Optimal feature vector for speech recognition of unequally segmented spoken digits

In this paper we describe a model obtained by applying the Discrete Wavelet Transform (DWT) to unequally segmented digits. Each signal is divided with a pre-determined segmentation into a maximum of five subwords. The purpose is speaker independent single digit recognition. The parameterization of the subwords is accomplished by measuring its energy contents after decomposing it with the DWT. This model uses one coefficient per subword and produces up to a 99% recognition rate. It is superior in its class due to the high reduction in the size of the feature vector and consequently in the speed of processing. Typically the reduction is 20:1 if compared with the traditional Melscale model. A successful attempt to classify vowels and accurately identify digits visibly using the proposed model is undertaken. A Radial Basis Function Artificial Neural Network (RBF-ANN) is employed for the recognition tasks and for the comparison of the proposed model with the Fourier one [2]. We use orthogonal wavelets from the Daubechies set. Also the performances of some biorthogonal wavelets are included.