{
    "paper_id": "O08-1014",
    "header": {
        "generated_with": "S2ORC 1.0.0",
        "date_generated": "2023-01-19T08:02:19.187531Z"
    },
    "title": "Study of the Improved Normalization Techniques of Energy-Related Features for Robust Speech Recognition",
    "authors": [
        {
            "first": "Chi-An",
            "middle": [],
            "last": "\u6f58\u5409\u5b89",
            "suffix": "",
            "affiliation": {},
            "email": ""
        },
        {
            "first": "",
            "middle": [],
            "last": "Pan",
            "suffix": "",
            "affiliation": {},
            "email": ""
        },
        {
            "first": "Wen-Hsiang",
            "middle": [],
            "last": "\u675c\u6587\u7965",
            "suffix": "",
            "affiliation": {},
            "email": ""
        },
        {
            "first": "",
            "middle": [],
            "last": "Tu",
            "suffix": "",
            "affiliation": {},
            "email": ""
        }
    ],
    "year": "",
    "venue": null,
    "identifiers": {},
    "abstract": "The rapid development of speech processing techniques has made themselves successfully applied in more and more applications, such as automatic dialing, voice-based information retrieval, and identity authentication. However, some unexpected variations in speech signals deteriorate the performance of a speech processing system, and thus relatively limit its application range. Among these variations, the environmental mismatch caused by the embedded noise in the speech signal is the major concern of this paper. In this paper, we provide a more rigorous mathematical analysis for the effects of the additive noise on two energy-related speech features, i.e. the logarithmic energy (logE) and the zeroth cepstral coefficient (c0). Then based on these effects, we propose a new feature compensation scheme, named silence feature normalization (SFN), in order to improve the noise robustness of the above two features for speech recognition. It is shown that, regardless of its simplicity in implementation, SFN brings about very significant improvement in noisy speech recognition, and it behaves better than many well-known feature normalization approaches. Furthermore, SFN can be easily integrated with other noise robustness techniques to achieve an even better recognition accuracy.",
    "pdf_parse": {
        "paper_id": "O08-1014",
        "_pdf_hash": "",
        "abstract": [
            {
                "text": "The rapid development of speech processing techniques has made themselves successfully applied in more and more applications, such as automatic dialing, voice-based information retrieval, and identity authentication. However, some unexpected variations in speech signals deteriorate the performance of a speech processing system, and thus relatively limit its application range. Among these variations, the environmental mismatch caused by the embedded noise in the speech signal is the major concern of this paper. In this paper, we provide a more rigorous mathematical analysis for the effects of the additive noise on two energy-related speech features, i.e. the logarithmic energy (logE) and the zeroth cepstral coefficient (c0). Then based on these effects, we propose a new feature compensation scheme, named silence feature normalization (SFN), in order to improve the noise robustness of the above two features for speech recognition. It is shown that, regardless of its simplicity in implementation, SFN brings about very significant improvement in noisy speech recognition, and it behaves better than many well-known feature normalization approaches. Furthermore, SFN can be easily integrated with other noise robustness techniques to achieve an even better recognition accuracy.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Abstract",
                "sec_num": null
            }
        ],
        "body_text": [
            {
                "text": ")\u3001(b) \u8207(c)\u5206\u5225\u8868\u793a\u4e00\u4e7e\u6de8\u8a9e\u97f3\u8a0a\u865f(Aurora-2.0 \u8cc7\uf9be\u5eab\u4e2d\u7684\"MAH_1390A\"\u6a94)\u7684\u6ce2\u5f62\u5716\u3001\u5c0d\uf969 \u80fd\uf97e(logE)\u66f2\u7dda\u5716\u8207\u7b2c\uf9b2\u7dad\u5012\u983b\u8b5c\u7279\u5fb5\u4fc2\uf969(c0)\u66f2\u7dda\u5716\uff1b\u800c(b)\u8207(c)\u4e2d\u7d05\u8272\u5be6\u7dda\u3001\uf93d\u8272\u865b \u7dda\u8207\uf923\u8272\u9ede\u7dda\u5247\u5206\u5225\u70ba\u4e7e\u6de8\u8a9e\u97f3\u3001\u8a0a\u96dc\u6bd4 15dB \u7684\u8a9e\u97f3\u53ca\u8a0a\u96dc\u6bd4 5dB \u7684\u8a9e\u97f3\u6240\u5c0d\u61c9\u7684\u66f2 \u7dda\u3002\u7531\u9019\u4e09\u5f35\u5716\u4e2d\uff0c\u53ef\u4ee5\u5f88\u660e\u986f\u5730\u770b\u51fa\uff0c\u5728\u6709\u8a9e\u97f3\u5b58\u5728\u7684\u5340\u57df\uff0clogE \u8207 c0 \u7279\u5fb5\u503c\u8f03\u5927\uff0c \u8f03\uf967\u5bb9\uf9e0\u53d7\u5230\u96dc\u8a0a\u7684\u5f71\u97ff\u800c\u5931\u771f\uff0c\u800c\u4e14\u96a8\u6642\u9593\u4e0a\u4e0b\u632f\u76ea\u7684\u60c5\u6cc1\u8f03\u70ba\u660e\u986f\uff1b\u53cd\u4e4b\uff0c\u5728\u7121\u8a9e \u97f3\u5b58\u5728\u7684\u5340\u6bb5\uff0c\u5176\u7279\u5fb5\u503c\u524d\u5f8c\u8b8a\u5316\u8f03\u5e73\u7de9\uff0c\u4e14\u53d7\u5230\u96dc\u8a0a\u7684\u5e72\u64fe\u5f8c\uff0c\u5176\u503c\u6703\u5f88\u660e\u986f\u5730\u88ab\u6539 \u8b8a\u8a31\u591a\u3002\u63a5\u4e0b\uf92d\uff0c\u6211\u5011\u5c31\u4ee5\u8f03\u56b4\u8b39\u7684\uf969\u5b78\uf9e4\uf941\uff0c\u5c0d\u4ee5\u4e0a\uf978\u7a2e\u5931\u771f\u73fe\u8c61\u52a0\u4ee5\u5206\u6790\u8207\u63a2\u8a0e\u3002 \u9996\u5148\uff0c\u6211\u5011\u63a2\u8a0e\u52a0\u6210\u6027\u96dc\u8a0a\u5c0d\u65bc logE \u7279\u5fb5\u7684\u5f71\u97ff\u3002\u5047\u8a2d\u4e00\u6bb5\u53d7\u52a0\u6210\u6027\u96dc\u8a0a\u5e72\u64fe\u7684\u8a9e\u97f3 (noisy speech)\u4e2d\uff0c\u7b2cn \u500b\u97f3\u6846\u7684\u8a0a\u865f [ ] n x m \u53ef\u8868\u793a\u70ba\uff1a [ ] [ ] [ ] n n n x m s m d m = + \uff0c \u5f0f(2-1) \u5176\u4e2d [ ] n s m \u8207 [ ] n d m \u5206\u5225\u8868\u793a\u7b2cn \u500b\u97f3\u6846\u4e4b\u4e7e\u6de8\u8a9e\u97f3\u8a0a\u865f(clean speech)\u4ee5\u53ca\u96dc\u8a0a(noise)\uff0c \u5247\u6b64\u97f3\u6846\u4e4b logE \u7279\u5fb5\u503c\u53ef\u7528\u4e0b\u5f0f\u8868\u793a\uff1a ( ) [ ] ( ) ( ) 2 2 2 log [ ] log [ ] [ ] x m n m n m n E n x m s m d m = \u2248 + \u2211 \u2211 \u2211 ( ) [ ] ( ) ( ) [ ] ( ) ( ) log exp exp d s E n E n = + \uff0c \u5f0f(2-2) \u5176\u4e2d ( ) [ ] x E n \u3001 ( ) [ ] s E n \u8207 ( ) [ ] d E n \u5206\u5225\u70ba [ ] n x m \u3001 [ ] n s m \u4ee5\u53ca [ ] n d m \u6240\u5c0d\u61c9\u4e4b logE \u7279\u5fb5\u503c\u3002 \u56e0\u6b64\uff0c\u53d7\u5230\u96dc\u8a0a\u5e72\u64fe\u6240\u5c0e\u81f4\u96dc\u8a0a\u8a9e\u97f3\u8207\u4e7e\u6de8\u8a9e\u97f3\u8a0a\u865f\uf978\u8005\u9593 logE \u7279\u5fb5\u7684\u5dee\uf962 [ ] E n \u0394 \u53ef \u7528\u4e0b\u5f0f\u8868\u793a\uff1a [ ] ( ) [ ] ( ) [ ] ( ) [ ] ( ) [ ] ( ) ( ) log 1 exp d x s s E n E n E n E n E n \u0394 = \u2212 \u2248 + \u2212 \u3002 \u5f0f(2-3) \u7531\u5f0f(2-3)\u53ef\u89c0\u5bdf\u51fa\uff0c\uf974\u5728\u76f8\u540c\u7684\u96dc\u8a0a\u80fd\uf97e( ( ) [ ] d E n )\u4e0b\uff0c\u6b64\u5dee\uf962\u503c [ ] E n \u0394 \u8207\u4e7e\u6de8\u8a9e\u97f3\u8a0a\u865f \u4e4b ( ) [ ] s E n \uf978\u8005\u5448\u73fe\u8ca0\u76f8\u95dc\u7684\u95dc\u4fc2\uff0c\u7576 ( ) [ ] s E n \u6108\u5927\u6642\uff0c [ ] E n \u0394 \u6108\u5c0f\uff0c\u53cd\u4e4b\u5247\u6108\u5927\u3002\u6839\u64da\u4e0a \u8ff0\u7684\u63a8\u5c0e\uff0c\u53ef\u4ee5\u770b\u51fa\u4e00\u96dc\u8a0a\u8a9e\u97f3\u8a0a\u865f\u4e2d\uff0c\u542b\u6709\u8a9e\u97f3\u6210\u4efd\u7684\u97f3\u6846( ( ) [ ] s E n \u8f03\u5927)\u76f8\u8f03\u65bc\u7d14\u96dc \u8a0a\u97f3\u6846( ( ) [ ] s E n \u8f03\u5c0f)\u800c\u8a00\uff0c\u5176 logE \u7279\u5fb5\u88ab\u96dc\u8a0a\u5f71\u97ff\u7684\u60c5\u6cc1\u8f03\u5c0f(\u5373\u5931\u771f\uf97e [ ] E n \u0394 \u8f03\u5c0f)\u3002 \u63a5\u4e0b\uf92d\uff0c\u6211\u5011\u63a2\u8a0e\u52a0\u6210\u6027\u96dc\u8a0a\u5c0d\u65bc\u8a9e\u97f3\u8a0a\u865f\u7684 logE \u7279\u5fb5\u5e8f\uf99c\u65bc\u8abf\u8b8a\u983b\u8b5c(modulation spectrum)\u4e0a\u7684\u5f71\u97ff\u3002\u9996\u5148\uff0c\u6211\u5011\u5c07\u5f0f(2-2)\u4ee5\u6cf0\uf952\u7d1a\uf969(Taylor series)\u5c55\u958b\uff0c\u5176\u5c55\u958b\u7684\u4e2d\u5fc3 \u9ede\u8a2d\u5b9a\u70ba ( ) [ ] ( ) [ ] ( ) ( ) , 0 ,0 d s E n E n = \uff0c\u5c55\u958b\u968e\u5c64\u70ba 2 \u968e\uff0c\u5982\u5f0f(2-4)\u6240\u793a\uff1a \u5716\u4e8c\u3001\u5728\uf967\u540c SNR \u4e0b\uff0c\u4e00\u8a9e\u97f3\u8a0a\u865f\u4e4b\u6ce2\u5f62\u5716\u53ca\u80fd\uf97e\u76f8\u95dc\u7279\u5fb5\u6642\u9593\u5e8f\uf99c\u5716\uff0c\u5176\u4e2d(a)\u70ba\u4e7e \u6de8\u8a9e\u97f3\u6ce2\u5f62\u3001(b)\u70ba logE \u7279\u5fb5\u66f2\u7dda\u3001(c)\u70ba c0 \u7279\u5fb5\u66f2\u7dda ( ) [ ] ( ) [ ] ( ) ( ) [ ] ( ) ( ) log exp exp d x s E n E n E n \u2248 + ( ) [ ] ( ) [ ] ( ) ( ) [ ] ( ) ( ) [ ] ( ) ( ) [ ] ( ) [ ] ( ) 2 2 1 1 log 2 2 8 d d d s s s E n E n E n E n E n E n \u2248 + + + + \u2212 . \u5f0f(2-4) \u56e0\u6b64\uff0c\uf974\u5c07\u4e0a\u5f0f(2-4)\u53d6\u5085\uf9f7\uf96e\u8f49\u63db\uff0c\u5247\u6b64\u96dc\u8a0a\u8a9e\u97f3\u7684\u5c0d\uf969\u80fd\uf97e\u5e8f\uf99c ( ) [ ] { } x E n \u7684\u8abf\u8b8a \u983b\u8b5c\u53ef\u7528\u4e0b\u5f0f\u8868\u793a\uff1a ( ) ( ) ( ) ( ) ( ) ( ) 1 2 log 2 2 X j S j D j \u03c9 \u03c0 \u03b4 \u03c9 \u03c9 \u03c9 \u2248 + + ( ) ( ) ( ) ( ) ( ) ( ) ( ) 1 16 S j S j D j D j S j D j \u03c9 \u03c9 \u03c9 \u03c9 \u03c9 \u03c9 \u03c0 + * + * \u2212 * \uff0c \u5f0f(2-5) \u5f0f\u4e2d ( ) X j\u03c9 \u3001 ( ) S j\u03c9 \u4ee5\u53ca ( ) D j\u03c9 \u5206\u5225\u70ba\u96dc\u8a0a\u8a9e\u97f3\u4e4blogE\u5e8f\uf99c ( ) [ ] { } x E n \u3001\u4e7e\u6de8\u8a9e\u97f3\u4e4blogE \u5e8f\uf99c ( ) [ ] { } s E n \u8207\u96dc\u8a0a\u4e4blogE\u5e8f\uf99c ( ) [ ] { } d E n \u7684\u8abf\u8b8a\u983b\u8b5c\u3002\u5047\u8a2d ( ) [ ] { } s E n \u8207 ( ) [ ] { } d E n \uf978\u5e8f\uf99c \u7686 \u70ba \u4f4e \u901a (low-pass) \u8a0a \u865f \uff0c \u4e14 s B \u8207 d B \u70ba \u5176 \u76f8 \u5c0d \u61c9 \u4e4b \u983b \u5bec (bandwidth) \uff0c \u5247 \u5f0f (2-5) \u4e2d ( ) ( ) D j D j \u03c9 \u03c9 * \u8207 ( ) ( ) S j D j \u03c9 \u03c9 * \uf978\u9805\u7684\u983b\u5bec\u5206\u5225\u70ba 2 d B \u8207 s d B B + \uff1b\u9019\u610f\u5473\u8457\u96dc\u8a0a\u8a9e\u97f3 \u4e4blogE\u5e8f\uf99c ( ) [ ] { } x E n \u76f8\u8f03\u65bc\u96dc\u8a0a\u7684logE\u5e8f\uf99c ( ) [ ] { } d E n \u5c07\u64c1\u6709\uf901\u5927\u7684\u983b\u5bec\u3002\u63db\u8a00\u4e4b\uff0c\u5c0dlogE \u5e8f\uf99c\u800c\u8a00\uff0c\u96dc\u8a0a\u8a9e\u97f3\u6bd4\u96dc\u8a0a\u64c1\u6709\u8f03\u591a\u9ad8\u983b\u7684\u8abf\u8b8a\u983b\u8b5c\u6210\u4efd\uff1b\u9019\uf965\u53ef\u4ee5\u89e3\u91cb\u70ba\u4f55\u5728\u4e00\u96dc\u8a0a (a) (a) (b) (c) \u8a9e\u97f3\u8a0a\u865f\u4e2d\u542b\u6709\u8a9e\u97f3\u7684\u5340\u6bb5\uff0c\u6bd4\u8d77\u7d14\u96dc\u8a0a\u7684\u5340\u6bb5\u770b\u8d77\uf92d\u632f\u76ea\u60c5\u5f62(fluctuating)\uf901\u70ba\u660e\u986f\u3002 \u63a5\u8457\u6211\u5011\u63a2\u8a0e\u52a0\u6210\u6027\u96dc\u8a0a\u5c0d\u65bc c0 \u7279\u5fb5\u7684\u5f71\u97ff\u3002\u5047\u8a2d\u96dc\u8a0a\u8a9e\u97f3\u4e2d\u7b2cn \u500b\u97f3\u6846\u7684 c0 \u7279 \u5fb5\u503c\u4ee5 ( ) [ ] 0 x c n \u505a\u8868\u793a\uff0c\u800c ( ) [ ] 0 s c n \u8207 ( ) [ ] 0 d c n \u5206\u5225\u8868\u793a\u6b64\u97f3\u6846\u4e4b\u6240\u542b\u4e7e\u6de8\u8a9e\u97f3\u8a0a\u865f\u53ca\u7d14\u96dc\u8a0a \u7684 c0 \u7279\u5fb5\u503c\uff0c\u5247\u5b83\u5011\u53ef\u88ab\u63a8\u5c0e\u5982\u4e0b\u4e09\u5f0f\uff1a ( ) [ ] ( ) [ ] ( ) ( ) ( ) ( ) 0 log , log [ , ] [ , ] d x x s k k c n M k n M k n M k n = \u2248 + \u2211 \u2211 \uff0c \u5f0f(2-6) ( ) ( ) ( ) 0 [ ] log [ , ] s s k c n M k n = \u2211 \uff0c \u5f0f(2-7) ( ) ( ) ( ) 0 [ ] log [ , ] d d k c n M k n = \u2211 \uff0c \u5f0f(2-8) \u5176\u4e2d\uff0c ( ) [ , ] x M k n \u3001 ( ) [ , ] s M k n \u8207 ( ) [ , ] d M k n \u5206\u5225\u70ba\u5f0f(2-1)\u4e2d\u96dc\u8a0a\u8a9e\u97f3\u8a0a\u865f [ ] n x m \u3001\u4e7e\u6de8\u8a9e\u97f3 \u8a0a\u865f [ ] n s m \u4ee5\u53ca\u96dc\u8a0a [ ] n d m \u65bc\u8f49\u63db\u6210\u6885\u723e\u5012\u983b\u8b5c\u7279\u5fb5\u6642\uff0c\u7b2ck \u500b\u6885\u723e\uf984\u6ce2\u5668\u7684\u8f38\u51fa\u503c\u3002\u56e0 \u6b64\u6211\u5011\u53ef\u63a8\u5c0e\u51fa\uff0c\u7531\u65bc\u52a0\u6210\u6027\u96dc\u8a0a\u5e72\u64fe\u6240\u5c0e\u81f4\u96dc\u8a0a\u8a9e\u97f3\u8207\u4e7e\u6de8\u8a9e\u97f3\u8a0a\u865f\uf978\u8005\u4e4b c0 \u7279\u5fb5 \u503c\u7684\u5dee\uf962 [ ] 0 c n \u0394 \u5982\u4e0b\u5f0f\u6240\u793a\uff1a [ ] ( ) [ ] ( ) [ ] ( ) ( ) 0 0 0 [ , ] log 1 [ , ] d x s k s M k n c n c n c n M k n \u239b \u239e \u239f \u239c \u239f \u0394 = \u2212 \u2248 + \u239c \u239f \u239c \u239f \u239c \u239d \u23a0 \u2211 1 log 1 [ , ] k SNR k n \u239b \u239e \u239f \u239c \u239f = + \u239c \u239f \u239c \u239f \u239c \u239d \u23a0 \u2211 \uff0c \u5f0f(2-9) \u5f0f\u4e2d [ , ] SNR k n \u5b9a\u7fa9\u70ba\u7b2cn \u500b\u97f3\u6846\u4e2d\u7b2ck \u7dad\u6885\u723e\u983b\u5e36\u7684\u8a0a\u96dc\u6bd4\uff0c\u5373 ( ) ( ) [ , ] [ , ] [ , ] s d M k n SNR k n M k n = . \u5f0f(2-10) \u7531\u5f0f(2-9)\u53ef\u770b\u51fa\uff0c\uf974\u591a\uf969\u6885\u723e\u983b\u5e36\u7684\u8a0a\u96dc\u6bd4 [ , ] SNR k n \u90fd\u6bd4\u8f03\u5927\u6642\uff0c\u5dee\uf962\u503c 0 [ ] c n \u0394 \u4e5f \u76f8\u5c0d\u8b8a\u5c0f\uff0c\u56e0\u6b64\u9019\u53ef\u7d04\uf976\u89e3\u91cb\u542b\u8a9e\u97f3\u4e4b\u97f3\u6846(SNR \u8f03\u5927)\u76f8\u5c0d\u65bc\u7d14\u96dc\u8a0a\u97f3\u6846(SNR \u8f03\u5c0f)\u800c \u8a00\uff0cc0 \u7279\u5fb5\u503c\u8f03\uf967\uf9e0\u53d7\u5230\u5f71\u97ff\u7684\u73fe\u8c61\u3002 \u4ee5\u4e0b\u6211\u5011\u5c07\u63a2\u8a0e\u52a0\u6210\u6027\u96dc\u8a0a\u5c0d\u65bc c0 \u7279\u5fb5\u5e8f\uf99c\u4e4b\u8abf\u8b8a\u983b\u8b5c(modulation spectrum)\u4e0a\u7684 \u5f71\u97ff\u3002\u9996\u5148\u70ba\uf9ba\u63a8\u5c0e\u8d77\ufa0a\uff0c\u6211\u5011\u5c07\u5f0f(2-6)\u3001\u5f0f(2-7)\u8207\u5f0f(2-8)\u6539\u5beb\u6210\u4e0b\uf99c\u4e09\u5f0f\uff1a ( ) ( ) ( ) ( ) ( ) ( ) ( ) 0 [ ] [ , ] log exp [ , ] exp [ , ] x x s d k k c n M k n M k n M k n = \u2248 + \u2211 \u2211 , \u5f0f(2-11) ( ) ( ) 0 [ ] [ , ] s s k c n M k n = \u2211 , \u5f0f(2-12) ( ) ( ) 0 [ ] [ , ] d d k c n M k n = \u2211 , \u5f0f(2-13) \u5176\u4e2d ( ) ( ) ( ) [ , ] log [ , ] x x M k n M k n = \u3001 ( ) ( ) ( ) [ , ] log [ , ] s s M k n M k n = \u3001 ( ) ( ) ( ) [ , ] log [ , ] d d M k n M k n = \u3002\uf9d0\u4f3c \u5c07\u5f0f(2-11)\u8207\u5f0f(2-2)\u4f5c\u6bd4\u8f03\uff0c\u53ef\u770b\u51fa\u96dc\u8a0a\u8a9e\u97f3\u3001\u4e7e\u6de8\u8a9e\u97f3\u8207\u7d14\u96dc\u8a0a\u4e09\u8005\u7684\u95dc\u4fc2\u5728 logE \u8207 c0 \uf978\u7279\u5fb5\u4e2d\u5341\u5206\uf9d0\u4f3c\uff0c\u56e0\u6b64\u85c9\u7531\u524d\u9762\u4e4b\u5f0f(2-4)\u8207\u5f0f(2-5)\u5c0d\u65bc logE \u7279\u5fb5\u5e8f\uf99c\u4e4b\u8abf\u8b8a\u983b\u8b5c \u7684\u63a8\u5c0e\uff0c\u6211\u5011\u53ef\u4ee5\u767c\u73fe\u5c0d\u6bcf\u500b\u6885\u723e\uf984\u6ce2\u5668\u8f38\u51fa\u7684\u5c0d\uf969\u503c\u5e8f\uf99c ( ) [ ] { } , x M k n \u800c\u8a00\uff0c\u5176\u983b\u5bec\u4ecd \u662f\u5927\u65bc ( ) [ ] { } , d M k n \uff0c\u4e5f\u5c31\u662f\uf96f ( ) [ ] { } 0 x c n \u6bd4\u8d77 ( ) [ ] { } 0 d c n \u5c07\u64c1\u6709\uf901\u5927\u7684\u983b\u5bec\uff0c\u56e0\u6b64\uff0c\uf9d0\u4f3c logE \u7279\u5fb5\u7684\u7d50\u679c\uff0c\u6211\u5011\u540c\u6a23\u6b78\u7d0d\u51fa\u96dc\u8a0a\u8a9e\u97f3\u4e4b c0 \u7279\u5fb5\u5e8f\uf99c\u6bd4\u7d14\u96dc\u8a0a\u4e4b c0 \u7279\u5fb5\u5e8f\uf99c\u64c1\u6709\u8f03\u591a \u9ad8\u983b\u7684\u8abf\u8b8a\u983b\u8b5c\u6210\u4efd\uff0c\u4ea6\u5373\u524d\u8005\u6bd4\u5f8c\u8005\u6709\uf901\u660e\u986f\u7684\u4e0a\u4e0b\u632f\u76ea\u73fe\u8c61\u3002 \u5716\u4e09(a)\u8207\u5716\u4e09(b)\u5206\u5225\u70ba\u4e00\uf906\u8a9e\u97f3\u8a0a\u865f\u4e4b logE \u7279\u5fb5\u53ca c0 \u7279\u5fb5\u7684\u529f\uf961\u983b\u8b5c\u5bc6\ufa01(power spectral density, PSD) \u66f2 \u7dda \u5716 \uff0c \u5176 \u4e2d \u7684 \u8a9e \u97f3 \u8a0a \u865f \u53ca \u96dc \u8a0a \u70ba Aurora-2.0 \u8cc7 \uf9be \u5eab \u4e2d \u7684 \"FAC_5Z31ZZ4A\"\u6a94\u8207\u4eba\u8072\u96dc\u8a0a(babble noise)\uff0c\u8a0a\u96dc\u6bd4\u70ba 15dB\u3002\u7531\u9019\uf978\u5716\u6211\u5011\u53ef\u4ee5\u5f88\u660e \u986f\u5730\u770b\u51fa\uff0c\u96dc\u8a0a\u8a9e\u97f3\u76f8\u5c0d\u65bc\u7d14\u96dc\u8a0a\u800c\u8a00\uff0c\u5176 logE \u7279\u5fb5\u5e8f\uf99c\u8207 c0 \u7279\u5fb5\u5e8f\uf99c\u90fd\u6709\u8f03\u5927\u7684\u983b \u5bec\uff0c\u6b64\u4ea6\u9a57\u8b49\uf9ba\u6211\u5011\u4e4b\u524d\u7684\u63a8\u5c0e\u3002 \u5716\u4e09\u3001\u80fd\uf97e\u76f8\u95dc\u7279\u5fb5\u4e4b\u529f\uf961\u983b\u8b5c\u5bc6\ufa01\u5716\uff0c(a)\u70ba logE \u7279\u5fb5\u3001(b)\u70ba c0 \u7279\u5fb5 \u7d9c\u5408\u4e0a\u8ff0\u7684\u63a8\u5c0e\u53ca\u5716\uf9b5\uff0c\u6211\u5011\u9a57\u8b49\uf9ba\u4e00\u6bb5\u96dc\u8a0a\u8a9e\u97f3\u4e2d\u542b\u6709\u8a9e\u97f3\u7684\u97f3\u6846\u5176 logE \u7279\u5fb5 \u8207",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "",
                "sec_num": null
            },
            {
                "text": "\u6839\u64da\u5f0f(2-15)\u6240\u5f97\u4e4b [ ] y n \uff0c\u6211\u5011\u53ef\u4f5c\u4e00\u6bb5\u8a0a\u865f\u4e2d\u8a9e\u97f3\u8207\u975e\u8a9e\u97f3\u97f3\u6846\u7684\u5224\u5225\uff0c\u4e26\u9032\u800c \u5c07 \u5176 \u975e \u8a9e \u97f3 \u7684 \u97f3 \u6846 \u505a \u6b63 \u898f \u5316 \u8655 \uf9e4 \uff0c \u6b64 \u5373 \u70ba \u975c \u97f3 \u7279 \u5fb5 \u6b63 \u898f \u5316 \u6cd5 I (silence feature normalization I, SFN-I)\uff0c\u5176\u5f0f\u5982\u4e0b\uff1a SFN-I: [ ] [ ] ( ) [ ] [ ] if log if x n y n x n y n \u03b8 \u03b5 \u03b4 \u03b8 > \u23a7 \u23aa \u23aa = \u23a8 + \u23aa \u2264 \u23aa \u23a9 \uff0c \u5f0f(2-16) \u5176\u4e2d \u03b8 \u3001 \u03b5 \u8207 \u03b4 \u5206\u5225\u70ba\u9580\u6abb\u503c\u3001\u4e00\u6975\u5c0f\u7684\u6b63\uf969\u4ee5\u53ca\u4e00\u5e73\u5747\u503c\u70ba 0 \u4e14\u8b8a\uf962\uf969\u5f88\u5c0f\u7684\u96a8\u6a5f\u8b8a \uf969\uff0c [ ] x n \u70ba\u7d93\u904e SFN-I \u8655\uf9e4\u5f8c\u6240\u5f97\u5230\u7684\u65b0\u7279\u5fb5\uf96b\uf969\u3002\u5176\u9580\u6abb\u503c \u03b8 \u8a08\u7b97\u5f0f\u5982\u4e0b\uff1a [ ] 1 1 N n y n N \u03b8 = = \u2211 \uff0c \u5f0f(2-17) \u5f0f\u4e2d N \u70ba\u6b64\u6bb5\u8a9e\u97f3\u7684\u97f3\u6846\u7e3d\uf969\u3002\u56e0\u6b64\uff0c\u9580\u6abb\u503c\u5373\u70ba\u6574\u6bb5\u8a9e\u97f3\u6240\u6709 [ ] y",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "",
                "sec_num": null
            },
            {
                "text": "EQUATION",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [
                    {
                        "start": 0,
                        "end": 8,
                        "text": "EQUATION",
                        "ref_id": "EQREF",
                        "raw_str": "1/ 1 exp if if 1/ 1 exp y n y n w n y n y n \u03b8 \u03b2\u03c3 \u03b8 \u03b8 \u03b8 \u03b2\u03c3 \u23a7 \u23aa + \u2212 \u2212 \u23aa > \u23aa \u23aa = \u23a8 \u23aa \u2264 \u23aa + \u2212 \u2212 \u23aa \u23aa \u23a9 \uff0c \u5f0f(2-19) (a) (b) (c) (d) \u5716\u4e94\u3001\u975c\u97f3\u7279\u5fb5\u6b63\u898f\u5316\u6cd5 I \u8655\uf9e4\u524d((a)\u8207(b))\u8207\u8655\uf9e4\u5f8c((c)\u8207(d))\u80fd\uf97e\u76f8\u95dc\u7279\u5fb5\u5e8f\uf99c\u66f2\u7dda \u5716\uff0c\u5176\u4e2d(a)\u8207(c)\u70ba logE \u7279\u5fb5\u5e8f\uf99c\u66f2\u7dda\uff0c(b)\u8207(d)\u70ba c0 \u7279\u5fb5\u5e8f\uf99c\u66f2\u7dda \u5176\u4e2d [ ] y n \u5982\u524d\u4e00\u7bc0\u4e4b\u5f0f(2-15)\u6240\u793a\uff0c\u70ba [ ] { } x n \u901a\u904e\u4e00\u9ad8\u901a\uf984\u6ce2\u5668\u4e4b\u8f38\u51fa\u503c\uff0c\u03b8 \u70ba\u9580\u6abb\u503c\u3001 1 \u03c3 \u8207 2 \u03c3 \u5206\u5225\u70ba [ ] [ ] { } y n y n \u03b8 > (\u5927\u65bc\u9580\u6abb\u503c \u03b8 \u4e4b\u6240\u6709\u7684 [ ] y n )\u4ee5\u53ca [ ] [ ] { } y n y n \u03b8 \u2264 (\u5c0f\u65bc \u6216\u7b49\u65bc\u9580\u6abb\u503c \u03b8 \u4e4b\u6240\u6709\u7684 [ ] y n )\u6240\u5c0d\u61c9\u4e4b\u6a19\u6e96\u5dee\u3001\u03b2 \u70ba\u4e00\u5e38\uf969\u3002SFN-II \u4e4b\u9580\u6abb\u503c \u03b8 \u8ddf SFN-I \u76f8\u540c\uff0c\u8a08\u7b97\u5f0f\u5982\u4e0b\u6240\u793a\uff1a ( ) [ ] 1 1 N n N y n \u03b8 = = \u2211 , \u5f0f(2-20) \u5f0f\u4e2d N \u70ba\u6b64\u6bb5\u8a9e\u97f3\u4e2d\u97f3\u6846\u7e3d\uf969\u3002 \u5f0f(2-19)\u7684\u6b0a\u91cd\u503c [ ] w n \u5982\u5716\uf9d1\u6240\u793a\uff0c\u5176\u4e2d\u5047\u8a2d 0 \u03b8 = \u3001 1 1 \u03c3 = \u3001 2 3 \u03c3 = \u4ee5\u53ca 0.1 \u03b2 = \u3002 \u7531\u5716\uf9d1\u53ef\u4ee5\u767c\u73fe\uff0c\u6b0a\u91cd\u503c\u51fd\uf969 [ ] w n \u70ba\u4e00\u500b\u5de6\u53f3\uf967\u5c0d\u7a31\u4e4b\u905e\u589e\u7684 