Lamarckian vs darwinian evolution for the adaptation to acoustical environment change

Abstract

The adaptation to the changes of environment is crucial to improve automatic speech recognition systems’ robustness in various conditions of use. We investigate the adaptation of such systems using evolutionary algorithms. Our systems are based on neural networks. Their adaptation abilities rely on their capacity to learn and to evolve. Within the framework of this work, we study both main methods concerning hybridization of training and evolution, namely the Lamarckian and Darwinian evolution. We show that the knowledge inheritance of a generation to another is much faster and more powerful for the adaptation to a set of acoustic environments changes.