Evolution, learning and speech recognition in changing acoustic environments

Abstract

In this paper, we apply Evolutionary Algorithms (EA) to evolve Automatic Speech Recognition Systems (ASRSs) in order to adapt them to acoustic environment changes. The general framework relates to the Evolutionary paradigm and it addresses the problem of robustness of speech recognition as a two level process. First, some initial ASRSs based on feedforward Artificial Neural Networks (ANNs) are designed and trained with an initial speech corpus. Second, the ASRSs are tested in Virtual Acoustic Environments (VAEs) in which we playback some speech test data. By using Evolutionary Operators as mutation, crossover and selection, the adaptation of initial ASRSs to a new VAE is achieved. The VAE includes different real world noises and are physical models of real rooms (1 floor, 1 ceiling and 4 walls) thanks to image methods of sound propagation in small rooms.