{
    "paper_id": "O13-1002",
    "header": {
        "generated_with": "S2ORC 1.0.0",
        "date_generated": "2023-01-19T08:03:28.341072Z"
    },
    "title": "Sub-band modulation spectrum factorization in robust speech recognition",
    "authors": [
        {
            "first": "Hao-Teng",
            "middle": [],
            "last": "\u8303\u9865\u9a30",
            "suffix": "",
            "affiliation": {},
            "email": ""
        },
        {
            "first": "",
            "middle": [],
            "last": "Fan",
            "suffix": "",
            "affiliation": {},
            "email": ""
        },
        {
            "first": "Yi-Zhang",
            "middle": [],
            "last": "\u8521\u76ca\u5f70",
            "suffix": "",
            "affiliation": {},
            "email": ""
        },
        {
            "first": "",
            "middle": [],
            "last": "Cai",
            "suffix": "",
            "affiliation": {},
            "email": ""
        }
    ],
    "year": "",
    "venue": null,
    "identifiers": {},
    "abstract": "This paper proposes a novel scheme that enhance the modulation spectrum of speech features in noise speech recognition via non-negative matrix factorization (NMF). In the presented approach, we apply NMF to obtain a set of non-negative basis spectra vectors which derived from the clean speech to represent the important components for speech recognition. The difference compared to the conventional NMF-based scheme that leverages iterative search to update the full-band modulation spectra is two: first, we apply the orthogonal projection to update the low sub-band modulation spectra. Second, we process the low half-band of the",
    "pdf_parse": {
        "paper_id": "O13-1002",
        "_pdf_hash": "",
        "abstract": [
            {
                "text": "This paper proposes a novel scheme that enhance the modulation spectrum of speech features in noise speech recognition via non-negative matrix factorization (NMF). In the presented approach, we apply NMF to obtain a set of non-negative basis spectra vectors which derived from the clean speech to represent the important components for speech recognition. The difference compared to the conventional NMF-based scheme that leverages iterative search to update the full-band modulation spectra is two: first, we apply the orthogonal projection to update the low sub-band modulation spectra. Second, we process the low half-band of the",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Abstract",
                "sec_num": null
            }
        ],
        "body_text": [
            {
                "text": "modulation spectrum rather than the full-band. The presented new process improves the computation efficiency without the cost of degarded recognition performance. In the Aurora-2 database and task, the presented new NMF-based approach can achieve the average error reduction rate of over 58% relative to the baseline MFCC.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "",
                "sec_num": null
            },
            {
                "text": "Keywords: nonnegative matrix factorization, modulation spectrum, speech recognition, noise robustness.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "\u95dc\u9375\u8a5e\uff1a\u975e\u8ca0\u77e9\u9663\u5206\u89e3\u6cd5\u3001\u5f37\u5065\u6027\u3001\u8abf\u8b8a\u983b\u8b5c\u3001\u8a9e\u97f3\u8fa8\u8b58\u3002",
                "sec_num": null
            },
            {
                "text": "\u5e38\u898b\u7684\u8a9e\u97f3\u7279\u5fb5(speech features)\u5f37\u5065\u6027\u6280\u8853\u4e2d\uff0c\u6709\u4e00\u5927\u985e\u5225\u662f\u5c0d\u65bc\u8a9e\u97f3\u7279\u5fb5\u7684\u7d71\u8a08 \u503c \u505a \u6b63 \u898f \u5316 (statistics normalization) \uff0c \u5982 \u5012 \u983b \u8b5c \u5e73 \u5747 \u503c \u6b63 \u898f \u5316 \u6cd5 (cepstral mean normalization, CMN)[2]\u3001\u5012\u983b\u8b5c\u5e73\u5747\u503c\u8207\u8b8a\uf962\uf969\u6b63\u898f\u5316\u6cd5(cepstral mean and variance normalization, CMVN) [3] \u3001\u5012\u983b\u8b5c\u589e\u76ca\u6b63\u898f\u5316\u6cd5(cepstral gain normalization, CGN) [4] \u3001\u76f8 \u5c0d\u983b\u8b5c\u6cd5(RelAtive SpecTra, RASTA) [5] \u3001\u5012\u983b\u8b5c\u5e73\u5747\u503c\u8207\u8b8a\u7570\u6578\u6b63\u898f\u5316\u7d50\u5408\u81ea\u56de\u6b78\u52d5\u614b \u5e73 \u5747 \u6ffe \u6ce2 \u5668 \u6cd5 (cepstral mean and variance normalization plus auto-regressive-moving average filtering, MVA) [6] \u3001\u7d71\u8a08\u5716\u7b49\u5316\u6cd5(histogram equalization, HEQ) [7] [8]\u3001\u6642\u9593\u5e8f\u5217 \u7d50\u69cb\u6b63\u898f\u5316\u6cd5(temporal structure normalization, TSN) [9] \u7b49\uff0c\u9019\u4e9b\u6b63\u898f\u5316\u901a\u5e38\u662f\u4f5c\u7528\u65bc\u8a9e \u97f3\u7279\u5fb5\u7684\u6642\u9593\u5e8f\u5217(temporal sequence)\u4e0a\u3002 \u672c\u7bc7\u8ad6\u6587\u4e2d\uff0c\u8207\u4e0a\u8ff0\u6b63\u898f\u5316\u6cd5\u4e3b\u8981\u4e0d\u540c\u7684\u662f\uff0c\u6211\u5011\u5c0d\u65bc\u7279\u5fb5\u6642\u9593\u5e8f\u5217\u7684\u5085\u7acb\u8449\u8f49 \u63db\u3001\u5373\u5176\u8abf\u8b8a\u983b\u8b5c(modulation spectrum)\u4f5c\u5f37\u5065\u6027\u66f4\u65b0\uff0c\u5176\u5b83\u5e38\u898b\u7684\u8abf\u8b8a\u983b\u8b5c\u8655\u7406\u6280\u8853 \u6709 \u983b \u8b5c \u7d71 \u8a08 \u5716 \u7b49 \u5316 \u6cd5 (spectral histogram equalization, SHE) [10] \u3001 \u5f37 \u5ea6 \u6bd4 \u7387 \u7b49 \u5316 \u6cd5 (magnitude ratio equalization, MRE) [10] \u3001 \u8abf \u8b8a \u983b \u8b5c \u66ff \u4ee3 \u6cd5 (modulation spectrum replacement, MSR)\u8207\u8abf\u8b8a\u983b\u8b5c\u6ffe\u6ce2\u6cd5(modulation spectrum filtering, MSF) [11] \u7b49\u3002\u4f5c\u7528\u65bc \u8abf\u8b8a\u983b\u8b5c\u4e0a\u5176\u53ef\u80fd\u7684\u597d\u8655\u662f\uff0c\u6211\u5011\u53ef\u4ee5\u76f4\u63a5\u91dd\u5c0d \u4e0d \u540c \u7684 \u983b \u7387 \u6210 \u5206 \u52a0 \u4ee5 \u8655 \u7406 \u3002 \u7531 N. Kanedera ",
                "cite_spans": [
                    {
                        "start": 209,
                        "end": 212,
                        "text": "[3]",
                        "ref_id": "BIBREF1"
                    },
                    {
                        "start": 258,
                        "end": 261,
                        "text": "[4]",
                        "ref_id": "BIBREF2"
                    },
                    {
                        "start": 295,
                        "end": 298,
                        "text": "[5]",
                        "ref_id": "BIBREF3"
                    },
                    {
                        "start": 427,
                        "end": 430,
                        "text": "[6]",
                        "ref_id": "BIBREF4"
                    },
                    {
                        "start": 468,
                        "end": 471,
                        "text": "[7]",
                        "ref_id": "BIBREF5"
                    },
                    {
                        "start": 527,
                        "end": 530,
                        "text": "[9]",
                        "ref_id": "BIBREF7"
                    },
                    {
                        "start": 719,
                        "end": 723,
                        "text": "[10]",
                        "ref_id": "BIBREF8"
                    },
                    {
                        "start": 776,
                        "end": 780,
                        "text": "[10]",
                        "ref_id": "BIBREF8"
                    },
                    {
                        "start": 880,
                        "end": 884,
                        "text": "[11]",
                        "ref_id": "BIBREF9"
                    },
                    {
                        "start": 942,
                        "end": 950,
                        "text": "Kanedera",
                        "ref_id": null
                    }
                ],
                "ref_spans": [],
                "eq_spans": [],
                "section": "\u4e00\u3001\u7dd2\u8ad6 \u7576\u4e00\u5957\u8a9e\u97f3\u8fa8\u8b58\u7cfb\u7d71[1]\u61c9\u7528\u5728\u5be6\u969b\u74b0\u5883\u4e0b\u6642\uff0c\u74b0\u5883\u7684\u4e0d\u5339\u914d\u3001\u8a9e\u8005\u8b8a\u7570\u6027\u53ca\u767c\u97f3\u7684\u8b8a \u7570\u6027\u901a\u5e38\u6703\u9020\u6210\u8fa8\u8b58\u6548\u80fd\u7684\u4f4e\u843d\uff0c\u70ba\u4e86\u964d\u4f4e\u9019\u4e9b\u8b8a\u7570\u6027\u6240\u9020\u6210\u7684\u5f71\u97ff\u800c\u767c\u5c55\u7684\u6280\u8853\uff0c\u4e00 \u822c\u800c\u8a00\u7d71\u7a31\u70ba\u5f37\u5065\u6027\u6280\u8853(robustness techniques)\u3002",
                "sec_num": null
            },
            {
                "text": "EQUATION",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [
                    {
                        "start": 0,
                        "end": 8,
                        "text": "EQUATION",
                        "ref_id": "EQREF",
