A level dependent compression scheme for automatic speaker independent recognition

In this paper, we introduce a new approach to automatic speaker independent speech recognition via a level dependent compression scheme that employs wavelet analysis in parameterizing subwords of the speech signals for the digits 0,1,...,9 and the utterance " oh". Initially, every signal is divided automatically into a maximum of five spectrally stable segments known as subwords, using the spectral variation function (SVF). Each subword is then subjected to the new compression scheme and parametrized by using the discrete wavelet transform scale (DWTS). Energy vectors are finally chosen to represent these subwords. A radial basis functions (RBF) artificial neural network (ANN) is used for the training, testing and recognition of the digits. This approach proved to maintain a constant and competitive recognition rate with different analyzing wavelet functions