5408\u983b\u8b5c\u8207\u97fb\uf9d8\u8a0a\u606f\u3002\u6587\u5b57\uf967\u7279\u5b9a\u689d\u4ef6\u4e0b\uff0c\u6709\u4e09\u7a2e\u6a21\u7d44\u7528\u4f5c\u8a9e\u8005\u78ba\u8a8d\u7cfb\u7d71\u7684\u5efa\u69cb\uff0c\u5305\u62ec\u76ee\u524d\u88ab\u8996\u70ba \u662f\u4ee5 MAP-GMM \u5b8c\u6210\u7cfb\u7d71\u5728\u983b\u8b5c\u7279\u5fb5\u7684\u4e3b\u9ad4\u3002\uf9dd\u7528 MAP-GMM \u53d6\u4ee3\u539f\u672c\u7684\u6587\u5b57\u9650\u5b9a\u8a9e\u8005\u9ad8\u65af \uf978\u5206\u9418\u53ca\u4e09\u5341\u79d2\u7684\u8a13\uf996\u8207\u6e2c\u8a66\u8a9e\uf9be\uf97e\uff0c\u6240\u4ee5\u6211\u5011\u5c07\u5617\u8a66\u8457\u5957\u7528\u6b64\u65b9\u6cd5\u5728\u5c6c\u65bc\u8072\u8abf\u578b\u7684\u8a9e\u8a00\uff0c\u7279\u5225
\u6a19\u6e96\u4f5c\u6cd5\u7684 a maximum a posteriori (MAP)-adapted GMM (MAP-GMM) [10] \u3001\u97f3\u9ad8\u8207\u80fd\uf97e\u4e4b\u9ad8\u65af \u6df7\u5408\u6a21\u578b\uff0c\u539f\u56e0\u5728\u65bc\u5be6\u969b\u61c9\u7528\u60c5\u6cc1\u4e2d\uf967\u53ef\u80fd\u8981\u6c42\u4f7f\u7528\u8005\u5728\u8a3b\u518a\u6642\uf93f\u88fd\u5927\uf97e\u7684\u8a9e\u97f3\uff0c\u4ee5\u81f4\u65bc\u6bcf\u4e00\u500b \u662f\u5728\u6f22\u8a9e\u8a9e\u8a00\u4e0a\u7684\u8868\u73fe\u5c1a\u672a\u80fd\u660e\u78ba\u5730\u5f97\u77e5\u3002
\u6df7\u5408\u6a21\u578b\uff0c\u4ee5\u53ca\u6f5b\u5728\u97fb\uf9d8\u5206\u6790\u6a21\u7d44\u3002\u800c\u6587\u5b57\u7279\u5b9a\u5247\u6709\u53e6\u5916\u4e09\u7a2e\u6a21\u7d44\uf92d\u69cb\u6210\uff0c\u5305\u62ec\u6587\u5b57\u9650\u5b9a\u7684\u8a9e\u8005 \u4eba\u7684\u8a13\uf996\u8a9e\uf9be\u53ef\u80fd\u6709\u4e00\u4e9b\u8072\u5b78\u7279\u6027\u6c92\u88ab\u6db5\u84cb\u5230\uff0c\u5728\u6e2c\u8a66\u6642\u53ef\u80fd\u6703\u9020\u6210\u7cfb\u7d71\u6548\u80fd\u4e0b\ufa09\uff0c\u4e26\u4e14\u6587\u5b57\uf967 \u95dc\u65bc\u6f5b\u5728\u97fb\uf9d8\u5206\u6790\u7684\u57fa\u672c\u69cb\u60f3\u662f\uf9dd\u7528 PLSA \u6982\uf9a3\u627e\u51fa\u4e00\u500b\u4f4e\u7dad\ufa01\u7684\u97fb\uf9d8\u8cc7\u8a0a\u7a7a\u9593\u4ee5\u8868\u793a\u8a9e
\u9ad8\u65af\u6df7\u5408\u6a21\u578b\uff0c\u96b1\u85cf\u5f0f\u99ac\u53ef\u592b\u6a21\u578b(hidden Markov model, HMM)\u4ee5\u53ca\u97f3\u9ad8\u8207\u80fd\uf97e\u4e4b\u9ad8\u65af\u6df7\u5408\u6a21 \u7279\u5b9a\u7684\u78ba\u8a8d\u662f\u7121\u6cd5\u9650\u5236\u6e2c\u8a66\u8a9e\u8005\uf96f\u8a71\u7684\u5167\u5bb9\uff0c\u6240\u4ee5\u5efa\uf9f7\u51fa\uf92d\u7684\u8a9e\u8005\u6a21\u578b\uf967\u50c5\u8981\u80fd\u4ee3\u8868\u8a72\u8a3b\u518a\u8a9e\u8005 \u5716\u4e00\u3001\u6587\u5b57\uf967\u7279\u5b9a\u8a9e\u8005\u9a57\u8b49\u65b9\u6cd5\u4e4b\u65b9\u584a\u5716\u3002 \u8005\u7684\u7279\u6027\u6240\u5728\u4f4d\u7f6e\uff0c\u4e3b\u8981\u662f\u70ba\uf9ba\u64f7\u53d6\u51fa\u91cd\u8981\u7684\u97fb\uf9d8\u7dda\uf96a\uf92d\u9451\u5225\u8a9e\u8005\u4e4b\u9593\u7684\uf967\u540c\uff0c\u518d\u8005\u662f\u8b93\u8a9e\u8005\u7279
\u578b\u3002\u800c\u5f8c\u7aef\u6539\uf97c\u578b\u7684\u6e2c\u8a66\u5206\uf969\u6b63\u898f\u5316(test normalization, T-norm) [11] \u5247\u53ef\u4ee5\u5c0d\u5206\uf969\u4f5c\u8abf\u6574\u3002\u6700\u5f8c \u7684\u7279\u6027\uff0c\u9084\u8981\u80fd\u5920\u6db5\u62ec\u5728\uf967\u540c\u8072\u5b78\u60c5\u6cc1\u4e0b\u7684\u8a9e\u8005\u8b8a\uf962\u6027\u3002 \u5b9a\u7684\u97fb\uf9d8\uf9fa\u614b N-gram \u8a9e\u8005\u6a21\u578b\u80fd\uf901\u53ef\u9760\u7684\u5efa\uf9f7\u3002\u5716\u4e09\u662f\u6f5b\u5728\u97fb\uf9d8\u5206\u6790\u65b9\u6cd5\u5728\u8a9e\u8005\u9a57\u8b49\u61c9\u7528\u7684\u65b9
\u6211\u5011\uf9dd\u7528 MIT \uf9f4\u80af\u5be6\u9a57\u5ba4\u6240\u767c\u5c55\u7684 LNKnet [12] \u8edf\u9ad4\u505a\uf967\u540c\u6a21\u7d44\u5206\uf969\u4e0a\u7684\u7d50\u5408\u3002 \u70ba\uf9ba\u514b\u670d\u96fb\u8a71\u8a71\u7b52\u8207\u901a\u9053\uf967\u5339\u914d\u7684\u5f71\u97ff\uff0c\u97fb\uf9d8\u8a0a\u606f\u7684\u4f7f\u7528\u4ecd\u662f\u6211\u5011\u4e3b\u8981\u8003\uf97e\uff0c\u96d6\u7136\u4f7f\u7528\u9ad8\u65af \u584a\u5716\uff0c\u9996\u5148\u5fc5\u9808\u628a\u8f38\u5165\u8a9e\uf906\u7684\u97fb\uf9d8\u8ecc\u8de1\u7d93\u7531 Tokenization \u81ea\u52d5\u8f49\u63db\u6210\u97fb\uf9d8\uf9fa\u614b\u5e8f\uf99c\uff0c\u4e26\u5728\u8a13\uf996\u968e
\u5728 MVA \u5c0d\u983b\u8b5c\u7279\u5fb5\u7684\u8655\uf9e4\u4e3b\u8981\u662f\u5c07\u7279\u5fb5\u5411\uf97e\u4f5c\u4e00\u7a2e\u6b63\u898f\u5316\uff0c\u96d6\u7136\u8fd1\uf98e\uf92d\u6709\u5f88\u591a\u7279\u5fb5\u6b63\u898f\u5316 \u6df7\u5408\u6a21\u578b\u53ef\u4ee5\u7528\uf92d\u7d71\u8a08\u97fb\uf9d8\u8a0a\u606f\uff0c\u4f46\u4e00\u822c\u53ea\u80fd\u88dc\u6349\u5230\u97f3\u9ad8\u8207\u80fd\uf97e\u8b8a\u5316\u7b49\u77ed\u7a0b\u7684\u97fb\uf9d8\u8a0a\u606f\uff0c\u5176\u6240\u5f97 \u6bb5\u4e2d\u5efa\uf9f7\u8d77 N-gram \u8a9e\u8005\u95dc\u4fc2\u77e9\u9663(co-occurrence matrix)\uff0c\u76ee\u7684\u662f\u96c6\u5408\u6bcf\u4f4d\u8a9e\u8005\u7684\u97fb\uf9d8\ufa08\u70ba\u7279\u6027\uf92d
\u7684\u65b9\u6cd5\uff0c\u5982 feature warping [13] \u53ca histogram equalization(HEQ) [14]\ufa26\u80fd\u6709\u5f88\u597d\u6548\u679c\uff0c\u4f46\u662f MVA \u5230\u7684\u6539\u5584\u5e45\ufa01\u4ecd\u7136\u6709\u9650\uff0c\u800c\u5c0d\u65bc\u88dc\u6349\u8f03\u9577\u7a0b\u97fb\uf9d8\u8a0a\u606f\u8b8a\u5316\u7684\u65b9\u6cd5\u901a\u5e38\u6709 DHMM \u548c N-gram \uf978\u7a2e\uff0c \u5b78\u7fd2\u97fb\uf9d8\uf9fa\u614b\u8cc7\u8a0a\u548c\u8a9e\u8005\u4e4b\u9593\u7684\u76f8\u4e92\u95dc\u4fc2\u3002
\u7684\uf97c\u597d\u8868\u73fe\u8207\u7c21\u55ae\u4f7f\u7528\u662f\u6211\u5011\u5728\u6b64\u512a\u5148\u8003\uf97e\u7684\u539f\u56e0\u3002\u800c\u6587\u5b57\uf967\u7279\u5b9a\u8a9e\u8005\u9a57\u8b49\u4e2d\u7684 MAP-GMM \u662f \u53ef\u662f\ufa26\u9700\u4f7f\u7528\u5927\uf97e\u7684\u8a13\uf996\u8207\u6e2c\u8a66\u8a9e\uf9be\uff0c\u5c0d\u6b64\u6211\u5011\u63d0\u51fa\u6f5b\u5728\u97fb\uf9d8\u5206\u6790\u7684\u65b9\u6cd5\u662f\u80fd\u5728\u6709\u9650\u7684\u8a9e\uf9be\u60c5\u6cc1 \u5716\u56db\u5247\uf96f\u660e Tokenization \u5982\u4f55\u81ea\u52d5\u6a19\u8a18\u53ca\u8f49\u63db\u6210\u97fb\uf9d8\uf9fa\u614b\u5e8f\uf99c\u3002\u7531 piece-wise curve fitting \u5148
1. \u5e8f\uf941 \u900f\u904e\u901a\u7528\u80cc\u666f\u6a21\u578b(universal background model, UBM)\u8abf\u9069\u51fa\u8a9e\u8005\u500b\u5225\u7684\u9ad8\u65af\u6df7\u5408\u6a21\u578b\uff0c\u4f7f\u6bcf\u500b\u8a9e \u4e0b\u5f97\u5230\u53ef\u9760\u7684\u97fb\uf9d8\u8a0a\u606f\u3002 \u628a\u6bcf\u4e00\u6bb5\u50b3\u5165\u7684\u97fb\uf9d8\u8ecc\u8de1\u64f7\u53d6\u5176\u97fb\uf9d8\u7279\u5fb5\u5411\uf97e\uff0c\u4e14\u8a31\u591a\u9130\u8fd1\u7684\u5340\u6bb5\u5c07\uf905\uf99a\u6210\u4e00\u500b\u9f90\u5927\u7684\u97fb\uf9d8\u7279\u5fb5
