{
    "paper_id": "O08-1006",
    "header": {
        "generated_with": "S2ORC 1.0.0",
        "date_generated": "2023-01-19T08:02:32.261342Z"
    },
    "title": "\u4e00\u500b \u4e00\u500b \u4e00\u500b \u4e00\u500b\u7d50\u5408 \u7d50\u5408 \u7d50\u5408 \u7d50\u5408 SVM \u8207 \u8207 \u8207 \u8207 Eigen-MLLR \u65b0 \u65b0 \u65b0 \u65b0\u7684 \u7684 \u7684 \u7684\u591a\u8a9e\u8005 \u591a\u8a9e\u8005 \u591a\u8a9e\u8005 \u591a\u8a9e\u8005\u7dda\u4e0a \u7dda\u4e0a \u7dda\u4e0a \u7dda\u4e0a\u8abf\u9069\u67b6\u69cb \u8abf\u9069\u67b6\u69cb \u8abf\u9069\u67b6\u69cb \u8abf\u9069\u67b6\u69cb\u61c9\u7528\u65bc \u61c9\u7528\u65bc \u61c9\u7528\u65bc \u61c9\u7528\u65bc \u6cdb\u5728 \u6cdb\u5728 \u6cdb\u5728 \u6cdb\u5728\u8a9e\u97f3\u8fa8\u8b58\u7cfb\u7d71 \u8a9e\u97f3\u8fa8\u8b58\u7cfb\u7d71 \u8a9e\u97f3\u8fa8\u8b58\u7cfb\u7d71 \u8a9e\u97f3\u8fa8\u8b58\u7cfb\u7d71 A New On-Line Multi-Speaker Adaptation Architecture Combining SVM with Eigen-MLLR for Ubiquitous Speech Recognition System",
    "authors": [
        {
            "first": "\u65bd\u4f2f\u5b9c",
            "middle": [],
            "last": "Po",
            "suffix": "",
            "affiliation": {},
            "email": ""
        },
        {
            "first": "Yi",
            "middle": [],
            "last": "Shih",
            "suffix": "",
            "affiliation": {},
            "email": ""
        },
        {
            "first": "Yuan-Ning",
            "middle": [],
            "last": "\u6797\u82d1\u5be7",
            "suffix": "",
            "affiliation": {},
            "email": ""
        },
        {
            "first": "",
            "middle": [],
            "last": "Lin",
            "suffix": "",
            "affiliation": {},
            "email": ""
        },
        {
            "first": "Jhing-Fa",
            "middle": [],
            "last": "Wang",
            "suffix": "",
            "affiliation": {},
            "email": "wangjf@mail.ncku.edu.tw"
        }
    ],
    "year": "",
    "venue": null,
    "identifiers": {},
    "abstract": "This work presents a novel architecture using SVM and Eigen-MLLR for rapid on-line multi-speaker adaptation in ubiquitous speech recognition. The recognition performance in speaker independent system is better than in conventional speaker dependence system, and the key point is speaker adaptation techniques. The adaptation approach is on the basis of combine SVM and Eigen-MLLR, generating a classification model and building parameters vector-space for all speakers' individual training data. While in recognition, to find test speaker classification by SVM and look for MLLR parameters matrix correspond to speaker classification, then the MLLR parameters matrix and original acoustic model will integrate into speaker dependent model. Last, we estimate the adapted MLLR transformation matrix set by weighting function with recognition result, the present MLLR matrix, and Eigenspace. The estimate result will be used to update the MLLR matrices in adaptation phase. The experimental results show that the proposed method can improve 5% to 8% speech recognition accuracy with speaker adaptation.",
    "pdf_parse": {
        "paper_id": "O08-1006",
        "_pdf_hash": "",
        "abstract": [
            {
                "text": "This work presents a novel architecture using SVM and Eigen-MLLR for rapid on-line multi-speaker adaptation in ubiquitous speech recognition. The recognition performance in speaker independent system is better than in conventional speaker dependence system, and the key point is speaker adaptation techniques. The adaptation approach is on the basis of combine SVM and Eigen-MLLR, generating a classification model and building parameters vector-space for all speakers' individual training data. While in recognition, to find test speaker classification by SVM and look for MLLR parameters matrix correspond to speaker classification, then the MLLR parameters matrix and original acoustic model will integrate into speaker dependent model. Last, we estimate the adapted MLLR transformation matrix set by weighting function with recognition result, the present MLLR matrix, and Eigenspace. The estimate result will be used to update the MLLR matrices in adaptation phase. The experimental results show that the proposed method can improve 5% to 8% speech recognition accuracy with speaker adaptation.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Abstract",
                "sec_num": null
            }
        ],
        "body_text": [
            {
