Acoustic modeling using deep belief network for Bangla speech recognition

Most of the Speech Recognition (SR) systems use Hidden Markov Model (HMM) for acoustic modeling and Gaussian Mixture Model (GMM) for state modeling. Artificial Neural Network (ANN) based methods are also found as a good replacement of GMMs in SR system development. This paper presents a method for Bangla SR using Deep Belief Network (DBN) which is probabilistic generative ANN composed by multiple layers of restricted Boltzmann machine along with HMM. At first Mel Frequency Cepstral Coefficients is used extract features from the speech data. Then DBN is trained with these feature vectors to calculate each of the phoneme states. After that Viterbi decoder is used to determine the resulting hidden state sequence that generates the word. The training of DBN is performed in two steps. At first generative pre-training is used to train the network layer by layer. In the second step, enhanced gradient is used to slightly adjust the model parameters to make it more accurate. Total 840 utterances (20 utterances for each of 42 speakers) of the words are used in this study. The proposed method is shown satisfactory recognition accuracy and outperformed other prominent existing methods.