S \u5f62\u66f2\u7dda(sigmoid curve)\uff0c\u5176\u503c\u4ecb\u65bc 0 \u548c 1 \u4e4b\u9593\u3002\u6b64\u6b0a\u91cd\u503c\u6240\u4ee3\u8868\u7684\u610f\u7fa9\u8207 SFN-I \u6cd5\u76f8\u4f3c\uff0c\u6211\u5011\u5e0c\u671b\u65b0\u5f97 \u5230\u7684\u80fd\uf97e\u76f8\u95dc\u7279\u5fb5 [ ] x n \u80fd\u5728\u539f\u59cb\u7279\u5fb5\u503c\u5f88\u5927\u6642\uff0c\u5118\uf97e\u7dad\u6301\uf967\u8b8a\uff1b\u800c\u539f\u59cb\u503c\u8f03\u5c0f\u6642\uff0c\u5247\u4f7f \u5176\u8b8a\u5f97\uf901\u5c0f\u3002SFN-II \u6cd5\u548c SFN-I \u6cd5\uf967\u540c\u4e4b\u8655\u5728\u65bc\uff0cSFN-II \u6cd5\u5177\u6709\"\u8edf\u5f0f\"\u7684\u8a9e\u97f3\u7aef\u9ede\u5075\u6e2c \u6c7a\u7b56(soft-decision VAD)\uff0c\u800c SFN-I \u6cd5\u5247\u70ba\"\u786c\u5f0f\"\u7684\u8a9e\u97f3\u7aef\u9ede\u5075\u6e2c\u6c7a\u7b56(hard-decision VAD)\uff1b\u56e0\u6b64 SFN-II \u6cd5\u76f8\u8f03\u65bc SFN-I \u6cd5\u800c\u8a00\uff0c\u5176 VAD \u5224\u5b9a\u932f\u8aa4\u7684\u5f71\u97ff\u53ef\u80fd\u76f8\u5c0d\uf92d\u5f97\u6bd4 \u8f03\u5c0f\uff0c\u6548\u80fd\u4e5f\u6703\u6bd4\u8f03\u597d\uff0c\u9019\u63a8\u60f3\u5c07\u6703\u5728\u4e4b\u5f8c\u7684\u7ae0\u7bc0\u9a57\u8b49\u3002 \u5716\uf9d1\u3001\u6b0a\u91cd\u503c\u51fd\uf969 [ ] w n \u66f2\u7dda\u793a\u610f\u5716 \u5716\u4e03\u70ba SFN-II \u6cd5\u8655\uf9e4\u524d\u8207\u8655\uf9e4\u5f8c\u80fd\uf97e\u76f8\u95dc\u7279\u5fb5\u4e4b\u66f2\u7dda\u5716\u3002\u8207\u4e4b\u524d\u7684\u5716\u4e09\uf9d0\u4f3c\uff0c(a)\u8207(b) \u5206\u5225\u70ba\u539f\u59cb\u7684 logE \u7279\u5fb5\u5e8f\uf99c\u4ee5\u53ca c0 \u7279\u5fb5\u5e8f\uf99c\u66f2\u7dda\uff1b(c)\u8207(d)\u5206\u5225\u70ba\u7d93\u904e\u975c\u97f3\u7279\u5fb5\u6b63\u898f \u5316\u6cd5 II \u8655\uf9e4\u5f8c\u6240\u5f97\u5230\u4e4b logE \u5e8f\uf99c\u4ee5\u53ca c0 \u5e8f\uf99c\u66f2\u7dda\uff0c\u5176\u4e2d\u7d05\u8272\u5be6\u7dda\u662f\u5c0d\u61c9\u81f3\u4e7e\u6de8\u8a9e\u97f3 (Aurora-2.0 \u8cc7\uf9be\u5eab\u4e2d\u7684\"FAK_3Z82A\"\u6a94)\u3001\uf93d\u8272\u865b\u7dda\u8207\uf923\u8272\u9ede\u7dda\u5247\u5206\u5225\u70ba\u5c0d\u61c9\u81f3\u8a0a\u96dc\u6bd4 15dB \u8207 5dB \u7684\u96dc\u8a0a\u8a9e\u97f3\u3002\u5f88\u660e\u986f\u5730\uff0c\u7d93\u7531 SFN-II \u8655\uf9e4\u904e\u5f8c\u4e4b\u96dc\u8a0a\u8a9e\u97f3\u7684\u80fd\uf97e\u76f8\u95dc\u7279\u5fb5\uff0c \u7686\uf9d0\u4f3c SFN-I \u6cd5\u7684\u6548\u679c\uff0c\u53ef\u4ee5\uf901\u8da8\u8fd1\u65bc\u539f\u59cb\u4e7e\u6de8\u8a9e\u97f3\u4e4b\u7279\u5fb5\uff0c\u6709\u6548\ufa09\u4f4e\u96dc\u8a0a\u9020\u6210\u7684\u5931\u771f\u3002 (a)",
                        "eq_num": "( ) ( ) [ ] ( ) ( ) ( ) [ ] [ ] 1 2"
                    }
                ],
                "section": "",
                "sec_num": null
            }
        ],
        "back_matter": [],
        "bib_entries": {
            "BIBREF0": {
                "ref_id": "b0",
                "title": "Discriminative Analysis for Feature Reduction in Automatic Speech Recognition",
                "authors": [
                    {
                        "first": "E",
                        "middle": [
                            "L"
                        ],
                        "last": "Bocchieri",
                        "suffix": ""
                    },
                    {
                        "first": "J",
                        "middle": [
                            "G"
                        ],
                        "last": "Wilpon",
                        "suffix": ""
                    }
                ],
                "year": null,
                "venue": "1992 International Conference on Acoustics, Speech, and Signal Processing",
                "volume": "",
                "issue": "",
                "pages": "",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Bocchieri , E. L., and Wilpon, J. G., \"Discriminative Analysis for Feature Reduction in Automatic Speech Recognition\", 1992 International Conference on Acoustics, Speech, and Signal Processing (ICASSP 1992).",
                "links": null
            },
            "BIBREF1": {
                "ref_id": "b1",
                "title": "An Energy Search Approach to Variable Frame Rate Front-End Processing for Robust ASR",
                "authors": [
                    {
                        "first": "Julien",
                        "middle": [],
                        "last": "Epps",
                        "suffix": ""
                    },
                    {
                        "first": "Eric",
                        "middle": [
                            "H C"
                        ],
                        "last": "Choi",
                        "suffix": ""
                    }
                ],
                "year": 2005,
                "venue": "European Conference on Speech Communication and Technology",
                "volume": "",
                "issue": "",
                "pages": "",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Julien Epps and Eric H.C. Choi, \"An Energy Search Approach to Variable Frame Rate Front-End Processing for Robust ASR\", 2005 European Conference on Speech Communication and Technology (Interspeech 2005-Eurospeech).",
                "links": null
            },
            "BIBREF2": {
                "ref_id": "b2",
                "title": "Log-Energy Dynamic Range Normalization for Robust Speech Recognition",
                "authors": [
                    {
                        "first": "Weizhong",
                        "middle": [],
                        "last": "Zhu",
                        "suffix": ""
                    },
                    {
                        "first": "O'",
                        "middle": [],
                        "last": "Douglas",
                        "suffix": ""
                    },
                    {
                        "first": "",
                        "middle": [],
                        "last": "Shaughnessy",
                        "suffix": ""
                    }
                ],
                "year": 2005,
                "venue": "2005 International Conference on Acoustics, Speech, and Signal Processing",
                "volume": "",
                "issue": "",
                "pages": "",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Weizhong Zhu and Douglas O'Shaughnessy, \"Log-Energy Dynamic Range Normalization for Robust Speech Recognition\", 2005 International Conference on Acoustics, Speech, and Signal Processing (ICASSP 2005).",
                "links": null
            },
            "BIBREF3": {
                "ref_id": "b3",
                "title": "On the Study of Energy-Based Speech Feature Normalization and Application to Voice Activity Detection",
                "authors": [
                    {
                        "first": "Hung-Bin",
                        "middle": [],
                        "last": "Chen",
                        "suffix": ""
                    }
                ],
                "year": 2007,
                "venue": "",
                "volume": "",
                "issue": "",
                "pages": "",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Hung-Bin Chen, \"On the Study of Energy-Based Speech Feature Normalization and Application to Voice Activity Detection\", M.S. thesis, National Taiwan Normal University, Taiwan, 2007.",
                "links": null
            },
            "BIBREF4": {
                "ref_id": "b4",
                "title": "Silence Energy Normalization for Robust Speech Recognition in Additive Noise Environments",
                "authors": [
                    {
                        "first": "C-F",
                        "middle": [],
                        "last": "Tai",
                        "suffix": ""
                    },
                    {
                        "first": "J-W",
                        "middle": [],
                        "last": "Hung",
                        "suffix": ""
                    }
                ],
                "year": null,
                "venue": "2006 International Conference on Spoken Language Processing",
                "volume": "",
                "issue": "",
                "pages": "",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "C-F. Tai and J-W. Hung, \"Silence Energy Normalization for Robust Speech Recognition in Additive Noise Environments\", 2006 International Conference on Spoken Language Processing (Interspeech 2006-ICSLP).",
                "links": null
            },
            "BIBREF5": {
                "ref_id": "b5",
                "title": "Energy Contour Enhancement for Noisy Speech Recognition",
                "authors": [
                    {
                        "first": "Tai-Hwei",
                        "middle": [],
                        "last": "Hwang",
                        "suffix": ""
                    },
                    {
                        "first": "Sen-Chia",
                        "middle": [],
                        "last": "Chang",
                        "suffix": ""
                    }
                ],
                "year": 2004,
                "venue": "2004 International Symposium on Chinese Spoken Language Processing",
                "volume": "",
                "issue": "",
                "pages": "",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Tai-Hwei Hwang and Sen-Chia Chang, \"Energy Contour Enhancement for Noisy Speech Recognition\", 2004 International Symposium on Chinese Spoken Language Processing (ISCSLP 2004).",
                "links": null
            },
            "BIBREF6": {
                "ref_id": "b6",
                "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
            },
            "BIBREF7": {
                "ref_id": "b7",
                "title": "Multiband and Adaptation Approaches to Robust Speech Recognition",
                "authors": [
                    {
                        "first": "S",
                        "middle": [],
                        "last": "Tiberewala",
                        "suffix": ""
                    },
                    {
                        "first": "H",
