This paper deals with phoneme recognition based on a hybrid Multi-Layer Perceptron (MLP)/ hidden Markov model system. The effects of the combination of multiple feature sets and the use of a new wide acoustic context procedure on the training of a MLP are investigated.
Experimental results show that the contribution of specific features to phoneme recognition, when used in combination with standard MFCC features was about 1.3% of accuracy improvement. The proposed acoustic context window widening leads to FER relative improvements of 2.8%. Relative improvements of 3.3% and 8.5%, respectively, on accuracy and correctness rates, were obtained if both proposals are included in the training of the phoneme recognition system.