| { |
| "paper_id": "O09-1001", |
| "header": { |
| "generated_with": "S2ORC 1.0.0", |
| "date_generated": "2023-01-19T08:10:54.243963Z" |
| }, |
| "title": "Likelihood Ratio Based Discriminant Analysis for Large Vocabulary Continuous Speech Recognition", |
| "authors": [ |
| { |
| "first": "Hung-Shin", |
| "middle": [], |
| "last": "\u674e\u9d3b\u6b23", |
| "suffix": "", |
| "affiliation": { |
| "laboratory": "", |
| "institution": "National Taiwan Normal University", |
| "location": {} |
| }, |
| "email": "hungshin@live.com" |
| }, |
| { |
| "first": "", |
| "middle": [], |
| "last": "Lee", |
| "suffix": "", |
| "affiliation": { |
| "laboratory": "", |
| "institution": "National Taiwan Normal University", |
| "location": {} |
| }, |
| "email": "" |
| }, |
| { |
| "first": "Berlin", |
| "middle": [], |
| "last": "\u9673\u67cf\u7433", |
| "suffix": "", |
| "affiliation": { |
| "laboratory": "", |
| "institution": "Normal University", |
| "location": { |
| "country": "Taiwan" |
| } |
| }, |
| "email": "berlin@ntnu.edu.tw" |
| }, |
| { |
| "first": "", |
| "middle": [], |
| "last": "Chen", |
| "suffix": "", |
| "affiliation": { |
| "laboratory": "", |
| "institution": "Normal University", |
| "location": { |
| "country": "Taiwan" |
| } |
| }, |
| "email": "" |
| } |
| ], |
| "year": "", |
| "venue": null, |
| "identifiers": {}, |
| "abstract": "", |
| "pdf_parse": { |
| "paper_id": "O09-1001", |
| "_pdf_hash": "", |
| "abstract": [], |
| "body_text": [ |
| { |
| "text": "\u6240\u9650\u5236\u7684\u53c3\u6578 \u5b50\u7a7a\u9593\uff0c\u5247\u76f8\u4f3c\u5ea6\u6bd4\u7387\u6aa2\u5b9a\u91dd\u5c0d\u865b\u7121\u5047\u8a2d H 0 \u548c\u5c0d\u7acb\u5047\u8a2d H 1 \u4e4b\u9593\u7684\u6a19\u6e96\u70ba \u2126 L L LR sup sup \u03c9 =", |
| "cite_spans": [], |
| "ref_spans": [], |
| "eq_spans": [], |
| "section": "", |
| "sec_num": null |
| }, |
| { |
| "text": "(1) ", |
| "cite_spans": [], |
| "ref_spans": [], |
| "eq_spans": [], |
| "section": "", |
| "sec_num": null |
| }, |
| { |
| "text": "\u5176\u4e2d\uff0c L \u8868\u793a\u8a13\u7df4\u6a23\u672c(sample)\u8cc7\u6599\u7684\u76f8\u4f3c\u5ea6\uff0c S L sup \u5247\u8868\u793a\u4ee5 S \u70ba\u53c3\u6578\u5b50\u7a7a\u9593\u6642\u7684\u6700 \u5927\u76f8\u4f3c\u5ea6\u3002\u7531\u5f0f(1)\u53ef\u770b\u51fa\uff0c\u76f8\u4f3c\u5ea6\u6bd4\u7387\u6aa2\u5b9a\u662f\u7531\u5169\u500b\u90e8\u5206\u7d44\u6210\uff1a\u6700\u5927\u76f8\u4f3c\u5ea6\u8207\u6bd4\u7387\u3002 \u4f7f\u7528\u6700\u5927\u5316\u76f8\u4f3c\u5ea6\u4f30\u8a08\u6cd5(maximum likelihood estimation, MLE)\u7684\u7528\u610f\u5728\u65bc\u627e\u51fa\u6700\u9069\u5408 \u5169\u500b\u7d71\u8a08\u5047\u8a2d\u6216\u6700\u5177\u4ee3\u8868\u6027\u7684\u53c3\u6578\u4f30\u8a08\u91cf\u3002\u800c\u76f8\u4f3c\u5ea6\u6bd4\u7387\u5176\u80cc\u5f8c\u7684\u908f\u8f2f\u5247\u5728\u65bc\uff0c\u82e5\u6211\u5011 \u4e0d\u8003\u616e\u4efb\u4f55\u4fe1\u5fc3\u5ea6\u91cf\u6e2c(confidence measure)\u4e14\u865b\u7121\u5047\u8a2d H 0 \u7d55\u5c0d\u70ba\u771f(true)\uff0c\u5247\u5728\u5b8c\u6574\u53c3 \u6578\u7a7a\u9593 \u2126 \u4e2d\u7684\u6700\u5927\u76f8\u4f3c\u5ea6\u53c3\u6578\u4f30\u6e2c\u5fc5\u5b9a\u767c\u751f\u5728\u53c3\u6578\u5b50\u7a7a\u9593\u70ba\u03c9 \u7684\u60c5\u6cc1\uff1b\u56e0\u6b64\uff0c \u03c9 L sup \u8207 \u2126 L sup \u5fc5\u5b9a\u6703\u975e\u5e38\u63a5\u8fd1\uff0c\u4f7f LR \u8da8\u8fd1\u65bc 1\u3002\u53cd\u4e4b\uff0c\u82e5 H 0 \u7d55\u5c0d\u70ba\u5047(false)\uff0c\u5247\u6700\u5927\u76f8\u4f3c\u5ea6 \u767c\u751f\u7684\u53c3\u6578\u7a7a\u9593\u5fc5\u5b9a\u4e0d\u662f\u03c9 \uff1b\u56e0\u6b64\uff0c \u03c9 L sup \u5c07\u6703\u9060\u5c0f\u65bc \u2126 L sup \u3002 (\u4e8c)", |
| "cite_spans": [], |
| "ref_spans": [], |
| "eq_spans": [], |
| "section": "", |
| "sec_num": null |
| }, |
| { |
| "text": ": 1 0 H H \u56e0\u6b64\uff0c\u6211\u5011\u6240\u627e\u5230\u7684\u5b50\u7a7a\u9593 d n\u00d7 \u211c \u2208 \u0398 \uff0c\u5fc5\u9808\u76e1\u53ef\u80fd\u5730\u63a8\u7ffb\u4e0d\u5177\u9451\u5225\u6027\u7684\u865b\u7121\u5047\u8a2d H 0 \uff0c \u4e5f \u5c31 \u662f \u4f7f \u5176 \u76f8 \u4f3c \u5ea6 \u6700 \u5c0f \u3002 \u666e \u904d \u5316 \u76f8 \u4f3c \u5ea6 \u6bd4 \u7387 \u9451 \u5225 \u5206 \u6790 (generalized likelihood ratio discriminant analysis, GLRDA)\u76ee\u6a19\u51fd\u5f0f\u4fbf\u53ef\u5beb\u6210 ) ( sup ) ( sup ) ( ) ( GLRDA GLRDA \u0398 \u0398 \u0398 \u0398 \u5b8c\u6574\u7684\u53c3\u6578\u7a7a\u9593 \u53c3\u6578\u5b50\u7a7a\u9593 \u6240\u6709\u985e\u5225\u6bcd\u9ad4\u5747\u76f8\u540c\u7684 L L LR J = = (2) \u8f49\u63db\u77e9\u9663 \u0398 \u4fbf\u53ef\u85c9\u7531\u6700\u5c0f\u5316 ) ( GLRDA \u0398 J \u6c42\u5f97\u3002 (\u4e09)\u540c\u65b9\u5dee\u6027(Homoscedasticity) \u4e00\u822c\u4f86\u8aaa\uff0c\u6211\u5011\u6703\u4ee5\u985e\u5225\u6bcd\u9ad4\u4e4b\u671f\u671b\u503c\u5411\u91cf\u7684\u4f30\u8a08\u91cf\u6240\u5f62\u6210\u7684\u7a7a\u9593\u4f5c\u70ba\u5224\u65b7\u6240\u6709\u985e \u5225\u6bcd\u9ad4\u662f\u5426\u76f8\u540c\u7684\u53c3\u6578\u7a7a\u9593\u3002\u82e5\u6240\u6709\u985e\u5225\u6bcd\u9ad4\u5177\u540c\u65b9\u5dee\u6027(homoscedasticity)\uff0c\u4e5f\u5c31\u662f\u6bcf \u4e00\u985e\u5225\u6bcd\u9ad4\u5177\u6709\u76f8\u540c\u7684\u5171\u8b8a\u7570\u77e9\u9663\uff0c\u5247\u4ee4 i \u03bc \u70ba\u6bcf\u4e00\u985e\u5225\u6bcd\u9ad4 i C \u7684\u671f\u671b\u503c\u5411\u91cf\uff0c i \u03a3 \u70ba\u6bcf \u4e00\u985e\u5225\u6bcd\u9ad4 i C \u7684\u5171\u8b8a\u7570\u77e9\u9663\uff0c homo 0 H \u548c homo 1 H \u53ef\u8a2d\u5b9a\u70ba\uff1a \uf0ee \uf0ed \uf0ec = = = \u3002 \uff0c \uff0c\u4e14 \u6bcf\u4e00\u985e\u5225 : \u3002 \uff0c \uff0c \u6bcf\u4e00\u985e\u5225 : homo 1 homo 0 \u4e0d\u53d7\u4efb\u4f55\u9650\u5236 \u5c0d\u65bc \u5c0d\u65bc i i i i i i C H C H \u03bc \u03a3 \u03a3 \u03bc \u03bc \u03a3 \u03a3 homo 0 H \u4ee3\u8868\u4e86\u4e00\u7a2e\u6975\u7aef\u7684\u60c5\u6cc1\uff0c\u82e5\u5b83\u70ba\u771f\uff0c\u5247\u6240\u6709\u985e\u5225\u6bcd\u9ad4\u5e7e\u8fd1\u5b8c\u5168\u91cd\u758a\uff0c\u4e5f\u5c31\u6c92\u6709\u4efb \u4f55\u9451\u5225\u6027\u3002\u56e0\u6b64\uff0c\u666e\u904d\u5316\u76f8\u4f3c\u5ea6\u6bd4\u7387\u9451\u5225\u5206\u6790(GLRDA)\u7684\u4efb\u52d9\u5373\u5728 homo 0 H \u6700\u4e0d\u53ef\u80fd\u70ba\u771f \u7684\u60c5\u6cc1\u4e0b\uff0c\u627e\u51fa\u6700\u5408\u9069\u7684\u6295\u5f71\u5b50\u7a7a\u9593\u3002 \u4ee5\u4e0b\u7684\u547d\u984c\u8b49\u660e\u4e86\u82e5\u6bcf\u4e00\u985e\u5225\u6bcd\u9ad4\u5747\u9075\u5faa\u9ad8\u65af\u5206\u4f48(Gaussian distribution)\uff0c\u5247\u7dda\u6027 \u9451\u5225\u5206\u6790(LDA)\u8f49\u63db\u77e9\u9663\uff0c\u7b49\u540c\u65bc\u5c07 homo 0 H \u8207 homo 1 H \u7f6e\u65bc\u666e\u904d\u5316\u76f8\u4f3c\u5ea6\u6bd4\u7387\u9451\u5225\u5206\u6790 (GLRDA)\u7684\u6846\u67b6\u4e0b\u6c42\u89e3\u3002 \u547d\u984c\u4e00\uff1a\u82e5\u6bcf\u4e00\u985e\u5225\u6bcd\u9ad4 i C \u90fd\u5177\u6709\u9ad8\u65af\u5206\u4f48\uff0c\u5247\u6700\u5c0f\u5316\u666e\u904d\u5316\u76f8\u4f3c\u5ea6\u6bd4\u7387\u9451\u5225\u5206\u6790 (GLRDA)\u7684\u76ee\u6a19\u51fd\u5f0f ) ( sup ) ( sup ) ( homo 1 homo 0 homo GLRDA \u0398 \u0398 \u0398 H H L L J = (3) \u7b49\u540c\u65bc\u6700\u5927\u5316\u7dda\u6027\u9451\u5225\u5206\u6790(LDA)\u7684\u76ee\u6a19\u51fd\u5f0f | | | | ) ( LDA \u0398 S \u0398 \u0398 S \u0398 \u0398 W T B T J = (4) \u5176\u4e2d\uff0c n n B \u00d7 \u211c \u2208 S \u548c n n W \u00d7 \u211c \u2208 S", |