                        "middle": [],
                        "last": "Hermansky",
                        "suffix": ""
                    }
                ],
                "year": 1997,
                "venue": "European Conference on Speech Communication and Technology",
                "volume": "",
                "issue": "",
                "pages": "",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "S. Tiberewala and H. Hermansky, \"Multiband and Adaptation Approaches to Robust Speech Recognition\", 1997 European Conference on Speech Communication and Technology (Eurospeech 1997)",
                "links": null
            },
            "BIBREF8": {
                "ref_id": "b8",
                "title": "MVA Processing of Speech Features",
                "authors": [
                    {
                        "first": "C-P",
                        "middle": [],
                        "last": "Chen",
                        "suffix": ""
                    },
                    {
                        "first": "J-A",
                        "middle": [],
                        "last": "Bilmes",
                        "suffix": ""
                    }
                ],
                "year": 2006,
                "venue": "Speech, and Language Processing",
                "volume": "",
                "issue": "",
                "pages": "",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "C-P. Chen and J-A. Bilmes, \"MVA Processing of Speech Features\", IEEE Trans. on Audio, Speech, and Language Processing, 2006",
                "links": null
            },
            "BIBREF9": {
                "ref_id": "b9",
                "title": "Non-Linear Transformations of the Feature Space for Robust Speech Recognition",
                "authors": [
                    {
                        "first": "A",
                        "middle": [],
                        "last": "Torre",
                        "suffix": ""
                    },
                    {
                        "first": "J",
                        "middle": [],
                        "last": "Segura",
                        "suffix": ""
                    },
                    {
                        "first": "C",
                        "middle": [],
                        "last": "Benitez",
                        "suffix": ""
                    },
                    {
                        "first": "A",
                        "middle": [
                            "M"
                        ],
                        "last": "Peinado",
                        "suffix": ""
                    },
                    {
                        "first": "A",
                        "middle": [
                            "J"
                        ],
                        "last": "Rubio",
                        "suffix": ""
                    }
                ],
                "year": 2002,
                "venue": "2002 International Conference on Acoustics, Speech and Signal Processing",
                "volume": "",
                "issue": "",
                "pages": "",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "A. Torre, J. Segura, C. Benitez, A. M. Peinado, and A. J. Rubio, \"Non-Linear Transformations of the Feature Space for Robust Speech Recognition\", 2002 International Conference on Acoustics, Speech and Signal Processing (ICASSP 2002)",
                "links": null
            }
        },
        "ref_entries": {
            "TABREF0": {
                "content": "<table><tr><td colspan=\"2\">\u672c\uf941\u6587\u5176\u5b83\u7ae0\u7bc0\u6982\u8981\u5982\u4e0b\uff1a\u5728\u7b2c\u4e8c\u7ae0\u4e2d\uff0c\u6211\u5011\u5148\u4e3b\u8981\u5c07\u5c0d\u80fd\uf97e\u76f8\u95dc\u7279\u5fb5\u53d7\u96dc\u8a0a\u5f71\u97ff\u7684\u6548</td></tr><tr><td colspan=\"2\">\u61c9\uff0c\u505a\u9032\u4e00\u6b65\u7684\u5206\u6790\u8207\u63a2\u8a0e\uff0c\u63a5\u8457\u4ecb\u7d39\u672c\uf941\u6587\u6240\u65b0\u63d0\u51fa\u7684\u4e4b\u975c\u97f3\u7279\u5fb5\u6b63\u898f\u5316\u6cd5(SFN)\uff1b</td></tr><tr><td colspan=\"2\">\u7b2c\u4e09\u7ae0\u5305\u542b\uf9ba\u5404\u7a2e\u91dd\u5c0d\u80fd\uf97e\u76f8\u95dc\u7279\u5fb5\u4e4b\u8655\uf9e4\u6280\u8853\u7684\u8a9e\u97f3\u8fa8\uf9fc\u5be6\u9a57\uf969\u64da\u53ca\u76f8\u95dc\u8a0e\uf941\uff0c\u5176\u4e2d</td></tr><tr><td colspan=\"2\">\u9664\uf9ba\u4ecb\u7d39\u8a9e\u97f3\u8fa8\uf9fc\u5be6\u9a57\u74b0\u5883\u5916\uff0c\u4e3b\u8981\u662f\u8a55\u4f30\u975c\u97f3\u7279\u5fb5\u6b63\u898f\u5316\u6cd5\u7684\u6548\u80fd\uff0c\u4e26\u8207\u5176\u4ed6\u65b9\u6cd5\u4f5c</td></tr><tr><td colspan=\"2\">\u6bd4\u8f03\uff0c\u85c9\u6b64\u9a57\u8b49\u6211\u5011\u6240\u63d0\u51fa\u65b0\u65b9\u6cd5\u80fd\u6709\u6548\u63d0\u5347\u80fd\uf97e\u76f8\u95dc\u7279\u5fb5\u5728\u96dc\u8a0a\u74b0\u5883\u4e0b\u7684\u5f37\u5065\u6027\u3002\u5728</td></tr><tr><td colspan=\"2\">\u7b2c\u56db\u7ae0\u4e2d\uff0c\u6211\u5011\u5617\u8a66\u5c07\u6240\u63d0\u7684\u65b0\u65b9\u6cd5\u7d50\u5408\u5176\u5b83\u7684\u5f37\u5065\u6027\u7279\u5fb5\u6280\u8853\uff0c\u5c0d\u6b64\uf9d0\u7684\u7d50\u5408\u4f5c\u8fa8\uf9fc</td></tr><tr><td colspan=\"2\">\u5be6\u9a57\u6240\u5f97\u5230\u7684\u8fa8\uf9fc\uf961\u52a0\u4ee5\u63a2\u8a0e\u8207\u5206\u6790\uff0c\u4ee5\u9a57\u8b49\u6211\u5011\u6240\u63d0\u51fa\u7684\u975c\u97f3\u7279\u5fb5\u6b63\u898f\u5316\u6cd5\u662f\u5426\u8207\u5176 \u4e00\u3001\u7dd2\uf941 \u8fd1\uf98e\uf92d\u79d1\u6280\u767c\u5c55\u8fc5\u901f\uff0c\u4f46\u662f\u81ea\u52d5\u8a9e\u97f3\u8fa8\uf9fc\u4ecd\u7136\u662f\u4e00\u9580\u76f8\u7576\u5177\u6709\u6311\u6230\u6027\u7684\u8ab2\u984c\u3002\u901a\u5e38 \u5b83\u6280\u8853\u6709\uf97c\u597d\u7684\u52a0\u6210\u6027\u3002\u7b2c\u4e94\u7ae0\u5247\u70ba\u672c\uf941\u6587\u7d50\uf941\u8207\u672a\uf92d\u5c55\u671b\u3002</td></tr><tr><td colspan=\"2\">\u4e00\u81ea\u52d5\u8a9e\u97f3\u8fa8\uf9fc\u7cfb\u7d71\u5728\uf967\u53d7\u5916\u5728\u96dc\u8a0a\u5e72\u64fe\u7684\u7814\u7a76\u5ba4\u74b0\u5883\u4e0b\uff0c\u90fd\u53ef\u4ee5\u7372\u5f97\u6975\u9ad8\u7684\u8fa8\uf9fc\u6548 \u4e8c\u3001\u975c\u97f3\u7279\u5fb5\u6b63\u898f\u5316\u6cd5 \u80fd\uff0c\u4f46\uf974\u662f\u61c9\u7528\u5230\u5be6\u969b\u7684\u74b0\u5883\u4e2d\uff0c\u7cfb\u7d71\u8fa8\uf9fc\u6548\u80fd\u5247\u901a\u5e38\u6703\u5927\u5e45\ufa09\u4f4e\uff0c\u9019\u4e3b\u8981\u662f\u88ab\u73fe\u5be6\u74b0 \u9996\u5148\uff0c\u6211\u5011\u5728\u7b2c\u4e00\u7bc0\u4e2d\uff0c\u91dd\u5c0d\u8a9e\u97f3\u80fd\uf97e\u76f8\u95dc\u7279\u5fb5\uff1a\u5c0d\uf969\u80fd\uf97e(logarithmic energy, logE) \u5883\u4e2d\u8a31\u591a\u7684\u8b8a\uf962\u6027(variation)\u6240\u5f71\u97ff\u3002\u800c\u8a9e\u97f3\u8fa8\uf9fc\u7684\u8b8a\uf962\u6027\u7a2e\uf9d0\u7e41\u591a\uff0c\uf9b5\u5982\u8a13\uf996\u74b0\u5883\u8207 \u8857 \u9053 \u4e7e \u6de8 \u8a9e \u97f3 \u8a0a \u865f \u96dc \u8a0a \u8a9e \u97f3 \u8a0a \u865f \u8207\u7b2c\uf9b2\u7dad\u5012\u983b\u8b5c\u4fc2\uf969(c0)\u53d7\u5230\u74b0\u5883\u96dc\u8a0a\u5e72\u64fe\u7684\u8b8a\uf962\u73fe\u8c61\u505a\u8f03\u6df1\u5165\u7684\u89c0\u5bdf\u5206\u6790\u8207\u63a2\u8a0e\uff0c\u63a5 \u8457\u5728\u7b2c\u4e8c\u7bc0\u4e2d\uff0c\u6211\u5011\u6839\u64da\u9019\u4e9b\u7d50\u679c\uff0c\u63d0\u51fa\u975c\u97f3\u7279\u5fb5\u6b63\u898f\u5316\u6cd5\u7684\u65b0\u5f37\u5065\u6027\u6280\u8853\u3002 (\u4e00)\u5c0d\uf969\u80fd\uf97e\u7279\u5fb5\u53ca\u7b2c\uf9b2\u7dad\u5012\u983b\u8b5c\u7279\u5fb5\u4fc2\uf969\u53d7\u52a0\u6210\u6027\u96dc\u8a0a\u5e72\u64fe\u4e4b\u73fe\u8c61\u7684\u63a2\u8a0e \u6e2c\u8a66\u74b0\u5883\u9593\u5b58\u5728\u7684\u74b0\u5883\uf967\u5339\u914d(environmental mismatch) \u6a5f \u5834 \u52a0\u6210\u6027\u96dc\u8a0a\u5c0d\u65bc\u80fd\uf97e\u76f8\u95dc\u7279\u5fb5(logE \u8207 c0)\u9020\u6210\u7684\u6548\u61c9\u53ef\u7531\u5716\u4e8c\u770b\u51fa\u7aef\u502a\u3002\u5716\u4e8c(a</td></tr><tr><td>\u52a0 \u6210 \u6027 \u96dc \u8a0a</td><td>\u647a \u7a4d \u6027 \u96dc \u8a0a</td></tr><tr><td colspan=\"2\">\u5716\u4e00\u3001\u4e7e\u6de8\u8a9e\u97f3\u53d7\u96dc\u8a0a\u5e72\u64fe\u4e4b\u793a\u610f\u5716</td></tr><tr><td colspan=\"2\">\u672c\uf941\u6587\u662f\u4ee5\u4e0a\u8ff0\u6240\u63d0\u53ca\u7684\u74b0\u5883\uf967\u5339\u914d\u4e2d\u7684\u52a0\u6210\u6027\u96dc\u8a0a\u56e0\u7d20\uff0c\u4f5c\u70ba\u4e3b\u8981\u63a2\u8a0e\u7684\u4e3b\u984c\uff0c</td></tr><tr><td colspan=\"2\">\u4ee5\u671f\u5c07\u52a0\u6210\u6027\u96dc\u8a0a\u5c0d\u8a9e\u97f3\u8fa8\uf9fc\u7684\u5f71\u97ff\ufa09\u4f4e\u3002\u5728\u7279\u5fb5\uf96b\uf969\u62bd\u53d6\u6b65\u9a5f\u6642\uff0c\u6211\u5011\u7d93\u5e38\u8a08\u7b97\u8a9e\u97f3</td></tr><tr><td colspan=\"2\">\u7684\u80fd\uf97e\u503c\u4f5c\u70ba\u7279\u5fb5\u4e4b\u4e00\uff1b\u6839\u64da\u904e\u53bb\u7684\u6587\u737b\u6307\u51fa[1][2]\u7686\u50be\u5411\u65bc\u5c07\u975e\u8a9e\u97f3\u90e8\u5206\u7684\u5c0d\uf969\u80fd\uf97e\uf969\u503c\u8abf\u4f4e\uff0c\u4e26\u5c07\u8a9e\u97f3\u90e8\u5206\u7684\u5c0d\uf969\u80fd\uf97e\u503c\u4fdd\u6301\uf967\u8b8a\uff1b\u5176</td></tr><tr><td colspan=\"2\">\u4e3b\u8981\u7684\u539f\u56e0\u662f\u4e00\u6bb5\u8a9e\u97f3\u7279\u5fb5\u4e2d\uff0c\u80fd\uf97e\u8f03\u4f4e\u7684\u90e8\u5206\u901a\u5e38\u6703\u6bd4\u80fd\uf97e\u8f03\u9ad8\u7684\u90e8\u5206\uf901\u5bb9\uf9e0\u53d7\u5230\u96dc</td></tr><tr><td colspan=\"2\">\u8a0a\u7684\u5f71\u97ff\u3002\u672c\uf941\u6587\u4f9d\u64da\u524d\u4eba\u6240\u767c\u8868\u7684\u6587\u737b\u52a0\u4ee5\u6539\u9032\uff0c\u4e14\u91dd\u5c0d\u8a9e\u97f3\u8a0a\u865f\u80fd\uf97e\u76f8\u95dc\u7684\u7279\u5fb5\u5982</td></tr><tr><td colspan=\"2\">\u4f55\u53d7\u5230\u96dc\u8a0a\u5f71\u97ff\uff0c\u4ee5\u8f03\u56b4\u8b39\u7684\uf969\u5b78\uf9e4\uf941\u52a0\u4ee5\u5206\u6790\uff0c\u4e26\u63d0\u51fa\u4e00\u5957\u65b0\u7684\u5f37\u5065\u6280\u8853\uff0c\u7a31\u70ba\u300c\u975c</td></tr><tr><td colspan=\"2\">\u97f3\u7279\u5fb5\u6b63\u898f\u5316\u6cd5\u300d (silence feature normalization, SFN)\uff0c\u6b64\u65b9\u6cd5\u53ef\u4ee5\u6709\u6548\u5730\ufa09\u4f4e\u52a0\u6210\u6027\u96dc</td></tr><tr><td colspan=\"2\">\u8a0a\u5c0d\u8a9e\u97f3\u80fd\uf97e\u76f8\u95dc\u7279\u5fb5\u7684\u5e72\u64fe\uff0c\u9032\u800c\u63d0\u9ad8\u7cfb\u7d71\u7684\u8fa8\uf9fc\u6548\u80fd\u3002</td></tr></table>",
                "html": null,
