语音识别中的两级MEL域滤波器组维纳滤波方法

欧洲电信标准化协会(European Telecommunications Standards Institute,简称ETSI)于2002年10月发布了分布式语音识别的鲁棒性前端标准。该标准参数的鲁棒性远优于MFCC参数。为了能够在低运算资源的设备上实现鲁棒性前端,在ETSI标准的核心两级维纳滤波算法的基础上,我们提出了一种新方法以提高算法效率。我们首先在Mel域滤波器组幅度上构造维纳滤波器,然后对维纳滤波器系数进行平滑。最后,将维纳滤波器直接应用到Mel域滤波器组幅度上。实验表明,新方法在保持ETSI两级维纳滤波算法出色性能的同时,大大地降低了运算量。

Two-stage Mel-scale Filterbank Wiener Filtering in Speech Recognition

In Oct. 2002, European Telecommunications Standards Institute (ETSI) standardized its robust feature extraction algorithm for distributed speech recognition. Its parameter is more robust than MFCC. In order to use the robust feature in low computational resource device, on the basis of two-stage Wiener filtering algorithm in ETSI Standard, we propose a novel approach to improve the efficiency. Wiener filter is first designed using Mel-scale filterbank amplitude of the speech signal, then the wiener filter coefficients are smoothed. Finally, the wiener filter is applied directly on Mel-scale filterbank amplitude. The experiments show that the new approach can reduce the computational requirement significantly while preserving the excellent performance of the ETSI two-stage wiener filtering algorithm.