首页>
外国专利>
Noise under the voice recognition device and noise under the speech recognition method
Noise under the voice recognition device and noise under the speech recognition method
展开▼
机译:语音识别装置下的噪音和语音识别方法下的噪音
展开▼
页面导航
摘要
著录项
相似文献
摘要
PROBLEM TO BE SOLVED: To suppress the degradation of recognizing performance caused by the fluctuation of environmental noise and the fluctuation of distance between noise source and a voice input microphone by obtaining estimated superposed noise spectrum using a noise model and a noiseless voice model. ;SOLUTION: An estimated S/N computing means 16 computes an estimated S/N 130 to each vector of a noise superposed voice feature vector time series 120 using a noiseless voice model 122 stored in a memory 13 and a noise model 121 stored in a memory 14. A noise superposed voice model generating means 17 synthesizes the noiseless voice model 122 and the noise model 121 according to the estimated S/N 130 to generate a noise superposed voice model 131. A noise spectrum generating means 20 generates estimated superposed noise spectrum 5 using the noise superposed voice feature vector time series 120, the estimated S/N 130, weight data 133 which is the output of a maximum likelihood selecting type weight data computing means 19, and the noise model 121, and outputs it to a noise spectrum eliminating means.;COPYRIGHT: (C)1997,JPO
展开▼
机译:要解决的问题:通过使用噪声模型和无噪声语音模型获得估计的叠加噪声谱,抑制环境噪声波动以及噪声源与语音输入麦克风之间的距离波动导致的识别性能下降。 ;解决方案:估计信噪比计算装置16使用存储在存储器13中的无噪声语音模型122和存储在存储器13中的噪声模型121来计算噪声叠加语音特征矢量时间序列120的每个矢量的估计信噪比130。存储器14。噪声叠加语音模型生成装置17根据估计的S / N 130合成无噪声语音模型122和噪声模型121,以生成噪声叠加语音模型131。噪声谱生成装置20生成估计的叠加噪声谱。使用噪声叠加语音特征矢量时间序列120,估计的S / N 130,作为最大似然选择类型权重数据计算装置19的输出的权重数据133和噪声模型121,将其输出到图5所示的噪声。频谱消除手段。;版权:(C)1997,日本特许厅
展开▼