| "cite_spans": [], |
| "ref_spans": [], |
| "eq_spans": [], |
| "section": "", |
| "sec_num": null |
| }, |
| { |
| "text": "\u5206\u5225\u4ee3\u8868\u985e\u5225\u9593\u6563\u4f48\u77e9\u9663(between-class scatter matrix)\u8207\u985e \u5225\u5167\u6563\u4f48\u77e9\u9663(within-class scatter matrix) [ ", |
| "cite_spans": [ |
| { |
| "start": 80, |
| "end": 81, |
| "text": "[", |
| "ref_id": null |
| } |
| ], |
| "ref_spans": [], |
| "eq_spans": [], |
| "section": "", |
| "sec_num": null |
| }, |
| { |
| "text": "\u0398 \u0398 \u0398 H H L L J \u2212 = (5) \u800c ) ( log homo 0 \u0398 H L \u548c ) ( log homo 1 \u0398 H L \u53ef\u88ab\u9032\u4e00\u6b65\u5206\u5225\u8868\u793a\u70ba\u6a23\u672c\u6240\u6709\u8cc7\u6599 N 1 x \u5728\u6240\u5c6c\u65bc\u9ad8\u65af\u5206 \u4f48\u4e4b\u985e\u5225\u6bcd\u9ad4\u4e0b\u7684\u76f8\u4f3c\u5ea6\uff1a ( ) \uf0e5 = \u2212 \u2212 + + \u2212 \u2212 \u2212 = = C i i i i T i i N H n d N g p L | | log ) trace( ) ( ) ( 2 ) , ( ) , , , ( log ) ( log 1 1 1 homo 0 \u03a3 S \u03a3 \u03bc m \u03a3 \u03bc m \u0398 \u03a3 \u03bc x \u0398 (6) ( ) \uf0e5 = \u2212 \u2212 + + \u2212 \u2212 \u2212 = = C i i i i i T i i i i N H n d N g p L | | log ) trace( ) ( ) ( 2 ) , ( ) , }, { , ( log ) ( log 1 1 1 homo 1 \u03a3 S \u03a3 \u03bc m \u03a3 \u03bc m \u0398 \u03a3 \u03bc x \u0398 (7) \u5176\u4e2d\uff0c ) 2 log( ) 2 ( ) , ( \u03c0 Nd d N g \u2212 = \uff0cN \u70ba\u6a23\u672c\u6240\u6709\u8cc7\u6599\u7e3d\u6578\uff0c i n \u70ba\u985e\u5225 i C \u7684\u8cc7\u6599\u6578\uff0cC \u70ba\u985e\u5225\u7e3d\u6578\uff0cd \u70ba\u6295\u5f71\u5f8c\u4e4b\u7279\u5fb5\u5b50\u7a7a\u9593\u7684\u7dad\u5ea6(\u6216\u7279\u5fb5\u6578) \uff1b i m \u548c i S \u5206\u5225\u70ba\u7d93\u904e \u0398 \u8f49\u63db \u5f8c\u7684\u6a23\u672c\u671f\u671b\u503c\u5411\u91cf\u8207\u5171\u8b8a\u7570\u77e9\u9663\uff0c\u800c \u03bc \u3001 } { i \u03bc \u548c \u03a3 \u5247\u662f\u6839\u64da\u7d71\u8a08\u5047\u8a2d homo 0 H \u8207 homo 1 H \u800c \u8a2d\u5b9a\u4e4b\u7d93\u904e \u0398 \u8f49\u63db\u5f8c\u7684\u6bcd\u9ad4\u671f\u671b\u503c\u5411\u91cf\u8207\u5171\u8b8a\u7570\u77e9\u9663\uff0c\u5373\u6211\u5011\u6240\u8981\u4f30\u8a08\u7684\u53c3\u6578\u3002 \u6b32\u6c42\u5f97\u5728\u5047\u8a2d homo 0 H \u4e0b\u7684\u6700\u5927\u76f8\u4f3c\u5ea6\u4f30\u8a08\u91cf homo 0 \u03bc \u548c homo 0 \u03a3 \uff0c\u53ef\u5c07\u5f0f(6)\u5206\u5225\u5c0d \u03bc \u548c \u03a3 \u504f\u5fae\u5206\uff0c\u4e26\u4ee4\u5176\u70ba 0\uff0c\u53ef\u5f97\uff1a ( ) ( ) m m \u03bc \u03bc m \u03a3 \u03bc \u03a3 S \u03a3 \u03bc m \u03a3 \u03bc m \u03bc \u0398~0 ) (~| | log ) trace( ) ( ) ( 2 ) ( log homo 0 1 1 1 homo 0 = = \uf0de = \u2212 \u2212 = \u2202 \uf0f7 \uf0f8 \uf0f6 \uf0e7 \uf0e8 \uf0e6 + + \u2212 \u2212 \u2212 \u2202 = \u2202 \u2202 \uf0e5 \uf0e5 \uf0e5 = = \u2212 = \u2212 \u2212 C i i i i C i i i i C i i i i T i i H N n n n L (8) ( ) T B W C i i T i i i C i i i i C i i T i i i C i i i i C i i i C i i i i T i i H N n N n n n n n L S S S m m m m S \u03a3 \u03a3 m m \u03a3 m m \u03a3 \u03a3 S \u03a3 \u03a3 \u03a3 \u03a3 S \u03a3 \u03bc m \u03a3 \u03bc m \u03a3 \u0398) )( ( 0 ) ( ) (~| | log ) trace( ) ( ) ( 2 ) ( log homo 0 1 1 1 1 1 1 1 homo 0 1 homo 0 homo 0 = + = \uf0f7 \uf0f8 \uf0f6 \uf0e7 \uf0e8 \uf0e6 \u2212 \u2212 + \uf0f7 \uf0f8 \uf0f6 \uf0e7 \uf0e8 \uf0e6 = \uf0de = \u2212 \u2212 + + \u2212 = \u2202 \uf0f7 \uf0f8 \uf0f6 \uf0e7 \uf0e8 \uf0e6 + + \u2212 \u2212 \u2212 \u2202 = \u2202 \u2202 \uf0e5", |
| "cite_spans": [], |
| "ref_spans": [], |
| "eq_spans": [], |
| "section": "", |
| "sec_num": null |
| }, |
| { |
| "text": "EQUATION", |
| "cite_spans": [], |
| "ref_spans": [], |
| "eq_spans": [ |
| { |
| "start": 0, |
| "end": 8, |
| "text": "EQUATION", |
| "ref_id": "EQREF", |
| "raw_str": "2 | | log 2 ) , ( ) , , , ( log max ) ( log sup homo 0 homo 0 1 homo 0 Nd N d N g p L T N H \u2212 \u2212 = = S \u0398 \u03a3 \u03bc x \u0398 (10) \u540c\u7406\uff0c\u6b32\u6c42\u5f97\u5728\u5047\u8a2d homo 1 H \u4e0b\u7684\u6700\u5927\u76f8\u4f3c\u5ea6\u4f30\u8a08\u91cf } { homo , 1 i \u03bc \u548c homo 1 \u03a3 \uff0c\u53ef\u5c07\u5f0f(7)\u5206\u5225 \u5c0d i \u03bc \u548c \u03a3 \u504f\u5fae\u5206\uff0c\u4e26\u4ee4\u5176\u70ba 0\uff0c\u53ef\u5f97\uff1a ( ) ( ) i i i i i i C i i i i i T i i i i H n n L m \u03bc \u03bc m \u03a3 \u03bc \u03a3 S \u03a3 \u03bc m \u03a3 \u03bc m \u03bc \u0398 0 ) (~| | log ) trace( ) ( ) ( 2 ) ( log homo , 1 1 1 1 homo 1 = \uf0de = \u2212 = \u2202 \uf0f7 \uf0f8 \uf0f6 \uf0e7 \uf0e8 \uf0e6 + + \u2212 \u2212 \u2212 \u2202 = \u2202 \u2202 \u2212 = \u2212 \u2212 \uf0e5 (11) ( ) W C i i i i C i i i i C i i i C i i i i i T i i i H N n n n n L S S \u03a3 \u03a3 S \u03a3 \u03a3 \u03a3 \u03a3 S \u03a3 \u03bc m \u03a3 \u03bc m \u03a3 \u0398~0~| | log ) trace( ) ( ) ( 2 ) ( log homo 1 1 1 1 1 homo , 1 1 homo , 1 homo 1 = \uf0f7 \uf0f8 \uf0f6 \uf0e7 \uf0e8 \uf0e6 = \uf0de = + \u2212 = \u2202 \uf0f7 \uf0f8 \uf0f6 \uf0e7 \uf0e8 \uf0e6 + + \u2212 \u2212 \u2212 \u2202 = \u2202 \u2202 \uf0e5 \uf0e5 \uf0e5 \uf0e5 = = \u2212 \u2212 = \u2212 = \u2212 \u2212 \u5c07 i i m \u03bch omo , 1 = \u548c W S \u03a3h omo 1 = \u4ee3\u5165\u5f0f(7)\uff0c\u53ef\u5f97\u5728\u5047\u8a2d homo 1 H \u4e0b\u7684\u6700\u5927\u5c0d\u6578\u76f8\u4f3c\u5ea6\uff1a 2 | | log 2 ) , ( ) , }, { , ( log max ) ( log sup homo 1 homo , 1 1 homo 1 Nd N d N g p L W i N H \u2212 \u2212 = = S \u0398 \u03a3 \u03bc x \u0398", |
| "eq_num": "(13)" |
| } |
| ], |
| "section": "", |
| "sec_num": null |
| }, |
| { |
| "text": "\u6700\u5f8c\uff0c\u5c07\u5f0f(10)\u8207\u5f0f(13)\u4ee3\u5165\u5f0f(5)\uff0c\u53ef\u5f97\u5728\u540c\u65b9\u5dee\u6027\u5047\u8a2d\u4e0b\u7684\u5c0d\u6578\u76f8\u4f3c\u5ea6\u6bd4\u7387\uff1a ", |
| "cite_spans": [], |
| "ref_spans": [], |
| "eq_spans": [], |
| "section": "", |
| "sec_num": null |
| }, |
| { |
| "text": "EQUATION", |
| "cite_spans": [], |
| "ref_spans": [], |
| "eq_spans": [ |
| { |
| "start": 0, |
| "end": 8, |
| "text": "EQUATION", |
| "ref_id": "EQREF", |
