Speaker adaptation for speech recognition systems. Multiple linear regression and multilayer perceptrons

摘要

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