Multi-input Multi-output Beta Wavelet Network: Modeling of Acoustic Units for Speech Recognition

IIn this paper, we propose a novel architecture of wavelet network called Multi-input Multi-output Wavelet Network MIMOWN as a generalization of the old architecture of wavelet network. This newel prototype was applied to speech recognition application especially to model acoustic unit of speech. The originality of our work is the proposal of MIMOWN to model acoustic unit of speech. This approach was proposed to overcome limitation of old wavelet network model. The use of the multi-input multi-output architecture will allows training wavelet network on various examples of acoustic units. The development of robust systems for speech recognition is now one of the main issues of language processing. Many systems were based on concepts that were relatively close. However, despite those progresses, the performances of current systems remain much lower than the capacity of the human auditory system. In this respect, the prospects for improvement are strong. Economic issues related to speech recognition make this area a sector in constant evolution and pushes us to constantly imagine new approaches in order to be able to at least match or exceed the performance of our hearing. To solve some problems of modelling and recognition of speech, we suggest a new method based on wavelet networks (21) that exploit the intrinsic properties of the speech signal, while the statistical models are concerned with statistical properties of the speech signal (2). These models are similar to neural network for the structure and the training approach. But, training algorithms for wavelet network require a smaller number of iterations when compared with neural network. Wavelet network model, the single-input single-output wavelet network, was introduced, firstly, by Zhang and Benveniste in 1992 (8). This model was inspired from neural network architecture as a combination of neuronal contraption and wavelets as activation functions (20). Those models have supplied an access to all frequency of signal thanks to the use of wavelets in the hidden layer of each neurone. It has more advantages than common networks such as faster convergence, avoiding local minimum, easy decision and adaptation of structure (4). Despite the contribution of these models in different fields of pattern recognition (4), (5), (7), they remained limited in the field of modelling. These prototypes cannot instil entities at different occurrences. To overcome these limitations, we advanced a new model for the training of several instances of a single entity at the same time. These models are called multi- input multi-output wavelet network MIMOWN (23)(24).