Speaker-independent consonant classification with distinctive features

We study the problem of classifying stop and nasal consonants in continuous speech independently of the speaker. We consider some acoustic parameters computed from the auditory spectrogram, and other parameters computed from the speech waveform. The classification algorithm uses a recurrent multi-layer perceptron (MLP) with localized connections. The design of the classifier is motivated by knowledge in phonetics and in pattern recognition. We report experiments for the TIMIT database, using 343 speakers in the training set and 77 different speakers in the test set. Good performance is obtained when many acoustic parameters are fed to the MLP, and when the MLP desired outputs represent context-dependent articulatory features. Classification is performed by Principal Component Analysis of the MLP outputs. Refinement of the design parameters yield increasingly better performance on the test set, ranging from 45% errors for a perceptron to 23.3% errors for the best MLP