                        "raw_str": "V = [v , ...v ]\uff0c\u5176\u4e2d j v \u70ba\u77e9\u9663 V \u4e4b\u7b2c j \u884c\uff0c\u800c\u77e9\u9663 V \u7684\u5c3a\u5bf8\u70ba N\u00d7M\u3002\u85c9\u7531 NMF \u5206\u89e3 V \uff0c\u5f97\u5230 W \u53ca H \u5169\u500b\u975e\u8ca0\u77e9\u9663\uff0c\u5982\u4e0b\u5f0f\u8868\u793a\uff1a N r r \u1e3e\u00bb V W H (1) W \u53ca H \u5c3a\u5bf8\u5206\u5225\u70ba N\u00d7r \u8207 r\u00d7M\uff0cr \u53ef\u6c7a\u5b9a W \u53ca H \u5169\u77e9\u9663\u5c3a\u5bf8(\u4e00\u822c\u800c\u8a00 r \u9060\u5c0f\u65bc N \u8207 M)\uff0c\u5176\u4e2d W \u5e36\u6709 v \u7684\u884c\u5411\u91cf\u7684\u7d9c\u5408\u8cc7\u8a0a\uff0c\u7576\u6211\u5011\u6539\u5beb\u6210 \u00bb v Wh \u6642\uff0c\u5247 v \u3001 h \u548c V \u3001H \u5c07\u5448\u73fe\u884c\u76f8\u95dc\u7684\u95dc\u4fc2\uff0c\u63db\u53e5\u8a71\u8aaa\uff0cv \u53ef\u8996\u70ba W \u4e4b\u884c\u5411\u91cf\u7684\u7dda\u6027\u7d44\u5408\u800c h \u70ba\u5176\u6b0a\u91cd\u6bd4\uff0c\u7136 \u800c\u6700\u4e3b\u8981\u7684\u76ee\u7684\u5728\u65bc\u4f7f W \u8207 H \u5169\u77e9\u9663\u7684\u4e58\u7a4d\u903c\u8fd1\u65bc V \uff0c\u5982\u6b64\u4e00\u4f86\u624d\u80fd\u5f97\u5230\u4e00\u7d44\u53ef\u9760\u7684 \u57fa\u5e95\uff0c\u9019\u500b\u904e\u7a0b\u6211\u5011\u4ee5\u4e0b\u5f0f\u8868\u793a\uff1a ( ) ( ) 2 , ,",
                        "eq_num": ", ( ) min"
                    }
                ],
                "section": "\u4e00\u3001\u7dd2\u8ad6 \u7576\u4e00\u5957\u8a9e\u97f3\u8fa8\u8b58\u7cfb\u7d71[1]\u61c9\u7528\u5728\u5be6\u969b\u74b0\u5883\u4e0b\u6642\uff0c\u74b0\u5883\u7684\u4e0d\u5339\u914d\u3001\u8a9e\u8005\u8b8a\u7570\u6027\u53ca\u767c\u97f3\u7684\u8b8a \u7570\u6027\u901a\u5e38\u6703\u9020\u6210\u8fa8\u8b58\u6548\u80fd\u7684\u4f4e\u843d\uff0c\u70ba\u4e86\u964d\u4f4e\u9019\u4e9b\u8b8a\u7570\u6027\u6240\u9020\u6210\u7684\u5f71\u97ff\u800c\u767c\u5c55\u7684\u6280\u8853\uff0c\u4e00 \u822c\u800c\u8a00\u7d71\u7a31\u70ba\u5f37\u5065\u6027\u6280\u8853(robustness techniques)\u3002",
                "sec_num": null
            },
            {
                "text": "v \uf025 \u8868\u793a) \uff0c\u6211 \u5011\u5229\u7528\u524d\u6b65\u9a5f\u6240\u5f97\u7684\u77e9\u9663 W \uff0c\u5c0d\u6b64\u5411\u91cf v \uf025 \u4ee5 NMF \u7684\u8fed\u4ee3\u6cd5\u4f5c\u903c\u8fd1\u9032\u800c\u6c42\u5f97\u6700\u5f8c\u7684 ( ) L h \uff0c\u800c ( ) L h \u5373\u70ba v \uf025 \u8207 W \u4e4b\u9593\u7684\u6b0a\u91cd\u6bd4\u95dc\u4fc2\uff0c\u6574\u500b\u904e\u7a0b\u5982\u4e0b\uff1a \u521d\u59cb\uff1a \u4efb\u610f\u6307\u5b9a\u4e00\u975e\u8ca0\u5411\u91cf ( ) 0 h \u3002 \u8fed\u4ee3\uff1a\u524d\u5f8c\u5169\u5411\u91cf ( ) j h \u8207 ( ) 1 j + h \u7684\u95dc\u4fc2\u70ba\uff1a ( ) ( ) 1 ( ) ( ) ( ) T j j k k k T j k + = W v h h W Wh \uf025 (3) \u5176\u4e2d\u5c0d\u5411\u91cf\u4f7f\u7528\u4e0b\u6a19 k \u4ee3\u8868\u6b64\u5411\u91cf\u4e2d\u7684\u7b2c k \u9805\u3002 \u7d42\u6b62\uff1a\u5728\u591a\u6b21\u8fed\u4ee3\u4e4b\u5f8c\uff0c\u85c9\u7531\u6700\u5f8c\u4e00\u6b21\u8fed\u4ee3\u6240\u5f97\u7684\u5411\u91cf ( ) L h (\u5047\u8a2d\u8fed\u4ee3\u7e3d\u6b21\u6578\u70ba L) \uff0c\u65b0\u8abf\u8b8a\u983b\u8b5c\u5f37\u5ea6\u70ba\uff1a ( ) \u00a2 L v = Wh \uf025 . (4) \u6b64\u65b0\u7684\u8abf\u8b8a\u983b\u8b5c\u5f37\u5ea6\u914d\u5408\u539f\u59cb\u7684\u76f8\u4f4d\u6210\u5206(phase part)\uff0c\u7d93\u904e\u53cd\u5085\u7acb\u8449\u8f49\u63db\u5c31\u53ef\u5f97\u5230\u65b0\u7684 \u7279\u5fb5\u6642\u9593\u5e8f\u5217\u3002 (\u4e09)NMF \u4f7f\u7528\u65bc\u8abf\u8b8a\u983b\u8b5c\u4e4b\u6295\u5f71\u5f0f\u66f4\u65b0 \u5728\u9019\u88e1\uff0c\u6211\u5011\u70ba\u524d\u8ff0\u7684 NMF \u8abf\u8b8a\u983b\u8b5c\u66f4\u65b0\u6280\u8853\uff0c\u63d0\u51fa\u4e86\u5169\u500b\u964d\u4f4e\u8a08\u7b97\u8907\u96dc\u5ea6\u7684\u4fee \u6b63\u6b65\u9a5f\uff0c\u5206\u8ff0\u5982\u4e0b\uff1a I. \u4f7f\u7528\u6b63\u4ea4\u6295\u5f71(orthogonal projection)\u66ff\u4ee3\u539f\u59cb\u7684\u8fed\u4ee3\u6cd5\uff1a \u6839\u64da\u4e4b\u524d\u7684\u63cf\u8ff0\uff0c\u65b0\u8abf\u8b8a\u983b\u8b5c\u5f37\u5ea6\u7531\u65bc \u00a2 v \uf025 \u5fc5\u9808\u4ee5\u8fed\u4ee3\u65b9\u5f0f\u8f3e\u8f49\u6c42\u5f97\uff0c\u56e0\u6b64\u6975\u53ef\u80fd\u5f71 \u97ff\u6f14\u7b97\u6cd5\u7684\u904b\u7b97\u8907\u96dc\u5ea6\u3002\u56e0\u6b64\uff0c\u6211\u5011\u5e0c\u671b\u80fd\u627e\u51fa\"\u55ae\u6b21\"\u7684\u904b\u7b97\u6cd5\u4f86\u6c42\u53d6\u65b0\u8abf\u8b8a\u983b\u8b5c\u5f37 \u5ea6\u3002\u5982\u5f0f(4)\u6240\u793a\uff0c\u65b0\u8abf\u8b8a\u983b\u8b5c\u5f37\u5ea6 \u00a2 v \uf025 \u662f\u7531\u4e7e\u6de8\u8a9e\u97f3\u6240\u5f97\u7684\u57fa\u5e95\u77e9\u9663 W \u4e2d\u6bcf\u4e00\u500b\u884c\u5411\u91cf \u4f5c\u7dda\u6027\u52a0\u6210(linear combination)\u800c\u5f97\uff0c\u63db\u8a00\u4e4b\uff0c \u00a2 v \uf025 \u5fc5\u843d\u5728\u57fa\u5e95\u77e9\u9663 W \u4e4b\u884c\u7a7a\u9593(column space)\u4e2d\u3002\u6839\u64da\u7dda\u6027\u4ee3\u6578\u7684\u77e5\u8b58\uff0c\u6211\u5011\u76f4\u63a5\u63a1\u7528\u539f\u59cb\u8abf\u8b8a\u983b\u8b5c\u5f37\u5ea6 v \uf025 \u6295\u5f71\u65bc\u57fa\u5e95\u77e9\u9663 W \u4e4b\u884c\u7a7a\u9593\u4e4b\u5206\u91cf\uff0c\u4f5c\u70ba\u65b0\u7684\u8abf\u8b8a\u983b\u8b5c\u5f37\u5ea6\uff0c\u53ef\u8868\u793a\u70ba\uff1a proj \u00a2 = = T -1 T W v ( v ) W ( W W ) W v \uf025 (5) \u6216 proj \u00a2 = = T W v ( v ) B Bv \uf025 (6) \u5176\u4e2d\uff0c\u77e9\u9663 B \u5305\u542b\u4e86\u57fa\u5e95\u77e9\u9663 W \u4e4b\u884c\u7a7a\u9593\u7684\u6b63\u4ea4\u57fa\u5e95(orthogonal basis)\uff0c\u63a5\u4e0b\u4f86\u6211\u5011\u5c07 \u63a1\u7528(6)\u5f0f\u4f86\u6c42\u5f97\u65b0\u7684\u8abf\u8b8a\u983b\u8b5c\u5f37\u5ea6\u800c\u4e0d\u662f(5)\u5f0f\u7684\u4e3b\u8981\u539f\u56e0\uff0c\u662f\u56e0\u70ba W \u6709\u53ef\u80fd\u70ba\u7a00\u758f\u77e9 \u9663(sparse matrix)\u4e14\u79e9\u6578\u5c0f\u65bc r\uff0c\u82e5\u6b64\u60c5\u6cc1\u4e00\u65e6\u767c\u751f\uff0c\u5247 T -1 (W W) \u9805\u5c07\u5448\u73fe\u7121\u89e3\u7684\u60c5\u5f62\uff0c \u56e0\u6b64\u63a1\u7528(6)\u5f0f\u5c07\u53ef\u907f\u514d\u6b64\u554f\u984c\uff0c\u4e26\u4e14\u5e36\u6709\u984d\u5916\u7684\u597d\u8655\u70ba\u77e9\u9663 B \u53ef\u5728\u66f4\u65b0\u6bcf\u4e00\u53e5\u8a9e\u97f3\u8abf \u8b8a\u983b\u8b5c\u524d\u5c31\u4e8b\u5148\u7531\u77e9\u9663 W \u6c42\u5f97\uff0c\u76f8\u8f03\u4e4b\u4e0b\u8907\u96dc\u5ea6\u76f8\u5c0d\u8f03\u4f4e\u3002\u63a1\u7528\u6295\u5f71\u6cd5\u7684\u6700\u5927\u512a\u9ede\u5728 \u65bc\u7121\u9808\u96a8\u8457\u4e0d\u540c\u7684\u8a9e\u53e5\u800c\u91cd\u65b0\u8a08\u7b97\uff0c\u6240\u4ee5\u4e26\u4e0d\u6703\u5f71\u97ff\u66f4\u65b0\u6bcf\u4e00\u53e5\u7279\u5fb5\u4e4b\u904b\u7b97\u8907\u96dc\u5ea6\u3002\u540c \u6642\uff0c\u5728\u7b49\u5f0f(6)\u7684\u904b\u7b97\u4e2d\uff0c\u6211\u5011\u53ef\u4ee5\u5148\u8a08\u7b97\u5411\u91cf T B v \uff0c\u518d\u5c07\u5176\u5de6\u4e58\u4e0a\u77e9\u9663\uff0c\u9019\u6a23\u7684\u4f5c\u6cd5 \u6703\u9060\u6bd4\u76f4\u63a5\u5c07\u4e8b\u5148\u7b97\u597d\u7684\u6295\u5f71\u77e9\u9663 T BB \u4e58\u4e0a\u5411\u91cf v \u4f86\u7684\u5c11\u3002\u4f8b\u5982\uff0c\u77e9\u9663 B \u7684\u5c3a\u5bf8(\u81f3\u591a)",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "\u4e00\u3001\u7dd2\u8ad6 \u7576\u4e00\u5957\u8a9e\u97f3\u8fa8\u8b58\u7cfb\u7d71[1]\u61c9\u7528\u5728\u5be6\u969b\u74b0\u5883\u4e0b\u6642\uff0c\u74b0\u5883\u7684\u4e0d\u5339\u914d\u3001\u8a9e\u8005\u8b8a\u7570\u6027\u53ca\u767c\u97f3\u7684\u8b8a \u7570\u6027\u901a\u5e38\u6703\u9020\u6210\u8fa8\u8b58\u6548\u80fd\u7684\u4f4e\u843d\uff0c\u70ba\u4e86\u964d\u4f4e\u9019\u4e9b\u8b8a\u7570\u6027\u6240\u9020\u6210\u7684\u5f71\u97ff\u800c\u767c\u5c55\u7684\u6280\u8853\uff0c\u4e00 \u822c\u800c\u8a00\u7d71\u7a31\u70ba\u5f37\u5065\u6027\u6280\u8853(robustness techniques)\u3002",
                "sec_num": null
            },
            {
                "text": "Proceedings of the Twenty-Fifth Conference on Computational Linguistics and Speech Processing (ROCLING 2013) (i,low) \u8207 NMF (f,low) \u7686\u662f\u66f4\u65b0\u534a\u983b\u5e36\u7684\u983b\u8b5c\u5f37\u5ea6 (\u76f8 \u7576\u65bc\u66f4\u65b0 256 \u500b\u4f4e\u983b\u7387\u9ede) \uff0c\u5982\u679c\u6211\u5011\u5c07\u66f4\u65b0\u7bc4\u570d\u7e2e\u6e1b\u81f3\u5168\u983b\u5e36\u7684 1/3 \u8207 1/4(\u5206\u5225\u70ba \u4e09\u500b\u8868\u4e2d\u6240\u5217\u4e4b 170 \u9ede\u8207 128 \u9ede) \uff0c\u5176\u8fa8\u8b58\u7387\u76f8\u5c0d\u65bc\u5168\u983b\u5e36\u8207\u534a\u983b\u5e36\u7684\u66f4\u65b0\u6cd5\u800c\u8a00\uff0c \u78ba\u5be6\u6703\u9010\u6f38\u8b8a\u4f4e\uff0c\u4f46\u662f\u8b8a\u4f4e\u7684\u5e45\u5ea6\u4e26\u672a\u5f88\u986f\u8457\uff0cPCA \u6cd5\u4e5f\u662f\u5982\u6b64\uff0c\u9019\u4ee3\u8868\u4e86\u5c0d\u65bc NMF \u8207 PCA \u6cd5\u800c\u8a00\uff0c\u6211\u5011\u53ef\u4ee5\u66f4\u65b0\u6bd4\u4e00\u534a\u66f4\u5c11\u7684\u983b\u7387\u9ede\uff0c\u5c31\u8db3\u4ee5\u8da8\u8fd1\u66f4\u65b0\u5168\u90e8\u983b\u7387\u9ede\u4e4b \u6548\u80fd\u3002 5. \u6211\u5011\u6240\u63d0\u7684\u4e09\u7a2e\u65b0\u65b9\u6cd5\uff1aNMF (i,low) \u3001NMF (p,f) \u8207 NMF (p,low) \u82e5\u8207\u539f\u59cb\u7684 NMF (i,f) \u6cd5\u76f8\u8f03\uff0c \u5168\u983b\u5f0f\u7684 NMF (p,f) \u3001\u534a\u983b\u5f0f\u7684 NMF (i,low) ",
                "cite_spans": [
                    {
                        "start": 109,
                        "end": 116,
                        "text": "(i,low)",
                        "ref_id": null
                    },
                    {
                        "start": 423,
                        "end": 430,
                        "text": "(i,low)",
                        "ref_id": null
                    }
                ],
                "ref_spans": [],
                "eq_spans": [],