\u8a9e\u8005\u9a57\u8b49\u5728\u73fe\u4eca\u7684\u8a9e\u97f3\u8655\uf9e4\u4e2d\u70ba\u91cd\u8981\u7684\u5206\u652f\u7814\u7a76\u9805\u76ee\u4e4b\u4e00 [1] \uff0c\u76ee\u524d\u6709\u76f8\u7576\u591a\u7684\u7814\u7a76\uf967\u65b7\u5730\u6301 \u8005\u6a21\u578b\u6240\u542b\u84cb\u7684\u8072\u5b78\u7279\u6027\uf901\u5177\u5b8c\u6574\u6027\uff0c\u5982\u6b64\u5c0d\u65bc\u6587\u5b57\u5167\u5bb9\u7684\u8b8a\uf962\u6027\u5c31\u80fd\u5ee3\u6cdb\u63a5\u7d0d\u3002\u800c\u97fb\uf9d8\u7279\u5fb5\u7531 \u5728\u6587\u5b57\u7279\u5b9a\u689d\u4ef6\u4e0b\uff0c\u6709\u53e6\u5916\u4e09\u7a2e\u6a21\u7d44\u7528\u4f5c\u8a9e\u8005\u9a57\u8b49\u7cfb\u7d71\u7684\u5efa\u69cb\uff0c\u5982\u5716\u4e8c\u6240\u793a\uff0c\u5305\u62ec\u7372\u77e5\u97f3\u9ad8 supervector\uff0c\u800c\u8003\uf97e\u5230\u97f3\u7bc0\u70ba\u6700\u5c0f\u7684\u97fb\uf9d8\u55ae\u4f4d\uff0c\u6240\u4ee5\u63a1\u7528\u4e94\u7a2e\u97f3\u7bc0\u5c64\u6b21\u7684\u97fb\uf9d8\u7279\u5fb5\uf96b\uf969\uff0c\u5305\u62ec\u4e00
\u7e8c\u767c\u5c55\u4e2d\u3002\u5c24\u5176\u5f9e 1996 \uf98e\u958b\u59cb\uff0cNIST \u6a5f\u69cb\u6bcf\uf98e\ufa26\u6703\u85c9\u7531\u8209\u8fa6\u8a9e\u8005\u8fa8\u8a8d\u8a55\u4f30(speaker recognition \uf978\u65b9\u9762\u8457\u624b\uff0c\u77ed\u7a0b\u97fb\uf9d8\u7528\u9ad8\u65af\u6df7\u5408\u6a21\u578b\u5c0d\u80fd\uf97e\u8207\u97f3\u9ad8\u8ecc\u8de1\u5efa\u7f6e\u5176\u52d5\u614b\u8b8a\u5316\u6a21\u578b\uff0c\u9577\u7a0b\u5247\u7528\u6240\u63d0\u51fa \u8207\u80fd\uf97e\u8ecc\u8de1\u52d5\u614b\u8b8a\u5316\u7684\u9ad8\u65af\u6df7\u5408\u6a21\u578b\u3001\u6a21\u578b\u5316\u6885\u723e\u983b\uf961\u5012\u983b\u8b5c\u4fc2\uf969\u66ab\u614b\u8ecc\u8de1\u6240\u7528\u7684\u96b1\u85cf\u5f0f\u99ac\u53ef\u592b \u500b\u6bcd\u97f3\u5340\u6bb5\u7684\u97f3\u9ad8\u659c\uf961(pitch slope)\u548c\u9577\ufa01\u7684\u5ef6\u9577\u8b8a\u5316(lengthening factor)\u3001\uf978\u500b\u6bcd\u97f3\u9593\u7684\u5c0d\uf969\u80fd\uf97e
evaluation, SRE)\uf92d\u63d0\u4f9b\u4e00\u500b\u5171\u540c\u7684\u6e2c\u8a66\u5e73\u53f0 [2] \uff0c\u4ee5\u4fc3\u9032\u8a9e\u8005\u8fa8\u8a8d\u6280\u8853\u6f14\u9032\u53ca\u5404\u7a2e\u6f14\u7b97\u65b9\u6cd5\u7684\u5be6 \u7684\u6f5b\u5728\u97fb\uf9d8\u5206\u6790\uf901\u6709\u6548\u5730\u5f97\u77e5\u97fb\uf9d8\ufa08\u70ba\uff0c\u5176\u4e3b\u8981\u662f\u5c07\u8a9e\u8005\u9a57\u8b49\u554f\u984c\u8f49\u63db\u70ba\uf9d0\u4f3c\u6587\u4ef6\u6aa2\uf96a \u6a21\u578b\u4ee5\u53ca\u7d71\u8a08\u6885\u723e\u983b\uf961\u5012\u983b\u8b5c\u4fc2\uf969\u5206\u4f48\u7684\u9ad8\u65af\u6df7\u5408\u6a21\u578b\u3002\u4e00\u822c\uf92d\uf96f\uff0c\u8a9e\u8005\u9a57\u8b49\u7cfb\u7d71\ufa26\u6703\u63a1\u7528\u6885\u723e \u5dee\u548c\u97f3\u9ad8\u8df3\u8e8d(pitch jump)\u4ee5\u53ca\uf978\u500b\u97f3\u7bc0\u4e4b\u9593\u7684\u66ab\u505c\u9577\ufa01(pause duration)\u3002\u6b64\u5916\u70ba\uf9ba\u79fb\u9664\u8a9e\uf906\u767c\u97f3
\u7528\u6027\uff0c\uf901\u8b93\u5168\u4e16\u754c\u6700\u65b0\u7a4e\u7684\u60f3\u6cd5\u5f97\u4ee5\u5728\u7af6\u8cfd\uf9e8\u7372\u5f97\u9a57\u8b49\u3002\u76f8\u8f03\u65bc\u5916\u570b\u8a9e\u8a00\uff0c\u6f22\u8a9e\u7684\u8a9e\u8005\u8fa8\u8a8d\u7af6\u8cfd (document retrieval)\u7684\u554f\u984c\uff0c\u7d71\u8a08\u51fa\u97fb\uf9d8\u5e8f\uf99c\u7684\u7d44\u5408\u4e26\u5efa\uf9f7\u97fb\uf9d8\u7a7a\u9593(prosody space)\uff0c\u518d\u900f\u904e \u983b\uf961\u5012\u983b\u8b5c\u4fc2\uf969\u8207\u9ad8\u65af\u6df7\u5408\u6a21\u578b\u7684\u642d\u914d\uff0c\u5176\u4e2d\u6885\u723e\u983b\uf961\u5012\u983b\u8b5c\u4fc2\uf969\u5df2\u7d93\u5c07\u8a9e\u97f3\u7684\u983b\u8b5c\u7279\u5fb5\u505a\uf9ba\uf97c \u5167\u5bb9(context-information)\u5c0d\u97fb\uf9d8\u8b8a\u5316\u7684\u5f71\u97ff\uff0c\u5fc5\u9808\u5c07\u97fb\uf9d8\u7279\u5fb5\uf96b\uf969\u505a\u6b63\u898f\u5316\u7684\u52d5\u4f5c\uff0c\u85c9\u7531\u6574\u500b
\u9084\u5728\u8d77\u6b65\u968e\u6bb5\uff0c\u5728 2006 \uf98e\u8209\u8fa6\u7684\u4e2d\u6587\u53e3\u8a9e\u8a9e\u8a00\u8655\uf9e4\u570b\u969b\u6703\u8b70(ISCSLP)\u4e2d\uff0c\u9996\ufa01\u5efa\uf9f7\uf9ba\u6f22\u8a9e\u8a9e\u8a00 probabilistic latent semantic analysis(PLSA)[15-16] \u7684\u7a7a\u9593\u7dad\ufa01\u7c21\u5316\u5f8c\uf92d\u5448\u73fe\u8a9e\u8005\u7684\u97fb\uf9d8\u6a21\u578b\u3002 \u597d\u63cf\u8ff0\uff0c\u7136\u5f8c\u900f\u904e\u7531\u8a31\u591a\u9ad8\u65af\u5bc6\ufa01\u51fd\uf969\u7d44\u6210\u7684\u9ad8\u65af\u6df7\u5408\u6a21\u578b\uf92d\u8868\u793a\u8a9e\u8005\u7279\u6027\u7684\u5206\u4f48\uff0c\u800c\u9019\uf9e8\u4e26\uf967 \u8a13\uf996\u8a9e\uf9be\u6240\u7d71\u8a08\u51fa\uf92d\u4e4b\u97fb\uf9d8\u7279\u5fb5\uf96b\uf969\u7684\u5e73\u5747\u503c\u53ca\u6a19\u6e96\u5dee\uff0c\u79fb\u53bb\u4efb\u4f55\u975e\u97fb\uf9d8\u7279\u6027\u7684\u5f71\u97ff\u3002\u65bc\u662f\u4e00\u500b
\u7684\u8a9e\u8005\u7af6\u8cfd\u6a5f\u5236 [3] \uff0c\u8b93\u6b64\uf9b4\u57df\u7684\u7814\u7a76\u4eba\u54e1\u80fd\u5920\u540c\u6642\u5728\u64c1\u6709\u4e00\u6a23\u7684\u8cc7\u6e90\u4e0b\uff0c\u900f\u904e\u4e2d\u6587\u8a9e\u8a00\u8cc7\u6e90 \u6587\u5b57\u7279\u5b9a\u8a9e\u8005\u9a57\u8b49\u4efb\u52d9\u5c0d\u4f7f\u7528\u8005\u7684\uf96f\u8a71\u5167\u5bb9\u662f\u6709\u5176\u9650\u5236\uff0c\u6240\u4ee5\u5c0d\u8a9e\u97f3\u4e8b\u4ef6\u4e4b\u8072\u5b78\u8b8a\u5316\u6709\u8a73\u7d30 \u5982\u6587\u5b57\uf967\u7279\u5b9a\u4e2d\u4f7f\u7528 MAP-GMM \uf92d\u5efa\uf9f7\u8a9e\u8005\u7279\u5b9a\u6a21\u578b\uff0c\u56e0\u70ba\u85c9\u52a9\u901a\u7528\u80cc\u666f\u6a21\u578b\u88dc\u5f37\u7684\u8072\u5b78\u7279\u6027 \u5716\u4e8c\u3001\u6587\u5b57\u7279\u5b9a\u8a9e\u8005\u9a57\u8b49\u65b9\u6cd5\u4e4b\u65b9\u584a\u5716\u3002 \u4ee5\u5411\uf97e\uf97e\u5316\u70ba\u57fa\u790e\uff0c\u900f\u904e Expectation-Maximization(EM)\u6f14\u7b97\u6cd5\u8a13\uf996\u597d\u7684\u97fb\uf9d8\u6a21\u578b\uf965\u53ef\u4ee5\u81ea\u52d5\u5730
\uf997\u76df(Chinese Corpus Consortium, CCC) [4] \u6240\u63d0\u4f9b\u7684\u8cc7\uf9be\u5eab\uff0c\ufa00\u78cb\u6f22\u8a9e\u8a9e\u8005\u7684\u8fa8\u8a8d\u6280\u8853\u8207\u7814\u7a76\u3002 \u8003\u616e\u7684\u96b1\u85cf\u5f0f\u99ac\u53ef\u592b\u6a21\u578b\u662f\u5fc5\u9700\u7684\uff0c\u9019\u6a23\u624d\u80fd\u5584\u7528\u7cfb\u7d71\u5c0d\u4f7f\u7528\u8005\u5148\u5929\u7684\u9650\u5236\u689d\u4ef6\u3002\u7576\u7136\u9ad8\u65af\u6df7\u5408 \u53cd\u800c\u6703\u5c0d\u6587\u5b57\u7279\u5b9a\u7522\u751f\u56f0\u64fe\uff0c\u9020\u6210\u8a9e\u8005\u6a21\u578b\u7121\u6cd5\u91dd\u5c0d\u6587\u5b57\u7279\u5b9a\u4efb\u52d9\u9032\ufa08\u9a57\u8b49\u3002 \u628a\u8f38\u5165\u8a9e\uf906\u6240\u69cb\u6210\u7684 supervector \u4f5c\u7b26\u865f\u7684\u6a19\u8a18\uff0c\u4e26\u4e14\u518d\u8f49\u63db\u6210\u4e00\uf99a\uf905\u7684\u97fb\uf9d8\uf9fa\u614b\u5e8f\uf99c\u3002
