| { |
| "paper_id": "O09-1004", |
| "header": { |
| "generated_with": "S2ORC 1.0.0", |
| "date_generated": "2023-01-19T08:10:49.472122Z" |
| }, |
| "title": "\u5f37\u5065\u6027\u8a9e\u97f3\u8fa8\uf9fc\u4e2d\u5206\u983b\u6bb5\u8abf\u8b8a\u983b\u8b5c\u88dc\u511f\u4e4b\u7814\u7a76 A Study of Sub-band Modulation Spectrum Compensation for Robust Speech Recognition", |
| "authors": [ |
| { |
| "first": "Sheng-Yuan", |
| "middle": [], |
| "last": "\u9ec3\u52dd\u6e90", |
| "suffix": "", |
| "affiliation": {}, |
| "email": "" |
| }, |
| { |
| "first": "", |
| "middle": [], |
| "last": "Huang", |
| "suffix": "", |
| "affiliation": {}, |
| "email": "" |
| }, |
| { |
| "first": "Wen-Hsiang", |
| "middle": [], |
| "last": "\u675c\u6587\u7965", |
| "suffix": "", |
| "affiliation": {}, |
| "email": "" |
| }, |
| { |
| "first": "", |
| "middle": [], |
| "last": "Tu", |
| "suffix": "", |
| "affiliation": {}, |
| "email": "" |
| } |
| ], |
| "year": "", |
| "venue": null, |
| "identifiers": {}, |
| "abstract": "In this paper, we propose a novel scheme in performing feature statistics normalization techniques for robust speech recognition. In the proposed approach, the processed temporal-domain feature sequence is first converted into the modulation spectral domain. The magnitude part of the modulation spectrum is decomposed into non-uniform sub-band", |
| "pdf_parse": { |
| "paper_id": "O09-1004", |
| "_pdf_hash": "", |
| "abstract": [ |
| { |
| "text": "In this paper, we propose a novel scheme in performing feature statistics normalization techniques for robust speech recognition. In the proposed approach, the processed temporal-domain feature sequence is first converted into the modulation spectral domain. The magnitude part of the modulation spectrum is decomposed into non-uniform sub-band", |
| "cite_spans": [], |
| "ref_spans": [], |
| "eq_spans": [], |
| "section": "Abstract", |
| "sec_num": null |
| } |
| ], |
| "body_text": [ |
| { |
| "text": "segments, and then each sub-band segment is individually processed by the well-known normalization methods, like mean normalization (MN), mean and variance normalization (MVN) and histogram equalization (HEQ). Finally, we reconstruct the feature stream with all the modified sub-band magnitude spectral segments and the original phase spectrum using the inverse DFT. With this process, the components that correspond to more important modulation spectral bands in the feature sequence can be processed separately. For the Aurora-2 clean-condition training task, the new proposed sub-band spectral MN, MVN and HEQ provide relative error rate reductions of 18.66% and 23.58% over the conventional temporal MVN and HEQ, respectively. \u4e00\u3001\u7c21\u4ecb \u96d6\u7136\u8a9e\u97f3\u79d1\u6280\u9032\u6b65\u8fc5\u901f\uff0c\u4f46\u81ea\u52d5\u8a9e\u97f3\u8fa8\uf9fc(automatic speech recognition, ASR) [ ", |
| "cite_spans": [ |
| { |
| "start": 790, |
| "end": 791, |
| "text": "[", |
| "ref_id": null |
| } |
| ], |
| "ref_spans": [], |
| "eq_spans": [], |
| "section": "", |
| "sec_num": null |
| }, |
| { |
| "text": "(2-1)\u8868\u793a\uff1a ( ) [ ] { } ;1 , 1 m x n n N m M \u2264 \u2264 \u2264 \u2264 , \u5f0f(2-1) \u5176\u4e2d M \u70ba\u4e00\u8a9e\u97f3\u7279\u5fb5\u5411\uf97e\u4e2d\u7279\u5fb5\u500b\uf969\uff0c N \u8868\u793a\u70ba\u6b64\u55ae\u4e00\u8a9e\uf906\u7684\u97f3\u6846\u7e3d\uf969\u3002\u6bcf\u500b\u7279\u5fb5\u5e8f \uf99c ( ) [ ] { } m x n \u7d93\u6b63\u898f\u5316\u8655\uf9e4\u5f8c\uff0c\u4ee5 ( ) [ ] { } m x n \u8868\u793a\uff0c\u6211\u5011\u5e0c\u671b\u65b0\u7684\u7279\u5fb5\u5e8f\uf99c ( ) [ ] { } m x n \u76f8\u5c0d\u65bc \u539f\u59cb\u7279\u5fb5\u5e8f\uf99c\u800c\u8a00\uff0c\uf901\u5177\u6709\u5f37\u5065\u6027\uff0c\u4f7f\u8fa8\uf9fc\u6548\u679c\u6709\u660e\u986f\u5730\u63d0\u5347\u3002\u5728\u4e4b\u5f8c\u7684\u6558\u8ff0\uff0c\u70ba\uf9ba\u7cbe \u7c21\u7b26\u865f\u7684\u6a19\u793a\uff0c\u6211\u5011\uf96d\uf976\uf9ba\u4e0a\u6a19 ( ) \" \" m \u7b26\u865f\u3002 2. \u5c07\u7279\u5fb5\u5e8f\uf99c [ ] { } ;1 x n n N \u2264 \u2264 \u7d93 N \u9ede\uf9ea\u6563\u5085\uf9f7\uf96e\u8f49\u63db(discrete Fourier transform, DFT) \u5f8c\u5f97\u5230\u5176\u8abf\u8b8a\u983b\u8b5c [ ] { } X k \uff0c\u5982\u4e0b\u5f0f\u3002 [ ] [ ] 2 1 0 , 0 2 nk N j K n N X k x n e k \u03c0 \u2212 \u2212 = \u23a1 \u23a4 = \u2264 \u2264 \u23a2 \u23a5 \u23a2 \u23a5 \u23a2 \u23a5 \u2211 \u5f0f(2-2) \u5047\u8a2d [ ] { } x n \u7684\u97f3\u6846\u53d6\u6a23\u983b\uf961(frame rate)\u70ba s F Hz\uff0c\u5247\u5728\u5176\u8abf\u8b8a\u983b\u8b5c\u57df\u4e0a [ ] { } X k \u7684\u983b\uf961\u7bc4 \u570d\u70ba 0, 2 s F \u23a1 \u23a4 \u23a2 \u23a5 \u23a2 \u23a5 \u23a3 \u23a6 \uff1b\u800c\u7531\u65bc [ ] X k \u70ba\u4e00\u8907\uf969\uff0c\u6211\u5011\u4ee5\u6975\u5ea7\u6a19(polar form)\u8868\u793a [ ] X k \u5982\u4e0b\u5f0f\uff1a [ ] [ ] k j X k A k e \u03b8 = \u5f0f(2-3) \u5176\u4e2d [ ] A k \u662f [ ] X k \u7684\u5f37\ufa01\u6210\u4efd\uff0c [ ] k \u03b8 \u662f [ ] X k \u7684\u76f8\u4f4d\u6210\u4efd\uff0c\u63a5\u4e0b\uf92d\u6211\u5011\u53ea\u91dd\u5c0d\u5f37\ufa01\u6210\u4efd [ ] { } A k \u4f5c\u8abf\u6574\uff0c\u800c\u4fdd\uf9cd\u76f8\u4f4d\u6210\u4efd [ ] { } k \u03b8 \uf967\u8b8a\u3002 3. \u5c07\u4e0a\u4e00\u6b65\u9a5f\u8abf\u8b8a\u983b\u8b5c\u7684\u5f37\ufa01\u6210\u5206 [ ];0 2 N A k k \u23a7 \u23ab \u23a1 \u23a4 \u23aa \u23aa \u23aa \u23aa \u2264 \u2264 \u23a2 \u23a5 \u23a8 \u23ac \u23aa \u23aa \u23a2 \u23a5 \u23a2 \u23a5 \u23aa \u23aa \u23a9 \u23ad \u4ee5\uf967\u7b49\ufa00(non-uniform)\u4e14\u500d\u983b (octave)\u7684\u65b9\u5f0f\uff0c\ufa00\u5272\u6210 L \u500b\u983b\u6bb5\uff0c\u6bcf\u500b\u983b\u6bb5\u7684\u7bc4\u570d\u5982\u4e0b\u5f0f(2-4)\u6240\u793a\uff1a 1 2 1 1 1 1 0, , if 1. 2 2 2 2 , , if 2, 3,..., . 