                "text": "(x) (x, x ) L i i i i f t K \u03b3 \u03be = = + \u2211 (1) t i \u8868\u793a\u70ba\u7406\u60f3\u7684\u8f38\u51fa\u503c\uff0c 1 0 L i i i t \u03b3 = = \u2211 \uff0c\u4e26\u4e14 0 i \u03b3 > \u3002\u5411\u91cf\u7fa4 Xi",
                "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": "A B A f x A B B f x A B \u03b8 \u03b8 = = (3) \u5728\u9019\u88e1\u03b8\u6307\u7684\u662f\u6240\u6709\u8a9e\u97f3\u53c3\u6578\u6a21\u578b\u7684\u96c6\u5408\uff0cx \u5247\u8868\u793a\u8abf\u9069\u8a9e\u6599\u7684\u89c0\u6e2c\u503c\u3002 \u5728\u8a31\u591a\u7684\u8a9e\u8005\u8abf\u9069\u6cd5\u7576\u4e2d\uff0c\u6700\u5927\u76f8\u4f3c\u5ea6\u7dda\u6027\u8ff4\u6b78\u7684\u65b9\u6cd5\u88ab\u5ee3\u6cdb\u7684\u61c9\u7528\u5728\u5feb\u901f\u8a9e\u8005\u8abf\u9069\uff0c \u4f8b\u5982\u5b83\u53ea\u9700\u8981\u4e9b\u8a31\u7684\u8a9e\u6599\u4fbf\u53ef\u4ee5\u5c0d\u6a21\u578b\u53c3\u6578\u505a\u8abf\u9069\u3002\u5728\u6700\u5927\u76f8\u4f3c\u5ea6\u7dda\u6027\u8ff4\u6b78\u4e2d\uff0c\u975e\u8a9e\u8005 \u7368\u7acb\u7684\u6a21\u578b\u53c3\u6578\u53ef\u4ee5\u6839\u64da\u4e00\u500b\u6216\u591a\u500b\u4eff\u5c04\u8f49\u63db\u65b9\u5f0f(affine transformations)\u4f86\u9054\u5230\u8abf\u9069 \u7684 \u76ee \u7684 \u3002 \u6700 \u5927 \u76f8 \u4f3c \u5ea6\u7dda \u6027 \u8ff4 \u6b78 \u8abf \u9069 \u6cd5 \u662f \u4f7f\u7528 \u4eff \u5c04 \u8f49 \u63db \u65b9 \u5f0f \u4f86 \u8abf\u9069 \u9ad8 \u65af \u6df7 \u5408 \u6a21 \u578b (Gaussian mixture model, GMM)\u4e2d\u6240\u6709\u6df7\u5408\u5143\u4ef6\u7684\u5e73\u5747\u503c(mean) \uff0c\u800c\u540c\u4e00\u500b\u4eff\u5c04\u8f49 \u63db\u6cd5\u5247\u53ef\u4ee5\u63d0\u4f9b\u6240\u6709\u7684\u6df7\u5408\u5143\u4ef6\u5171\u4eab\uff0c\u8868\u793a\u6cd5\u5982\u4e0b\uff1a \u02c6A b i i i \u00b5 \u00b5 = + \u2200 (4) i \u00b5 \u8868\u793a\u5728 GMM \u4e2d\u9084\u6c92\u88ab\u8abf\u9069\u904e\u7684\u5e73\u5747\u503c\uff0c\u800c\u02c6i \u00b5 \u8868\u793a\u5df2\u7d93\u8abf\u9069\u904e\u7684\u5e73\u5747\u503c\u3002 \u5728\u773e\u591a\u7684\u8cc7\u6599\u91cf\u7576\u4e2d\uff0c\u6df7\u5408\u7684\u5143\u4ef6\u53ef\u4ee5\u88ab\u6b78\u985e\u6210\u591a\u500b\u985e\u5225\uff0c\u4e14\u4e0d\u540c\u7684\u4eff\u5c04\u8f49\u63db\u65b9\u5f0f\u53ef\u4ee5 \u5728\u4e0d\u540c\u7684\u985e\u5225\u4e2d\u88ab\u5171\u7528\uff0c\u5ef6\u4f38(4)\u7684\u8868\u793a\u6cd5\uff1a 1 1 1,\u00c2 b class i i i \u00b5 \u00b5 \u00b5 = + \u2200 \u2208 (5) 2 2 2,\u00c2 b class i i i \u00b5 \u00b5 \u00b5 = + \u2200 \u2208",
                        "eq_num": "(6)"
                    }
                ],
                "section": "",
                "sec_num": null
            },
            {
                "text": "\uff1a ) \uff1a ) \uff1a ) \uff1a \u5728\u7d93\u904e\u8fa8\u8b58\u968e\u6bb5\u53d6\u5f97\u8fa8\u8b58\u7d50\u679c\u5f8c\uff0c\u6703\u5c07\u8fa8\u8b58\u7d50\u679c\u7576\u6210\u70ba\u8abf\u9069\u8a9e\u6599 (adaptation data) \uff0c \u5c0d \u65bc \u6bcf \u4e00 \u500b \u6e2c \u8a66 \u8a9e \u8005 \u6240 \u63d0 \u4f9b \u7684 \u8abf \u9069 \u8a9e \u6599 \u6703 \u4f7f \u7528 \u6700 \u5927 \u76f8 \u4f3c \u5ea6 \u4f30 \u6e2c (",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "",
                "sec_num": null
            }
        ],
        "back_matter": [],
        "bib_entries": {
            "BIBREF0": {
                "ref_id": "b0",
                "title": "Fast speaker adaptation using eigenspace-based maximum likelihood linear regression",
                "authors": [
                    {
                        "first": "K",
                        "middle": [],
                        "last": "Chen",
                        "suffix": ""
                    }
                ],
                "year": 2000,
                "venue": "Proc. ICSLP",
                "volume": "",
                "issue": "",
                "pages": "",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "K. Chen et al, \"Fast speaker adaptation using eigenspace-based maximum likelihood linear regression,\" in Proc. ICSLP, Beijing, Oct. 2000.",
                "links": null
            },
            "BIBREF1": {
                "ref_id": "b1",
                "title": "Speaker Adaptation: Techniques and Challenges",
                "authors": [
                    {
                        "first": "P",
                        "middle": [
                            "C"
                        ],
                        "last": "Woodland",
                        "suffix": ""
                    }
                ],
                "year": 2000,
                "venue": "Proc. IEEE Workshop on Automatic Speech Recognition and Understanding",
                "volume": "",
                "issue": "",
