{
    "paper_id": "O08-1007",
    "header": {
        "generated_with": "S2ORC 1.0.0",
        "date_generated": "2023-01-19T08:02:28.091714Z"
    },
    "title": "\u8abf\u8b8a\u983b\u8b5c\u6b63\u898f\u5316\u6cd5\u4f7f\u7528\u65bc\u5f37\u5065\u8a9e\u97f3\u8fa8\uf9fc\u4e4b\u7814\u7a76 Study of Modulation Spectrum Normalization Techniques for Robust Speech Recognition",
    "authors": [
        {
            "first": "Chih-Cheng",
            "middle": [],
            "last": "\u738b\u81f4\u7a0b",
            "suffix": "",
            "affiliation": {},
            "email": ""
        },
        {
            "first": "",
            "middle": [],
            "last": "Wang",
            "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": "The performance of an automatic speech recognition system is often degraded due to the embedded noise in the processed speech signal. A variety of techniques have been proposed to deal with this problem, and one category of these techniques aims to normalize the temporal statistics of the speech features, which is the main direction of our proposed new approaches here. In this thesis, we propose a series of noise robustness approaches, all of which attempt to normalize the modulation spectrum of speech features. They include equi-ripple temporal filtering (ERTF), least-squares spectrum fitting (LSSF) and magnitude spectrum interpolation (MSI). With these approaches, the mismatch between the modulation spectra for clean and noise-corrupted speech features is reduced, and thus the resulting new features are expected to be more noise-robust. Recognition experiments implemented on Aurora-2 digit database show that the three new approaches effectively improve the recognition accuracy under a wide range of noise-corrupted environment. Moreover, it is also shown that they can be successfully",
    "pdf_parse": {
        "paper_id": "O08-1007",
        "_pdf_hash": "",
        "abstract": [
            {
                "text": "The performance of an automatic speech recognition system is often degraded due to the embedded noise in the processed speech signal. A variety of techniques have been proposed to deal with this problem, and one category of these techniques aims to normalize the temporal statistics of the speech features, which is the main direction of our proposed new approaches here. In this thesis, we propose a series of noise robustness approaches, all of which attempt to normalize the modulation spectrum of speech features. They include equi-ripple temporal filtering (ERTF), least-squares spectrum fitting (LSSF) and magnitude spectrum interpolation (MSI). With these approaches, the mismatch between the modulation spectra for clean and noise-corrupted speech features is reduced, and thus the resulting new features are expected to be more noise-robust. Recognition experiments implemented on Aurora-2 digit database show that the three new approaches effectively improve the recognition accuracy under a wide range of noise-corrupted environment. Moreover, it is also shown that they can be successfully",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Abstract",
                "sec_num": null
            }
        ],
        "body_text": [
            {
                "text": "combined with some other noise robustness approaches, like CMVN and MVA, to achieve a more excellent recognition performance. [2] \u3001 \u5012 \u983b \u8b5c \u5e73 \u5747 \u8207 \u8b8a \uf962 \uf969 \u6b63 \u898f \u5316 \u6cd5 (cepstral mean and variance normalization, CMVN) [3] \u3001\u76f8\u5c0d\u983b\u8b5c\u6cd5(RelAtive SpecTrAl, RASTA) [4] \u3001\u5012\u983b\u8b5c\u5e73\u5747\u8207\u8b8a \uf962\uf969\u6b63\u898f\u5316\u5316\u7d50\u5408\u81ea\u52d5\u56de\u6b78\u52d5\u614b\u5e73\u5747\uf984\u6ce2\u5668\u6cd5(cepstral mean and variance normalization plus auto-regressive-moving-average filtering, MVA) [5] \u3001\u5012\u983b\u8b5c\u589e\u76ca\u6b63\u898f\u5316\u6cd5 (cepstral gain normalization, CGN) [6] \u3001\u8cc7\uf9be\u5c0e\u5411\u6642\u9593\u5e8f\uf99c\uf984\u6ce2\u5668\u6cd5(data-driven temporal filter design) [7] \u7b49\u3002\u4ee5\u4e0a\u9019\u4e9b\u65b9\u6cd5\u7686\u662f\u5728\u8a9e\u97f3\u7279\u5fb5\u7684\u6642\u9593\u5e8f\uf99c\u57df(temporal domain)\u4f5c\u8655\uf9e4\uff0c\u6839 \u64da\u8a9e\u97f3\u8a0a\u865f\u8207\u96dc\u8a0a\u5728\u6642\u9593\u5e8f\uf99c\u57df\u4e0a\uf967\u540c\u7684\u7279\u6027\uff0c\u5f37\u8abf\u51fa\u8a9e\u97f3\u7684\u6210\u5206\uff0c\u800c\u6291\u5236\u96dc\u8a0a\u7684\u5f71\u97ff\u3002 \u8fd1\uf92d\uff0c\u65b0\u52a0\u5761\u5927\u5b78\u4e4b\uf9e1\u6d77\u6d32\u535a\u58eb\u7814\u7a76\u5718\u968a\uff0c\u65b0\u63a8\u51fa\uf9ba\u4e00\u5957\u6642\u9593\u5e8f\uf99c\uf984\u6ce2\u5668\u8a2d\u8a08\u7684\u65b0\u65b9 \u6cd5\uff0c\u7a31\u70ba\u300e\u6642\u9593\u5e8f\uf99c\u7d50\u69cb\u6b63\u898f\u5316\u6cd5\u300f(temporal structure normalization, TSN) [ ",
                "cite_spans": [
                    {
                        "start": 126,
                        "end": 129,
                        "text": "[2]",
                        "ref_id": "BIBREF1"
                    },
                    {
                        "start": 207,
                        "end": 210,
                        "text": "[3]",
                        "ref_id": "BIBREF2"
                    },
                    {
                        "start": 244,
                        "end": 247,
                        "text": "[4]",
                        "ref_id": "BIBREF3"
                    },
                    {
                        "start": 371,
                        "end": 374,
                        "text": "[5]",
                        "ref_id": "BIBREF4"
                    },
                    {
                        "start": 421,
                        "end": 424,
                        "text": "[6]",
                        "ref_id": "BIBREF5"
                    },
                    {
                        "start": 475,
                        "end": 478,
                        "text": "[7]",
                        "ref_id": "BIBREF6"
                    },
                    {
                        "start": 657,
                        "end": 658,
                        "text": "[",
                        "ref_id": null
                    }
                ],
                "ref_spans": [],
                "eq_spans": [],
                "section": "",
                "sec_num": null
            },
            {
                "text": "{ } s n \u8207\u6e2c\u8a66\u8a9e\uf9be\u540c\u4e00\u7dad\u7279\u5fb5\u5e8f \uf99c [ ] { } x n \uff0c\u5148\u6c42\u53d6\u5176\u529f\uf961\u983b\u8b5c\u5bc6\ufa01\uff0c\u5206\u5225\u4ee5 ( ) { } SS k P \u03c9 \u8207 ( ) { } XX k P \u03c9 \u8868\u793a\u3002\u63a5\u8457\u5c07\u8a13\uf996\u8a9e \uf9be\u6240\u6709\uf906\u5b50\u540c\u4e00\u7dad\u7684\u529f\uf961\u983b\u8b5c\u5bc6\ufa01\u4f5c\u5e73\u5747\uff0c\u6240\u5f97\u5373\u70ba\uf96b\u8003\u529f\uf961\u983b\u8b5c\u5bc6\ufa01\uff0c\u5982\u4e0b\u6240\u793a\uff1a ( ) ( ) { }, SS k SS k P E P \u03c9 \u03c9 = (\u5f0f 2.1) \u5728 TSN \u6cd5\u4e2d\u6240\u4f7f\u7528\u7684\uf984\u6ce2\u5668\uff0c\u5176\u521d\u59cb\u7684\u5f37\ufa01\u983b\u8b5c\u8a2d\u5b9a\u5982\u4e0b\u5f0f\u6240\u793a\uff1a ( ) ( ) ( ), k S S k X X",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "",
                "sec_num": null
            },
            {
                "text": "( ) 1 0 1 , 0 1 k M j m k k h m H j e m M M \u03c9 \u03c9 \u2212 \u2212 = = \u2264 \u2264 \u2212 \u2211 . (\u5f0f 2.3) 2\u3001\u6f22\uf95f\u7a97\u5316\u8655\uf9e4\uff1a [ ] [ ] [ ] h m h m w m = \u22c5 , (\u5f0f 2.4) \u5176\u4e2d [ ] 0.5 1 cos 2 , 0 1 1 m w m m M M \u03c0 \u239b \u239e \u239b \u239e \u239f \u239c \u239f \u239c = \u2212 \u2264 \u2264 \u2212 \u239f \u239f \u239c \u239c \u239f\u239f \u239c \u239d \u23a0 \u239d \u23a0 \u2212 . 