\u8a9e\u8005\u9a57\u8b49\u6280\u8853\u5728\u73fe\u5be6\u751f\u6d3b\u4e2d\u53ef\u4ee5\u6709\u8a31\u591a\u7684\u61c9\u7528\uff0c\uf9b5\u5982\u53ef\u4ee5\u85c9\u7531\u96fb\u8a71\uf99a\u63a5\u5230\u9280\ufa08\u6216\u662f\u4fe1\u7528\u5361\u7b49 \u6a21\u578b\u5728\u983b\u8b5c\u4e0a\u5c0d\u8a9e\u8005\u7279\u6027\u7684\u63cf\u8ff0\u4ecd\u662f\uf967\u53ef\u6216\u7f3a\u7684\u89d2\u8272\uff0c\u56e0\u70ba\u7528\u9ad8\u65af\u5bc6\ufa01\u51fd\uf969\u8868\u793a\u8a9e\u8005\u7684\u8072\u5b78\uf9d0\u5225 \u53e6\u5916\u5716\u4e8c\u7684\u6587\u5b57\u7279\u5b9a\u8a9e\u8005\u9a57\u8b49\u4e2d\uff0c\u6211\u5011\u53ef\u77e5\u6e2c\u8a66\u8a9e\u8005\uf96f\u8a71\u5167\u5bb9\u662f\u6709\u9650\u5236\u6027\u7684\uff0c\u5b83\u5fc5\u9808\u7b26\u5408\u5ba3 \u6700\u5f8c\u503c\u5f97\u4e00\u63d0\u7684\u662f\u7cfb\u7d71\uf967\uf941\u7279\u5b9a\u6216\uf967\u7279\u5b9a\u7684\u4efb\u52d9\uff0c\u5c0d\u53d6\u81ea\u65bc\u6885\u723e\u983b\uf961\u5012\u983b\u8b5c\u4fc2\uf969\u7684\u7279\u5fb5\u5411\uf97e \u7372\u5f97\u97fb\uf9d8\uf9fa\u614b N-gram \u8a9e\u8005\u95dc\u4fc2\u77e9\u9663\u5f8c\uff0c\u7531\u65bc\u8a13\uf996\u8a9e\uf9be\u8207\u6e2c\u8a66\u8a9e\uf9be\u4e4b\u8cc7\uf9be\uf97e\u7684\uf967\u8db3\uff0c\u5728\u53d7\u6b64
\u5ba2\u670d\u4e2d\u5fc3\uff0c\u4e26\u76f4\u63a5\u900f\u904e\u4f7f\u7528\u8005\u7684\u8072\u97f3\uf92d\u9a57\u8b49\u8eab\u4efd\u4ee5\u5373\u6642\u63d0\u4f9b\uf965\uf9dd\u7684\u79c1\u4eba\u670d\u52d9\u3002\u7136\u800c\u4f7f\u7528\u8005\uf974\u4efb\u610f \u4ecd\u53ef\u53cd\u61c9\u51fa\u8a9e\u8005\u7279\u6027\u5206\u4f48\uff0c\u8207\u96b1\u85cf\u5f0f\u99ac\u53ef\u592b\u6a21\u578b\u5206\u5c6c\uf967\u540c\u89d2\ufa01\u7684\u5206\u6790\u3002\u800c\u5728\u97fb\uf9d8\u8a0a\u606f\u65b9\u9762\uff0c\u8003\u616e \u7a31\u8a9e\u8005\u5728\u7cfb\u7d71\u4e2d\u8a3b\u518a\u8a9e\uf9be\u7684\u8a9e\uf906\u5167\u5bb9\uff0c\u9664\u6b64\u5167\u5bb9\u5916\u7684\u8a9e\uf906\ufa26\u5c07\u4e00\uf9d8\u62d2\u7d55\uff0c\u5373\uf965\u662f\u771f\u5be6\u8a9e\u8005\uf96f\u51fa\uf967 \u6211\u5011\ufa26\uf9dd\u7528 MVA \u53bb\u9664\u90e8\u4efd\u901a\u9053\uf967\u5339\u914d\u554f\u984c\uff0c\u56e0\u70ba\u983b\u8b5c\u4e0a\u53d7\u901a\u9053\u9020\u6210\u7684\u504f\u79fb\uf97e\u76f8\u7576\u65bc\u6642\u9593\u4e0a\u7684\u65cb \u9650\u5236\u4e4b\u4e0b\u4ee5\u97fb\uf9d8\u8a0a\u606f\u6240\u5efa\u69cb\u51fa\u7684 N-gram \u8a9e\u8005\u6a21\u578b\u53ef\u80fd\uf967\u5920\u5177\u6709\u7d71\u8a08\u7279\u6027\uff0c\u6c92\u8fa6\u6cd5\u6e96\u78ba\u7684\u8a13\uf996\u51fa
\u4f7f\u7528\uf967\u540c\u7684\u96fb\u8a71\u8a71\u7b52\u6216\u901a\u9053\uff0c\u5247\u6703\u6709\u96fb\u8a71\u8a71\u7b52\u8207\u901a\u9053\u74b0\u5883\uf967\u5339\u914d\u554f\u984c\uff0c\u800c\u5c0e\u81f4\u50b3\u7d71\u4ee5\u983b\u8b5c\u7279\u5fb5\u70ba \u5230\u8a9e\uf9be\u9577\ufa01\u7684\u7f3a\u4e4f\uff0c\u50c5\u5c0d\u77ed\u7a0b\u97fb\uf9d8\u65b9\u9762\u4f7f\u7528\u9ad8\u65af\u6df7\u5408\u6a21\u578b\uf92d\u63cf\u8ff0\u97f3\u9ad8\u8207\u80fd\uf97e\u8ecc\u8de1\u7684\u52d5\u614b\u8b8a\u5316\u3002\u81f3 \u540c\u6a23\u7684\u5167\u5bb9\u4e5f\u662f\u7121\u6cd5\u63a5\u53d7\u7684\uff0c\uf9dd\u7528\u9019\u7a2e\u7cfb\u7d71\u4f7f\u7528\u4e0a\u7684\u9650\u5236\u689d\u4ef6\uff0c\u96b1\u85cf\u5f0f\u99ac\u53ef\u592b\u6a21\u578b\u6703\u662f\uf901\u9069\u5408\u7528 \u7a4d\u6027(convolutional)\u566a\u97f3\uff0c\u800c\u6885\u723e\u983b\uf961\u5012\u983b\u8b5c\u4fc2\uf969\u5c0d\u5e73\u5747\u503c\u7684\u524a\u6e1b\u6b63\u53ef\u4ee5\u5c0d\u6297\u65cb\u7a4d\u6027\u566a\u97f3\u4e0b\u4e4b\u5931 \u4ee3\u8868\u8a9e\u8005\u97fb\uf9d8\u7279\u6027\u7684\u8a9e\u8005\u6a21\u578b\uff0c\u6240\u4ee5\u5fc5\u9808\u518d\u7d93\u904e PLSA \u627e\u51fa\ufa09\u4f4e\u7dad\ufa01\u7684\u97fb\uf9d8\u8cc7\u8a0a\u7a7a\u9593\uff0c\u5982\u5716\u4e94\u6240
\u4e3b\u4e4b\u8a9e\u8005\u9a57\u8b49\u7cfb\u7d71\u6548\u80fd\ufa09\u4f4e\u3002\u70ba\uf9ba\u6539\u5584\u96fb\u8a71\u8a71\u7b52\u8207\u901a\u9053\uf967\u5339\u914d\u554f\u984c\uff0c\u8fd1\uf98e\uf92d\u6709\u8a31\u591a\u4eba\uf9dd\u7528\u97fb\uf9d8\u8a0a \u65bc\u7cfb\u7d71\u5f8c\u7aef\u6211\u5011\u4e5f\u8003\u616e\u5230\u5206\uf969\u7684\u8b8a\u5316\u6027\uff0c\uf92d\u81ea\u8a9e\u8005\u4e4b\u9593\uf96f\u8a71\u5167\u5bb9\u6216\u662f\u9577\ufa01\u7684\uf967\u540c\ufa26\u6703\u9020\u6210\u5f71\u97ff\uff0c \uf92d\u5efa\uf9f7\u6a21\u578b\u7684\u65b9\u6cd5\uff0c\u56e0\u70ba\u96b1\u85cf\u5f0f\u99ac\u53ef\u592b\u6a21\u578b\u5c0d\u6885\u723e\u983b\uf961\u5012\u983b\u8b5c\u4fc2\uf969\u4e4b\u66ab\u614b\u8ecc\u8de1\u53ef\u4ee5\u6709\u8a73\u7d30\u7684\u63cf \u771f\uff0c\u81f3\u65bc\u8b8a\uf962\uf969\u6b63\u898f\u5316\u8207\uf984\u6ce2\u5668\u7684\u4f7f\u7528\u5247\u5206\u5225\u53ef\u4ee5\u5c0d\u6297\u52a0\u6210\u6027(additive)\u8207\u9ad8\u529f\uf961\u52a0\u6210\u6027\u566a\u97f3\u4e0b\u7684 \u793a\uff0c\u800c\u5176\u5206\u89e3\u5b9a\u7fa9\u5982\u4e0b\uff0c
\u606f\uf92d\u5f37\u5316\u50b3\u7d71\u4ee5\u983b\u8b5c\u7279\u5fb5\u70ba\u57fa\u790e\u4e4b\u8a9e\u8005\u9a57\u8b49\u7cfb\u7d71 [5-8] \u7684\u6548\u80fd\uff0c\u97fb\uf9d8\u7279\u5fb5(prosodic feature)\uf967\u50c5 \u4e14\u8a13\uf996\u548c\u6e2c\u8a66\u74b0\u5883\u7684\uf967\u5339\u914d\uf901\u662f\u4e00\u5927\u4e3b\u56e0\uff0c\u6240\u4ee5\u4f7f\u7528\u6539\uf97c\u5f0f\u6e2c\u8a66\u5206\uf969\u6b63\u898f\u5316(modified test \u8ff0\uff0c\u800c\u9ad8\u65af\u6df7\u5408\u6a21\u578b\u4e2d\u4e26\u672a\u8003\u616e\u5230\u8a9e\u97f3\u4e8b\u4ef6\u7684\u8072\u5b78\u8b8a\u5316\u3002 \u5931\u771f\u3002\u5728\u6587\u5b57\uf967\u7279\u5b9a\u8a9e\u8005\u9a57\u8b49\u7cfb\u7d71\u5f8c\u6bb5\u7684\u5206\uf969\u65b9\u9762\uff0c\uf901\u900f\u904e\u6539\uf97c\u5f0f\u6e2c\u8a66\u5206\uf969\u6b63\u898f\u5316\u505a\u88dc\u511f\uff0c\u5c07\u540c
\u542b\u6709\u8a9e\u8005\u8a0a\u606f\u4e26\u5df2\u88ab\u8a8d\u5b9a\u662f\uf967\uf9e0\u53d7\u5230\u96fb\u8a71\u8a71\u7b52\u8207\u901a\u9053\uf967\u5339\u914d\u7684\u5f71\u97ff\uff0c\u800c\u4e14\u5728\u897f\u65b9\u8a9e\u8a00\u7684\u7814\u7a76\u4e2d\u4ea6 normalization, MT-norm)\uf92d\u8abf\u6574\u76ee\u6a19\u8a9e\u8005(target speaker)\u6a21\u578b\u7684\u5206\uf969\uff0c\uf925\u958b\u76ee\u6a19\u8a9e\u8005\u8207\u5192\u5145\u8a9e\u8005 \u4ee5\u983b\u8b5c\u7279\u5fb5\u70ba\u4e3b\u7684\u7cfb\u7d71\uf92d\uf96f\uff0c\u96b1\u85cf\u5f0f\u99ac\u53ef\u592b\u6a21\u578b\u8207\u9ad8\u65af\u6df7\u5408\u6a21\u578b\u7684\u7d50\u5408\u5df2\u7d93\u53ef\u4ee5\u7372\u5f97\u9084\uf967\u932f \u5115\u8a9e\u8005\u6a21\u578b(cohort model set)\u5206\uf969\u7684\u5e73\u5747\u503c\u8207\u8b8a\uf962\uf969\uf92d\u8abf\u6574\u76ee\u6a19\u8a9e\u8005\u6a21\u578b\u7684\u5206\uf969\uff0c\u7d93\u7531\u6e1b\u53bb\u5e73\u5747
