Improving robustness of connectionist speech recognition systems by genetic algorithms

We present an approach which limits significantly the drop of performances related to automatic speech recognition systems (ASRSs) caused by acoustic environment changes. We propose to combine principal component analysis (PCA) and genetic algorithms (GA) in order to transform the noisy acoustic environment into a predefined and well-known (canonical) environment. The idea consists in projecting the noisy speech parameters onto the optimal subspace generated by the genetically modified principal components of the canonical environment. The results show that in noisy and changing environments, the proposed PCA/GA optimized system achieves high recognition rate compared to the baseline system.