                "pages": "85--90",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "P. C. Woodland, \"Speaker Adaptation: Techniques and Challenges\", Proc. IEEE Workshop on Automatic Speech Recognition and Understanding, pp.85-90, 2000.",
                "links": null
            },
            "BIBREF2": {
                "ref_id": "b2",
                "title": "Kernel Eigenspace-Based MLLR Adaptation",
                "authors": [
                    {
                        "first": "Brian Kan-Wing",
                        "middle": [],
                        "last": "Mak",
                        "suffix": ""
                    },
                    {
                        "first": "Roger Wend-Huu",
                        "middle": [],
                        "last": "Hsiao",
                        "suffix": ""
                    }
                ],
                "year": 2007,
                "venue": "IEEE Trans. on Audio, Speech, and Language Processing",
                "volume": "15",
                "issue": "3",
                "pages": "",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Brian Kan-Wing Mak, and Roger Wend-Huu Hsiao, \"Kernel Eigenspace-Based MLLR Adaptation\", IEEE Trans. on Audio, Speech, and Language Processing, vol. 15, no. 3, Mar 2007.",
                "links": null
            },
            "BIBREF4": {
                "ref_id": "b4",
                "title": "Mean and variance adaptation within the MLLR framework",
                "authors": [
                    {
                        "first": "M",
                        "middle": [
                            "J F"
                        ],
                        "last": "Gales",
                        "suffix": ""
                    },
                    {
                        "first": "P",
                        "middle": [
                            "C"
                        ],
                        "last": "Woodland",
                        "suffix": ""
                    }
                ],
                "year": 1996,
                "venue": "Computer Speech and Language",
                "volume": "10",
                "issue": "4",
                "pages": "249--264",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "M.J.F.Gales and P.C.Woodland, \"Mean and variance adaptation within the MLLR framework\", Computer Speech and Language, vol. 10, no. 4, pp. 249-264, 1996.",
                "links": null
            },
            "BIBREF5": {
                "ref_id": "b5",
                "title": "The HTK book",
                "authors": [
                    {
                        "first": "S",
                        "middle": [],
                        "last": "Young",
                        "suffix": ""
                    }
                ],
                "year": null,
                "venue": "",
                "volume": "",
                "issue": "",
                "pages": "",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "S.Young et al, \"The HTK book,\" http://htk.eng.cam.ac.uk.",
                "links": null
            },
            "BIBREF6": {
                "ref_id": "b6",
                "title": "A multi-class MLLR kernel for SVM speaker recognition",
                "authors": [
                    {
                        "first": "N",
                        "middle": [],
                        "last": "Zahi",
                        "suffix": ""
                    },
                    {
                        "first": "William",
                        "middle": [
                            "M"
                        ],
                        "last": "Karam",
                        "suffix": ""
                    },
                    {
                        "first": "",
                        "middle": [],
                        "last": "Campbell",
                        "suffix": ""
                    }
                ],
                "year": 2008,
                "venue": "",
                "volume": "",
                "issue": "",
                "pages": "",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Zahi N. Karam and William M. Campbell, \"A multi-class MLLR kernel for SVM speaker recognition,\" in ICASSP 2008.",
                "links": null
            },
            "BIBREF7": {
                "ref_id": "b7",
                "title": "Eigen-MLLR Coefficients as New Feature Parameters for Speaker Identification",
                "authors": [
                    {
                        "first": "J",
                        "middle": [
                            "C"
                        ],