| "raw_str": "1 | | | | 1 log 2 | | | | | | log 2 ) | | | log(| | | log ( 2 ) | | log | | (log 2 2 | | log 2 ) , ( 2 | | 2 ) , ( ) ( log homo GLRDA + = + = + \u2212 = \u2212 = \uf0f7 \uf0f8 \uf0f6 \uf0e7 \uf0e8 \uf0e6 \u2212 \u2212 \u2212 \uf0f7 \uf0f8 \uf0f6 \uf0e7 \uf0e8 \uf0e6 \u2212 \u2212 = W B W B W W B W T W W T N N N N Nd N d N g Nd N d N g J S S S S S S S S S S S S \u0398 (14) \u0398 \u53ef\u7d93\u7531\u6700\u5c0f\u5316\u5f0f(14)\u4f86\u6c42\u51fa\u3002\u56e0\u70ba\u5c0d\u6578\u51fd\u6578\u70ba\u55ae\u8abf\u905e\u589e(monotonically increasing)\u51fd \u6578\uff0c\u6240\u4ee5 \u0398 \u4e5f\u53ef\u85c9\u7531\u6700\u5927\u5316\u5f0f(14)\u4e2d\u7684 | | | | \u0398 S \u0398 \u0398 S \u0398 W T B T \u6c42\u5f97\uff0c\u800c\u6b64\u9805\u6b63\u597d\u662f\u7dda\u6027\u9451\u5225 \u5206\u6790(LDA)\u7684\u76ee\u6a19\u51fd\u5f0f(\u5f0f(4)) \u3002 \u25a0 (\u56db)\u7570\u65b9\u5dee\u6027(Heteroscedasticity) \u73fe\u5728\uff0c\u6211\u5011\u8003\u616e\u7570\u65b9\u5dee\u6027\u7684\u7d71\u8a08\u5047\u8a2d[17]\uff1a \uf0ee \uf0ed \uf0ec = \u3002 \u5747\u5448\u9ad8\u65af\u5206\u5e03\uff0c\u4e14 \u6bcf\u4e00\u985e\u5225 : \u3002 \u5747\u5448\u9ad8\u65af\u5206\u5e03\uff0c\u4e14 \u6bcf\u4e00\u985e\u5225 : heter 1 heter 0 \u5747\u4e0d\u53d7\u4efb\u4f55\u9650\u5236 \u8207 \u4e0d\u53d7\u4efb\u4f55\u9650\u5236 \uff0c i i i i i i C H C H \u03a3 \u03bc \u03a3 \u03bc \u03bc \u4eff\u7167\u547d\u984c\u4e00\u7684\u65b9\u5f0f\uff0c\u666e\u904d\u5316\u76f8\u4f3c\u5ea6\u6bd4\u7387\u9451\u5225\u5206\u6790(GLRDA)\u7684\u76ee\u6a19\u51fd\u5f0f\u53ef\u5beb\u6210\uff1a ) ( log sup ) ( log sup ) ( log heter 1 heter 0 heter GLRDA \u0398 \u0398 \u0398 H H L L J \u2212 = (15) \u800c ) ( log heter 0 \u0398 H L \u548c ) ( log heter 1 \u0398 H L \u53ef\u88ab\u5206\u5225\u9032\u4e00\u6b65\u8868\u793a\u70ba (12) ( ) \uf0e5 = \u2212 \u2212 + + \u2212 \u2212 \u2212 = = C i i i i i i i T i i i N H n d N g p L | | log ) trace( ) ( ) ( 2 ) , ( ) }, { , , ( log ) ( log 1 1 1 heter 0 \u03a3 S \u03a3 \u03bc m \u03a3 \u03bc m \u0398 \u03a3 \u03bc x \u0398 (16) ( ) \uf0e5 = \u2212 \u2212 + + \u2212 \u2212 \u2212 = = C i i i i i i i i T i i i i i N H n d N g p L | | log ) trace( ) ( ) ( 2 ) , ( ) }, { }, { , ( log ) ( log 1 1 1 heter 1 \u03a3 S \u03a3 \u03bc m \u03a3 \u03bc m \u0398 \u03a3 \u03bc x \u0398 (17) \u6b32\u6c42\u5f97\u5728\u5047\u8a2d heter 0 H \u4e0b\u7684\u6700\u5927\u76f8\u4f3c\u5ea6\u4f30\u8a08\u91cf heter 0 \u03bc \u548c heter , 0 i \u03a3 \uff0c\u53ef\u5c07\u5f0f(16)\u5206\u5225\u5c0d \u03bc \u548c i \u03a3 \u504f\u5fae\u5206\uff0c\u4e26\u4ee4\u5176\u70ba 0\uff0c\u53ef\u5f97\uff1a ( ) ( ) \uf0e5 \uf0e5 \uf0e5 \uf0e5 \uf0e5 \uf0e5 = \u2212 \u2212 = \u2212 = \u2212 \u2212 = \u2212 = \u2212 = \u2212 \u2212 \uf0f7 \uf0f8 \uf0f6 \uf0e7 \uf0e8 \uf0e6 \u2248 \uf0f7 \uf0f8 \uf0f6 \uf0e7 \uf0e8 \uf0e6 = \uf0de = \u2212 \u2212 = \u2202 \uf0f7 \uf0f8 \uf0f6 \uf0e7 \uf0e8 \uf0e6 + + \u2212 \u2212 \u2212 \u2202 = \u2202 \u2202 C i i i i i C i i i i C i i i i i C i i i i C i i i i i C i i i i i i i T i i H n n n n n n L m S S m \u03a3 \u03a3 \u03bc \u03bc m \u03a3 \u03bc \u03a3 S \u03a3 \u03bc m \u03a3 \u03bc m \u03bc \u03980 ) (~| | log ) trace( ) ( ) ( 2 ) ( log 1 1 1 1 1 1 heter 0 1 1 1 heter 0 (18) ( ) ( ) i i i T i i i i i i i i T i i i k i C i i i i i i i T i i i H n n L S B S \u03bc m \u03bc m \u03a3 \u03a3 \u03a3 S \u03a3 \u03a3 \u03bc m \u03bc m \u03a3 \u03a3 \u03a3 S \u03a3 \u03bc m \u03a3 \u03bc m \u03a3 \u0398) )( ( 0 ) )( ( 2 1~| | log ) trace( ) ( )", |
| "eq_num": "(" |
| } |
| ], |
| "section": "", |
| "sec_num": null |
| }, |
| { |
| "text": "EQUATION", |
| "cite_spans": [], |
| "ref_spans": [], |
| "eq_spans": [ |
| { |
| "start": 0, |
| "end": 8, |
| "text": "EQUATION", |
| "ref_id": "EQREF", |
| "raw_str": "= = C i i i i i N H n Nd d N g p L 1 heter , 0 heter 0 1 | | log 2 2 ) , ( ) }, { , , ( log max ) ( log sup heter 0 S B \u0398 \u03a3 \u03bc x \u0398 (20) \u6b32\u6c42\u5f97\u5728\u5047\u8a2d heter 1 H \u4e0b\u7684\u6700\u5927\u76f8\u4f3c\u5ea6\u4f30\u8a08\u91cf heter 1 \u03bc \u548c heter , 1 i \u03a3 \uff0c\u53ef\u5c07\u5f0f(17)\u5206\u5225\u5c0d i \u03bc \u548c i \u03a3 \u504f\u5fae\u5206\uff0c\u4e26\u4ee4\u5176\u70ba 0\uff0c\u53ef\u5f97\uff1a ( ) i i i i i i i C i i i i i i i i T i i i i H n n L m \u03bc \u03bc m \u03a3 \u03bc \u03a3 S \u03a3 \u03bc m \u03a3 \u03bc m \u03bc \u0398 0 ) (~| | log ) trace( ) ( ) ( 2 ) ( log heter ,", |
| "eq_num": "1 1 1" |
| } |
| ], |
| "section": "", |
| "sec_num": null |
| }, |
| { |
| "text": "EQUATION", |
| "cite_spans": [], |
| "ref_spans": [], |
| "eq_spans": [ |
| { |
| "start": 0, |
| "end": 8, |
| "text": "EQUATION", |
| "ref_id": "EQREF", |
| "raw_str": "1 heter 1 = \uf0de = \u2212 = \u2202 \uf0f7 \uf0f8 \uf0f6 \uf0e7 \uf0e8 \uf0e6 + + \u2212 \u2212 \u2212 \u2202 = \u2202 \u2202 \u2212 = \u2212 \u2212 \uf0e5 (21) (19) ( ) ( ) i i i i i i i i C i i i i i i i i T i i i i H n n L S \u03a3 \u03a3 \u03a3 S \u03a3 \u03a3 \u03a3 S \u03a3 \u03bc m \u03a3 \u03bc m \u03a3 \u0398 0 2~| | log ) trace( ) ( )", |
| "eq_num": "(" |
| } |
| ], |
| "section": "", |
| "sec_num": null |
| }, |
| { |
| "text": "= = C i i i C i i i C i i i i i i i T i i i i N H n Nd d N g d n d N g n d N g p L", |
| "cite_spans": [], |
| "ref_spans": [], |
| "eq_spans": [], |
| "section": "", |
| "sec_num": null |
| }, |
| { |
| "text": "\u2212 = \uf0f7 \uf0f8 \uf0f6 \uf0e7 \uf0e8 \uf0e6 \u2212 \u2212 \u2212 \uf0f7 \uf0f8 \uf0f6 \uf0e7 \uf0e8 \uf0e6 + \u2212 \u2212 = C i i i T i i C i T i i i p p i C i i i p p i C i i i i i C i i i C i i i i n n n n n Nd d N g n Nd d N", |
| "cite_spans": [], |
| "ref_spans": [], |
| "eq_spans": [], |
| "section": "", |
| "sec_num": null |
| }, |
| { |
| "text": "EQUATION", |
| "cite_spans": [], |
| "ref_spans": [], |
| "eq_spans": [ |
| { |
| "start": 0, |
| "end": 8, |
| "text": "EQUATION", |
| "ref_id": "EQREF", |
| "raw_str": "\uf0e5 = \u2212 \u2212 \u2212 + \u2212 = C i T i T i T T T i T i n G 1 heter 0 1 heter 0 H ) ( ) ( ) ( 1 log 2 ) ( \u03bc \u0398 m \u0398 \u0398 S \u0398 \u03bc \u0398 m \u0398 \u0398 (25) \u70ba\u4e86\u4f7f\u7528\u68af\u5ea6\u4e0b\u964d\u7b49\u905e\u8ff4\u5f0f\u7684\u6700\u4f73\u5316\u6280\u8853\u6c42\u89e3 \u0398 \uff0c\u5f0f(25)\u5c0d \u0398 \u7684\u4e00\u968e\u504f\u5c0e\u6578\u53ef\u5beb\u6210\uff1a \uf0e5 = \u2212 \u2212 \u2212 + + \u2212 \u2212 = \u2202 \u2202 C i i i i i i i i i n G 1 1 1 1 H ) trace( 1) ( ) ( B S S \u0398 B B S \u0398 S \u0398 \u0398 (26) \u5176\u4e2d\uff0c T i i i ) )( ( heter 0 heter 0 \u03bc m \u03bc m B \u2212 \u2212 = \uff0c \u0398 B \u0398 B i T i = \uff0c \u0398 S \u0398 S i T i = \u3002", |
| "eq_num": "(\u4e94)" |
| } |
| ], |
| "section": "", |
| "sec_num": null |
| }, |
| { |
| "text": "EQUATION", |
| "cite_spans": [], |
| "ref_spans": [], |
| "eq_spans": [ |
| { |
| "start": 0, |
| "end": 8, |
| "text": "EQUATION", |
| "ref_id": "EQREF", |
| "raw_str": "| | log | | log 2 | | log 2 ) ( 1 ) ( ) ( HLDA \u0398 \u0398 S \u0398 \u0398 S \u0398 \u0398 N n N J C i d i T d i d n T T d n + \u2212 \u2212 = \uf0e5 = \u2212 \u2212 (27) \u56e0\u70ba\u5728\u6b64\uff0c \u0398 \u70ba\u5168\u79e9\u77e9\u9663(full-rank matrix)\uff0c\u4e14 ] , [ ) ( ) ( d n d n n \u2212 \u00d7 = \u0398 \u0398 \u0398 \uff0c\u6211\u5011\u53ef\u8b49\u660e[18] ) ( ) ( ) ( ) ( ) ( ) ( d T T d T T d n T T d n d n T T d n d T T d T T d n T T d n d T T d T T \u0398 S \u0398 \u0398 S \u0398 \u0398 S \u0398 \u0398 S \u0398 \u0398 S \u0398 \u0398 S \u0398 \u0398 S \u0398 \u0398 S \u0398 \u0398 S \u0398 \u2212 = \uf0de + = \uf0de \u00d7 = \u2212 \u2212 \u2212 \u2212 T T T T N N N N N N N S \u0398 S \u0398 \u0398 \u0398 S \u0398 \u0398 \u2212 = \u2212 \u2212 \u2212 = \u2212", |