3\u3001\u76f4\uf9ca\u589e\u76ca\u6b63\u898f\u5316\uff1a [ ] [ ] 1 0 M m h m h m h m \u2212 \u2032= = \u23a1 \u23a4 \u2032 \u23a3 \u23a6 \u2211 . (\u5f0f 2.5) \uf96b\u8003\u4e4b\u529f\uf961\u983b\u8b5c\u5bc6\ufa01 [ ] x n ( ) SS P \u03c9 ( ) H \u03c9 IDFT ( ) h n Hanning Window DC Gain Normalization ( ) h n [ ] y n \u8655\uf9e4\u524d\u7684\u7279\u5fb5\u5e8f\uf99c \u8655\uf9e4\u5f8c\u7684\u7279\u5fb5\u5e8f\uf99c \u6642\u9593\u5e8f\uf99c\uf984\u6ce2\u5668 ( ) XX P \u03c9 \u4e7e\u6de8\u8a9e\uf906\u7279\u5fb5 \uf96b\uf969\u5e8f\uf99c \u529f\uf961 \u983b\u8b5c\u5bc6\ufa01 \u5176\u4e2d M \u70ba\uf984\u6ce2\u5668\u9577\ufa01\u3002\u5f0f(2.5)\u4e4b [ ] h m \u5373\u70ba TSN \u6240\u6c42\u5f97\u4e4b\u6642\u9593\u5e8f\uf99c\uf984\u6ce2\u5668\u7684\u8108\u885d\u97ff\u61c9\u3002 (\u4e8c)TSN \u6cd5\u6548\u679c\u76f8\u95dc\u8a0e\uf941 \u5728 TSN \u4e4b\u6587\u737b[8]\u4e2d\uff0c\u6240\u7528\u7684\u539f\u59cb\u7279\u5fb5\uf96b\uf969\u7686\u70ba\u7d93\u904e CMVN \u6cd5\u6216 MVA \u6cd5\u6240\u8655\uf9e4\u5f8c \u4e4b\u6885\u723e\u5012\u983b\u8b5c\u7279\u5fb5\uf96b\uf969(MFCC)\u3002\u9019\uf9e8\u6211\u5011\u7279\u5225\u5c07 TSN \u6cd5\u904b\u7528\u5728\u672a\u7d93\u8655\uf9e4\u4e4b\u6885\u723e\u5012\u983b\u8b5c \u7279\u5fb5\uf96b\uf969\u4e0a\uff0c\u89c0\u5bdf\u5176\u6539\u9032\u6548\u679c\u3002\u5176\u4e2d\u6211\u5011\u628a\u539f\u59cb TSN \u6cd5\u547d\u540d\u70ba TSN-1\uff0c\u800c\u628a\uf96d\uf976\uf9ba\u76f4 \uf9ca\u589e\u76ca\u6b63\u898f\u5316\u6b65\u9a5f\u7684 TSN \u6cd5\uff0c\u547d\u540d\u70ba TSN-2\u3002\u5716\u4e09\u70ba\u539f\u59cb\u7b2c\u4e00\u7dad\u6885\u723e\u5012\u983b\u8b5c\u4fc2\uf969(c 1 ) \u5e8f\uf99c\u7684\u529f\uf961\u983b\u8b5c\u5bc6\ufa01\u66f2\u7dda\u5716\uff0c\u5716\u56db\u70ba\u539f\u59cb c 1 \u5e8f\uf99c\u7d93 TSN-1 \u6cd5\u8655\uf9e4\u5f8c\u7684\u529f\uf961\u983b\u8b5c\u5bc6\ufa01\u66f2 \u7dda\u5716\uff0c\u5716\u4e94\u70ba\u539f\u59cb c 1 \u5e8f\uf99c\u7d93 TSN-2 \u6cd5\u8655\uf9e4\u5f8c\u7684\u529f\uf961\u983b\u8b5c\u5bc6\ufa01\u66f2\u7dda\u5716\u3002\u9019\u4e9b\u5716\u90fd\u4f7f\u7528\uf9ba AURORA 2 \u8cc7\uf9be\u5eab[9]\uf9e8\u7684 MAH_4625A \u8a9e\u97f3\u6a94\uff0c\u52a0\u5165\uf967\u540c\u8a0a\u96dc\u6bd4\u7684\u5730\u4e0b\u9435\u96dc\u8a0a\u3002\u5176\u4e2d \uf96b\u8003\u529f\uf961\u983b\u8b5c\u5bc6\ufa01\u70ba\u8a13\uf996\u8a9e\uf9be\u5eab\u4e4b\u6240\u6709 c 1 \u5e8f\uf99c\u4e4b\u529f\uf961\u983b\u8b5c\u5bc6\ufa01\u5e73\u5747\u800c\u5f97\u3002 \u5716\u4e09\u3001\uf967\u540c\u8a0a\u96dc\u6bd4\u4e4b\u4e0b\uff0c\u539f\u59cb c 1 \u5e8f\uf99c\u4e4b\u529f\uf961\u983b\u8b5c\u5bc6\ufa01\u66f2\u7dda\u5716 \u5716\u56db\u3001\uf967\u540c\u8a0a\u96dc\u6bd4\u4e4b\u4e0b\uff0c\u539f\u59cb c 1 \u5e8f\uf99c\u7d93 TSN-1 \u8655\uf9e4\u5f8c\u4e4b\u529f\uf961\u983b\u8b5c\u5bc6\ufa01\u66f2\u7dda\u5716 \u5716\u4e94\u3001\uf967\u540c\u8a0a\u96dc\u6bd4\u4e4b\u4e0b\uff0c\u539f\u59cb c 1 \u5e8f\uf99c\u7d93 TSN-2 \u8655\uf9e4\u5f8c\u4e4b\u529f\uf961\u983b\u8b5c\u5bc6\ufa01\u66f2\u7dda\u5716 Hz Hz Hz \u9996\u5148\uff0c\u5f9e\u5716\u4e09\u53ef\u4ee5\u660e\u986f\u770b\u51fa\uff0c\u96dc\u8a0a\u6703\u9020\u6210 c 1 \u7279\u5fb5\u5e8f\uf99c\u5728\u529f\uf961\u983b\u8b5c\u5bc6\ufa01\u4e0a\u7684\u5931\u771f\uff0c \u6b64\u662f\u9020\u6210\u96dc\u8a0a\u74b0\u5883\u4e0b\uff0c\u8a9e\u97f3\u8fa8\uf9fc\u7cbe\u78ba\uf961\u4e0b\ufa09\u7684\u539f\u56e0\u4e4b\u4e00\u3002\u63a5\u8457\uff0c\u6211\u5011\u5f9e\u5716\u56db\u89c0\u5bdf\u5230\u539f\u59cb TSN \u6cd5(TSN-1)\u4f5c\u7528\u65bc\u539f\u59cb c 1 \u5e8f\uf99c\u6642\uff0c\u539f\u672c\u5728\u5716\u4e09\u6240\u770b\u5230\u4e4b\u529f\uf961\u983b\u8b5c\u5bc6\ufa01\u7684\u5931\u771f\u4e26\u672a \u88ab\u6709\u6548\u5730\u6539\u5584\uff0c\u4ea6\u5373\u5176\u6b63\u898f\u5316\u6548\u679c\u4e26\uf967\uf9e4\u60f3\uff0c\u53d7\u5230\u96dc\u8a0a\u5f71\u97ff\u7684 c 1 \u5e8f\uf99c\uff0c\u7576\u8a0a\u96dc\u6bd4(SNR) \u8d8a\u4f4e\u6642\uff0c\u504f\u79fb\uf96b\u8003\u529f\uf961\u983b\u8b5c\u5bc6\ufa01\u7684\uf97e\u8d8a\u660e\u986f\u3002\u6700\u5f8c\uff0c\u5f9e\u5716\u4e94\u53ef\u4ee5\u770b\u51fa\uff0c\u7d93\u904e\uf96d\u537b\u76f4\uf9ca\u589e \u76ca\u6b63\u898f\u5316\u6b65\u9a5f\u7684 TSN-2 \u6cd5\u8655\uf9e4\u5f8c\uff0c\uf967\u540c\u8a0a\u96dc\u6bd4\u4e0b\u7684 c 1 \u7279\u5fb5\u5e8f\uf99c\u5176\u529f\uf961\u983b\u8b5c\u5bc6\ufa01\u5f7c\u6b64\u5341 \u5206\u63a5\u8fd1\uff0c\u4ea6\u5373 TSN-2 \u6cd5\u53ef\u4ee5\u6709\u6548\u6b63\u898f\u5316\u53d7\u96dc\u8a0a\u5e72\u64fe\u4e4b\u539f\u59cb c 1 \u5e8f\uf99c\u7684\u529f\uf961\u983b\u8b5c\u5bc6\ufa01\uff0c\u5176 \ufa09\u4f4e\u5931\u771f\u7684\u6548\u80fd\u9060\u6bd4 TSN-1 \uf92d\u7684\u597d\u3002\u7531\u6b64\u6211\u5011\u63a8\uf941\uff0c\u539f\u59cb",
                "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": "( ) XX k P \u03c9 \u3001 ( ) SS k P \u03c9 \u548c ( ) k H \u03c9 \u6c42\u53d6\u65b9\u5f0f\u90fd\u548c\u524d\u4e00\u7ae0\u6240\u8ff0\u4e4b\u539f\u59cb \u8655\uf9e4\u5f8c\u7684\u7279\u5fb5\u5e8f\uf99c \u4e7e\u6de8\u8a9e\uf906\u7279\u5fb5\uf96b\uf969 \u5e8f\uf99c ( ) SS P \u03c9 ( ) H \u03c9 ( ) XX P \u03c9 [ ] x n [ ] y n \u6642\u9593\u5e8f\uf99c\uf984\u6ce2\u5668 \u8655\uf9e4\u524d\u7684\u7279\u5fb5\u5e8f\uf99c ( ) h n \u7b49\uf992\u6ce2 \uf984\u6ce2\u5668\u8a2d\u8a08 ( ) h n \uf96b\u8003\u4e4b\u529f\uf961\u983b\u8b5c\u5bc6\ufa01 \u529f\uf961 \u983b\u8b5c\u5bc6\ufa01 TSN \u6cd5\u76f8\u540c\uff0c\u4f46\u662f\uf984\u6ce2\u5668\u4fc2\uf969 [ ] { } h n \u662f\u4ee5\u7b49\uf992\u6ce2\uf984\u6ce2\u5668\u8a2d\u8a08\u6cd5[10]\u6c42\u5f97\uff0c\u6b64\u65b9\u6cd5\u662f\uf9dd\u7528 \u6240\u8b02\u7684\u6700\u5c0f\u5316\u6700\u5927\u8aa4\u5dee\u6e96\u5247(minimax criterion)\uf92d\u6c42\u53d6\u4e00\u6700\u4f73\u7684\uf984\u6ce2\u5668\u983b\uf961\u97ff\u61c9\uff0c\u5982\u4e0b\u5f0f \u6240\u793a\uff1a ( ) ( ) ( ) ( ) ( ) ( ) arg min max k k k k k H H W H D \u03c9 \u03c9 \u03c9 \u03c9 \u03c9 \u03c9 = \u2212 , (\u5f0f 3.1) \u5176\u4e2d ( ) k W \u03c9 \u70ba\u6b0a\u91cd\u503c\uff0c ( ) k H \u03c9 \u70ba\u6700\u4f73\u5316\uf984\u6ce2\u5668\u4e4b\u983b\uf961\u97ff\u61c9\uff0c ( ) k D \u03c9 \u70ba\uf96b\u8003\u7684\u983b\uf961\u97ff\u61c9\uff0c ( ) k D \u03c9 \u53ef\u8868\u793a\u5982\u4e0b\u5f0f\uff1a ( ) ( ) ( ) SS k k XX k P D P \u03c9 \u03c9 \u03c9 = (\u5f0f 3.2) \u7531\u6b64\u6cd5\u5f97\u5230\u7684\uf984\u6ce2\u5668\u4fc2\uf969 [ ] { } h n \uff0c\u6703\u81ea\u52d5\u7b26\u5408\u524d\u5f8c\u5c0d\u7a31(symmetric)\u7684\u6027\u8cea\uff0c\u56e0\u6b64\u5176\u76f8\u4f4d \u97ff\u61c9\u662f\u7dda\u6027\u7684(linear phase)[10]\uff0c\u4e26\uf967\u6703\u4f7f\u539f\u59cb\u7279\u5fb5\u5e8f\uf99c\u7684\u8abf\u8b8a\u983b\u8b5c\u7522\u751f\u76f8\u4f4d\u5931\u771f\u7684\u60c5 \u5f62\uff0c\u540c\u6642\uff0c\u56e0\u70ba\uf984\u6ce2\u5668\u672c\u8eab\u662f\u6839\u64da\u6700\u4f73\u5316\u6e96\u5247\u8a2d\u8a08\uff0c\u6240\u4ee5\u6211\u5011\u9810\u671f\u5b83\u6703\u6bd4 TSN \u6cd5\u6240\u5f97 \u4e4b\uf984\u6ce2\u5668\u6548\u679c\uf92d\u7684\u597d\u3002 (\u4e8c)\u6700\u5c0f\u5e73\u65b9\u983b\u8b5c\u64ec\u5408\u6cd5(least-squares spectrum fitting, LSSF) \u5728\u9019\u65b9\u6cd5\uf9e8\uff0c\u6211\u5011\u91dd\u5c0d\u6bcf\u4e00\u500b\u5f85\u6b63\u898f\u5316\u7684 N \u9ede\u7279\u5fb5\u6642\u9593\u5e8f\uf99c [ ] { } 0 1 x n n N \u2264 \u2264 \u2212 \u5148\u5b9a\u7fa9\u4e00 2P \u9ede\u7684\uf96b\u8003\u8abf\u8b8a\u983b\u8b5c\uff0c\u4f5c\u70ba\u6b64\u7279\u5fb5\u5e8f\uf99c\u7684\u8abf\u8b8a\u983b\u8b5c\u6b63\u898f\u5316\u7684\u76ee\u6a19\uff1a ( ) ( ) ( ) ( ) exp , 0 2 1, k k X k Y Y j k P \u03c9 \u03c9 \u03b8 \u03c9 = \u2264\u2264 \u2212 (\u5f0f 3.3) \u5176\u4e2d\u7684\u5f37\ufa01\u6210\u4efd ( ) k Y \u03c9 \u4ee5\u4e0b\u5f0f\u8868\u793a\uff1a ( ) ( ) ( ) ( ) k k S S k X X k Y X P P \u03c9 \u03c9 \u03c9 \u03c9 = (\u5f0f 3.4) \u5176\u4e2d\uff0c ( ) SS k P \u03c9 \u8207\u5982\u524d\u7ae0\u7684\u5f0f(2.1)\u4e2d\u6240\u5b9a\u7fa9\uff0c\u5373 ( ) SS k P \u03c9 \u70ba\u6240\u6709\u8a13\uf996\u8a9e\uf9be\u7279\u5fb5\u8207 [ ] { } x n \u540c\u4e00\u7dad\u5e8f\uf99c\u7684\u529f\uf961\u983b\u8b5c\u5bc6\ufa01\u5e73\u5747\u800c\u5f97\uff0c ( ) XX k P \u03c9 \u70ba\u539f\u59cb\u7279\u5fb5\u5e8f\uf99c [ ] { } x n \u7684\u529f\uf961\u983b \u8b5c\u5bc6\ufa01\u3002\u800c\u5f37\ufa01\u6210\u4efd ( ) k X \u03c9 \u548c\u76f8\u89d2\u6210\u4efd ( ) X k \u03b8 \u03c9 \u70ba [ ] { } x n \u7d93\u904e 2P \u9ede\u4e4b\uf9ea\u6563\u5085\uf9f7\uf96e\u8f49\u63db (discrete Fourier transform, DFT)\u6240\u5f97\u5230\u3002\u503c\u5f97\u6ce8\u610f\u7684\u662f\uff0c\u7279\u5fb5\u9577\ufa01 N \u6703\u96a8\u8457\uf967\u540c\u7684\u8a9e\uf906 \u800c\uf967\u540c\uff0c\u4f46\u662f\u9019\uf9e8\u7684 DFT \u53d6\u6a23\u9ede\uf969 2P \u5247\u8a2d\u70ba\u4e00\u56fa\u5b9a\u503c\uff0c\u4e5f\u5c31\u662f\uf96b\u8003\u8abf\u8b8a\u983b\u8b5c\u7684\u9577\ufa01\u5c0d \u65bc\u6bcf\u4e00\u500b\u8a9e\uf906\u90fd\u662f\u76f8\u540c\u7684\u3002 \u7531\u5f0f(3.3)\u8207\u5f0f(3.4)\u53ef\u77e5\uff0c\u6211\u5011\u5e0c\u671b\u6bcf\u4e00\u500b\uf901\u65b0\u5f8c\u7684\u7279\u5fb5\u5e8f\uf99c\uff0c\u5176\u8abf\u8b8a\u983b\u8b5c\u7684\u5f37\ufa01\u6210 \u4efd\u80fd\u8da8\u65bc\u4e00\u81f4\uff0c\u800c\u76f8\u4f4d\u6210\u4efd\u5247\u7531\u539f\u59cb\u7684\u7279\u5fb5\u5e8f\uf99c [ ] { } x n \u800c\uf92d\u3002\u63a5\u4e0b\uf92d\uff0c\u6211\u5011\uf9dd\u7528\u6700\u5c0f\u5e73 \u65b9\u5316(least-squares)[10]\u7684\u6700\u4f73\u5316\u6e96\u5247\u6c42\u53d6\u4e00\u65b0\u7684\u7279\u5fb5\uf96b\uf969\u5e8f\uf99c\uff0c\u4f7f\u65b0\u7684\u7279\u5fb5\u5e8f\uf99c [ ] { } y n \u7684\u8abf\u8b8a\u983b\u8b5c\u903c\u8fd1\u5982\u5f0f(3.3)\u7684\uf96b\u8003\u8abf\u8b8a\u983b\u8b5c\uff0c\u5982\u4e0b\u5f0f\u6240\u793a\uff1a [ ] [ ] { } [ ] ( ) ( ) 2 2 2 1 1 2 0 1 0 0 min , 2 nk P N j P k y n n N k n y n y n e Y P N \u03c0 \u03c9 \u2212 \u2212 \u2212 \u2264 \u2264 \u2212 = = = \u2212 \u2265 \u2211 \u2211 (\u5f0f 3.5) \u5176\u4e2d 2P \u70ba DFT \u53d6\u6a23\u9ede\uf969\uff0c N \u70ba\u6b64\u7279\u5fb5\u5e8f\uf99c\u7684\u9ede\uf969\u3002 \u85c9\u7531\u77e9\u9663\u8207\u5411\uf97e\u8868\u793a\u6cd5\uff0c\u6211\u5011\u53ef\u5c07\u5f0f(3.5)\u6539\u5beb\u70ba\u4e0b\u5f0f\uff1a 2 min W = \u2212 y y y Y (\u5f0f 3.6) \u5176\u4e2dW \u662f 2P N \u00d7 \u7684\u77e9\u9663\uff0c\u5176\u7b2c( , ) m n \u9805\u5982\u4e0b\u6240\u793a\uff1a 2 exp , 2 mn mn W j P \u03c0 \u239b \u239e \u239f \u239c = \u2212 \u239f \u239c \u239f \u239d \u23a0 \u800c y \u3001 y \u8207 Y \u5247\u5b9a\u7fa9\u70ba\uff1a [ ] [ ] [ ] 0 1 1 , T y y y n \u23a1 \u23a4 = \u2212 \u23a2 \u23a5 \u23a3 \u23a6 y [ ] [ ] [ ] 0 1 1 , T y y y N \u23a1 \u23a4 = \u2212 \u23a2 \u23a5 \u23a3 \u23a6 y ( ) ( ) ( ) 0 1 2 1 , T P Y Y Y \u03c9 \u03c9 \u03c9 \u2212 \u23a1 \u23a4 = \u23a2 \u23a5 \u23a3 \u23a6 Y \u7531\u65bc y \u70ba\u5be6\uf969\u5411\uf97e\uff0c\u6545\u5f0f(3.6)\u53ef\u6539\u5beb\u70ba\uff1a (",
