Development of robust acoustic model for hands-free speech recognition

(1)免提语音识别的鲁棒声学参数研究在免提语音识别中，由于扬声器与麦克风之间的距离变化以及混响引起的加性噪声和卷积噪声的变化极大地降低了识别性能。为了提高对这些干扰的鲁棒性，本项目研究了一个新的特征参数“广义动态倒谱（DyMFGC）”，该参数基于对数尺度和线性尺度之间的广义对数尺度上的前向掩蔽。首先，将所提出的前向掩模应用于mel频率滤波器组频谱。此外，将该前向掩模应用于mel-LPC谱，该谱通过一种简单有效的时域技术来估计mel-频率轴上的全轮询模型。数字识别测试在扬声器和麦克风之间距离为20 ~ 200cm的条件下，在相对安静和较小的办公室环境中进行。在白噪声环境下，DyMFGC在对数谱上优于动态倒谱和倒谱均值归一化的MFCC，在距离源1m的范围内保持90% ~ 95%的词精度。(2) HMM对免提语音的适应研究本项目还开发了一种基于奇异值分解（SVD）和有效秩估计的最大似然线性回归（MLLR）技术。这种技术允许我们将其应用于任何大小的回归类，也可以扩展二阶回归。通过说话人自适应的初步测试表明，在大词汇量语音识别中，基于奇异值分解的MLLR的识别准确率略高于传统的MLLR。此外，二阶回归提高了对加性噪声条件的自适应精度。在另一项研究中，我们将并行模型组合（PMC）扩展到分段单元输入HMM，以使其适应由加性噪声和/或混响环境退化的语音。该方法在加性噪声环境下的识别性能优于原PMC，但在混响环境下的识别效果较差。少

( 1 ) A Study on Robust Acoustic Parameters for Hands-free speech recognitionIn hands-free speech recognition, the variation of additive and convolutional noises due to the variable distance between speaker and microphone as well as reverberation extremely degrades recognition performance. In order to improve the robustness to these disturbances, this project examined a new feature parameter, "a generalized dynamic cepstrum ( DyMFGC ) ," based on the forward masking on the generalized logarithmic scale between the logarithmic and the linear scales. First, the forward masking proposed is applied to a mel-frequency filter bank spectra. Furthermore, this forward masking was applied to a mel-LPC spectra, which is derived by a simple and efficient time domain technique to estimate an all-poll model on a mel-frequency axis.Digit recognition tests are carried out under the conditions that the distance between speaker and microphone is from 20 to 200cm in a relatively quiet and small size offi … More ce environments. Under white noise environments, the DyMFGC outperforms the dynamic cepstrum on the logarithmic spectrum and MFCC with cepstral mean normalization, and maintains the word accuracy of 90 % to 95 % within a 1m distance from a source.(2) A Study on a HMM Adaptation to Hands-free SpeechA part of this project also developed a Maximum Likelihood Linear Regression ( MLLR ) technique based on a singular value decomposition ( SVD ) and an effective rank estimation. This technique allows us to apply it to any size of regression classes and also to extend the second order regression. A preliminary test by speaker adaptation shows that the SVD-based MLLR achieves slightly higher recognition accuracy than the conventional MLLR in large vocabulary speech recognition. Furthermore, the second order regression improves adaptation accuracy for additive noise conditions.In another study, we extended the parallel model combination ( PMC ) to the segmental unit input HMM to adapt it to speech degraded by additive noise and/or reverberant environments. This method gives better recognition performance than the original PMC in the additive noise environments, but is not so effective to the reverberant environments. Less