| "eq_num": "(" |
| } |
| ], |
| "section": "", |
| "sec_num": null |
| }, |
| { |
| "text": "EQUATION", |
| "cite_spans": [], |
| "ref_spans": [], |
| "eq_spans": [ |
| { |
| "start": 0, |
| "end": 8, |
| "text": "EQUATION", |
| "ref_id": "EQREF", |
| "raw_str": "= = \u03bc \u03bc \u03a3 \u03a3 i i H homo 0 m T S | | log 2 T N S \u2212 \uf0ee \uf0ed \uf0ec = \u7121\u9650\u5236 : homo 1 i i H \u03bc \u03a3 \u03a3 i m W S | | log 2 W N S \u2212 \uf0ee \uf0ed \uf0ec = \u03bc \u03bc \u03a3 i i H \u7121\u9650\u5236 : heter 0 \uf0e5 \uf0e5 = \u2212 \u2212 = \u2212 \uf0f7 \uf0f8 \uf0f6 \uf0e7 \uf0e8 \uf0e6 C i i i i i C i i i i n n m S S1 1 1 i i S B+ \uf0e5 = + \u2212 C i i i i n 1 | | log 2 S B \uf0ee \uf0ed \uf0ec \u7121\u9650\u5236 \u7121\u9650\u5236 : : heter 1 i i H \u03bc \u03a3 i m i S \uf0e5 = \u2212 C i i i n 1 | | log 2 S | | log 2 | | log 2 | | log 2 ) ( ) ( log sup ) ( log sup 1 HLDA homo 0 heter 1 T L d T T d L C i d i T d i N N n J H H S \u0398 S \u0398 \u0398 S \u0398 \u0398 \u0398 \u0398 \u2212 \uf0f7 \uf0f8 \uf0f6 \uf0e7 \uf0e8 \uf0e6 \u2212 \u2212 \u2212 = \uf0e5 = \uf034 \uf034 \uf034 \uf033 \uf034 \uf034 \uf034 \uf032 \uf031 \uf034 \uf034 \uf034 \uf033 \uf034 \uf034 \uf034 \uf032 \uf031 (30) \u5f88\u660e\u986f\u5730\uff0c\u82e5\u4e0d\u8003\u616e\u5e38\u6578 | | log ) 2 ( T N S \u2212 \uff0c\u5247\u5f0f", |
| "eq_num": "(" |
| } |
| ], |
| "section": "", |
| "sec_num": null |
| }, |
| { |
| "text": "\u82e5\u985e\u5225\u914d\u5c0d 1 C \u8207 2 C \u70ba\u6700\u6df7\u6dc6\u4e4b\u914d\u5c0d\uff0c\u5247\u865b\u7121\u5047\u8a2d\u53ef\u8a2d\u5b9a\u70ba 2 1 \u03bc \u03bc = \uff1b\u800c\u985e\u5225\u914d\u5c0d 2 C \u8207 3 C \u70ba \u6b21 \u6df7 \u6dc6 \u4e4b \u914d \u5c0d \uff0c \u5247 \u865b \u7121 \u5047 \u8a2d \u53ef \u589e \u52a0 3 2 \u03bc \u03bc = \u3002 \u56e0 \u6b64 \uff0c \u6574 \u500b \u865b \u7121 \u5047 \u8a2d \u53ef \u5408 \u4f75 \u70ba 3 2 1 \u03bc \u03bc \u03bc = =", |
| "cite_spans": [], |
| "ref_spans": [], |
| "eq_spans": [], |
| "section": "", |
| "sec_num": null |
| }, |
| { |
| "text": "\u3002\u4e5f\u5c31\u662f\u8aaa\uff0c\u6211\u5011\u5fc5\u9808\u5728\u985e\u5225\u914d\u5c0d\u96c6\u5408\u4e2d\u627e\u5230\u6240\u6709\u76f8\u95dc\u7684\u985e\u5225\u914d\u5c0d\u4ee5\u7d44\u6210 \u6df7\u6dc6\u7fa4\u805a(confusable cluster)\uff0c\u5982\u5716\u4e8c\u3002\u82e5\u6211\u5011\u628a\u6240\u6709\u985e\u5225\u8996\u70ba\u5716\u5f62(graph)\u4e2d\u7684\u9ede (vertex)\uff0c\u800c\u7531\u6613\u65bc\u6df7\u6dc6\u4e4b\u985e\u5225\u914d\u5c0d\u6240\u5efa\u7acb\u7684\u95dc\u4fc2\u8996\u70ba\u5169\u9ede\u4e4b\u9593\u7684\u908a(edge)\uff0c\u5247\u6df7\u6dc6\u7fa4\u805a \u7684\u7522\u751f\u53ef\u88ab\u8996\u70ba\u5c0b\u627e\u5716\u5f62(graph)\u4e2d\u6240\u6709\u7684\u9023\u901a\u5b50\u5716(connected subgraph)\u3002\u6240\u4ee5\uff0c\u6211\u5011\u53ef \u4ee5\u4f7f\u7528\u4e00\u4e9b\u5716\u8ad6\u4e2d\u7684\u6f14\u7b97\u6cd5\uff0c\u5982\u6eff\u6c34\u586b\u5145\u6f14\u7b97\u6cd5(flood fill algorithm) [19] \uff0c\u4f86\u89e3\u6c7a\u9019\u500b\u554f \u984c\u3002 \u56e0\u6b64\uff0c\u6211\u5011\u53ef\u4ee5\u5c07\u7570\u65b9\u5dee\u6027\u4e4b\u666e\u904d\u5316\u76f8\u4f3c\u5ea6\u6bd4\u7387\u9451\u5225\u5206\u6790(GLRDA)\u6539\u826f\u6210\u57fa\u65bc\u6df7 \u6dc6 \u8cc7 \u8a0a \u4e4b \u666e \u904d \u5316 \u76f8 \u4f3c \u5ea6 \u6bd4 \u7387 \u9451 \u5225 \u5206 \u6790 (confusion information based GLRDA, CI-GLRDA)\uff1a\u4ee4 ", |
| "cite_spans": [ |
| { |
| "start": 230, |
| "end": 234, |
| "text": "[19]", |
| "ref_id": "BIBREF18" |
| } |
| ], |
| "ref_spans": [], |
| "eq_spans": [], |
| "section": "", |
| "sec_num": null |
| }, |
| { |
| "text": "} { : k G G \u70ba\u6240\u6709\u6839\u64da\u524d K \u7d44\u6613\u65bc\u6df7\u6dc6\u4e4b\u985e\u5225\u914d\u5c0d\uff0c\u4e26\u5229\u7528\u6eff\u6c34\u586b\u5145\u6f14\u7b97 \u6cd5\u6c42\u51fa\u4e4b\u7fa4\u805a\u7684\u96c6\u5408\uff0c\u5247\u5176\u865b\u7121\u5047\u8a2d\u8207\u5c0d\u7acb\u5047\u8a2d\u53ef\u8a2d\u5b9a\u5982\u4e0b\uff1a C 1 C 2 C 4 C 3 C 6 C 7 C 9 C 8 G 1 G 2 G 3 \u5716\u4e8c\u3001\u985e\u5225\u914d\u5c0d\u8207\u7fa4\u805a\u5f62\u6210\u793a\u610f\u5716 \u7df4\u5de5\u5177\u63a1\u7528 SRI", |
| "cite_spans": [], |
| "ref_spans": [], |
| "eq_spans": [], |
| "section": "", |
| "sec_num": null |
| } |
| ], |
| "back_matter": [ |
| { |
| "text": "\uf0ee \uf0ed \uf0ec = \u2208 \u3002 \u5747\u5448\u9ad8\u65af\u5206\u5e03\uff0c\u4e14 \u6bcf\u4e00\u985e\u5225 : ", |
| "cite_spans": [], |
| "ref_spans": [], |
| "eq_spans": [], |
| "section": "annex", |
| "sec_num": null |
| } |
| ], |
| "bib_entries": { |
| "BIBREF0": { |
| "ref_id": "b0", |
| "title": "Pattern Recognition and Neural Networks", |
| "authors": [ |
| { |
| "first": "B", |
| "middle": [ |
| "D" |
| ], |
| "last": "Ripley", |
| "suffix": "" |
| } |
| ], |
| "year": 1996, |
| "venue": "", |
| "volume": "", |
| "issue": "", |
| "pages": "", |
| "other_ids": {}, |
| "num": null, |
| "urls": [], |
| "raw_text": "B. D. Ripley, Pattern Recognition and Neural Networks. New York: Cambridge University Press, 1996.", |
| "links": null |
| }, |
| "BIBREF1": { |
| "ref_id": "b1", |
| "title": "Feature extraction and dimensionality reduction algorithms and their applications in vowel recognition", |
| "authors": [ |
| { |
| "first": "X", |
| "middle": [], |
| "last": "Wang", |
| "suffix": "" |
| }, |
| { |
| "first": "K", |
| "middle": [ |
| "K" |
| ], |
| "last": "Paliwal", |
| "suffix": "" |
| } |
| ], |
| "year": 2003, |
| "venue": "Pattern Recognition", |
| "volume": "36", |
| "issue": "", |
| "pages": "2429--2439", |
| "other_ids": {}, |
| "num": null, |
| "urls": [], |
| "raw_text": "X. Wang and K. K. Paliwal, \"Feature extraction and dimensionality reduction algorithms and their applications in vowel recognition,\" Pattern Recognition, vol. 36, pp. 2429-2439, 2003.", |
| "links": null |
| }, |
| "BIBREF2": { |
| "ref_id": "b2", |
| "title": "fMPE: discriminatively trained features for speech recogntion", |
| "authors": [ |
| { |
| "first": "D", |
| "middle": [], |
| "last": "Povey", |
| "suffix": "" |
| } |
| ], |
| "year": 2005, |
| "venue": "Proc. ICASSP", |
| "volume": "", |
| "issue": "", |
| "pages": "961--964", |
| "other_ids": {}, |
| "num": null, |
| "urls": [], |
| "raw_text": "D. Povey, et al., \"fMPE: discriminatively trained features for speech recogntion,\" in Proc. ICASSP, 2005, pp. 961-964.", |
| "links": null |
| }, |
| "BIBREF3": { |
| "ref_id": "b3", |
| "title": "Dimensionality reduction using MCE-optimized LDA transformation", |
| "authors": [ |
| { |
| "first": "X.-B", |
| "middle": [], |
| "last": "Li", |
| "suffix": "" |
| } |
| ], |
| "year": 2004, |
| "venue": "Proc. ICASSP", |
| "volume": "", |
| "issue": "", |
| "pages": "137--140", |
| "other_ids": {}, |
| "num": null, |
| "urls": [], |
| "raw_text": "X.-B. Li, et al., \"Dimensionality reduction using MCE-optimized LDA transformation,\" in Proc. ICASSP, 2004, pp. 137-140.", |
| "links": null |
| }, |
