4 - Adaptation au locuteur de systèmes de reconnaissances. Régression linéaire multiple et perceptrons multicouches

Interspeaker variability is a major source of errors in automatic speech recognition . This paper describes a series of experiments, conducted at TELECOM Paris by the « Pattern Recognition and Speech Processing » Group, for controlling some aspects of this variability, thus allowing for the adaptation of speech recognition systems to new users . The firsi experiments are based on a linear data analysis technique multiple linear regression (MLR) . The second set uses multilayer perceptrons, and yields slightly better results, because non linear phenomena are taken into account. The average improvement of recognition scores is 16 % with the second approach, versus 15 % with the first one . Those techniques can also be used for the adaptation of recognizers to new acoustical environments and recording conditions .