Sparse Wavelet Decomposition and Filter Banks with CNN Deep Learning for Speech Recognition

In this paper, the speech recognition algorithms using CNN deep learning based on the sparse discrete wavelet decomposition (SDWD) and bandpass filter banks (BPFB) are proposed. The proposed algorithms consist of three stages. First, speech signal is decomposed into sub-band signals according to the Mel filter bank frequency specification using the SDWD or BPFB. The power values from sub-bands form a feature vector for the speech frame. Cascading feature vectors for consecutive speech frames constructs a two-dimension feature image. Secondly, each obtained feature image is subject to flipping operations in order to reduce edge effect when using the standard CNN. Finally, the CNN deep learning is adopted for training and recognition. The experimental results demonstrate that our proposed SDWD-CNN and BPFB-CNN outperforms the support vector machine (SVM), K-nearest neighbors (KNN), and random forest (RF) algorithms.