| "BIBREF4": { |
| "ref_id": "b4", |
| "title": "The statistical utilization of multiple measurements", |
| "authors": [ |
| { |
| "first": "R", |
| "middle": [ |
| "A" |
| ], |
| "last": "Fisher", |
| "suffix": "" |
| } |
| ], |
| "year": 1938, |
| "venue": "Annals of Eugenics", |
| "volume": "8", |
| "issue": "", |
| "pages": "376--386", |
| "other_ids": {}, |
| "num": null, |
| "urls": [], |
| "raw_text": "R. A. Fisher, \"The statistical utilization of multiple measurements,\" Annals of Eugenics, vol. 8, pp. 376-386, 1938.", |
| "links": null |
| }, |
| "BIBREF5": { |
| "ref_id": "b5", |
| "title": "Heteroscedastic discriminant analysis and reduced rank HMMs for improved speech recognition", |
| "authors": [ |
| { |
| "first": "N", |
| "middle": [], |
| "last": "Kumar", |
| "suffix": "" |
| }, |
| { |
| "first": "A", |
| "middle": [ |
| "G" |
| ], |
| "last": "Andreou", |
| "suffix": "" |
| } |
| ], |
| "year": 1998, |
| "venue": "Speech Communication", |
| "volume": "26", |
| "issue": "", |
| "pages": "283--297", |
| "other_ids": {}, |
| "num": null, |
| "urls": [], |
| "raw_text": "N. Kumar and A. G. Andreou, \"Heteroscedastic discriminant analysis and reduced rank HMMs for improved speech recognition,\" Speech Communication, vol. 26, pp. 283-297, 1998.", |
| "links": null |
| }, |
| "BIBREF6": { |
| "ref_id": "b6", |
| "title": "Maximum likelihood discriminant feature spaces", |
| "authors": [ |
| { |
| "first": "G", |
| "middle": [], |
| "last": "Saon", |
| "suffix": "" |
| } |
| ], |
| "year": 2000, |
| "venue": "Proc. ICASSP", |
| "volume": "", |
| "issue": "", |
| "pages": "1129--1132", |
| "other_ids": {}, |
| "num": null, |
| "urls": [], |
| "raw_text": "G. Saon, et al., \"Maximum likelihood discriminant feature spaces,\" in Proc. ICASSP, 2000, pp. 1129-1132.", |
| "links": null |
| }, |
| "BIBREF7": { |
| "ref_id": "b7", |
| "title": "Optimal feature sub-space selection based on discriminant analysis", |
| "authors": [ |
| { |
| "first": "K", |
| "middle": [], |
| "last": "Demuynck", |
| "suffix": "" |
| } |
| ], |
| "year": 1999, |
| "venue": "Proc. Eurospeech", |
| "volume": "", |
| "issue": "", |
| "pages": "1311--1314", |
| "other_ids": {}, |
| "num": null, |
| "urls": [], |
| "raw_text": "K. Demuynck, et al., \"Optimal feature sub-space selection based on discriminant analysis \" in Proc. Eurospeech, 1999, pp. 1311-1314.", |
| "links": null |
| }, |
| "BIBREF8": { |
| "ref_id": "b8", |
| "title": "Linear discriminant feature extraction using weighted classification confusion information", |
| "authors": [ |
| { |
| "first": "H.-S", |
| "middle": [], |
| "last": "Lee", |
| "suffix": "" |
| }, |
| { |
| "first": "B", |
| "middle": [], |
| "last": "Chen", |
| "suffix": "" |
| } |
| ], |
| "year": 2008, |
| "venue": "Proc.Interspeech", |
| "volume": "", |
| "issue": "", |
| "pages": "2254--2257", |
| "other_ids": {}, |
| "num": null, |
| "urls": [], |
| "raw_text": "H.-S. Lee and B. Chen, \"Linear discriminant feature extraction using weighted classification confusion information,\" in Proc.Interspeech, 2008, pp. 2254-2257.", |
| "links": null |
| }, |
| "BIBREF9": { |
| "ref_id": "b9", |
| "title": "Improved linear discriminant analysis considering empirical pairwise classification error rates", |
| "authors": [ |
| { |
| "first": "H.-S", |
| "middle": [], |
| "last": "Lee", |
| "suffix": "" |
| }, |
| { |
| "first": "B", |
| "middle": [], |
| "last": "Chen", |
| "suffix": "" |
| } |
| ], |
| "year": 2008, |
| "venue": "Proc. ISCSLP", |
| "volume": "", |
| "issue": "", |
| "pages": "149--152", |
| "other_ids": {}, |
| "num": null, |
| "urls": [], |
| "raw_text": "H.-S. Lee and B. Chen, \"Improved linear discriminant analysis considering empirical pairwise classification error rates,\" in Proc. ISCSLP, 2008, pp. 149-152.", |
| "links": null |
| }, |
| "BIBREF10": { |
| "ref_id": "b10", |
| "title": "Empirical error rate minimization based linear discriminant analysis", |
| "authors": [ |
| { |
| "first": "H.-S", |
| "middle": [], |
| "last": "Lee", |
| "suffix": "" |
| }, |
| { |
| "first": "B", |
| "middle": [], |
| "last": "Chen", |
| "suffix": "" |
| } |
| ], |
| "year": 2009, |
| "venue": "Proc. ICASSP", |
| "volume": "", |
| "issue": "", |
| "pages": "", |
| "other_ids": {}, |
| "num": null, |
| "urls": [], |
| "raw_text": "H.-S. Lee and B. Chen, \"Empirical error rate minimization based linear discriminant analysis,\" in Proc. ICASSP, 2009.", |
| "links": null |
| }, |
| "BIBREF11": { |
| "ref_id": "b11", |
| "title": "Stereo-based stochastic mapping with discriminative training for noise robust speech recognition", |
| "authors": [ |
| { |
| "first": "X", |
| "middle": [], |
| "last": "Cui", |
| "suffix": "" |
| } |
| ], |
| "year": 2009, |
| "venue": "Proc. ICASSP", |
| "volume": "", |
| "issue": "", |
| "pages": "2933--2936", |
| "other_ids": {}, |
| "num": null, |
| "urls": [], |
| "raw_text": "X. Cui, et al., \"Stereo-based stochastic mapping with discriminative training for noise robust speech recognition,\" in Proc. ICASSP, 2009, pp. 2933-2936.", |
| "links": null |
| }, |
| "BIBREF12": { |
| "ref_id": "b12", |
| "title": "Principles of Multivariate Analysis: A User's Perspective", |
| "authors": [ |
| { |
| "first": "W", |
| "middle": [ |
| "J" |
| ], |
| "last": "Krzanowski", |
| "suffix": "" |
| } |
| ], |
| "year": 1988, |
| "venue": "", |
| "volume": "", |
| "issue": "", |
| "pages": "", |
| "other_ids": {}, |
| "num": null, |
| "urls": [], |
| "raw_text": "W. J. Krzanowski, Principles of Multivariate Analysis: A User's Perspective. New York: Oxford University Press, 1988.", |
| "links": null |
| }, |
| "BIBREF13": { |
| "ref_id": "b13", |
| "title": "Acoustic and phonetic confusions in accented speech recognition", |
| "authors": [ |
| { |
| "first": "Y", |
| "middle": [], |
| "last": "Liu", |
| "suffix": "" |
| }, |
| { |
| "first": "P", |
| "middle": [], |
| "last": "Fung", |
| "suffix": "" |
| } |
| ], |
| "year": 2005, |
| "venue": "Proc. Interspeech", |
| "volume": "", |
| "issue": "", |
| "pages": "3033--3036", |
| "other_ids": {}, |
| "num": null, |
| "urls": [], |
| "raw_text": "Y. Liu and P. Fung, \"Acoustic and phonetic confusions in accented speech recognition,\" in Proc. Interspeech, 2005, pp. 3033-3036.", |
| "links": null |
| }, |
| "BIBREF14": { |
| "ref_id": "b14", |
| "title": "Generalized LRT-based voice activity detector", |
| "authors": [ |
| { |
| "first": "J", |
| "middle": [ |
| "M" |
| ], |
| "last": "G\u00f3rriz", |
| "suffix": "" |
| } |
| ], |
| "year": 2006, |
| "venue": "IEEE Signal Processing Letters", |
| "volume": "13", |
| "issue": "", |
| "pages": "636--639", |
| "other_ids": {}, |
| "num": null, |
| "urls": [], |
| "raw_text": "J. M. G\u00f3rriz, et al., \"Generalized LRT-based voice activity detector,\" IEEE Signal Processing Letters, vol. 13, pp. 636-639, 2006.", |
| "links": null |
| }, |
| "BIBREF15": { |
| "ref_id": "b15", |