                "section": "\u4e00\u3001\u7dd2\u8ad6 \u7576\u4e00\u5957\u8a9e\u97f3\u8fa8\u8b58\u7cfb\u7d71[1]\u61c9\u7528\u5728\u5be6\u969b\u74b0\u5883\u4e0b\u6642\uff0c\u74b0\u5883\u7684\u4e0d\u5339\u914d\u3001\u8a9e\u8005\u8b8a\u7570\u6027\u53ca\u767c\u97f3\u7684\u8b8a \u7570\u6027\u901a\u5e38\u6703\u9020\u6210\u8fa8\u8b58\u6548\u80fd\u7684\u4f4e\u843d\uff0c\u70ba\u4e86\u964d\u4f4e\u9019\u4e9b\u8b8a\u7570\u6027\u6240\u9020\u6210\u7684\u5f71\u97ff\u800c\u767c\u5c55\u7684\u6280\u8853\uff0c\u4e00 \u822c\u800c\u8a00\u7d71\u7a31\u70ba\u5f37\u5065\u6027\u6280\u8853(robustness techniques)\u3002",
                "sec_num": null
            },
            {
                "text": "\u70ba N\u00d7r\uff0c\u77e9\u9663 T BB \u7684\u5c3a\u5bf8\u70ba N\u00d7N\uff0c\u6545\u82e5\u76f4\u63a5\u628a T BB \u53f3\u4e58\u5c3a\u5bf8\u70ba N\u00d71 \u7684\u5411\u91cf v \uff0c\u6240\u9700\u7684 \u4e58\u6cd5\u6578\u76ee\u70ba N 2 \uff1b\u7136\u800c\uff0c\u5148\u8a08\u7b97 T B v \uff0c\u518d\u8a08\u7b97 T B(B v) \u6240\u9700\u7684\u4e58\u6cd5\u6578\u76ee\u70ba Nr\uff0bNr=2Nr\uff0c \u800c\u5be6\u969b\u904b\u7528\u4e0a\uff0c2Nr \u901a\u5e38\u4f4e\u65bc N 2 \uff0c\u9019\u662f\u56e0\u70ba\u7531\u65bc\u5728 NMF \u904b\u7b97\u4e2d\uff0c\u6c42\u53d6\u7684\u57fa\u5e95\u77e9\u9663 W \u5176 \u884c\u5411\u91cf\u500b\u6578 r \u901a\u5e38\u9060\u4f4e\u65bc\u5176\u5c3a\u5bf8 N\u3002 (\u5728\u6211\u5011\u7684\u5be6\u9a57\u8a2d\u5b9a\u4e2d\uff0cN=513, r=10\uff0c\u5247 2Nr=5120 < 65536= N 2 ) \uff0c\u6574\u9ad4\u904e\u7a0b\u5982\u4e0b\u5716\u4e00\u6240\u793a\u3002 \u5716\u4e00 \u6295\u5f71\u5f0f\u8abf\u8b8a\u983b\u8b5c\u66f4\u65b0\u6cd5\u6d41\u7a0b\u5716 II. \u4f7f\u7528\u5b50\u983b\u5e36\u66f4\u65b0\u66ff\u4ee3\u539f\u59cb\u7684\u5168\u983b\u5e36\u66f4\u65b0\uff1a \u5982\u540c\u5148\u524d\u6240\u63d0\u5230\u7684\uff0c\u4e0d\u540c\u983b\u5e36\u7684\u8a9e\u97f3\u7279\u5fb5\u6642\u9593\u5e8f\u5217\u8abf\u8b8a\u983b\u8b5c\u5c0d\u8a9e\u97f3\u8fa8\u8b58\u7684\u91cd\u8981\u6027\u4e26 \u4e0d\u4e00\u81f4\uff0c\u5716\u4e8c\u70ba\u5229\u7528 NMF \u6240\u6c42\u5f97\u7684\u57fa\u5e95\u77e9\u9663 W \u4e4b\u8abf\u8b8a\u983b\u8b5c\u5f37\u5ea6\u5206\u5e03\u5716\uff0c\u6211\u5011\u53ef\u89c0\u5bdf\u5230 \u5176\u5206\u4f48\u5340\u57df\u5927\u591a\u96c6\u4e2d\u65bc 10 Hz \u4e4b\u524d\uff0c\u7531\u6b64\u53ef\u77e5\u4f4e\u983b\u5e36\u7684\u6210\u5206\u5c24\u5176\u91cd\u8981\uff0c\u7b49\u91cf\u7684\u96dc\u8a0a\u5e72\u64fe \u82e5\u5b58\u5728\u65bc\u4f4e\u983b\u5e36\uff0c\u76f8\u8f03\u65bc\u5b58\u5728\u65bc\u9ad8\u983b\u5e36\uff0c\u5c0d\u65bc\u8a9e\u97f3\u8fa8\u8b58\u7684\u7cbe\u78ba\u5ea6\u5f71\u97ff\u66f4\u5927\uff0c\u539f\u59cb\u7684 NMF \u66f4\u65b0\u6cd5\u5247\u662f\u5c0d\u65bc\u5168\u983b\u5e36\u4e00\u4f75\u66f4\u65b0\uff0c\u800c\u5728\u9019\u88e1\u7684\u65b0\u6b65\u9a5f\u88e1\uff0c\u6211\u5011\u53ea\u91dd\u5c0d\u524d\u534a\u6bb5\u5b50\u983b\u5e36\u7684\u8abf \u8b8a\u983b\u8b5c\u4f5c\u66f4\u65b0\uff0c\u76f8\u8f03\u65bc\u539f\u65b9\u6cd5\uff0c\u4e0d\u50c5\u53ef\u4ee5\u964d\u4f4e\u4e00\u534a\u7684\u8907\u96dc\u5ea6\uff0c\u6211\u5011\u4e5f\u9810\u671f\u9019\u6a23\u7684\u8655\u7406\u4e26 \u4e0d\u6703\u5f71\u97ff\u5176\u5f8c\u7684\u8fa8\u8b58\u7cbe\u78ba\u5ea6\uff0c\u6b64\u9ede\u53ef\u5728\u4e4b\u5f8c\u7684\u5be6\u9a57\u6578\u64da\u4e2d\u8b49\u5be6\u3002\u63a5\u4e0b\u4f86\u6211\u5011\u5c07\u539f\u59cb\u7684 NMF \u66f4\u65b0\u6cd5\u4ee5 NMF (i,f) \u8868\u793a\uff0c\u5176\u4e2d\u7684\u4ee3\u865f\"i\"\u8207\"f\"\u5206\u5225\u4ee3\u8868\u4e86\u8fed\u4ee3(iteration)\u8207\u5168\u983b\u5e36 (full-band)\u5169\u500b\u8a5e) \uff0c\u53ca\u672c\u8ad6\u6587\u63d0\u51fa\u7684\u5169\u7a2e\u6539\u826f\u5f0f NMF \u6cd5\u7684\u6392\u5217\u7d44\u5408\uff0c\u5206\u5225\u4ee5 NMF (p,f) \u3001 NMF (i,low) \u8207 NMF (p,low) \u8868\u793a\uff0c\u5176\u4e2d\u7684\u4ee3\u865f\"p\"\u8207\"low\"\u5206\u5225\u4ee3\u8868\u4e86\u6295\u5f71 projection \u8207\u4f4e\u983b\u5e36 low-band \u5169\u500b\u8a5e) \u3002",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "\u4e00\u3001\u7dd2\u8ad6 \u7576\u4e00\u5957\u8a9e\u97f3\u8fa8\u8b58\u7cfb\u7d71[1]\u61c9\u7528\u5728\u5be6\u969b\u74b0\u5883\u4e0b\u6642\uff0c\u74b0\u5883\u7684\u4e0d\u5339\u914d\u3001\u8a9e\u8005\u8b8a\u7570\u6027\u53ca\u767c\u97f3\u7684\u8b8a \u7570\u6027\u901a\u5e38\u6703\u9020\u6210\u8fa8\u8b58\u6548\u80fd\u7684\u4f4e\u843d\uff0c\u70ba\u4e86\u964d\u4f4e\u9019\u4e9b\u8b8a\u7570\u6027\u6240\u9020\u6210\u7684\u5f71\u97ff\u800c\u767c\u5c55\u7684\u6280\u8853\uff0c\u4e00 \u822c\u800c\u8a00\u7d71\u7a31\u70ba\u5f37\u5065\u6027\u6280\u8853(robustness techniques)\u3002",
                "sec_num": null
            },
            {
                "text": "(i,f) \u3001NMF (i,low) \u3001NMF (p,f) \u8207 NMF (p,low) )\u7684\u904b\u7b97\u91cf\u8207\u6240\u9700 \u6642\u9593\u3001\u85c9\u6b64\u6bd4\u8f03\u5b83\u5011\u7684\u8907\u96dc\u5ea6\u3002 \u5728\u53c3\u6578\u7684\u8a2d\u5b9a\u4e0a\uff0c\u6c42\u53d6\u8abf\u8b8a\u983b\u8b5c\u7684 FFT \u9ede\u6578\u70ba 1024 \u9ede\uff0c\u7531\u65bc\u5171\u8edb\u5c0d\u7a31\u7279\u6027\u7684\u95dc\u4fc2\uff0c\u56e0 \u6b64\u5f0f(1)\u4e2d\u7684\u8abf\u8b8a\u983b\u8b5c\u9ede\u6578 N=513\uff0c\u57fa\u5e95\u77e9\u9663 W \u7684\u884c\u5411\u91cf\u6578 r=10\uff0c\u8fed\u4ee3\u5f0f NMF \u6cd5(NMF (i,f) \u8207 NMF (i,low) )\u4e4b\u8fed\u4ee3\u6578 L \u8a2d\u70ba 100\uff0c\u5247\u56db\u7a2e\u6f14\u7b97\u6cd5 NMF (i,f) \u3001NMF (i,low) \u3001NMF (p,f) \u8207 NMF (p,",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "\u4e00\u3001\u7dd2\u8ad6 \u7576\u4e00\u5957\u8a9e\u97f3\u8fa8\u8b58\u7cfb\u7d71[1]\u61c9\u7528\u5728\u5be6\u969b\u74b0\u5883\u4e0b\u6642\uff0c\u74b0\u5883\u7684\u4e0d\u5339\u914d\u3001\u8a9e\u8005\u8b8a\u7570\u6027\u53ca\u767c\u97f3\u7684\u8b8a \u7570\u6027\u901a\u5e38\u6703\u9020\u6210\u8fa8\u8b58\u6548\u80fd\u7684\u4f4e\u843d\uff0c\u70ba\u4e86\u964d\u4f4e\u9019\u4e9b\u8b8a\u7570\u6027\u6240\u9020\u6210\u7684\u5f71\u97ff\u800c\u767c\u5c55\u7684\u6280\u8853\uff0c\u4e00 \u822c\u800c\u8a00\u7d71\u7a31\u70ba\u5f37\u5065\u6027\u6280\u8853(robustness techniques)\u3002",
                "sec_num": null
            },
            {
                "text": "[1] \u738b\u5c0f\u5ddd, \"\u8a9e\u97f3\u8a0a\u865f\u8655\u7406,\" \u5168\u83ef\u79d1\u6280\u5716\u66f8, 2004.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "\u53c3\u8003\u6587\u737b",
                "sec_num": null
            },
            {
                "text": "[2] S. Furui, \"Cepstral analysis technique for automatic speaker verification,\" IEEE Trans.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "\u53c3\u8003\u6587\u737b",
                "sec_num": null
            },
            {
                "text": "Proceedings of the Twenty-Fifth Conference on Computational Linguistics and Speech Processing (ROCLING 2013) ",
                "cite_spans": [
                    {
                        "start": 94,
                        "end": 108,
                        "text": "(ROCLING 2013)",
                        "ref_id": null
                    }
                ],
                "ref_spans": [],
                "eq_spans": [],
                "section": "\u53c3\u8003\u6587\u737b",
                "sec_num": null
            },
            {
                "text": "Proceedings of the Twenty-Fifth Conference on Computational Linguistics and Speech Processing(ROCLING 2013)",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "",
                "sec_num": null
            }
        ],
        "back_matter": [],
        "bib_entries": {
            "BIBREF1": {
                "ref_id": "b1",
                "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": "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
            },
            "BIBREF2": {
                "ref_id": "b2",
                "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": "1021--1024",
                "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), pp.1021-1024, 2004.",
                "links": null
            },
            "BIBREF3": {
                "ref_id": "b3",
                "title": "RASTA Processing of speech features",
                "authors": [
                    {
                        "first": "H",
                        "middle": [],
                        "last": "Hermansky",
                        "suffix": ""
                    },
                    {
                        "first": "N",
                        "middle": [],
                        "last": "Morgan",
                        "suffix": ""
                    }
                ],
                "year": 1994,
                "venue": "IEEE Trans-actions on Industrial Electronics",
                "volume": "",
                "issue": "",
                "pages": "",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "H. Hermansky and N. Morgan, \"RASTA Processing of speech features,\" IEEE Trans-actions on Industrial Electronics, IEEE Trans. On Speech and Audio Processing, 1994.",
                "links": null
            },
            "BIBREF4": {
                "ref_id": "b4",
                "title": "MVA Processing of speech features",
                "authors": [
                    {
                        "first": "C",
                        "middle": [],
                        "last": "Chen",
                        "suffix": ""