2 2 2 2 s L s s L L F F F L \u2212 \u2212 \u2212 \u2212 \u2212 \u23a7\u23a1 \u23a4 \u239b \u239e \u23aa \u23aa \u239f \u239c \u23a2 \u23a5 = \u239f \u23aa \u239c \u239f \u239c \u239d \u23a0 \u23aa\u23a2 \u23a5 \u23a3 \u23a6 \u23aa \u23a8 \u23a1 \u23a4 \u23aa \u239b \u239e \u239b \u239e \u23aa \u239f \u239f \u239c \u239c \u23a2 \u23a5 = \u23aa \u239f \u239f \u239c \u239c \u239f \u239f \u239c \u239c \u23aa \u23a2 \u23a5 \u239d \u23a0 \u239d \u23a0 \u23aa \u23a3 \u23a6 \u23a9 \u5f0f(2-4) \u7531\u4e0a\u5f0f\u53ef\u4ee5\u5f97\u77e5\uff0c\u8abf\u8b8a\u983b\u8b5c\u4f4e\u983b\u5e36\u7684\u90e8\u5206\u88ab\ufa00\u5272\u6210\u8f03\u591a\u500b\u983b\u6bb5\uff0c\u4e14\u6bcf\u500b\u983b\u6bb5\u7684\u9577\ufa01\u8f03 \u77ed\uff0c\u76f8\u5c0d\u5730\uff0c\u9ad8\u983b\u7684\u90e8\u5206\u88ab\ufa00\u5272\u6210\u8f03\u5c11\u7684\u983b\u6bb5\uff0c\u4e14\u6bcf\u500b\u983b\u6bb5\u7684\u9577\ufa01\u8f03\u9577\u3002\u5728\u5c07 [ ] { } A k \u4f5c \u4e0a\u8ff0\u7684\u983b\u6bb5\ufa00\u5272\u5f8c\uff0c\u6211\u5011\u4ee5 { } ' A k \u23a1 \u23a4 \u23a3 \u23a6 \u8868\u793a\u5176\u4e2d\u7684\u7b2c L \u500b\u983b\u6bb5\u3002\u6b64\u5c0d\u65bc\u983b\u6bb5\uf967\u7b49\ufa00\u7684\u539f\u56e0\uff0c \u5728\u65bc\u6211\u5011\u4e4b\u524d\u6240\u63d0\uff0c\u4f4e\u8abf\u8b8a\u983b\u5e36\u5c0d\u65bc\u8a9e\u97f3\u8fa8\uf9fc\u8f03\u70ba\u91cd\u8981\uff0c\uf9e4\u61c9\u5206\u8f03\u591a\u7684\u983b\u6bb5\uf92d\u500b\u5225\u8655 \uf9e4\uff0c\u800c\u9ad8\u8abf\u8b8a\u983b\u5e36\u76f8\u5c0d\u800c\u8a00\u8f03\uf967\u91cd\u8981\uff0c\u6240\u4ee5\u53ef\u5c07\u8f03\u5927\u7684\u983b\u6bb5\u7bc4\u570d\u4e00\u4f75\u8655\uf9e4\u3002 4. \u6211\u5011\u5c07\u4e0a\u4e00\u6b65\u9a5f\u6240\u5f97\u4e4b\uf967\u540c\u983b\u6bb5\u7684\u5f37\ufa01\u983b\u8b5c { } ' A k \u23a1 \u23a4 \u23a3 \u23a6 \u4f5c\u7d71\u8a08\u6b63\u898f\u5316\u8655\uf9e4\u3002\u6211\u5011\u4f7f\u7528\u7684 \u6b63\u898f\u5316\u6cd5\u5206\u5225\u70ba\uff1a\u5e73\u5747\u503c\u6b63\u898f\u5316\u6cd5(MN)\u3001\u5e73\u5747\u503c\u8207\u8b8a\uf962\uf969\u6b63\u898f\u5316\u6cd5(MVN)\u8207\u7d71\u8a08\u5716\u7b49 \u5316\u6cd5(HEQ)\uff0c\u8655\uf9e4\u5f8c\u7684\u7279\u5fb5\u5373\u4ee5 { } ' A k \u23a1 \u23a4 \u23a3 \u23a6 \u8868\u793a\u3002\u8a73\u7d30\u5730\uf96f\uff0c\u5e73\u5747\u503c\u6b63\u898f\u5316\u6cd5(MN)\u5728\u6b64\u7684 \u8a08\u7b97\u65b9\u5f0f\u4ee5\u4e0b\u5f0f(2-5)\u8868\u793a\uff1a ' ' , , s a A k A k \u03bc \u03bc \u23a1 \u23a4 \u23a1 \u23a4 = \u2212 + \u23a3 \u23a6 \u23a3 \u23a6 , \u5f0f(2-5) \u5176\u4e2d\uff0c ,s \u03bc \u70ba\u55ae\u4e00(single)\u8a9e\uf906\u4e4b\u5206\u983b\u6bb5\u5f37\ufa01\u983b\u8b5c\u7684\u5e73\u5747\u503c\uff0c ,a \u03bc \u70ba\u5168\u90e8(all)\u8a13\uf996\u8a9e\uf906\u4e4b\u5206 \u983b\u6bb5\u5f37\ufa01\u983b\u8b5c\u7684\u5e73\u5747\u503c\u3002 \u5e73\u5747\u503c\u8207\u8b8a\uf962\uf969\u6b63\u898f\u5316\u6cd5(MVN)\u5728\u6b64\u7684\u8a08\u7b97\u65b9\u5f0f\u4ee5\u5f0f(2-6)\u8868\u793a\uff1a ' , ' , , , s a a s A k A k \u03bc \u03c3 \u03bc \u03c3 \u239b \u239e \u23a1 \u23a4 \u2212 \u239f \u239c \u23a3 \u23a6 \u239f \u23a1 \u23a4 \u239c = \u22c5 + \u239f \u239c \u23a3 \u23a6 \u239f \u239c \u239f \u239d \u23a0 \u5f0f(2-6) \u5176\u4e2d\uff0c ,s \u03bc \u70ba\u55ae\u4e00\u8a9e\uf906\u4e4b\u5206\u983b\u6bb5\u5f37\ufa01\u983b\u8b5c\u7684\u5e73\u5747\u503c\uff0c ,s \u03c3 \u70ba\u55ae\u4e00\u8a9e\uf906\u4e4b\u5206\u983b\u6bb5\u5f37\ufa01\u983b\u8b5c \u7684\u6a19\u6e96\u5dee\uff0c ,a \u03bc \u70ba\u5168\u90e8\u8a13\uf996\u8a9e\uf906\u4e4b\u5206\u983b\u6bb5\u5f37\ufa01\u983b\u8b5c\u7684\u5e73\u5747\u503c\uff0c ,a \u03c3 \u70ba\u5168\u90e8\u8a13\uf996\u8a9e\uf906\u4e4b\u5206 \u983b\u6bb5\u5f37\ufa01\u983b\u8b5c\u7684\u6a19\u6e96\u5dee\u3002 \u7d71\u8a08\u5716\u7b49\u5316\u6cd5(HEQ)\u5728\u6b64\u7684\u8a08\u7b97\u65b9\u5f0f\u4ee5\u5f0f(2-7)\u8868\u793a\uff1a ( ) ( ) ' 1 ' , , a s A k F F A k \u2212 \u23a1 \u23a4 \u23a1 \u23a4 = \u23a3 \u23a6 \u23a3 \u23a6 \u5f0f(2-7) \u5176\u4e2d ( ) ,s F i \u70ba\u55ae\u4e00\u8a9e\uf906\u4e4b\u5206\u983b\u6bb5\u5f37\ufa01\u983b\u8b5c\u7684\u6a5f\uf961\u5206\u4f48\uff0c ( ) ,a F i \u70ba\u5168\u90e8\u8a13\uf996\u8a9e\uf906\u4e4b\u5206\u983b\u6bb5 \u5f37\ufa01\u983b\u8b5c\u7684\u6a5f\uf961\u5206\u4f48\u3002 5. \u5728\u8655\uf9e4\u5b8c\u6bcf\u4e00\u983b\u6bb5\u4e4b\u5f8c\uff0c\u6211\u5011\u5c07\u5404\u983b\u6bb5\u7684\u5f37\ufa01\u983b\u8b5c { } ' A k \u23a1 \u23a4 \u23a3 \u23a6 \u7167\u5176\u983b\uf961\u5927\u5c0f\u9806\u5e8f\u91cd\u65b0\uf905 \u63a5\u8d77\uf92d\uff0c\u5f97\u5230\u65b0\u7684\u5168\u983b\u6bb5\u5f37\ufa01\u983b\u8b5c [ ];0 2 N A k k \u23a7 \u23ab \u23a1 \u23a4 \u23aa \u23aa \u23aa \u23aa \u2264 \u2264 \u23a2 \u23a5 \u23a8 \u23ac \u23aa \u23aa \u23a2 \u23a5 \u23a2 \u23a5 \u23aa \u23aa \u23a9 \u23ad \uff0c\u6b64\u5373\u70ba\u7d71\u8a08\u6b63\u898f\u5316\u6cd5\u8655\uf9e4\u5f8c\u7684 \u8abf\u8b8a\u983b\u8b5c\u4e4b\u5f37\ufa01\u6210\u4efd\uff0c\u63a5\u8457\u5c07 [ ] { } A k \u88dc\u56de\u5f0f(2-3)\u4e2d\u7684\u539f\u672c\u76f8\u4f4d\u6210\u5206 [ ] { } k \u03b8 \uff0c\u518d\u7d93\u9006\u8f49\u63db \uf9ea\u6563\u5085\uf9f7\uf96e\u8f49\u63db(inverse discrete Fourier transform, IDFT)\u6240\u5f97\u65b0\u7684\u7279\u5fb5 [ ] x n \uff0c\u5982\u4e0b\u5f0f(2-8) \u8868\u793a\uff1a [ ] [ ] [ ] ( ) 2 1 0 1 , 0 1. nk N j j k N k x n A k e e n N N \u03c0 \u03b8 \u2212 = = \u2264\u2264 \u2212 \u2211 \u5f0f(2-8) \u7531 \u65bc \u7279 \u5fb5 \u5e8f \uf99c \u7d93 \u5085 \uf9f7 \uf96e \u8f49 \u63db \u5f8c \uff0c \u5177 \u6709 \u5de6 \u53f3 \u5c0d \u7a31 \u7684 \u7279 \u6027 \uff0c \u5373 [ ] [ ] A k A N k = \u2212 \u8207 [ ] [ ] k N k \u03b8 \u03b8 = \u2212 \u2212 \uff0c \u56e0 \u6b64 \u6211 \u5011 \u53ef \u85c9 \u6b64 \u63a8 \u5f97 \u5f0f (2-8) \u6240 \u9700 \u7528 \u5230 \u7684 [ ] { } A k \u8207 [ ] { } k \u03b8 \u5728 1 2 N k N \u23a1 \u23a4 < \u2264 \u2212 \u23a2 \u23a5 \u23a2 \u23a5 \u23a2 \u23a5 \u7684\u6bcf\u4e00\u9805\u3002 \u5728\u6b65\u9a5f 2 \u4e2d\uff0c\uf974\u6211\u5011\u672a\u5c0d\u8abf\u8b8a\u983b\u8b5c\u7684\u8a9e\u97f3\u7279\u5fb5\u4f5c\u5206\u983b\u6bb5\u8655\uf9e4\uff0c\u5373\u5206\u6bb5\uf969 1 L = \uff0c\u63a5 \u8457\u5728\u6b65\u9a5f 3", |
| "cite_spans": [], |
| "ref_spans": [], |
| "eq_spans": [], |
| "section": "", |
| "sec_num": null |
| }, |
| { |
| "text": "\u4e4b SHE \u6280\u8853\uff0c\u6211\u5011\u5247\u4ee5 FB-SHE \uf92d\u8868\u793a\u3002\u4ee5\u4e0b\u5c07\u91dd\u5c0d\u9019\u4e9b\u5206\u983b\u6bb5\u6b63\u898f\u5316\u6cd5\u7684\u7279\u9ede\u52a0\u4ee5 \u8a0e\uf941\uff1a (1) \u7d93\u7531 SB-SMN \u8207 SB-SMVN \u7684\u65b9\u6cd5\u8655\uf9e4\u4e4b\u5f8c\uff0c\u6240\u5f97\u7684\u8abf\u8b8a\u983b\u8b5c\u5f37\ufa01\u4e4b\u90e8\u4efd\uf969\u503c\u53ef\u80fd \u70ba\u8ca0\u503c\uff0c\u6b64\u660e\u986f\u9055\u53cd\u983b\u8b5c\u5f37\ufa01\u5fc5\u7136\u975e\u8ca0\u7684\u689d\u4ef6\uff0c\u56e0\u6b64\u7576\u8ca0\u503c\u7684\u60c5\u5f62\u51fa\u73fe\u6642\uff0c\u6211\u5011\u5c07\u5176\u503c \u91cd\u8a2d\u70ba 0\u3002 (2) \u5728 SB-SMN \u8207 SB-SMVN \u65b9\u6cd5\u4e2d\uff0c\uf967\u540c\u983b\u6bb5\u5177\u6709\u5404\u81ea\u7684\u76ee\u6a19\u5e73\u5747\u503c\u6216\u76ee\u6a19\u8b8a\uf962\uf969\uff0c \u540c\u6a23\u5730\uff0c\u5728 SB-SHE \u4e2d\uff0c\uf967\u540c\u983b\u6bb5\u4f7f\u7528\uf967\u540c\u7684\u76ee\u6a19\u6a5f\uf961\u5206\u4f48\u4f5c\u6b63\u898f\u5316\u904b\u7b97\u3002\u9019\u6a23\u7684\u4f5c\u6cd5\uff0c \u53ef\u4ee5\u4fdd\uf9cd\uf967\u540c\u983b\u6bb5\u7684\u983b\u8b5c\u5f37\ufa01\u4e4b\u9593\u5dee\uf962\u6027\u3002 (3) \u5728\u9019\u4e9b\u5206\u983b\u6bb5\u8abf\u8b8a\u983b\u8b5c\u6b63\u898f\u5316\u6cd5\u4e2d\uff0c\u5168\u983b\u6bb5\u7684\u9577\ufa01\u7b49\u65bc\u5404\u5206\u983b\u6bb5\u4e4b\u9577\ufa01\u7684\u548c\uff0c\u6240\u4ee5 \u589e\u52a0\u5206\u983b\u6bb5\u7684\uf969\u76ee\u4e26\uf967\u6703\u660e\u986f\u589e\u52a0\u904b\u7b97\u4e0a\u7684\u8907\u96dc\ufa01\u3002\u7136\u800c\u5b50\u983b\u6bb5\u7684\uf969\u76ee\uf967\u80fd\u904e\u591a\uff0c\u5426\u5247 \u5728\u4f4e\u983b\u7684\u5b50\u983b\u6bb5\u7684 [ ] A k \u9805\uf969\u5c07\u904e\u5c11\u751a\u81f3\u70ba\uf9b2\uff0c\u5982\u6b64\u660e\u986f\u6703\u5f71\u97ff\u55ae\u4e00\u983b\u6bb5\u6240\u6c42\u53d6\u4e4b\u7d71\u8a08 \u503c(\u5982\u5e73\u5747\u503c\u3001\u8b8a\uf962\uf969\u8207\u6a5f\uf961\u5206\u4f48\u7b49)\u7684\u7cbe\u78ba\u6027\u3002\u8209\uf9b5\uf96f\u660e\uff0c\u5047\u8a2d\u55ae\u4e00\u8a9e\u97f3\u7279\u5fb5\u5e8f\uf99c\u7684 \u7e3d\u9ede\uf969\u70ba N \u9ede\u6642\uff0c\u7531\u65bc\u6b64(\u5be6\uf969)\u7279\u5fb5\u5e8f\uf99c\u7d93 N \u9ede\u5085\uf9f7\uf96e\u8f49\u63db\u5f8c\uff0c\u5176\u983b\u8b5c\u7684\u5f37\ufa01\u6210\u4efd \u5177\u6709\u5de6\u53f3\u5c0d\u7a31\u7684\u7279\u6027\uff0c\u56e0\u6b64\u5be6\u969b\u4f7f\u7528\u7684\u983b\u8b5c\u9ede\uf969\u70ba\u7e3d\uf969\u7684\u4e00\u534a\uff0c\u5373\u70ba 2 N \u23a1 \u23a4 \u23a2 \u23a5 \u23a2 \u23a5 \u23a2 \u23a5 \u9ede\uff0c\u5982\u679c\u6211\u5011 \u4ee5\uf967\u7b49\ufa00\u7684\u65b9\u5f0f\ufa00\u5272\u6574\u500b\u983b\u5e36\uff0c\u6240\u5f97\u7684\u983b\u6bb5\uf967\u80fd\u7121\u9650\u5236\u5730\u589e\u591a\uff1b\uf9b5\u5982\u7576\u6211\u5011\ufa00 L \u500b\u5b50\u983b \u6bb5\u6642\uff0c\u6240\u5f97\u7684\u6bcf\u500b\u5b50\u983b\u6bb5\u9ede\uf969\u7531\u591a\u5230\u5c11\u5206\u5225\u70ba\uff1a ( ) 1 , ,..., 4 8 2 L N N N + \u23a2 \u23a5 \u23a2 \u23a5 \u23a2 \u23a5 \u23a2 \u23a5 \u23a2 \u23a5 \u23a2 \u23a5 \u23a2 \u23a5 \u23a2 \u23a5 \u23a2 \u23a5 \u23a3 \u23a6 \u23a3 \u23a6 \u23a3 \u23a6 \uff0c\u6240\u4ee5\u70ba\uf9ba\u6eff\u8db3\u6700 \u5c11\u7684\u90a3\u4e00\u500b\u5b50\u983b\u6bb5\u7684\u9ede\uf969\uf967\u70ba\uf9b2\uff0c\u5373\u6bcf\u500b\u983b\u6bb5\u7684\u8cc7\uf9be\uf97e\u81f3\u5c11\u6709\u4e00\u9ede\uff0c\u6211\u5011\u9808\u6eff\u8db3 ( ) 1 2 L N + \u2265 \u7684\u689d\u4ef6\uff0c\u7531\u6b64\u63a8\u77e5\uff0c\uf974 60 N = \uff0c\u6700\u591a\u53ea\u80fd\ufa00 5 \u500b\u983b\u6bb5\uff0c\u800c\uf974 30 N = \uff0c\u5247\u6700 \u591a\u53ea\u80fd\ufa00 4 \u500b\u5b50\u983b\u6bb5\uff0c\uf974\u9032\u4e00\u6b65\u8981\u6c42\uf974\u8981\u6c42\u6bcf\u500b\u5b50\u983b\u6bb5\u9ede\uf969\uf967\u80fd\u592a\u5c11\uff0c\u5247\u5b50\u983b\u6bb5\uf969\u76ee\u9650 \u5236\u5c07\u6703\uf901\u56b4\u683c\u3002 (\u4e8c)", |
| "cite_spans": [], |
| "ref_spans": [], |
| "eq_spans": [], |
| "section": "", |
| "sec_num": null |
| } |
| ], |
| "back_matter": [], |
| "bib_entries": { |
| "BIBREF1": { |
| "ref_id": "b1", |
| "title": "Cepstral analysis technique for automatic speaker verification", |
| "authors": [ |
| { |
| "first": "Sadaoki", |
| "middle": [], |
| "last": "Furui", |
| "suffix": "" |
| } |
| ], |
| "year": 1981, |
| "venue": "IEEE Trans. on Acoustics, Speech and Signal Processing", |
| "volume": "", |
| "issue": "", |
| "pages": "254--272", |
| "other_ids": {}, |
| "num": null, |
| "urls": [], |
| "raw_text": "Sadaoki Furui, \"Cepstral analysis technique for automatic speaker verification\", IEEE Trans. on Acoustics, Speech and Signal Processing, pp.254-272, 1981", |
| "links": null |
| }, |
| "BIBREF2": { |
| "ref_id": "b2", |
| "title": "Cepstral domain segmental feature vector normalization for noise robust speech recognition", |
| "authors": [ |
| { |
| "first": "Olli", |
| "middle": [], |
| "last": "Viikki", |
| "suffix": "" |
| }, |
| { |
| "first": "Kari", |
| "middle": [], |
| "last": "Laurila", |
| "suffix": "" |
| } |
| ], |
| "year": 1998, |
| "venue": "Speech Communication", |
| "volume": "25", |
| "issue": "", |
| "pages": "133--147", |
| "other_ids": {}, |
| "num": null, |
| "urls": [], |
| "raw_text": "Olli Viikki and Kari Laurila, ''Cepstral domain segmental feature vector normalization for noise robust speech recognition '', Speech Communication, vol. 25, pp.133-147, 1998", |
| "links": null |
| }, |
| "BIBREF3": { |
| "ref_id": "b3", |
| "title": "Histogram equalization of speech representation for robust speech recognition", |
| "authors": [ |
| { |
| "first": "Antonio", |
| "middle": [ |
| "M" |
| ], |
| "last": "\u00c1ngel De La Torre", |
| "suffix": "" |
| }, |
| { |
| "first": "Jos\u00e9", |
| "middle": [ |
| "C" |
| ], |
| "last": "Peinado", |
| "suffix": "" |
| }, |
| { |
| "first": "", |
| "middle": [], |
| "last": "Segura", |
| "suffix": "" |
| }, |
| { |
| "first": "L", |
| "middle": [], |
| "last": "Jos\u00e9", |
| "suffix": "" |
| }, |
| { |
| "first": "Ma", |
| "middle": [ |
| "Carmen" |
| ], |
| "last": "P\u00e9rez-C\u00f3rdoba", |
| "suffix": "" |
| }, |
| { |
| "first": "Antonio", |
| "middle": [ |
| "J" |
| ], |
| "last": "Ben\u00edtez", |
| "suffix": "" |
| }, |
| { |
| "first": "", |
| "middle": [], |
| "last": "Rubio", |
| "suffix": "" |
| } |
| ], |
| "year": 2005, |
| "venue": "IEEE Trans. on Speech and Audio Processing", |
| "volume": "", |
| "issue": "", |
| "pages": "355--366", |
| "other_ids": {}, |
| "num": null, |
| "urls": [], |
| "raw_text": "\u00c1ngel de la Torre, Antonio M. Peinado, Jos\u00e9 C. Segura, Jos\u00e9 L. P\u00e9rez-C\u00f3rdoba, Ma Carmen Ben\u00edtez, and Antonio J. Rubio, \"Histogram equalization of speech representation for robust speech recognition\", IEEE Trans. on Speech and Audio Processing, pp.355-366, 2005", |
| "links": null |
| }, |
| "BIBREF4": { |
| "ref_id": "b4", |
| "title": "Modulation spectrum equalization for robust speech recognition", |
| "authors": [ |
| { |
| "first": "Chang-Wen", |
| "middle": [], |
| "last": "Liang-Che Sun", |
| "suffix": "" |
| }, |
| { |
| "first": "Lin-Shan", |
| "middle": [], |
| "last": "Hsu", |
| "suffix": "" |
| }, |
| { |
| "first": "", |
| "middle": [], |
| "last": "Lee", |
| "suffix": "" |
| } |
| ], |
| "year": 2007, |
| "venue": "Proc. IEEE Workshop on Automatic Speech Recognition & Understanding (ASRU)", |
| "volume": "", |
| "issue": "", |
| "pages": "81--86", |
| "other_ids": {}, |
| "num": null, |
| "urls": [], |
| "raw_text": "Liang-che Sun, Chang-wen Hsu and Lin-shan Lee, \"Modulation spectrum equalization for robust speech recognition\", in Proc. IEEE Workshop on Automatic Speech Recognition & Understanding (ASRU), pp.81-86, 2007", |
| "links": null |