| "title": "Introduction to Statistical Pattern Recognition", |
| "authors": [ |
| { |
| "first": "K", |
| "middle": [], |
| "last": "Fukunaga", |
| "suffix": "" |
| } |
| ], |
| "year": 1990, |
| "venue": "", |
| "volume": "", |
| "issue": "", |
| "pages": "", |
| "other_ids": {}, |
| "num": null, |
| "urls": [], |
| "raw_text": "K. Fukunaga, Introduction to Statistical Pattern Recognition, 2nd ed. New York: Academic Press, 1990.", |
| "links": null |
| }, |
| "BIBREF16": { |
| "ref_id": "b16", |
| "title": "Canonical variate analysis with unequal covariance matricesgeneralizations of the usual solution", |
| "authors": [ |
| { |
| "first": "N", |
| "middle": [ |
| "A" |
| ], |
| "last": "Campbell", |
| "suffix": "" |
| } |
| ], |
| "year": 1984, |
| "venue": "Mathematical Geology", |
| "volume": "16", |
| "issue": "", |
| "pages": "109--124", |
| "other_ids": {}, |
| "num": null, |
| "urls": [], |
| "raw_text": "N. A. Campbell, \"Canonical variate analysis with unequal covariance matrices - generalizations of the usual solution,\" Mathematical Geology, vol. 16, pp. 109-124, 1984.", |
| "links": null |
| }, |
| "BIBREF17": { |
| "ref_id": "b17", |
| "title": "Linear discriminant analysis using a generalized mean of class covariances and its application to speech recognition", |
| "authors": [ |
| { |
| "first": "M", |
| "middle": [], |
| "last": "Sakai", |
| "suffix": "" |
| } |
| ], |
| "year": 2008, |
| "venue": "IEICE Trans. Information and Systems", |
| "volume": "", |
| "issue": "", |
| "pages": "478--487", |
| "other_ids": {}, |
| "num": null, |
| "urls": [], |
| "raw_text": "M. Sakai, et al., \"Linear discriminant analysis using a generalized mean of class covariances and its application to speech recognition,\" IEICE Trans. Information and Systems, vol. E91-D, pp. 478-487, 2008.", |
| "links": null |
| }, |
| "BIBREF18": { |
| "ref_id": "b18", |
| "title": "Computer Graphics: Principles and Practice in C", |
| "authors": [ |
| { |
| "first": "J", |
| "middle": [ |
| "D" |
| ], |
| "last": "Foley", |
| "suffix": "" |
| } |
| ], |
| "year": 1995, |
| "venue": "", |
| "volume": "", |
| "issue": "", |
| "pages": "", |
| "other_ids": {}, |
| "num": null, |
| "urls": [], |
| "raw_text": "J. D. Foley, et al., Computer Graphics: Principles and Practice in C, 2nd ed.: Addison-Wesley, 1995.", |
| "links": null |
| }, |
| "BIBREF19": { |
| "ref_id": "b19", |
| "title": "MATBN: A mandarin Chinese broadcast news corpus", |
| "authors": [ |
| { |
| "first": "H.-M", |
| "middle": [], |
| "last": "Wang", |
| "suffix": "" |
| } |
| ], |
| "year": 2005, |
| "venue": "International Journal of Computational Linguistics and Chinese Language Processing", |
| "volume": "10", |
| "issue": "", |
| "pages": "219--235", |
| "other_ids": {}, |
| "num": null, |
| "urls": [], |
| "raw_text": "H.-M. Wang, et al., \"MATBN: A mandarin Chinese broadcast news corpus,\" International Journal of Computational Linguistics and Chinese Language Processing, vol. 10, pp. 219-235, 2005.", |
| "links": null |
| }, |
| "BIBREF20": { |
| "ref_id": "b20", |
| "title": "Lightly supervised and data-driven approaches to mandarin broadcast news transcription", |
| "authors": [ |
| { |
| "first": "B", |
| "middle": [], |
| "last": "Chen", |
| "suffix": "" |
| } |
| ], |
| "year": 2004, |
| "venue": "Proc. ICASSP", |
| "volume": "", |
| "issue": "", |
| "pages": "777--780", |
| "other_ids": {}, |
| "num": null, |
| "urls": [], |
| "raw_text": "B. Chen, et al., \"Lightly supervised and data-driven approaches to mandarin broadcast news transcription,\" in Proc. ICASSP, 2004, pp. 777-780.", |
| "links": null |
| }, |
| "BIBREF21": { |
| "ref_id": "b21", |
| "title": "SRI Language Modeling Toolkit", |
| "authors": [ |
| { |
| "first": "A", |
| "middle": [], |
| "last": "Stolcke", |
| "suffix": "" |
| } |
| ], |
| "year": null, |
| "venue": "", |
| "volume": "", |
| "issue": "", |
| "pages": "", |
| "other_ids": {}, |
| "num": null, |
| "urls": [], |
| "raw_text": "A. Stolcke, SRI Language Modeling Toolkit (Version 1.5.2).", |
| "links": null |
| }, |
| "BIBREF22": { |
| "ref_id": "b22", |
| "title": "Maximum likelihood modeling with Gaussian distributions for classification", |
| "authors": [ |
| { |
| "first": "R", |
| "middle": [ |
| "A" |
| ], |
| "last": "Gopinth", |
| "suffix": "" |
| } |
| ], |
| "year": 1998, |
| "venue": "Proc. ICASSP", |
| "volume": "", |
| "issue": "", |
| "pages": "661--664", |
| "other_ids": {}, |
| "num": null, |
| "urls": [], |
| "raw_text": "R. A. Gopinth, \"Maximum likelihood modeling with Gaussian distributions for classification,\" in Proc. ICASSP, 1998, pp. 661-664.", |
| "links": null |
| }, |
| "BIBREF23": { |
| "ref_id": "b23", |
| "title": "Improved vocabulary-independent sub-word HMM modelling", |
| "authors": [ |
| { |
| "first": "L", |
| "middle": [], |
| "last": "Wood", |
| "suffix": "" |
| } |
| ], |
| "year": 1991, |
| "venue": "Proc. ICASSP", |
| "volume": "", |
| "issue": "", |
| "pages": "181--184", |
| "other_ids": {}, |
| "num": null, |
| "urls": [], |
| "raw_text": "L. Wood, et al., \"Improved vocabulary-independent sub-word HMM modelling,\" in Proc. ICASSP, 1991, pp. 181-184.", |
| "links": null |
| }, |
| "BIBREF24": { |
| "ref_id": "b24", |
| "title": "Discriminant analysis and supervised vector quantization for continuous speech recognition", |
| "authors": [ |
| { |
| "first": "G", |
| "middle": [], |
| "last": "Yu", |
| "suffix": "" |
| } |
| ], |
| "year": 1990, |
| "venue": "Proc. ICASSP", |
| "volume": "", |
| "issue": "", |
| "pages": "685--688", |
| "other_ids": {}, |
| "num": null, |
| "urls": [], |
| "raw_text": "G. Yu, et al., \"Discriminant analysis and supervised vector quantization for continuous speech recognition,\" in Proc. ICASSP, 1990, pp. 685-688.", |
| "links": null |
| } |
| }, |
| "ref_entries": { |
| "FIGREF4": { |
| "type_str": "figure", |
| "num": null, |
| "uris": null, |
