DEEP BELIEF NETWORKS FOR PHONEME RECOGNITION IN CONTINUOUS TAMIL SPEECH–AN ANALYSIS

Abstract

. A combination of Gaussian Mixture Model and Hidden Markov Model has been used successfully in building acoustic models for speech recognition. These models have dominated this area for nearly three decades. Re-entry of neural networks in many clustering, classification and pattern recognition problems have triggered current researchers to focus in making use of its power in the area of speech recognition. This article compares the performance of Bernoulli-Bernoulli Deep Belief Networks (BBDBN) and Gaussian-Bernoulli Deep Belief Networks (GBDBN) on phoneme recognition of spoken speech in Tamil. In addition to that the impact of feature representation in the performance of acoustic model is also studied by using three different datasets built using different feature representation for the phoneme samples extracted from the continuous Tamil speech.