                        "last": "Nick",
                        "suffix": ""
                    },
                    {
                        "first": "Wei-Ho",
                        "middle": [],
                        "last": "Wang",
                        "suffix": ""
                    },
                    {
                        "first": "Lin-Shan",
                        "middle": [],
                        "last": "Tsai",
                        "suffix": ""
                    },
                    {
                        "first": "",
                        "middle": [],
                        "last": "Lee",
                        "suffix": ""
                    }
                ],
                "year": 2001,
                "venue": "European Conference on Speech Communication and Technology",
                "volume": "",
                "issue": "",
                "pages": "",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Nick J.C.Wang, Wei-Ho Tsai and Lin-shan Lee, \"Eigen-MLLR Coefficients as New Feature Parameters for Speaker Identification,\" 2001 European Conference on Speech Communication and Technology, Aalborg, Denmark, Sept 2001",
                "links": null
            },
            "BIBREF8": {
                "ref_id": "b8",
                "title": "Critical Band Subspace-Based Speech Enhancement Using SNR and Auditory Masking Aware Technique",
                "authors": [
                    {
                        "first": "Wang",
                        "middle": [],
                        "last": "Jia-Ching",
                        "suffix": ""
                    },
                    {
                        "first": "Lee",
                        "middle": [],
                        "last": "Hsiao-Ping",
                        "suffix": ""
                    },
                    {
                        "first": "Wang",
                        "middle": [],
                        "last": "Jhing-Fa",
                        "suffix": ""
                    },
                    {
                        "first": "Yang",
                        "middle": [],
                        "last": "Chung-Hsien",
                        "suffix": ""
                    }
                ],
                "year": 2007,
                "venue": "IEICE Transactions on Information and Systems. vol",
                "volume": "90",
                "issue": "7",
                "pages": "1055--1062",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Wang Jia-Ching, Lee Hsiao-Ping, Wang Jhing-Fa, and Yang Chung-Hsien, \"Critical Band Subspace-Based Speech Enhancement Using SNR and Auditory Masking Aware Technique\", IEICE Transactions on Information and Systems. vol 90; number 7, pages 1055-1062, July 2007",
                "links": null
            },
            "BIBREF9": {
                "ref_id": "b9",
                "title": "Fast speaker adaptation using eigenspace-based maximum-likelihood linear regression",
                "authors": [
                    {
                        "first": "K",
                        "middle": [
                            "T"
                        ],
                        "last": "Chen",
                        "suffix": ""
                    },
                    {
                        "first": "W",
                        "middle": [
                            "W"
                        ],
                        "last": "Liau",
                        "suffix": ""
                    },
                    {
                        "first": "H",
                        "middle": [
                            "M"
                        ],
                        "last": "Wang",
                        "suffix": ""
                    },
                    {
                        "first": "L",
                        "middle": [
                            "S"
                        ],
                        "last": "Lee",
                        "suffix": ""
                    }
                ],
                "year": 2000,
                "venue": "Proc. ICSLP",
                "volume": "3",
                "issue": "",
                "pages": "742--745",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "K. T. Chen, W. W. Liau, H. M. Wang, and L. S. Lee, \"Fast speaker adaptation using eigenspace-based maximum-likelihood linear regression,\" in Proc. ICSLP, 2000, vol. 3, pp. 742-745.",
                "links": null
            }
        },
        "ref_entries": {
            "TABREF1": {
