Building Acoustic Model for Phoneme Recognition using PSO-DBN

Abstract

Deep neural networks has shown its power in generous classification problems including speech recognition. This paper proposes to enhance the power of deep belief network (DBN) further by pre-training the neural network using particle swarm optimisation (PSO). The objective of this work is to build an efficient acoustic model with deep belief networks for phoneme recognition with much better computational complexity. The result of using PSO for pre-training the network drastically reduces the training time of DBN and also decreases the phoneme error rate (PER) of the acoustic model built to classify the phonemes. Three variations of PSO namely, the basic PSO, second generation PSO (SGPSO) and the new model PSO (NMPSO) are applied in pre-training the DBN to analyse their performance on phoneme classification. It is observed that the basic PSO is performing comparably better to other PSOs considered in this work, most of the time.