Grant: Discovery Grants Program - Individual

Two of the principal problems in speech technology are speaker recognition (who is speaking?) and speech recognition (what is being said?). These problems have resisted solution because of the numerous types of variability in speech. For example, the performance of speaker recognition systems degrades as a result of channel and microphone effects and careful modeling is needed to distinguish these nuisance effects from aspects of the speech signal which are useful in discriminating between speakers. In speech recognition on the other hand, variation between speakers is considered to be a nuisance and the only type of variability which is of real interest is phonetic variability. This proposal is concerned with developing new statistical methods of modeling speaker and channel variation using `hidden variables' which I call speaker and channel factors. The techniques that I am developing are purely statistical so it is hard to say exactly what the hidden variables represent but it is reasonable to surmise that they capture features such as a speaker's age and sex, electrical noise in transmission channels and ordinary background acoustic noise.