Speech recognition using three channel redundant wavelet filterbank

Although Wavelet Transform has multi resolution properties, it is not optimized for speech recognition. This paper presents redundant discrete wavelet-based speech representations, which owing to much less shift-sensitivity are better for speech recognition tasks, in contrast with two-channel Discrete wavelet transform. However, this improvement is at the expense of higher redundancy. In this paper, three types of wavelet features are presented, including a combination of critically sampled Discrete Wavelet and 3-channel redundant filter banks with down-sampling by 2. For comparing different methods, time-delay neural networks are implemented. Using different mother wavelets for feature extraction improve speech recognition rates. It is shown that redundant three-channel wavelet filter banks work better in speech recognition.