A method for speaker normalization and adaption using connectionist networks is developed. A speaker-specific linear transformation of observations of the speech signal is computed using second-order network units. Classification is accomplished by a multilayer feedforward network that operates on the normalized speech data. The network is adapted for a new talker by modifying the transformation parameters while leaving the classifier fixed. This is accomplished by backpropagating classification error through the classifier to the second-order transformation units. This method was evaluated for the classification of ten vowels for 76 speakers using the first two formant values of the Peterson-Barney data. The results suggest that rapid speaker adaptation resulting in high classification accuracy can be accomplished by this method.

Second-order recurrent networks that recognize simple finite state languages over {0,1}* are induced from positive and negative examples. Using the complete gradient of the recurrent network and sufficient training examples to constrain the definition of the language to be induced, solutions are obtained that correctly recognize strings of arbitrary length.

A modular connectionist network model for capturing invariant relationships between the acoustic signal and phonetic categories is developed. The model addresses variations in the acoustic manifestation of phonemes as a function of loudness, speaker identity, speaking rate, and phonetic context. These sources of variation are referred to separate specialized network modules. Each module transforms the representation of its input signal in order to normalize the effect of different source variables. Components of the model have been tested on isolated words for speaker-adaptive vowel recognition (97%) and context-dependent vowel recognition (99.7%). A model integrating amplitude-normalization, speaker-normalization, and context-modulation for continuous speech recognition is under development.>