New low rate wavelet models for the recognition of single spoken digits

This paper describes three models acquired by applying various wavelet analysis techniques to subwords for the purpose of speaker independent single digit recognition. We emphasize the parameterization of the subwords according to a Mel scale in the cases of the sampled continuous wavelet transform (SCWT) and the wavelet packet decomposition (WPD). When using the discrete wavelet transform (DWT), a logarithmic segmentation is obtained and with it comes a very low parameter representation with a reduction of 3:1 when compared with the other two introduced models and with the Mel scale model. The DWT model has advantage over the other two due to its simplicity and fast implementation. Previous work by Phillips, Tosuner and Robertson (1995), based on preprocessing using traditional Fourier transform (FT) followed by a radial basis functions artificial neural network (RBF-ANN) yielded recognition in 90% range. Our results show that these new models outperformed the Mel scale model.