Speaker normalization using second-order connectionist networks
Raymond L. Watrous · The Journal of the Acoustical Society of America · 1990
A method for speaker-adaptive classification of vowels using connectionist networks is developed. A speaker-specific linear transformation of observations of the speech signal is computed using second-order connectionist network units. Vowel classification is accomplished by a multilayer network, which operates on the transformed speech data. The network may be adapted for a new talker by modifying the transformation parameters while leaving the classifier network fixed. This is accomplished by back-propagating classification error through the classifier to the transformation parameters. A model of this type was evaluated for 75 speakers using the first two formant values of the Peterson/Barney data (ten vowels). When the speaker-dependent transformation and nonlinear classifier were simultaneously optimized, a vowel recognition accuracy of 96.8% was obtained. However, when adapted from average speaker parameters, the network was shown to be very sensitive to initial values. When the linear transformation was optimized separately by minimizing the mean-squared error between vowels across talkers, and the classifier optimized for this normalized data, better results were obtained. Recognition accuracy using the normalized data was 91.7%; when adapted to each speaker, the accuracy improved to 96.6%.