                        "eq_num": ") ( ) ( ) 2 2"
                    }
                ],
                "section": "",
                "sec_num": null
            },
            {
                "text": "\u6240\u4ee5\uff0c\u5f0f(3.8)\u4e2d\u7684 y \u5373\u70ba LSSF \u6cd5\u6240\u6c42\u5f97\u4e4b\u65b0\u7279\u5fb5\uf96b\uf969\u5e8f\uf99c [ ] { } y n \uff0c\u5176 2P \u9ede\u4e4b DFT \u548c \u5f0f(3.3)\u7684\uf96b\u8003\u8abf\u8b8a\u983b\u8b5c\u4e4b\u9593\u5177\u6709\u6700\u5c0f\u5e73\u65b9\u8aa4\u5dee\u7684\uf97c\u597d\u6027\u8cea\u3002 (\u4e09)\u5f37\ufa01\u983b\u8b5c\u5167\u63d2\u6cd5(magnitude spectrum interpolation, MSI) \u5728\u6b64\u65b9\u6cd5\u4e2d\uff0c\u6211\u5011\u70ba\u6bcf\u4e00\u500b\u5f85\u6b63\u898f\u5316\u7684 N \u9ede\u7279\u5fb5\u5e8f\uf99c [ ] { } 0 1 x n n N \u2264 \u2264 \u2212 \uff0c\u5b9a\u7fa9 \uf9ba\u4e00\u500b N \u9ede\u7684\uf96b\u8003\u8abf\u8b8a\u983b\u8b5c\uff0c\u4f5c\u70ba\u6b64\u7279\u5fb5\u5e8f\uf99c\u4e4b\u8abf\u8b8a\u983b\u8b5c\u6b63\u898f\u5316\u7684\u76ee\u6a19\uff0c\u5982\u4e0b\u5f0f\u6240\u793a\uff1a ( ) ( ) ( ) ( ) exp , 0 1 X k k k Y Y j k N \u03c9 \u03c9 \u03b8 \u03c9 \u2032 \u2032 \u2032 \u2032 = \u2264 \u2264 \u2212 (\u5f0f 3.9) \u5176\u4e2d\u76f8\u4f4d\u6210\u4efd ( ) X k \u03b8 \u03c9 \u2032 \u70ba [ ] x n \u53d6 N \u9ede\u7684 DFT \u6240\u5f97\u3002MSI \u6cd5\u8ddf\u524d\u7bc0\u4e4b LSSF \u6cd5\u7684\u6700\u5927\uf967\u540c \u4e4b\u8655\uff0c\u5728\u65bc\u6b64\u6642\u6211\u5011\u662f\u4f7f\u7528\u4e00\u500b\u8ddf\u539f\u59cb\u7279\u5fb5\u5e8f\uf99c\u9577\ufa01\u76f8\u540c\u7684\uf96b\u8003\u8abf\u8b8a\u983b\u8b5c\uff0c\u800c\u7531\u65bc\uf967\u540c \u8a9e\uf906\u7684\u7279\u5fb5\u5e8f\uf99c\uff0c\u5176\u9ede\uf969 N \u4e5f\u96a8\u4e4b\uf967\u540c\uff0c\u6211\u5011\uf967\u80fd\u5982\u524d\u9762\u7684 LSSF \u6cd5\u4e2d\uff0c\u76f4\u63a5\u62ff 2P \u9ede \u7684\uf96b\u8003\u529f\uf961\u983b\u8b5c\u5bc6\ufa01 ( ) { } 0 2 1 SS k P k P \u03c9 \u2264 \u2264 \u2212 (\u5982\u5f0f(2.1)\u6240\u793a)\uf92d\u6c42\u53d6\u5f0f(3.9)\u4e2d\u7684 N \u9ede\u983b \u8b5c\u5f37\ufa01 ( ) k Y \u03c9 \u2032 \u3002\u7136\u800c\uff0c\u7531\u65bc\u539f\u59cb 2P \u9ede\u7684\uf96b\u8003\u983b\u8b5c\u5176\u6db5\u84cb\u983b\uf961\u7bc4\u570d\u8207\u6b32\u6c42\u7684 ( ) k Y \u03c9 \u2032 \u983b \uf961\u7bc4\u570d\u76f8\u540c\uff0c\u5728\u9019\uf9e8\uff0c\u6211\u5011\u4f7f\u7528\u7dda\u6027\u5167\u63d2(linear interpolation)[10]\u7684\u65b9\u6cd5\uff0c\u85c9\u7531\u5f0f(3.4) \u4e2d \u6240 \u793a \u7684 \u4e4b 2P \u9ede \u7684 ( ) { } 0 2 1 k Y k P \u03c9 \u2264 \u2264 \u2212 \uf92d \u6c42 \u53d6 \u5f0f (3.9) \u4e2d N \u9ede \u7684 ( ) { } 0 1 k Y k N \u03c9 \u2032 \u2032 \u2264 \u2264 \u2212 \u4e4b\u8fd1\u4f3c\u503c\u3002\u4f46\u662f\u5f0f(3.9)\u7684 ( ) { } k Y \u03c9 \u2032 \u70ba\u4e00\u5be6\uf969\u5e8f\uf99c\u4e4b\uf9ea\u6563\u5085\uf9f7\uf96e \u8f49\u63db\uff0c\u5176\u5f37\ufa01\u6210\u4efd ( ) { } k Y \u03c9 \u2032 \u5fc5\u9808\u7b26\u5408\u5de6\u53f3\u5c0d\u7a31\u7684\u6027\u8cea\uff0c\u5373 ( ) ( ) k Nk Y Y \u03c9 \u03c9 \u2032 \u2032 \u2212 = \uff0c\u56e0\u6b64\u6211 \u5011 \u5148 \uf9dd \u7528 ( ) { } k Y \u03c9 \u7684 \u5de6 \u534a \u90e8 \u57f7 \ufa08 \u5167 \u63d2 \u6cd5 \uff0c \u6c42 \u53d6 ( ) { } k Y \u03c9 \u2032 \u7684 \u5de6 \u534a \u90e8 ( ) 0 2 k N Y k \u03c9 \u2032 \u23a7 \u23ab \u23a2 \u23a5 \u23aa \u23aa \u23aa \u23aa \u2032 \u2264 \u2264 \u23a2 \u23a5 \u23a8 \u23ac \u23aa \u23aa \u23a2 \u23a5 \u23a3 \u23a6 \u23aa \u23aa \u23a9 \u23ad \uff0c \u518d \uf9dd \u7528 \u5de6 \u53f3 \u5c0d \u7a31 \u7684 \u6027 \u8cea \uff0c \u6c42 \u53d6 ( ) { } k Y \u03c9 \u2032 \u53f3 \u534a \u90e8 ( ) 1 1 2 k N Y N k N \u03c9 \u2032 \u23a7 \u23ab \u23a2 \u23a5 \u23aa \u23aa \u23aa \u23aa \u2032 \u2212 \u2212 \u2264 \u2264 \u2212 \u23a2 \u23a5 \u23a8 \u23ac \u23aa \u23aa \u23a2 \u23a5 \u23a3 \u23a6 \u23aa \u23aa \u23a9 \u23ad \u3002\u5728\u5f97\u5230 ( ) { } 0 1 k Y k N \u03c9 \u2032 \u2032 \u2264 \u2264 \u2212 \u5f8c\uff0c\u6211\u5011\u5c31\u53ef\u4ee5\u76f4 \u63a5\u5c0d\u5f0f(3.9)\u7684 ( ) { } k Y \u03c9 \u2032 \u505a N",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "",
                "sec_num": null
            }
        ],
        "back_matter": [],
        "bib_entries": {
            "BIBREF0": {
                "ref_id": "b0",
                "title": "Maximum Likelihood Linear Regression for Speaker Adaptation of Continuous Density Hidden Markov Models",
                "authors": [
                    {
                        "first": "C",
                        "middle": [
                            "J"
                        ],
                        "last": "Leggetter",
                        "suffix": ""
                    },
                    {
                        "first": "P",
                        "middle": [
                            "C"
                        ],
                        "last": "Woodland",
                        "suffix": ""
                    }
                ],
                "year": 1995,
                "venue": "Computer Speech and Language",
                "volume": "",
                "issue": "",
                "pages": "",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "C. J. Leggetter and P. C. Woodland, \"Maximum Likelihood Linear Regression for Speaker Adaptation of Continuous Density Hidden Markov Models\", Computer Speech and Language,1995",
                "links": null
            },
            "BIBREF1": {
                "ref_id": "b1",
                "title": "Cepstral Analysis Technique for Automatic Speaker Verification",
                "authors": [
                    {
                        "first": "S",
                        "middle": [],
                        "last": "Furui",
                        "suffix": ""
                    }
                ],
                "year": 1981,
                "venue": "IEEE Trans. on Acoustics, Speech and Signal Processing",
                "volume": "",
                "issue": "",
                "pages": "",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "S. Furui, \"Cepstral Analysis Technique for Automatic Speaker Verification\", IEEE Trans. on Acoustics, Speech and Signal Processing, 1981",
                "links": null
            },
            "BIBREF2": {
                "ref_id": "b2",
                "title": "Multiband and Adaptation Approaches to Robust Speech Recognition",
                "authors": [
                    {
                        "first": "S",
                        "middle": [],
                        "last": "Tiberewala",
                        "suffix": ""
                    },
                    {
                        "first": "H",
                        "middle": [],
                        "last": "Hermansky",
                        "suffix": ""
                    }
                ],
                "year": 1997,
                "venue": "European Conference on Speech Communication and Technology",
                "volume": "",
                "issue": "",
                "pages": "",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "S. Tiberewala and H. Hermansky, \"Multiband and Adaptation Approaches to Robust Speech Recognition\", 1997 European Conference on Speech Communication and Technology (Eurospeech 1997)",
                "links": null
            },
            "BIBREF3": {
                "ref_id": "b3",
                "title": "RASTA Processing of Speech",
                "authors": [
                    {
                        "first": "H",
                        "middle": [],
                        "last": "Hermansky",
                        "suffix": ""
                    },
                    {
                        "first": "N",
                        "middle": [],
                        "last": "Morgan",
                        "suffix": ""
                    }
                ],
                "year": 1994,
                "venue": "IEEE Trans. on Speech and Audio Processing",
                "volume": "",
                "issue": "",
                "pages": "",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "H. Hermansky and N. Morgan, \"RASTA Processing of Speech\", IEEE Trans. on Speech and Audio Processing, 1994",
                "links": null
            },
            "BIBREF4": {
                "ref_id": "b4",
                "title": "MVA Processing of Speech Features",
                "authors": [
                    {
                        "first": "C-P",
                        "middle": [],
                        "last": "Chen",
                        "suffix": ""
                    },
                    {