                "text": "\u3001\u8a9e\u8005\u8b8a\uf962(speaker variation)\u4ee5\u53ca \u767c\u97f3\u7684\u8b8a\uf962(pronunciation variation)\u7b49\u3002\u5c0d\u65bc\u74b0\u5883\uf967\u5339\u914d\u800c\u8a00\uff0c\u5176\u76f8\u95dc\u7684\u8b8a\uf969\u53ef\u6982\uf976\u5206\u70ba \u4e0b\uf99c\u5e7e\u9805\uf9d0\u578b\uff1a\u52a0\u6210\u6027\u96dc\u8a0a(additive noise)\u3001\u647a\u7a4d\u6027\u96dc\u8a0a(convolutional noise)\u4ee5\u53ca\u983b\u5bec\u7684 \u9650\u5236(bandwidth limitation)\u7b49\u3002\u5716\u4e00\u70ba\u4e7e\u6de8\u8a9e\u97f3\u8a0a\u865f\u53d7\u5230\u96dc\u8a0a\u5e72\u64fe\u4e4b\u793a\u610f\u5716\u3002 \uff0c\u8a9e\u97f3\u8a0a\u865f\u7684\u80fd\uf97e\u7279\u5fb5(energy feature) \u6240\u5305\u542b\u7684\u8fa8\uf9fc\u8cc7\u8a0a\u5927\u904e\u65bc\u5176\u5b83\u7279\u5fb5\uff0c\u4e14\u80fd\uf97e\u7279\u5fb5\u7684\u8a08\u7b97\u8907\u96dc\ufa01\u5f88\u4f4e\u3002\u6240\u4ee5\u6839\u64da\u4e0a\u8ff0\u80fd\uf97e \u7279\u5fb5\u7684\u512a\u52e2\uff0c\u5728\u672c\uf941\u6587\u4e2d\uff0c\u6211\u5011\u7279\u5225\u5c0d\u5176\u5f37\u5065\u6027\u6280\u8853\u52a0\u4ee5\u5206\u6790\u3001\u8a0e\uf941\u8207\u767c\u5c55\u3002 \u8fd1\uf98e\uf92d\uff0c\u6709\u8a31\u591a\u6210\u529f\u7684\u5f37\u5065\u6027\u5c0d\uf969\u80fd\uf97e\u7279\u5fb5(logarithmic energy, logE)\u7684\u6280\u8853\u76f8\u7e7c\u88ab\u63d0 \u51fa \uff0c \uf9b5 \u5982 \uff0c \u5c0d \uf969 \u80fd \uf97e \u52d5 \u614b \u7bc4 \u570d \u6b63 \u898f \u5316 \u6cd5 (log-energy dynamic range normalization, LEDRN)[3]\u5176\u76ee\u6a19\u662f\u4f7f\u8a13\uf996\u8207\u6e2c\u8a66\u7684\u8a9e\u97f3\u8cc7\uf9be\u5176\u5c0d\uf969\u80fd\uf97e\u503c\u4e4b\u52d5\u614b\u7bc4\u570d\u4e00\u81f4\u5316\uff1b\u5c0d\uf969\u80fd \uf97e\u5c3a\ufa01\u91cd\u523b\u6cd5(log-energy rescaling normalization, LERN)[4]\u5247\u662f\u5c07\u5c0d\uf969\u80fd\uf97e\u7279\u5fb5\u4e58\u4e0a\u4e00 \u500b\u4ecb\u65bc 0 \u8207 1 \u9593\u7684\u6b0a\u91cd\u503c\uff0c\u8a66\u5716\u91cd\u5efa\u51fa\u4e7e\u6de8\u8a9e\u97f3\u7684\u5c0d\uf969\u80fd\uf97e\u7279\u5fb5\uff1b\u800c\u672c\u5be6\u9a57\u5ba4\u5148\u524d\u6240\u63d0 \u51fa\u7684\u975c\u97f3\u97f3\u6846\u5c0d\uf969\u80fd\uf97e\u6b63\u898f\u5316\u6cd5(silence energy normalization, SLEN)[5]\uff0c\u662f\u5c07\u5224\u5225\u70ba\u975e \u8a9e\u97f3\u97f3\u6846(non-speech frame)\u7684\u5c0d\uf969\u80fd\uf97e\u7279\u5fb5\u8a2d\u5b9a\u70ba\u4e00\u6975\u5c0f\u503c\u7684\u5e38\uf969\u3002\u4e0a\u8ff0\u7684\u4e09\u7a2e\u65b9\u6cd5\uff0c",
                "type_str": "table",
                "num": null
            },
            "TABREF2": {
                "content": "<table><tr><td/><td/><td/><td/><td/><td>n \u7684\u5e73\u5747\u503c\uff0c\u5176\u8a08</td></tr><tr><td colspan=\"6\">\u7b97\u5341\u5206\u7c21\uf965\uff0c\u4e14\u7121\u9700\u984d\u5916\u7279\u5225\u8a2d\u8a08\u4e4b\u8655\u3002</td></tr><tr><td colspan=\"6\">\u5f9e\u5f0f(2-16)\u770b\u51fa\uff0c\uf974 [ ] y n \u5927\u65bc\u9580\u6abb\u503c \u03b8 \uff0c\u5247\u5c07\u5176\u6240\u5c0d\u61c9\u4e4b\u97f3\u6846\u5224\u65b7\u70ba\u8a9e\u97f3\uff0c\u4e14\u539f\u7279</td></tr><tr><td colspan=\"6\">\u5fb5\uf96b\uf969\u4fdd\u6301\uf967\u8b8a\uff1b\u53cd\u4e4b\u5247\u5c07\u5176\u6b78\uf9d0\u70ba\u975e\u8a9e\u97f3\u97f3\u6846\uff0c\u4e26\u5c07\u539f\u7279\u5fb5\uf96b\uf969\u6b63\u898f\u5316\u6210\u4e00\u6975\u5c0f\u7684\u96a8</td></tr><tr><td colspan=\"6\">\u6a5f\u8b8a\uf969\uff1b\u76f8\u8f03\u65bc\u4e4b\u524d\u975c\u97f3\u97f3\u6846\u5c0d\uf969\u80fd\uf97e\u6b63\u898f\u5316\u6cd5(SLEN)[5]\u800c\u8a00\uff0c\u975c\u97f3\u7279\u5fb5\u6b63\u898f\u5316\u6cd5 I</td></tr><tr><td colspan=\"6\">\u53ef\u907f\u514d\u5c07\u975e\u8a9e\u97f3\u90e8\u4efd\u7684\u7279\u5fb5\u6b63\u898f\u5316\u70ba\u4e00\u5b9a\u503c\uff0c\u800c\u53ef\u80fd\u5c0e\u81f4\u4e4b\u5f8c\u6240\u8a13\uf996\u7684\u8072\u5b78\u6a21\u578b\u4e2d\u7684\u8b8a</td></tr><tr><td colspan=\"6\">\uf962\uf969(variance)\u8b8a\u70ba 0 { x n \u4e58\u4e0a\u4e00\u6b0a\u91cd\u503c(weight)\uff0c\u800c\u5f97\u5230\u65b0\u7279\u5fb5\u503c }</td></tr><tr><td colspan=\"6\">[ ] } x n \u3002SFN-II \u7684\u6f14\u7b97\u6cd5\u5982\u4e0b\u5f0f\u6240\u793a\uff1a {</td></tr><tr><td>SFN-II:</td><td>[ ] x n</td><td>=</td><td>[ ] [ ] w n x n</td><td>,</td><td>\u5f0f(2-18)</td></tr><tr><td>\u5176\u4e2d\uff0c</td><td/><td/><td/><td/><td/></tr><tr><td/><td/><td/><td>(</td><td>[ ]</td><td>)</td></tr><tr><td>[ ]</td><td/><td/><td/><td/><td/></tr></table>",
                "html": null,
                "text": "\u7684\u932f\u8aa4\u73fe\u8c61\u7522\u751f\u3002\u6211\u5011\u53ef\u4ee5\u900f\u904e\u5716\u4e94\uf92d\u89c0\u5bdf SFN-I \u6cd5\u7684\u4f5c\u7528\u3002\u5716\u4e94 \u4e2d\uff0c(a)\u8207(b)\u5206\u5225\u70ba\u539f\u59cb\u7684 logE \u7279\u5fb5\u5e8f\uf99c\u4ee5\u53ca c0 \u7279\u5fb5\u5e8f\uf99c\u66f2\u7dda\uff1b(c)\u8207(d)\u5206\u5225\u70ba\u7d93\u904e\u975c \u97f3\u7279\u5fb5\u6b63\u898f\u5316\u6cd5 I \u8655\uf9e4\u5f8c\u6240\u5f97\u5230\u4e4b logE \u7279\u5fb5\u5e8f\uf99c\u4ee5\u53ca c0 \u7279\u5fb5\u5e8f\uf99c\u66f2\u7dda\uff0c\u5176\u4e2d\u7d05\u8272\u5be6\u7dda \u662f\u5c0d\u61c9\u81f3\u4e7e\u6de8\u8a9e\u97f3(Aurora-2.0 \u8cc7\uf9be\u5eab\u4e2d\u7684\"FAK_3Z82A\"\u6a94)\u3001\uf93d\u8272\u865b\u7dda\u8207\uf923\u8272\u9ede\u7dda\u5247\u5206 \u5225\u70ba\u5c0d\u61c9\u81f3\u8a0a\u96dc\u6bd4 15dB \u8207 5dB \u7684\u96dc\u8a0a\u8a9e\u97f3\u3002\u7531\u9019\u4e9b\u5716\u660e\u986f\u5730\u770b\u51fa\uff0cSFN-I \u6cd5\u8655\uf9e4\u904e\u5f8c \u4e4b\u80fd\uf97e\u76f8\u95dc\u7279\u5fb5\u503c\u53ef\u4ee5\u8f03\u8da8\u8fd1\u65bc\u539f\u59cb\u4e7e\u6de8\u8a9e\u97f3\u8a0a\u865f\u4e4b\u7279\u5fb5\u503c\uff0c\u9054\u5230\ufa09\u4f4e\u5931\u771f\u7684\u76ee\u7684\u3002 (\u4e09)\u975c\u97f3\u7279\u5fb5\u6b63\u898f\u5316\u6cd5 II (silence feature normalization II, SFN-II) \u5728\u672c\u7bc0\u4e2d\uff0c\u6211\u5011\u5c07\u4ecb\u7d39\u7b2c\u4e8c\u7a2e\u6a21\u5f0f\u7684\u975c\u97f3\u7279\u5fb5\u6b63\u898f\u5316\u6cd5\uff0c\u7a31\u4e4b\u70ba\u300c\u975c\u97f3\u7279\u5fb5\u6b63\u898f\u5316 \u6cd5 II\u300d (silence feature normalization II, SFN-II)\uff0cSFN-II \u6cd5\u8207\u524d\u4e00\u7bc0\u4e4b SFN-I \u6cd5\u6700\u5927\u7684\u5dee \uf962\u5728\u65bc\uff0cSFN-II \u662f\u5c07\u539f\u80fd\uf97e\u76f8\u95dc\u7279\u5fb5 [ ]",
                "type_str": "table",
                "num": null
            },
            "TABREF3": {
                "content": "<table><tr><td>\u53ca\u975c\u97f3\u6a21\u578b\uff0c\u6bcf\u500b\uf969\u5b57\u6a21\u578b\u5305\u542b 16 \u500b\uf9fa\u614b(states)\uff0c\u800c\u6bcf\u500b\uf9fa\u614b\u662f\u7531 20 \u500b\u9ad8\u65af\u5bc6\ufa01\u51fd\uf969 \u56db\u3001\u975c\u97f3\u7279\u5fb5\u6b63\u898f\u5316\u6cd5\u8207\u5176\u5b83\u7279\u5fb5\u5f37\u5065\u6cd5\u7d50\u5408\u4e4b\u5be6\u9a57\u7d50\u679c\u8207\u8a0e\uf941 \u5fb5(logE \u8207 c0)\u56e0\u52a0\u6210\u6027\u96dc\u8a0a\u9020\u6210\u7684\u5931\u771f\u73fe\u8c61\u4f5c\u9069\u7576\u7684\u88dc\u511f\u3002SFN \u6cd5\uf9dd\u7528\uf9ba\u4e00\u500b\u9ad8\u901a\uf984\u6ce2</td></tr><tr><td>\u6df7\u5408(Gaussian mixtures)\u6240\u7d44\u6210\u3002 \u524d\u4e00\u7ae0\u4e4b\u4e00\u7cfb\uf99c\u7684\u5be6\u9a57\uff0c\u4e3b\u8981\u662f\u63a2\u8a0e\u5404\u7a2e\u80fd\uf97e\u76f8\u95dc\u7279\u5fb5\u8655\uf9e4\u6280\u8853\u6548\u80fd\uff0c\u9032\u800c\u7a81\u986f\u51fa \u5668\u53bb\u8655\uf9e4\u539f\u59cb\u80fd\uf97e\u76f8\u95dc\u7279\u5fb5\u5e8f\uf99c\uff0c\u4e26\u5c07\u901a\u904e\u6b64\u9ad8\u901a\uf984\u6ce2\u5668\u6240\u7684\u4e4b\u8f38\u51fa\u7279\u5fb5\u5e8f\uf99c\u62ff\uf92d\u4f5c\u8a9e</td></tr><tr><td>\u8868\u4e00\u3001\u672c\uf941\u6587\u4e2d\u6240\u4f7f\u7528\u4e4b\u8a9e\u97f3\u7279\u5fb5\uf96b\uf969\u8a2d\u5b9a \u6211\u5011\u6240\u65b0\u63d0\u51fa\u4e4b\u975c\u97f3\u7279\u5fb5\u6b63\u898f\u5316(SFN)\u6cd5\u7684\u512a\uf962\u8868\u73fe\uff0c\u9019\u4e9b\u5be6\u9a57\u4e2d\uff0c\u53ea\u6709 logE \u8207 c0 \uf978 \u97f3/\u975e\u8a9e\u97f3\u7684\u5206\uf9d0\uff0c\u4e26\u61c9\u7528\u7c21\u55ae\u4e14\u6709\u6548\u7684\u65b9\u6cd5\uf92d\u8655\uf9e4\u975e\u8a9e\u97f3\u90e8\u4efd\u7684\u7279\u5fb5\uff0c\u5c07\u96dc\u8a0a\u5c0d\u8a9e\u97f3</td></tr><tr><td>\u53d6\u6a23\u983b\uf961 \u7a2e\u80fd\uf97e\u76f8\u95dc\u7279\u5fb5\u88ab\u8655\uf9e4\uff0c\u5269\u9918\u7684\u6885\u723e\u5012\u983b\u8b5c\u7279\u5fb5\u4fc2\uf969(c1~c12)\u5247\u7dad\u6301\uf967\u8b8a\u3002\u5728\u9019\u4e00\u7ae0\u4e2d\uff0c 8kHz \u7279\u5fb5\u7684\u5e72\u64fe\ufa09\u4f4e\uff0c\u4ee5\u671f\u63d0\u5347\u8a13\uf996\u8207\u6e2c\u8a66\u74b0\u5883\u5339\u914d\ufa01\uff0c\u9032\u800c\u63d0\u5347\u96dc\u8a0a\u74b0\u5883\u4e0b\u7684\u8a9e\u97f3\u8fa8\uf9fc\uf961\u3002</td></tr><tr><td>(c) \u5716\u4e03\u3001\u975c\u97f3\u7279\u5fb5\u6b63\u898f\u5316\u6cd5 II \u8655\uf9e4\u524d((a)\u8207(b))\u8207\u8655\uf9e4\u5f8c((c)\u8207(d))\u80fd\uf97e\u76f8\u95dc\u7279\u5fb5\u5e8f\uf99c\u66f2\u7dda (d) \u5716\uff0c\u5176\u4e2d(a)\u8207(c)\u70ba logE \u7279\u5fb5\u5e8f\uf99c\u66f2\u7dda\uff0c(b)\u8207(d)\u70ba c0 \u7279\u5fb5\u5e8f\uf99c\u66f2\u7dda \u4e09\u3001\u80fd\uf97e\u76f8\u95dc\u7279\u5fb5\u8655\uf9e4\u6280\u8853\u4e4b\u5be6\u9a57\u7d50\u679c\u8207\u8a0e\uf941 (\u4e00) \u3001\u8a9e\u97f3\u8cc7\uf9be\u5eab\u7c21\u4ecb \u672c\uf941\u6587\u4e2d\u7684\u8a9e\u97f3\u8fa8\uf9fc\u5be6\u9a57\u6240\u4f7f\u7528\u7684\u8a9e\u97f3\u8cc7\uf9be\u5eab\u70ba\u6b50\u6d32\u96fb\u4fe1\u6a19\u6e96\u5354\u6703(European Telecommunication Standard Institute, ETSI)\u767c\ufa08\u7684 Aurora-2.0 \u8a9e\uf9be\u5eab[7]\u3002\u5b83\u662f\u4e00\u5957\u85c9\u7531 \u4ee5\u4eba\u5de5\u7684\u65b9\u5f0f\uf93f\u88fd\u7684\uf99a\u7e8c\u82f1\u6587\uf969\u5b57\u5b57\uf905\uff0c\u8a9e\u8005\u70ba\u7f8e\u570b\u6210\uf98e\u7537\uf981\uff0c\u52a0\u4e0a\u516b\u7a2e\u52a0\u6210\u6027\u96dc\u8a0a\uff0c \u5206\u5225\u70ba\u5730\u4e0b\u9435\u3001\u4eba\u8072\u3001\u6c7d\uf902\u3001\u5c55\u89bd\u9928\u3001\u9910\u5ef3\u3001\u8857\u9053\u3001\u6a5f\u5834\u3001\u706b\uf902\u7ad9\u7b49\uff0c\u4ee5\u53ca\uf967\u540c\u7a0b\ufa01\u7684 \u8a0a\u96dc\u6bd4\uff0c\u5206\u5225\u70ba 20dB\u300115dB\u300110dB\u30015dB\u30010dB \u4ee5\u53ca-5dB\uff0c\u9644\u52a0\u4e0a\u4e7e\u6de8(clean)\u8a9e\uf9be\u3002 (\u4e8c) \u3001\u7279\u5fb5\uf96b\uf969\u7684\u8a2d\u5b9a\u8207\u8fa8\uf9fc\u7cfb\u7d71\u7684\u8a13\uf996 \u672c\uf941\u6587\u6839\u64da Aurora-2.0 \u5be6\u9a57\u8a9e\uf9be\u5eab\u6a19\u6e96\u8a2d\u5b9a[7] \uff0c\u8a9e\u97f3\u7279\u5fb5\uf96b\uf969\u4e3b\u8981\u662f\u4f7f\u7528\u6885\u723e\u5012\u983b \u8b5c\u4fc2\uf969(mel-frequency cepstral coefficients, MFCC)\u53ca\u80fd\uf97e\u76f8\u95dc\u7279\u5fb5\uff0c\u9644\u52a0\u4e0a\u5176\u4e00\u968e\u5dee\uf97e\u8207 \u4e8c\u968e\u5dee\uf97e\u3002\u70ba\uf9ba\u5206\u6790\u80fd\uf97e\u76f8\u95dc\u7279\u5fb5\u7684\u5f71\u97ff\uff0c\u65bc\u672c\uf941\u6587\u4e2d\u63a1\u7528\uf978\u7d44\uf967\u540c\u7684\u7279\u5fb5\uf96b\uf969\uff1b\u7b2c\u4e00 \u7d44\u662f 12 \u7dad\u6885\u723e\u5012\u983b\u8b5c\u7279\u5fb5\u503c(c1\uff5ec12)\u52a0\u4e0a 1 \u7dad\u7684\u5c0d\uf969\u80fd\uf97e(logE)\uff0c\u53e6\u4e00\u7d44\u5247\u662f\u4f7f\u7528 12 \u7dad\u6885\u723e\u5012\u983b\u8b5c\u7279\u5fb5\u503c(c1\uff5ec12)\u52a0\u4e0a\u7b2c\uf9b2\u7dad\u5012\u983b\u8b5c\u7279\u5fb5\u4fc2\uf969(c0)\uff1b\u800c\u6bcf\u7d44\u7686\u6703\u518d\u52a0\u4e0a\u4e00\u968e \u8207\u4e8c\u968e\u5dee\uf97e\uff0c\u6545\uf978\u7d44\u7686\u7528\uf9ba 39 \u7dad\u7684\u7279\u5fb5\uf96b\uf969\u3002\u8a73\u7d30\u7684\u7279\u5fb5\uf96b\uf969\u8a2d\u5b9a\uff0c\u5982\u8868\u4e00\u6240\u793a\u3002 modified SFN-II 85.03 86.06 85.55 15.46 51.68 \u6211\u5011\uf9dd\u7528 HTK \u7a0b\u5f0f[8]\uf92d\u8a13\uf996\u8072\u5b78\u6a21\u578b\uff0c\u7522\u751f\uf9ba 11(oh, zero, one~nine)\u500b\uf969\u5b57\u6a21\u578b\u4ee5 \u97f3\u6846\u9577\ufa01(Frame Size) 25ms, 200 \u9ede \u97f3\u6846\u5e73\u79fb(frame Shift) 10ms, 80 \u9ede \u9810\u5f37\u8abf\uf984\u6ce2\u5668 1 1 0.97z \u2212 \u2212 \u8996\u7a97\u5f62\u5f0f \u6f22\u660e\u7a97(Hamming window) \u5085\uf9f7\uf96e\u8f49\u63db\u9ede\uf969 256 \u9ede \uf984\u6ce2\u5668\u7d44(filters) \u6885\u723e\u523b\ufa01\u4e09\u89d2\uf984\u6ce2\u5668\u7d44\uff0c \u5171 23 \u500b\u4e09\u89d2\uf984\u6ce2\u5668 \u7279\u5fb5\u5411\uf97e (feature vector) \u7b2c\u4e00\u7d44\uff1a { } 1 12 i c i \u2264 \u2264 , 0 c , 0 c \u0394 , 2 0 c \u0394 \u5171\u8a08 39 \u7dad (\u4e09)\u8a9e\u97f3\u8fa8\uf9fc\u5be6\u9a57\u7d50\u679c \u5728\u9019\u4e00\u7bc0\u4e2d\uff0c\u6211\u5011\u5c07\u57f7\ufa08\u5404\u7a2e\u91dd\u5c0d\u80fd\uf97e\u76f8\u95dc\u7279\u5fb5\u4e4b\u5f37\u5065\u6027\u6280\u8853\u7684\u8a9e\u97f3\u8fa8\uf9fc\uff0c\u4e26\u6bd4\u8f03 \u5176\u6548\u80fd\u3002\u9664\uf9ba\u6211\u5011\u6240\u65b0\u63d0\u51fa\u7684\u975c\u97f3\u7279\u5fb5\u6b63\u898f\u5316\u6cd5(SFN-I\u8207SFN-II)\u5916\uff0c\u6211\u5011\u540c\u6642\u5be6\u9a57 \uf9ba\u5e73\u5747\u8207\u8b8a\uf962\uf969\u6b63\u898f\u5316\u6cd5(mean and variance normalization, MVN)[9]\u3001\u5e73\u5747\u8207\u8b8a\uf962\uf969\u6b63\u898f \u5316\u9644\u52a0ARMA\uf984\u6ce2\u5668\u6cd5(MVN plus ARMA filtering, MVA)[10]\u3001\u7d71\u8a08\u5716\u7b49\u5316\u6cd5(histogram equalization, HEQ)[11] \u3001 \u5c0d \uf969 \u80fd \uf97e \u52d5 \u614b \u7bc4 \u570d \u6b63 \u898f \u5316 \u6cd5 (log-energy dynamic range normalization, LEDRN)[3]\u3001\u5c0d\uf969\u80fd\uf97e\u5c3a\ufa01\u91cd\u523b\u6cd5 (log-energy rescaling normalization, LERN)[4]\u8207\u975c\u97f3\u5c0d\uf969\u80fd\uf97e\u6b63\u898f\u5316\u6cd5(silence log-energy normalization, SLEN)[5] \uff0c\u503c\u5f97\u6ce8\u610f \u7684\u662f\uff0c\u539f\u59cb\u4e4bMVN\u3001MVA\u8207HEQ\u4e09\u65b9\u6cd5\u96d6\u662f\u8a2d\u8a08\u65bc\u6240\u6709\u7a2e\uf9d0\u7684\u7279\u5fb5\u4e0a\uff0c\u6211\u5011\u70ba\uf9ba\u8a55\u4f30 \u5176\u5728\u80fd\uf97e\u76f8\u95dc\u7279\u5fb5\u7684\u6548\u80fd\uff0c\u5728\u9019\uf9e8\u53ea\u5c07\u5b83\u5011\u904b\u7528\u65bclogE\u8207c0\u7279\u5fb5\u7684\u6b63\u898f\u5316\u4e0a\uff0c\u53e6\u5916\uff0c LEDRN\u6cd5\u6709\u5206\u7dda\u6027\u8207\u975e\u7dda\u6027\uf978\u7a2e\uff0c\u5728\u9019\uf9e8\u6211\u5011\u5206\u5225\u4ee5LEDRN-I\u8207LEDRN-II\u8868\u793a\uff0c\u800c LERN\u4ea6\u6709\uf978\u7a2e\u7248\u672c\uff0c\u6211\u5011\u5206\u5225\u4ee5LERN-I\u8207LERN-II\u8868\u793a\u3002 1\u3001\u91dd\u5c0d\u5c0d\uf969\u80fd\uf97e\u7279\u5fb5(logE)\u4e4b\u5f37\u5065\u5f0f\u8a9e\u97f3\u6280\u8853\u7d9c\u5408\u5206\u6790 \u6b64\u5c0f\u7bc0\u4e4b\u5be6\u9a57\u6240\u7528\u5230\u8a9e\u97f3\u7279\u5fb5\u70ba\u524d\u8ff0\u4e4b\u7b2c\u4e00\u7d44\u7684\u7279\u5fb5\uf96b\uf969\uff0c\u5373 12 \u7dad\u6885\u723e\u5012\u983b\u8b5c\u7279 \u5fb5\u503c(c1\uff5ec12)\u52a0\u4e0a 1 Method Set A Set B average AR RR (1) Baseline 71.98 67.79 69.89 \u2500 \u2500 (2) MVN 79.04 81.08 80.06 10.18 33.79 (3) MVA 80.53 82.64 81.59 11.70 38.85 (4) HEQ 83.91 85.79 84.85 14.97 49.69 (5) LEDRN-I 82.01 79.70 80.86 10.97 36.43 (6) LEDRN-II 77.21 75.53 76.37 6.49 21.53 (7) LERN-I 83.64 83.35 83.50 13.61 45.19 (8) LERN-II 82.71 81.94 82.33 12.44 41.31 (9) SLEN 84.87 85.27 85.07 15.19 50.42 (10) SFN-I 85.02 85.50 85.26 15.38 51.05 (11) SFN-II 85.67 86.32 86.00 16.11 53.49 2\u3001\u91dd\u5c0d\u7b2c\uf9b2\u7dad\u5012\u983b\u8b5c\u7279\u5fb5\u4fc2\uf969(c0)\u4e4b\u5f37\u5065\u5f0f\u8a9e\u97f3\u6280\u8853\u7d9c\u5408\u5206\u6790 \u6b64\u5c0f\u7bc0\u4e4b\u5be6\u9a57\u6240\u7528\u5230\u8a9e\u97f3\u7279\u5fb5\u70ba\u524d\u8ff0\u4e4b\u7b2c\u4e8c\u7d44\u7684\u7279\u5fb5\uf96b\uf969\uff0c\u5373 12 \u7dad\u6885\u723e\u5012\u983b\u8b5c\u7279 \u5fb5\u503c(c1\uff5ec12)\u52a0\u4e0a\u7b2c\uf9b2\u7dad\u5012\u983b\u8b5c\u7279\u5fb5\u4fc2\uf969(c0)\uff0c\u9644\u52a0\u5176\u4e00\u968e\u8207\u4e8c\u968e\u5dee\uf97e\uff0c\u5171 39 \u7dad\u3002\uf9d0 \u4f3c\u524d\u4e00\u5c0f\u7bc0\uff0c\u6211\u5011\u5c07\u539f\u59cb\u91dd\u5c0d logE \u7279\u5fb5\u7684\u5341\u7a2e\u7279\u5fb5\u5f37\u5065\u6027\u65b9\u6cd5\uff0c\u4f5c\u7528\u65bc c0 \u7279\u5fb5\u4e0a\uff0c\u5176 \u5b83 12 \u7dad\u7684\u6885\u723e\u5012\u983b\u8b5c\u4fc2\uf969\u5247\u7dad\u6301\uf967\u8b8a\u3002\u96d6\u7136\u76ee\u524d\u8655\uf9e4\u7684\u662f c0 \u7279\u5fb5\uff0c\u4f46\u70ba\uf9ba\u7c21\u660e\u8d77\ufa0a\uff0c \u9019\uf9e8\u6211\u5011\uf967\u5c07\u539f\u672c\u5404\u7a2e\u6280\u8853\u7684\u540d\u7a31\u4f5c\u4fee\u6539\uff0c\uf9b5\u5982 LEDRN \u6cd5\uff0c\u6211\u5011\u4e26\uf967\u7279\u5225\u5c07\u5176\u6539\u540d\u70ba c0-DRN \u6cd5\uff0c\u800c\u4ecd\u6cbf\u8972\u5176\u540d\uff0c\u5176\u4ed6\u65b9\u6cd5\u540d\u7a31\u4f9d\u6b64\uf9d0\u63a8\u3002 \u8868\u4e09\uf99c\u51fa\uf9ba\u57fa\u790e\u5be6\u9a57\u53ca\u9019\u5341\u7a2e\u65b9\u6cd5\u6240\u5f97\u4e4b\u5e73\u5747\u8fa8\uf9fc\uf961(20dB\u300115dB\u300110dB\u30015dB \u8207 0dB \u4e94\u7a2e\u8a0a\u96dc\u6bd4\u4e0b\u7684\u8fa8\uf9fc\uf961\u5e73\u5747) \uff0c\u800c\u5176\u4e2d\u7684 AR \u8207 RR \u5206\u5225\u70ba\u76f8\u8f03\u65bc\u57fa\u790e\u5be6\u9a57\u7d50\u679c\u4e4b\u7d55\u5c0d \u932f\u8aa4\ufa09\u4f4e\uf961\u548c\u76f8\u5c0d\u932f\u8aa4\ufa09\u4f4e\uf961\u3002\u5f9e\u8868\u4e09\u7684\uf969\u64da\uff0c\u6211\u5011\u53ef\u89c0\u5bdf\u5230\u4e0b\uf99c\u5e7e\u9ede\u73fe\u8c61\uff1a \u25cb 1 \uf9d0\u4f3c\u4e4b\u524d\u7684\u8868\u4e8c\u4e4b\u7d50\u679c\uff0c\u5404\u7a2e\u65b9\u6cd5\u4f5c\u7528\u65bc c0 \u7279\u5fb5\u6642\uff0c\u90fd\u80fd\u5e36\uf92d\u63d0\u6607\u8fa8\uf9fc\uf961\u7684\u6548\u679c\uff0c \u5176\u4e2d\uff0cLEDRN-I \u8207 LEDRN-II \u7684\u8868\u73fe\u6bd4\u5176\u4ed6\u65b9\u6cd5\u7a0d\u5dee\uff0c\u5c24\u5176\u662f LEDRN-II\uff0c\u53ea\u6709 3.57% \u4e4b \u7d55\u5c0d\u932f\u8aa4\ufa09\u4f4e\uf961(AR)\uff0c\u5176\u53ef\u80fd\u539f\u56e0\u5728\u65bc\uff0cLEDRN \u539f\u672c\u662f\u91dd\u5c0d logE \u7279\u5fb5\u6240\u8a2d\u8a08\uff0c\uf974\u6211\u5011 \u76f4\u63a5\u5c07\u5176\u5957\u7528\u65bc c0 \u7279\u5fb5\u8655\uf9e4\u4e0a\uff0c\u5176\u6240\u4f7f\u7528\u7684\uf96b\uf969\u4e26\u975e\u662f\u6700\u4f73\u5316\u800c\u5f97\uff0c\u5c0e\u81f4\u6548\u679c\uf967\u5f70\u3002 modified SFN-I 84.54 85.79 85.17 15.08 50.41 SFN-II 83.29 85.14 84.22 14.13 47.23 \u5728\u672c\uf941\u6587\u4e2d\uff0c\u6211\u5011\u63d0\u51fa\u4e00\u500b\u65b0\u7684\u8a9e\u97f3\u5f37\u5065\u6280\u8853 -\u300c\u975c\u97f3\u7279\u5fb5\u6b63\u898f\u5316\u6cd5\u300d(silence feature normalization, SFN)\uff0c\u6b64\u65b9\u6cd5\u57f7\ufa08\u4e0a\u5341\u5206\u7c21\uf9e0\u4e14\u6548\u679c\u512a\uf962\u3002\u5b83\u662f\u91dd\u5c0d\u80fd\uf97e\u76f8\u95dc\u7279 \u25cb 2 \u5011\u6240\u63d0\u7684\uf978\u500b\u65b0\u65b9\u6cd5\uff0c\u80fd\u6709\u6548\u5730\u63d0\u6607 c0 \u7279\u5fb5\u5728\u52a0\u6210\u6027\u96dc\u8a0a\u74b0\u5883\u4e0b\u7684\u5f37\u5065\u6027\u3002 \u8868\u4e09\u3001\u91dd\u5c0d c0 \u7279\u5fb5\u4e4b\u5f37\u5065\u5f0f\u8a9e\u97f3\u6280\u8853\u4e4b\u8fa8\uf9fc\uf961\u7684\u7d9c\u5408\u6bd4\u8f03\u8868(%) Method Set A Set B Average AR RR (1) Baseline 71.95 68.22 70.09 \u2500 \u2500 (2) MVN 80.80 82.95 81.88 11.79 39.41 (3) MVA 81.76 84.04 82.90 12.82 42.84 (4) HEQ 82.89 84.59 83.74 13.66 45.65 (5) LEDRN-I 79.04 77.36 78.20 8.11 27.13 (6) LEDRN-II 74.08 73.22 73.65 3.57 11.92 (7) LERN-I 83.81 83.65 83.73 13.65 45.61 (8) LERN-II 83.03 82.53 82.78 12.70 42.44 (9) SLEN 82.94 84.28 83.61 13.53 45.21 (10) SFN-I 83.04 84.70 83.87 13.79 46.08 (11) SFN-II 83.29 85.14 84.22 14.13 47.23 c0 \u4e0a\uff0c\u4e26\u5c0d\u8a9e\u97f3\u8207\u975e\u8a9e\u97f3\u97f3\u6846\u7684 c0 \u7279\u5fb5\u5e8f\uf99c\u6c42\u53d6\u5176\u5f0f(2-19)\u6240\u7528\u7684 \u6a19\u6e96\u5dee 1 \u03c3 \u8207 2 \u03c3 \uff0c\u7136\u5f8c\u4f5c\u5f0f(2-18)\u4e4b\u6b63\u898f\u5316\u8655\uf9e4\u3002\u6211\u5011\u5c07\u4ee5\u4e0a\u7684\u4fee\u6b63\u4f5c\u6cd5\u5206\u5225\u7a31\u4f5c\u91dd\u5c0d c0 \u7279\u5fb5\u4e4b\u4fee\u6b63\u5f0f SFN-I \u6cd5(modified SFN-I)\u8207\u4fee\u6b63\u5f0f SFN-II \u6cd5(modified SFN-II)\u3002 Method Set A Set B Average AR RR Baseline 71.95 68.22 70.09 \u2500 \u2500 SFN-I 83.04 84.70 83.87 13.79 46.08 \u6211\u5011\u5617\u8a66\u5c07\u4f5c\u7528\u65bc logE \u8207 c0 \u7279\u5fb5\u7684 SFN \u6cd5\u8207\u4f5c\u7528\u65bc c1~c12 \u4e4b\u6885\u723e\u5012\u983b\u8b5c\u7279\u5fb5\u4fc2\uf969\u7684 \u5f37\u5065\u6027\u6280\u8853\u52a0\u4ee5\u7d50\u5408\uff0c\u85c9\u4ee5\u89c0\u5bdf\uf978\u8005\u4e4b\u9593\u662f\u5426\u6709\u52a0\u6210\u6027\uff0c\u80fd\u9032\u4e00\u6b65\u6539\u9032\u8a9e\u97f3\u8fa8\uf9fc\uf961\u3002 \u5728\u9019\uf9e8\uff0c\u6211\u5011\u9078\u64c7\u4e4b\u524d\u6240\u63d0\u4e4b MVN[9]\u3001MVA[10]\u4ee5\u53ca HEQ[11]\u4e09\u7a2e\u5f37\u5065\u6027\u6280\u8853\uff0c \u5206\u5225\u4f5c\u7528\u65bc c1~c12 \u4e4b\u6885\u723e\u5012\u983b\u8b5c\u7279\u5fb5\u4fc2\uf969\u4e0a\uff0c\u800c\u5c07\u6211\u5011\u6240\u63d0\u4e4b SFN-I \u6216 SFN-II \u6cd5\u4f5c\u7528 \u65bc\u80fd\uf97e\u76f8\u95dc\u7279\u5fb5(logE \u6216 c0)\u4e0a\uff0c\u6211\u5011\u5c07\u5176\u4e0a\u8ff0\u6240\u6709\u7684\u5be6\u9a57\u7d50\u679c\u5206\u5225\u5f59\u6574\u6210\u8868\u4e94\u8207\u8868\uf9d1\u3002 \u91dd\u5c0d\u7b2c\u4e00\u7d44\u7279\u5fb5(logE, c1~c12)\u8655\uf9e4\u4e4b\u8868\u4e94\u7684\uf969\u64da\u4e2d\uff0c\uf99c(2)~(4)\u662f\uf9dd\u7528\u55ae\u4e00\u5f37\u5065\u6280\u8853 (MVN, MVA \u6216 HEQ)\u8655\uf9e4\u5168\u90e8\u7279\u5fb5\uf96b\uf969\u4e4b\u7d50\u679c\uff0c\u800c\uf99c(5)~(10)\u5247\u5206\u5225\u70ba\u975c\u97f3\u7279\u5fb5\u6b63\u898f\u5316 \u6cd5(SFN)\u7d50\u5408\u5176\u5b83\u65b9\u6cd5\u4e4b\u7d50\u679c\u3002\u7576\u6211\u5011\u5c07\uf99c(2)\u3001\uf99c(5)\u8207\uf99c(8)\u7684\u7d50\u679c\u76f8\u6bd4\u8f03\u3001\uf99c(3)\u3001\uf99c(6) \u8207\uf99c(9)\u7684\u7d50\u679c\u76f8\u6bd4\u8f03\uff0c\u53ca\uf99c(4)\u3001\uf99c(7)\u8207\uf99c(10)\u7684\u7d50\u679c\u76f8\u6bd4\u8f03\uff0c\u90fd\u53ef\u4ee5\u770b\u51fa\u5c07 SFN-I \u6216 SFN-II \u4f7f\u7528\u65bc logE \u7279\u5fb5\uff0c\u4e26\u7528\u5176\u4ed6\u65b9\u6cd5\u4f7f\u7528\u5728 c1\uff5ec12 \u7279\u5fb5\u4e0a\uff0c\u6240\u5f97\u5230\u7684\u8fa8\uf9fc\uf961\u6bd4\u55ae \u7368\u4f7f\u7528\u4e00\u7a2e\u65b9\u6cd5\u8655\uf9e4\u5168\u90e8\u7279\u5fb5\u7684\u8fa8\uf9fc\u7d50\u679c\u9ad8\u51fa\u8a31\u591a\uff0c\uf9b5\u5982\uf99c(9)\u4e4b\u300eSFN-II (logE) + MVA (c1~c12)\u300f\u6cd5\uff0c\u5176\u5e73\u5747\u8fa8\uf9fc\uf961\u9ad8\u9054 89.97%\uff0c\u8d85\u8d8a\uf9ba\uf99c(4)\u4e4b\u300eHEQ (logE, c1~c12)\u300f\u6cd5\u6240 \u5f97\u4e4b 87.