| "text": "Language Modeling Toolkit (SRILM)[22]\u3002\u5c31\u4ee5\u6885\u723e\u5012\u983b\u8b5c\u4fc2\u6578(MFCCs) \u70ba\u57fa\u790e\u5be6\u9a57(baseline)\u800c\u8a00\uff0c\u5176\u8a5e\u8fa8\u8b58\u6b63\u78ba\u7387(character accuracy)\u70ba 72.23%\u3002 (\u4e8c)\u5be6\u9a57\u7d50\u679c \u672c\u8ad6\u6587\u4e2d\u6240\u63d0\u5230\u7684\u91cd\u8981\u7279\u5fb5\u62bd\u53d6\u65b9\u6cd5\uff0c\u5982\u7dda\u6027\u9451\u5225\u5206\u6790(LDA)\u3001\u7570\u65b9\u5dee\u6027\u7dda\u6027\u9451\u5225 \u5206\u6790(HLDA)\u3001\u7570\u65b9\u5dee\u6027\u9451\u5225\u5206\u6790(HDA)\uff0c\u4ee5\u53ca\u6211\u5011\u6240\u63d0\u51fa\u7684\u666e\u904d\u5316\u76f8\u4f3c\u5ea6\u6bd4\u7387\u9451\u5225\u5206 \u6790(GLRDA)\uff0c\u90fd\u662f\u4f5c\u7528\u5728 162 \u7dad(n = 162)\u7684\u8d85\u7d1a\u5411\u91cf(super-vector)\u4e0a\u9032\u884c\u964d\u7dad\u8655\u7406\u3002\u6b64 \u8d85\u7d1a\u5411\u91cf\u662f\u7531\u9023\u7e8c 9 \u500b\u97f3\u6846\u4e4b\u6885\u6c0f\u6ffe\u6ce2\u5668\u7d44(Mel-frequency filterbank)\u6240\u8f38\u51fa\u7684 18 \u7dad\u7279 \u5fb5\u5411\u91cf\u4e32\u63a5\u800c\u6210\uff0c\u76ee\u7684\u5728\u65bc\u6355\u6349\u97f3\u6846\u9593\u7684\u52d5\u614b\u8cc7\u8a0a\u3002\u800c\u76ee\u6a19\u7dad\u5ea6\u5247\u8a2d\u5b9a\u70ba 39 \u7dad(d = 39)\uff0c \u5176\u76ee\u7684\u5247\u5728\u65bc\u4f7f\u6211\u5011\u80fd\u5728\u5b50\u7a7a\u9593\u7dad\u5ea6\u56fa\u5b9a\u7684\u60c5\u6cc1\u4e0b\uff0c\u5b9a\u6027\u5730\u6bd4\u8f03\u5404\u7a2e\u65b9\u6cd5\u7684\u512a\u52a3\u3002\u5206\u985e \u7684\u6700\u5c0f\u55ae\u4f4d\u5247\u662f\u4ee5\u96b1\u85cf\u5f0f\u99ac\u53ef\u592b\u6a21\u578b(HMMs)\u4e2d\u7684\u72c0\u614b(state)\u70ba\u4e3b\uff0c\u4e26\u7d93\u7531\u4e00\u500b\u8fa8\u8b58\u6548\u679c \u8f03\u9ad8\u7684\u7cfb\u7d71\u91dd\u5c0d\u6bcf\u4e00\u8a13\u7df4\u8a9e\u53e5\u9032\u884c\u5f37\u5236\u6821\u6e96(forced alignment)\uff0c\u5f9e\u800c\u7522\u751f\u8a9e\u53e5\u4e2d\u7684\u985e\u5225 (\u97f3\u7d20\u548c\u72c0\u614b)\u5206\u754c\u3002\u800c\u6df7\u6dc6\u8cc7\u8a0a\u7684\u7372\u5f97\u5373\u4ee5\u6b64\u70ba\u6b63\u78ba\u7b54\u6848\uff0c\u91dd\u5c0d\u6bcf\u4e00\u97f3\u6846\u9032\u884c\u985e\u5225\u6bd4 \u5c0d\u800c\u5f97\u3002\u7531\u65bc\u9019\u4e9b\u7279\u5fb5\u8f49\u63db\u65b9\u6cd5\u6240\u64f7\u53d6\u51fa\u7684\u7279\u5fb5\u5411\u91cf\u4e26\u4e0d\u6703\u4f7f\u5f97\u5404\u500b\u985e\u5225\u7684\u5171\u8b8a\u7570\u77e9\u9663 \u70ba\u5c0d\u89d2\u5316\uff0c\u6703\u9020\u6210\u5f8c\u7aef\u96b1\u85cf\u5f0f\u99ac\u53ef\u592b\u6a21\u578b(HMMs)\u53c3\u6578\u7684\u4f30\u8a08\u5931\u771f\uff0c\u6240\u4ee5\u6211\u5011\u5617\u8a66\u5728\u9019 \u4e9b\u65b9\u6cd5\u5f8c\u5404\u81ea\u52a0\u4e0a\u6700\u5927\u5316\u76f8\u4f3c\u5ea6\u7dda\u6027\u8f49\u63db(maximum likelihood linear transformation, MLLT)[23]\u3002 \u5728\u7570\u65b9\u5dee\u6027\u4e4b\u666e\u904d\u5316\u76f8\u4f3c\u5ea6\u6bd4\u7387\u9451\u5225\u5206\u6790(GLRDA)\u7684\u90e8\u5206\uff0c\u7531\u65bc\u5728\u76ee\u6a19\u51fd\u5f0f(25" |
| }, |
| "TABREF0": { |
| "content": "<table><tr><td>\u7dda\u6027\u8f49\u63db \u6839\u64da\u7d71\u8a08\u5f0f\u5047\u8a2d\u6aa2\u5b9a(statistical hypothesis testing)\u7684\u5b9a\u7fa9[13]\uff0c\u76f8\u4f3c\u5ea6\u6bd4\u7387\u6aa2\u5b9a(LRT) d n\u00d7 \u211c \u2208 \u0398</td></tr><tr><td>\u662f\u4e00\u7a2e\u5ee3\u70ba\u4f7f\u7528\u7684\u65b9\u6cd5\uff0c\u85c9\u8457\u5b83\u6211\u5011\u53ef\u7372\u5f97\u865b\u7121\u5047\u8a2d(null hypothesis)H 0 \u8207\u5b8c\u5168\u666e\u904d\u5316\u4e4b</td></tr><tr><td>\u5c0d\u7acb\u5047\u8a2d(alternative hypothesis)H 1 \u9593\u76f8\u4e92\u6bd4\u8f03\u7684\u6aa2\u5b9a\u7d71\u8a08\u91cf\u3002\u5728\u672c\u8ad6\u6587\u4e2d\uff0c\u865b\u7121\u5047\u8a2d</td></tr><tr><td>H 0 \u901a\u5e38\u8868\u793a\u4e0d\u5229\u65bc\u6211\u5011\u7684\u76ee\u6a19\u8a2d\u5b9a\uff0c\u6216\u6211\u5011\u4e0d\u9858\u898b\u5230\u7684\u60c5\u6cc1\uff0c\u5728\u9451\u5225\u5f0f\u7279\u5fb5\u64f7\u53d6\u4e0a\uff0c</td></tr><tr><td>\u5373\u70ba\u4f7f\u985e\u5225\u6bcd\u9ad4\u4e0d\u5177\u9451\u5225\u6027\u7684\u60c5\u6cc1\u3002\u503c\u5f97\u4e00\u63d0\u7684\u662f\uff0c\u865b\u7121\u5047\u8a2d\u548c\u5c0d\u7acb\u5047\u8a2d\u4e4b\u806f\u96c6(union)</td></tr><tr><td>\u6070\u70ba\u5b8c\u6574\u7684\u53c3\u6578\u7a7a\u9593\u3002</td></tr><tr><td>\u82e5</td></tr><tr><td>\u6a19\u6e96\u4e5f\u88ab\u5f15\u5165\u6b64\u7bc4\u7587\u4e2d\u4f7f\u7528[8]\u3002\u6b64\u5916\uff0c\u6700\u8fd1\u7684\u7814\u7a76\u8005\u958b\u59cb\u5c07\u6210\u5c0d\u7d93\u9a57\u932f\u8aa4\u7387(pairwise</td></tr><tr><td>empirical error rate)\u5217\u5165\u8003\u91cf\uff0c\u671f\u671b\u8fa8\u8b58\u5668\u5728\u524d\u7aef\u8655\u7406\u8207\u5f8c\u7aef\u5206\u985e\u968e\u6bb5\u7684\u4e0d\u4e00\u81f4\u6027\u80fd\u5920\u964d</td></tr><tr><td>\u4f4e\u81f3\u4e00\u5b9a\u7a0b\u5ea6[9-11]\u3002</td></tr><tr><td>\u5c0d\u65bc\u7368\u7acb\u65bc\u5206\u985e\u5668\u7684\u7bc4\u7587\u4f86\u8aaa\uff0c\u96d6\u7136\u5728\u985e\u5225\u5206\u96e2\u5ea6\u8207\u8fa8\u8b58\u7d50\u679c\u4e4b\u9593\u4ecd\u5b58\u6709\u8f03\u5927\u7684\u5dee</td></tr><tr><td>\u8ddd\uff0c\u4e5f\u5c31\u662f\u8f03\u9ad8\u7684\u985e\u5225\u5206\u96e2\u5ea6\uff0c\u4e26\u4e0d\u5fc5\u7136\u4fdd\u8b49\u6709\u8f03\u4f4e\u7684\u8fa8\u8b58\u932f\u8aa4\u7387\u3002\u4f46\u5728\u672c\u8ad6\u6587\u4e2d\uff0c\u6211</td></tr><tr><td>\u5011\u4ecd\u5c07\u7814\u7a76\u91cd\u9ede\u805a\u7126\u65bc\u6b64\uff0c\u539f\u56e0\u5728\u65bc\uff1a\u7576\u8a9e\u97f3\u7279\u5fb5\u64f7\u53d6\u5b8c\u5168\u8207\u5f8c\u7aef\u8072\u5b78\u6a21\u578b\u5206\u96e2\uff0c\u5c0d\u65bc</td></tr><tr><td>\u8f03\u8907\u96dc\u7684\u81ea\u52d5\u8a9e\u97f3\u8fa8\u8b58\u7cfb\u7d71\uff0c\u8072\u5b78\u6a21\u578b\u8a13\u7df4\u6a21\u5f0f\u7684\u6539\u8b8a\uff0c\u5c31\u8f03\u4e0d\u6703\u5f71\u97ff\u5230\u524d\u7aef\u7684\u8a0a\u865f\u8655</td></tr><tr><td>\u7406\uff0c\u4f7f\u5f97\u6b64\u7cfb\u7d71\u8f03\u6613\u65bc\u88ab\u5206\u6790\u89e3\u69cb\u3002\u800c\u7576\u67d0\u4e9b\u7cfb\u7d71\u7684\u8072\u5b78\u6a21\u578b\u6a5f\u5236\u662f\u56fa\u5b9a\u7684\uff0c\u6216\u662f\u4ee5\u786c</td></tr><tr><td>\u9ad4\u65b9\u5f0f\u5448\u73fe\uff0c\u90a3\u9ebc\u6211\u5011\u5c31\u80fd\u5728\u7121\u6cd5\u66f4\u52d5\u786c\u9ad4\u7684\u60c5\u6cc1\u4e0b\uff0c\u5c0d\u524d\u7aef\u8a0a\u865f\u8655\u7406\u9032\u884c\u7814\u7a76\u6216\u6539\u5584</td></tr><tr><td>[12]\u3002\u66f4\u91cd\u8981\u7684\u662f\uff0c\u6211\u5011\u76f8\u4fe1\u5728\u6b64\u7bc4\u7587\u4e2d\u6240\u8a2d\u8a08\u51fa\u7684\u65b9\u6cd5\uff0c\u80fd\u5920\u66f4\u5ee3\u6cdb\u5730\u61c9\u7528\u5728\u5176\u4ed6\u5716</td></tr><tr><td>\u578b\u8fa8\u8b58(pattern recognition)\u7684\u9818\u57df\uff0c\u5982\u4eba\u81c9\u8fa8\u8b58\u7b49\uff0c\u800c\u4e0d\u4fb7\u9650\u65bc\u7cfb\u7d71\u6a21\u578b\u8f03\u70ba\u8907\u96dc\u7684\u8a9e</td></tr><tr><td>\u97f3\u8fa8\u8b58\u3002</td></tr><tr><td>\u5728\u672c\u8ad6\u6587\u4e2d\uff0c\u6211\u5011\u63d0\u51fa\u4e86\u4e00\u500b\u5d84\u65b0\u7684\u9451\u5225\u5f0f\u7279\u5fb5\u64f7\u53d6\u65b9\u6cd5\uff0c\u7a31\u70ba\u666e\u904d\u5316\u76f8\u4f3c\u5ea6\u6bd4\u7387</td></tr><tr><td>\u9451\u5225\u5206\u6790(generalized likelihood ratio discriminant analysis, GLRDA)\uff0c\u5176\u65e8\u5728\u5229\u7528\u76f8\u4f3c\u5ea6</td></tr><tr><td>\u6bd4\u7387\u6aa2\u9a57(likelihood ratio test, LRT)\u7684\u6982\u5ff5\u4f86\u5c0b\u6c42\u4e00\u500b\u7dad\u5ea6\u8f03\u4f4e\u7684\u7279\u5fb5\u7a7a\u9593\u3002\u5728\u6b64\u5b50\u7a7a\u9593</td></tr><tr><td>\u4e2d\uff0c\u6211\u5011\u4e0d\u50c5\u8003\u616e\u4e86\u5168\u9ad4\u8cc7\u6599\u7684\u7570\u65b9\u5dee\u6027(heteroscedasticity)\uff0c\u5373\u6240\u6709\u985e\u5225\u6bcd\u9ad4\u4e4b\u5171\u8b8a\u7570</td></tr><tr><td>\u77e9\u9663\u53ef\u88ab\u5f48\u6027\u5730\u8996\u70ba\u76f8\u7570\uff0c\u4e26\u4e14\u5728\u5206\u985e\u4e0a\uff0c\u56e0\u8457\u6211\u5011\u4e5f\u5c07\u985e\u5225\u9593\u6700\u6df7\u6dc6\u4e4b\u60c5\u6cc1(\u7531\u865b\u7121</td></tr><tr><td>\u5047\u8a2d(null hypothesis)\u6240\u63cf\u8ff0)\u7684\u767c\u751f\u7387\u964d\u81f3\u6700\u4f4e\uff0c\u800c\u9054\u5230\u6709\u52a9\u65bc\u5206\u985e\u6b63\u78ba\u7387\u63d0\u5347\u7684\u6548</td></tr><tr><td>\u679c\u3002\u6b64\u5916\uff0c\u82e5\u6211\u5011\u5047\u8a2d\u6240\u6709\u985e\u5225\u6bcd\u9ad4\u5747\u9075\u5faa\u9ad8\u65af\u5206\u4f48(Gaussian distribution)\uff0c\u4e14\u91dd\u5c0d\u5176</td></tr><tr><td>\u5171\u8b8a\u7570\u77e9\u9663\u7d66\u4e88\u4e0d\u540c\u7684\u9650\u5236\uff0c\u5247\u666e\u904d\u5316\u76f8\u4f3c\u5ea6\u6bd4\u7387\u9451\u5225\u5206\u6790(GLRDA)\u53ef\u88ab\u5316\u7d04\u81f3\u50b3\u7d71</td></tr><tr><td>\u7684\u7dda\u6027\u9451\u5225\u5206\u6790\u8207\u6709\u540d\u7684\u7570\u65b9\u5dee\u6027\u7dda\u6027\u9451\u5225\u5206\u6790\u3002\u800c\u70ba\u4e86\u589e\u9032\u8072\u5b78\u7279\u5fb5\u7684\u5f37\u5065\u6027\uff0c\u6211\u5011</td></tr><tr><td>\u7684\u65b9\u6cd5\u66f4\u53ef\u9032\u4e00\u6b65\u5730\u8207\u8fa8\u8b58\u5668\u6240\u63d0\u4f9b\u7684\u7d93\u9a57\u6df7\u6dc6\u8cc7\u8a0a\u7d50\u5408\u3002</td></tr><tr><td>\u4e8c\u3001\u666e\u904d\u5316\u76f8\u4f3c\u5ea6\u6bd4\u7387\u9451\u5225\u5206\u6790</td></tr><tr><td>(\u4e00)\u80cc\u666f</td></tr></table>", |