                        "first": "J-A",
                        "middle": [],
                        "last": "Bilmes",
                        "suffix": ""
                    }
                ],
                "year": 2006,
                "venue": "Speech, and Language Processing",
                "volume": "",
                "issue": "",
                "pages": "",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "C-P. Chen and J-A. Bilmes, \"MVA Processing of Speech Features\", IEEE Trans. on Audio, Speech, and Language Processing, 2006",
                "links": null
            },
            "BIBREF5": {
                "ref_id": "b5",
                "title": "Cepstral Gain Normalization for Noise Robust Speech Recognition",
                "authors": [
                    {
                        "first": "S",
                        "middle": [],
                        "last": "Yoshizawa",
                        "suffix": ""
                    },
                    {
                        "first": "N",
                        "middle": [],
                        "last": "Hayasaka",
                        "suffix": ""
                    },
                    {
                        "first": "N",
                        "middle": [],
                        "last": "Wada",
                        "suffix": ""
                    },
                    {
                        "first": "Y",
                        "middle": [],
                        "last": "Miyanaga",
                        "suffix": ""
                    }
                ],
                "year": 2004,
                "venue": "2004 International Conference on Acoustics, Speech and Signal Processing",
                "volume": "",
                "issue": "",
                "pages": "",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "S. Yoshizawa, N. Hayasaka, N. Wada and Y. Miyanaga, \"Cepstral Gain Normalization for Noise Robust Speech Recognition\", 2004 International Conference on Acoustics, Speech and Signal Processing (ICASSP 2004)",
                "links": null
            },
            "BIBREF6": {
                "ref_id": "b6",
                "title": "Optimization of Temporal Filters for Constructing Robust Features in Speech Recognition",
                "authors": [
                    {
                        "first": "J-W",
                        "middle": [],
                        "last": "Hung",
                        "suffix": ""
                    },
                    {
                        "first": "L-S",
                        "middle": [],
                        "last": "Lee",
                        "suffix": ""
                    }
                ],
                "year": 2006,
                "venue": "Speech and Language Processing",
                "volume": "",
                "issue": "",
                "pages": "",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "J-W. Hung and L-S. Lee, \"Optimization of Temporal Filters for Constructing Robust Features in Speech Recognition\", IEEE Trans. on Audio, Speech and Language Processing, 2006",
                "links": null
            },
            "BIBREF7": {
                "ref_id": "b7",
                "title": "Temporal Structure Normalization of Speech Feature for Robust Speech Recognition",
                "authors": [
                    {
                        "first": "X",
                        "middle": [],
                        "last": "Xiao",
                        "suffix": ""
                    },
                    {
                        "first": "E-S",
                        "middle": [],
                        "last": "Chng",
                        "suffix": ""
                    },
                    {
                        "first": "Haizhou",
                        "middle": [],
                        "last": "Li",
                        "suffix": ""
                    }
                ],
                "year": 2007,
                "venue": "IEEE Signal Processing Letters",
                "volume": "14",
                "issue": "",
                "pages": "",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "X. Xiao, E-S. Chng, and Haizhou Li, \"Temporal Structure Normalization of Speech Feature for Robust Speech Recognition\", IEEE Signal Processing Letters, vol. 14, 2007",
                "links": null
            },
            "BIBREF8": {
                "ref_id": "b8",
                "title": "The AURORA Experimental Framework for the Performance Evaluations of Speech Recognition Systems under Noisy Conditions",
                "authors": [
                    {
                        "first": "H",
                        "middle": [
                            "G"
                        ],
                        "last": "Hirsch",
                        "suffix": ""
                    },
                    {
                        "first": "D",
                        "middle": [],
                        "last": "Pearce",
                        "suffix": ""
                    }
                ],
                "year": 2000,
                "venue": "Proceedings of ISCA IIWR ASR2000",
                "volume": "",
                "issue": "",
                "pages": "",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "H. G. Hirsch and D. Pearce, \"The AURORA Experimental Framework for the Performance Evaluations of Speech Recognition Systems under Noisy Conditions\", Proceedings of ISCA IIWR ASR2000, Paris, France, 2000",
                "links": null
            },
            "BIBREF9": {
                "ref_id": "b9",
                "title": "The McGraw-Hill International, 3 rd edition",
                "authors": [
                    {
                        "first": "S",
                        "middle": [
                            "K"
                        ],
                        "last": "Mitra",
                        "suffix": ""
                    }
                ],
                "year": 2006,
                "venue": "",
                "volume": "",
                "issue": "",
                "pages": "",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "S. K. Mitra, \"Digital Signal Processing\", The McGraw-Hill International, 3 rd edition, 2006",
                "links": null
            },
            "BIBREF10": {
                "ref_id": "b10",
                "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
            }
        },
        "ref_entries": {
            "TABREF0": {
                "text": "\u95dc\u9375\u8a5e\uff1a\u8a9e\u97f3\u8fa8\uf9fc\u3001\u8abf\u8b8a\u983b\u8b5c\u6b63\u898f\u5316\u3001\u5f37\u5065\u6027\u8a9e\u97f3\u7279\u5fb5\uf96b\uf969keyword: speech recognition, modulation spectrum, robust speech features \u4e00\u3001\u7dd2\uf941 \u81ea\u52d5\u8a9e\u97f3\u8fa8\uf9fc\u7cfb\u7d71(automatic speech recognition systems, ASR)\uff0c\u85c9\u7531\u591a\uf98e\uf92d\u5404\u65b9\u5b78\u8005\u7684",
                "num": null,
                "content": "<table><tr><td>\u7814\u7a76\u767c\u5c55\uff0c\u9010\u6f38\u9054\u5230\u5be6\u969b\u61c9\u7528\u7684\u968e\u6bb5\uff0c\u800c\u70ba\u4eba\uf9d0\u751f\u6d3b\u5e36\uf92d\uf901\u591a\u65b9\uf965\u8207\u5e6b\u52a9\uff0c\u96d6\u7136\u9084\uf967\u80fd</td></tr><tr><td>\u9054\u5230\u4e00\u500b\u5b8c\u7f8e\u7684\u5730\u6b65\uff0c\u4f46\u662f\u9019\u65b9\u9762\u7684\u6280\u8853\u4ecd\u4e00\u76f4\uf967\u65b7\u5730\u9032\u6b65\u7576\u4e2d\u3002</td></tr><tr><td>\u81ea\u52d5\u5316\u8a9e\u97f3\u8fa8\u8a8d\u4ecd\u6709\u8a31\u591a\u76f8\u7576\u5177\u6709\u6311\u6230\u6027\u7684\u7814\u7a76\u8ab2\u984c\uff0c\u7531\u65bc\u8a9e\u97f3\u7684\u8b8a\uf962\u6027\u592a\u591a\uff0c\uf9b5</td></tr><tr><td>\u5982\u6bcf\u4f4d\u8a9e\u8005\uf96f\u8a71\u7684\u65b9\u5f0f\u8207\u53e3\u6c23\u90fd\uf967\u4e00\u6a23\u3001\uf967\u540c\u8a9e\u8a00\u6709\uf967\u540c\u7684\u7279\u6027\u3001\u8a9e\u8005\u7576\u6642\uf96f\u8a71\u7684\u60c5</td></tr><tr><td>\u7dd2\u3001\u8a9e\u8005\u6240\u8655\u7684\u74b0\u5883\u662f\u5426\u6709\u5176\u4ed6\u96dc\u8a0a\u5e72\u64fe\u7b49\uff0c\u9019\u4e9b\u8b8a\uf962\u5c0d\u65bc\u8a9e\u97f3\u8fa8\uf9fc\u6548\u679c\u90fd\u6709\u5f71\u97ff\u3002\u5728</td></tr><tr><td>\u771f\u5be6\u61c9\u7528\u74b0\u5883\u4e0b\uff0c\u8a9e\u97f3\u8fa8\uf9fc\u7cfb\u7d71\u6240\u9047\u5230\u7684\u4e3b\u8981\u554f\u984c\u5176\u4e2d\uf978\u500b\uff0c\u5206\u5225\u70ba\uff1a</td></tr><tr><td>(\u4e00)\u8a9e\u8005\uf967\u5339\u914d(speaker mismatch)</td></tr><tr><td>\u8a9e\u8005\uf967\u5339\u914d\u7684\u554f\u984c\u662f\u56e0\u70ba\uf96f\u8a71\u8005\u5148\u5929\u689d\u4ef6(\u5982\u53e3\u8154\u5f62\uf9fa)\u8207\u5f8c\u5929\u7fd2\u6163(\u5982\uf96f\u8a71\u8154\u8abf)</td></tr><tr><td>\u7684\u5dee\uf962\u6240\u7522\u751f\u7684\u8b8a\uf962\u6027\uff0c\u56e0\u6b64\u7576\u4ee5\u7279\u5b9a\u8a9e\u8005\u6240\u8a13\uf996\u51fa\uf92d\u7684\u8072\u5b78\u6a21\u578b\uf92d\u8fa8\uf9fc\uf967\u5c6c\u65bc\u6b64\u7279\u5b9a</td></tr><tr><td>\u8a9e\u8005\u7684\u8a9e\u97f3\u6642\uff0c\u8fa8\uf9fc\u6548\u679c\u5e38\u6703\u660e\u986f\u4e0b\ufa09\uff0c\u800c\u8981\u514b\u670d\u9019\u4e00\uf9d0\u554f\u984c\u7684\u65b9\u6cd5\uff0c\u901a\u5e38\u662f\u4f7f\u7528\u6240\u8b02</td></tr><tr><td>\u7684\u8a9e\u8005\u8abf\u9069(speaker