\u6216 HEQ \u6cd5\u7684\u78ba\u5177\u6709\u52a0\u6210\u6027\u3002 \u8868\u4e94\u3001SFN \u6cd5\u4f5c\u7528\u5728 logE \u7279\u5fb5\u7d50\u5408\u5176\u5b83\u8a9e\u97f3\u5f37\u5065\u6280\u8853\u4f5c\u7528\u65bc c1\uff5ec12 \u7279\u5fb5\uf96b\uf969\u4e4b\u5e73\u5747 \u8fa8\uf9fc\uf961\u7684\u7d9c\u5408\u6bd4\u8f03\u8868(%) Method Set A Set B average AR RR (1) Baseline 71.98 67.79 69.89 \u2500 \u2500 (2) MVN (logE, c1~c12) 83.55 83.75 83.65 13.77 45.71 (3) MVA (logE, c1~c12) 86.69 86.89 86.79 16.91 56.13 (4) HEQ (logE, c1~c12) 87.15 87.72 87.44 17.55 58.28 (5) SFN-I (logE) + MVN (c1~c12) 87.33 87.81 87.57 17.69 58.72 (6) SFN-I (logE) + MVA (c1~c12) 88.40 88.84 88.62 18.74 62.21 (7) SFN-I (logE) + HEQ (c1~c12) 87.93 88.04 87.99 18.10 60.10 (8) SFN-II (logE) + MVN (c1~c12) 88.45 88.88 88.67 18.78 62.36 (9) SFN-II (logE) + MVA (c1~c12) 89.82 90.12 89.97 20.09 66.69 (10) SFN-II (logE) + HEQ (c1~c12) 89.29 89.33 89.31 19.43 64.50 \u91dd\u5c0d\u7b2c\u4e8c\u7d44\u7279\u5fb5(c0, c1~c12)\u8655\uf9e4\u4e4b\u8868\uf9d1\u7684\uf969\u64da\u4e2d\uff0c\uf99c(2)~(4)\u662f\uf9dd\u7528\u55ae\u4e00\u5f37\u5065\u6280\u8853 (MVN, MVA \u6216 HEQ)\u8655\uf9e4\u5168\u90e8\u7279\u5fb5\uf96b\uf969\u4e4b\u7d50\u679c\uff0c\u800c\uf99c(5)~(16)\u5247\u5206\u5225\u70ba\u975c\u97f3\u7279\u5fb5\u6b63\u898f\u5316 \u6cd5(SFN)\u7d50\u5408\u5176\u5b83\u65b9\u6cd5\u4e4b\u7d50\u679c\u3002\uf9d0\u4f3c\u8868\u4e94\u4e2d\uf99c(1)~(10)\u6240\u5448\u73fe\u7684\u7d50\u679c\uff0c\u5f9e\u8868\uf9d1\u4e2d\u4e4b\uf99c \uf9fc\uf961\u7d9c\u5408\u6bd4\u8f03\u8868(%) Method Set A Set B Average AR RR (1) Baseline 71.95 68.22 70.09 \u2500 \u2500 (11) modified SFN-I (c0) + MVN (c1~c12) 87.49 87.89 87.69 17.61 58.85 (12) modified SFN-I (c0) + MVA (c1~c12) 89.30 89.54 89.42 19.34 64.63 (13) modified SFN-I (c0) + HEQ (c1~c12) 88.10 88.39 88.25 18.16 60.71 (14) modified SFN-II (c0) + MVN (c1~c12) 88.25 88.33 88.29 18.21 60.86 (15) modified SFN-II (c0) + MVA (c1~c12) 89.87 89.98 89.93 19.84 66.32 (16) modified SFN-II (c0) + HEQ (c1~c12) 89.25 89.46 89.36 19.27 64.42 \u5c07\u8868\uf9d1\u4e4b\uf99c(11)~(16)\u7684\uf969\u64da\u8207\uf99c(1)~(10)\u76f8\u6bd4\u8f03\uff0c\u6211\u5011\u53ef\u4ee5\u660e\u986f\u770b\u51fa\u91dd\u5c0d c0 \u7279\u5fb5\u4e4b\u4fee\u6b63 \u5f0f \u4e94\u3001\u7d50\uf941 \u7531\u5be6\u9a57\uf969\u64da\u4e2d\u53ef\u767c\u73fe\uff0c\u5c31\u8655\uf9e4\u80fd\uf97e\u76f8\u95dc\u7279\u5fb5\u800c\u8a00\uff0cSFN \u6cd5\u6bd4\u57fa\u672c\u5be6\u9a57\u4ee5\u53ca\u8a31\u591a\u5f37\u5065 \u5f0f\u8a9e\u97f3\u6280\u8853\u5f97\u5230\uf901\u597d\u7684\u8fa8\uf9fc\uf961\uff1b\u7531\u6b64\u53ef\u77e5\u91dd\u5c0d\u80fd\uf97e\u76f8\u95dc\u7279\u5fb5\u505a\u9069\u7576\u7684\u88dc\u511f\uff0c\u5728\u7a69\u5b9a\u4ee5\u53ca \u975e\u7a69\u5b9a\u96dc\u8a0a\u74b0\u5883\u4e0b\u7686\u5f97\u5230\u5341\u5206\u986f\u8457\u7684\u8fa8\uf9fc\uf961\u63d0\u5347\uff0c\u986f\u793a\uf9ba\u80fd\uf97e\u76f8\u95dc\u7279\u5fb5\u6240\u542b\u7684\u8a9e\u97f3\u9451\u5225 \u8cc7\u8a0a\u662f\u5f71\u97ff\u8fa8\uf9fc\uf961\u7684\u4e00\u500b\u91cd\u8981\u6307\u6a19\u3002\u6b64\u5916\uff0c\u7576\u6211\u5011\u5c07 SFN \u6cd5\u8207\u5176\u5b83\u5f37\u5065\u5f0f\u8a9e\u97f3\u6280\u8853\u505a \u7d50\u5408\uff0c\u767c\u73fe\u5176\u8fa8\uf9fc\uf961\u6bd4\u55ae\u7368\u4f7f\u7528\u4e00\u7a2e\u5f37\u5065\u5f0f\u8a9e\u97f3\u6280\u8853\u6240\u5f97\u5230\u7684\u8fa8\uf9fc\uf961\uf901\u9ad8\uff0c\u5176\u4e2d\u53c8\u4ee5 (1)~(10)\u8868\uf9d1\u3001SFN \u6cd5\u4f5c\u7528\u5728 c0 \u7279\u5fb5\u7d50\u5408\u5176\u5b83\u8a9e\u97f3\u5f37\u5065\u6280\u8853\u4f5c\u7528\u65bc c1\uff5ec12 \u7279\u5fb5\uf96b\uf969\u4e4b\u5e73\u5747\u8fa8 MVN (c0, c1~c12) 85.03 85.54 85.29 15.20 50.81 SFN-II \u6cd5\u7d50\u5408 MVA \u6cd5\u5f97\u5230\u7684\u8fa8\uf9fc\uf961\u6700\u9ad8\uff0c\u53ef\u9054\u5230\u5c07\u8fd1 90%\u7684\u5e73\u5747\u8fa8\uf9fc\uf961\u3002 (2) (3) MVA (c0, c1~c12) 88.11 88.81 88.46 18.38 \u80fd\uf97e\u76f8\u95dc\u7279\u5fb5\u96d6\u7136\u5177\u9ad8\ufa01\u8a9e\u97f3\u9451\u5225\uf98a\uff0c\u4f46\u662f\u96dc\u8a0a\u5c0d\u5176\u5e72\u64fe\u7a0b\ufa01\u4e5f\u76f8\u5c0d\u5f88\u5927\uff0c\u56e0\u6b64\u80fd 61.42 (4) HEQ (c0, c1~c12) 86.99 88.13 87.56 17.48 \uf97e\u76f8\u95dc\u7279\u5fb5\u8655\uf9e4\u7684\u597d\u58de\uff0c\u5c07\u6703\u5f88\u76f4\u63a5\u5730\u5f71\u97ff\u5230\u7cfb\u7d71\u7684\u8fa8\uf9fc\u6548\u80fd\uff0c\u7531\u6b64\u53ef\u77e5\u80fd\uf97e\u76f8\u95dc\u7279\u5fb5 58.42 (5) SFN-I (c0) + MVN (c1~c12) 85.62 86.62 86.12 16.04 \u7684\u5f37\u5065\u5316\u8655\uf9e4\u5728\u672a\uf92d\u4ecd\u662f\u503c\u5f97\u63a2\u8a0e\u7684\u4e00\u5927\u8ab2\u984c\uff1b\u6211\u5011\u5e0c\u671b\u672a\uf92d\u53ef\u4ee5\u5c07\u6240\u767c\u5c55\u7684\u6280\u8853\uff0c\u64f4 53.60 (6) SFN-I (c0) + MVA (c1~c12) 87.38* 88.16* 87.77* 17.69 \u5c55\u6e2c\u8a66\u81f3\u5176\u5b83\u8f03\u5927\u5b57\u5f59\uf97e\u7684\u8a9e\u97f3\u8fa8\uf9fc\u7cfb\u7d71\u4e0a\uff0c\u63a2\u8a0e\u9019\uf9d0\u6280\u8853\u5728\uf967\u540c\u8907\u96dc\ufa01\u4e4b\u8a9e\u97f3\u8fa8\uf9fc\u7cfb 59.12 (7) SFN-I (c0) + HEQ (c1~c12) 85.95* 86.53* 86.24* 16.16 \u7d71\u7684\u6548\u80fd\u3002\u53e6\u5916\uff0c\u672a\uf92d\u6211\u5011\u4ecd\u53ef\u671d\u5411\u6d88\u9664\u52a0\u6210\u6027\u96dc\u8a0a\u7684\u65b9\u5411\u7e7c\u7e8c\u6df1\u5165\u7814\u7a76\uff0c\u4e5f\u53ef\u4ee5\u91dd\u5c0d 54.00 (8) SFN-II (c0) + MVN (c1~c12) 86.92 87.69 87.31 17.22 \u6d88\u9664\u901a\u9053\u6027\u96dc\u8a0a\u7684\u65b9\u6cd5\u53bb\u4f5c\u76f8\u95dc\u7684\u63a2\u8a0e\uff0c\u4e26\u5617\u8a66\u5c07\uf978\u8005\u7d50\u5408\uff0c\u4f7f\u5f97\u8a9e\u97f3\u8fa8\u8a8d\u7cfb\u7d71\u80fd\uf901\u6709 57.56 (9) SFN-II (c0) + MVA (c1~c12) 89.04 89.61 89.33 19.24 \u6548\u5730\ufa09\u4f4e\u5404\uf9d0\u96dc\u8a0a\u7684\u5e72\u64fe\uff0c\u800c\u64c1\u6709\uf9a8\u4eba\u6eff\u610f\u4e4b\u8fa8\uf9fc\uf961\u3002 64.32 (10) SFN-II (c0) + HEQ (c1~c12) 87.43 87.88* 87.66 17.57 58.73 \uf96b\u8003\u6587\u737b</td></tr></table>",
                "html": null,
                "text": "\u7dad\u7684\u5c0d\uf969\u80fd\uf97e(logE)\uff0c\u9644\u52a0\u5176\u4e00\u968e\u8207\u4e8c\u968e\u5dee\uf97e\uff0c\u5171 39 \u7dad\u3002\u800c\u9019\uf9e8 \u6240\u7528\u5230\u7684\u5341\u7a2e\u7279\u5fb5\u5f37\u5065\u6027\u65b9\u6cd5\uff0c\u7686\u662f\u55ae\u7d14\u8655\uf9e4 logE \u7279\u5fb5\uff0c\uf967\u8003\u616e\u5176\u5b83 12 \u7dad\u7684\u6885\u723e\u5012\u983b \u8b5c\u4fc2\uf969\uff0c\u8868\u4e8c\uf99c\u51fa\uf9ba\u57fa\u790e\u5be6\u9a57\u53ca\u9019\u5341\u7a2e\u65b9\u6cd5\u6240\u5f97\u4e4b\u5e73\u5747\u8fa8\uf9fc\uf961(20dB\u300115dB\u300110dB\u3001 5dB \u8207 0dB \u4e94\u7a2e\u8a0a\u96dc\u6bd4\u4e0b\u7684\u8fa8\uf9fc\uf961\u5e73\u5747) \uff0c\u5176\u4e2d AR \u8207 RR \u5206\u5225\u70ba\u76f8\u8f03\u57fa\u790e\u7d50\u679c\u4e4b\u7d55\u5c0d \u932f\u8aa4\ufa09\u4f4e\uf961(absolute error rate reduction)\u548c\u76f8\u5c0d\u932f\u8aa4\ufa09\u4f4e\uf961(relative error rate reduction)\u3002 \u5f9e\u8868\u4e8c\u7684\uf969\u64da\uff0c\u6211\u5011\u53ef\u89c0\u5bdf\u5230\u4e0b\uf99c\u5e7e\u9ede\u73fe\u8c61\uff1a \u25cb 1 \u539f\u59cb\u4f5c\u7528\u65bc\u6240\u6709\u7a2e\uf9d0\u7279\u5fb5\u4e4b MVN\u3001MVA \u8207 HEQ \u6cd5\u55ae\u7d14\u4f5c\u7528\u65bc logE \u7279\u5fb5\u6642\uff0c \u5176\u63d0\u4f9b\u7684\u6539\u9032\u6548\u679c\u4e5f\u5341\u5206\u660e\u986f\uff0c\u76f8\u8f03\u65bc\u57fa\u790e\u5be6\u9a57\u7d50\u679c\uff0c\u5206\u5225\u5177\u6709 10.18%\u300111.70%\u8207 14.97%\u7684\u8fa8\uf9fc\uf961\u63d0\u5347\u3002\u76f8\u5c0d\u65bc MVN \u800c\u8a00\uff0c\u7531\u65bc MVA \u591a\u4f7f\u7528\uf9ba\u4e00\u500b ARMA \u4f4e\u901a\uf984\u6ce2 \u5668\u4ee5\u5f37\u8abf\u8a9e\u97f3\u7684\u6210\u5206\uff0c\u800c HEQ \u984d\u5916\u5c0d\u8a9e\u97f3\u7279\u5fb5\u7684\u9ad8\u968e\u52d5\u5dee(higher-order moments)\u4f5c\u6b63 \u898f\u5316\uff0c\u6240\u4ee5\uf978\u8005\u6548\u679c\u7686\u6bd4 MVN \u9084\uf92d\u5f97\u597d\u3002 \u25cb 2 \u4ee5\u5f80\u6587\u737b\u6240\u63d0\u51fa\u4e4b\u91dd\u5c0d logE \u7279\u5fb5\u4f5c\u88dc\u511f\u7684\u5404\u7a2e\u65b9\u6cd5\uff1aLEDRN-I\u3001LEDRN-II\u3001 LERN-I \u3001 LERN-II \u8207 SLEN \uff0c \u90fd \u80fd \u5e36 \uf92d \u5341 \u5206 \u986f \u8457 \u7684 \u8fa8 \uf9fc \uf961 \u63d0 \u5347 \uff0c \u5176 \u4e2d \u7dda \u6027 LEDRN(LEDRN-I)\u660e\u986f\u512a\u65bc\u975e\u7dda\u6027 LEDRN(LEDRN-II)\uff0c\u5176\u5e73\u5747\u8fa8\uf9fc\uf961\u76f8\u5dee\uf9ba\u5927\u7d04 4%\uff0c \uf978\u7a2e\u7248\u672c\u7684 LERN(LERN-I \u8207 LERN-II)\uff0c\u6548\u679c\u5247\u5341\u5206\u63a5\u8fd1\uff0c\u4e14\u8868\u73fe\u512a\u65bc LEDRN\u3002\u800c\u672c \u5be6\u9a57\u5ba4\u904e\u53bb\u6240\u63d0\u51fa\u7684 SLEN \u6cd5\uff0c\u76f8\u5c0d\u65bc\u57fa\u790e\u5be6\u9a57\u7684\u5e73\u5747\u8fa8\uf9fc\uf961\u800c\u8a00\uff0c\u6709 15.19%\u7684\u63d0\u5347\uff0c \u660e\u986f\u512a\u65bc\u4e4b\u524d\u6240\u63d0\u4e4b LEDRN \u8207 LERN \u7b49\u65b9\u6cd5\u3002 \u25cb 3 \u672c\uf941\u6587\u6240\u63d0\u51fa\u7684\uf978\u7a2e\u975c\u97f3\u7279\u5fb5\u6b63\u898f\u5316\u6cd5\uff0cSFN-I \u8207 