| }, |
| "BIBREF5": { |
| "ref_id": "b5", |
| "title": "RASTA processing of speech", |
| "authors": [ |
| { |
| "first": "Hynek", |
| "middle": [], |
| "last": "Hermansky", |
| "suffix": "" |
| }, |
| { |
| "first": "Nelson", |
| "middle": [], |
| "last": "Morgan", |
| "suffix": "" |
| } |
| ], |
| "year": 1994, |
| "venue": "IEEE Trans. on Speech and Audio Processing", |
| "volume": "", |
| "issue": "", |
| "pages": "578--589", |
| "other_ids": {}, |
| "num": null, |
| "urls": [], |
| "raw_text": "Hynek Hermansky and Nelson Morgan, \"RASTA processing of speech\", IEEE Trans. on Speech and Audio Processing, pp.578-589, 1994", |
| "links": null |
| }, |
| "BIBREF6": { |
| "ref_id": "b6", |
| "title": "Multi-resolution RASTA filtering for TANDEM-based ASR", |
| "authors": [ |
| { |
| "first": "Hynek", |
| "middle": [], |
| "last": "Hermansky", |
| "suffix": "" |
| }, |
| { |
| "first": "Petr", |
| "middle": [], |
| "last": "Fousek", |
| "suffix": "" |
| } |
| ], |
| "year": null, |
| "venue": "2005 International Conference on Spoken Language Processing (Interspeech)", |
| "volume": "", |
| "issue": "", |
| "pages": "361--364", |
| "other_ids": {}, |
| "num": null, |
| "urls": [], |
| "raw_text": "Hynek Hermansky and Petr Fousek, \"Multi-resolution RASTA filtering for TANDEM-based ASR\", 2005 International Conference on Spoken Language Processing (Interspeech), pp.361-364", |
| "links": null |
| }, |
| "BIBREF7": { |
| "ref_id": "b7", |
| "title": "On the importance of various modulation frequencies for speech recognition", |
| "authors": [ |
| { |
| "first": "Noboru", |
| "middle": [], |
| "last": "Kanedera", |
| "suffix": "" |
| }, |
| { |
| "first": "Takayuki", |
| "middle": [], |
| "last": "Arai", |
| "suffix": "" |
| }, |
| { |
| "first": "Hynek", |
| "middle": [], |
| "last": "Hermansky", |
| "suffix": "" |
| }, |
| { |
| "first": "Misha", |
| "middle": [], |
| "last": "Pavel", |
| "suffix": "" |
| } |
| ], |
| "year": 1997, |
| "venue": "European Conference on Speech Communication and Technology", |
| "volume": "", |
| "issue": "", |
| "pages": "1079--1082", |
| "other_ids": {}, |
| "num": null, |
| "urls": [], |
| "raw_text": "Noboru Kanedera, Takayuki Arai, Hynek Hermansky, and Misha Pavel, \"On the importance of various modulation frequencies for speech recognition\", 1997 European Conference on Speech Communication and Technology (Eurospeech), pp. 1079-1082", |
| "links": null |
| }, |
| "BIBREF8": { |
| "ref_id": "b8", |
| "title": "The AURORA experimental framework for the performance evaluations of speech recognition systems under noisy conditions", |
| "authors": [ |
| { |
| "first": "David", |
| "middle": [], |
| "last": "Pearce", |
| "suffix": "" |
| }, |
| { |
| "first": "Hans-G\u00fcnter", |
| "middle": [], |
| "last": "Hirsch", |
| "suffix": "" |
| } |
| ], |
| "year": 2000, |
| "venue": "Proc. of ISCA IIWR ASR2000", |
| "volume": "", |
| "issue": "", |
| "pages": "181--188", |
| "other_ids": {}, |
| "num": null, |
| "urls": [], |
| "raw_text": "David Pearce and Hans-G\u00fcnter Hirsch, \"The AURORA experimental framework for the performance evaluations of speech recognition systems under noisy conditions\", in Proc. of ISCA IIWR ASR2000, Paris, France, pp.181-188, 2000", |
| "links": null |
| }, |
| "BIBREF9": { |
| "ref_id": "b9", |
| "title": "Transmission performance characteristics of pulse code modulation channels", |
| "authors": [], |
| "year": 1996, |
| "venue": "", |
| "volume": "", |
| "issue": "", |
| "pages": "", |
| "other_ids": {}, |
| "num": null, |
| "urls": [], |
| "raw_text": "ITU recommendation G.712, \"Transmission performance characteristics of pulse code modulation channels\", Nov. 1996", |
| "links": null |
| }, |
| "BIBREF10": { |
| "ref_id": "b10", |
| "title": "Probability and random processes with applications to signal processing", |
| "authors": [ |
| { |
| "first": "Henry", |
| "middle": [], |
| "last": "Stark", |
| "suffix": "" |
| }, |
| { |
| "first": "John", |
| "middle": [ |
| "W" |
| ], |
| "last": "Woods", |
| "suffix": "" |
| } |
| ], |
| "year": 2002, |
| "venue": "", |
| "volume": "", |
| "issue": "", |
| "pages": "", |
| "other_ids": {}, |
| "num": null, |
| "urls": [], |
| "raw_text": "Henry Stark, John W. Woods, \"Probability and random processes with applications to signal processing\", 3 rd Edition, Prentice-Hall, 2002", |
| "links": null |
| } |
| }, |
| "ref_entries": { |
| "FIGREF0": { |
| "text": "\u4f5c\u7d71\u8a08\u5716\u7b49\u5316\u6cd5(HEQ)\u7684\u6b63\u898f\u5316\u904b\u7b97\uff0c\u9019\u6a23\u7684\u904b\u7b97\u65b9\u5f0f\u76f8\u7576\u65bc[5]\u4e2d\u7684\u8abf\u8b8a\u983b \u8b5c\u7d71\u8a08\u5716\u7b49\u5316\u6cd5(SHE)\uff1b\u70ba\uf9ba\u4e4b\u5f8c\u8a0e\uf941\u65b9\uf965\u8d77\ufa0a\uff0c\u6211\u5011\u5c07\u4e0a\u8ff0\u5f0f(2-5)\u3001\u5f0f(2-6)\u3001\u5f0f(2-7) \u7684\u6b63\u898f\u5316\u65b9\u6cd5\u8655\uf9e4\uff0c\u5206\u5225\u547d\u540d\u70ba\uff1a\u5206\u983b\u6bb5\u8abf\u8b8a\u983b\u8b5c\u5e73\u5747\u503c\u6b63\u898f\u5316\u6cd5(sub-band spectral mean normalization, SB-SMN)\u3001\u5206\u983b\u6bb5\u8abf\u8b8a\u983b\u8b5c\u5e73\u5747\u503c\u8207\u8b8a\uf962\uf969\u6b63\u898f\u5316\u6cd5(sub-band spectral mean and variance normalization, SB-SMVN)\u8207\u5206\u983b\u6bb5\u8abf\u8b8a\u983b\u8b5c\u7d71\u8a08\u5716\u7b49\u5316\u6cd5 (sub-band spectral histogram equalization, SB-SHE)\uff0c\u800c\u6587\u737b[5]\u4e2d\u6240\u7528\u7684\u5168\u983b\u5e36(full-band)", |
| "num": null, |
| "type_str": "figure", |
| "uris": null |
| }, |
| "FIGREF1": { |
| "text": "\u5206\u983b\u6bb5\u8abf\u8b8a\u983b\u8b5c\u6b63\u898f\u5316\u6cd5\u5176\u521d\u6b65\u6548\u80fd\u7684\u8a0e\uf941\uff1a \u5728\u9019\uf9e8\uff0c\u6211\u5011\u5c07\u63a2\u8a0e\u672c\u7ae0\u6240\u63d0\u51fa\u7684\u4e09\u7a2e\u5206\u983b\u6bb5\u8abf\u8b8a\u983b\u8b5c\u6b63\u898f\u5316\u6cd5\u5728\u7279\u5fb5\u5e8f\uf99c\u4e4b\u529f\uf961 \u983b\u8b5c\u5bc6\ufa01(power spectral density, PSD)\ufa09\u4f4e\u5931\u771f\u7684\u6548\u679c\uff0c\u540c\u6642\uff0c\u628a\u9019\u4e9b\u65b9\u6cd5\u5448\u73fe\u7684\u7d50\u679c\uff0c \u548c\u7b2c\u4e8c\u7ae0\u6240\u4ecb\u7d39\u4e4b\u7279\u5fb5\u6642\u9593\u5e8f\uf99c\u57df(temporal domain)\u4e4b\u6b63\u898f\u5316\u6280\u8853\uff1a\u5012\u983b\u8b5c\u5e73\u5747\u503c\u6b63\u898f \u5316\u6cd5(CMN)\u3001\u5012\u983b\u8b5c\u5e73\u5747\u503c\u8207\u8b8a\uf962\uf969\u6b63\u898f\u5316\u6cd5(CMVN)\u3001\u7d71\u8a08\u5716\u7b49\u5316\u6cd5(HEQ)\u5206\u5225\u4f5c\u6bd4 \u8f03\u3002\u6211\u5011\uf9dd\u7528\uf9ba AURORA-2 \u8cc7\uf9be\u5eab[9]\uf9e8 