Dyadic Wavelet Features for Isolated Word Speaker Dependent Speech Recognition

摘要

This research examines the use of dyadic wavelet features for the recognition ofspeaker dependent isolated word speech. The features were generated using three different wavelet filters-Daubechies 4 coefficient (Db4), Daubechies 20 coefficient (Db20) and a 31 coefficient cubic spline; and three different window lengths-15ms, 8ms and 4ms. The accuracy of the standard and over-sampled dyadic wavelet methods were compared. The over-sampled dyadic wavelet method using the Db4 scaling function, with a maximum accuracy of 65.5%, was found to be the most accurate of the wavelet methods tested. The accuracy of this over-sampled dyadic Db4 wavelet method was compared to the accuracy of three Fourier feature methods: octave frequency bandwidth, equal bandwidth and Mel scaled bandwidth features. The dyadic wavelet methods did not perform as well as the Fourier methods. The maximum accuracy obtained for the wavelet methods was 65.5%, compared to the maximum accuracy of the octave bandwidth feature Fourier method of 85.6%. The combination of wavelet features and Fourier features was tested. The order of magnitude of the covariance matrices of each set were equalized and the resulting feature vector set classified. It was found that the recognition accuracy of the wavelet plus Fourier feature vectors, 74.0%, was lower than the recognition accuracy of the Fourier-only feature vectors, 84.5%. The inclusion of the wavelet features added information to the system that reduced the recognition effectiveness of the Fourier features.