                "content": "<table><tr><td>\u7684\u6c7a\u5b9a\u662f\u5728\u65bc\u6240\u5b9a\u7684\u6a19\u6e96(threshold)\u4e4b\u4e0a\u6216\u4e4b\u4e0b\u3002</td><td/></tr><tr><td>(\u4e09) \u3001\u7279\u5fb5\u5f0f\u6700\u5927\u76f8\u4f3c\u5ea6\u7dda\u6027\u8ff4\u6b78(Eigen-Maximum Likelihood Linear</td><td/></tr><tr><td>Regression, Eigen-MLLR)</td><td/></tr><tr><td colspan=\"2\">\u9996\u5148\u6211\u5011\u5f9e\u539f\u59cb\u7684 MLLR \u6f14\u7b97\u6cd5\u4f86\u770b [3] [5]\uff0c\u6b64\u65b9\u6cd5\u7684\u539f\u7406\u80cc\u5f8c\u662f\u900f\u904e\u5047\u8a2d\u6539\u8b8a\u5f8c\u7684</td></tr><tr><td>\u53c3\u6578\u6703\u548c\u539f\u672c\u7684\u57fa\u672c\u53c3\u6578\u9593\u552f\u4e00\u7dda\u6027\u8ff4\u6b78\u7684\u51fd\u6578\u95dc\u4fc2\uff1a\u5982\u4e0b</td><td/></tr><tr><td>Y AX B = +</td><td>(2)</td></tr><tr><td colspan=\"2\">\u82e5\u5c07\u6b64\u539f\u7406\u4f7f\u7528\u5728\u8a9e\u8005\u8abf\u9069\u4e2d\uff0c\u8a9e\u6599\u7684\u6bcf\u4e00\u500b\u7279\u5fb5\u5411\u91cf\u63a5\u4ee3\u8868\u8457\u8a9e\u8005\u8207\u56e0\u7a7a\u9593\u4e2d\u7684\u5176\u4e2d</td></tr><tr><td colspan=\"2\">\u4e00\u500b\u6a23\u672c\uff0c\u800c\u6211\u5011\u4e5f\u5047\u8a2d\u4e86\u5f85\u6c42\u53c3\u6578\u548c\u73fe\u6709\u53c3\u6578\u4e4b\u9593\u7684\u51fd\u6578\u95dc\u4fc2\uff0c\u56e0\u6b64\u53ef\u4ee5\u4fbf\u53ef\u4ee5\u4f7f\u7528</td></tr><tr><td>\u6700\u5927\u76f8\u4f3c\u5ea6(Maximum Likelihood, ML)\u4f30\u6e2c\u6cd5\u4f86\u6c42\u5f97 A\u3001B \u7684\u503c\uff0c\u5982\u4e0b\u6240\u793a\uff1a</td><td/></tr><tr><td>arg max ( | , , )</td><td/></tr><tr><td>arg max ( | , , )</td><td/></tr></table>",
                "text": "\u70ba\u85c9\u7531\u6700\u4f73\u5316\u8655\u7406 \u5f9e\u8a13\u7df4\u96c6\u5408\u4e2d\u7372\u5f97\u7684\u652f\u63f4\u5411\u91cf\u7fa4(Support Vectors) \u3002\u7406\u60f3\u7684\u8f38\u51fa\u503c\u85c9\u8457\u76f8\u95dc\u7684\u652f\u63f4\u5411\u91cf \u662f\u843d\u5728\u985e\u5225 1(\u5176\u503c\u70ba\uff0b1)\u6216\u985e\u5225 2(\u5176\u503c\u70ba-1) \u3002\u5c0d\u65bc\u5206\u985e\u4f86\u8aaa\uff0c\u6578\u503c f(x)\u6240\u5c6c\u985e\u5225",
                "num": null,
                "html": null,
                "type_str": "table"
            },
            "TABREF3": {
                "content": "<table><tr><td/><td colspan=\"2\">Recognition Phase SVM classification (Recognition Phase)</td><td>Test Data MLLR regression Matrix database</td></tr><tr><td/><td/><td/><td>(One of S speakers)</td></tr><tr><td colspan=\"4\">Adaptation Phase SVM classification Adaptation Data 2 3 \u6b63\u78ba\u8a9e\u53e5 139 136 137 142 142 135 142 141 142 138 Update Speech SI Model Find speaker database \u6e2c\u8a66\u8a9e\u8005 \u6e2c\u8a66\u8a9e\u8005 \u6e2c\u8a66\u8a9e\u8005 \u6e2c\u8a66\u8a9e\u8005 1 4 5 6 7 8 9 10 \u7e3d\u7d50\u679c \u7e3d\u7d50\u679c \u7e3d\u7d50\u679c \u7e3d\u7d50\u679c SI Model MLLR-MS \u6b63\u78ba\u8a9e\u53e5 \u6b63\u78ba\u8a9e\u53e5 \u6b63\u78ba\u8a9e\u53e5 1394</td></tr><tr><td>\u932f\u8aa4\u8a9e\u53e5 \u932f\u8aa4\u8a9e\u53e5 \u932f\u8aa4\u8a9e\u53e5 \u932f\u8aa4\u8a9e\u53e5 \u8fa8 \u8b58 \u6b63 \u78ba \u8fa8 \u8b58 \u6b63 \u78ba \u8fa8 \u8b58 \u6b63 \u78ba \u8fa8 \u8b58 \u6b63 \u78ba \u7387 \u7387 \u7387 \u7387(%)</td><td colspan=\"3\">Finding speaker MLLR matrix set MLLR full regression To find speaker MLLR-RMS 14 13 8 8 8 matrix estimation Maximum likelihood coordinate estimation MLLR-RMS By ML 92.6 90.6 91.3 94.6 94.6 90 94.6 94 94.6 92 recognition Adaptation Data weight 11 9 8 12</td><td>106 92.9</td></tr><tr><td/><td>MLLR matrix set</td><td>MLLR regression</td><td>(Adaptation Phase)</td></tr><tr><td/><td>(Adaptation Phase) Eigenspace</td><td>Matrix database</td></tr><tr><td/><td colspan=\"3\">(Training Phase) \u5716\u516d\u3001\u6cdb\u5728\u9ea5\u514b\u98a8\u9663\u5217\u793a\u610f\u5716</td></tr><tr><td/><td colspan=\"3\">\u5716\u4e94\u3001\u8abf\u9069\u968e\u6bb5\u67b6\u69cb\u5716</td></tr><tr><td colspan=\"4\">\u5be6\u9a57\u904e\u7a0b\u7e3d\u5171\u4f7f\u7528\u4e86 10 \u500b\u4eba\u70ba\u63d0\u4f9b\u8a13\u7df4\u8a9e\u6599\u4ee5\u53ca\u8abf\u9069\u7684\u8a9e\u8005\uff0c\u6bcf\u500b\u8a9e\u8005\u7684\u8a13\u7df4\u8a9e\u53e5\u90fd \u56db\u3001\u5be6\u9a57\u7d50\u679c \u70ba 15 \u53e5\u3002\u6240\u63a1\u53d6\u7684\u8fa8\u8b58\u7387\u8a08\u7b97\u4ee5\u6b63\u78ba\u7387\u70ba\u4e3b\uff0c\u7686\u4ee5\u767e\u5206\u6bd4\u8868\u793a\uff0c\u8a08\u7b97\u65b9\u5f0f\u5982\u4e0b\uff1a \u8868\u4e03\u3001\u672c\u8ad6\u6587\u6240\u63d0\u51fa\u7684\u8a9e\u8005\u8abf\u9069\u6cd5 150 \u53e5\u751f\u6d3b\u7528\u8a9e\u6e2c\u8a66</td></tr><tr><td colspan=\"4\">Maximum Likelihood estimate, ML)\u65b9\u6cd5\u4f86\u5c07\u8a9e\u8005\u5b9a\u4f4d\u5728\u7279\u5fb5\u7a7a\u9593\u4e2d\u7684\u56de\u6b78\u77e9\u9663\u3002\u6700\u5927\u76f8\u4f3c\u5ea6 \u4f30\u6e2c\u548c\u6700\u5927\u76f8\u4f3c\u5ea6\u7dda\u6027\u56de\u6b78\u6703\u4f7f\u7528\u8abf\u9069\u8a9e\u6599\u4f86\u500b\u5225\u505a\u4e00\u500b\u65b0\u7684\u4f30\u6e2c\u3002\u500b\u5225\u4f30\u6e2c\u5f8c\u6240 \u7372\u5f97\u7684\u7d50\u679c\u5c07\u6703\u548c\u5229\u7528 SVM \u6240\u627e\u51fa\u5c0d\u61c9\u8a9e\u8005\u7684 MLLR \u56de\u6b78\u77e9\u9663\uff0c\u4e09\u8005\u505a\u4e00\u500b weighting \u7684\u904b\u7b97\uff0c\u4e26\u5c07\u7372\u5f97\u7684\u7d50\u679c\u5c0d\u539f\u672c\u8a9e\u8005\u7684 MLLR \u56de\u6b78\u77e9\u9663\u4f5c\u66f4\u65b0\u3002\u65bc\u662f\u6211 \u6211\u5011\u5be6\u73fe\u5be6\u9a57\u7684\u7d50\u679c\u662f\u4ee5\u53f0\u7063\u53e3\u97f3\u4e2d\u6587\u8a9e\u6599 MAT-400 \u70ba\u57fa\u790e\u7684\u8a9e\u97f3\u6a21\u578b\u3002\u8abf\u9069\u8a9e\u6599\u5247 \u662f\u4ee5\u65e5\u5e38\u751f\u6d3b\u7528\u8a9e\u5927\u7d04 7~10 \u500b\u5b57\uff0c\u800c\u5728\u6e2c\u8a66\u6642\u5247\u4ee5\u4eba\u540d\u6216\u662f\u65e5\u5e38\u751f\u6d3b\u7528\u8a9e\u70ba\u4e3b\u3002\u8a9e\u97f3 = 100% \u00d7 \u8fa8\u8b58\u6b63\u78ba\u8a9e\u53e5\u53e5\u6578 \u8fa8\u8b58\u6b63\u78ba\u7387 (8) \u6e2c\u8a66\u8a9e\u8005 \u6e2c\u8a66\u8a9e\u8005 \u6e2c\u8a66\u8a9e\u8005 \u6e2c\u8a66\u8a9e\u8005 1 2 3 4 5 6 7 8 9 10 \u7e3d\u7d50\u679c \u7e3d\u7d50\u679c \u7e3d\u7d50\u679c \u7e3d\u7d50\u679c \u5168\u90e8\u8a9e\u53e5\u53e5\u6578 \u6b63\u78ba\u8a9e\u53e5 \u6b63\u78ba\u8a9e\u53e5 \u6b63\u78ba\u8a9e\u53e5 \u6b63\u78ba\u8a9e\u53e5 142 137 139 144 141 138 140 144 145 138 1408 \u7279\u5fb5\u7684\u652b\u53d6\u8a2d\u5b9a\u5247\u5b9a\u7fa9\u5728\u8868\u4e09\u3002\u6211\u5011\u4f7f\u7528 MLLR \u8abf\u9069\u88fd\u9020\u51fa\u64f4\u5f35\u6bcf\u4e00\u500b\u8a9e\u53e5\u7684\u57fa\u5e95 SVM \u7279\u5fb5\uff0c\u5c0d\u65bc\u9ad8\u65af\u6df7\u5408\u6a21\u578b\u7684\u6bcf\u4e00\u500b\u5177\u9ad4\u7684\u985e\u5225\u90fd\u8981\u505a\u8abf\u9069\u3002\u6211\u5011\u4fbf\u4f7f\u7528 HTK[6] \u4f86 \u5efa\u7acb\u4e0a\u8ff0\u7684\u5de5\u4f5c\u4e26\u4e14\u7522\u751f\u4e00\u905e\u8ff4\u5f0f MLLR \u4f86\u5efa\u7acb\u8f49\u63db\u5f0f\u3002 \u8868\u56db\u70ba\u5728\u5c1a\u672a\u6709\u4efb\u4f55\u8a9e\u8005\u8abf\u9069\u65b9\u6cd5\u7684\u6cdb\u5728\u8a9e\u97f3\u8fa8\u8b58\u7cfb\u7d71\u4e0b\u6240\u6e2c\u8a66\u7684\u7d50\u679c\uff0c\u53ef\u4ee5\u767c\u73fe\u521d\u59cb \u932f\u8aa4\u8a9e\u53e5 \u932f\u8aa4\u8a9e\u53e5 \u932f\u8aa4\u8a9e\u53e5 \u932f\u8aa4\u8a9e\u53e5 8 13 11 6 9 12 10 6 5 12 92 \u8072\u97f3\u6a21\u578b\u78ba\u5be6\u662f\u5c0d\u4efb\u4f55\u4e00\u500b\u8a9e\u8005\u90fd\u63d0\u4f9b\u76f8\u540c\u7684\u8fa8\u8b58\u6e96\u78ba\u6548\u80fd\uff0c\u7e3d\u5e73\u5747\u6e96\u78ba\u7387\u70ba 85.7%\u3002 \u8868\u56db\u3001\u672a\u6709\u4efb\u4f55\u8a9e\u8005\u8abf\u9069\u6cd5 150 \u53e5\u751f\u6d3b\u7528\u8a9e\u6e2c\u8a66 \u8fa8 \u8b58 \u6b63 \u78ba \u8fa8 \u8b58 \u6b63 \u78ba \u8fa8 \u8b58 \u6b63 \u78ba \u8fa8 \u8b58 \u6b63 \u78ba \u7387 \u7387 \u7387 \u7387(%) 94.7 91.3 92.7 96 94 92 93.3 96 96.7 92 93.9</td></tr><tr><td colspan=\"4\">\u5c0d\u65bc\u5df2\u7d93\u63d0\u4f9b\u8a13\u7df4\u8a9e\u6599\u7684\u76ee\u6a19\u8a9e\u8005\uff0c\u6bcf\u53e5\u8a71\u90fd\u6703\u7528\u4e00\u500b SVM \u7279\u5fb5\u5411\u91cf\u4f86\u4ee3\u8868\uff0c\u800c\u9078\u64c7 \u5011\u4fee\u6539\u4e86 [10] \u4e2d\u7684 Equation (5) \u589e\u52a0\u4e86\u4e00\u500b\u8abf\u9069\u4fe1\u5fc3\u5ea6\u7684\u4f30\u7b97\u3002\u4e0b\u5217\u70ba\u4e00\u500b\u66f4\u65b0 SVM \u7684\u985e\u5225\u5247\u6703\u900f\u904e\u76ee\u6a19\u8a9e\u8005\u7684\u7279\u5fb5\u5411\u91cf\u548c\u539f\u6709\u7684 SVM \u7279\u5fb5\u5411\u91cf\u985e\u5225\u4f86\u4f5c\u4e00\u6bd4\u91cd\u9078 \u6e2c\u8a66\u8a9e\u8005 \u6e2c\u8a66\u8a9e\u8005 \u6e2c\u8a66\u8a9e\u8005 \u6e2c\u8a66\u8a9e\u8005 1 2 3 4 5 6 7 8 9 10 \u7e3d\u7d50\u679c \u7e3d\u7d50\u679c \u7e3d\u7d50\u679c \u7e3d\u7d50\u679c \u503c\u7684\u904b\u7b97\u3002 \u64c7\u3002 \u6b63\u78ba\u8a9e\u53e5 \u6b63\u78ba\u8a9e\u53e5 \u6b63\u78ba\u8a9e\u53e5 \u6b63\u78ba\u8a9e\u53e5 130 124 126 133 129 126 127 131 134 125 1285 \u4e94\u3001\u7d50\u8ad6</td></tr><tr><td colspan=\"4\">( ) ( ) \u0174 (1 ) W n c present conf c m m \u03b3 \u03be \uf8f6 \uf8f7 \uf8f7 + \u2212 \u22c5 \uf8f7 \uf8f7 \uf8f8 (7) \u5728\u6211\u5011\u7684\u5be6\u9a57\u4e2d\uff0c\u4f7f\u7528\u4e86\u6cdb\u5728\u8a9e\u97f3\u8fa8\u8b58\u7cfb\u7d71(ubiquitous speech recognition system)\u4f86\u6e2c 1 1 1 1 W W c N M EIGEN c m n c conf S N n m n \u03bb \u03be \u03bb \u03b3 = = = = \uf8eb \u22c5 + \uf8ec \uf8ec = \u22c5 \uf8ec + \uf8ec \uf8ed \u2211 \u2211 \u8a66\uff0c\u7e3d\u5171\u4f7f\u7528\u4e86 6 \u652f\u5168\u5411\u6027\u9ea5\u514b\u98a8\u4f48\u5728\u623f\u9593\u7684\u5929\u82b1\u677f\uff0c\u5176\u6536\u97f3\u7bc4\u570d\u53ef\u4ee5\u6db5\u84cb\u4e86\u6574\u500b\u623f\u9593 \u5167\u90e8\uff0c\u5982\u5716\u516d\u3002\u7531\u65bc\u5176\u6392\u5217\u65b9\u5f0f\u4e26\u975e\u50b3\u7d71\u7684\u9ea5\u514b\u98a8\u9663\u5217\u65b9\u5f0f\uff0c\u7121\u6cd5\u4f7f\u7528\u50b3\u7d71\u7684\u96dc\u8a0a\u6d88\u9664 \u932f\u8aa4\u8a9e\u53e5 \u932f\u8aa4\u8a9e\u53e5 \u932f\u8aa4\u8a9e\u53e5 \u932f\u8aa4\u8a9e\u53e5 20 26 24 17 21 24 23 19 16 25 215 \u8fa8 \u8b58 \u6b63 \u78ba \u8fa8 \u8b58 \u6b63 \u78ba \u8fa8 \u8b58 \u6b63 \u78ba \u5728\u8a9e\u97f3\u8fa8\u8b58\u7cfb\u7d71\u4e2d\uff0c\u8a9e\u8005\u8abf\u9069\u7684\u5de5\u4f5c\u662f\u6574\u9ad4\u8fa8\u8b58\u6548\u80fd\u7684\u4e00\u500b\u5f88\u91cd\u8981\u7684\u74b0\u7bc0\uff0c\u800c\u4e14\u5c0d\u65bc\u5728 \u8fa8 \u8b58 \u6b63 \u78ba \u7387 \u7387 \u7387 \u7387(%) 86.6 82.6 84 88.7 86 84 84.7 87.3 89.3 83.3 \u591a\u4eba\u4f7f\u7528\u7684\u74b0\u5883\u4e0b\uff0c\u8a9e\u8005\u8abf\u9069\u6280\u8853\u5728\u5feb\u901f\u8abf\u9069\u4ee5\u53ca\u975e\u76e3\u7763\u5f0f\u8abf\u9069\u7684\u60c5\u6cc1\u4e0b\u5c31\u66f4\u5fc5\u9808\u52a0 85.7 \u5f37\uff0c\u9019\u5169\u65b9\u9762\u4e5f\u6703\u662f\u4eca\u5f8c\u8a9e\u8005\u8abf\u9069\u6280\u8853\u767c\u5c55\u7684\u91cd\u9ede\u3002\u4e00\u500b\u8abf\u9069\u6280\u8853\u4e0d\u50c5\u8981\u80fd\u5920\u9054\u6210\u5feb\u901f \u2211 \u2211 M \u8868\u793a\u70ba\u7279\u5fb5\u5411\u7684\u6578\u76ee\uff0cn \u8868\u793a\u70ba\u5728 N c \u4e2d\u7684\u4e00\u500b\u6df7\u5408\u5143\u4ef6\uff0cr n (m)\u8868\u793a\u5728\u6642\u9593\u70ba t \u6642\u7684\u89c0\u6e2c\u6a5f\u7387\uff0cW c present \uff0cW c EIGEN \uff0cW c \u6cd5\u4f86\u9054\u5230\u8f03\u597d\u7684\u6548\u679c\uff0c\u6240\u4ee5\u4fbf\u5c07 6 \u652f\u9ea5\u514b\u98a8\u4f7f\u7528\u591a\u901a\u9053\u6df7\u97f3(multi-channel mixer)\u7684 \u65b9\u5f0f\u6210\u55ae\u4e00\u8f38\u5165\uff0c\u518d\u4f7f\u7528\u5b50\u7a7a\u9593\u5f0f\u8a9e\u97f3\u589e\u5f37\u6f14\u7b97\u6cd5 [9](Subspace Speech Enhancement, Using SNR and Auditory Masking Aware Technique)\u5728\u8a9e\u97f3\u8a0a\u865f\u7684\u524d\u8655\u7406\u4f86\u6d88\u9664\u8a9e\u97f3\u96dc \u8a0a\uff0c\u518d\u958b\u59cb\u53d6\u8a9e\u97f3\u7279\u5fb5\u3002 \u8abf\u9069\uff0c\u4e14\u5fc5\u9808\u80fd\u5920\u5728\u53ea\u6709\u5c11\u91cf\u8a9e\u6599\u7684\u60c5\u5f62\u4e0b\u5c07\u73fe\u6709\u539f\u59cb\u7684\u8072\u5b78\u6a21\u578b\u8abf\u6574\u81f3\u66f4\u9069\u5408\u7576\u4e0b\u8a9e \u63a5\u8457\u6211\u5011\u4ee5\u50b3\u7d71\u7684 MAP \u8a9e\u8005\u8abf\u9069\u65b9\u5f0f\u4f86\u4f7f\u7528\u65bc\u7cfb\u7d71\uff0c\u8868\u4e94\u70ba\u4f7f\u7528 MAP \u8a9e\u8005\u8abf\u9069\u5f8c\u7684 \u8005\u7684\u72c0\u614b\uff0c\u7279\u5225\u662f\u5c0d\u65bc\u5728\u4e00\u500b\u7a7a\u9593\u4e2d\u7684\u56fa\u5b9a\u6210\u54e1\u66f4\u662f\u6709\u6240\u9700\u8981\uff0c\u5982\u5bb6\u5ead\u6210\u54e1\u3002\u5728\u672c\u8ad6\u6587 \u7d50\u679c\u3002\u53ef\u4ee5\u767c\u73fe\u5c0d\u65bc\u6bcf\u500b\u8a9e\u8005\u505a\u8abf\u9069\u4e4b\u5f8c\uff0c\u8fa8\u8b58\u7cfb\u7d71\u5c0d\u65bc\u6bcf\u500b\u4eba\u7684\u6e96\u78ba\u7387\u5e73\u5747\u90fd\u6709 \u63d0\u51fa\u7684\u67b6\u69cb\u4e2d\uff0c\u5229\u7528\u7279\u5fb5\u5f0f\u6700\u5927\u76f8\u4f3c\u5ea6\u7dda\u6027\u56de\u6b78(Eigen-MLLR)\u6240\u5efa\u7acb\u7684\u591a\u8a9e\u8005\u7279\u5fb5 2%~3%\u7684\u6e96\u78ba\u7387\u63d0\u5347\uff0c\u7e3d\u5e73\u5747\u6e96\u78ba\u7387\u70ba 88.2%\u3002 \u5411\u91cf\u7a7a\u9593\u7d50\u5408\u652f\u63f4\u5411\u91cf\u6a5f(SVM)\u7684\u5206\u985e\u4f86\u9054\u6210\u5feb\u901f\u591a\u8a9e\u8005\u8abf\u9069\u3002\u6e2c\u8a66\u8a9e\u8005\u7684\u8a9e\u53e5\u7d93 estimate \u5247\u5206\u5225\u70ba\u985e\u5225 c \u76ee\u524d\u7684\u56de\u6b78\u77e9\u9663\uff0c\u7531 ML \u4f30\u6e2c\u51fa\u4f86\u7684\u7279\u5fb5\u7a7a\u9593\u77e9\u9663\uff0c\u4ee5\u53ca\u7531 MLLR full regression \u4f30\u6e2c\u51fa\u4f86\u7684\u56de\u6b78\u77e9\u9663\u3002 W c \u70ba\u66f4\u65b0\u904e\u5f8c\u7684\u56de\u6b78\u77e9\u9663\u3002 conf \u03be \u70ba\u4fe1\u5fc3\u6bd4\u91cd\uff0c\u4fe1\u5fc3\u6bd4\u91cd\u5247\u662f\u4f9d\u7167\u8fa8\u8b58\u7d50\u679c\u800c\u5f97 \u7684\u3002\u589e\u52a0\u4fe1\u5fc3\u6bd4\u91cd\u53c3\u6578\u662f\u907f\u514d\u7576\u767c\u751f\u8fa8\u8b58\u7d50\u679c\u7522\u751f\u932f\u8aa4\u800c\u63a5\u53d7, \u4e5f\u5c31\u662f\u5c0d\u7684\u8a9e\u97f3\u8fa8 \u8868\u4e94\u3001\u4f7f\u7528 MAP \u8a9e\u8005\u8abf\u9069\u6cd5 150 \u53e5\u751f\u6d3b\u7528\u8a9e\u6e2c\u8a66 \u6e2c\u8a66\u8a9e\u8005 \u6e2c\u8a66\u8a9e\u8005 \u6e2c\u8a66\u8a9e\u8005 \u6e2c\u8a66\u8a9e\u8005 1 2 3 4 5 6 7 8 9 10 \u7e3d\u7d50\u679c \u7e3d\u7d50\u679c \u7e3d\u7d50\u679c \u7e3d\u7d50\u679c \u6b63\u78ba\u8a9e\u53e5 \u6b63\u78ba\u8a9e\u53e5 \u6b63\u78ba\u8a9e\u53e5 \u6b63\u78ba\u8a9e\u53e5 135 126 131 136 134 129 132 135 137 