\u6709\u5f88\u591a\u7684\u6587\u737b\u8b49\u5be6\u5176\u6548\u679c\u3002\u56e0\u6b64\u5728\u672c\uf941\u6587\u4e2d\u6211\u5011\u5c07\u8457\u91cd\u5728\u8a0e\uf941\u5982\u4f55\uf9dd\u7528\u97fb\uf9d8\u7279\u5fb5\uf92d\u5f37\u5316\u6f22\u8a9e\u8a9e\u8005 (impostor)\u4e4b\u9593\u7684\u5206\u4f48\uff0c\u9032\u800c\u6539\u5584\u6b63\u78ba\uf961\u4e26\uf901\u7c21\uf9e0\u7522\u751f\u9a57\u8b49\u6240\u7528\u7684\u9580\u6abb\u503c\u3002 \u7684\u7d50\u679c\uff0c\u7136\u800c\u5728\u8a13\uf996\u8207\u6e2c\u8a66\u74b0\u5883\uf967\u5339\u914d\u7684\uf9fa\u6cc1\u4e0b\uff0c\u4ecd\u9700\u52a0\u5165\uf967\u540c\u89c0\u9ede\u7684\u97fb\uf9d8\u8a0a\u606f\uf92d\u5f37\u5065\u7cfb\u7d71\uff0c\u56e0 \u503c\u53ef\u4ee5\u4f7f\u5192\u5145\u8a9e\u8005\u5206\uf969\u7684\u5206\u4f48\u4e2d\u5fc3\u79fb\u81f3\u539f\u9ede\uff0c\u800c\u9664\u4ee5\u8b8a\uf962\uf969\u5247\u80fd\u5c07\u5192\u5145\u8a9e\u8005\u5206\uf969\u5206\u4f48\u4e4b\u6a19\u6e96\u5dee\u9650 1
\u9a57\u8b49\u7cfb\u7d71\u7684\u6548\u80fd\uff0c\u4e3b\u8981\u662f\u8003\u616e\u5230\u6f22\u8a9e\u5c6c\u65bc\u4e00\u7a2e\u8072\u8abf(tonal)\u8a9e\u8a00\uff0c\u5176\u672c\u8cea\u4e0a\u4f9d\u8cf4\u8072\u8abf\u7684\uf967\u540c\uf92d\u5340\u5225 \u7531\u65bc\u983b\u8b5c\u7279\u5fb5\u8207\u97fb\uf9d8\u7279\u5fb5\u662f\u5448\u73fe\u8a0a\u865f\u4e2d\uf967\u540c\u7684\u8a0a\u606f\uff0c\u6240\u4ee5\u8003\u616e\u5176\u4e4b\u9593\u53ef\u80fd\u7684\u4e92\u88dc\u7279\u6027\uff0c\u5247\u6587 \u6b64\u6211\u5011\u8003\u616e\u97f3\u9ad8\u8207\u80fd\uf97e\u8ecc\u8de1\u7684\u52d5\u614b\u8b8a\u5316\uff0c\uf9dd\u7528\u6709\u8072\u97f3(voiced)\u7684\u5340\u6bb5\u4e2d\u53d6\u51fa\u6bcf\u4e00\u500b\u97f3\u6846\u7684\u5c0d\uf969(log) \u5b9a\u70ba\u4e00\uff0c\u9019\u6a23\u7684\u65b9\u6cd5\uf967\u50c5\u53ef\u4ee5\uf925\u958b\u76ee\u6a19\u8a9e\u8005\u8207\u5192\u5145\u8a9e\u8005\u4e4b\u9593\u7684\u5206\u4f48\u9032\u800c\u6539\u5584\u6b63\u78ba\uf961\uff0c\u9084\u80fd\u8b93\u6c7a\u5b9a
\u51fa\u540c\u97f3\uf962\u5b57\u8a5e\uff0c\u6545\u97fb\uf9d8\u7279\u5fb5\u5c0d\u6f22\u8a9e\u7684\u5f71\u97ff\u61c9\u8f03\u897f\u65b9\u8a9e\u8a00\u5f37\uf99f\u3002 \u5b57\uf967\u7279\u5b9a\u8207\u6587\u5b57\u7279\u5b9a\u8a9e\u8005\u9a57\u8b49\u7684\uf967\u540c\u6a21\u7d44\u5fc5\u9808\u6574\u5408\uff0c\u800c\u6211\u5011\u662f\u900f\u904e\u591a\u5c64\u611f\u77e5\u6a5f(multi-layer \u97f3\u9ad8\u53ca\u5c0d\uf969\u80fd\uf97e\uff0c\u4e26\u4f30\u8a08\u5c0d\uf969\u97f3\u9ad8\u53ca\u80fd\uf97e\u7684\u4e00\u968e\u5fae\u5206\uf92d\u5efa\uf9f7\u9ad8\u65af\u6df7\u5408\u6a21\u578b\uff0c\u800c\u7531\u65bc\u97fb\uf9d8\u7279\u5fb5\u662f\u6bd4 \u63a5\u53d7\u8207\u5426\u7684\u9580\u6abb\u503c\uf901\u5bb9\uf9e0\u7522\u751f\u3002\u53e6\u5916\uff0c\u591a\u5c64\u611f\u77e5\u6a5f\u53ef\u7528\uf92d\u7d50\u5408\u5404\u6a21\u7d44\u4e4b\u8a9e\u8005\u7684\u6e2c\u8a66\u5206\uf969\uff0c\u9032\u4e00\u6b65
\u4e00\u822c\uf92d\uf96f\u983b\u8b5c\u7279\u5fb5\u4ee3\u8868\u662f\u8f03\u77ed\u7a0b(short term)\u4e14\u4f4e\u968e\u5c64\u7684\u8072\u5b78\u8a0a\u606f\uff0c\ufa26\u662f\u548c\u767c\u97f3\u5668\u5b98\u76f8\u95dc\u7684\u5be6 perceptrons, MLPs)\u8207 development \u7684\u6e2c\u8a66\u8a9e\uf9be\uf92d\u6c7a\u5b9a\u9a57\u8b49\u7cfb\u7d71\u7684\u5408\u4f75\u65b9\u5f0f\uff0c\u5728\u7cfb\u7d71\u6c42\u5f97\u7684\u5206\uf969\u4e0a \u8f03\uf967\u53d7\u8a71\u7b52\u6216\u901a\u9053\u7684\u5f71\u97ff\uff0c\u6240\u4ee5\u53ef\u4ee5\u88dc\u5f37\u539f\u983b\u8b5c\u7cfb\u7d71\u7684\u7f3a\u5931\u3002 \u628a\u983b\u8b5c\u7cfb\u7d71\u53ca\u97fb\uf9d8\u7cfb\u7d71\u878d\u5408\uff0c\u4ee5\uf965\u5f37\u5316\u9a57\u8b49\u7cfb\u7d71\u4e4b\u6548\u80fd\u3002
\u9ad4\u7dda\uf96a\uff0c\u5176\u4e2d\u88ab\u5ee3\u6cdb\u4f7f\u7528\u7684\u6885\u723e\u983b\uf961\u5012\u983b\u8b5c\u4fc2\uf969(Mel-frequency cepstral coefficients, MFCCs)\u662f\u53ef \u4f5c\u975e\u7dda\u6027\u7d44\u5408\uff0c\u4ee5\u9054\u5230\uf9dd\u7528\u97fb\uf9d8\u7279\u5fb5\uf92d\u5f37\u5316\u6f22\u8a9e\u8a9e\u8005\u9a57\u8b49\u7cfb\u7d71\u4e4b\u76ee\u7684\u3002
\u4ee5\u64f7\u53d6\u4e26\u50b3\u9054\u51fa\u767c\u97f3\u8154\u9053(vocal tract)\u7684\u5206\u4f48\uff1b\u97fb\uf9d8\u7279\u5fb5\u5247\u901a\u5e38\u4f5c\u70ba\u8072\u9580\u8cc7\u8a0a(glottic source)\u7684\u7279\u5fb5 \u672c\u6587\u5167\u5bb9\u5b89\u6392\u5982\u4e0b\uff1a\u7b2c\u4e8c\u7ae0\u7bc0\u63cf\u8ff0\u5728\u6f22\u8a9e\u8a9e\u8a00\uf9e8\u6240\u4f7f\u7528\u7684\u5404\u7a2e\u65b9\u6cd5\uff0c\u4e26\u8a0e\uf941\u97fb\uf9d8\u7279\u5fb5\u5728\u7cfb\u7d71 3. \u6f5b\u5728\u97fb\uf9d8\u5206\u6790
\uf96b\uf969\uff0c\uf967\u50c5\u662f\u8f03\u9577\u7a0b(long term)\u4e14\u9ad8\u968e\u7684\u7279\u5fb5\u4e26\u542b\u6709\u8a9e\u8005\u672c\u8eab\u7279\u6b8a\u7684\u8a0a\u606f\uff0c\u5982\u97f3\u9ad8\u8ecc\u8de1\u53ca\u97f3\u8abf \u4e2d\u7684\u8f14\u52a9\u4f5c\u7528\uff1b\u7b2c\u4e09\u7bc0\u5247\u8a73\u7d30\uf96f\u660e\u6f5b\u5728\u97fb\uf9d8\u5206\u6790\u7684\u65b9\u6cd5\uff1b\u7b2c\u56db\u7bc0\u662f\u7cfb\u7d71\u904b\u7528\u5728 ISCSLP2006-SRE \u5728\u8a9e\u97f3\u8a0a\u865f\u4e2d\uff0c\u97fb\uf9d8\u8a0a\u606f\u7684\u52d5\u614b\u8b8a\u5316\u53d7\u5230\u5404\u7a2e\u6f5b\u85cf\u56e0\u7d20\u7684\u5f71\u97ff\uf901\u751a\u65bc\u8a9e\u8005\u672c\u8eab\u7279\u6027\uff0c\u8b6c\u5982\uf96f\u8a71\u901f
(intonation)\u7b49\uff0c\u56e0\u6b64\uf978\u8005\u5404\u662f\u5448\u73fe\u8a9e\u97f3\u8a0a\u865f\u4e2d\uf967\u540c\u7684\u8a0a\u606f\u3002\u5728\u97fb\uf9d8\u8a0a\u606f\u6539\u5584\uf967\u5339\u914d\u554f\u984c\u7684\u65b9\u6cd5 \u7684\u5be6\u9a57\u7d50\u679c\uff1b\u6700\u5f8c\u5247\u70ba\u7d50\uf941\u3002 \ufa01\u3001\u60c5\u7dd2\u8f49\u8b8a\u4ee5\u53ca\uf96f\u8a71\u5167\u5bb9\u7b49\u7b49\uff0c\u56e0\u6b64\u6211\u5011\u6240\u89c0\u5bdf\u5230\u97fb\uf9d8\u8ecc\u8de1\u8868\u8c61\u7684\u8b8a\uf962\uf97e\u662f\u76f8\u7576\u5927\u3002\u800c\u8ddf\u897f\u65b9
" }, "TABREF1": { "type_str": "table", "html": null, "num": null, "text": "False Rejection Rate, FR)\uff0c\u5373\u6b63\u78ba\u8a9e\u8005\u7684 \u5728\u6b64\u8a9e\uf9be\u5eab\u4e2d\uff0c\uf967\uf941\u662f\u6587\u5b57\u7279\u5b9a \u4f9b\u7684 development \u8207 evaluation \u8cc7\uf9be\u5eab\uff0c\u800c\u6b64\u8a9e\uf9be\u5eab\u6240\u6709\u7684\u8072\u97f3\u6a94\u6848\ufa26\u662f\u63a1\u7528 8kHz \u53d6\u6a23\u983b\uf961\uff0c\u4e14 \u70ba 16bits \u55ae\u8072\u9053\u7684 PCM \u683c\u5f0f\u3002\u81f3\u65bc evaluation \u7684\u8a9e\uf9be\u5eab\u4e2d\uff0c\u5176\u771f\u5be6\u8a9e\u8005\u8207\u5192\u5145\u8a9e\u8005\u7684\u6e2c\u8a66\u6a23\u672c", "content": "
4. \u6539\uf97c\u5f0f\u6e2c\u8a66\u5206\uf969\u6b63\u898f\u5316 5. \u6f22\u8a9e\u4e4b\u8a9e\u8005\u9a57\u8b49\u5be6\u9a57\u7d50\u679c \u76f8\u7b49\u932f\u8aa4\uf961\u662f\u4e00\u7a2e\u8a55\u4f30\u8a9e\u8005\u9a57\u8b49\u7cfb\u7d71\u7684\u65b9\u5f0f\uff0c\u6240\u8b02\u7684\u76f8\u7b49\u932f\u8aa4\uf961\u5c31\u662f\u932f\u8aa4\u62d2\u7d55\uf961\u8207\u932f\u8aa4\u63a5\u53d7 \u6642\u6240\u7528\u7684\u9ea5\u514b\u98a8\u7279\u6027\ufa26\u5728\u8a13\uf996\u904e\u7a0b\u4e2d\u9047\u904e\uff0c\u5982\u6b64\u73fe\u8c61\u5728\u96b1\u85cf\u5f0f\u99ac\u53ef\u592b\u6a21\u578b\u7cfb\u7d71\u4e2d\u5c07\u662f\uf967\u8b00\u800c\u5408\u3002 \u6b64\u5916\uff0c\u8868\u4e00\u8207\u8868\u4e8c\u5448\u73fe\u6587\u5b57\uf967\u7279\u5b9a\u8207\u6587\u5b57\u7279\u5b9a\u8a9e\u8005\u9a57\u8b49\u7cfb\u7d71\u7684\u6240\u6709\u7d50\u679c\u3002\u5f9e\u8868\u4e00\u4e2d\u7684\u6bd4\u8f03\u6211