                    },
                    {
                        "first": "",
                        "middle": [],
                        "last": "Bilmes",
                        "suffix": ""
                    }
                ],
                "year": 2006,
                "venue": "Speech and Language Processing",
                "volume": "",
                "issue": "",
                "pages": "257--270",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "C. Chen and Bilmes, \"MVA Processing of speech features,\" IEEE Trans. on Audio, Speech and Language Processing, pp.257-270, 2006.",
                "links": null
            },
            "BIBREF5": {
                "ref_id": "b5",
                "title": "Quantile based histogram equalization for noise robust large vocabulary speech recognition",
                "authors": [
                    {
                        "first": "F",
                        "middle": [],
                        "last": "Hilger",
                        "suffix": ""
                    },
                    {
                        "first": "H",
                        "middle": [],
                        "last": "Ney",
                        "suffix": ""
                    }
                ],
                "year": 2006,
                "venue": "Speech and Language Processing",
                "volume": "",
                "issue": "",
                "pages": "845--854",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "F.Hilger and H. Ney, \"Quantile based histogram equalization for noise robust large vocabulary speech recognition,\" IEEE Trans. on Audio, Speech and Language Processing, pp.845-854, 2006.",
                "links": null
            },
            "BIBREF6": {
                "ref_id": "b6",
                "title": "Histogram equalization of speech representation for robust speech recognition",
                "authors": [
                    {
                        "first": "A",
                        "middle": [],
                        "last": "Torre",
                        "suffix": ""
                    },
                    {
                        "first": "A",
                        "middle": [
                            "M"
                        ],
                        "last": "Peinado",
                        "suffix": ""
                    },
                    {
                        "first": "J",
                        "middle": [
                            "C"
                        ],
                        "last": "Segura",
                        "suffix": ""
                    },
                    {
                        "first": "J",
                        "middle": [
                            "L"
                        ],
                        "last": "Perez-Cordoba",
                        "suffix": ""
                    },
                    {
                        "first": "M",
                        "middle": [
                            "C"
                        ],
                        "last": "Benitez",
                        "suffix": ""
                    },
                    {
                        "first": "A",
                        "middle": [
                            "J"
                        ],
                        "last": "Rubio",
                        "suffix": ""
                    }
                ],
                "year": 2005,
                "venue": "IEEE Trans. Speech Audio Processing",
                "volume": "13",
                "issue": "3",
                "pages": "355--366",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "A. Torre, A. M. Peinado, J. C. Segura, J. L. Perez-Cordoba, M. C. Benitez, and A. J. Rubio, \"Histogram equalization of speech representation for robust speech recognition,\" IEEE Trans. Speech Audio Processing, vol. 13, no. 3, pp. 355-366, May 2005.",
                "links": null
            },
            "BIBREF7": {
                "ref_id": "b7",
                "title": "Normalizing the speech modulation spectrum for robust speech recognition",
                "authors": [
                    {
                        "first": "X",
                        "middle": [],
                        "last": "Xiao",
                        "suffix": ""
                    },
                    {
                        "first": "E",
                        "middle": [
                            "S"
                        ],
                        "last": "Chug",
                        "suffix": ""
                    },
                    {
                        "first": "H",
                        "middle": [],
                        "last": "Li",
                        "suffix": ""
                    }
                ],
                "year": 2007,
                "venue": "IEEE International Conference on Acoustics, Speech and Signal Processing",
                "volume": "",
                "issue": "",
                "pages": "",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "X. Xiao, E. S. Chug and H. Li, \"Normalizing the speech modulation spectrum for robust speech recognition,\" IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP 2007), 2007.",
                "links": null
            },
            "BIBREF8": {
                "ref_id": "b8",
                "title": "Modulation spectrum equalization for robust speech recognition",
                "authors": [
                    {
                        "first": "L",
                        "middle": [
                            "C"
                        ],
                        "last": "Sun",
                        "suffix": ""
                    },
                    {
                        "first": "C",
                        "middle": [
                            "W"
                        ],
                        "last": "Hsu",
                        "suffix": ""
                    },
                    {
                        "first": "L",
                        "middle": [
                            "S"
                        ],
                        "last": "Lee",
                        "suffix": ""
                    }
                ],
                "year": 2007,
                "venue": "Automatic Speech Recognition and Understanding (ASRU2007)",
                "volume": "",
                "issue": "",
                "pages": "",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "L. C. Sun, C. W. Hsu and L. S. Lee, \"Modulation spectrum equalization for robust speech recognition,\" 2007 Automatic Speech Recognition and Understanding (ASRU2007).",
                "links": null
            },
            "BIBREF9": {
                "ref_id": "b9",
                "title": "Study of modulation spectrum normalization for robust speech recognition",
                "authors": [
                    {
                        "first": "C",
                        "middle": [
                            "C"
                        ],
                        "last": "Lai",
                        "suffix": ""
                    }
                ],
                "year": 2011,
                "venue": "",
                "volume": "",
                "issue": "",
                "pages": "",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "C. C. Lai, \"Study of modulation spectrum normalization for robust speech recognition,\" Master's thesis, National Chi Nan University, Taiwan, July. 2011.",
                "links": null
            },
            "BIBREF10": {
                "ref_id": "b10",
                "title": "On the importance of various modulation frequencies for speech recognition",
                "authors": [
                    {
                        "first": "N",
                        "middle": [],
                        "last": "Kanedera",
                        "suffix": ""
                    },
                    {
                        "first": "T",
                        "middle": [],
                        "last": "Arai",
                        "suffix": ""
                    },
                    {
                        "first": "H",
                        "middle": [],
                        "last": "Hermansky",
                        "suffix": ""
                    },
                    {
                        "first": "M",
                        "middle": [],
                        "last": "Pavel",
                        "suffix": ""
                    }
                ],
                "year": 1997,
                "venue": "",
                "volume": "",
                "issue": "",
                "pages": "",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "N. Kanedera, T. Arai, H. Hermansky, and M. Pavel, \"On the importance of various modulation frequencies for speech recognition,\" Eurospeech 1997.",
                "links": null
            },
            "BIBREF11": {
                "ref_id": "b11",
                "title": "Learning the parts of objects by non-negative matrix factorization",
                "authors": [