SFN-II\uff0c\u76f8\u5c0d\u65bc\u57fa\u790e\u5be6\u9a57\u7d50 \u679c\u800c\u8a00\uff0c\u5e73\u5747\u8fa8\uf9fc\uf961\u5206\u5225\u63d0\u5347\uf9ba 15.38%\u8207 16.11%\uff0c\u76f8\u5c0d\u932f\u8aa4\ufa09\u4f4e\uf961\u90fd\u5728 50%\u4ee5\u4e0a\uff0c\u76f8 \u8f03\u65bc\u4e4b\u524d\u6240\u63d0\u7684\u5404\u7a2e\u65b9\u6cd5\uff0cSFN-I \u8207 SFN-II \u90fd\u6709\uf901\u512a\uf962\u7684\u8868\u73fe\uff0c\u6b64\u9a57\u8b49\uf9ba\u6211\u5011\u6240\u63d0\u7684\uf978 \u500b\u65b0\u65b9\u6cd5\uff0c\u90fd\u80fd\u6709\u6548\u5730\u63d0\u6607 logE \u7279\u5fb5\u5728\u52a0\u6210\u6027\u96dc\u8a0a\u74b0\u5883\u4e0b\u7684\u5f37\u5065\u6027\uff0c\u4e14\u512a\u65bc\u76ee\u524d\u8a31\u591a \u8457\u540d\u7684 logE \u7279\u5fb5\u6b63\u898f\u5316\u6280\u8853\u3002\u6b64\u5916\uff0c\u6211\u5011\u4e5f\u767c\u73fe\uff0cSFN-II \u6240\u5f97\u4e4b\u8fa8\uf9fc\uf961\u6bd4 SFN-I \uf901\u597d\uff0c \u6b64\u53ef\u80fd\u539f\u56e0\u5982\u4e4b\u524d\u6240\u8ff0\uff0c\u7531\u65bc SFN-II \u5728\u8a9e\u97f3\u5075\u6e2c(voice activity detection)\u7684\u6c7a\u7b56\u6a5f\u5236\u8207 SFN-I \u4e26\uf967\u76f8\u540c\uff0c\u8a9e\u97f3\u5075\u6e2c\u4e4b\u932f\u8aa4\u5728 SFN-II \u4e2d\u76f8\u5c0d\u5f71\u97ff\u8f03\u5c0f\uff0c\u800c\u4f7f\u5176\u76f8\u5c0d\u8868\u73fe\u8f03\u4f73\u3002 \u8868\u4e8c\u3001\u91dd\u5c0d logE \u7279\u5fb5\u4e4b\u5f37\u5065\u5f0f\u8a9e\u97f3\u6280\u8853\u4e4b\u8fa8\uf9fc\uf961\u7684\u7d9c\u5408\u6bd4\u8f03\u8868(%) \u4e09\u7a2e\u539f\u672c\u4f5c\u7528\u65bc\u6240\u6709\u7a2e\uf9d0\u7279\u5fb5\u4e4b\u65b9\u6cd5\uff1aMVN\u3001MVA \u8207 HEQ \u6cd5\uff0c\u55ae\u7d14\u4f5c\u7528\u65bc c0 \u7279\u5fb5 \u6642\uff0c\u4ecd\u7136\u4ee5 HEQ \u8868\u73fe\u6700\u597d\uff0cMVA \u6cd5\u6b21\u4e4b\uff0cMVN \u6cd5\u8f03\u5dee\uff0c\u4f46\u5f7c\u6b64\u8868\u73fe\u7684\u5dee\u8ddd\u4e26\u672a\u5982\u4e4b \u524d\u5728\u8868\u4e8c\uf92d\u5f97\u660e\u986f\u3002\u6b64\u5916\uff0cLERN-I\u3001LERN-II \u8207 SLEN \u90fd\u6709\u5341\u5206\u986f\u8457\u7684\u6539\u9032\u6548\u679c\uff0c\u552f \u8207\u8868\u4e8c\u7684\uf969\u64da\uf967\u540c\u4e4b\u8655\uff0c\u5728\u65bc\u4e09\u7a2e\u65b9\u6cd5\u7684\u6548\u80fd\u5341\u5206\u63a5\u8fd1\uff0c\u800c LERN-I \uf976\u512a\u65bc SLEN\u3002 \u25cb 3 \u672c\uf941\u6587\u6240\u63d0\u51fa\u7684\uf978\u7a2e\u975c\u97f3\u7279\u5fb5\u6b63\u898f\u5316\u6cd5\uff0cSFN-I \u8207 SFN-II\uff0c\u76f8\u5c0d\u65bc\u57fa\u790e\u5be6\u9a57\u7d50\u679c\u800c\u8a00\uff0c \u5e73\u5747\u8fa8\uf9fc\uf961\u5206\u5225\u63d0\u5347\uf9ba 13.79%\u8207 14.13%\uff0c\u76f8\u5c0d\u932f\u8aa4\ufa09\u4f4e\uf961\u7d04\u70ba 46%\uff0c\uf9d0\u4f3c\u8868\u4e8c\u7684\u7d50\u679c\uff0c SFN-II \u4ecd\u7136\u512a\u65bc SFN-I\uff0c\u4e14\u9019\uf978\u7a2e\u65b9\u6cd5\u4e4b\u8868\u73fe\u4ecd\u512a\u65bc\u5176\u4ed6\u6240\u6709\u7684\u65b9\u6cd5\u3002\u6b64\u7d50\u679c\u9a57\u8b49\uf9ba\u6211 \u96d6\u7136 SFN \u6cd5\u6709\u6548\u5730\ufa09\u4f4e\u96dc\u8a0a\u5c0d c0 \u9020\u6210\u7684\u5931\u771f\uff0c\u9032\u800c\u63d0\u6607\u8fa8\uf9fc\uf961\uff0c\u4f46\u7576\u6211\u5011\u6bd4\u8f03\u8868 \u4e8c\u8207\u8868\u4e09\u6642\uff0c\u767c\u73fe\u7121\uf941\u662f SFN-I \u6216 SFN-II\uff0c\u4f5c\u7528\u65bc logE \u7279\u5fb5\u53ef\u5f97\u5230\u7684\u8fa8\uf9fc\uf961\u6703\u9ad8\u65bc\u4f5c \u7528\u5728 c0 \u7279\u5fb5\u6240\u5f97\u4e4b\u8fa8\uf9fc\uf961\uff1b\u7531\u6b64\uff0c\u6211\u5011\u63a8\u65b7\u7531 logE \u7279\u5fb5\u6240\u5f97\u4e4b SFN-I \u6cd5\u8207 SFN-II \u6cd5 \u5176\u4e2d\u7684\u8a9e\u97f3\u7aef\u9ede\u5075\u6e2c(VAD)\u7d50\u679c\uff0c\u53ef\u80fd\u6703\u6bd4\u7531 c0 \u6240\u5f97\u7d50\u679c\uf92d\u7684\u597d\u3002\u6839\u64da\u6b64\u63a8\u60f3\uff0c\u6211\u5011 \u5c07\u539f\uf92d\u91dd\u5c0d c0 \u7279\u5fb5\u7684\uf978\u7a2e SFN \u6cd5\u7a0d\u4f5c\u4fee\u6539\u3002\u65bc SFN-I \u4e2d\uff0c\u6211\u5011\u5148\uf9dd\u7528 logE \u5c0d\u97f3\u6846\u505a \u8a9e\u97f3/\u975e\u8a9e\u97f3\u7684\u5206\uf9d0\uff0c\u518d\u5c07\u6b64\u5224\u5225\u7d50\u679c\u5957\u7528\u65bc c0 \u4e0a\uff0c\u5c0d\u975e\u8a9e\u97f3\u97f3\u6846\u7684 c0 \u505a\u5982\u5f0f(2-16)\u4e4b \u6b63\u898f\u5316\u8655\uf9e4\uff1b\u800c SFN-II \u4e5f\u662f\uf9dd\u7528\u76f8\u540c\u7684\u65b9\u5f0f\uff0c\u5148\uf9dd\u7528 logE \u5c0d\u97f3\u6846\u505a\u8a9e\u97f3/\u975e\u8a9e\u97f3\u7684\u5206\uf9d0\uff0c \u518d\u5c07\u5176\u7d50\u679c\u8f49\u63db\u81f3 \u91dd\u5c0d c0 \u7279\u5fb5\u4e4b\u4fee\u6b63\u5f0f SFN-I \u6cd5\u8207\u4fee\u6b63\u5f0f SFN-II \u6cd5\uff0c\u5176\u6240\u5f97\u4e4b\u5e73\u5747\u8fa8\uf9fc\uf961\u5982\u8868\u56db\u6240 \u793a\uff0c\u5982\u6211\u5011\u6240\u9810\u671f\u7684\uff0c\u4fee\u6b63\u5f0f SFN \u6cd5\u76f8\u5c0d\u65bc\u539f\u59cb SFN \u6cd5\uff0c\u80fd\u6709\uf901\u9032\u4e00\u6b65\u7684\u6539\u9032\u6548\u679c\uff0c \u5c0d SFN-I \u800c\u8a00\uff0c\u524d\u8005\u76f8\u8f03\u65bc\u5f8c\u8005\u984d\u5916\u63d0\u6607\uf9ba 1.29%\u7684\u5e73\u5747\u8fa8\uf9fc\uf961\uff0c\u800c\u5c0d SFN-II \u800c\u8a00\uff0c \u524d\u8005\u76f8\u8f03\u65bc\u5f8c\u8005\u984d\u5916\u63d0\u6607\uf9ba\u7684 1.33%\u5e73\u5747\u8fa8\uf9fc\uf961\u3002\u6b64\u7d50\u679c\u90e8\u5206\u9a57\u8b49\uf9ba\u6211\u5011\u7684\u63a8\u60f3\uff0c\u5373\uf9dd \u7528 logE \u7279\u5fb5\uf92d\u57f7\ufa08\u8a9e\u97f3\u7aef\u9ede\u5075\u6e2c(VAD)\uff0c\u5176\u6548\u679c\u6703\u6bd4 c0 \u7279\u5fb5\uf92d\u7684\u597d\u3002 \u8868\u56db\u3001\u91dd\u5c0d c0 \u7279\u5fb5\u4e4b\u539f\u59cb SFN \u6cd5\u8207\u4fee\u6b63\u5f0f SFN \u6cd5\u4e4b\u8fa8\uf9fc\uf961\u6bd4\u8f03\u8868(%) 44%\u7684\u5e73\u5747\u8fa8\uf9fc\uf961\u3002\u540c\u6642\uff0c\u6211\u5011\u4e5f\u770b\u51fa SFN-II \u7684\u6548\u80fd\u666e\u904d\u512a\u65bc SFN-I\uff0c\u6b64\u7d50\u679c \u8ddf\u524d\u4e00\u7ae0\u7684\u7d50\uf941\u662f\u4e00\u81f4\u7684\u3002\u800c\u7576\u6211\u5011\u5c07\u8868\u4e94\u8207\u8868\u4e8c\u7684\uf969\u64da\u76f8\u6bd4\u8f03\u6642\uff0c\u4e5f\u53ef\u4ee5\u770b\u51fa\uff0c\u4f7f\u7528 SFN \u8655\uf9e4 logE \u7279\u5fb5\u7d50\u5408\u4f7f\u7528 MVN\u3001MVA \u6216 HEQ \u6cd5\u984d\u5916\u8655\uf9e4 c1\uff5ec12 \u7279\u5fb5\uff0c\u53ef\u4ee5\u6bd4\u55ae \u7368\u4f7f\u7528 SFN \u8655\uf9e4 logE \u7279\u5fb5\u5f97\u5230\uf901\u4f73\u7684\u8fa8\uf9fc\u6548\u679c\uff0c\u6b64\u7d50\u679c\u9a57\u8b49\uf9ba SFN \u6cd5\u8207 MVN\u3001MVA \u8207\u8868\u4e09\u7684\uf969\u64da\u76f8\u8f03\uff0c\u4f7f\u7528 SFN \u8655\uf9e4 c0 \u7279\u5fb5\u7d50\u5408\u4f7f\u7528 MVN\u3001MVA \u6216 HEQ \u6cd5\u984d \u5916\u8655\uf9e4 c1\uff5ec12 \u7279\u5fb5\uff0c\u53ef\u4ee5\u6bd4\u55ae\u7368\u4f7f\u7528 SFN \u8655\uf9e4 c0 \u7279\u5fb5\u5f97\u5230\uf901\u4f73\u7684\u6548\u80fd\uff0c\u7136\u800c\u6211\u5011\u767c \u73fe\uff0c\u5c07 SFN-I \u6216 SFN-II \u4f7f\u7528\u65bc c0 \u7279\u5fb5\uff0c\u4e26\u7528\u5176\u4ed6\u65b9\u6cd5\u4f7f\u7528\u5728 c1\uff5ec12 \u7279\u5fb5\u6642\uff0c\u6240\u5f97\u5230 \u7684\u8fa8\uf9fc\uf961\u4e26\u975e\u7e3d\u662f\u512a\u65bc\u55ae\u7368\u4f7f\u7528\u4e00\u7a2e\u65b9\u6cd5\u8655\uf9e4\u5168\u90e8\u7279\u5fb5\u7684\u8fa8\uf9fc\u7d50\u679c (\u9019\u4e9b\u8f03\u5dee\u7684\uf969\u64da\u5728 \u8868\u4e2d\u4ee5*\u865f\u52a0\u4ee5\u8a3b\u8a18) \uff0c\uf9b5\u5982\uf99c(6)\u4e4b\u300eSFN-I (c0) + MVA (c1~c12)\u300f\u6cd5\uff0c\u5176\u5e73\u5747\u8fa8\uf9fc\uf961\u70ba 87.77%\uff0c\u76f8\u8f03\u65bc\uf99c(3)\u4e4b\u300eMVA (c0, c1~c12)\u300f\u6cd5\u6240\u5f97\u4e4b 88.46%\uf92d\u5f97\u5dee\u3002\u6b64\u73fe\u8c61\u7684\u53ef\u80fd\u539f \u56e0\uff0c\u5728\u524d\u4e00\u7ae0\u5df2\u7d93\u63d0\u5230\uff0c\u5373\uf9dd\u7528 c0 \u7279\u5fb5\u57f7\ufa08 SFN \u6cd5\u4e2d\u7684\u8a9e\u97f3\u7aef\u9ede\u5075\u6e2c(VAD)\u6703\u6bd4\u8f03\uf967 \u7cbe\u78ba\uff0c\u9032\u800c\ufa09\u4f4e SFN \u7684\u6548\u80fd\u3002\u56e0\u6b64\uff0c\uf9d0\u4f3c\u524d\u4e00\u7ae0\uff0c\u5728\u9019\uf9e8\u6211\u5011\u4f7f\u7528\u91dd\u5c0d c0 \u7279\u5fb5\u4e4b\u4fee\u6b63 \u5f0f\u7684 SFN \u6cd5\uff0c\uf92d\u8207 MVN\u3001MVA \u6216 HEQ \u6cd5\u4f5c\u7d50\u5408\uff0c\u9019\u4e9b\u7d50\u679c\uf99c\u65bc\u8868\uf9d1\u7684\uf99c(11)~(16)\u4e2d\u3002 SFN \u6cd5(modified SFN-I \u8207 modified SFN-II)\uff0c\u6bd4\u539f\u59cb SFN \u6cd5\u7684\u6548\u80fd\u9ad8\u51fa\u8a31\u591a\uff0c\u4e14\u8207 MVN\u3001MVA \u6216 HEQ \u4e00\u4f75\u4f7f\u7528\u5f8c\uff0c\u5176\u7d50\u679c\u5fc5\u7136\u512a\u65bc MVN\u3001MVA \u6216 HEQ \u8655\uf9e4\u6240\u6709\u7279\u5fb5 \u7684\u7d50\u679c\uff0c\u5176\u4e2d\u4ee5\uf99c(15)\u4e4b\u300emodified SFN-II (c0) + MVA (c1~c12)\u300f\u6cd5\u6240\u5f97\u5230\u7684\u5e73\u5747\u8fa8\uf9fc \uf961\u6700\u9ad8\uff0c\u70ba 89.93%\uff0c\u8207\u4e4b\u524d\u8868\u4e94\u4e2d\u6700\u4f73\u8fa8\uf9fc\uf961 89.97%(\uf99c(9)\u7684\u300eSFN-II (logE) + MVA (c1~c12)\u300f\u6cd5)\u5341\u5206\u63a5\u8fd1\uff0c\u6b64\u7d50\u679c\u660e\u986f\u9a57\u8b49\uf9ba\u4fee\u6b63\u5f0f SFN \u6cd5\u78ba\u5be6\uf901\u9032\u4e00\u6b65\u6539\u9032\uf9ba c0 \u7279 \u5fb5\u5728\u52a0\u6210\u6027\u96dc\u8a0a\u74b0\u5883\u4e0b\u7684\u5f37\u5065\u6027\u3002 \u7531\u7b2c\u4e09\u7ae0\u8207\u7b2c\u56db\u7ae0\u4e4b\u5168\u90e8\u7684\u5be6\u9a57\uf969\u64da\u4e2d\uff0c\u6211\u5011\u53ef\u4ee5\u5145\u5206\u9a57\u8b49\u6240\u63d0\u51fa\u7684\uf978\u7a2e\u975c\u97f3\u7279\u5fb5 \u6b63\u898f\u5316\u6cd5(SFN-I \u8207 SFN-II)\u5c0d\u65bc\u80fd\uf97e\u76f8\u95dc\u7279\u5fb5\u5177\u6709\uf97c\u597d\u7684\u5f37\u5065\u5316\u6548\u679c\uff0c\u800c SFN-II \u6240\u5f97\u5230 \u7684\u8fa8\uf9fc\uf961\u7e3d\u662f\u6bd4 SFN-I \u9ad8\uff0c\u5176\u53ef\u80fd\u539f\u56e0\u5982\u7b2c\u4e8c\u7ae0\u6240\u9673\u8ff0\uff0c\u56e0\u70ba SFN-II \u6cd5\u5177\u6709\"\u8edf\u5f0f\"\u6c7a \u7b56\u4e4b\u8a9e\u97f3\u7aef\u9ede\u5075\u6e2c(soft-decision voice activity detection)\u7684\u6a5f\u5236\uff0c\u76f8\u8f03\u65bc SFN-I \u6cd5\"\u786c\u5f0f\" \u6c7a\u7b56\u4e4b\u8a9e\u97f3\u7aef\u9ede\u5075\u6e2c(hard-decision voice activity detection)\u7684\u6a5f\u5236\uff0c\u524d\u8005\u7684\u8a9e\u97f3/\u975e\u8a9e\u97f3\u5224 \u5225\u932f\u8aa4\u6240\u9020\u6210\u7684\u5f71\u97ff\u76f8\u5c0d\u8f03\u5c0f\u3002\u7136\u800c\uff0c\u7e3d\u62ec\u800c\u8a00\uff0cSFN-I \u6cd5 SFN-II \u6cd5\u7684\u5171\u540c\u512a\u9ede\u5728\u65bc\u57f7 \ufa08\u4e0a\u5341\u5206\u7c21\uf9e0(\u5373\u8907\u96dc\ufa01\u6975\u4f4e)\u4e14\u6548\u679c\u5f88\u512a\uf962\uff0c\u56e0\u6b64\u6975\u5177\u5be6\u7528\u7684\u50f9\u503c\u3002",
                "type_str": "table",
                "num": null
            }
        }
    }
}