\u5728\u6539\uf97c\u5f0f\u6e2c\u8a66\u5206\uf969\u6b63\u898f\u5316\u7684\u539f\uf9e4\u662f\uf9dd\u7528\u4e00\u7fa4\u76f8\u4f3c\u65bc\u76ee\u6a19\u8a9e\u8005\u6a21\u578b\u7684\u540c\u5115\u8a9e\u8005\u6a21\u578b\uff0c\u4f30\u8a08\u51fa\u76f8\u4f3c\u65bc 5.1. ISCSLP2006-SRE \u8a9e\uf9be\u5eab \uf961\u76f8\u7b49\u6642\u7684\u6a5f\uf961\u503c\uff0c\u4f46\u5728\u67d0\u4e9b\u7279\u6b8a\u7684\u60c5\u5f62\u4e2d\uff0c\u932f\u8aa4\u62d2\u7d55\u8207\u932f\u8aa4\u63a5\u53d7\u7684\u5f8c\u679c\u548c\u91cd\u8981\u6027\u4e26\uf967\u76f8\u7b49\u3002\u8209 \u6587\u5b57\u7279\u5b9a\u8a9e\u8005\u9a57\u8b49\u4e2d\uff0c\u6709\u6587\u5b57\u9650\u5b9a\u7684\u8a9e\u8005\u9ad8\u65af\u6df7\u5408\u6a21\u578b\u53ca\u96b1\u85cf\u5f0f\u99ac\u53ef\u592b\u6a21\u578b\uf978\u7a2e\u6a21\u7d44\u88ab\u7528\uf92d \u5011\u767c\u73fe\u97fb\uf9d8\u7279\u5fb5\u7684\u4f5c\u7528\u5fc5\u9808\u5efa\u69cb\u5728\u983b\u8b5c\u7279\u5fb5\u7684\u7cfb\u7d71\u4e0a\uff0c\u5118\u7ba1 22.7%\u548c 17.7%\u7684\u76f8\u7b49\u932f\u8aa4\uf961\u53ca 0.272
\u6bcf\u4f4d\uf967\u540c\u76ee\u6a19\u8a9e\u8005\u7684\u5192\u5145\u8a9e\u8005\uff0c\u9019\u548c\u539f\u59cb\u6e2c\u8a66\u5206\uf969\u6b63\u898f\u5316\u65b9\u6cd5\u6709\u6240\uf967\u540c\uff0c\u56e0\u70ba\u6b63\u898f\u5316\u6240\u7528\u7684\uf96b\uf969 \u8207\u6587\u5b57\uf967\u7279\u5b9a\u7684\u8a9e\u8005\u9a57\u8b49\u4efb\u52d9\uff0c\ufa26\u662f\uf92d\u81ea\u4e2d\u6587\u8a9e\u8a00\u8cc7\u6e90\uf997\u76df\u6240\u63d0 \uf9b5\uf92d\uf96f\uff0c\u8a9e\u8005\u9a57\u8b49\u61c9\u7528\u5728\uf90a\u878d\u4ea4\uf9e0\u7684\u60c5\u6cc1\uff0c\u70ba\uf9ba\u907f\u514d\u5192\uf9b4\u76dc\u7528\uff0c\u56e0\u6b64\u932f\u8aa4\u63a5\u53d7\u7684\u6a5f\uf961\u5fc5\u9808\u6e1b\u81f3\u6700 \u5efa\u69cb\u9a57\u8b49\u7cfb\u7d71\uff0c\u5305\u542b 16 \u6df7\u5408\uf969\u7684\u9ad8\u65af\u6df7\u5408\u6a21\u578b\u53ca 8 \u6df7\u5408\uf969\u7684\u96b1\u85cf\u5f0f\u99ac\u53ef\u592b\u6a21\u578b\uff0c\u800c\u96b1\u85cf\u5f0f\u99ac\u53ef \u8207 0.223 \u6c7a\u7b56\u6210\u672c\u51fd\uf969\uf967\u80fd\u548c\u983b\u8b5c\u7279\u5fb5\u7684 4.0%\u548c 0.047 \u76f8\u6bd4\uff0c\u4f46\uf978\u8005\u5408\u4f75\u5f8c\u7684\u76f8\u7b49\u932f\u8aa4\uf961\u53ca\u6c7a\u7b56
\uf967\u518d\u7531\u540c\u4e00\u7d44\u540c\u5115\u8a9e\u8005\u6a21\u578b\u5f97\u5230\uff0c\u800c\u662f\u91dd\u5c0d\u6bcf\u500b\u76ee\u6a19\u8a9e\u8005\u6a21\u578b\u627e\u51fa\u5404\u5225\u5c0d\u61c9\u7684\u540c\u5115\u8a9e\u8005\u6a21\u578b\uff0c\u5982 \u4f4e\u3002\u800c\u6c7a\u7b56\u6210\u672c\u51fd\uf969\u5247\u662f\u88ab\u5b9a\u7fa9\u6210\u4e00\u7a2e\u932f\u8aa4\u6a5f\uf961\u7684\u52a0\u6b0a\u7e3d\u548c\uff0c\u5982\u4e0b\u6240\u793a\u3002 \u592b\u6a21\u578b\u7684\uf9fa\u614b\uf969\u76ee\u662f\u6839\u64da\u8a3b\u518a\u8a9e\uf9be\u4e2d\u6587\u5b57\u7684\u591a\u5be1\uf92d\u505a\u8abf\u6574\u3002\u5f9e\u5716\u5341\u7684\u7d50\u679c\u770b\u5230\uff0c\u96d6\u7136\uf978\u7a2e\u65b9\u6cd5\u7d50 \u6210\u672c\u51fd\uf969\u70ba 3.8%\u548c 0.045\uff0c\u78ba\u5be6\u6539\u5584\uf9ba\u6587\u5b57\uf967\u7279\u5b9a\u8a9e\u8005\u9a57\u8b49\u7cfb\u7d71\u6b63\u78ba\uf961\u3002\u800c\u8868\u4e8c\u4e2d\uf901\u986f\u793a\u6587\u5b57\u7279
( , ) i j ( ) s I O S \u03bb \u628a\u9580\u6abb\u503c\u63d0\u9ad8\uff0c\u5247\u932f\u8aa4\u62d2\u7d55\uf961\u5c07\u6703\u63d0\u9ad8\uff0c\u800c\u932f\u8aa4\u63a5\u53d7\uf961\u5247\u6703\ufa09\u4f4e\uff1b\uf974\u9580\u6abb\u503c\ufa09\u4f4e\uff0c\u5247\u932f\u8aa4\u62d2\u7d55\uf961 i d i w A PLSA P(D|Z) P(Z) P(W|Z) MT-norm score ( ) I T O S \u03bc \u672a\u6b63\u898f\u5316\u4e4b\u983b\u8b5c\u7279\u5fb5\u4e0b\u5c0d\u65bc\u7cfb\u7d71\u7684\u5f71\u97ff\uff0c\u76f8\u5c0d\u7684\u6587\u5b57\u7279\u5b9a\u8a9e\u8005\u9a57\u8b49\u4e5f\u662f\u5982\u6b64\u3002 \u5728\u983b\u8b5c\u7279\u5fb5\u7684\u5be6\u9a57\u4e2d\uff0c\u91dd\u5c0d\u7cfb\u7d71\u524d\u7aef\u8655\uf9e4\u5247\u8003\uf97e cepstral mean normalization(CMN)\u3001MVA \u53ca MV \u03c3 = Test Message \u8005\u7684\u5206\uf969\u9ad8\u65bc\u9580\u6abb\u503c\u9020\u6210\u63a5\u53d7\u7684\u932f\u8aa4\uf961\u3002FA\u3001FR \u9019\uf978\u7a2e\u932f\u8aa4\uf961\u662f\u4e00\u7a2e\u53d6\u6368(tradeoff)\u7684\u95dc\u4fc2\uff0c\uf974 \u97fb\uf9d8\u7279\u5fb5\u7684\u8f14\u52a9\u4f5c\u7528\uff0c\u7528\u4ee5\u5f37\u5316\u4f7f\u7528\u983b\u8b5c\u7279\u5fb5\u4e4b\u8a9e\u8005\u9a57\u8b49\u7cfb\u7d71\u7684\u6548\u80fd\uff0c\u56e0\u6b64\u6c92\u6709\u8a0e\uf941\u97fb\uf9d8\u7279\u5fb5\u5728 5.3.2 \u6587\u5b57\u7279\u5b9a\u8a9e\u8005\u9a57\u8b49\u7d50\u679c \u2212 \u5206\uf969\u5c0f\u65bc\u9580\u6abb\u503c\u9020\u6210\u62d2\u7d55\u7684\u932f\u8aa4\uf961\u3002\u53e6\u4e00\u7a2e\u662f\u932f\u8aa4\u63a5\u53d7\uf961(False Acceptance Rate, FA)\uff0c\u5373\u4eff\u5192\u8a9e \u5148\u505a\u7279\u5fb5\u6b63\u898f\u5316\u7684\u8655\uf9e4\uff0c\u56e0\u70ba\u6211\u5011\u4e3b\u8981\u662f\u5148\u6c7a\u5b9a\u597d\u983b\u8b5c\u7279\u5fb5\u65b9\u9762\u9a57\u8b49\u6548\u679c\u6700\u597d\u7684\u67b6\u69cb\u5f8c\uff0c\u518d\u85c9\u7531 prosody P d w \u5716\u4e09\u3001\u6f5b\u5728\u8a9e\u610f\u5206\u6790\u65b9\u6cd5\u8f14\u52a9\u7cfb\u7d71\u4e4b\u65b9\u584a\u5716\u3002 \u5716\u56db\u3001\u81ea\u52d5\u6a19\u8a18\u53ca\u8f49\u63db\u6210\u97fb\uf9d8\uf9fa\u614b\u5e8f\uf99c\u7684\u65b9\u584a\u5716\u3002 high dimensional prosody space Latent \u6b64\u624d\u80fd\u627e\u51fa\u6bcf\u500b\u76ee\u6a19\u8a9e\u8005\u6a21\u578b\u771f\u6b63\u7684\u5192\u5145\u8a9e\u8005\u7fa4\uff0c\u9019\u6a23\u7684\u65b9\u5f0f\u4ea6\u53ef\u4ee5\u5e36\uf92d\u6e1b\u5c11\u904b\u7b97\uf97e\u7684\u597d\u8655\uff0c\u56e0 \u70ba\u5c0d\u65bc\u548c\u76ee\u6a19\u8a9e\u8005\u6a21\u578b\u8f03\uf967\u76f8\u4f3c\u7684\u540c\u5115\u8a9e\u8005\u6a21\u578b\u53ef\u4ee5\uf967\u518d\u8003\u616e\u5176\u5f71\u97ff\u3002\u800c\u6211\u5011\u7684\u540c\u5115\u8a9e\u8005\u6a21\u578b\u4f30 \u6e2c\u662f\u6839\u64da\u8a13\uf996\u8a9e\uf9be\u5c0d\u6bcf\u500b\u8a9e\u8005\u7279\u5b9a\u6a21\u578b\uf97e\u6e2c log likelihood score \u800c\u5f97\uff0c\u5982\u5716\uf9d1\u6240\u793a\uff0c\u9019\u548c [11] \u4e2d \uf9dd\u7528\u8ddd\uf9ea\u7684\u4f30\u6e2c\u65b9\u5f0f\u662f\u6709\u6240\uf967\u540c\u7684\uff0c\u4e3b\u8981\u662f\u85c9\u7531 log likelihood score \u7684\u9ad8\u4f4e\uf92d\u6c7a\u5b9a\u540c\u5115\u8a9e\u8005\u6a21 \u578b\uff0c\u4e26\u9078\u51fa\u524d\u9762 K \u500b\u540c\u5115\u8a9e\u8005\u6a21\u578b\uf92d\u8a08\u7b97\uf96b\uf969\u3002\u63a5\u8457\u6bcf\u4f4d\uf967\u540c\u76ee\u6a19\u8a9e\u8005\u4f9d\u64da\u76f8\u5c0d\u61c9\u7684\u540c\u5115\u8a9e\u8005 \u6a21\u578b\u8a08\u7b97\u51fa\u5206\uf969\u7684\u5e73\u5747\u503c\u8207\u8b8a\uf962\uf969\u4f5c\u70ba\u8abf\u6574\u76ee\u6a19\u8a9e\u8005\u5206\uf969\u7684\uf96b\uf969\uff0c\u5176\u5b9a\u7fa9\u5982\u4e0b\uff0c -I s T I S S \u03bb \u03bc \u03c3 = (2) \u5176\u4e2d s S \u03bb \u70ba\u6e2c\u8a66\u8a9e\uf9be\u8207\u8a9e\u8005\u6a21\u578b s \u03bb \u6240\u8a08\u7b97\u7684 log likelihood score\uff0c I \u03bc \u8207 I \u03c3 \u5206\u5225\u4ee3\u8868\u6e2c\u8a66\u8a9e\uf9be\u76f8 \u5c0d\u65bc\u540c\u5115\u8a9e\u8005\u6a21\u578b\u5206\uf969\u7684\u5e73\u5747\u503c\u8207\u8b8a\uf962\uf969\uff0c \u70ba\u7d93\u904e\u6e2c\u8a66\u6b63\u898f\u5316\u5f8c\u7684\u5206\uf969\u3002\u5f9e(2)\u5f0f\u4e2d\u770b\u5230\u6e2c\u8a66 \u6b63\u898f\u5316\u6e1b\u53bb T S I \u03bc \uff0c\u9019\u52d5\u4f5c\u53ef\u4ee5\u5c07\u5192\u5145\u8a9e\u8005\u5206\uf969\u7684\u5206\u4f48\u4e4b\u4e2d\u5fc3\u79fb\u81f3\u539f\u9ede\uff0c\u4ea6\u5373\u540c\u6642\uf925\u5927\u76ee\u6a19\u8a9e\u8005\u8207 \u5192\u5145\u8a9e\u8005\u7684\u5206\uf969\u5206\u4f48\uff0c\u800c\u9664\u4ee5 I \u03c3 \u5247\u53ef\u4ee5\u5c07\u5192\u5145\u8a9e\u8005\u7684\u5206\uf969\u5206\u4f48\u4e4b\u6a19\u6e96\u5dee\u9650\u5b9a\u70ba\u4e00\uff0c\u9032\u800c\u63d0\u5347\u6b63 \u78ba\uf961\u3002\u800c\u6574\u500b\u6539\uf97c\u5f0f\u6e2c\u8a66\u5206\uf969\u6b63\u898f\u5316\u7684\u67b6\u69cb\u5247\u5982\u5716\u4e03\u6240\u793a\u3002 \u5716\uf9d1\u3001\u540c\u5115\u8a9e\u8005\u6a21\u578b\u7684\u4f30\u6e2c\u3002 Likelihood computation Choose K highest ( , ) cohort cohort \u03bc \u03c3 Cohort Model Cohort Model \u8005\u9a57\u8b49\u4e4b\u8a9e\uf9be\u5eab PR3C2005\uff0c\u8a9e\uf9be\u5eab\u5305\u542b\uf9ba\u7537\uf981\u6027\u5404 5 \u4f4d\u7684\u500b\u5225\u8cc7\uf9be\uf97e\uff0c\u6bcf\u500b 2C2005-1000\uff0c\u8a9e\uf9be\u5eab\u53ea\u5305\u542b\uf9ba 300 \u4f4d\u7537\u6027\u8a9e\u8005\uff0c\u6bcf\u4f4d\u8a9e\u8005 \u7684\u983b\u8b5c\u7279\u5fb5\u70ba\u4e3b\u4e4b\u8a9e\u8005\u9a57\u8b49\u7cfb\u7d71\ufa26\u7528 39 \u7dad\u7684\u6885\u723e\u983b\uf961\u5012\u983b\u8b5c\u4fc2\uf969\u4f5c\u70ba\u7279\u5fb5\uf96b \u8b5c\u7279\u5fb5\u8207 5.1.1 \u6587\u5b57\u7279\u5b9a\u8a9e development \u7684\u8cc7\uf9be\u53d6\u81ea\u65bc CCC-V \u4eba\u7684\u8072\u97f3\u900f\u904e\u4e09\u7a2e\uf967\u540c\u9ea5\u514b\u98a8\u901a\u9053\uf92d\u7372\u5f97\uff0c\u5206\u5225\u7528\"micl\"\u3001\"micr\"\u53ca\"micu\"\u4e09\u7a2e\u7b26\u865f\uf92d\u8868\u793a\uff0c\u5176 \u4e2d\u6bcf\u4f4d\u8a9e\u8005\u5728\u5206\u5225\u901a\u9053\u4e0a\u6709\u4e94\u7a2e\uf906\u5b50\u6703\u91cd\u8907\uf93f\u88fd 4 \u904d\uff0c\u800c\u53e6\u5916\u4e8c\u5341\u4e00\u7a2e\uf906\u5b50\u53ea\u6703\u5404\uf93f\u88fd\u4e00\u6b21\u3002\u5c0d \u8a9e\uf906\u9577\ufa01\u7d04\u6709 4.5 \u79d2\uff0c\u6700\u5f8c\u7528\uf92d\u8a66\u9a57\u7684\uf906\u5b50\u5171\u6709 11181 \u500b\u4e14\u5e73\u5747\u6642\u9593\u9577\ufa01\u70ba 5.2 \u79d2\u3002\u503c\u5f97\u4e00\u63d0\u7684 \u662f\u6bcf\uf906\u767c\u8a71\u958b\u59cb\ufa26\u6709\u5f88\u9577\u7684\u975c\u97f3\uff0c\u6b64\u5916\uff0c\u67d0\u4e9b\u96d6\u7136\u7531\u76f8\u540c\u8a9e\u8005\u6240\u767c\u51fa\u4f46\u537b\u70ba\uf967\u540c\u8a9e\uf906\u5167\u5bb9\u7684\uf906 \u5b50\uff0c\u6211\u5011\u61c9\u8a72\u8996\u70ba\u5192\u5145\u8a9e\u8005\u4e26\u52a0\u4ee5\u62d2\u7d55\u6389\u3002 5.1.2 \u6587\u5b57\uf967\u7279\u5b9a\u8a9e\u8005\u9a57\u8b49\u4e4b\u8a9e\uf9be\u5eab development \u7684\u8cc7\uf9be\u53d6\u81ea\u65bc CCC-VPR \u542b\u6709\uf978\u7a2e\u8a9e\uf906\uff0c\u5206\u5225\u7531\u96fb\u8a71\u7dda(PTSN)\u53ca\u624b\u6a5f\u901a\u9053(GSM)\u6240\u88fd\u6210\uff0c\u6240\u4ee5\u7e3d\u5171\u6709 600 \u500b\uf906\u5b50\u5728\u5167\u3002 evaluation \u7684\u90e8\u4efd\u5247\u5171\u6709 800 \u4f4d\u8a3b\u518a\u8a9e\u8005\uff0c\u6bcf\u4e00\u4f4d\ufa26\u53ea\u6703\u6709\u4e00\uf906\u5f9e\u96fb\u8a71\u7dda\u6216\u662f\u624b\u6a5f\u901a\u9053\u6240\u63d0\u4f9b\u7684 \u8a9e\uf9be\uff0c\u5e73\u5747\u7528\uf92d\u8a3b\u518a\u7684\u8a9e\uf906\u9577\ufa01\u7d04\u6709 36.2 \u79d2\uff0c\u6700\u5f8c\u7528\uf92d\u8a66\u9a57\u7684\uf906\u5b50\u5171\u6709 11800 \u500b\u4e14\u5e73\u5747\u6642\u9593\u9577 \ufa01\u70ba 15.9 \u79d2\u3002 5.2. \u5be6\u9a57\u689d\u4ef6 \u672c\u6587\u4e2d\u5c0d\u65bc\u6240\u6709 \uf969\uff0c\u5305\u62ec\u524d 13 \u7dad\u5012\u983b\u8b5c\u4fc2\uf969(\u5305\u542b C 0 )\u53ca\u5176\u5dee\u5206\u503c \u0394-MFCCs \u8207\u4e8c\u6b21\u5dee\u5206\u503c \u03942-MFCCs\uff0c\u81f3\u65bc\u97f3\u9ad8 \u8207\u80fd\uf97e\u7684\u8ecc\u8de1\u5247\u662f\u85c9\u7531 snack \u8edf\u9ad4\u5957\u4ef6\u4e2d\u7684 ESPS \u97f3\u9ad8\u64f7\u53d6\u6f14\u7b97\u6cd5\uf92d\u6c42\u5f97 [17] \uff0c\u540c\u6642\u4e5f\u8a08\u7b97\u5176 \u5dee\u5206\u503c\u53ca\u4e8c\u6b21\u5dee\u5206\u503c\uff0c\u6700\u5f8c\u5247\u5c07\u97f3\u9ad8\u8207\u80fd\uf97e\uf99a\u540c\u6885\u723e\u983b\uf961\u5012\u983b\u8b5c\u4fc2\uf969\u4e00\u4f75\u4f5c\u70ba\u4f7f\u7528\u3002 \u53e6\u5916\u5728\u9a57\u8b49\u7cfb\u7d71\u7684\u5408\u4f75\u65b9\u5f0f\uff0c\u6211\u5011\u662f\u904b\u7528\u5171\u6709 120 \u500b\u96b1\u85cf\u7bc0\u9ede\u7684\u591a\u5c64\u611f\u77e5\u6a5f\uff0c\u5c07\u983b \u97fb\uf9d8\u7279\u5fb5\u5728 evaluation \u6e2c\u8a66\u8a9e\uf9be\u4e0a\u6240\u5f97\u5230\u7684\u5206\uf969\u4f5c\u4e00\u7d50\u5408\u3002\u9019\u90e8\u4efd\u7684\u6b65\u9a5f\u662f\u9808\u5148\u628a development \u7684\u8a9e\uf9be\u5340\u5206\u70ba\u8a13\uf996\u548c\u6e2c\u8a66\u4f7f\u7528\uff0c\u5176\u8a13\uf996\u8a9e\uf9be\u90e8\u5206\u7528\uf92d\u8a13\uf996\u51fa\u5404\u500b\u7cfb\u7d71\u7684\u6a21\u578b\uf96b\uf969\uff0c\u800c\u5176\u6e2c\u8a66\u8a9e\uf9be \u7684\u90e8\u4efd\u5247\u7528\uf92d\u5f97\u53d6\u6bcf\u500b\u7cfb\u7d71\u7684\u8fa8\uf9fc\u7d50\u679c\uff0c\u63a5\u8457\u7e7c\u7e8c\u518d\uf9dd\u7528\u5176\u6e2c\u8a66\u8a9e\uf9be\u7684\u90e8\u4efd\u5efa\u7f6e\u51fa\u591a\u5c64\u611f\u77e5\u6a5f\u7684 \u5404\u9805\uf96b\uf969\uff0c\u7136\u5f8c\uf965\u53ef\u5c07\u5404\u500b\u7cfb\u7d71\u7684\u878d\u5408\uf96b\uf969\u56fa\u5b9a\u5957\u7528\u5230\u4e4b\u5f8c evaluation \u6e2c\u8a66\u8a9e\uf9be\u6240\u5f97\u5230\u7684\u5206\uf969 \u4e0a\uff0c\u800c\u591a\u5c64\u611f\u77e5\u6a5f\u7684\u5404\u9805\uf96b\uf969\u662f\u7531\u5be6\u9a57\uf969\u64da\u6240\u5f97\uff0c\u56e0\u6b64\u5728\u5f8c\u9762\u7684\u6587\u5b57\uf967\u7279\u5b9a\u8207\u6587\u5b57\u7279\u5b9a\u8a9e\u8005\u9a57\u8b49 \u5be6\u9a57\u7d50\u679c\uff0c\u6240\u5448\u73fe\u7684\u662f\u7cfb\u7d71\u4e4b\u6700\u4f73\u6548\u679c\u3002 \u8a9e\u8005\u9a57\u8b49\u7cfb\u7d71\u7684\u932f\u8aa4\uf961\u6709\uf978\u7a2e\uff1a\u4e00\u7a2e\u662f ( ) 1 DET Miss Miss Target False False Target C C P P C P P = \u22c5 \u22c5 + \u22c5 \u22c5 \u2212 \u679c\u7684\u66f2\u7dda\u8da8\u52e2\u6709\u6240\uf967\u540c\uff0c\u4f46\u9ad8\u65af\u6df7\u5408\u6a21\u578b\u6700\u4f73\u7684\u76f8\u7b49\u932f\u8aa4\uf961\u53ca\u6c7a\u7b56\u6210\u672c\u51fd\uf969\u5206\u5225\u70ba 2.9%\u548c \u5b9a\u8a9e\u8005\u9a57\u8b49\u7684\u97fb\uf9d8\u7279\u5fb5\u5c0d\u65bc\u7cfb\u7d71\u5f37\u5316\u6709\u5f88\uf967\u932f\u7684\u5e6b\u52a9\uff0c\u8b93\u7cfb\u7d71\u5f9e\u983b\u8b5c\u7279\u5fb5\u6700\u4f73\u7d50\u679c\u7684 1.9%\u8207 (3) \u5176\u4e2d \u3002 \u5448\u73fe\u932f\u8aa4\u62d2\u7d55\uf961\u53ca\u932f\u8aa4\u63a5\u53d7\uf961\u7684\u65b9\u5f0f\u5247\u4f7f\u7528\u5075\u6e2c\u932f\u8aa4\u4ea4\uf9e0\u66f2\u7dda\u5716(Detection Error M \u9a57\u8b49\u7d50\u679c\u6703\u4f5c\u70ba\u5404\u9805\u6a21\u7d44\u6bd4\u8f03\u7684\u6a19\u6e96\uff0c\u5176\u901a\u7528\u80cc\u666f\u6a21\u578b\u4e00\uf9d8 10, 1, 0.05 Miss False Target C C P = = = \u53e6\u5916\uff0c Tradeoff Curve, DET Curve)\uff0c\u6b64\u7a2e\u65b9\u5f0f\u662f\u5047\u8a2d\u76ee\u6a19\u8a9e\u8005\u548c\u4eff\u5192\u8a9e\u8005\u7684\u5c0d\uf969\u76f8\u4f3c\ufa01\u6bd4\u5206\uf969\u70ba\uf978\u5404\u500b \uf967\u540c\u7684\u9ad8\u65af\u5206\u4f48\uff0c\u96a8\u8457\u9580\u6abb\u503c\u7684\u8b8a\u5316\u8868\u73fe\u51fa\u76f8\u5c0d\u61c9\u932f\u8aa4\u62d2\u7d55\uf961\u53ca\u932f\u8aa4\u63a5\u53d7\uf961\u7684\u66f2\u7dda\u8b8a\u5316\u3002 5.3.1 \u6587\u5b57\uf967\u7279\u5b9a\u8a9e\u8005\u9a57\u8b49\u7d50\u679c \u9996\u5148\uff0c\u983b\u8b5c\u7279\u5fb5\u70ba\u4e3b\u7684 MAP-GM \u70ba\u6240\u6709\u8a3b\u518a\u8a9e\u8005\u8a13\uf996\u8a9e\uf9be\u96c6\u6210\u4e26\u7528 1024 \u6df7\u5408\uf969\u7d44\u6210\uff0c\u800c\u8a9e\u8005\u7279\u5b9a\u7684\u9ad8\u65af\u6df7\u5408\u6a21\u578b\u662f\uf9dd\u7528\u5176\u8a3b\u518a \u8a9e\uf9be\u5411\u901a\u7528\u80cc\u666f\u6a21\u578b\u8abf\u9069\u5f97\u5230\u3002\u91dd\u5c0d\u7cfb\u7d71\u524d\u7aef\u6709\uf978\u7a2e\u7279\u5fb5\u6b63\u898f\u5316\u7684\u65b9\u6cd5\u6703\u4f5c\u70ba\u8003\uf97e\uff0c\u5305\u62ec cepstral mean and variance normalization (MV)\u53ca MVA\uff0c\u9019\uf978\u7a2e\u65b9\u5f0f\u7684\u7d50\u679c\u5728\u5716\u516b\u53ef\u770b\u51fa\uff0cMVA \u7684\u6548\u679c\u660e \u986f\u6bd4 MV \u597d\uf9ba\u8a31\u591a\uff0c\u56e0\u6b64\u6211\u5011\u5c07 MAP-GMM \u548c MVA \u7684\u7d44\u5408\u4f5c\u70ba\u6700\u57fa\u672c\u7684\u8a9e\u8005\u9a57\u8b49\u7cfb\u7d71\u3002\u63a5\u8457 \u7cfb\u7d71\u5f8c\u7aef\u7684\u6539\uf97c\u5f0f\u6e2c\u8a66\u5206\uf969\u6b63\u898f\u5316\u5247\u4ee5 320 \u500b\u76f8\u4f3c\u8a9e\u8005\u70ba\u4e3b\uff0c\u7531\u5716\u516b\u53ef\u770b\u5230\u6b64\u65b9\u6cd5\u5c0d\u9a57\u8b49\u7d71\u78ba\u5be6 \u6709\u5f88\u5927\u5f71\u97ff\u4e26\u5927\u5e45\u6539\u5584 MAP-GMM \u7684\u7d50\u679c\uff0c\u7531\u6b64\u53ef\u77e5 MVA \u8207\u6539\uf97c\u5f0f\u5206\uf969\u6b63\u898f\u5316\u6cd5\uf967\u50c5\u76f8\u7576\u6709 \u6548\u4e14\u662f\u4e92\u76f8\u88dc\u511f\u3002\u6240\u4ee5\u6211\u5011\u5728\u6587\u5b57\uf967\u7279\u5b9a\u8a9e\u8005\u9a57\u8b49\u4e2d\uff0c\u4ee5 MVA \u8207\u6539\uf97c\u5f0f\u5206\uf969\u6b63\u898f\u5316\u6cd5\u548c MAP-GMM \u7684\u7d50\u5408\u65b9\u5f0f\uff0c\u4f5c\u70ba\u983b\u8b5c\u7279\u5fb5\u65b9\u9762\u6700\u4f73\u7684\u67b6\u69cb\uff0c\u723e\u5f8c\u518d\u52a0\u5165\u97fb\uf9d8\u7279\u5fb5\u7684\u8f14\u52a9\u3002 \u5728\u5716\u4e5d\u4e2d\u6211\u5011\u770b\u5230\uf978\u7a2e\u97fb\uf9d8\u6a21\u578b\u5316\u7684\u65b9\u6cd5\u88ab\u7528\uf92d\u548c\u983b\u8b5c\u7279\u5fb5\u6700\u4f73\u4e4b\u6548\u679c\u505a\u4e00\u7d50\u5408\u3002\u4ee5\u9ad8\u65af\u6df7 \u5408\u6a21\u578b\u65b9\u5f0f\uf92d\uf96f\uff0c\u8a72\u8a9e\u8005\u7279\u5b9a 64 \u6df7\u5408\uf969\u7684\u97fb\uf9d8\u6a21\u578b\u662f\u76f4\u63a5\u7531\u5176\u8a3b\u518a\u8a9e\uf9be\u8a13\uf996\u800c\u6210\uff0c\u800c\u975e\u900f\u904e\u80cc \u666f\u6a21\u578b\uf92d\u8abf\u9069\uff0c\u5176\u9a57\u8b49\u7d50\u679c\u7684\u76f8\u7b49\u932f\u8aa4\uf961\u53ca\u6c7a\u7b56\u6210\u672c\u51fd\uf969\u5206\u5225\u70ba 17.7%\u548c 0.223\uff1b\u53e6\u5916\u6f5b\u5728\u97fb\uf9d8 \u5206\u6790\u65b9\u5f0f\u5247\u4f7f\u7528\u5230 bi-gram \u6a21\u578b\u53ca 11 \u500b\uf9fa\u614b\u7684\u5411\uf97e\uf97e\u5316(8 \u500b\u70ba\u97f3\u9ad8\u8207\u80fd\uf97e\u4f7f\u7528\uff0c3 \u500b\u70ba pause segments \u6240\u7528)\uff0c\u53ef\u4ee5\u8b93\u6f5b\u5728\u97fb\uf9d8\u7a7a\u9593\u4e2d\u7684\u6587\u4ef6\u5927\u5c0f\u5f9e 112(11*11-9)\u500b\u7dad\ufa01\u6e1b\u5c11\u81f3 30 \u500b\uff0c\u9019\u8868\u793a \uf96f\u6bcf\u4f4d\u8a9e\u8005\u5176 N-gram \u6a21\u578b\u7684\u5e73\u5747\uf96b\uf969\uf97e\u53ef\u5f9e 112 \ufa09\u5230\u50c5\u50c5\u53ea\u6709 34.2 \u7dad\uff0c\u800c\u5176\u6240\u5e36\uf92d\u7684\u597d\u8655\u662f\u5927 \u5e45\u7684\u7c21\u5316\uf9ba\u7cfb\u7d71\u7684\u8907\u96dc\ufa01\uff0c\u5176\u9a57\u8b49\u7d50\u679c\u7684\u76f8\u7b49\u932f\u8aa4\uf961\u53ca\u6c7a\u7b56\u6210\u672c\u51fd\uf969\u5206\u5225\u70ba 22.7%\u548c 0.272\u3002 \u5728\u97fb\uf9d8\u7279\u5fb5\u65b9\u9762\uff0c\u97f3\u9ad8\u8207\u80fd\uf97e\u4e4b\u9ad8\u65af\u6df7\u5408\u6a21\u578b\u53ca\u6f5b\u5728\u97fb\uf9d8\u5206\u6790\u7684\u76f8\u7b49\u932f\u8aa4\uf961\u5206\u5225\u70ba 17.7%\u548c 22.7%\uff0c\u9019\u6a23\u7684\u7d50\u679c\u4ee5\u5f37\u5316\u983b\u8b5c\u7279\u5fb5\u70ba\u4e3b\u7684\u8f14\u52a9\u89d2\ufa01\uf92d\u770b\u5df2\u7d93\u662f\u5f88\uf967\u932f\u7684\uff0c\u800c\u5c07\u9019\uf978\u7a2e\u97fb\uf9d8\u6a21\u578b \u5316\u7684\u65b9\u6cd5\u8207\u983b\u8b5c\u7279\u5fb5\u6700\u4f73\u7d50\u679c\u5408\u4f75\u5f8c\uff0c\u5206\u5225\u80fd\u8b93\u9a57\u8b49\u7cfb\u7d71\u518d\u5f9e\u76f8\u7b49\u932f\u8aa4\uf961 4.0%\u53ca\u6c7a\u7b56\u6210\u672c\u51fd\uf969 0.047 \u6539\u5584\u81f3 3.8%\u8207 0.045 \u4ee5\u53ca 3.8%\u8207 0.050\uff0c\u53ef\ufa0a\u97fb\uf9d8\u7279\u5fb5\u5c0d\u983b\u8b5c\u7279\u5fb5\u7684\u7cfb\u7d71\u78ba\u5be6\u662f\u7522\u751f\u8f14\u52a9 \u7684\u6548\ufa17\u3002\u6b64\u5916\uff0c\u5728\u5716\u4e00\u53ca\u5716\u4e5d\u7684\u7d50\u679c\u53ef\u770b\u5230\u6587\u5b57\uf967\u7279\u5b9a\u7684\u8a9e\u8005\u9a57\u8b49\u4e2d\uff0c\u5c0d\u65bc\u97fb\uf9d8\u7279\u5fb5\u7684\u4f7f\u7528\u4e26\u672a \u5716\u516b\u3001\u6587\u5b57\uf967\u7279\u5b9a\u8a9e\u8005\u9a57\u8b49\u7cfb\u7d71\uff0c\u5728\u983b\u8b5c\u7279\u5fb5\u4e0a\u4f7f\u7528\uf967\u540c\u524d\u5f8c\u7aef\u8655\uf9e4\u65b9\u5f0f\u7684 DET \u66f2\u7dda\u5716\u3002 \u5716\u4e5d\u3001\u5305\u62ec 5 \u7a2e\uf967\u540c\u6587\u5b57\uf967\u7279\u5b9a\u8a9e\u8005\u9a57\u8b49\u7cfb\u7d71\u4e4b DET \u66f2\u7dda\u5716\u3002 0.038\uff0c\u548c\u96b1\u85cf\u5f0f\u99ac\u53ef\u592b\u6a21\u578b 2.9%\u548c 0.034 \u7684\u8868\u73fe\u537b\u662f\u5dee\uf967\u591a\u3002\u5716\u5341\u4ea6\u53ef\u770b\ufa0a\u6709\u8da3\u7684\u7d50\u679c\u662f\u7576\u9019 \uf978\u500b\u7cfb\u7d71\u7d50\u5408\u5f8c\u6703\u5c0d\u7d50\u679c\u7522\u751f\u5f37\u52c1\u7684\u6539\u5584\uff0c\u76f8\u7b49\u932f\u8aa4\uf961\u53ca\u6c7a\u7b56\u6210\u672c\u51fd\uf969\u5206\u5225\u70ba 1.9%\u548c 0.023\uff0c\u800c \u9019\u6216\u8a31\u5c31\u662f\u5c0d\u9019\uf978\u7a2e\u7cfb\u7d71\u7684\u4e92\u88dc\u6027\u505a\uf9ba\u6700\u4f73\u7684\u9a57\u8b49\uff0c\u56e0\u70ba\u9ad8\u65af\u6df7\u5408\u6a21\u578b\u50c5\u9700\u5c11\uf97e\u8a9e\uf9be\uf965\u80fd\u5c0d\u6bcf\u500b \u97f3\u6846\u7684\u5012\u983b\u8b5c\u4fc2\uf969\u5206\u4f48\u4f5c\u6a21\u578b\u5316\uff0c\u53cd\u89c0\u96b1\u85cf\u5f0f\u99ac\u53ef\u592b\u6a21\u578b\u591a\uf97e\u9700\u6c42\u624d\u80fd\u4ed4\u7d30\u63cf\u7e6a\u51fa\u5012\u983b\u8b5c\u4fc2\uf969\u7684 \u66ab\u614b\u8ecc\u8de1\uff0c\u53ef\ufa0a\uf978\u8005\u6240\u9577\uf967\u540c\u65bc\u8a9e\uf9be\uf97e\u6240\u4f9b\u61c9\u7684\u5927\u5c0f\u3002 \u6587\u5b57\uf967\u7279\u5b9a\u7684\u4efb\u52d9\uff0c\u6211\u5011\u53ea\u904b\u7528\uf9ba\u97f3\u9ad8\u53ca\u80fd\uf97e\u7684 8 \u6df7\u5408\uf969\u9ad8\u65af\u6df7\u5408\u6a21\u578b\u3002\u5716\u5341\u4e00\u986f\u793a\u97fb\uf9d8\u7279\u5fb5\u548c \u4e0a\u9762\uf978\u5957\u7cfb\u7d71\u7684\u5408\u4f75\u7d50\u679c\uff0c\u8207\u6587\u5b57\u9650\u5b9a\u8a9e\u8005\u9ad8\u65af\u6df7\u5408\u6a21\u578b\u7684\u7d50\u5408\u662f\u76f8\u7b49\u932f\u8aa4\uf961\u53ca\u6c7a\u7b56\u6210\u672c\u51fd\uf969\u5206 \u5225\u70ba 2.9%\u548c 0.034\uff0c\u800c\u96b1\u85cf\u5f0f\u99ac\u53ef\u592b\u6a21\u578b\u5247\u662f 3.1%\u548c 0.034\uff0c\u96d6\u7136\u7531\u6b64\u770b\u5230\u6548\u679c\u6c92\u80fd\u6709\u986f\u8457\u7684\u6539 \u5584\uff0c\uf967\u904e\u6709\u8da3\u7684\u5728\u65bc\u6211\u5011\u5c07\u6240\u6709\u7684\u65b9\u6cd5\u7d50\u5408\u5f8c\uff0c\u6574\u500b\u7cfb\u7d71\u7684\u932f\u8aa4\uf961\u53c8\u518d\u4e0b\ufa09\u4e9b\u8a31\uff0c\u5982\u5716\u5341\u4e00\u6240\u793a\uff0c \u800c\u9019\u4f3c\u4e4e\u53c8\u8b49\u5be6\uf9ba\u97fb\uf9d8\u7279\u5fb5\u5c0d\u983b\u8b5c\u7279\u5fb5\u7684\u4e92\u88dc\u5728\u6587\u5b57\u7279\u5b9a\u7684\u7cfb\u7d71\u4e0a\u4ecd\u820a\u662f\u5f88\u6709\u6548\u679c\u7684\u3002 \u53e6\u5916\u5728\u5716\u4e8c\u53ca\u5716\u5341\u7684\u7d50\u679c\uf967\u96e3\u767c\u73fe\uff0c\u6587\u5b57\u7279\u5b9a\u8a9e\u8005\u9a57\u8b49\u7cfb\u7d71\u4e26\u672a\u5448\u73fe\u6539\uf97c\u5f0f\u5206\uf969\u6b63\u898f\u5316\u7684\u7d50 \u679c\uff0c\u9019\u662f\u56e0\u70ba\u6539\uf97c\u5f0f\u5206\uf969\u6b63\u898f\u5316\u65b9\u6cd5\u4e2d\uff0c\uf967\uf941\u6211\u5011\u4f7f\u7528\u591a\u5c11\u76f8\u4f3c\u8a9e\u8005\u7684\uf969\uf97e\uff0c\ufa26\u672a\u80fd\u8b93\u7cfb\u7d71\u6709\u6240 \u6539\u5584\u3002\u5982\u6b64\u7684\u7d50\u679c\u548c\u6587\u5b57\uf967\u7279\u5b9a\u8a9e\u8005\u9a57\u8b49\u96d6\uf967\u4e00\u81f4\uff0c\u4f46\u7531\u65bc\u6587\u5b57\u7279\u5b9a\u8a9e\u8005\u9a57\u8b49\u7684\u6e2c\u8a66\u8a9e\uf9be\u901a\u9053\u7279 \u6027\ufa26\u662f\u5728\u8a13\uf996\u968e\u6bb5\u770b\u904e\u7684\uff0c\u56e0\u6b64\u6539\uf97c\u5f0f\u5206\uf969\u6b63\u898f\u5316\u7121\u6cd5\u5728\u6587\u5b57\u7279\u5b9a\u8a9e\u8005\u9a57\u8b49\u4e0a\u6709\u6240\u8ca2\u737b\uff0c\u53ef\u80fd\u5c31 \u662f\u6e2c\u8a66\u8a9e\uf9be\u8207\u8a13\uf996\u6642\u901a\u9053\u4e00\u81f4\u7684\u95dc\u4fc2\uff0c\u6545\u91dd\u5c0d\u6b64\u8a9e\uf9be\u5eab\u7684\u6587\u5b57\u7279\u5b9a\u8a9e\u8005\u9a57\u8b49\u4e26\uf967\u9069\u5408\u4f7f\u7528\u6539\uf97c\u5f0f \u5206\uf969\u6b63\u898f\u5316\u3002 \u5716\u5341\u4e00\u3001\u5305\u62ec 6 \u7a2e\uf967\u540c\u6587\u5b57\u7279\u5b9a\u8a9e\u8005\u9a57\u8b49\u7cfb\u7d71\u4e4b DET \u66f2\u7dda\u5716\u3002 5.3.3 \u6587\u5b57\uf967\u7279\u5b9a\u8207\u6587\u5b57\u7279\u5b9a\u4e4b\u8a9e\u8005\u9a57\u8b49\u7d50\u679c\u6bd4\u8f03 \u5728\u6b64\u6211\u5011\u5c07\u878d\u5408\u6587\u5b57\uf967\u7279\u5b9a\u8207\u6587\u5b57\u7279\u5b9a\u8a9e\u8005\u9a57\u8b49\u7cfb\u7d71\u4e2d\u6240\u6709\u7684\u6a21\u7d44\uff0c\u56e0\u70ba\uf967\u50c5\u53ea\u8003\u616e\u983b\u8b5c\u7279\u5fb5\u8207 \u97fb\uf9d8\u7279\u5fb5\u4e4b\u9593\u7684\u4e92\u88dc\u7279\u6027\uff0c\uf967\u540c\u6a21\u7d44\u4e4b\u9593\u7684\u76f8\u95dc\u6027\u4ea6\u53ef\u70ba\u6709\u6548\u7684\u7279\u5fb5\uff0c\u6240\u4ee5\u518d\u85c9\u7531\u591a\u5c64\u611f\u77e5\u6a5f\uf967 \u540c\u65bc\u4e00\u822c\u7dda\u6027\u7d44\u5408\u7684\u65b9\u5f0f\uff0c\u5c07\u6240\u6709\u6a21\u7d44\u4f5c\u5168\u9762\u6027\u5408\u4f75\uff0c\u7d50\u679c\u5982\u5716\u5341\u4e8c\u6240\u793a\u3002 0.023\uff0c\u5927\u5e45\u6539\u5584\u81f3 1.5%\u8207 0.020\u3002 \u8868\u4e00\u30018 \u7a2e\uf967\u540c\u6587\u5b57\uf967\u7279\u5b9a\u8a9e\u8005\u9a57\u8b49\u7cfb\u7d71\u4e4b\u7d50\u679c\u6bd4\u8f03\u3002 ERR (%) DCF (1) LPA 22.7 0.272 \u932f\u8aa4\u62d2\u7d55\uf961(\u6bd4\uf9b5\u70ba 1 \u6bd4 20\u3002 \u65bc evaluation \u7684\u90e8\u4efd\u5247\u5171\u6709 591 \u4f4d\u8a3b\u518a\u8a9e\u8005\uff0c\u6bcf\u4e00\u4f4d\ufa26\u6709\u76f8\u540c\u5167\u5bb9\u7684\u4e09\u500b\uf906\u5b50\uff0c\u5e73\u5747\u7528\uf92d\u8a3b\u518a\u7684 5.3. \u6587\u5b57\uf967\u7279\u5b9a\u8207\u6587\u5b57\u7279\u5b9a\u8a9e\u8005\u9a57\u8b49\u5be6\u9a57\u7d50\u679c \u7576\u983b\u8b5c\u7279\u5fb5\u70ba\u4e3b\u7684\u6700\u4f73\u7cfb\u7d71\u5efa\uf9f7\u597d\u4e4b\u5f8c\uff0c\u518d\uf92d\u5c31\u662f\u95dc\u65bc\u97fb\uf9d8\u7279\u5fb5\u8207\u983b\u8b5c\u7279\u5fb5\u7684\u7d50\u5408\uff0c\u76f8\u8f03\u65bc (2) Pitch GMM 17.7 0.223
\u5c07\u6703\ufa09\u4f4e\uff0c\u800c\u932f\u8aa4\u63a5\u53d7\uf961\u5247\u6703\u63d0\u9ad8\u3002\u6240\u4ee5\u7cfb\u7d71\u6700\u5f8c\u6548\u80fd\u7684\uf97e\u6e2c\u4e3b\u8981\u662f\u900f\u904e\u76f8\u7b49\u932f\u8aa4\uf961(equal error \u4e09\u7a2e\u65b9\u5f0f\uff0c\u5f9e\u5716\u5341\u7684\u7d50\u679c\u53ef\u77e5 MVA \u7684\u9a57\u8b49\u7d50\u679c\uf976\u52dd CMN\uff0c\u7136\u800c\u6211\u5011\u767c\u73fe\u5728\u6587\u5b57\u7279\u5b9a\u4efb\u52d9\uf9e8\uff0c \u5716\u5341\u4e8c\u3001\u7d50\u5408\u7cfb\u7d71\u6240\u6709\u6a21\u7d44\u7684 DET \u66f2\u7dda\u5716\u3002
\u5373\u53ef\u5b8c\u6210\u9a57\u8b49\u7684\u4efb\u52d9\u3002 rate, EER)\u53ca\u6c7a\u7b56\u6210\u672c\u51fd\uf969(decision cost function, DCF)\uf92d\u8861\uf97e\u3002 \u5716\u4e94\u3001\u97fb\uf9d8\u7279\u5fb5\u7a7a\u9593\ufa09\u7dad\u3002 \u5716\u4e03\u3001\u6539\uf97c\u5f0f\u6e2c\u8a66\u5206\uf969\u6b63\u898f\u5316\u65b9\u584a\u5716\u3002 MV \u7684\u8868\u73fe\uf901\u512a\u65bc MVA\uff0c\u9019\u53ef\u80fd\u662f\u56e0\u70ba\u5728\u6587\u5b57\u7279\u5b9a\u7684\u8a9e\uf9be\u5eab\u662f\u7531\u9ea5\u514b\u98a8\uf93f\u88fd\u800c\u6210\uff0c\u4e14\u6240\u6709\u6e2c\u8a66 \u5716\u5341\u3001\u6587\u5b57\u7279\u5b9a\u8a9e\u8005\u9a57\u8b49\u7cfb\u7d71\uff0c\u5728\u983b\u8b5c\u7279\u5fb5\u4e0a\u4f7f\u7528\uf967\u540c\u8655\uf9e4\u65b9\u5f0f\u7684 DET \u66f2\u7dda\u5716\u3002
" } } } }