                    {
                        "first": "D",
                        "middle": [
                            "D"
                        ],
                        "last": "Lee",
                        "suffix": ""
                    },
                    {
                        "first": "H",
                        "middle": [
                            "S"
                        ],
                        "last": "Seung",
                        "suffix": ""
                    }
                ],
                "year": 1999,
                "venue": "Nature",
                "volume": "401",
                "issue": "",
                "pages": "788--791",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "D. D. Lee and H. S. Seung, \"Learning the parts of objects by non-negative matrix factorization,\" Nature, 401:788-791, 1999.",
                "links": null
            },
            "BIBREF12": {
                "ref_id": "b12",
                "title": "Algorithms for non-negative matrix factorization",
                "authors": [
                    {
                        "first": "D",
                        "middle": [
                            "D"
                        ],
                        "last": "Lee",
                        "suffix": ""
                    },
                    {
                        "first": "H",
                        "middle": [
                            "S"
                        ],
                        "last": "Seung",
                        "suffix": ""
                    }
                ],
                "year": 2000,
                "venue": "Ad-vances in Neural Information Processing Systems 13",
                "volume": "",
                "issue": "",
                "pages": "",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "D. D. Lee and H. S. Seung, \"Algorithms for non-negative matrix factorization,\" Ad-vances in Neural Information Processing Systems 13, 2000.",
                "links": null
            },
            "BIBREF13": {
                "ref_id": "b13",
                "title": "Non-negative matrix factorization as noise-robust feature extractor for speech recognition",
                "authors": [
                    {
                        "first": "B",
                        "middle": [],
                        "last": "Schuller",
                        "suffix": ""
                    },
                    {
                        "first": "F",
                        "middle": [],
                        "last": "Weninger",
                        "suffix": ""
                    },
                    {
                        "first": "M",
                        "middle": [],
                        "last": "Wollmer",
                        "suffix": ""
                    },
                    {
                        "first": "Y",
                        "middle": [],
                        "last": "Sun",
                        "suffix": ""
                    },
                    {
                        "first": "G",
                        "middle": [],
                        "last": "Rigoll",
                        "suffix": ""
                    }
                ],
                "year": 2010,
                "venue": "ICASSP",
                "volume": "",
                "issue": "",
                "pages": "",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "B. Schuller, F. Weninger, M. Wollmer, Y. Sun, and G. Rigoll, \"Non-negative matrix factorization as noise-robust feature extractor for speech recognition,\" ICASSP 2010.",
                "links": null
            },
            "BIBREF14": {
                "ref_id": "b14",
                "title": "Compressed sensing",
                "authors": [
                    {
                        "first": "David",
                        "middle": [
                            "L"
                        ],
                        "last": "Donoho",
                        "suffix": ""
                    }
                ],
                "year": 2006,
                "venue": "IEEE Transactions on Information Theory",
                "volume": "52",
                "issue": "4",
                "pages": "1289--1306",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "David L. Donoho, \"Compressed sensing,\" IEEE Transactions on Information Theory, 52(4): 1289-1306, 2006.",
                "links": null
            },
            "BIBREF15": {
                "ref_id": "b15",
                "title": "Modulation spectrum factorization for robust speech recognition",
                "authors": [
                    {
                        "first": "Wen-Yi",
                        "middle": [],
                        "last": "Chu",
                        "suffix": ""
                    },
                    {
                        "first": "Jeih-Weih",
                        "middle": [],
                        "last": "Hung",
                        "suffix": ""
                    },
                    {
                        "first": "Berlin",
                        "middle": [],
                        "last": "Chen",
                        "suffix": ""
                    }
                ],
                "year": 2011,
                "venue": "APSIPA",
                "volume": "",
                "issue": "",
                "pages": "",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Wen-Yi Chu, Jeih-weih Hung, and Berlin Chen, \"Modulation spectrum factorization for robust speech recognition,\" APSIPA 2011.",
                "links": null
            },
            "BIBREF16": {
                "ref_id": "b16",
                "title": "Proceedings of the Twenty-Fifth Conference on Computational Linguistics and Speech Processing",
                "authors": [],
                "year": null,
                "venue": "",
                "volume": "",
                "issue": "",
                "pages": "",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Proceedings of the Twenty-Fifth Conference on Computational Linguistics and Speech Processing (ROCLING 2013)",
                "links": null
            },
            "BIBREF17": {
                "ref_id": "b17",
                "title": "Exploiting principal component analysis in modulation spectrum enhancement for robust speech recognition",
                "authors": [
                    {
                        "first": "Jan-Yee",
                        "middle": [],
                        "last": "Lee",
                        "suffix": ""
                    },
                    {
                        "first": "Jeih-Weih",
                        "middle": [],
                        "last": "Hung",
                        "suffix": ""
                    }
                ],
                "year": 2011,
                "venue": "Fuzzy Systems and Knowledge Discovery (FSKD)",
                "volume": "3",
                "issue": "",
                "pages": "",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Jan-Yee Lee, Jeih-weih Hung, \"Exploiting principal component analysis in modulation spectrum enhancement for robust speech recognition,\" Fuzzy Systems and Knowledge Discovery (FSKD), vol 3, pp -1951, 2011.",
                "links": null
            },
            "BIBREF18": {
                "ref_id": "b18",
                "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
            },
            "BIBREF19": {
                "ref_id": "b19",
                "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
            },
            "BIBREF20": {
                "ref_id": "b20",
                "title": "Proceedings of the Twenty-Fifth Conference on Computational Linguistics and Speech Processing",
                "authors": [],
                "year": null,
                "venue": "",
                "volume": "",
                "issue": "",
                "pages": "",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "The hidden Markov model toolkit (HTK): http://htk.eng.cam.ac.uk Proceedings of the Twenty-Fifth Conference on Computational Linguistics and Speech Processing (ROCLING 2013)",
                "links": null
            }
        },
        "ref_entries": {
            "TABREF1": {
                "num": null,
