Speaker adaptation using multi-layer feed-forward automata and canonical correlation analysis

Intra and inter-speaker variability is a major source of error in Automatic Speech Recognition (ASR). This paper reports on a series of experiments that allow us to control some aspects of this variability and therefore perform an adaptation of standard recognizers to new speakers. The first experiments concern spectral transformation using linear approaches based on some powerful data analysis techniques. In a second series of experiments spectral transformations using Multi-layer Feed Forward Automata (MLFFA) are introduced. Compared with linear techniques, the MLFFA approach offers some improvement.