Speech phoneme recognition using wavelets and artificial neural networks
Curtis Alan Tesdahl · Iowa State University Digital Repository (Iowa State University) · 1998
Speech recognition systems have explored the use of Artificial Neural Networks (ANN) to recognize phonemes from extracted speech for several years. Of the systems that performed well to date, the data sets lacked diversity. Further, most systems used the Short-Time Fast Fourier Transform (STFFT) as the front-end spectral preprocessor. It is envisioned that the use of a Wavelet Transform could offer higher performance on a diverse data set in terms of recognition and in terms of reduced computations. The few papers that have explored this approach have limited or confined the topic in terms of the following: a limited data set; no direct comparison to the popular STFFT; and limited comparisons to various Wavelets. Original work is presented in terms of addressing the limitations previous work. This is accomplished by the use of various Wavelet Transforms (Wavelet Filters) as preprocessors for an ANN, making direct comparisons to the STFFT, and making direct comparisons to various Hamming Filters. The work of others is also tabulated and presented. The wavelet-based preprocessor presented in this thesis performed well on a diverse data set. Several Wavelet Transforms were evaluated. Although most of them performed well, two of them namely Daubechies 6 and Daubechies 20 offered results that were better than those obtained using the STFFT or Hamming Filters. By inference of previous work and the results of this thesis, a conclusion can be drawn that the strategy in this thesis would outperform a comparable Hidden Markov Model (HMM). The results obtained were superior to those obtained by others. Further, the wavelet preprocessor can be implemented in a parallel environment to obtain improvements in computational speed. Paralleled computations are not possible in the HMM due to its sequential nature and are at best limited in the FFT.