                "content": "<table><tr><td>W, H</td><td>=</td><td>W,H \uf025 \uf025</td><td>i \u00e5 m</td><td>i V</td><td>m</td><td>-</td><td>WH \uf025 \uf025</td><td>i</td><td>m</td><td>(2)</td></tr><tr><td colspan=\"10\">(\u4e8c)NMF \u4f7f\u7528\u65bc\u5168\u983b\u5e36\u8abf\u8b8a\u983b\u8b5c\u4e4b\u8fed\u4ee3\u5f0f\u66f4\u65b0</td><td/></tr><tr><td colspan=\"11\">\u7121\u8ad6\u662f\u6b64\u5c0f\u7bc0\u5c07\u8981\u4ecb\u7d39\u7684\u8fed\u4ee3\u5f0f\u66f4\u65b0\u6cd5\u6216\u662f\u4e0b\u4e00\u5c0f\u7bc0\u7684\u6295\u5f71\u6cd5\uff0c\u6240\u63a1\u7528\u7684\u7686\u662f\u8207\u58d3</td></tr><tr><td colspan=\"11\">\u7e2e\u611f\u77e5(compressed sensing)[16]\u76f8\u540c\u7684\u6982\u5ff5\uff0c\u7c21\u55ae\u4f86\u8aaa\uff0c\u6240\u8b02\u7684\u58d3\u7e2e\u611f\u77e5\u4e26\u4e0d\u76f4\u63a5\u5c0d\u8a0a</td></tr><tr><td colspan=\"11\">\u865f\u76f4\u63a5\u505a\u63a1\u96c6\uff0c\u800c\u662f\u7d93\u7531\u5c07\u4fe1\u865f\u6295\u5f71\u81f3\u4e00\u7d44\u6ce2\u5f62\u4e0a\uff0c\u5f97\u5230\u4e00\u7d44\u58d3\u7e2e\u6578\u64da\u5f8c\uff0c\u518d\u85c9\u7531\u6700\u4f73</td></tr><tr><td colspan=\"10\">\u5316\u7684\u65b9\u5f0f\u9032\u884c\u89e3\u78bc\uff0c\u9032\u800c\u4f30\u8a08\u51fa\u539f\u59cb\u8a0a\u865f\u7684\u91cd\u8981\u8a0a\u606f\u3002</td><td/></tr><tr><td colspan=\"11\">\u5728\u6587\u737b[18]\u4e2d\uff0c\u9996\u5148\u63d0\u53ca\u5229\u7528NMF\u65bc\u8a9e\u97f3\u5012\u983b\u8b5c\u7279\u5fb5\u8abf\u8b8a\u983b\u8b5c\u7684\u66f4\u65b0\u4e0a\u3002\u5728\u6b64\uff0c\u6211</td></tr><tr><td colspan=\"5\">\u5011\u7c21\u8981\u5730\u4ecb\u7d39\u5176\u66f4\u65b0\u6b65\u9a5f\uff1a</td><td/><td/><td/><td/><td/><td/></tr><tr><td colspan=\"11\">\u6b65\u9a5f I. \u5c0d\u65bc\u7279\u5b9a\u9805\u4e4b\u8a9e\u97f3\u7279\u5fb5(\u5982\u7b2c\u4e00\u7dad\u5012\u983b\u8b5c\u7279\u5fb5)\u800c\u8a00\uff0c\u5c07\u7528\u4ee5\u8a13\u7df4\u8072\u5b78\u6a21\u578b\u7684\u6bcf\u4e00</td></tr><tr><td colspan=\"11\">\u53e5\u4e7e\u6de8\u8a9e\u97f3\u7279\u5fb5\u6642\u9593\u5e8f\u5217\u4f5c\u4f5c\u96e2\u6563\u5085\u7acb\u8449\u8f49\u63db(discrete Fourier transform, DFT)\uff0c\u5f97\u5230\u5176</td></tr><tr><td colspan=\"11\">\u8abf\u8b8a\u983b\u8b5c\u5e8f\u5217\uff0c\u5c07\u9019\u4e9b\u4e0d\u540c\u8a9e\u53e5\u6240\u5c0d\u61c9\u4e4b\u8abf\u8b8a\u983b\u8b5c\u5e8f\u5217\u7684\u5f37\u5ea6(magnitude)\u6392\u6210\u4e00\u500b\u77e9\u9663</td></tr><tr><td colspan=\"7\">V \u7684\u6bcf\u4e00\u884c(column)\uff0c\u56e0\u6b64\u82e5 V</td><td/><td/><td/><td/></tr></table>",
                "text": "\u7684\u5c3a\u5bf8\u70ba N\u00d7M\uff0c\u4ee3\u8868\u4e86\u6211\u5011\u5171\u6709 M \u53e5\u8a9e\u97f3\uff0c\u800c\u5176\u983b\u7387 \u9ede\u6578\u70ba N\u3002 \u6b65\u9a5f II. \u5229\u7528\u524d\u8ff0\u4e4b NMF \u6cd5\u5206\u89e3\u77e9\u9663 V \uff0c\u5373\u6c42\u53d6\u7b49\u5f0f(1)\u4e2d\u7684\u5169\u500b\u77e9\u9663 W \u8207 H \u3002\u5176\u4e2d\u77e9 \u9663 W \u5305\u542b\u4e86 r \u500b\u5c3a\u5bf8\u70ba N\u00d71 \u7684\u884c\u5411\u91cf(column vector)\uff0c\u9019 r \u500b\u884c\u5411\u91cf\u5305\u542b\u4e86\u6bcf\u4e00\u53e5\u4e7e\u6de8 \u8a9e\u97f3\u8abf\u8b8a\u983b\u8b5c\u5f37\u5ea6\u4e4b\u57fa\u5e95\u5411\u91cf\uff0c\u800c H \u5247\u662f\u4ee3\u8868\u4e86\u6bcf\u4e00\u53e5\u4e7e\u6de8\u8a9e\u97f3\u7684\u6b0a\u91cd\u3002 \u6b65\u9a5f III. \u5c0d\u65bc\u8a13\u7df4\u8207\u6e2c\u8a66\u7684\u8a9e\u53e5\u5176\u7279\u5fb5\u6642\u9593\u5e8f\u5217\u7684\u8abf\u8b8a\u983b\u8b5c\u5f37\u5ea6(\u4ee5\u5411\u91cf",
                "type_str": "table",
                "html": null
            },
            "TABREF2": {
                "num": null,
                "content": "<table><tr><td>\u5716\u4e8c \u57fa\u5e95\u77e9\u9663 W \u4e4b\u8abf\u8b8a\u983b\u8b5c\u5f37\u5ea6\u5206\u5e03\u5716 (\u56db)\u8fed\u4ee3\u6cd5\u53ca\u6295\u5f71\u6cd5\u4e4b\u521d\u6b65\u6548\u80fd\u8a0e\u8ad6 \u6b64\u5c0f\u7bc0\u4e3b\u8981\u5728\u65bc\u6bd4\u8f03\u8fed\u4ee3\u6cd5\u8207\u6295\u5f71\u6cd5\u7684\u8abf\u8b8a\u983b\u8b5c\u5931\u771f\u6539\u5584\u7a0b\u5ea6\uff0c\u85c9\u7531\u529f\u7387\u983b\u8b5c\u5bc6\u5ea6 (power spectral density, PSD)\u5716\u4f86\u8a55\u4f30\u9019\u5169\u500b\u65b9\u6cd5\u7684\u6548\u80fd\u3002\u5728\u6b64\u6211\u5011\u63a1\u7528 AURORA 2.0[19] \u8cc7\u6599\u5eab\u4e2d\u7684 MAH_27O6571A \u8a9e\u97f3\u6a94\uff0c\u52a0\u4e0a\u4e0d\u540c\u8a0a\u96dc\u6bd4(SNR)\u7684\u5730\u4e0b\u9435 (subway) \u96dc\u8a0a\uff0c \u4f7f\u7528\u7684 NMF \u6cd5\uff0c\u5176\u53c3\u6578 r \u8a2d\u70ba 10\u3002\u5716\u4e09\u7684(a)(b)(c)(d)\u5206\u5225\u4ee3\u8868\u7d93\u904e\u5168\u983b\u5e36\u8fed\u4ee3\u6cd5\u3001\u5168 \u983b\u5e36\u6295\u5f71\u6cd5\u3001\u4f4e\u983b\u8fed\u4ee3\u6cd5\u53ca\u4f4e\u983b\u6295\u5f71\u6cd5\u8655\u7406\u5f8c\u7684\u7b2c 1 \u7dad\u7279\u5fb5\u5e8f\u5217\u7684\u529f\u7387\u983b\u8b5c\u5bc6\u5ea6\u5716\u3002 \u6839\u64da\u56db\u500b\u529f\u7387\u8abf\u8b8a\u983b\u8b5c\u5bc6\u5ea6\u5716\u7684\u7d50\u679c\uff0c\u6211\u5011\u53ef\u767c\u73fe\u56db\u7a2e\u7d50\u679c\u90fd\u8868\u73fe\u51fa\u76f8\u7576\u597d\u7684\u5931\u771f\u6539\u5584 \u6027\u80fd\uff0c\u7121\u8ad6\u662f\u7d93\u904e\u8fed\u4ee3\u6cd5\u6216\u6295\u5f71\u6cd5\u8655\u7406\u904e\u5f8c\u7684\u529f\u7387\u983b\u8b5c\u5bc6\u5ea6\u5716\u90fd\u660e\u986f\u96c6\u4e2d\u65bc\u8abf\u8b8a\u983b\u7387\u7bc4 \u570d[0, 25Hz]\u4e4b\u9593\uff0c\u56e0\u6b64\uff0c\u91dd\u5c0d\u6b64\u6bb5\u8abf\u8b8a\u983b\u7387\u7bc4\u570d\u9032\u884c\u8655\u7406\u7684\u6548\u679c\u53ef\u76f8\u7576\u63a5\u8fd1\u65bc\u5168\u983b\u5e36\u8655 \u7406\uff0c\u5728\u4e4b\u5f8c\u7684\u8fa8\u8b58\u7387\u5be6\u9a57\u4e2d\u4e5f\u5c07\u8b49\u660e\u6b64\u9ede\u3002\u6b64\u5916\uff0c\u5728\u5716 3.1(c)(d)\u4e2d\u53ef\u89c0\u5bdf\u5230\u7d93\u904e\u534a\u983b\u8655 \u7406\u5f8c\u7684\u5169\u7a2e\u65b9\u6cd5\u5728 25Hz \u524d\u7684\u5931\u771f\u6539\u5584\u7a0b\u5ea6\u7686\u76f8\u7576\u826f\u597d\u3002 (a) (b) (c) (d) \u5716\u4e09 \u5404\u7a2e\u65b9\u6cd5\u4f5c\u7528\u65bc\u4e0d\u540c\u8a0a\u96dc\u6bd4\u4e0b\u7684 c1 \u7279\u5fb5\u5e8f\u5217\u4e4b\u529f\u7387\u983b\u8b5c\u5bc6\u5ea6\u5716\uff1a(a)\u5168\u983b\u5e36\u8fed\u4ee3\u6cd5 (b)\u5168\u983b\u5e36\u6295\u5f71\u6cd5 (c)\u4f4e\u983b\u5e36\u8fed\u4ee3\u6cd5 (d)\u4f4e\u983b\u5e36\u6295\u5f71\u6cd5 \u6bcf\u500b\u72c0\u614b\u7531 3 \u500b\u9ad8\u65af\u6df7\u5408\u51fd\u6578\u7d44\u6210\u3002 \u56db\u3001\u5be6\u9a57\u6578\u64da\u8207\u8a0e\u8ad6 \u672c\u7bc0\u5c07\u7531\u56db\u90e8\u5206\u6240\u7d44\u6210\u3002 (\u4e00) \u57fa\u65bc\u8fed\u4ee3\u65b9\u5f0f\u4e4b NMF \u983b\u8b5c\u5f37\u5ea6\u66f4\u65b0\u6cd5\u5176\u8fa8\u8b58\u7387\u5be6\u9a57\u7d50\u679c\u8207\u8a0e\u8ad6 \u5728\u672c\u7bc0\u4e2d\uff0c\u6211\u5011\u5217\u51fa\u5169\u7a2e\u57fa\u65bc\u8fed\u4ee3\u65b9\u5f0f\u4e4b NMF \u983b\u8b5c\u5f37\u5ea6\u66f4\u65b0\u6cd5\uff1aNMF (i,f) (\u5168\u983b\u5e36\u66f4\u65b0) \u8207 NMF (i,low) (\u534a\u983b\u5e36\u66f4\u65b0)\u6240\u5f97\u7684\u5be6\u9a57\u7d50\u679c\u4e26\u52a0\u4ee5\u8a0e\u8ad6\uff0c\u6211\u5011\u8b8a\u5316\u5f0f(1)\u6240\u8a2d\u5b9a\u7684\u77e9\u9663\u884c \u5411\u91cf\u6578 r\uff0c\u4f86\u89c0\u5bdf\u5176\u5e36\u4f86\u7684\u5f71\u97ff\u3002\u70ba\u4e86\u6bd4\u8f03\u65b9\u4fbf\u8d77\u898b\uff0c\u6211\u5011\u5148\u65bc\u8868\u4e00\u8207\u8868\u4e8c\u5206\u5225\u5217\u51fa\u5404 \u7a2e\u65b9\u6cd5\u65bc\u300c\u540c\u4e00\u7d44\u5225\u3001\u4e0d\u540c\u8a0a\u96dc\u6bd4\u300d\u8207\u300c\u540c\u4e00\u8a0a\u96dc\u6bd4\u3001\u4e0d\u540c\u7d44\u5225\u300d\u4e4b\u5e73\u5747\u8fa8\u8b58\u7387\uff0c\u4e14\u5148 \u56fa\u5b9a\u53c3\u6578 r=10\uff0c\u63a5\u8457\u8b8a\u5316\u53c3\u6578 r=5, 10, 15\uff0c\u89c0\u5bdf NMF (i,f) (\u5168\u983b\u5e36\u66f4\u65b0) \u8207 NMF (i,low) (\u534a \u983b\u5e36\u66f4\u65b0)\u5169\u6cd5\u7684\u7e3d\u5e73\u5747\u8fa8\u8b58\u7387\uff0c\u5448\u73fe\u65bc\u8868\u4e09\u4e2d\u3002 \u4e0d\u540c\u7d44\u5225\u4e4b\u4e0b\u3001\u53d6 5 \u7a2e\u8a0a\u96dc\u6bd4(20 dB, 15 dB, 10 dB, 5 dB \u8207 0 dB)\u4e4b\u8fa8\u8b58\u7387(%)\u5e73\u5747\u6bd4\u8f03 Set A Set B Set C Avg RR MVN 73.81 75.02 75.08 74.55 36.76 NMF (i,f) 82.65 83.94 81.75 82.99 57.73 NMF (i,low) 82.80 84.08 81.89 83.13 58.08 \u8868\u4e8c\u3001\u4f7f\u7528 10 \u500b\u57fa\u5e95\u983b\u8b5c\u5411\u91cf(r=10)\u6642\uff0c \u539f\u59cb NMF (i,f) \u6cd5\u8207\u65b0\u63d0\u51fa\u4e4b NMF (i,low) \u6cd5\u5728\u4e0d \u540c\u7d44\u5225\u4e4b\u4e0b\u3001\u53d6 5 \u7a2e\u8a0a\u96dc\u6bd4(20 dB, 15 dB, 10 dB, 5 dB \u8207 0 dB)\u4e4b\u8fa8\u8b58\u7387(%)\u5e73\u5747\u6bd4\u8f03 Clean 20 dB 15 dB 10 dB 5 dB 0 dB -5 dB MVN 99.08 96.58 93.07 84.56 65.24 33.73 13.39 NMF (i,f) 98.71 96.57 94.48 89.52 78.23 55.11 26.64 NMF (i,low) 98.76 96.74 94.65 89.74 78.61 54.87 25.60 \u8868\u4e09\u3001\u4f7f\u7528\u4e09\u7a2e\u57fa\u5e95\u983b\u8b5c\u5411\u91cf\u6578\u7684\u8a2d\u5b9a(r=5, 10, 15)\u6642\uff0c \u7e3d\u5e73\u5747\u8fa8\u8b58\u7387(%)\u4e4b\u6bd4\u8f03 r = 5 r = 10 r = 15 NMF (i,f) 83.36 82.99 82.87 NMF (i,low) 83.19 83.13 83.09 (\u4e8c) \u57fa\u65bc\u6b63\u4ea4\u6295\u5f71\u65b9\u5f0f\u4e4b NMF \u983b\u8b5c\u5f37\u5ea6\u66f4\u65b0\u6cd5\u5176\u8fa8\u8b58\u7387\u5be6\u9a57\u7d50\u679c\u8207\u8a0e\u8ad6 \u5728\u672c\u7bc0\u4e2d\uff0c\u6211\u5011\u5217\u51fa\u4e86\u6240\u65b0\u63d0\u51fa\u4e4b\u5169\u7a2e\u57fa\u65bc\u6b63\u4ea4\u6295\u5f71\u65b9\u5f0f\u4e4b NMF \u983b\u8b5c\u5f37\u5ea6\u66f4\u65b0\u6cd5\uff1a NMF (p,f) (\u5168\u983b\u5e36\u66f4\u65b0) \u8207 NMF (p,low) (\u534a\u983b\u5e36\u66f4\u65b0)\u6240\u5f97\u7684\u5be6\u9a57\u7d50\u679c\u4e26\u52a0\u4ee5\u8a0e\u8ad6\uff0c\u985e\u4f3c\u524d \u4e00\u7bc0\uff0c\u6211\u5011\u8b8a\u5316\u5f0f(1)\u6240\u8a2d\u5b9a\u7684\u77e9\u9663\u884c\u5411\u91cf\u6578 r\uff0c\u4f86\u89c0\u5bdf\u5176\u5e36\u4f86\u7684\u5f71\u97ff\u3002\u4e14\u70ba\u4e86\u65b9\u4fbf\u8d77\u898b\uff0c \u6211\u5011\u5148\u65bc\u8868\u56db\u8207\u8868\u4e94\u4e2d\u5206\u5225\u5217\u51fa\u5404\u7a2e\u65b9\u6cd5\u65bc\u300c\u540c\u4e00\u7d44\u5225\u3001\u4e0d\u540c\u8a0a\u96dc\u6bd4\u300d\u8207\u300c\u540c\u4e00\u8a0a\u96dc\u6bd4\u3001 \u4e0d\u540c\u7d44\u5225\u300d\u4e4b\u5e73\u5747\u8fa8\u8b58\u7387\uff0c\u4e14\u5148\u56fa\u5b9a\u53c3\u6578 r=10\uff0c\u63a5\u8457\u8b8a\u5316\u53c3\u6578 r=5. 