| "text": "\u6458\u8981 \u5728\u8fd1\u5341\u5e74\u4f86\u6240\u767c\u5c55\u51fa\u7684\u81ea\u52d5\u8a9e\u97f3\u8fa8\u8b58(automatic speech recognition, ASR)\u6280\u8853\u4e2d\uff0c\u4ecd \u6709\u8a31\u591a\u7814\u7a76\u8005\u5617\u8a66\u50c5\u85c9\u7531\u524d\u7aef\u8655\u7406\u4f86\u7522\u751f\u5177\u6709\u9451\u5225\u6027\u7684\u8a9e\u97f3\u7279\u5fb5\uff0c\u800c\u7368\u7acb\u65bc\u5f8c\u7aef\u6a21\u578b\u8a13 \u7df4\u8207\u5206\u985e\u5668\u7279\u6027\u3002\u672c\u8ad6\u6587\u5373\u5728\u6b64\u601d\u7dad\u4e0b\u63d0\u51fa\u5d84\u65b0\u7684\u9451\u5225\u5f0f\u7279\u5fb5\u8f49\u63db\u65b9\u6cd5\uff0c\u7a31\u70ba\u666e\u904d\u5316\u76f8 \u4f3c\u5ea6\u6bd4\u7387\u9451\u5225\u5206\u6790(generalized likelihood ratio discriminant analysis, GLRDA)\uff0c\u5176\u65e8\u5728\u5229 \u7528\u76f8\u4f3c\u5ea6\u6bd4\u7387\u6aa2\u9a57(likelihood ratio test)\u7684\u6982\u5ff5\u5c0b\u6c42\u4e00\u500b\u7dad\u5ea6\u8f03\u4f4e\u7684\u7279\u5fb5\u7a7a\u9593\u3002\u5728\u6b64\u5b50\u7a7a \u9593\u4e2d\uff0c\u6211\u5011\u4e0d\u50c5\u8003\u616e\u4e86\u5168\u9ad4\u8cc7\u6599\u7684\u7570\u65b9\u5dee\u6027(heteroscedasticity)\uff0c\u5373\u6240\u6709\u985e\u5225\u4e4b\u5171\u8b8a\u7570\u77e9 \u9663\u53ef\u88ab\u5f48\u6027\u5730\u8996\u70ba\u76f8\u7570\uff0c\u4e26\u4e14\u5728\u5206\u985e\u4e0a\uff0c\u56e0\u8457\u6211\u5011\u4e5f\u5c07\u985e\u5225\u9593\u6700\u6df7\u6dc6\u4e4b\u60c5\u6cc1(\u7531\u865b\u7121\u5047 \u8a2d(null hypothesis)\u6240\u63cf\u8ff0)\u7684\u767c\u751f\u7387\u964d\u81f3\u6700\u4f4e\uff0c\u800c\u9054\u5230\u6709\u52a9\u65bc\u5206\u985e\u6b63\u78ba\u7387\u63d0\u5347\u7684\u6548\u679c\u3002 \u540c\u6642\uff0c\u6211\u5011\u4e5f\u8b49\u660e\u4e86\u50b3\u7d71\u7684\u7dda\u6027\u9451\u5225\u5206\u6790(linear discriminant analysis, LDA)\u8207\u6709\u540d\u7684\u7570 \u65b9\u5dee\u6027\u7dda\u6027\u9451\u5225\u5206\u6790(heteroscedastic linear discriminant analysis, HLDA)\u53ef\u88ab\u8996\u70ba\u6211\u5011\u6240 \u63d0\u51fa\u4e4b\u666e\u904d\u5316\u76f8\u4f3c\u5ea6\u6bd4\u7387\u9451\u5225\u5206\u6790(GLRDA)\u7684\u5169\u7a2e\u7279\u4f8b\u3002\u6b64\u5916\uff0c\u70ba\u4e86\u589e\u9032\u8a9e\u97f3\u7279\u5fb5\u7684 \u5f37\u5065\u6027\uff0c\u6211\u5011\u6240\u63d0\u51fa\u7684\u65b9\u6cd5\u66f4\u53ef\u9032\u4e00\u6b65\u5730\u8207\u8fa8\u8b58\u5668\u6240\u63d0\u4f9b\u7684\u5be6\u969b\u6df7\u6dc6\u8cc7\u8a0a\u7d50\u5408\uff0c\u800c\u7372\u5f97 \u5728\u4e2d\u6587\u5927\u8a5e\u5f59\u9023\u7e8c\u8a9e\u97f3\u8fa8\u8b58\u7684\u5be6\u9a57\u4e2d\uff0c\u76f8\u8f03\u65bc\u4ee5\u4e0a\u5169\u7a2e\u50b3\u7d71\u65b9\u6cd5\u66f4\u9ad8\u7684\u8fa8\u8b58\u6b63\u78ba\u7387\u3002 \u95dc\u9375\u8a5e\uff1a\u8a9e\u97f3\u8fa8\u8b58\u3001\u7279\u5fb5\u64f7\u53d6\u3001\u76f8\u4f3c\u5ea6\u6bd4\u7387\u3001\u9451\u5225\u5206\u6790\u3001\u6df7\u6dc6\u8cc7\u8a0a \u4e00\u3001\u7dd2\u8ad6 \u70ba\u4e86\u964d\u4f4e\u8a08\u7b97\u91cf\u8207\u6a21\u578b\u7684\u8907\u96dc\u5ea6\uff0c\u8a9e\u97f3\u7279\u5fb5\u8f49\u63db(feature transformation)\u5728\u81ea\u52d5\u8a9e\u97f3 \u8fa8\u8b58(automatic speech recognition, ASR)\u4e2d\u626e\u6f14\u4e86\u5f88\u91cd\u8981\u7684\u89d2\u8272\u3002\u5b83\u7684\u76ee\u6a19\u5728\u65bc\u5c0b\u6c42\u4e00\u500b \uff0c\u5c07\u539f\u6709\u5728 n \u7dad\u7a7a\u9593\u7684\u8072\u5b78\u7279\u5fb5\u5411\u91cf\uff0c\u6295\u5f71\u81f3 d \u7dad\u7684\u5b50\u7a7a\u9593(d < n)\uff0c \u4f7f\u5f97\u65b0\u7684\u7279\u5fb5\u5728\u8cc7\u6599\u985e\u5225\u9593\u5177\u6709\u8f03\u597d\u7684\u9451\u5225\u529b[1]\u3002\u800c\u5728\u5be6\u52d9\u4e0a\uff0c\u8a9e\u97f3\u7279\u5fb5\u8f49\u63db\u7684\u6280\u8853 \u53ef \u88ab \u5206 \u70ba \u5169 \u7a2e \u7bc4 \u7587 [2] \uff1a \u76f8 \u4f9d \u65bc \u5206 \u985e \u5668 (classifier-dependent) \u8207 \u7368 \u7acb \u65bc \u5206 \u985e \u5668 (classifier-independent)\u3002\u5728\u76f8\u4f9d\u65bc\u5206\u985e\u5668\u7684\u7bc4\u7587\u4e2d\uff0c\u5982\u67d0\u4e9b\u57fa\u65bc\u6700\u5c0f\u97f3\u7d20\u932f\u8aa4(minimum phone error, MPE)[3]\u8207\u6700\u5c0f\u5206\u985e\u932f\u8aa4(minimum classification error, MCE)[4]\u7684\u9451\u5225\u5f0f\u65b9 \u6cd5\uff0c\u8f49\u63db\u77e9\u9663\u662f\u7d50\u5408\u8072\u5b78\u6a21\u578b(acoustic models)\u4e2d\u7684\u53c3\u6578\u4f30\u6e2c\u6216\u662f\u5206\u985e\u5668\u6240\u638c\u63e1\u7684\u5206\u985e\u898f \u5247\u4e00\u4f75\u6c42\u5f97\u3002\u76f8\u5c0d\u5730\uff0c\u7368\u7acb\u65bc\u5206\u985e\u5668\u7684\u7bc4\u7587\u5247\u662f\u57fa\u65bc\u5404\u7a2e\u4e0d\u540c\u7684\u985e\u5225\u5206\u96e2\u5ea6\u6a19\u6e96\uff0c\u7279\u5225 \u662f\u5e7e\u4f55\u5206\u96e2\u5ea6\uff0c\u5728\u8072\u5b78\u6a21\u578b\u8a13\u7df4\u4e4b\u524d\u5c31\u4f9d\u64da\u65e2\u6709\u7684\u985e\u5225\u7d71\u8a08\u8cc7\u8a0a\u6c42\u51fa\u8f49\u63db\u77e9\u9663\u3002\u4f8b\u5982\uff0c \u7dda\u6027\u9451\u5225\u5206\u6790(linear discriminant analysis, LDA)\u5373\u8a66\u5716\u6700\u5927\u5316\u985e\u5225\u9593\u7684\u5e73\u5747\u99ac\u6c0f\u8ddd\u96e2\u5e73 \u65b9(squared Mahalanobis distance)[5]\uff1b\u800c\u505a\u70ba\u7dda\u6027\u9451\u5225\u5206\u6790(LDA)\u7684\u666e\u904d\u5316\uff0c\u7570\u8cea\u6027\u9451\u5225 \u6027\u5206\u6790(heteroscedastic linear discriminant analysis, HLDA)\u5247\u5728\u6700\u5927\u5316\u76f8\u4f3c\u5ea6(maximum likelihood)\u7684\u6846\u67b6\u4e0b\uff0c\u8655\u7406\u6bcf\u4e00\u985e\u5225\u5177\u6709\u76f8\u7570\u4e4b\u5171\u8b8a\u7570\u77e9\u9663\u7684\u60c5\u5f62[6]\u3002\u53e6\u5916\uff0c\u7570\u8cea\u6027\u9451 \u5225\u5206\u6790(heteroscedastic discriminant analysis, HDA)\u500b\u5225\u5730\u8003\u616e\u4e86\u6bcf\u4e00\u985e\u5225\u7684\u5206\u4f48\uff0c\u800c\u7522 \u751f\u65b0\u7684\u76ee\u6a19\u51fd\u5f0f[7]\uff1b\u70ba\u4e86\u4fdd\u7559\u6bd4\u7dda\u6027\u9451\u5225\u5206\u6790(LDA)\u548c\u7570\u8cea\u6027\u9451\u5225\u6027\u5206\u6790(HLDA)\u66f4\u591a \u7684\u9451\u5225\u8cc7\u8a0a\uff0c\u6700\u5927\u5316\u4ea4\u4e92\u8cc7\u8a0a(maximum mutual information, MMI)\u548c\u6700\u5c0f\u5206\u985e\u932f\u8aa4(MCE)", |
| "html": null, |
| "type_str": "table", |
| "num": null |
| } |
| } |
| } |
| } |