adaptation)\u6280\u8853\u3002\u4e5f\u5c31\u662f\u5c07\u539f\u672c\u8a13\uf996\u51fa\uf92d\u7684\u8072\u5b78\u6a21\u578b\u8abf\u9069\u6210\u63a5\u8fd1\u7576</td></tr><tr><td>\u4e0b\u8a9e\u8005\u4e4b\u8a9e\u97f3\u7279\u6027\u7684\u6a21\u578b[1]\uff0c\u5982\u6b64\uf965\u53ef\u63d0\u9ad8\u8fa8\uf9fc\uf961\u3002</td></tr><tr><td>(\u4e8c)\u74b0\u5883\uf967\u5339\u914d(environment mismatch)</td></tr><tr><td>\u74b0\u5883\uf967\u5339\u914d\u7684\u554f\u984c\u662f\u56e0\u70ba\u8a9e\u97f3\u8fa8\uf9fc\u7cfb\u7d71\u8a13\uf996\u74b0\u5883\u8207\u6211\u5011\u5be6\u9a57\u6216\u61c9\u7528\u6642\u7684\u74b0\u5883\uf967\u540c</td></tr><tr><td>\u6240\u81f4\uff0c\u5176\u8b8a\uf962\u56e0\u5b50\u4e3b\u8981\u5305\u542b\uf9ba\u52a0\u6210\u6027\u96dc\u8a0a(additive noise)\uff0c\u5982\uf902\u7ad9\u56db\u5468\u7684\u96dc\u8a0a\u3001\u5608\u96dc\u8857</td></tr><tr><td>\u9053\u7684\u4eba\u8072\u6216\uf902\u8072\u7b49\uff0c\u53ca\u647a\u7a4d\u6027\u96dc\u8a0a(convolutional noise)\uff0c\u5982\uf967\u540c\u7684\u6709\u7dda\u6216\u7121\u7dda\u96fb\u8a71\u7dda\uf937</td></tr><tr><td>\u6216\u9ea5\u514b\u98a8\u6240\u9020\u6210\u7684\u901a\u9053\u6548\u61c9\u7b49\uff0c\u8a9e\u97f3\u8fa8\uf9fc\u7cfb\u7d71\u5e38\u6703\u56e0\u9019\u4e9b\u96dc\u8a0a\u7684\u5f71\u97ff\u4f7f\u8fa8\uf9fc\uf961\ufa09\u4f4e\u3002\u4e0b</td></tr><tr><td>\u5716\u4e00\u70ba\u4e7e\u6de8\u8a9e\u97f3\u53d7\u96dc\u8a0a\u5e72\u64fe\u4e4b\u793a\u610f\u5716\u3002</td></tr><tr><td>\u5716\u4e00\u3001\u4e7e\u6de8\u8a9e\u97f3\u53d7\u96dc\u8a0a\u5e72\u64fe\u4e4b\u793a\u610f\u5716</td></tr><tr><td>\u5728\u8af8\u591a\ufa09\u4f4e\u96dc\u8a0a\u5f71\u97ff\u3001\u6539\u9032\u8a9e\u97f3\u7279\u5fb5\u7684\u5f37\u5065\u6027\u6280\u8853\u4e2d\uff0c\u6709\u4e00\u5927\uf9d0\u7684\u65b9\u6cd5\u5176\u76ee\u6a19\u662f\u627e\u51fa\u4e00</td></tr><tr><td>\u5f37\u5065\u8a9e\u97f3\u7279\u5fb5\u8868\u793a\u5f0f(robust speech feature representation)\uff0c\ufa09\u4f4e\u8a9e\u97f3\u7279\u5fb5\u5c0d\u96dc\u8a0a\u7684\u654f\u611f</td></tr><tr><td>\ufa01\uff0c\u4f7f\u96dc\u8a0a\u7522\u751f\u7684\u5931\u771f\u8b8a\u5c0f\u3002\u6b64\uf9d0\u8457\u540d\u7684\u65b9\u6cd5\u5305\u62ec\uf9ba\u5012\u983b\u8b5c\u5e73\u5747\u6d88\u53bb\u6cd5(cepstral mean</td></tr><tr><td>subtraction, CMS)</td></tr></table>",
                "html": null,
                "type_str": "table"
            },
            "TABREF5": {
                "text": "1\u6cd5\u70ba\u7b2c\u4e8c\u7ae0\u4e2d\u6240\u4ecb\u7d39\u4e4b\u539f\u59cb TSN \u6cd5\uff0c\u800c TSN-2 \u6cd5\u5247\u662f\u5c07\u539f\u59cb TSN \u6cd5\u4e2d\u76f4\uf9ca\u589e\u76ca\u6b63\u898f\u5316\u6b65\u9a5f\uf96d\uf976\u6240\u5f97\u7684\u4fee\u6b63\u6cd5\uff0cTSN-1\u8207 TSN-2\u6240\u5f97\u4e4b \u6642\u9593\u5e8f\uf99c\uf984\u6ce2\u5668\u9577\ufa01\u7686\u8a2d\u70ba21\uff0c\u6b64\u503c\u662f\u76f4\u63a5\uf96b\u8003 TSN \u6cd5\u7684\u6587\u737b[8]\u800c\uf92d\u3002ERTF \u6cd5\u6240\u5f97 \u7684\u6642\u9593\u5e8f\uf99c\uf984\u6ce2\u5668\u9577\ufa01\u70ba21\uff0c\u800c LSSF \u8207 MSI \u6cd5\u6240\u7528\u7684 DFT \u9ede\uf969 2P (\u5982\u5f0f(3.3)\u6240\u793a)\u5247 \u74b0\u5883\u800c\u8a00\uff0c\u5b83\u5011\u5206\u5225\u4f7f\u8fa8\uf9fc\uf961\u63d0\u5347\uf9ba 12.99%\u300111.91%\u8207 11.15%\uff0c\u5c0d Set B \u74b0\u5883 \u800c\u8a00\uff0c\u8fa8\uf9fc\uf961\u5206\u5225\u63d0\u5347\uf9ba 18.61%\u300117.90%\u8207 17.05%\uff0c\u5728 Set C \u74b0\u5883\u4e0b\uff0c\u8fa8\uf9fc\uf961\u5206\u5225\u63d0 \u5347\uf9ba 6.52%\u30015.90%\u8207 5.46%\u3002\u9019\u4e09\u7a2e\u65b9\u6cd5\u4e2d\uff0c\u53c8\u4ee5 ERTF \u6cd5\u7684\u8868\u73fe\u6700\u597d\uff0c\u660e\u986f\u512a\u65bc LSSF \u6cd5\u8207 MSI \u6cd5\uff0c\u4f46\u5b83\u5011\u6240\u80fd\u9054\u5230\u7684\u76f8\u5c0d\u932f\u8aa4\ufa09\u4f4e\uf961\u90fd\u9ad8\u9054 40%\u4ee5\u4e0a\uff0c\u660e\u986f\u512a\u65bc TSN-1 \u6cd5 \u8207 TSN-2 \u6cd5\u3002\u53e6\u5916\uff0c\u503c\u5f97\u4e00\u63d0\u7684\u662f\uff0cTSN-2 \u6cd5\u5728 Set C \u4e2d\u7684\u6548\u679c\u6bd4 TSN-1 \u8207\u57fa\u790e\u5be6\u9a57 \u5dee\uff0c\u4f46 ERTF\u3001LSSF \u8207 MSI \u6cd5\u537b\u672a\u6709\u9019\u6a23\u7684\uf967\uf97c\u7d50\u679c\u3002 \u25cb 3 \uf978\u7a2e\u76ee\u524d\u5ee3\u70ba\u4eba\u7528\u7684\u7279\u5fb5\u6b63\u898f\u5316\u6280\u8853\uff1aCMVN \u6cd5\u8207 MVA \u6cd5\uff0c\u5c0d\u8fa8\uf9fc\uf961\u7684\u63d0\u5347\u90fd\u5341 \u5206\u660e\u986f\uff0cCMVN \u7684\u6548\u80fd\u8207\u6211\u5011\u6240\u63d0\u7684\u4e09\u7a2e\u65b0\u65b9\u6cd5\u5927\u81f4\u76f8\u540c\uff0c\u4f46\u7d50\u5408\uf9ba CMVN \u8207 ARMA \uf984\u6ce2\u8655\uf9e4\u7684 MVA \u6cd5\u5176\u6548\u80fd\u53c8\u6bd4 CMVN \u6cd5\uf92d\u7684\u597d\uff0c\u57fa\u65bc\u9019\u6a23\u7684\u89c0\u5bdf\uff0c\u5728\u4e0b\uf978\u5c0f\u7bc0\u4e2d\uff0c \u6211\u5011\u5c07\u8a66\u8457\u628a\u5404\u7a2e\u8abf\u8b8a\u983b\u8b5c\u6b63\u898f\u5316\u6cd5\u8207 CMVN \u6cd5\u6216 MVA \u6cd5\u52a0\u4ee5\u6574\u5408\uff0c\u63a2\u8a0e\u662f\u5426\u80fd\u5e36 \uf92d\u8fa8\uf9fc\uf961\u4e0a\uf901\u986f\u8457\u7684\u9032\u6b65\u3002 \u7576\u6211\u5011\u4f7f\u7528 LSSF \u6cd5\u8207 MSI \u6cd5\u6642\uff0c\u6211\u5011\u6703\u5c07\u539f\u59cb\u70ba N \u9ede\u7684\u7279\u5fb5\u5e8f\uf99c\u8f49\u63db\u6210 2P \u9ede\u4e4b \u529f\uf961\u983b\u8b5c\u5bc6\ufa01\u6216\uf9ea\u6563\u983b\u8b5c\uff0c\u7136\u800c\u7531\u65bc\u901a\u5e38 MSI \u6cd5\u7684\u6548\u679c\uff0c\u6211\u5011\u7a31\u9019\u6a23\u4fee\u6539\u7d50\u679c\u5206\u5225\u70ba\u4fee\u6b63\u5f0f LSSF \u6cd5(modified LSSF)\u8207\u4fee\u6b63\u5f0f MSI \u6cd5(modified MSI)\u3002 \u5716\u5341\u70ba\u539f\u59cb\u8207\u4fee\u6b63\u5f0f LSSF \u8207 MSI \u4f5c\u7528\u65bc\u539f\u59cb MFCC \u7279\u5fb5\u4e4b\u5e73\u5747\u8fa8\uf9fc\uf961\u9577\u689d\u5716\u3002 \u7531\u6b64\u5716\u4e2d\u53ef\u4ee5\u770b\u51fa\u4fee\u6b63\u5f0f LSSF \u6cd5\u76f8\u8f03\u65bc\u539f\u59cb LSSF \u6cd5\u800c\u8a00\uff0c\u5e73\u5747\u8fa8\uf9fc\uf961\u6709 0.67%\u7684\u63d0 \u5347\uff0c\u800c\u4fee\u6b63\u5f0f MSI \u76f8\u8f03\u65bc\u539f\u59cb MSI \u800c\u8a00\uff0c\u5728\u5e73\u5747\u8fa8\uf9fc\uf961\u4e0a\u6709 0.92%\u7684\u63d0\u5347\u3002\u7531\u6b64\u6211\u5011 \u9a57\u8b49\uf9ba\uff0c\u5728\u4fee\u6b63\u6cd5\u4e2d\u6240\u4f5c\u7684\u7a97\u5316\u8655\uf9e4\u78ba\u5be6\u80fd\u6709\u6548\u6539\u9032 LSSF \u6cd5\u8207 MSI \u6cd5\u7684\u6548\u80fd\u3002 \u8b5c\u6b63\u898f\u5316\u6cd5\u8207 CMVN \u6cd5\u4f5c\u7d50\u5408\uff0c\u610f\u5373\u539f\u59cb MFCC \u7279\u5fb5\u5148\u7d93\u904e CMVN \u6cd5\u8655\uf9e4\u5f8c\uff0c\u518d\u4ee5 \u5404\u7a2e\u8abf\u8b8a\u983b\u8b5c\u6b63\u898f\u5316\u6cd5\u5206\u5225\u4f5c\u8655\uf9e4\u3002\u4ee5\u4e0b\u6211\u5011\u6e2c\u8a66\u9019\u6a23\u7684\u7d50\u5408\u662f\u5426\u6709\u52a0\u6210\u6027\u7684\u6548\u679c\u3002\u5728 \u8868\u4e8c\u4e2d\uff0c\u6211\u5011\u6574\uf9e4\uf9ba CMVN \u6cd5\u5206\u5225\u7d50\u5408 TSN-1\u3001TSN-2\u3001ERTF\u3001LSSF\u3001MSI \u53ca ARMA \uf984\u6ce2\u6cd5(MVA)[5]\u5404\u65b9\u6cd5\u6240\u5f97\u7684\u5e73\u5747\u8fa8\uf9fc\uf961\uff0c\u5176\u4e2d AR \u8207 RR \u5206\u5225\u70ba\u76f8\u8f03\u65bc\u55ae\u4e00 CMVN \u7d50\u679c\u4e4b\u7d55\u5c0d\u932f\u8aa4\ufa09\u4f4e\uf961(absolute error rate reduction)\u548c\u76f8\u5c0d\u932f\u8aa4\ufa09\u4f4e\uf961(relative error rate reduction)\u3002 \u8868\u4e8c\u3001\u5404\u8abf\u8b8a\u983b\u8b5c\u8655\uf9e4\u6cd5\u4f5c\u7528\u65bc CMVN \u8655\uf9e4\u5f8c\u4e4b MFCC \u7279\u5fb5\u6240\u5f97\u4e4b\u8fa8\uf9fc\uf961(%) \u6cd5\u4f5c\u7528\u65bc CMVN \u8655\uf9e4\u904e\u7684 MFCC \u7279\u5fb5\uff0c\u5176\u6539\u9032\u8fa8\uf9fc\uf961\u7684\u6548\u80fd\u5341\u5206\u986f\u8457\uff0c\u76f8 \u8f03\u65bc\u55ae\u4e00 CMVN \u6cd5\u800c\u8a00\uff0c\u5728 Set A\u3001Set B \u8207 Set C \u74b0\u5883\u4e0b\u5206\u5225\u5177\u6709 4.39%\u30014.47%\u8207 3.43% \u7684\u6574\u9ad4\u8fa8\uf9fc\uf961\u9084\u662f\u6bd4 TSN-1 \uf92d\u7684\u597d\uff0c\u518d\u4e00\u6b21\u9a57\u8b49\u539f TSN \u6cd5\u4e2d\u76f4 \uf9ca\u589e\u76ca\u6b63\u898f\u5316\u7684\u6b65\u9a5f\u61c9\u8a72\u662f\uf967\u5fc5\u8981\u7684\u3002 \u25cb 2 \u7576\u6211\u5011\u6240\u63d0\u51fa\u7684\u65b0\u65b9\u6cd5 ERTF\u3001LSSF \u8207 MSI \u6cd5\u4f5c\u7528\u65bc CMVN \u8655\uf9e4\u5f8c\u4e4b MFCC \u7279\u5fb5 \u6642\uff0c\u76f8\u8f03\u65bc\u55ae\u4e00 CMVN \u6240\u5f97\u7684\u8fa8\uf9fc\uf961\u800c\u8a00\uff0c\u7686\u5e36\uf92d\u5341\u5206\u986f\u8457\u7684\u6539\u5584\uff0c\uf9b5\u5982\u5728 Set A \u74b0 \u5883\u4e0b\uff0c\u9019\u4e09\u7a2e\u65b9\u6cd5\u5206\u5225\u5177\u6709 4.58%\u30014.09%\u8207 4.56%\u7684\u8fa8\uf9fc\uf961\u63d0\u6607\uff0c\u6b64\u986f\u793a\uf9ba\u9019\u4e09\u7a2e\u65b0 \u65b9\u6cd5\u8207 CMVN \u6709\uf97c\u597d\u7684\u52a0\u6210\u6027\u3002\u800c\u4e09\u500b\u65b0\u65b9\u6cd5\u4e2d\uff0cERTF \u548c MSI \u6cd5\u8868\u73fe\u7684\u90fd\u6bd4 TSN-1 \u6216 TSN-2 \u6cd5\uf901\u597d\uff0c\u96d6\u7136 LSSF \u6cd5\u8868\u73fe\u7a0d\uf967\u5982\u9810\u671f\uff0c\u4f46\u662f\u53ef\u80fd\u539f\u56e0\u5728\u65bc\u524d\u4e00\u7bc0\u6240\u8a0e\uf941\u5230\u7684\uff0c \u539f\u59cb LSSF \u6cd5\u548c MSI \u6cd5\u53ef\u80fd\u6703\u7522\u751f\u983b\u8b5c\u907a\uf94e(leakage)\u73fe\u8c61\u4e4b\uf9d0\u7684\uf967\uf97c\u6548\u61c9\uff0c\u56e0\u6b64\u5728\u5f8c \u9762\uff0c\u6211\u5011\u5c07\u4ee5\u4fee\u6b63\u5f0f LSSF \u6cd5\u8207 MSI \u6cd5\u7d50\u5408 CMVN \u6cd5\uf92d\u63a2\u8a0e\u5176\u53ef\u80fd\u7684\u6539\u9032\u6548\u679c\u3002 \u25cb 3 \u5728\u4e4b\u524d\u7684\u8868\u4e00\uf969\u64da\u986f\u793a\uff0c\u7576\u4f5c\u7528\u65bc\u539f\u59cb MFCC \u7279\u5fb5\u6642\uff0cERTF \u8868\u73fe\u6bd4 LSSF \u548c MSI \u6cd5\uf901\u597d\u3002\u4f46\u662f\u5728\u9019\uf9e8\u6211\u5011\u767c\u73fe\u7576\u9019\u4e9b\u65b9\u6cd5\u548c CMVN \u6cd5\u7d50\u5408\u6642\uff0c\u5176\u6548\u679c\u8b8a\u5f97\u5341\u5206\u63a5\u8fd1\uff0c \u9019\u4e5f\u610f\u5473\u8457 CMVN \u6cd5\u78ba\u5be6\u5df2\u5c0d\u539f\u59cb MFCC \u7279\u5fb5\u4f5c\uf9ba\u5341\u5206\u6709\u6548\u7684\u5f37\u5065\u6027\u8655\uf9e4\uff0c\u800c\u4f7f\u5f8c\u7e8c \u7684\u6539\u9032\u6280\u8853\uff0c\u5176\u9032\u6b65\u7684\u7a7a\u9593\u76f8\u5c0d\u8b8a\u5c0f\u3002 \u5982\u4e4b\u524d\u6240\u63d0\u5230\u7684\uff0c\u539f\u59cb LSSF \u6cd5\u548c MSI \u6cd5\u53ef\u80fd\u6709\u983b\u8b5c\u907a\uf94e(leakage)\u7684\u7f3a\u9ede\uff0c\u56e0\u6b64\u6211 \u5011\u9019\uf9e8\u4f7f\u7528\u4e4b\u524d\u6240\u8ff0\u7684\u4fee\u6b63\u5f0f LSSF \u8207 MSI \u6cd5\uff0c\u4f5c\u7528\u65bc CMVN \u8655\uf9e4\u5f8c\u7684 MFCC \u7279\u5fb5\uff0c \u5373\u5728\u6b64\uf978\u65b9\u6cd5\u88dc\uf9b2\u7684\u7a0b\u5e8f\u524d\u5148\u5c07\u539f\u59cb N \u9ede\u7684 CMVN \u6cd5\u8655\uf9e4\u5f8c\u4e4b MFCC \u7279\u5fb5\u5e8f\uf99c\u4e58\u4e0a\u4e00 \u6f22\uf95f\u7a97(Hanning window)\uff0c\u89c0\u5bdf\u9019\u6a23\u7684\u64cd\u4f5c\u662f\u5426\u53ef\u9032\u4e00\u6b65\u63d0\u5347\u539f\u59cb LSSF \u6cd5\u8207 MSI \u6cd5\u7d50 \u5408 CMVN \u6cd5\u7684\u6548\u679c\u3002 \u5716\u5341\u4e00\u70ba\u539f\u59cb\u8207\u4fee\u6b63\u5f0f LSSF \u8207 MSI \u4f5c\u7528\u65bc CMVN \u6cd5\u8655\uf9e4\u5f8c MFCC \u7279\u5fb5\u4e4b\u5e73\u5747\u8fa8\uf9fc\uf961 \u9577\u689d\u5716\u3002\u7531\u6b64\u5716\u53ef\u4ee5\u770b\u51fa\uff0c\u5728\u7d50\u5408 CMVN \u6cd5\u5f8c\uff0c\u4fee\u6b63\u5f0f LSSF \u6cd5\u76f8\u8f03\u65bc\u539f\u59cb LSSF \u6cd5\u800c \u8a00\uff0c\u6709 0.85%\u4e4b\u5e73\u5747\u8fa8\uf9fc\uf961\u7684\u63d0\u5347\uff0c\u540c\u6a23\u5730\uff0c\u4fee\u6b63\u5f0f MSI \u6cd5\u76f8\u5c0d\u65bc\u539f\u59cb MSI \u6cd5\u800c\u8a00\uff0c \u6709 0.47%\u4e4b\u5e73\u5747\u8fa8\uf9fc\uf961\u7684\u63d0\u5347\uff0c\u4e8c\u8005\u5e73\u5747\u8fa8\uf9fc\uf961\u7686\u8d85\u904e 90%\u3002\u6b64\u5916\uff0c\u7576\u8207\u8868\u4e8c\u7684\uf969\u64da\u6bd4 \u8f03\uff0c\u6211\u5011\u770b\u5230\u9019\uf978\u7a2e\u4fee\u6b63\u5f0f\u65b9\u6cd5\u7d50\u5408 CMVN \u6cd5\u5f8c\u5728\u7e3d\u5e73\u5747\u8fa8\uf9fc\uf961\u4e0a\u7686\u660e\u986f\u512a\u65bc\u8207 CMVN \u6cd5\u7d50\u5408\u7684 TSN-1 \u6cd5(89.49%)\u8207 TSN-2 \u6cd5(89.76%)\uff0c\u4ee5\u4e0a\u7d50\u679c\u90fd\u986f\u793a\uf9ba\u9019\u6a23\u7684\u4fee \u6b63\u78ba\u5be6\u80fd\u6709\u6548\u6539\u9032\u539f\u65b9\u6cd5\u7684\u7f3a\u9ede\uff0c\u800c\u63d0\u5347\u5176\u6548\u80fd\u3002 \u80fd\u5920\u5c0d\u96dc\u8a0a\u74b0\u5883\u4e0b\u7684\u8a9e\u97f3\u7279\u5fb5\u6709\u660e\u986f\u7684\u5f37\u5065\u5316\u6548\u679c\uff0c\u800c\u5e36\uf92d\u5341\u5206\u986f\u8457\u7684\u8fa8\uf9fc\uf961\u63d0\u5347\uff0c\u4e14 \u5176\u6548\u80fd\u512a\u65bc CMVN\uff0c\u56e0\u6b64\u5728\u9019\uf9e8\uff0c\u6211\u5011\u5c07\u5404\u7a2e\u8abf\u8b8a\u983b\u8b5c\u6b63\u898f\u5316\u6cd5\u8207 MVA \u6cd5\u4f5c\u7d50\u5408\uff0c\u4e5f \u5c31\u662f\u628a\u9019\u4e9b\u6b63\u898f\u5316\u6cd5\u4f5c\u7528\u65bc\u7d93 MVA \u6cd5\u8655\uf9e4\u5f8c\u4e4b MFCC \u7279\u5fb5\u4e0a\uff0c\u4ee5\u6aa2\u8996\u9019\u4e9b\u6b63\u898f\u5316\u6cd5\u8207 MVA \u6cd5\u662f\u5426\u6709\u52a0\u6210\u6027\u3002\u5be6\u9a57\u4e2d\u6211\u5011\u8a2d\u5b9a MVA \u6cd5\u4e2d\u7684 ARMA \uf984\u6ce2\u5668\u968e\uf969\u70ba 2(\uf96b\u7167[5])\u3002 \u5728\u4e0b\u8868\u4e09\u4e2d\uff0c\u6211\u5011\uf99c\u51fa\uf9ba MVA \u6cd5\u5206\u5225\u7d50\u5408 TSN-1\u3001TSN-2\u3001ERTF\u3001LSSF \u8207 MSI \u5404\u65b9 \u6cd5\u6240\u5f97\u7684\u5e73\u5747\u8fa8\uf9fc\uf961\uff0c\u5176\u4e2d AR \u8207 RR \u5206\u5225\u70ba\u76f8\u8f03\u65bc\u55ae\u4e00 MVA \u6cd5\u4e4b\u7d50\u679c\u7684\u7d55\u5c0d\u932f\u8aa4\ufa09 \u4f4e\uf961(absolute error rate reduction)\u548c\u76f8\u5c0d\u932f\u8aa4\ufa09\u4f4e\uf961(relative error rate reduction)\u3002 \u8868\u4e09\u3001\u5404\u8abf\u8b8a\u983b\u8b5c\u8655\uf9e4\u6cd5\u4f5c\u7528\u65bc MVA \u8655\uf9e4\u5f8c\u4e4b MFCC \u7279\u5fb5\u6240\u5f97\u4e4b\u8fa8\uf9fc\uf961(%) \u5728\u7d50\u5408 MVA \u5f8c\uff0c\u5176\u6548\u80fd\u6709\u660e\u986f\u7684\u63d0\u6607\uff0c\u800c TSN-1 \u548c TSN-2 \u4e4b\u9593\u7684\u5dee\uf962\u96d6\u7136\uf967\u660e\u986f\uff0c\u4f46\u662f\uf96d\uf976\u76f4\uf9ca\u589e\u76ca\u6b63\u898f\u5316\u6b65\u9a5f\u7684 TSN-2 \u6cd5\u4ecd\u7136\u8868\u73fe\u6bd4\u8f03\u597d\uff0c \u76f8\u5c0d\u65bc\u55ae\u4e00 MVA \u6cd5\u7684\u7d50\u679c\u800c\u8a00\uff0c\u7d50\u5408\uf9ba MVA \u6cd5\u4e4b TSN-1 \u5728\u7e3d\u5e73\u5747\u8fa8\uf9fc\uf961\u4e0a\u63d0\u5347 1.36%\uff0c\u800c TSN-2 \u63d0\u5347\uf9ba 1.52%\u3002\u6b64\u5916\uff0c\u7d50\u5408 MVA \u6cd5\u4e4b\u5f8c\uff0c\u6211\u5011\u63d0\u51fa\u7684 ERTF\u3001LSSF \u8207 MSI \u4e09\u500b\u65b9\u6cd5\u4ecd\u512a\u65bc TSN-1 \u8207 TSN-2\uff0c\u800c\u5176\u4e2d\u4ee5 MSI \u6cd5\u6700\u597d\uff0c\u5728\u8fa8\uf9fc\uf961\u4e0a\u63d0\u5347 1.71%\uff0c \u5176\u6b21\u70ba LSSF \u6cd5\uff0c\u63d0\u5347\uf9ba 1.67%\uff0c ERTF \u6cd5\u5247\u63d0\u5347\uf9ba 1.59%\u3002\u5118\u7ba1\u5982\u6b64\uff0c\u6211\u5011\u53ef\u660e\u986f\u770b \u51fa\uff0c\u9019\u4e9b\u65b9\u6cd5\u5728\u7d50\u5408 MVA \u6cd5\u5f8c\uff0c\u6240\u5e36\uf92d\u7684\u8fa8\uf9fc\uf961\u63d0\u5347\u7a0b\ufa01\u76f8\u5c0d\u800c\u8a00\u90fd\u5df2\u5341\u5206\u63a5\u8fd1\u3002 \u5982\u540c\u524d\u7bc0\u6240\u63cf\u8ff0\u4e4b\u539f\u59cb LSSF \u6cd5\u8207 MSI \u6cd5\u7684\u53ef\u80fd\u7f3a\u9ede\uff0c\u5728\u9019\uf9e8\uff0c\u6211\u5011\u540c\u6a23\u5730\u6e2c\u8a66\u4fee \u6b63\u5f0f LSSF \u6cd5\u8207 MSI \u6cd5\u7d50\u5408 MVA \u6cd5\u7684\u6548\u679c\uff0c\u5373\u5728\u539f\u59cb LSSF \u6cd5\u6216 MSI \u6cd5\u4e4b\u88dc\uf9b2\u7684\u7a0b\u5e8f \u524d\u5148\u5c07\u539f\u59cb N \u9ede\u4e4b MVA \u6cd5\u8655\uf9e4\u5f8c\u4e4b MFCC \u7279\u5fb5\u5e8f\uf99c\u4e58\u4e0a\u4e00\u6f22\uf95f\u7a97(Hanning window)\uff0c \u89c0\u5bdf\u9019\u6a23\u7684\u64cd\u4f5c\u80fd\u5426\u5e36\uf92d\u9032\u6b65\u3002 \u5716\u5341\u4e8c\u70ba\u539f\u59cb\u8207\u4fee\u6b63\u5f0f LSSF \u8207 MSI \u4f5c\u7528\u65bc MVA \u6cd5\u8655\uf9e4\u5f8c MFCC \u7279\u5fb5\u4e4b\u5e73\u5747\u8fa8\uf9fc\uf961\u9577 \u689d\u5716\u3002\u7531\u6b64\u5716\u53ef\u4ee5\u770b\u51fa\uff0c\u5728\u7d50\u5408 MVA \u6cd5\u7684\u524d\u63d0\u4e0b\uff0c\u4fee\u6b63\u5f0f LSSF \u6cd5\u76f8\u8f03\u65bc\u539f\u59cb LSSF \u6cd5\u800c\u8a00\uff0c\u6709 0.65%\u5e73\u5747\u8fa8\uf9fc\uf961\u7684\u63d0\u5347\uff0c\u800c\u4fee\u6b63\u5f0f MSI \u6cd5\u76f8\u5c0d\u65bc\u539f\u59cb MSI \u6cd5\u800c\u8a00\uff0c\u6709 0.48% \u5e73\u5747\u8fa8\uf9fc\uf961\u7684\u63d0\u5347\uff0c\u56e0\u6b64\uff0c\u6211\u5011\u9a57\u8b49\uf9ba\uf978\u7a2e\u4fee\u6b63\u5f0f\u65b9\u6cd5\u90fd\u80fd\u4f7f\u539f\u59cb\u65b9\u6cd5\u9032\u4e00\u6b65\u63d0\u5347\u6548\u80fd\u3002 \u5716\u5341\u4e8c\u3001\u539f\u59cb\u548c\u4fee\u6b63\u5f0f LSSF \u8207 MSI \u6cd5\u4f5c\u7528\u65bc MVA \u6cd5\u8655\uf9e4\u5f8c MFCC \u7279\u5fb5\u4e4b\u5e73\u5747\u8fa8\uf9fc\uf961 \u4e94\u3001\u7d50\uf941 \u5728\u4f5c\u7528\u65bc\u539f\u59cb MFCC \u7279\u5fb5\u6642\uff0c\u6211\u5011\u767c\u73fe\uff0c\u539f\u59cb TSN \u6cd5(TSN-1)\u7684\u76f4\uf9ca\u589e\u76ca\u6b63\u898f\u5316\u6b65\u9a5f\u662f \u9020\u6210\u5176\u6548\u679c\uf967\u5f70\u7684\u539f\u56e0\u4e4b\u4e00\uff0c\u632a\u53bb\u6b64\u6b65\u9a5f\u6240\u5f97\u4e4b TSN-2 \u6cd5\u5373\u53ef\u6709\u5341\u5206\u986f\u8457\u7684\u8868\u73fe\uff0c\u800c \u6211\u5011\u63d0\u51fa\u7684\u4e09\u7a2e\u65b0\u65b9\u6cd5\uff0c\u76f8\u8f03\u65bc TSN-1 \u8207 TSN-2\uff0c\u90fd\u80fd\u6709\uf901\u4f73\u7684\u6548\u679c\uff0c\u800c\u5176\u4e2d\u53c8\u4ee5 ERTF \u6cd5\u4e4b\u8868\u73fe\u6700\u597d\uff0c\u7531\u65bc ERTF \u8207 TSN-2 \u53ea\u6709\u5728\u8a2d\u8a08\u6642\u9593\u5e8f\uf99c\uf984\u6ce2\u5668\u7684\u7a0b\u5e8f\u4e0a\u6709\u5dee\u5225\uff0c\u9019\u8868 \u793a\u6211\u5011 ERTF \u8a2d\u8a08\u51fa\uf92d\u7684\uf984\u6ce2\u5668\uff0c\u6bd4\u8d77 TSN-2 \u6cd5\u7684\uf984\u6ce2\u5668\uf901\u7cbe\u78ba\u5730\u5c0d\u7279\u5fb5\u4e4b\u8abf\u8b8a\u983b\u8b5c\u4f5c \u6b63\u898f\u5316\u3002\u800c\u7576\u6211\u5011\u5c07\u9019\u4e9b\u65b9\u6cd5\u4f5c\u7528\u65bc CMVN \u6cd5\u6216 MVA \u6cd5\u8655\uf9e4\u5f8c\u7684 MFCC \u7279\u5fb5\u6642\uff0c\u767c \u73fe\u5b83\u5011\u76f8\u8f03\u65bc\u55ae\u4e00 CMVN \u6cd5\u6216 MVA \u6cd5\u800c\u8a00\uff0c\u80fd\u5e36\uf92d\uf901\u4f73\u7684\u8fa8\uf9fc\uf961\uff0c\u4e14\u6211\u5011\u6240\u63d0\u51fa\u4e4b \u4e09\u7a2e\u65b0\u65b9\u6cd5\u7684\u8868\u73fe\u5e7e\u4e4e\u4ecd\u7136\u512a\u65bc TSN-1 \u6cd5\u8207 TSN-2 \u6cd5\u3002\u6b64\u5916\uff0c\u6211\u5011\u63a2\u8a0e LSSF \u6cd5\u8207 MSI \u6cd5\u53ef\u80fd\u5b58\u5728\u4e4b\u983b\u8b5c\u907a\uf94e(leakage)\u7684\u7f3a\u9ede\uff0c\u800c\u63d0\u51fa\u76f8\u5c0d\u61c9\u7684\u4fee\u6b63\u65b9\u6cd5\uff0c\u767c\u73fe\u9019\u4e9b\u4fee\u6b63\u6cd5\u80fd \uf901\u9032\u4e00\u6b65\u6539\u5584\u539f\u59cb LSSF \u6cd5\u8207 MSI \u6cd5\u7684\u6548\u80fd\u3002 \uf974\u5c31\u4e09\u7a2e\u65b0\u65b9\u6cd5\u5f7c\u6b64\u4f5c\u6bd4\u8f03\uff0cERTF \u6cd5\u8207 LSSF \u6cd5\u904b\u7b97\u8907\u96dc\ufa01\u8f03\u5927\uff0cMSI \u6cd5\u5247\u76f8\u5c0d\u8f03\u5c0f\uff0c \u96d6\u7136 ERTF \u6cd5\u5c0d\u539f\u59cb MFCC \u7279\u5fb5\u800c\u8a00\uff0c\u8868\u73fe\u6bd4 LSSF \u6cd5\u8207 MSI \u6cd5\uf92d\u5f97\u597d\uff0c\u4f46\u7576\u5b83\u5011\u4f5c \u7528\u65bc CMVN \u6cd5\u6216 MVA \u6cd5\u8655\uf9e4\u904e\u5f8c\u7684 MFCC \u7279\u5fb5\u6642\uff0c\u5176\u6548\u80fd\u7684\u5dee\uf962\u6027\u5df2\u7d93\u5f88\u5c0f\uff0c\u9019\u610f \u5473\u8457\u904b\u7b97\u8907\u96dc\ufa01\u8f03\u5c0f\u7684 MSI \u6cd5\u76f8\u5c0d\u65bc ERTF \u6cd5\u8207 LSSF \u6cd5\u800c\u8a00\uff0c\u53ef\u80fd\u6709\uf901\u4f73\u7684\u61c9\u7528\u6027\u3002 \uf9d1\u3001\uf96b\u8003\u6587\u737b",