10, 15\uff0c\u89c0\u5bdf NMF (p,f) (\u5168\u983b\u5e36\u66f4\u65b0) \u8207 NMF (p,low) (\u534a\u983b\u5e36\u66f4\u65b0)\u5169\u6cd5\u7684\u7e3d\u5e73\u5747\u8fa8\u8b58\u7387\uff0c\u5448\u73fe\u65bc\u8868\u516d\u4e4b\u4e2d\u3002\u5f9e\u9019 \u5e7e\u500b\u8868\u4e2d\uff0c\u6211\u5011\u5f97\u5230\u4e86\u8207\u4e0a\u4e00\u7bc0\u5341\u5206\u985e\u4f3c\u7684\u7d50\u8ad6\uff0c\u4ea6\u5373\uff1a Set B Set C Avg RR MVN 73.81 75.02 75.08 74.55 36.76 NMF (p,f) 82.66 83.82 81.63 82.92 57.56 NMF (p,low) 82.65 83.97 81.90 83.03 57.84 \u8868\u4e94\u3001\u4f7f\u7528 10 \u500b\u57fa\u5e95\u983b\u8b5c\u5411\u91cf(r=10)\u6642\uff0c \u539f\u59cb NMF (p,f) \u6cd5\u8207\u65b0\u63d0\u51fa\u4e4b NMF (p,low) \u6cd5\u5728\u4e0d \u540c\u7d44\u5225\u4e4b\u4e0b\u3001\u53d6 5 \u7a2e\u8a0a\u96dc\u6bd4(20 dB, 15 dB, 10 dB, 5 dB \u8207 0 dB)\u4e4b\u8fa8\u8b58\u7387\u5e73\u5747\u6bd4\u8f03 Clean 20 dB 15 dB 10 dB 5 dB 0 dB -5 dB MVN 99.08 96.58 93.07 84.56 65.24 33.73 13.39 NMF (p,f) 98.73 96.61 94.48 89.53 78.20 54.70 25.92 NMF (p,low) 98.74 96.65 94.57 89.72 78.55 54.72 25.07 \u8868\u516d\u3001\u4f7f\u7528\u4e09\u7a2e\u57fa\u5e95\u983b\u8b5c\u5411\u91cf\u6578\u7684\u8a2d\u5b9a(r=5, 10, 15)\u6642\uff0c \u7e3d\u5e73\u5747\u8fa8\u8b58\u7387\u4e4b\u6bd4\u8f03 r = 5 r = 10 r = 15 NMF (p,f) 83.33 82.92 83.20 NMF (p,low) 83.34 83.03 82.97 (\u4e09)\u5404\u7a2e\u5f37\u5065\u6027\u65b9\u6cd5\u4e4b\u6548\u80fd\u6bd4\u8f03 \u63a5\u4e0b\u4f86\u6211\u5011\u5c07\u5f59\u6574\u524d\u8ff0\u4e0d\u540c\u7684\u8abf\u8b8a\u983b\u8b5c\u66f4\u65b0\u65b9\u6cd5\u4e4b\u5be6\u9a57\u7d50\u679c\uff0c\u9664\u4e86\u6211\u5011\u5148\u524d\u63d0\u53ca\u7684\u56db\u7a2e \u57fa\u65bc NMF \u7684\u8abf\u8b8a\u983b\u8b5c\u66f4\u65b0\u6cd5\uff1aNMF (i, f) \u6b64\u5916\uff0c\u6211\u5011\u4e5f\u9032\u4e00\u6b65\u53bb\u5206\u6790\uff0c\u5728 NMF \u6cd5\u8207 PCA \u6cd5\u4e2d\uff0c\u50c5\u66f4\u65b0\u6574\u6bb5\u983b\u5e36\u4e4b\u524d 1/3(\u7d04\u524d 170 \u500b\u983b\u7387\u9ede)\u8207\u524d 1/4(\u524d 128 \u500b\u983b\u7387\u9ede)\u4e4b\u983b\u8b5c\u5f37\u5ea6\u5c0d\u65bc\u8fa8\u8b58\u7387\u4e4b\u5f71\u97ff\u70ba\u4f55\u3002\u9019\u4e9b \u5be6\u9a57\u6578\u64da\u7686\u5f59\u6574\u65bc\u8868\u4e03\u3002\u85c9\u7531\u8868\u4e03\u53ef\u4ee5\u89c0\u5bdf\u5230\u4e0b\u5217\u5e7e\u9ede\uff1a 1. Proceedings Set A 1.</td></tr></table>",
                "text": "Proceedings of the Twenty-Fifth Conference on Computational Linguistics and Speech Processing (ROCLING 2013) Proceedings of the Twenty-Fifth Conference on Computational Linguistics and Speech Processing (ROCLING 2013) MIRS \u5169\u7a2e\u901a\u9053\u6a19\u6e96\uff0c\u7531\u570b\u5bb6\u96fb\u4fe1\u806f\u76df (international telecommunication Union, ITU)[20]\u6240\u8a02\u5b9a\u800c\u6210\u3002 \u4e0a\u8ff0\u4e4b\u7528\u4ee5\u8a13\u7df4\u8207\u6e2c\u8a66\u7684\u8a9e\u53e5\uff0c\u6211\u5011\u5148\u5c07\u5176\u8f49\u63db\u6210\u6885\u723e\u5012\u983b\u8b5c\u7279\u5fb5\u53c3\u6578(mel-frequency cepstral coefficients, MFCC)\uff0c\u4f5c\u70ba\u4e4b\u5f8c\u5404\u7a2e\u5f37\u5065\u6027\u65b9\u6cd5\u7684\u57fa\u790e\u7279\u5fb5(baseline feature)\uff0c\u5efa \u69cb MFCC \u7279\u5fb5\u7684\u904e\u7a0b\u4e3b\u8981\u662f\u6839\u64da AURORA 2.0[19]\u8cc7\u6599\u5eab\u4e2d\u7684\u8a2d\u5b9a\uff0c\u6700\u7d42 MFCC \u7279\u5fb5 \u5305\u542b\u4e86 13 \u7dad\u7684\u975c\u614b\u7279\u5fb5(static features)\u9644\u52a0\u4e0a\u5176\u4e00\u968e\u5dee\u5206\u8207\u4e8c\u968e\u5dee\u5206\u7684\u52d5\u614b\u7279\u5fb5 (dynamic features)\uff0c\u5171 39 \u7dad\u7279\u5fb5\u3002\u503c\u5f97\u4e00\u63d0\u7684\u662f\uff0c\u672c\u8ad6\u6587\u4e4b\u5f8c\u6240\u63d0\u7684\u5f37\u5065\u6027\u6280\u8853\uff0c\u7686\u662f \u4f5c\u7528\u65bc 13 \u7dad\u7684\u975c\u614b\u7279\u5fb5\u4e0a\uff0c\u518d\u7531\u66f4\u65b0\u5f8c\u7684\u975c\u614b\u7279\u5fb5\u6c42\u53d6 26 \u7dad\u7684\u52d5\u614b\u7279\u5fb5\u3002\u540c\u6642\uff0c\u70ba\u4e86 \u521d\u6b65\u964d\u4f4e\u96dc\u8a0a\u5c0d\u65bc\u8a9e\u97f3\u7279\u5fb5\u7684\u5e72\u64fe\uff0c\u6211\u5011\u5148\u4ee5\u7c21\u6613\u4f46\u6548\u679c\u5353\u8457\u7684 MVN \u6cd5\u8655\u7406 MFCC \u7279\u5fb5\uff0c\u4e4b\u5f8c\u518d\u914d\u642d\u6240\u65b0\u63d0\u51fa\u7684\u5404\u7a2e\u57fa\u65bc NMF \u7684\u6f14\u7b97\u6cd5\uff0c\u85c9\u6b64\u5f97\u5230\u66f4\u4f73\u7684\u8fa8\u8b58\u7cbe\u78ba\u5ea6\u3002 \u7576\u8b8a\u5316\u57fa\u5e95\u983b\u8b5c\u5411\u91cf\u6578 r \u6642\uff0c\u53ef\u767c\u73fe\u5c31\u7e3d\u5e73\u5747\u8fa8\u8b58\u7387\u800c\u8a00\uff0cr \u503c\u7684\u4e09\u7a2e\u8a2d\u5b9a(5, 10, 15) \u6240\u5c0d\u61c9\u7684\u5169\u7a2e NMF \u6cd5\u6548\u679c\u4e5f\u662f\u5341\u5206\u63a5\u8fd1\uff0c\u8f03\u5c0f\u7684 r \u503c\u5c0d\u61c9\u4e4b\u5e73\u5747\u8fa8\u8b58\u7387\u751a\u81f3\u662f\u8f03\u597d\u7684\uff0c \u9019\u986f\u793a\u4e86\u6211\u5011\u53ef\u4ee5\u4f7f\u7528\u5c11\u91cf\u7684\u57fa\u5e95\u983b\u8b5c\uff0c\u5c31\u53ef\u4f7f\u9019\u5169\u7a2e NMF \u6cd5\u767c\u63ee\u5176\u8fd1\u4e4e\u6700\u4f73\u7684\u6548\u679c\u3002 \u8868\u4e00\u3001\u4f7f\u7528 10 \u500b\u57fa\u5e95\u983b\u8b5c\u5411\u91cf(r=10)\u6642\uff0c \u539f\u59cb NMF(i,f)\u6cd5\u8207\u65b0\u63d0\u51fa\u4e4b NMF(i,low)\u6cd5\u5728 \u5229\u7528\u6b63\u4ea4\u6295\u5f71\u65b9\u5f0f\u4e4b\u5168\u983b\u8655\u7406\u6cd5 NMF (p,f) \u8207\u534a\u983b\u8655\u7406\u6cd5 NMF (p,low) \uff0c\u7121\u8ad6\u65bc\u4e0d\u540c\u985e\u5225 \u6216\u662f\u4e0d\u540c\u7a0b\u5ea6\u7684\u96dc\u8a0a\u74b0\u5883\uff0c\u5c0d\u65bc MVN \u9810\u8655\u7406\u4e4b MFCC \u8a9e\u97f3\u7279\u5fb5\u90fd\u6709\u5341\u5206\u63a5\u8fd1\u7684\u8fa8\u8b58 \u7387\u6539\u5584\u6548\u80fd\u3002 of the Twenty-Fifth Conference on Computational Linguistics and Speech Processing (ROCLING 2013) \u8868\u56db\u3001\u4f7f\u7528 10 \u500b\u57fa\u5e95\u983b\u8b5c\u5411\u91cf(r=10)\u6642\uff0c \u539f\u59cb NMF (p,f) \u6cd5\u8207\u65b0\u63d0\u51fa\u4e4b NMF (p,low) \u6cd5\u5728\u4e0d \u540c\u7d44\u5225\u4e4b\u4e0b\u3001\u53d6 5 \u7a2e\u8a0a\u96dc\u6bd4(20 dB, 15 dB, 10 dB, 5 dB \u8207 0 dB)\u4e4b\u8fa8\u8b58\u7387\u5e73\u5747\u6bd4\u8f03 \u3001NMF (i, low) \u3001NMF (p, f) \u8207 NMF (i, low) \u4e4b\u5916\uff0c\u5728\u9019\u88e1 \u6211\u5011\u4e5f\u540c\u6642\u4f5c\u4e86\u5169\u500b\u7279\u5fb5\u6642\u9593\u5e8f\u5217\u66f4\u65b0\u6cd5\uff1aHEQ \u8207 MVA\uff0c\u53ca\u57fa\u65bc PCA \u7684\u8abf\u8b8a\u983b\u8b5c\u66f4 \u65b0\u6cd5\uff0c\u70ba\u4e86\u7cbe\u7c21\u6bd4\u8f03\u8d77\u898b\uff0c\u5728\u5404\u7a2e NMF \u6cd5\u8207 PCA \u6cd5\u4e2d\uff0c\u57fa\u5e95\u5411\u91cf\u6578\u76ee r \u56fa\u5b9a\u70ba 10\u3002 MVN \u6cd5\u5728\u8655\u7406 MFCC \u7279\u5fb5\u4e0a\uff0c\u53ef\u4f7f\u5e73\u5747\u8fa8\u8b58\u7387\u5f9e 59.75%\u5927\u5e45\u63d0\u6607\u81f3 74.55%\uff0c\u800c HEQ \u53ca MVA \u66f4\u53ef\u5206\u5225\u63d0\u4f9b 22.46%\u53ca 19.00%\u7684\u7d55\u5c0d\u932f\u8aa4\u7387\u6539\u5584\u3002 2. PCA \u6cd5\u7121\u8ad6\u8655\u7406\u5168\u983b\u5e36\u6216\u4f4e\u983b\u5e36\uff0c\u90fd\u80fd\u6709\u5341\u5206\u986f\u8457\u7684\u8fa8\u8b58\u7387\u6539\u9032\uff0c\u4e14\u8ddf NMF \u7684\u6548\u80fd \u5e7e\u4e4e\u4e0d\u76f8\u4e0a\u4e0b\uff0c\u751a\u81f3\u65bc\u66f4\u9ad8\uff0c\u6b64\u539f\u56e0\u6975\u53ef\u80fd\u662f\u96d6\u7136 PCA \u4e26\u672a\u52a0\u5165\u300c\u6240\u6c42\u5f97\u4e4b\u8abf\u8b8a\u983b \u8b5c\u5f37\u5ea6\u5fc5\u4e0d\u70ba\u8ca0\u300d\u7684\u9650\u5236\uff0c\u4f46\u5176\u5be6\u5f97\u5230\u7684\u983b\u8b5c\u5f37\u5ea6\u4ecd\u7136\u7d55\u5927\u90e8\u5206\u90fd\u662f\u5927\u65bc\u6216\u7b49\u65bc\u96f6\uff0c \u76f8\u4f4d\u5931\u771f\u7684\u53ef\u80fd\u6027\u6975\u5c0f\uff0c\u6240\u4ee5\u6548\u80fd\u512a\u8d8a\u3002\u9019\u5c07\u5728\u4e4b\u5f8c\u7684\u7ae0\u7bc0\u4f5c\u8aaa\u660e\u3002 3. HEQ \u8207 MVA \u76f8\u5c0d\u65bc MVN \u6cd5\u80fd\u5920\u5e36\u4f86\u66f4\u4f73\u7684\u8fa8\u8b58\u7387\uff0c\u4f46\u672c\u8ad6\u6587\u6240\u8a0e\u8ad6\u4e4b\u56db\u7a2e NMF \u983b\u8b5c\u66f4\u65b0\u6cd5\u5247\u660e\u986f\u90fd\u512a\u65bc HEQ \u8207 MVA\u3002 4. \u4e4b\u524d\u6240\u8a0e\u8ad6\u7684\u5169\u7a2e\u4f4e\u983b\u5e36\u66f4\u65b0\u6cd5 NMF",
                "type_str": "table",
                "html": null
            },
            "TABREF4": {
                "num": null,
