Binary Phoneme Classification Using Fixed and Adaptive Segment Based Neural Network Approach

This paper addresses the problem of binary phoneme classification via a neural net segment-based approach. Phoneme groups are categorized based on articulatory information. For an efficient segmental acoustic properties capture, the phoneme associated with a speech segment is represented using MFCC's features extracted from different portions of that segment as well as its duration. These portions are obtained with fixed or variable size analysis. The classification is done with a Multi-Layer Perceptron trained using the Mackay's Bayesian approach. Experimental results obtained from the Otago speech corpus favourites the use of fixed segmentation strategies over adaptive ones for resolving consonants/vowels, Fricatives/non fricatives, nasals/non nasals and stops/non stops binary classification problems