                "num": null,
                "content": "<table><tr><td colspan=\"8\">(\u56db)\u8abf\u8b8a\u983b\u8b5c\u6b63\u898f\u5316\u4e4b\u65b0\u65b9\u6cd5\u7684\u6548\u679c\u8a0e\uf941 \u9019\u5c0f\u7bc0\u5c07\u7c21\u55ae\u5c55\u793a\u672c\u7ae0\u7bc0\u6240\u63d0\u51fa\u7684\u4e09\u7a2e\u65b0\u65b9\u6cd5\u5c0d\u539f\u59cb MFCC \u7279\u5fb5\u5e8f\uf99c\u4e4b\u8abf\u8b8a\u983b\u8b5c \u6b63\u898f\u5316\u7684\u6548\u679c\uff0c\u5716\u4e03\u3001\u5716\u516b\u8207\u5716\u4e5d\u5206\u5225\u70ba\u539f\u59cb\u7b2c\u4e00\u7dad\u6885\u723e\u5012\u983b\u8b5c\u4fc2\uf969(c 1 )\u5e8f\uf99c\u5206\u5225\u7d93 ERTF \u6cd5\u3001LSSF \u6cd5\u8207 MSI \u6cd5\u8655\uf9e4\u5f8c\u7684\u529f\uf961\u983b\u8b5c\u5bc6\ufa01\u66f2\u7dda\u5716\u3002\u8207\u524d\u4e00\u7ae0\u7684\u5716\u4e09\u3001\u5716\u56db\u548c \u5716\u4e94\u76f8\u540c\uff0c\u9019\uf9e8\u6211\u5011\u6240\u4f7f\u7528\u7684\u662f AURORA 2 \u8cc7\uf9be\u5eab[9]\uf9e8\u7684 MAH_4625A \u8a9e\u97f3\u6a94\uff0c\u7136\u5f8c \u52a0\u5165\uf967\u540c\u8a0a\u96dc\u6bd4(SNR)\u7684\u5730\u4e0b\u9435(subway)\u96dc\u8a0a\u3002 \u96dc\u8a0a\u6bd4\u4e0b\u7684\u8a9e\u97f3\u7279\u5fb5\u4e4b\u8abf\u8b8a\u983b\u8b5c\u3002 \u25cb 2 LSSF \u6cd5\u548c MSI \u6cd5\u90fd\u662f\u76f4\u63a5\u5728\u7279\u5fb5\u7684\u8abf\u8b8a\u983b\u8b5c\u57df(modulation spectral domain)\u4e0a \u6b63\u898f\u5316\u5176\u5f37\ufa01\u6210\u4efd\uff0c\u5f9e\u5716\u516b\u548c\u5716\u4e5d\u53ef\u770b\u51fa\u9019\uf978\u7a2e\u65b9\u6cd5\u8207 ERTF \u6cd5\uf9d0\u4f3c\uff0c\u80fd\u5c07\u53d7\u96dc\u8a0a\u5e72\u64fe \u7684\u8a9e\u97f3\u4e4b\u529f\uf961\u983b\u8b5c\u5bc6\ufa01\u66f2\u7dda\uff0c\u903c\u8fd1\uf96b\u8003\u7684\u529f\uf961\u983b\u8b5c\u5bc6\ufa01\u66f2\u7dda\uff0c\u4f7f\u9019\u4e9b\u66f2\u7dda\u4e4b\u9593\u7684\u5dee\uf962\u660e \u986f\u8f03\u4f4e\uff0c\u4ee3\u8868\uf9ba\u9019\uf978\u500b\u8abf\u8b8a\u983b\u8b5c\u5f37\ufa01\u6b63\u898f\u5316\u6cd5\u4e5f\u80fd\u6709\u6548\u5730\u5f37\u5065\u8a9e\u97f3\u7279\u5fb5\u3002\u5176\u4e2d\uff0cMSI \u5728\u8abf\u8b8a\u983b\u8b5c\u4e0a\u5931\u771f\u7684\u73fe\u8c61\uff0c\u6211\u5011\u5728\u4e0b\u500b\u7ae0\u7bc0\uff0c\u5c07\u6703\u4ee5\u8fa8\uf9fc\u5be6\u9a57\uf969\u64da\u8b49\u5be6\u9019\u4e9b\u65b9\u6cd5\u7684\u6548\u80fd\u3002 LSSF 84.37 86.21 84.72 85.10 11.90 44.40 83 \u5f9e\u4e0a\u8ff0\u4e09\u500b\u5716\u4e2d\uff0c\u53ef\u770b\u51fa\u6211\u5011\u6240\u63d0\u7684\u4e09\u500b\u65b0\u65b9\u6cd5\u90fd\u80fd\u6709\u6548\u5730\ufa09\u4f4e\u96dc\u8a0a\u6240\u9020\u6210\u4e4b\u8a9e\u97f3\u7279\u5fb5 ERTF 85.45 86.92 85.34 85.90 12.70 47.39 \u6cd5\u662f\u4e09\u500b\u65b9\u6cd5\u4e2d\u8a08\u7b97\u8907\u96dc\ufa01\u6700\u4f4e\u7684\u6280\u8853\uff0c\u56e0\u6b64\u6709\uf901\u5927\u7684\u61c9\u7528\u50f9\u503c\u3002 \u8868\u4e00\u3001\u5404\u7a2e\u7279\u5fb5\u5e8f\uf99c\u8655\uf9e4\u6280\u8853\u4e4b\u8fa8\uf9fc\uf961(%) Method Set A Set B Set C average AR 87 RR Baseline 72.46 68.31 78.82 73.20 --TSN-1 73.61 70.44 77.19 73.75 0.55 2.05 TSN-2 80.29 82.36 75.82 79.49 6.29 23.47 84 85 85.77 \u8fa8 86 85.34 \u8b58 85.10 \u7387 84.42 (%)</td></tr><tr><td>MSI</td><td>82</td><td>83.61</td><td>85.36</td><td>84.28</td><td>84.42</td><td>11.22</td><td>41.87</td></tr><tr><td colspan=\"8\">\u56db\u3001\u8abf\u8b8a\u983b\u8b5c\u6b63\u898f\u5316\u6cd5\u8207\u5404\u7a2e\u7279\u5fb5\u6642\u9593\u5e8f\uf99c\u6b63\u898f\u5316\u6280\u8853\u6cd5\u4e4b\u8fa8\uf9fc\u5be6\u9a57\u7d50\u679c \u8207\u8a0e\uf941 CMVN 85.03 85.56 85.60 85.40 12.20 45.52 CMVN+ARMA(MVA) 88.12 88.81 88.50 88.48 15.28 57.01 Method Set A Set B Set C average AR RR 81 MVA 88.12 88.81 88.50 88.48 \u2500 \u2500</td></tr><tr><td colspan=\"8\">\u9ede\u7684\u53cd\uf9ea\u6563\u5085\uf9f7\uf96e\u8f49\u63db(inverse discrete Fourier transform, IDFT)\uff0c\u4ee5\u6c42\u5f97\u65b0\u7684\u7279\u5fb5\u5e8f\uf99c [ ] { } y n \uff0c\u5982\u4e0b\u5f0f\u6240\u793a\uff1a [ ] ( ) 2 1 0 1 , 0 1 nk N j N k k y n Y e n N N \u03c0 \u03c9 \u2032 \u2212 \u2032 \u2032= = \u2264 \u2264 \u2212 \u2211 . \u5f0f(3.10) \u4ee5\u4e0a\u7684\u65b9\u6cd5\uff0c\u5373\u7a31\u70ba\u5f37\ufa01\u983b\u8b5c\u5167\u63d2\u6cd5(magnitude spectrum interpolation, MSI)\u3002 \u5716\u4e03\u3001\uf967\u540c\u8a0a\u96dc\u6bd4\u4e4b\u4e0b\uff0c\u539f\u59cb c 1 \u5e8f\uf99c\u7d93 ERTF \u6cd5\u8655\uf9e4\u5f8c\u4e4b\u529f\uf961\u983b\u8b5c\u5bc6\ufa01\u66f2\u7dda\u5716 \u5716\u516b\u3001\uf967\u540c\u8a0a\u96dc\u6bd4\u4e4b\u4e0b\uff0c\u539f\u59cb c 1 \u5e8f\uf99c\u7d93 LSSF \u6cd5\u8655\uf9e4\u5f8c\u4e4b\u529f\uf961\u983b\u8b5c\u5bc6\ufa01\u66f2\u7dda\u5716 \u5716\u4e5d\u3001\uf967\u540c\u8a0a\u96dc\u6bd4\u4e4b\u4e0b\uff0c\u539f\u59cb c 1 \u5e8f\uf99c\u7d93 MSI \u6cd5\u8655\uf9e4\u5f8c\u4e4b\u529f\uf961\u983b\u8b5c\u5bc6\ufa01\u66f2\u7dda\u5716 \u5c07\u5716\u4e03\u3001\u5716\u516b\u3001\u8207\u5716\u4e5d\u914d\u5408\u524d\u4e00\u7ae0\u4e4b\u5716\u4e09\u3001\u5716\u56db\u8207\u5716\u4e94\u76f8\u6bd4\u8f03\uff0c\u6211\u5011\u6709\u4ee5\u4e0b\uf978\u9ede\u8a0e\uf941\uff1a \u25cb 1 \u7531\u65bc ERTF \u6cd5\u548c TSN \u6cd5\u540c\u6a23\u662f\u8a2d\u8a08\u4e00\u6642\u9593\u5e8f\uf99c\uf984\u6ce2\u5668\uff0c\u4f5c\u7528\u65bc\u7279\u5fb5\uf96b\uf969\u5e8f\uf99c \u4e0a\uff0c\u6211\u5011\u5148\u6bd4\u8f03\u9019\uf978\u7a2e\u65b9\u6cd5\u7684\u6548\u80fd\u3002\u5f9e\u5716\u4e03\u4e2d\u53ef\u770b\u51fa ERTF \u6cd5\u80fd\u540c\u6642\u4f7f\u5f97\u4e7e\u6de8\u8a9e\u97f3\u8207\u53d7 \u96dc\u8a0a\u5e72\u64fe\u7684\u8a9e\u97f3\u7684\u529f\uf961\u983b\u8b5c\u5bc6\ufa01\u66f2\u7dda\uff0c\u903c\u8fd1\uf96b\u8003\u7684\u529f\uf961\u983b\u8b5c\u5bc6\ufa01\u66f2\u7dda\uff0c\u6709\u6548\ufa09\u4f4e\u5716\u4e09\u6240 \u986f\u793a\u4e4b\uf967\u540c\u8a0a\u96dc\u6bd4\u4e0b\u7279\u5fb5\u5e8f\uf99c\u4e4b\u529f\uf961\u983b\u8b5c\u5bc6\ufa01\u7684\u5931\u771f\uff0c\u76f8\u8f03\u65bc\u5716\u56db\u6240\u986f\u793a\u4e4b\u539f\u59cb TSN \u6cd5\u7684\u6548\u679c\u6709\u660e\u986f\u6539\u5584\uff0c\u4e14\u8207\u5716\u4e94\u4e4b TSN-2 \u6cd5\u7684\u6548\u679c\u5341\u5206\u63a5\u8fd1\uff0c\u6b64\u4ee3\u8868\u6211\u5011\u4f7f\u7528\u7b49\uf992\u6ce2 \uf984\u6ce2\u5668\u8a2d\u8a08\u6cd5 (equi-ripple filter design)\uf92d\u8a2d\u8a08\u6642\u9593\u5e8f\uf99c\uf984\u6ce2\u5668\uff0c\u53ef\u4ee5\u6709\u6548\u5730\u6b63\u898f\u5316\uf967\u540c Hz Hz Hz \u672c\u7ae0\u7bc0\u4e3b\u8981\u662f\u5c07\u6211\u5011\u63d0\u51fa\u7684\u4e09\u7a2e\u8abf\u8b8a\u983b\u8b5c\u6b63\u898f\u5316\u6cd5\uff1a\u7b49\uf992\u6ce2\u6642\u9593\u5e8f\uf99c\uf984\u6ce2\u5668 (equi-ripple temporal filtering, ERTF)\u3001\u6700\u5c0f\u5e73\u65b9\u983b\u8b5c\u64ec\u5408\u6cd5(least-squares spectrum fitting, LSSF)\u548c\u5f37\ufa01\u983b\u8b5c\u5167\u63d2\u6cd5(magnitude spectrum interpolation, MSI)\u904b\u7528\u65bc\u96dc\u8a0a\u74b0\u5883\u4e0b\u7684\u8a9e \u97f3\u8fa8\uf9fc\uff0c\u85c9\u6b64\u89c0\u5bdf\u5206\u6790\u5176\u7d50\u679c\uff0c\u540c\u6642\u6211\u5011\u4e5f\u6703\u5c07\u5b83\u5011\u8207\u5176\u4ed6\u7279\u5fb5\u6642\u9593\u5e8f\uf99c\u6b63\u898f\u5316\u6cd5\u7684\u6548 \u679c\u4f5c\u6bd4\u8f03\u3002\u6700\u5f8c\uff0c\u6211\u5011\u5617\u8a66\u5c07\u9019\u4e9b\u65b0\u65b9\u6cd5\u8207\u5176\u4ed6\u65b9\u6cd5\u4e92\u76f8\u7d50\u5408\uff0c\uf92d\u89c0\u5bdf\u9019\u6a23\u7684\u7d50\u5408\u662f\u5426 \u80fd\uf92d\uf901\u9032\u4e00\u6b65\u7684\u6548\u80fd\u63d0\u5347\u3002 (\u4e00)\u5be6\u9a57\u74b0\u5883\u8207\u5be6\u9a57\u67b6\u69cb\u8a2d\u5b9a \u672c\uf941\u6587\u4e2d\u6240\u63a1\u7528\u7684\u8a9e\u97f3\u8cc7\uf9be\u5eab\u70ba\u6b50\u6d32\u96fb\u4fe1\u6a19\u6e96\u5354\u6703(European Telecommunication Standard Institute, ETSI)\u6240\u767c\ufa08\u4e4b\u8a9e\uf9be\u5eab\uff1aAURORA 2.0[9]\uff0c\u5167\u5bb9\u662f\u4ee5\u7f8e\u570b\u6210\uf98e\u7537\uf981\u6240\uf93f \u88fd\u7684\u4e00\u7cfb\uf99c\uf99a\u7e8c\u7684\u82f1\u6587\uf969\u5b57\u5b57\uf905\uff0c\u8a9e\u97f3\u672c\u8eab\u4e26\u52a0\u4e0a\u5404\u7a2e\u52a0\u6210\u6027\u96dc\u8a0a\u8207\u901a\u9053\u6548\u61c9\u7684\u5e72\u64fe\u3002 \u52a0\u6210\u6027\u96dc\u8a0a\u5171\u6709\u516b\u7a2e\uff0c\u5206\u5225\u70ba\u5730\u4e0b\u9435\u3001\u4eba\u8072\u3001\u6c7d\uf902\uff0c\u5c55\u89bd\u6703\u9928\u3001\u9910\u5ef3\u3001\u8857\u9053\u3001\u98db\u6a5f\u5834\u548c \u706b\uf902\u7ad9\u96dc\u8a0a\u7b49\uff0c\u901a\u9053\u6548\u61c9\u5247\u6709\uf978\u7a2e\uff0c\u5206\u5225\u70ba G712 \u8207 MIRS[11]\u3002\u96dc\u8a0a\u542b\uf97e\u7684\u5927\u5c0f\u5305\u542b\uf9ba \u4e7e\u6de8\u7121\u96dc\u8a0a\u7684\uf9fa\u614b\uff0c\u4ee5\u53ca\uf9d1\u7a2e\uf967\u540c\u8a0a\u96dc\u6bd4(signal-to-noise ratio, SNR)\uf9fa\u614b\uff0c\u5206\u5225\u662f 20dB\u3001 15dB\u300110dB\u30015dB\u30010dB \u8207-5dB\uff0c\u56e0\u6b64\u6211\u5011\u53ef\u4ee5\u89c0\u5bdf\uf967\u540c\u7684\u96dc\u8a0a\u74b0\u5883\u5c0d\u65bc\u8a9e\u97f3\u8fa8\uf9fc\u7684\u5f71 \u97ff\u3002\u56e0\u96dc\u8a0a\u7279\u6027\u7684\uf967\u540c\uff0c\u6e2c\u8a66\u74b0\u5883\u53ef\u5206\u70ba Set A\u3001Set B \u8207 Set C \u4e09\u7d44[9]\u3002 \u8072\u5b78\u6a21\u578b\u662f\u57f7\ufa08\u96b1\u85cf\u5f0f\u99ac\u53ef\u592b\u6a21\u578b\u5de5\u5177(hidden Markov model tool kit, HTK)[12]\u8a13 \uf996\u6240\u5f97\uff0c\u5305\u542b 11 \u500b\uf969\u5b57\u6a21\u578b(zero, one, two, \u2026, nine \u53ca oh)\u4ee5\u53ca\u975c\u97f3(silence)\u6a21\u578b\uff0c\u6bcf\u500b \uf969\u5b57\u6a21\u578b\u5305\u542b 16 \u500b\uf9fa\u614b\uff0c\u5404\uf9fa\u614b\u5305\u542b 20 \u500b\u9ad8\u65af\u5bc6\ufa01\u6df7\u5408\u3002 (\u4e8c)\u8abf\u8b8a\u983b\u8b5c\u6b63\u898f\u5316\u6cd5\u4f5c\u7528\u65bc\u6885\u723e\u5012\u983b\u8b5c\u7279\u5fb5\uf96b\uf969\u4e4b\u5be6\u9a57\u7d50\u679c \u672c \u7ae0 \u7bc0 \u6240 \u6709 \u5be6 \u9a57 \u6240 \u4f7f \u7528 \u7684 \u8a9e \u97f3 \u7279 \u5fb5 \u70ba \u6885 \u723e \u5012 \u983b \u8b5c \u4fc2 \uf969 (mel-frequency cepstral coefficients, MFCC) \uff0c\u6211\u5011\u63a1\u7528\u7684 MFCC \u7279\u5fb5\uf96b\uf969\u70ba13\u7dad(c0~c12) \uff0c\u52a0\u4e0a\u5176\u4e00\u968e\u5dee\uf97e(delta) \u548c\u4e8c\u968e\u5dee\uf97e(delta-delta)\uff0c\u7e3d\u5171\u70ba39\u7dad\u7279\u5fb5\uf96b\uf969\u3002\u57fa\u672c\u5be6\u9a57(baseline experiment)\u662f\u4ee5\u539f\u59cb MFCC \u7279\u5fb5\uf96b\uf969\u4f5c\u70ba\u8a13\uf996\u8207\u6e2c\u8a66\uff0cTSN-\u56fa\u5b9a\u70ba1024\u3002\u4e0b\u8868\u4e00\u4e2d\uff0c\u6211\u5011\u7d9c\u5408\uf9ba TSN-1\u3001TSN-2\u3001ERTF\u3001LSSF\u3001MSI\uff0c\u53ca\u8457\u540d\u7684 \u7279\u5fb5\u6b63\u898f\u5316\u6280\u8853 CMVN[3]\u548c MVA[5]\uff0c\u5176\u5404\u5225\u4f5c\u7528\u65bc\u539f\u59cb MFCC \u7279\u5fb5\uf96b\uf969\u6240\u5f97\u7684\u5e73\u5747 \u8fa8\uf9fc\uf961(20dB\u300115dB\u300110dB\u30015dB \u82070dB \u4e94\u7a2e\u8a0a\u96dc\u6bd4\u4e0b\u7684\u8fa8\uf9fc\uf961\u5e73\u5747) \uff0c\u5176\u4e2d AR \u8207 RR \u5206\u5225\u70ba\u76f8\u8f03\u65bc\u57fa\u672c\u5be6\u9a57\u7d50\u679c\u4e4b\u7d55\u5c0d\u932f\u8aa4\ufa09\u4f4e\uf961(absolute error rate reduction)\u548c\u76f8\u5c0d\u932f \u8aa4\ufa09\u4f4e\uf961(relative error rate reduction)\u3002 \u7531\u8868\u4e00\u7684\uf969\u64da\uff0c\u6211\u5011\u53ef\u770b\u51fa\u4ee5\u4e0b\u5e7e\u9ede\u73fe\u8c61\uff1a \u25cb 1 \u539f\u59cb TSN \u6cd5(TSN-1)\u5c0d MFCC \u7279\u5fb5\u5728\u96dc\u8a0a\u74b0\u5883\u4e0b\u7684\u8fa8\uf9fc\uf961\u7684\u6539\u9032\u4e26\uf967\u662f\u5f88\u660e\u986f\uff0c\u53ea\u9032 \u6b650.55%\uff0c\u7136\u800c TSN-2\u6cd5\u5e36\uf92d\u5341\u5206\u660e\u986f\u7684\u8fa8\uf9fc\uf961\u63d0\u5347(Set C \u9664\u5916)\uff0c\u5728 Set A \u548c Set B \u74b0\u5883 \u4e0b\uff0c\u5e73\u5747\u8fa8\uf9fc\uf961\u76f8\u5c0d\u65bc TSN-1\u800c\u8a00\u5206\u5225\u6539\u9032\uf9ba8%\u820714%\u5de6\u53f3\u3002\u5982\u6b64\u770b\u51fa\uff0c\u85c9\u7531\uf96d \uf976\u76f4\uf9ca\u589e\u76ca\u6b63\u898f\u5316\u7684\u6b65\u9a5f\uff0cTSN-2 \u6bd4 TSN-1 \u5177\u6709\uf901\u4f73\u7684\u7279\u5fb5\u8abf\u8b8a\u983b\u8b5c\u6b63\u898f\u5316\u7684\u6548 \u679c\uff0c\u9019\u4e5f\u547c\u61c9\uf9ba\u5728\u7b2c\u4e8c\u7ae0\u7684\u5716\u4e09\uff0c\u539f\u59cb TSN \u6cd5\u7121\u6cd5\u6709\u6548\ufa09\u4f4e\u5404\u96dc\u8a0a\u74b0\u5883\u4e0b\uff0c\u539f\u59cb\u8a9e\u97f3 MFCC \u7279\u5fb5\u6642\u9593\u5e8f\uf99c\u4e4b\u529f\uf961\u983b\u8b5c\u5bc6\ufa01\u66f2\u7dda\u7684\uf967\u5339\u914d\u73fe\u8c61\u3002 \u25cb 2 ERTF\u3001LSSF \u8207 MSI \u6cd5\u4e09\u7a2e\u65b0\u65b9\u6cd5\u5728\u5404\u7a2e\uf967\u540c\u7684\u96dc\u8a0a\u74b0\u5883\u4e0b\u7686\u80fd\u660e\u986f\u63d0\u5347\u8fa8\uf9fc\uf961\uff0c \u5c0d Set A 2P N &gt; \uff0c\u6211\u5011\u6703\u4ee5\u88dc\uf9b2\u7684\u65b9\u5f0f\u5148\u5c07\u539f\u59cb\u7684 N \u9ede\u7684\u7279\u5fb5\u5e8f\uf99c\u8b8a\u9577\u70ba 2P \u9ede\uff0c\u610f\u5373\u591a\u88dc\uf9ba 2P N \u2212 \u500b\uf9b2\u9ede\uff0c\u9019\u6a23\u7684\u4f5c\u6cd5\u5bb9\uf9e0\u7522\u751f\u975e\uf9b2\u503c\u7684 \u9ede\u8207\uf9b2\u503c\u7684\u9ede\u4e4b\u9593\u8a0a\u865f\u503c\uf967\uf99a\u7e8c\u7684\u60c5\u5f62\uff0c\u800c\u5f15\u9032\uf9ba\uf967\u5fc5\u8981\u7684\u9ad8\u983b\u6210\u4efd\uff0c\u9019\u6548\u61c9\uf9d0\u4f3c\u65bc\u76f4 \u63a5\u65bc\u4e00\u8a0a\u865f\u52a0\u4e0a\u77e9\u5f62\u7a97\u6240\u9020\u6210\u983b\u8b5c\u907a\uf94e(leakage)[10]\u7684\u7f3a\u9ede\uff0c\u56e0\u6b64\uff0c\u6211\u5011\u9019\uf9e8\u5728 LSSF \u8207 MSI \u6cd5\u4e4b\u88dc\uf9b2\u7684\u7a0b\u5e8f\u524d\uff0c\u5148\u5c07\u539f\u59cb\u7684 N \u9ede\u7684\u7279\u5fb5\u5e8f\uf99c\u4e58\u4e0a\u4e00\u6f22\uf95f\u7a97(Hanning window)[10]\uff0c\uf92d\ufa09\u4f4e\u4e0a\u8ff0\u53ef\u80fd\u7684\uf967\uf97c\u6548\u61c9\uff0c\u89c0\u5bdf\u9019\u6a23\u7684\u64cd\u4f5c\u662f\u5426\u53ef\u9032\u4e00\u6b65\u63d0\u5347 LSSF \u6cd5 LSSF \u4fee\u6b63\u5f0fLSSF \u4fee\u6b63\u5f0fMSI MSI \u5716\u5341\u3001\u539f\u59cb\u548c\u4fee\u6b63\u5f0f LSSF \u8207 MSI \u4f5c\u7528\u65bc\u539f\u59cb MFCC \u7279\u5fb5\u4e4b\u5e73\u5747\u8fa8\uf9fc\uf961 (\u4e09)\u8abf\u8b8a\u983b\u8b5c\u6b63\u898f\u5316\u6cd5\u7d50\u5408\u5012\u983b\u8b5c\u5e73\u5747\u8207\u8b8a\uf962\uf969\u6b63\u898f\u5316\u6cd5\u4e4b\u5be6\u9a57\u7d50\u679c \u524d\u9762\u63d0\u5230\uff0c\u5012\u983b\u8b5c\u5e73\u5747\u8207\u8b8a\uf962\uf969\u6b63\u898f\u5316\u6cd5(cepstral mean and variance normalization, MVA+TSN-1 89.58 90.19 89.74 89.84 1.36 11.80 MVA+TSN-2 89.81 90.34 89.84 99.00 1.52 13.19 MVA+ERTF 89.75 90.81 89.64 90.07 1.59 13.80 MVA+LSSF 89.63 90.87 89.94 90.14 1.67 14.49 MVA+MSI 89.71 90.91 89.94 90.19 1.71 14.84 \u8207 80 CMVN)[3]\u5c0d\u96dc\u8a0a\u74b0\u5883\u4e0b\u7684\u8a9e\u97f3\u8fa8\uf9fc\uf961\u6709\u660e\u986f\u7684\u6539\u9032\uff0c\u56e0\u6b64\u9019\uf9e8\u6211\u5011\u5617\u8a66\u5c07\u5404\u7a2e\u8abf\u8b8a\u983b Method Set A Set B Set C average AR RR CMVN 85.03 85.56 85.60 85.40 \u2500 \u2500 CMVN+TSN-1 89.42 90.03 89.03 89.49 4.10 28.05 CMVN+TSN-2 89.59 90.36 89.34 89.76 4.36 29.90 CMVN+ERTF 89.61 90.67 89.28 89.85 4.45 30.52 CMVN+LSSF 89.12 90.17 89.16 89.48 4.09 27.98 CMVN+MSI 89.59 90.56 89.60 89.92 4.52 30.95 CMVN+ARMA(MVA) 88.12 88.81 88.50 88.48 3.08 21.09 \u7531\u8868\u4e8c\u7684\uf969\u64da\uff0c\u6211\u5011\u53ef\u770b\u51fa\u4ee5\u4e0b\u5e7e\u9ede\u73fe\u8c61\uff1a \u25cb 1 TSN-1 \u7684\u8fa8\uf9fc\uf961\u6539\u5584\uff0c\u6b64\u7d50\u679c\u5341\u5206\u543b\u5408\u5728 TSN \u6cd5\u7684\u539f\u59cb\u6587\u737b[8]\uf9e8\u4e4b\u7d50\u679c\uff0c\u76f8\u8f03\u65bc\u8868\u4e00\u6240\u5448\u73fe \u4e4b TSN-1 \u4e26\u672a\u660e\u986f\u6539\u5584\u53d7\u96dc\u8a0a\u5f71\u97ff\u4e4b\u539f\u59cb MFCC \u7279\u5fb5\u7684\u73fe\u8c61\uff0c\u5728\u9019\uf9e8\uff0cTSN-1 \u6cd5\u80fd\u6709 \u660e\u986f\u6539\u9032\u4e4b\u6548\u80fd\u7684\u539f\u56e0\u53ef\u80fd\u5728\u65bc\uff0cCMVN \u6cd5\u5df2\u4e8b\u5148\u6709\u6548\u5730\ufa09\u4f4e\u539f\u59cb MFCC \u7279\u5fb5\u53d7\u96dc\u8a0a \u5f71\u97ff\u6240\u9020\u6210\u4e4b\u8abf\u8b8a\u983b\u8b5c\u4e0a\u4e0b\u504f\u79fb\u7684\u5931\u771f\uff0c\u56e0\u6b64 TSN-1 \u80fd\u55ae\u7d14\u8655\uf9e4\u8abf\u8b8a\u983b\u8b5c\u6b63\u898f\u5316\u7684\u90e8 \u4efd\uff0c\u800c\u5e36\uf92d\u8fa8\uf9fc\uf961\u7684\u6539\u5584\u3002\u53e6\u5916\uff0c\u6211\u5011\u4e5f\u767c\u73fe\u5230\uff0cTSN-1 \u548c TSN-2 \u6240\u5f97\u7d50\u679c\u4e4b\u9593\u7684\u5dee \u8ddd\u8b8a\u5f97\u8f03\u5c0f\uff0c\u4f46 TSN-2 85 86 87 88 89 90 91 92 \u8fa8\u8b58 \u7387 (%) 89.48 90.33 89.92 90.39 LSSF +CMVN \u4fee\u6b63\u5f0fLSSF +CMVN MSI +CMVN \u4fee\u6b63\u5f0fMSI +CMVN \u5716\u5341\u4e00\u3001\u539f\u59cb\u548c\u4fee\u6b63\u5f0f LSSF \u8207 MSI \u4f5c\u7528\u65bc CMVN \u6cd5\u8655\uf9e4\u5f8c MFCC \u7279\u5fb5\u4e4b\u5e73\u5747\u8fa8\uf9fc\uf961 (\u56db)\u8abf\u8b8a\u983b\u8b5c\u6b63\u898f\u5316\u6cd5\u7d50\u5408\u5012\u983b\u8b5c\u5e73\u5747\u8207\u8b8a\uf962\uf969\u6b63\u898f\u5316\u7d50\u5408\u81ea\u52d5\u56de\u6b78\u52d5 \u614b\u5e73\u5747\uf984\u6ce2\u5668\u6cd5\u4e4b\u5be6\u9a57\u7d50\u679c \u524d\u9762\u63d0\u5230\uff0c\u5012\u983b\u8b5c\u5e73\u5747\u8207\u8b8a\uf962\uf969\u6b63\u898f\u5316\u7d50\u5408\u81ea\u52d5\u56de\u6b78\u52d5\u614b\u5e73\u5747\uf984\u6ce2\u5668\u6cd5(MVA)[5] 86 87 88 89 90 91 92 \u8fa8\u8b58 \u7387(%) 90.14 90.79 90.19 90.67 \u7531\u8868\u4e09\u53ef\u770b\u51fa\uff0cTSN-1 85 \u4fee\u6b63\u5f0fMSI LSSF \u4fee\u6b63\u5f0fLSSF MSI +MVA +MVA +MVA +MVA</td></tr></table>",
                "html": null,
                "type_str": "table"
            }
        }
    }
}