                "content": "<table><tr><td/><td/><td/><td/><td/><td>low)</td></tr><tr><td colspan=\"6\">\u5c0d\u65bc\u55ae\u4e00\u8abf\u8b8a\u983b\u8b5c\u5411\u91cf\u4e4b\u66f4\u65b0\u6240\u9700\u7684\u7684\u4e58\u6cd5\u904b\u7b97\u6578\u76ee\u53ca\u904b\u4f5c\u65bc MATLAB \u7a0b\u5f0f\u6240\u9700\u7684\u6642 \u4f46\u6211\u5011\u767c\u73fe\uff0c\u7531\u65bc\u4e0a\u8ff0\u5169\u7a2e\u65b9\u6cd5\u5f97\u5230\u8ca0\u503c\u983b\u8b5c\u5f37\u5ea6\u7684\u6a5f\u7387\u76f8\u7576\u4f4e\uff0c\u56e0\u6b64\u53ef\u80fd\u5c0d\u8fa8\u8b58\u6027\u80fd</td></tr><tr><td colspan=\"5\">\u9593\u8a73\u5217\u65bc\u8868\u516b\uff0c\u5728\u6b64\uff0c\u57f7\u884c\u6642\u9593\u6539\u5584\u7387\u7684\u5b9a\u7fa9\u5982\u4e0b\uff1a \u7684\u5f71\u97ff\u4e5f\u5c31\u4e0d\u660e\u986f\u3002\u6b64\u7d50\u679c\u5df2\u5728\u524d\u9762\u7684\u5be6\u9a57\u7ae0\u7bc0\u4e2d\u5448\u73fe\u3002</td></tr><tr><td colspan=\"2\">\u57f7\u884c\u6642\u9593\u6539\u5584\u7387</td><td colspan=\"3\">\u00d7100% \u8868\u516b \u7570\u5e38\u4e4b\u8ca0\u503c\u983b\u8b5c\u5f37\u5ea6\u6578\u91cf\u6bd4\u8f03\u8868</td><td>(7)</td></tr><tr><td>\u65b9\u6cd5</td><td colspan=\"5\">\u8868\u516b \u56db\u7a2e\u57fa\u65bc NMF \u4e4b\u8abf\u8b8a\u983b\u8b5c\u66f4\u65b0\u6cd5\u7684\u8907\u96dc\u5ea6\u6bd4\u8f03 NMF (i,f) NMF (p,f) PCA (p,f)</td></tr><tr><td colspan=\"2\">\u6b63\u503c\u5e73\u5747\u6578\u91cf</td><td/><td>513</td><td>512.53</td><td>512.24</td></tr><tr><td colspan=\"4\">\u4e58\u6cd5\u7e3d\u6578 (\u4ee3\u6578\u8868\u793a\u8207\u53c3\u6578 \u8ca0\u503c\u5e73\u5747\u6578\u91cf 0 \u65b9\u6cd5 \u5e36\u5165\u5f8c\u7684\u5be6\u969b\u503c) NMF (i,f) L(r 2 +2r)+2Nr 22260 \u8ca0\u503c\u7387(\u8ca0\u503c\u5e73\u5747 \u6578\u91cf/\u7e3d\u6578\u91cf) 0%</td><td>\u5728 MATLAB \u57f7\u884c 0.47 \u6240\u9700\u6642\u9593(msec) 5.52 0.0009162%</td><td>\u57f7\u884c\u6642\u9593\u6539\u5584\u7387 0.7596 (%) -0.0015%</td></tr><tr><td>NMF (p,f)</td><td>2Nr</td><td/><td>10260</td><td>3.30</td><td>40.22</td></tr><tr><td colspan=\"3\">NMF (i,low) L(r 2 +2r)+Nr NMF (p,low) Nr \u4e94\u3001\u7d50\u8ad6</td><td>17130 5130</td><td>4.45 1.98</td><td>19.38 64.13</td></tr><tr><td colspan=\"6\">\u672c\u8ad6\u6587\u7684\u91cd\u9ede\u5728\u65bc\u6539\u5584\u539f\u59cb\u57fa\u65bc\u8fed\u4ee3\u65b9\u5f0f\u3001\u975e\u8ca0\u77e9\u9663\u5206\u89e3(NMF)\u4e4b\u5168\u983b\u5e36\u8abf\u8b8a\u983b\u8b5c</td></tr><tr><td colspan=\"6\">\u66f4\u65b0\u6cd5\u4e4b\u8907\u96dc\u5ea6\uff0c\u540c\u6642\u4fdd\u6709\u5176\u5c0d\u8a9e\u97f3\u7279\u5fb5\u7684\u96dc\u8a0a\u5f37\u5065\u5316\u6548\u80fd\uff0c\u6240\u63d0\u51fa\u4e4b\u65b9\u6cd5\u8207\u539f\u59cb\u6cd5\u6700</td></tr><tr><td colspan=\"6\">\u7531\u8868\u516b\u6211\u5011\u53ef\u770b\u51fa\uff0c\u6211\u5011\u63d0\u51fa\u7684\u4e09\u7a2e\u65b0\u65b9\u6cd5\uff1aNMF (p,f) \u3001NMF (i,low) \u8207 NMF (p,low) \uff0c\u76f8 \u5927\u4e0d\u540c\u5728\u65bc\uff0c\u6211\u5011\u63a1\u53d6\u4e00\u6b21\u6027\u7684\u6b63\u4ea4\u6295\u5f71\u65b9\u5f0f\u4f86\u6c42\u53d6\u5728\u57fa\u5e95\u983b\u8b5c\u77e9\u9663\u4e4b\u5c55\u958b\u7a7a\u9593(the</td></tr><tr><td colspan=\"6\">spanned subspace)\u7684\u65b0\u8abf\u8b8a\u983b\u8b5c\u5f37\u5ea6\uff0c\u800c\u539f\u59cb\u7684\u8fed\u4ee3\u65b9\u5f0f\u5247\u662f\u5229\u7528\u9010\u6b21\u903c\u8fd1\u7684\u65b9\u5f0f\u4f86\u6c42 \u5c0d\u65bc\u539f\u59cb\u7684\u8fed\u4ee3\u5f0f\u5168\u983b\u5e36\u8655\u7406\u4e4b NMF (i,f) \u6cd5\u53ef\u5f97\u5230\u8f03\u4f4e\u7684\u904b\u7b97\u8907\u96dc\u5ea6(\u8f03\u5c11\u7684\u4e58\u6cd5\u6578 \u5f97\u57fa\u5e95\u983b\u8b5c\u4e4b\u6b0a\u91cd\u3002\u540c\u6642\uff0c\u6211\u5011\u6839\u64da\u8a9e\u97f3\u8abf\u8b8a\u983b\u8b5c\u5176\u4e3b\u8981\u8fa8\u8b58\u8cc7\u8a0a\u90fd\u96c6\u4e2d\u5728\u4f4e\u983b\u5340\u57df\u7684 \u76ee)\u53ca\u57f7\u884c\u6642\u9593\uff0c\u5176\u4e2d\uff0c\u6295\u5f71\u65b9\u5f0f\u53ef\u6e1b\u5c11 L(r 2 +2r)\u6b21\u7684\u4e58\u6cd5\u904b\u7b97\uff0c\u534a\u983b\u8655\u7406\u5f8c\u5247\u53ef\u6e1b\u5c11 \u77ad\u89e3\uff0c\u63d0\u51fa\u4e86\u55ae\u7368\u66f4\u65b0\u4f4e\u983b\u8abf\u8b8a\u983b\u8b5c\u7684\u6a21\u5f0f\uff0c\u540c\u662f\u63d0\u5347\u8abf\u8b8a\u983b\u8b5c\u66f4\u65b0\u6cd5\u7684\u57f7\u884c\u6548\u7387\u3002\u5c31 Nr \u6b21\u7684\u904b\u7b97\u91cf\u3002\u6b64\u5916\uff0c\u5be6\u969b\u7a0b\u5f0f\u904b\u4f5c\u7684\u6642\u9593\u4e5f\u90fd\u6709\u76f8\u7576\u5927\u7684\u6539\u5584\uff0c\u5176\u4e2d NMF (p,low) \u53ef\u6e1b \u5be6\u969b\u904b\u884c\u65bc MATLAB \u7a0b\u5f0f\u4e2d\u3001\u57f7\u884c\u6642\u9593\u7684\u6539\u5584\u7a0b\u5ea6\u4f86\u770b\uff0c\u6295\u5f71\u6cd5\u53ef\u964d\u4f4e 40.22%\u7684\u57f7\u884c \u5c11\u7d04 64%\u7684\u57f7\u884c\u6642\u9593\u3002 \u6642\u9593\uff0c\u534a\u983b\u8655\u7406\u5247\u53ef\u964d\u4f4e 19.38\uff05\u4ee5\u4e0a\u7684\u6642\u9593\uff1b\u7279\u5225\u7684\u662f\uff0c\u6211\u5011\u767c\u73fe\u5728 Aurora-2 \u8cc7\u6599\u5eab</td></tr><tr><td colspan=\"6\">\u5728\u6211\u5011\u4e4b\u524d\u7684\u8fa8\u8b58\u5be6\u9a57\u88e1\uff0c\u4e3b\u8981\u662f\u6bd4\u8f03\u56db\u7a2e\u57fa\u65bc NMF \u7684\u983b\u8b5c\u5f37\u5ea6\u66f4\u65b0\u6280\u8853\uff0c\u4f46\u56e0\u6211\u5011 \u7684\u8fa8\u8b58\u5be6\u9a57\u4e2d\uff0c\u4e0a\u8ff0\u5169\u7a2e\u6539\u5584\u904b\u7b97\u8907\u96dc\u5ea6\u4e4b\u65b9\u5f0f\u5e7e\u4e4e\u4e0d\u6703\u5f71\u97ff NMF \u8abf\u8b8a\u983b\u8b5c\u66f4\u65b0\u6cd5\u5c0d</td></tr><tr><td colspan=\"6\">\u6240\u63d0\u51fa\u7684\u6b63\u4ea4\u6295\u5f71\u6cd5\u8207\u7dda\u6027\u4ee3\u6578\u7406\u8ad6\u4e2d\u7684 PCA \u6cd5\u5bc6\u5207\u76f8\u95dc\uff0c\u56e0\u6b64\u5be6\u9a57\u88e1\u6211\u5011\u4e5f\u540c\u6642\u6aa2 \u61c9\u4e4b\u8fa8\u8b58\u7cbe\u78ba\u5ea6\uff0c\u4ecd\u80fd\u63d0\u4f9b\u539f\u59cb MVN \u9810\u8655\u7406\u5f8c\u7684 MFCC \u7279\u5fb5\u986f\u8457\u7684\u8fa8\u8b58\u7387\u63d0\u5347\u3002</td></tr><tr><td colspan=\"6\">\u8996\u57fa\u65bc PCA \u7684\u983b\u8b5c\u5f37\u5ea6\u66f4\u65b0\u6280\u8853\u7684\u6548\u679c\uff0c\u7c21\u55ae\u4f86\u8aaa\uff0cPCA \u4ea6\u662f\u5982\u540c NMF \u4e00\u822c\uff0c\u5c0d\u65bc \u5728\u672a\u4f86\u7684\u5c55\u671b\u4e2d\uff0c\u6211\u5011\u53ef\u5c07\u6295\u5f71\u6cd5\u8207\u5176\u4ed6\u7279\u5fb5\u6642\u9593\u5e8f\u5217\u57df\u5f37\u5065\u6280\u8853\u505a\u7d50\u5408\uff0c\u6216\u662f\u85c9 \u5f0f(1)\u7684\u8cc7\u6599\u77e9\u9663 V \u6c42\u53d6\u4e00\u7d44\u57fa\u5e95\u77e9\u9663 W\uff0c\u4e26\u7b26\u5408\u5f0f(2)\u4e4b\u6700\u4f73\u5316\u6e96\u5247\u3002\u4f46\u8ddf NMF \u4e0d\u540c \u7531 NMF \u627e\u51fa\u5404\u7a2e\u96dc\u8a0a\u7684\u57fa\u5e95\uff0c\u518d\u85c9\u7531\u983b\u8b5c\u6d88\u53bb\u6cd5(spectral subtraction, SS)\u6216\u5176\u4ed6\u6d88\u566a\u6cd5</td></tr><tr><td colspan=\"6\">\u7684\u9ede\u5728\u65bc\uff0cPCA \u4e26\u7121\u9650\u5236\u9650\u5236\u77e9\u9663 W \u4e4b\u884c\u5411\u91cf\u5176\u4e2d\u5143\u7d20\u5fc5\u70ba\u975e\u8ca0\u5be6\u6578\uff0c\u7d93\u7531 PCA \u6240\u6c42 \u9054\u5230\u6291\u5236\u96dc\u8a0a\u6216\u63d0\u5347\u8fa8\u8b58\u7387\u7684\u6548\u679c\u3002\u6b64\u5916\uff0c\u4e5f\u53ef\u9032\u4e00\u6b65\u5728\u5176\u4ed6\u8cc7\u6599\u5eab\u4e0a\u8655\u7406(\u5982\u4e2d\u6587\u6578</td></tr><tr><td colspan=\"6\">\u51fa\u4e4b\u57fa\u5e95\u77e9\u9663\uff0c\u5176\u6240\u5305\u542b\u7684\u884c\u5411\u91cf\u6070\u70ba\u8cc7\u6599\u77e9\u9663 V \u6240\u5c0d\u61c9\u4e4b\u5171\u8b8a\u7570\u77e9\u9663(convariance \u5b57\u8a9e\u97f3\u6216\u662f\u66f4\u591a\u5b57\u5f59\u7684\u8cc7\u6599\u5eab)\uff0c\u4f7f\u5176\u5728\u73fe\u5be6\u5c64\u9762\u4e2d\u80fd\u6709\u66f4\u591a\u5be6\u969b\u7684\u61c9\u7528\u3002</td></tr><tr><td colspan=\"6\">matrix)\u5176 r \u500b\u56fa\u6709\u5411\u91cf(eigenvector)\uff0c\u524d\u9019\u4e9b\u56fa\u6709\u5411\u91cf\u5206\u5225\u5c0d\u61c9\u81f3\u5171\u8b8a\u7570\u77e9\u9663\u5f9e\u5927\u81f3\u5c0f\u6392</td></tr><tr><td colspan=\"6\">\u5e8f\u4e4b\u524d r \u500b\u56fa\u6709\u503c(eigenvalue)\u3002\u5728\u7121\u975e\u8ca0\u7684\u9650\u5236\u4e0b\uff0cPCA \u6bd4 NMF \u80fd\u5920\u9054\u5230\u66f4\u5c0f\u7684\u5e73\u65b9</td></tr><tr><td colspan=\"6\">\u8aa4\u5dee\u548c(\u5982\u5f0f(2)\u6240\u793a)\uff0c\u4f46\u6f5b\u5728\u7f3a\u9ede\u662f\uff0cPCA \u6c42\u51fa\u4e4b\u57fa\u5e95\u5411\u91cf\u8207\u4e4b\u5f8c\u4f7f\u7528\u6b63\u4ea4\u6295\u5f71\u65b9</td></tr><tr><td colspan=\"6\">\u5f0f\u6c42\u53d6\u51fa\u7684\u8abf\u8b8a\u983b\u8b5c\u5f37\u5ea6\u53ef\u80fd\u51fa\u73fe\u8ca0\u503c\uff0c\u6b64\u9055\u80cc\u4e86\u983b\u8b5c\u5f37\u5ea6\u5fc5\u70ba\u975e\u8ca0\u503c\u7684\u524d\u63d0\u3002</td></tr><tr><td colspan=\"6\">\u4ee5\u4e0b\uff0c\u6211\u5011\u5c07\u8a0e\u8ad6\u5404\u7a2e\u65b9\u6cd5(\u5305\u542b PCA \u6cd5)\u5728\u66f4\u65b0\u983b\u8b5c\u5f37\u5ea6\u6642\uff0c\u6240\u53ef\u80fd\u7522\u751f\u8ca0\u503c</td></tr></table>",
                "text": "Proceedings of the Twenty-Fifth Conference on Computational Linguistics and Speech Processing (ROCLING 2013)\u7684\u60c5\u5f62\uff0c\u5176\u4e2d\uff0c\u539f\u59cb\u5229\u7528\u8fed\u4ee3\u65b9\u5f0f\u7684 NMF (i,f) \u6cd5\u4e26\u4e0d\u6703\u9020\u6210\u8ca0\u503c\u7684\u7522\u751f\uff0c\u800c\u5229\u7528\u6b63\u4ea4\u6295 \u5f71\u65b9\u5f0f\u7684 NMF (p,f) \u6cd5\u5728\u6c42\u5f97\u6b63\u4ea4\u5316\u77e9\u9663 B \u6642\u53ef\u80fd\u7522\u751f\u8ca0\u503c\uff0c\u6b64\u5916\uff0c\u5982\u524d\u6240\u8ff0\uff0cPCA\u6cd5\u4e5f \u7121\u6cd5\u4fdd\u8b49\u6240\u6c42\u53d6\u51fa\u7684\u983b\u8b5c\u5f37\u5ea6\u5fc5\u4e0d\u5c0f\u65bc\u96f6\u3002\u5728\u6b64\uff0c\u6211\u5011\u7d71\u8a08\u5728 Aurora-2 \u8cc7\u6599\u5eab\u4e4b\u4e09\u500b \u6e2c\u8a66\u96c6\u88e1\uff0c\u85c9\u7531\u4e0d\u540c\u65b9\u6cd5\u5176\u66f4\u65b0\u5f8c\u7684\u983b\u8b5c\u5f37\u5ea6\u4e2d\u6240\u6709\u8ca0\u6578\u7e3d\u6578\uff0c\u4e26\u4e14\u5e73\u5747\u5f8c\u5f97\u5230\u6bcf\u4e00\u53e5 \u4e2d\u6bcf\u4e00\u7dad\u7279\u5fb5\u4e4b\u8ca0\u6578\u5e73\u5747\u6578\u91cf\uff0c\u5217\u65bc\u8868\u516b\u3002\u7531\u8868\u516b\u4e2d\u53ef\u767c\u73fe\u5230\uff0c\u7d93\u7531\u8fed\u4ee3\u516c\u5f0f\u6240\u6c42\u5f97\u4e4b \u983b\u8b5c\u5f37\u5ea6\u7531\u65bc NMF \u6cd5\u5206\u89e3\u975e\u8ca0\u77e9\u9663\u7684\u672c\u8cea\uff0c\u78ba\u5be6\u4e0d\u6703\u7522\u751f\u8ca0\u503c\uff0c\u4f46 NMF \u6295\u5f71\u6cd5\u8207 PCA \u7686\u6709\u5c11\u8a31\u7684\u6a5f\u6703\u5f97\u5230\u8ca0\u503c\u7684\u983b\u8b5c\u5f37\u5ea6\uff0c\u5176\u4e2d PCA \u5f97\u5230\u8ca0\u983b\u8b5c\u5f37\u5ea6\u7684\u6bd4\u7387\u7565\u9ad8\u65bc NMF \u6295\u5f71\u6cd5\u7d04 0.0006%\uff0c\u96d6\u7136\u983b\u8b5c\u5f37\u5ea6\u70ba\u8ca0\u503c\u4e26\u4e0d\u5408\u7406\u3001\u6703\u5f15\u5165\u76f8\u4f4d\u7684\u5931\u771f(\u589e\u52a0\u76f8\u4f4d \u03c0)\uff0c",
                "type_str": "table",
                "html": null
            }
        }
    }
}