128 1323 \u904e Likelihood estimate)\u4e09\u8005\u4f86\u4f7f\u7528\u6bd4\u91cd\u53d6\u6c7a\u91cd\u65b0\u8a08\u7b97\uff0c\u4e26\u4e14\u5c07\u7d50\u679c\u4f86\u5373\u6642\u66f4\u65b0\u6e2c\u8a66\u8a9e\u8005\u76f8</td></tr><tr><td colspan=\"4\">\u8b58\u6210\u932f\u7684, \u6216\u932f\u7684\u8a9e\u97f3\u8fa8\u8b58\u6210\u5c0d\u7684, \u4f86\u5c0e\u81f4\u6574\u500b\u8abf\u9069\u6a21\u578b\u8d8a\u4f86\u8d8a\u5dee\u3002\u6574\u500b\u8abf\u9069\u904e\u7a0b \u932f\u8aa4\u8a9e\u53e5 \u932f\u8aa4\u8a9e\u53e5 \u932f\u8aa4\u8a9e\u53e5 \u932f\u8aa4\u8a9e\u53e5 15 24 19 14 16 21 18 15 13 22 177 \u5c0d\u61c9\u7684 MLLR \u56de\u6b78\u77e9\u9663\u53c3\u6578\u3002\u5728\u672c\u8ad6\u6587\u4e2d\u4e5f\u767c\u73fe\uff0c\u82e5\u53ef\u4ee5\u518d\u52a0\u5f37\u8a9e\u97f3\u589e\u5f37\u8655\u7406\u7684\u6f14\u7b97</td></tr><tr><td colspan=\"4\">\u5982\u5716\u4e94\u3002 \u8fa8 \u8b58 \u6b63 \u78ba \u8fa8 \u8b58 \u6b63 \u78ba \u8fa8 \u8b58 \u6b63 \u78ba \u8fa8 \u8b58 \u6b63 \u78ba \u7387 \u7387 \u7387 \u7387(%) \u6cd5\u964d\u4f4e\u66f4\u591a\u96dc\u8a0a\u4ee5\u53ca\u63d0\u5347\u8a0a\u865f\u5f37\u5ea6\uff0c\u5247\u5c0d\u6574\u500b\u8a9e\u8005\u8abf\u9069\u548c\u8fa8\u8b58\u6548\u80fd\u518d\u9032\u4e00\u6b65\u63d0\u5347\u3002\u5728\u672a 90 84 87.3 90.7 89.3 86 88 90 91.3 85.3 88.2 \u4f86\u7684\u8a9e\u97f3\u8fa8\u8b58\u74b0\u5883\u4e2d\uff0c\u5e0c\u671b\u80fd\u5920\u589e\u52a0\u66f4\u591a\u7684\u9ea5\u514b\u98a8\uff0c\u4f7f\u7684\u80fd\u9054\u5230\u5bec\u5bb9\u5ea6\u66f4\u9ad8\u7684\u6cdb\u5728\u8a9e\u97f3</td></tr><tr><td colspan=\"4\">\u4f7f\u7528\u74b0\u5883\uff0c\u4e5f\u53ef\u4ee5\u96a8\u6642\u52a0\u5165\u65b0\u7684\u8a9e\u8005\u8b93\u7cfb\u7d71\u53ef\u4ee5\u81ea\u6211\u66f4\u65b0\u4ee5\u53ca\u4f5c\u8abf\u9069\uff0c\u4e5f\u5e0c\u671b\u9019\u9805\u6280\u8853</td></tr><tr><td colspan=\"4\">\u80fd\u7d50\u5408\u5176\u4ed6\u7684\u61c9\u7528\u5230\u66f4\u5ee3\u6cdb\u7684\u5c64\u9762\uff0c\u8b93\u4eba\u5011\u751f\u6d3b\u53ef\u4ee5\u85c9\u8457\u6578\u4f4d\u5316\u66f4\u4fbf\u5229\u3002</td></tr></table>",
                "text": "\u8868\u516d\u5247\u4f7f\u7528\u50b3\u7d71 MLLR \u8a9e\u8005\u8abf\u9069\u6cd5\u4f7f\u7528\u65bc\u8fa8\u8b58\u7cfb\u7d71\u7684\u8fa8\u8b58\u7d50\u679c\uff0c\u76f8\u8f03\u65bc MAP \u8abf\u9069\u6cd5\uff0c \u6709\u8457\u66f4\u9ad8\u7684\u8fa8\u8b58\u7387\uff0c\u6703\u9020\u6210\u9019\u6a23\u7684\u7d50\u679c\u6700\u6709\u53ef\u80fd\u7684\u539f\u56e0\u662f 15 \u53e5\u7684\u8abf\u9069\u8a9e\u6599\u5c0d MAP \u4f86 \u8aaa\u662f\u975e\u5e38\u7684\u5c11\u91cf\uff0c\u800c\u4f7f\u5f97 MAP \u771f\u6b63\u7684\u512a\u9ede\u6839\u672c\u9084\u4f86\u4e0d\u53ca\u767c\u63ee\uff0c\u76f8\u8f03\u65bc MLLR\uff0c\u53cd\u800c\u80fd \u5728\u5c11\u91cf\u8a9e\u6599\u6642\u5c31\u986f\u793a\u51fa\u76f8\u7576\u51fa\u8272\u7684\u8868\u73fe\uff0c\u6bd4\u672a\u8abf\u9069\u7684\u6b63\u78ba\u7387\u591a\u51fa 5%~8%\uff0c\u6bd4 MAP \u591a \u51fa 3%~5%\u7684\u8fa8\u8b58\u6b63\u78ba\u7387\u3002 \u8868\u516d\u3001\u4f7f\u7528 MLLR \u8a9e\u8005\u8abf\u9069\u6cd5 150 \u53e5\u751f\u6d3b\u7528\u8a9e\u6e2c\u8a66 \u8868\u4e03\u7684\u5be6\u9a57\u7d50\u679c\u5247\u70ba\u672c\u8ad6\u6587\u6240\u63d0\u51fa\u61c9\u7528\u7279\u5fb5\u5f0f MLLR \u8207 SVM \u7684\u8a9e\u8005\u8abf\u9069\u67b6\u69cb\u3002\u8207 MAP \u8abf\u9069\u6cd5\u4f86\u6bd4\u8f03\uff0c\u7531\u65bc\u85c9\u8457 MLLR \u7684\u7279\u6027\u95dc\u4fc2\uff0c\u9084\u662f\u53ef\u4ee5\u85c9\u8457\u5c11\u91cf\u8a9e\u6599\u800c\u9054\u5230\u660e\u986f\u7684\u6548 \u679c\uff0c\u4e14\u5229\u7528 SVM \u4f86\u76f4\u63a5\u9078\u64c7\u76f8\u5c0d\u61c9\u7684\u7279\u5fb5\u53c3\u6578\u77e9\u9663\uff0c\u5c11\u6389\u4e86\u91cd\u65b0\u7531\u8a9e\u6599\u8a08\u7b97\u53c3\u6578\u7684\u904b \u7b97\u91cf\uff0c\u4e5f\u6bd4\u50b3\u7d71\u7684 MLLR \u5e73\u5747\u63d0\u5347\u4e86 1%\u5de6\u53f3\u8fa8\u8b58\u7387\uff0c\u76f8\u8f03\u65bc\u672a\u8abf\u9069\u7684\u8fa8\u8b58\u7cfb\u7d71\u66f4\u63d0\u5347 \u4e86 8%\u5e73\u5747\u8fa8\u8b58\u7387\uff0c\u76f8\u8f03\u65bc MAP \u5e73\u5747\u63d0\u5347\u4e86 4%~5%\u8fa8\u8b58\u7387\u3002 SVM \u5206\u985e\u5b8c\u7562\u4e4b\u5f8c\uff0c\u4fbf\u6703\u5728 MLLR \u56de\u6b78\u77e9\u9663\u7fa4\u4e2d\u627e\u51fa SVM \u985e\u5225\u76f8\u5c0d\u61c9\u7684\u56de\u6b78\u77e9\u9663 \u4e26\u4e14\u8207\u521d\u59cb\u6a21\u578b\u7d50\u5408\u6210\u8a9e\u8005\u7279\u5b9a\u6a21\u578b\uff0c\u518d\u9032\u884c\u8a9e\u97f3\u8fa8\u8b58\u3002\u7136\u5f8c\u5c07\u8a9e\u97f3\u8fa8\u8b58\u7684\u7d50\u679c\u5229\u7528 MLLR \u56de\u6b78\u77e9\u9663\u4f30\u6e2c (MLLR regression matrix estimate) \u4ee5\u53ca\u6700\u5927\u76f8\u4f3c\u5ea6\u4f30\u6e2c (Maximum",
                "num": null,
                "html": null,
                "type_str": "table"
            }
        }
    }
}