EUROSPEECH 2003 - INTERSPEECH 2003
In the learning process of speech modeling, many choices or settings are defined "a priori" or are resulting from years of experimental work. In this paper, instead, a global learning scheme is proposed based on a Distributed Genetic Algorithm combined with a standard speech-modeling algorithm. The speech recognition models are now created out of a predefined space of solutions. Furthermore, this global scheme enables to learn the speech models as well as the best feature extraction module. Experimental validation is performed on the task of discovering the Wavelet Packet best basis decomposition, knowing that the "a priori" reference is the mel-scaled subband decomposition. Two experiments are presented, a reference system using a simulated fitness and a second one that uses the speech recognition performance as fitness value. In the latter, each element of the space is a connectionist system defined by a Wavelet topology and its associated Neural Network.
Bibliographic reference. Kommer, Robert van / Hirsbrunner, Beat (2003): "Distributed genetic algorithm to discover a wavelet packet best basis for speech recognition", In EUROSPEECH-2003, 453-456.