MIP_28826Z4A \u8a9e\u97f3\u6a94\uff0c\u52a0\u5165\uf967\u540c\u8a0a\u96dc\u6bd4(SNR) \u7684\u4eba\u8072(babble)\u96dc\u8a0a\uff0c\u518d\u7d93\u5404\u7a2e\u6b63\u898f\u5316\u6cd5\u52a0\u4ee5\u8655\uf9e4\uff0c\u6700\u5f8c\u6c42\u53d6\u5176\u529f\uf961\u983b\u8b5c\u5bc6\ufa01\u3002\u9996\u5148\uff0c \u5716\u4e00\u70ba\u4e00\u7cfb\uf99c\u4e4b\u5e73\u5747\u503c\u6b63\u898f\u5316\u6cd5(MN)\u4f5c\u7528\u65bc\u7b2c\u4e00\u7dad\u5012\u983b\u8b5c\u7279\u5fb5(the first cepstral coefficient, c 1 )\u5e8f\uf99c\u6240\u5f97\u4e4b\u529f\uf961\u983b\u8b5c\u5bc6\ufa01\u5716\u3002\u5728\u5716\u4e00\u4e2d\uff0c\u85c9\u7531\u5716(a)\u6211\u5011\u767c\u73fe\uff0c\u96dc\u8a0a\u7684\u5b58\u5728 \u4f7f\u4e7e\u6de8\u8a9e\u97f3\u8207\u96dc\u8a0a\u8a9e\u97f3\u7522\u751f\u660e\u986f\u7684 PSD \u5931\u771f\uff0c\u800c\u5716(b)\u4e2d\u6240\u7528 CMN \u6cd5\uff0c\u5373\u6642\u57df\u578b MN \u6cd5(temporal CMN)\uff0c\u53ef\u7a0d\u5fae\ufa09\u4f4e\u6b64\u5931\u771f\uff0c\u800c\u983b\u57df\u578b\u7684 MN \u6cd5\u4e2d\uff0c\u7531\u5716(c)\u81f3\u5716(f)\u767c\u73fe,\u5c07 \u5168\u983b\u5e36\u9010\u6f38\u7d30\u5206\u81f3 2 \u5230 4 \u500b\u5b50\u983b\u6bb5(\u5206\u5225\u4ee5 SB-SMN (2) , SB-SMN (3) \u8207 SB-SMN (4) \u8868\u793a\uff0c \u4e0b\u6a19\u62ec\u865f\u4e2d\u7684\uf969\u5b57\u8868\u793a\u5206\u983b\u6bb5\u7684\u500b\uf969) \uff0c\u6b64 PSD \u5931\u771f\u9010\u6f38\ufa09\u4f4e\uff0c\u5176\u4e2d\u4ee5\u5716(f)\u7d93 SB-SMN (4) \u8655\uf9e4\u5f8c\uff0cPSD \u7684\u5931\u771f\u7a0b\ufa01\u6700\u5c0f\u3002\u7531\u6b64\uf96f\u660e\u5206\u983b\u6bb5\u8abf\u8b8a\u983b\u8b5c\u5e73\u5747\u503c\u6b63\u898f\u5316\u6cd5\u5c0d\u65bc\ufa09\u4f4e\u56e0\u96dc \u8a0a\u6240\u9020\u6210\u7684 PSD \u5931\u771f\u6709\u660e\u986f\u7684\u5e6b\u52a9\u3002 (a) unprocessed c1 (b) Temporal-CMN (c) FB-SMN (d) SB-SMN (2) (e) SB-SMN (3) (f) SB-SMN (4) \u5716\u4e00 \u5e73\u5747\u503c\u6b63\u898f\u5316\u6cd5\u4f5c\u7528\u65bc\uf967\u540c\u8a0a\u96dc\u6bd4\u4e0b\u8a9e\u97f3\u4e4b\u539f\u59cb 1 c \u7279\u5fb5\u5e8f\uf99c\uff0c\u5176\u8abf\u8b8a\u983b\u8b5c\u66f2\u7dda \u5716\uff1a(a)\u539f\u59cb 1 c \u7279\u5fb5\u5e8f\uf99c\uff0c(b)\u6642\u57df\u578b MN \u6cd5-CMN\uff0c(c)\u983b\u57df\u578b\u4e4b\u5168\u983b\u5e36 MN \u6cd5-FB-SMN\uff0c (d)\u983b\u57df\u578b\u4e4b\u5206\u983b\u6bb5 MN \u6cd5-SB-SMN (2) \uff0c(e)\u983b\u57df\u578b\u4e4b\u5206\u983b\u6bb5 MN \u6cd5-SB-SMN (3) \uff0c(f)\u983b \u57df\u578b\u4e4b\u5206\u983b\u6bb5 MN \u6cd5-SB-SMN (4) \u63a5\u8457\uff0c\u5716\u4e8c\u70ba\u4e00\u7cfb\uf99c\u4e4b\u5e73\u5747\u503c\u8207\u8b8a\uf962\uf969\u6b63\u898f\u5316\u6cd5(MVN)\u4f5c\u7528\u65bc\u7b2c\u4e00\u7dad\u5012\u983b\u8b5c\u7279\u5fb5(the first cepstral coefficient, c 1 )\u5e8f\uf99c\u6240\u5f97\u4e4b\u529f\uf961\u983b\u8b5c\u5bc6\ufa01\u5716\u3002\u5728\u5716\u4e8c\u4e2d\uff0c\u85c9\u7531\u5716(b)\u6211\u5011\u767c\u73fe\uff0c \u50b3\u7d71\u7684 CMVN \u6cd5\uff0c\u5373\u6642\u57df\u578b MVN \u6cd5(temporal CMVN)\uff0c\u76f8\u8f03\u65bc\u5716\u4e00\u7684\u5716(b)\u4e4b CMN \u800c \u8a00\uff0c\ufa09\u4f4e PSD \u5931\u771f\u7684\u6548\u61c9\uf901\u597d\uff0c\u610f\u5473\uf9ba\u984d\u5916\u8655\uf9e4\u7279\u5fb5\u7684\u8b8a\uf962\uf969\u78ba\u5be6\u662f\u6709\u5e6b\u52a9\u7684\u3002\u800c\u5728 \u5404\u7a2e\u983b\u57df\u578b\u7684 MVN \u6cd5\u4e2d\uff0c\u7531\u5716(c)\u81f3\u5716(f)\u767c\u73fe\uff0c\uf9d0\u4f3c MN \u6cd5\u7684\u6548\u679c\uff0c\u7576\u6211\u5011\u5c07\u5168\u983b\u5e36\u9010 \u6f38\u7d30\u5206\u81f3 2 \u5230 4 \u500b\u5b50\u983b\u6bb5(\u5206\u5225\u4ee5 SB-SMVN (2) , SB-SMVN (3) \u8207 SB-SMVN (4) \u8868\u793a\uff0c\u4e0b\u6a19 \u62ec\u865f\u4e2d\u7684\uf969\u5b57\u8868\u793a\u5206\u983b\u6bb5\u7684\u500b\uf969) \uff0cPSD \u5931\u771f\u4e5f\u9010\u6f38\ufa09\u4f4e\uff0c\u5176\u4e2d\u4ee5\u5716(f)\u7d93 SB-SMVN (4) \u8655\uf9e4\u5f8c\uff0c\u5c0d\u65bc PSD \u7684\u5931\u771f\u7684\ufa09\u4f4e\u6548\u679c\u6700\u597d\u3002\u7531\u6b64\u4ea6\uf96f\u660e\uf9ba\u5206\u983b\u6bb5\u8abf\u8b8a\u983b\u8b5c\u5e73\u5747\u503c\u8207\u8b8a \uf962\uf969\u6b63\u898f\u5316\u6cd5\u5c0d\u65bc\ufa09\u4f4e\u56e0\u96dc\u8a0a\u6240\u9020\u6210\u7684 PSD \u5931\u771f\u4e5f\u6709\u660e\u986f\u5e6b\u52a9\uff0c\u540c\u6642\u5c07\u5716\u4e8c\u8207\u5716\u4e00\u6bd4 Hz Hz Hz Hz Hz Hz \u8f03\u5f8c\uff0c\u53ef\u4ee5\u660e\u986f\u770b\u51fa MVN \u6cd5\u5728\ufa09\u4f4e PSD \u5931\u771f\u7684\u6548\u80fd\u4e0a\u512a\u65bc MN \u6cd5\uff0c\u6b64\u543b\u5408\u6211\u5011\u4e00\u822c \u5c0d\u9019\uf978\uf9d0\u65b9\u6cd5\u4e4b\u6548\u80fd\u7684\u8a8d\u77e5\u3002 (a) un-processed c 1 (b) Temporal-CMVN (c) FB-SMVN (d) SB-SMVN (2) (e) SB-SMVN (3) (f) SB-SMVN (4) \u5716\u4e8c \u5e73\u5747\u503c\u8207\u8b8a\uf962\uf969\u6b63\u898f\u5316\u6cd5\u4f5c\u7528\u65bc\uf967\u540c\u8a0a\u96dc\u6bd4\u4e0b\u8a9e\u97f3\u4e4b\u539f\u59cb 1 c \u7279\u5fb5\u5e8f\uf99c\uff0c\u5176\u8abf\u8b8a\u983b \u8b5c\u66f2\u7dda\u5716\uff1a(a)\u539f\u59cb 1 c \u7279\u5fb5\u5e8f\uf99c\uff0c(b)\u6642\u57df\u578b MVN \u6cd5-CMVN\uff0c(c)\u983b\u57df\u578b\u4e4b\u5168\u983b\u5e36 MVN \u6cd5-FB-SMVN\uff0c (d)\u983b\u57df\u578b\u4e4b\u5206\u983b\u6bb5 MVN \u6cd5-SB-SMVN (2) \uff0c(e)\u983b\u57df\u578b\u4e4b\u5206\u983b\u6bb5 MVN \u6cd5-SB-SMVN (3) \uff0c(f)\u983b\u57df\u578b\u4e4b\u5206\u983b\u6bb5 MVN \u6cd5-SB-SMVN (4) \u6700\u5f8c\uff0c\u5716\u4e09\u70ba\u4e00\u7cfb\uf99c\u4e4b\u7d71\u8a08\u5716\u7b49\u5316\u6cd5(HEQ)\u4f5c\u7528\u65bc\u7b2c\u4e00\u7dad\u5012\u983b\u8b5c\u7279\u5fb5(the first cepstral coefficient, c1)\u5e8f\uf99c\u6240\u5f97\u4e4b\u529f\uf961\u983b\u8b5c\u5bc6\ufa01\u5716\u3002\u6211\u5011\u5c07\u5716\u4e09\u8207\u5716\u4e00\u548c\u5716\u4e8c\u6bd4\u8f03\uff0c\u53ef\u660e\u986f\u770b\u51fa \u6b63\u898f\u5316\u6574\u500b\u6a5f\uf961\u5206\u4f48\u7684 HEQ \u6cd5\uff0c\u660e\u986f\u5728\ufa09\u4f4e PSD \u7684\u5931\u771f\u4e0a\u512a\u65bc\u53ea\u6b63\u898f\u5316\u5e73\u5747\u503c\u7684 MN \u6cd5\u8207\u6b63\u898f\u5316\u5e73\u5747\u503c\u8207\u8b8a\uf962\uf969\u7684 MVN \u6cd5(\u7121\uf941\u662f\u6642\u57df\u578b\u6216\u983b\u57df\u578b\u7684\u7686\u662f\u5982\u6b64) \uff0c\u6b64\u5916\uff0c \u6211\u5011\uf974\u6bd4\u8f03\u4e09\u7a2e\u5168\u983b\u5f0f\u7684\u65b9\u6cd5( \u5716\u4e00(c) \u7684 FB-SMN, \u5716\u4e8c(c) \u7684 FB-SMVN \u8207\u5716\u4e09 (c)FB-SHE)\uff0c\u53ef\u767c\u73fe FB-SHE \u76f8\u5c0d\u65bc FB-SMN \u8207 FB-SMVN \u800c\u8a00\uff0c\u5f9e\u4f4e\u983b\u5230\u9ad8\u983b\u7684 PSD \u5931\u771f\u90fd\u6709\u660e\u986f\ufa09\u4f4e\uff0c\u800c\uf967\u662f\u50cf FB-SMN \u8207 FB-SMVN \u76f8\u5c0d\u53ea\u6709\u6e1b\u5c11\u4f4e\u983b\u6210\u5206\u7684 PSD \u5931 \u6240\u63d0\u4e4b FB-SHE \u7684\uf97c\u597d\u6548\u80fd\u3002\u7136\u800c\u5728\u6211\u5011\u6240\u63d0\u51fa\u7684\u5404 \u7a2e\u5206\u983b\u6bb5 SHE(SB-SHE)\u6cd5\u4e2d\uff0c\u660e\u986f\u770b\u51fa\u5b83\u5011\u7686\u6bd4 FB-SHE \u5728\u6e1b\u4f4e PSD \u5931\u771f\u7684\u6548\u80fd\uf92d\u5f97 \u597d\uff0c\u7531\u5716(c)\u81f3\u5716(f)\u767c\u73fe\uff0c\uf9d0\u4f3c\u4e4b\u524d MN \u8207 MVN \u6cd5\u7684\u6548\u679c\uff0c\u7576\u6211\u5011\u5c07\u983b\u6bb5\u5f9e\u5168\u983b\u6bb5\u9010\u6f38 \u7d30\u5206\u81f3 2 \u5230 4 \u500b\u983b\u6bb5(\u5206\u5225\u4ee5 SB-HEQ (2) , SB-HEQ (3) \u8207 SB-HEQ (4) \u8868\u793a\uff0c\u4e0b\u6a19\u62ec\u865f\u4e2d\u7684 \uf969\u5b57\u8868\u793a\u5206\u983b\u6bb5\u7684\u500b\uf969) \uff0cPSD \u5931\u771f\u9010\u6f38\ufa09\u4f4e\uff0c\u5176\u4e2d\u4ee5\u5716(f)\u7d93 SB-HEQ (4) \u7279\u5fb5\u5e8f\uf99c\uff0c(b)\u6642\u57df\u578b HEQ \u6cd5-Temporal HEQ\uff0c(c)\u983b\u57df\u578b\u4e4b\u5168\u983b\u5e36 HEQ \u6cd5-FB-HEQ\uff0c(d)\u983b\u57df\u578b\u4e4b\u5206\u983b\u6bb5 HEQ \u6cd5-SB-HEQ (2) \uff0c(e)\u983b\u57df\u578b\u4e4b\u5206\u983b\u6bb5 HEQ \u6cd5-SB-HEQ (3) \uff0c(f)\u983b\u57df\u578b\u4e4b\u5206\u983b\u6bb5 HEQ \u6cd5-SB-HEQ (4)", |
| "num": null, |
| "type_str": "figure", |
| "uris": null |
| }, |
| "TABREF1": { |
| "content": "<table><tr><td colspan=\"2\">Hz Hz Hz \u672c \uf941 \u6587 \u4f7f \u7528 \u7684 \u8a9e \u97f3 \u8cc7 \uf9be \u5eab \u70ba \u6b50 \u6d32 \u96fb \u4fe1 \u6a19 \u6e96 \u5354 \u6703 (European Telecommunication Hz Hz Hz (\u4e00) \u8a9e\u97f3\u8cc7\uf9be\u5eab\u7c21\u4ecb Standard Institute, ETSI)\u6240\u767c\ufa08\u7684 Aurora-2 \u8a9e\u97f3\u8cc7\uf9be\u5eab[9]\uff0c\u5b83\u662f\u4e00\u5957\u4ee5\u4eba\u5de5\u65b9\u5f0f\uf93f\u88fd\u7684 \uf99a\u7e8c\u82f1\u6587\uf969\u5b57\u5b57\uf905\uff0c\u8a9e\u8005\u7531\u7f8e\u570b\u6210\uf98e\u7537\uf981\u6240\u7d44\u6210\uff0c\u52a0\u4e0a\u516b\u7a2e\uf92d\u6e90\uf967\u540c\u7684\u96dc\u8a0a\uff0c\u5206\u5225\u70ba\uff1a \u5730\u4e0b\u9435\u3001\u4eba\u8072\u3001\u6c7d\uf902\u3001\u5c55\u89bd\u9928\u3001\u9910\u5ef3\u3001\u8857\u9053\u3001\u6a5f\u5834\u3001\u706b\uf902\u7ad9\u7b49\uff0c\u4e26\u4ee5\uf967\u540c\u7a0b\ufa01\u7684\u8a0a\u96dc\u6bd4 (signal-to-noise ratio, SNR)\u52a0\u5165\u96dc\u8a0a\uff0c\u5206\u5225\u70ba\uff1aclean\u300120 dB\u300115 dB\u300110 dB\u30015 dB\u30010 dB \u8207 -5 dB \uff1b \u5176 \u901a \u9053 \u6548 \u61c9 \u5206 \u5225 \u70ba G.712 \u8207 MIRS \uff0c \u5176 \u70ba \u570b \u969b \u96fb \u4fe1 \uf997 \u76df (International Telecommunication Union, ITU)[10]\u6240\u8a02\uf9f7\u7684\uf978\u500b\u901a\u9053\u6a19\u6e96\u3002 (\u4e8c) \u8a9e\u97f3\u7279\u5fb5\uf96b\uf969\u8a2d\u5b9a\u53ca\u8072\u5b78\u6a21\u578b \u672c\uf941\u6587\u4e4b\u76f8\u95dc\u8a9e\u97f3\u8fa8\uf9fc\u5be6\u9a57\u6240\u4f7f\u7528\u7279\u5fb5\uf96b\uf969\u70ba\u6885\u723e\u5012\u983b\u8b5c\u4fc2\uf969(MFCC)\uff0c\u9644\u52a0\u4e0a\u5176 \u4e00\u968e\u5dee\uf97e\u8207\u4e8c\u968e\u5dee\uf97e\uff0c\u5176\u8a73\u7d30\u8a9e\u97f3\u7279\u5fb5\uf96b\uf969\u8a2d\u5b9a\u5206\u5225\u5728\u8868\u4e00\u8868\u793a\u3002 \u53d6\u6a23\u983b\uf961 8 kHz \u97f3\u6846\u9577\ufa01(frame size) 25 ms, 200 \u9ede \u97f3\u6846\u5e73\u79fb(frame shift) 10 ms, 80 \u9ede \u9810\u5f37\u8abf\uf984\u6ce2\u5668 1 1 0.97z \u2212 \u2212 \u8996\u7a97\u5f62\u5f0f(window) \u6f22\u660e\u7a97(Hamming window) \u5085\uf9f7\uf96e\u8f49\u63db\u9ede\uf969(T) 256 \u9ede \uf984\u6ce2\u5668\u7d44(filter bank) \u6885\u723e\u523b\ufa01\u4e09\u89d2\uf984\u6ce2\u5668\u7d44\uff0c\u5171 23 \u500b\u4e09\u89d2\uf984\u6ce2\u5668 \u7279\u5fb5\u5411\uf97e(feature vector) MFCC 13 \u7dad(c1~c12, log-energy) + MFCC 13 \u7dad+ MFCC 13 \u7dad\uff0c\u5171 39 \u7dad \u8868\u4e00 \u672c\uf941\u6587\u6240\u4f7f\u7528\u8a9e\u97f3\u7279\u5fb5\uf96b\uf969\u8a2d\u5b9a \u6211 \u5011 \u662f \u4ee5 \u96b1 \u85cf \u5f0f \u99ac \u53ef \u592b \u6a21 \u578b (hidden Markov model, HMM)[11] \u4f5c \u70ba \u8072 \u5b78 \u6a21 \u578b (acoustic models)\u7684\u578b\u5f0f\u3002\u5305\u542b11\u500b\uf969\u5b57\u6a21\u578b(zero, one, two, \u2026, nine \u53caoh)\u4ee5\u53ca\u975c\u97f3 (silence)\u6a21\u578b\uff0c\u6bcf\u500b\uf969\u5b57\u6a21\u578b\u5305\u542b16\u500b\uf9fa\u614b\uff0c\u5404\uf9fa\u614b\u5305\u542b20\u500b\u9ad8\u65af\u5bc6\ufa01\u6df7\u5408\u3002 (\u4e09) \u8a9e\u97f3\u8fa8\uf9fc\u5be6\u9a57\u7d50\u679c \u5728\u9019\u4e00\u7bc0\u4e2d\uff0c\u6211\u5011\u5c07\u5404\u7a2e\u8abf\u8b8a\u983b\u8b5c\u6b63\u898f\u5316\u6cd5\u4e4b\u5be6\u9a57\u7d50\u679c\u7d9c\u5408\u6574\uf9e4\u6210\u8868\u4e8c\uff0c\u5176\u4e2d\u7d55\u5c0d \u932f\u8aa4\ufa09\u4f4e\uf961(absolute error rate reduction, AR)\u8207\u76f8\u5c0d\u932f\u8aa4\ufa09\u4f4e\uf961 1 (relative error rate reduction 1, RR 1 )\u5206\u5225\u70ba\u65b0\u8fa8\uf9fc\uf961\u8207\u57fa\u790e\u5be6\u9a57\u8fa8\uf9fc\uf961(baseline)\u6bd4\u8f03\u4e0b\uff0c\u6240\u5f97\u5230\u7684\u7d55\u5c0d\u6539\u5584 \uf961\u8207\u76f8\u5c0d\u6539\u5584\uf961\uff0c\u76f8\u5c0d\u932f\u8aa4\u6539\u5584\uf961 2 (relative error rate reduction 2, RR 2 )\uff0c\u5b83\u662f\u5206\u983b\u6bb5\u6280 \u8853\u76f8\u8f03\u65bc\u5168\u983b\u6bb5\u6280\u8853\u800c\u8a00\u6240\u5f97\u5230\u7684\u76f8\u5c0d\u932f\u8aa4\u6539\u5584\uf961\uff0c\u5176\u8a08\u7b97\u65b9\u5f0f\u5206\u5225\u7531\u5f0f(3-1)\u3001\u5f0f (3-2)\u3001\u5f0f(3-3)\u6240\u793a\uff1a ( ) ( ) % 100% AR = \u2212 \u00d7 \u65b0\u8fa8\u8b58\u7387 \u57fa\u790e\u5be6\u9a57\u8fa8\u8b58\u7387 \u5f0f(3-1) ( ) 1 % 100% 100% RR \u239b \u239e \u2212 \u239f \u239c = \u00d7 \u239f \u239c \u239f \u239f \u239c \u2212 \u239d \u23a0 \u65b0\u8fa8\u8b58\u7387 \u57fa\u790e\u5be6\u9a57\u8fa8\u8b58\u7387 \u57fa\u790e\u5be6\u9a57\u8fa8\u8b58\u7387 \u5f0f(3-2) \u7531\u8868\u4e8c\u89c0\u5bdf\u4e2d\uff0c\u6211\u5011\u53ef\u4ee5\u5f97\u5230\u4ee5\u4e0b\u5e7e\u9ede\u7d50\u679c\uff1a 1. \u6211\u5011\u6240\u65b0\u63d0\u51fa\u4e4b\u5404\u7a2e\u5206\u983b\u6bb5\u8abf\u8b8a\u983b\u8b5c\u6b63\u898f\u5316\u6cd5\u76f8\u8f03\u65bc\u57fa\u672c\u5be6\u9a57\u800c\u8a00\uff0c\u7686\u80fd\u4f7f\u8fa8\uf9fc\uf961 \u660e\u986f\u63d0\u5347\uff0c\u5f9e RR 1 \u9ad8\u968e\u7684\u52d5\u5dee\u4f5c\u6b63\u898f\u5316\u8655\uf9e4\uff0c\u4f7f\u5f97 SB-SHE \u6cd5\u6709\u8f03\u512a\uf962\u7684\u8868\u73fe\u3002 2. \u7531\u65bc\u8abf\u8b8a\u983b\u8b5c\u4e2d\u4f4e\u983b\u90e8\u5206(1~16Hz)\u4f54\u6709\u8f03\u591a\u91cd\u8981\u7684\u8a9e\u97f3\u6210\u4efd\uff0c\u6240\u4ee5\u6211\u5011\u8457\u91cd\u65bc\u5c07\u4f4e \u983b\u90e8\u5206\ufa00\u5272\u958b\uf92d\u5206\u5225\u4f5c\u6b63\u898f\u5316\u8655\uf9e4\uff0c\u5f9e\u8868\u4e8c\u53ef\u4ee5\u6e05\u695a\u767c\u73fe\uff0c\u7576\u4f4e\u983b\u90e8\u4efd\ufa00\u5272\u8d8a\u7d30\uff0c\u80fd\u6709 \u6548\u63d0\u5347\u8a9e\u97f3\u8fa8\uf9fc\u6548\u80fd\uff0c\u800c\u4e09\u7a2e\u5206\u983b\u6bb5\u8abf\u8b8a\u983b\u8b5c\u88dc\u511f\u6280\u8853\u7686\u4ee5\u5206\u5272\u56db\u500b\u983b\u6bb5\u7684\u6548\u679c\u6700\u70ba\u512a \u8d8a\u3002\u76f8\u5c0d\u65bc\u5168\u983b\u6bb5\u5f0f\u7684\u65b9\u6cd5\u800c\u8a00\uff0c\u5206\u983b\u6bb5\u5f0f\u7684\u65b9\u6cd5\u5176\u76f8\u5c0d\u932f\u8aa4\u6539\u5584\uf961(RR 2 )\u70ba\uff1aSB-SMN (4) \u7684 8.31%\uff0cSB-SMVN (4) \u7684 32.64%\uff0cSB-SHE (4) \u7684 7.56%\u3002 Method Set A Set B Set C average AR RR 1 RR 2 Baseline 71.98 67.79 78.28 71.56 \u2500 \u2500 \u2500 FB-SMN 77.43 76.26 78.05 77.08 5.52 19.41 \u2500 SB-SMN (2) 77.87 77.26 78.36 77.72 6.16 21.66 2.79 SB-SMN (3) 78.21 76.37 80.82 77.99 6.43 22.61 3.97 SB-SMN (4) 79.12 77.26 82.20 78.99 7.43 26.13 8.31 FB-SMVN 79.03 81.19 78.29 79.75 8.19 28.80 \u2500 SB-SMVN (2) 80.06 81.97 79.28 80.67 9.11 32.03 4.54 SB-SMVN (3) 80.84 82.59 80.89 81.55 9.99 35.13 8.89 SB-SMVN (4) 85.94 87.06 85.79 86.36 14.80 52.04 32.64 FB-SHE 89.71 90.03 88.27 89.55 17.99 63.26 \u2500 SB-SHE (2) 89.76 90.09 88.40 89.62 18.06 63.50 0.67 SB-SHE (3) 90.13 90.47 88.68 89.98 18.42 64.77 4.11 SB-SHE (4) 90.59 90.69 89.13 90.34 18.78 66.03 7.56 \u8868\u4e8c \u8abf\u8b8a\u983b\u8b5c\u7d71\u8a08\u6b63\u898f\u5316\u6cd5\u4e4b\u5be6\u9a57\u8fa8\uf9fc\uf961(%)\u7d9c\u5408\u6bd4\u8f03\u8868 (\u56db) \u8abf\u8b8a\u983b\u8b5c\u6b63\u898f\u5316\u6cd5\u7d50\u5408\u6642\u57df\u578b\u7279\u5fb5\u6b63\u898f\u5316\u6cd5\u4e4b\u5be6\u9a57\u7d50\u679c \u5728\u672c\u7bc0\u4e2d\uff0c\u6211\u5011\u5148\u5c07\u539f\u59cb MFCC \u7279\u5fb5\u7d93\u5404\u5f0f\u6642\u57df\u578b\u7279\u5fb5\u7d71\u8a08\u6b63\u898f\u5316\u6cd5\u8655\uf9e4\u5f8c\uff0c\u518d \u4f5c\u8abf\u8b8a\u983b\u8b5c\u7d71\u8a08\u6b63\u898f\u5316\u6cd5\u7684\u8655\uf9e4\u3002\u5728\u4ee5\u4e0b\u5404\u9805\u5c07\u5448\u73fe\u4e26\u8a0e\uf941\u5404\u5f0f\u8abf\u8b8a\u983b\u8b5c\u6b63\u898f\u5316\u7d50\u5408\u6642 \u57df\u578b\u7279\u5fb5\u6b63\u898f\u5316\u6cd5\u4e4b\u5be6\u9a57\u7d50\u679c\u3002 1. \u8abf\u8b8a\u983b\u8b5c\u5e73\u5747\u503c\u6b63\u898f\u5316\u6cd5\u7d50\u5408\u6642\u57df\u578b\u7279\u5fb5\u7d71\u8a08\u6b63\u898f\u5316\u6cd5\u4e4b\u5be6\u9a57\u7d50\u679c \u5be6\u9a57\u7d50\u679c\u8a0e\uf941\uff1a (1) \u8868\u4e09\u4e2d\uff0c\u8abf\u8b8a\u983b\u8b5c\u5e73\u5747\u503c\u6b63\u898f\u5316\u6cd5\u7d50\u5408\u6642\u57df\u578b\u7279\u5fb5\u7d71\u8a08\u6b63\u898f\u5316\u6cd5\uff0c\u8207\u5176\u4e2d\u55ae\u4e00\u7279\u5fb5 \u6b63\u898f\u5316\u6cd5\u6bd4\u8f03\uff0c\u5e7e\u4e4e\u7686\u80fd\u6709\u6548\u63d0\u5347\u8a9e\u97f3\u8fa8\uf9fc\u6548\u80fd\u3002\u8209\uf9b5\u800c\u8a00\uff1aSB-SMN (4) Method Set A Set B Set C average AR RR 1 RR 2 Baseline 71.98 67.79 78.28 71.56 \u2500 \u2500 \u2500 CMN 80.69 83.41 80.09 81.66 \u2500 \u2500 \u2500 FB-SMN 82.34 84.06 81.61 82.88 1.22 6.65 \u2500 SB-SMN (2) 80.89 82.22 80.24 81.29 -0.37 -2.02 -9.29 SB-SMN (3) 83.67 84.63 82.69 83.86 2.20 12.00 5.72 CMN SB-SMN (4) 88.09 88.64 86.63 88.02 6.36 34.68 30.02 CMVN 83.55 83.75 81.57 83.23 \u2500 \u2500 \u2500 FB-SMN 87.81 88.18 86.27 87.65 4.42 26.36 \u2500 SB-SMN (2) 89.08 89.39 87.39 88.87 5.64 33.63 9.88 SB-SMN (3) 89.63 89.97 88.37 89.51 6.28 37.45 15.06 CMVN SB-SMN (4) 89.86 90.09 88.20 89.62 6.39 38.10 15.95 MVA 86.69 86.89 84.98 86.43 \u2500 \u2500 \u2500 FB-SMN 88.89 89.19 87.54 88.74 2.31 17.02 \u2500 SB-SMN (2) 89.87 90.17 88.77 89.77 3.34 24.61 9.15 SB-SMN (3) 90.08 90.51 88.94 90.02 3.59 26.46 11.37 MVA SB-SMN (=4) 90.36 90.59 88.94 90.17 3.74 27.56 12.70 HEQ 86.90 87.73 87.56 87.36 \u2500 \u2500 \u2500 FB-SMN 89.10 89.70 89.20 89.36 2.00 15.82 \u2500 SB-\u8868\u4e09 SMN \u6cd5\u7d50\u5408\u6642\u57df\u578b\u7279\u5fb5\u6b63\u898f\u5316\u6cd5\u4e4b\u5be6\u9a57\u7d9c\u5408\u6bd4\u8f03\u8868 2. \u8abf\u8b8a\u983b\u8b5c\u5e73\u5747\u503c\u8207\u8b8a\uf962\uf969\u6b63\u898f\u5316\u6cd5\u7d50\u5408\u6642\u57df\u578b\u7279\u5fb5\u6b63\u898f\u5316\u6cd5\u4e4b\u5be6\u9a57\u7d50\u679c \u5be6\u9a57\u7d50\u679c\u8a0e\uf941\uff1a (1) \u8868\u56db\u4e2d\uff0c\u8abf\u8b8a\u983b\u8b5c\u5e73\u5747\u503c\u8207\u8b8a\uf962\uf969\u6b63\u898f\u5316\u6cd5\u7d50\u5408\u6642\u57df\u4e0a\u7279\u5fb5\u6b63\u898f\u5316\u6cd5\u8207\u55ae\u4e00\u4f5c\u7279\u5fb5 \u6b63\u898f\u5316\u6cd5\u6bd4\u8f03\uff0c\u7686\u80fd\u6709\u6548\u63d0\u5347\u8a9e\u97f3\u8fa8\uf9fc\u6548\u80fd\u3002\u8209\u800c\u8a00\u4e4b\uff1aSB-SMVN (4) Method Set A Set B Set C average AR RR 1 RR 2 Baseline 71.98 67.79 78.28 71.56 \u2500 \u2500 \u2500 CMN 80.69 83.41 80.09 81.66 \u2500 \u2500 \u2500 FB-SMVN 87.38 87.73 85.61 87.17 5.51 30.04 \u2500 SB-SMVN (2) 88.60 88.89 86.83 88.36 6.70 36.53 9.28 SB-SMVN (3) 89.77 89.90 88.17 89.50 7.84 42.75 18.16 CMN SB-SMVN (4) 90.01 90.35 88.43 CMVN 83.55 83.75 81.57 83.23 \u2500 \u2500 \u2500 FB-SMVN 87.33 87.80 85.47 87.15 3.92 23.38 \u2500 SB-SMVN (2) MVA 86.69 86.89 84.98 86.43 \u2500 \u2500 \u2500 FB-HEQ 86.90 87.73 87.56 87.36 \u2500 \u2500 \u2500 FB-\u8868\u56db SMVN \u6cd5\u7d50\u5408\u6642\u57df\u578b\u7279\u5fb5\u6b63\u898f\u5316\u6cd5\u4e4b\u5be6\u9a57\u7d9c\u5408\u6bd4\u8f03\u8868 3. \u8abf\u8b8a\u983b\u8b5c\u7d71\u8a08\u5716\u7b49\u5316\u6cd5\u7d50\u5408\u6642\u57df\u578b\u7279\u5fb5\u6b63\u898f\u5316\u6cd5\u4e4b\u5be6\u9a57\u7d50\u679c \u5be6\u9a57\u7d50\u679c\u8a0e\uf941\uff1a \u5728\u672a\uf92d\u5c55\u671b\u4e2d\uff0c\u6211\u5011\u5c07\u9032\u4e00\u6b65\u7814\u7a76\u5206\u983b\u6bb5\u8abf\u8b8a\u983b\u8b5c\u7d71\u8a08\u6b63\u898f\u5316\u6cd5\u4e2d\u7684\uf9e4\uf941\u57fa\u790e\uff0c\u4e26 \u5e0c\u671b\u80fd\u85c9\u7531\uf901\u56b4\u8b39\u7684\uf969\u5b78\u5206\u6790\u8207\u63a8\u5c0e\uff0c\u6c42\u53d6\u9019\u4e9b\u65b9\u6cd5\u4e2d\u6700\u4f73\u7684\u5206\u983b\u6bb5\uf969\u76ee\u3002\u6b64\u5916\uff0c\u6211\u5011 \u4e5f\u5e0c\u671b\u76f8\u95dc\u5be6\u9a57\uf967\u50c5\u5728\uf969\u5b57\u8fa8\uf9fc\u4e0a\u8655\uf9e4\uff0c\u4e5f\u64f4\u5c55\u81f3\u5176\u4ed6\u8f03\u5927\u5b57\u5f59\uf97e\u7684\u8a9e\u97f3\u8fa8\uf9fc\uff0c\u63a2\u8a0e\u9019 \u4e00\u7cfb\uf99c\u5206\u983b\u6bb5\u8abf\u8b8a\u983b\u8b5c\u7d71\u8a08\u6b63\u898f\u5316\u6cd5\u5728\uf967\u540c\u8907\u96dc\ufa01\u4e4b\u8a9e\u97f3\u8fa8\uf9fc\u7cfb\u7d71\u7684\u6548\u80fd\uff0c\u6216\u662f\u61c9\u7528\u65bc \u5176\u4ed6\uf9d0\u578b\u7684\u5e72\u64fe\u5931\u771f\u74b0\u5883\uff0c\u9032\u4e00\u6b65\u9a57\u8b49\u9019\u4e9b\u65b0\u65b9\u6cd5\u7684\u6548\u80fd\u8207\u5be6\u7528\u6027\uff0c\u4ee5\u4e0a\u5404\u9ede\u90fd\u662f\u672a\uf92d \u80fd\u5920\u5617\u8a66\u7814\u7a76\u767c\u5c55\u7684\u65b9\u5411\u3002\u671f\u76fc\u5c07\uf92d\u8a9e\u97f3\u8fa8\uf9fc\u4e4b\u6548\u80fd\u80fd\u5920\uf901\u52a0\u63d0\u5347\uff0c\u4e26\u4e14\u666e\u904d\u61c9\u7528\u65bc\u65e5 \u5e38\u751f\u6d3b\uff0c\u8b93\u4eba\u5011\u8f15\u9b06\u5730\uf9dd\u7528\u8a9e\u97f3\u8207\u96fb\u8166\u6216 3C \u7522\u54c1\u9032\ufa08\u4e92\u52d5\uff0c\uf9a8\u751f\u6d3b\u80fd\u5920\uf901\uf965\uf9dd\uff0c\u4f7f\u8a9e \u97f3\u8fa8\uf9fc\u4e4b\u767c\u5c55\u517c\u5177\uf9e4\uf941\u6027\u8207\u5be6\u7528\u6027\u3002 (1) \u97f3\u5f37\u5065\u6027\u6280\u8853\u6709\uf97c\u597d\u7684\u52a0\u6210\u6027\uff0c\u5f97\u4ee5\u660e\u986f\u6539\u5584\u96dc\u8a0a\u74b0\u5883\u4e0b\u7684\u8a9e\u97f3\u8fa8\uf9fc\u6548\u80fd\u3002 \uf96b\u8003\u6587\u737b</td></tr><tr><td>\u4e09\u3001\u5206\u983b\u6bb5\u8abf\u8b8a\u983b\u8b5c\u7d71\u8a08\u6b63\u898f\u5316\u6cd5\u4e4b\u5be6\u9a57\u7d50\u679c\u53ca\u5206\u6790\u8a0e\uf941 ( ) 2 % 100% 100% RR \u239b \u239e \u2212 \u239f \u239c = \u00d7 \u239f \u239c \u239f \u239f \u239c \u2212 \u239d \u23a0 \u5206\u983b\u6bb5\u6cd5\u8fa8\u8b58\u7387 \u5168\u983b\u6bb5\u6cd5\u8fa8\u8b58\u7387 \u5168\u983b\u6bb5\u6cd5\u8fa8\u8b58\u7387</td><td>\u5f0f(3-3)</td></tr></table>", |
| "html": null, |
| "text": "\u7684\uf969\u64da\u770b\u51fa\uff0c\u5b83\u5011\u81f3\u5c11\u80fd\u6709 19.00%\u7684\u76f8\u5c0d\u932f\u8aa4\ufa09\u4f4e\uf961\uff1b\u5176\u4e2d SB-SHE \u6cd5\u7684\u8fa8\uf9fc\u6548\u679c\u6bd4 SB-SMN \u6cd5\u53ca SB-SMVN \u6cd5\uf901\u512a\u8d8a\uff0c\u53ef\u80fd\u4e4b\u539f\u56e0\u5982\u6211\u5011\u9810\u671f\u7684\uff0cSB-SMN \u6cd5\u53ca SB-SMVN \u6cd5\u53ea\u5c0d\u4e00\u968e\u52d5\u5dee\u6216\u4e00\u968e\u53ca\u4e8c\u968e\u52d5\u5dee\u4f5c\u6b63\u898f\u5316\uff0c\u800c SB-SHE \u6cd5\u80fd\u540c\u6642\u5c0d\uf901 \u7d50\u5408 CMN \u7684 \u8fa8\uf9fc\uf961\u70ba 88.02%\uff0c\u6bd4\u8d77 CMN \u7684\u8fa8\uf9fc\uf961 81.66%\u8207 SB-SMN (L=4) \u7684\u8fa8\uf9fc\uf961 78.99%\uff0c\u90fd\u6709 \u76f8\u7576\u660e\u986f\u7684\u6539\u5584\u3002\u60df\u7368\u5728 SB-SMN (2) \u7d50\u5408 CMN \u60c5\u6cc1\u4e0b\uff0c\u7121\u6cd5\u9032\u4e00\u6b65\u63d0\u5347\u8fa8\uf9fc\uf961\uff0c\u9019\u53ef \u80fd\u662f\u8a72\u7d50\u5408\u65b9\u5f0f\u7684\u5206\u983b\u6bb5\u4e4b\u5e73\u5747\u503c\u7121\u6cd5\u6709\u6548\u903c\u8fd1\u8a13\uf996\u8a9e\uf906\u4e4b\u5206\u983b\u6bb5\u5f37\ufa01\u983b\u8b5c\u7684\u5e73\u5747 \u503c\uff0c\u5c0e\u81f4\u8fa8\uf9fc\uf961\u660e\u986f\u4e0b\ufa09\u3002 (2) \u5f9e\u8868\u4e09\u4e5f\u53ef\u4ee5\u6e05\u695a\u767c\u73fe\uff0c\u7576\u4f4e\u983b\u90e8\u4efd\ufa00\u5272\u8d8a\u7d30\uff0c\u80fd\u6709\u6548\u63d0\u5347\u8a9e\u97f3\u8fa8\uf9fc\u6548\u80fd\uff0c\u800c SMN \u6cd5\u7d50\u5408\u5176\u4ed6\u7279\u5fb5\u6b63\u898f\u5316\u6cd5\u7686\u4ee5\u5206\u5272\u56db\u500b\u983b\u6bb5\u7684\u6548\u679c\u6700\u70ba\u512a\u8d8a\u3002\u5176\u4e2d SB-SMN (4) \u7d50\u5408 CMN \u7684 RR 2 \u70ba 30.02%\uff0cSB-SMN (4) \u7d50\u5408 CMVN \u7684 RR 2 \u70ba 15.95%\uff0cSB-SMN (4) \u7d50\u5408 MVA \u7684 RR 2 \u70ba 12.70%\uff0cSB-SMN (4) \u7d50\u5408 HEQ \u7684 RR 2 \u70ba 4.98%\u3002 (3) SB-SMN (L=4) \u5206\u5225\u8207 CMVN\u3001MVA \u53ca HEQ \u7d50\u5408\uff0c\u4f7f\u8fa8\uf9fc\uf961\u5e7e\u4e4e\u9054\u5230 90.00%\uff1b\u800c\u6b64 \u4ee3\u8868\u6211\u5011\u7528\u76f8\u5c0d\u7c21\u55ae\u7684\u4e00\u968e\u7d71\u8a08\u6b63\u898f\u5316\u6cd5(SB-SMN)\u7d50\u5408 CMVN\u3001MVA \u53ca HEQ\uff0c\u5373\u53ef \u9054\u5230\u5341\u5206\u7a81\u51fa\u7684\u6548\u679c\u3002 SMN (L=2) 89.20 89.81 89.30 89.46 2.10 16.61 0.94 SB-SMN (L=3) 89.15 89.89 89.35 89.48 2.12 16.77 1.13 HEQ SB-SMN (L=4) 89.54 90.28 89.82 \u7d50\u5408 CMVN \u7684\u8fa8 \uf9fc\uf961\u70ba 89.87%\uff0c\u6bd4\u8d77 CMVN \u7684\u8fa8\uf9fc\uf961 83.23%\u8207 SB-SMVN (4) \u7684\u8fa8\uf9fc\uf961 86.36%\uff0c\u90fd\u6709\u76f8 \u7576\u660e\u986f\u7684\u6539\u5584\u3002 (2) \u5f9e\u8868\u56db\u4e5f\u53ef\u4ee5\u6e05\u695a\u767c\u73fe\uff0c\u7576\u4f4e\u983b\u90e8\u4efd\ufa00\u5272\u8d8a\u7d30\uff0c\uf901\u80fd\u6709\u6548\u63d0\u5347\u8a9e\u97f3\u8fa8\uf9fc\u6548\u80fd\uff0c\u800c SMVN \u6cd5\u7d50\u5408\u5176\u4ed6\u7279\u5fb5\u6b63\u898f\u5316\u6cd5\u7686\u4ee5\u5206\u5272\u56db\u500b\u983b\u6bb5\u7684\u6548\u679c\u6700\u70ba\u512a\u8d8a\u3002\u5176\u4e2d SB-SMVN (4) \u7d50\u5408 CMN \u7684 RR 2 \u70ba 20.73%\uff0cSB-SMVN (4) \u7d50\u5408 CMVN \u7684 RR 2 \u70ba 21.17%\uff0cSB-SMVN (4) \u7d50\u5408 MVA \u7684 RR 2 \u70ba 16.99%\uff0cSB-SMVN (4) \u7d50\u5408 HEQ \u7684 RR 2 \u70ba 9.80%\u3002 88.54 88.84 86.80 88.31 5.08 30.29 9.03 SB-SMVN (3) 89.72 89.90 88.03 89.45 6.22 37.09 17.90 CMVN SB-SMVN (4) 90.11 90.37 88.42 SMVN 88.18 88.60 86.39 87.99 1.56 11.50 \u2500 SB-SMVN (2) 89.27 89.49 87.58 89.02 2.59 19.09 8.58 SB-SMVN (3) 89.69 89.95 88.36 89.52 3.09 22.77 12.74 MVA SB-SMVN (4) 90.27 90.45 88.71 SMVN 89.05 89.55 89.26 89.29 1.93 15.27 \u2500 SB-SMVN (2) 89.33 89.81 89.40 89.54 2.18 17.25 2.33 SB-SMVN (3) 89.95 90.35 89.91 90.11 2.75 21.76 7.66 HEQ SB-SMVN (4) 90.19 90.59 90.17 \uf9d0\u4f3c\u8868\u4e09\u3001\u8868\u56db\uff0c\u5f9e\u8868\u4e94\u770b\u51fa\uff0c\u8abf\u8b8a\u983b\u8b5c\u7d71\u8a08\u5716\u7b49\u5316\u6cd5\u7d50\u5408\u6642\u57df\u578b\u7279\u5fb5\u6b63\u898f\u5316\u6cd5\u7686 \u512a\u65bc\u500b\u5225\u7279\u5fb5\u6b63\u898f\u5316\u6cd5\u3002\u8209\uf9b5\u800c\u8a00\uff1aSB-SHE (L=4) \u7d50\u5408 CMVN \u7684\u8fa8\uf9fc\uf961\u70ba 90.18%\uff0c\u6bd4\u8d77 CMVN \u7684\u8fa8\uf9fc\uf961 83.23%\u6709\u76f8\u7576\u660e\u986f\u7684\u6539\u5584\u3002\u4f46 SB-SHE \u7d50\u5408 CMVN \u4e4b\u5be6\u9a57\u7d50\u679c\u8207 SB-SHE \u4f5c\u5728 MFCC \u7279\u5fb5\u4e4b\u5be6\u9a57\u7d50\u679c\u6bd4\u8f03\uff0c\u5728\u67d0\u4e9b\u7d50\u5408\u65b9\u5f0f\u4e0b\uff0c\u8fa8\uf9fc\u7d50\u679c\u6703\u6709\u4e9b\u5fae\u7684\u4e0b \ufa09\uff0c\u9019\u53ef\u80fd\u662f\u5be6\u9a57\u7684\u8aa4\u5dee\u7bc4\u570d\uff0c\u6216\u662f\u904e\ufa01\u6b63\u898f\u5316\u7684\uf967\uf97c\u6548\u61c9\u3002 (2) \u5f9e\u8868\u4e94\u4e5f\u53ef\u4ee5\u6e05\u695a\u767c\u73fe\uff0c\u7576\ufa00\u5272\u7684\u5b50\u983b\u6bb5\uf969\u76ee\u8d8a\u591a\uff0c\u8d8a\u80fd\u6709\u6548\u63d0\u5347\u8a9e\u97f3\u8fa8\uf9fc\u6548\u80fd\uff0c \u800c SHE \u6cd5\u7d50\u5408\u5176\u4ed6\u7279\u5fb5\u6b63\u898f\u5316\u6cd5\u7686\u4ee5\u5206\u5272\u56db\u500b\u983b\u6bb5\u7684\u6548\u679c\u6700\u70ba\u512a\u8d8a\u3002\u5176\u4e2d SB-SHE (4) \u7d50\u5408 CMN \u7684 RR 2 \u70ba 6.45%\uff0cSB-SHE (4) \u7d50\u5408 CMVN \u7684 RR 2 \u70ba 7.97%\uff0cSB-SHE (4) \u7d50\u5408", |
| "type_str": "table", |
| "num": null |
| } |
| } |
| } |
| } |
