Speech recognition in real world: Artificial neural networks and robustness

This paper presents an empirical modeling of the role of environment for Automatic Speech Recognition systems in real world, taken in the framework of an Artificial Life (AL) methodology. Environment is modeled as an active system which triggers the shift between the training and testing states of Automatic Speech Recognition Systems (ASRSs) which are built from Artificial Neural Networks. First an initial set of ASRSs are created to recognize speech under the constraints of an unpredictable acoustic world (defined by the different kind of noises present in it). The training of the ASRSs starts and goes on until ASRSs no longer decrease their error of classification in the current acoustic environment because of noises. This moment is detected by the reactive environment and the structure of the ASRSs are changed. The simulation performed with mathematical models of real rooms as environment showed that our system could be used as a prediction tool of ASRSs performances for the study of any speech perceiver based on Artificial Neural Networks or on Hidden Markov Models. Moreover it is shown that on a task of French digits recognition, the new method performs better than conventional ones.