Layered neural nets applied in the recognition of voiceless unaspirated stops

The authors present the application of layered neural nets in the recognition of the syllable-initial voiceless unaspirated stops in Mandarin speech. The input to the neural net is a feature vector with components derived from the burst spectrum, the formant transitions and the voice-onset time. The function of the neural net is to classify the places of articulation of these stop consonants. The authors compare the neural-net classifier with the Bayes classifier based on the assumption of single multivariate Gaussian distribution for each consonant model. The effects of the number of hidden layers and also the number of the hidden units are investigated. A method to minimise the degradation of the performance of an existing neural-net classifier when one of the hidden processing units misses is also proposed.<>