Hybridization of two stage Multilayer Neural Networks based Bangla ASR incorporating dynamic parameters

This paper presents a hybridization of Multilayer Neural Network-based Bangla phoneme recognition method for Automatic Speech Recognition (ASR) incorporating dynamic parameters. The method consists of four stages: at first stage, a multilayer neural network (MLN) converts acoustic features, mel frequency cepstral coefficients (MFCCs), into phoneme probabilities. Phoneme probabilities from the first stage are inserted into second stage MLN for obtaining more accurate phoneme probabilities with reduced context where the third stage computes dynamic (velocity (A) and acceleration (AA)) parameters from the phoneme probabilities by using three point linear regressions (LRs). Finally, the phoneme probabilities, dynamic parameters, A and AA, and the input MFCCs, combined as hybrid features, are fed into a hidden Markov model (HMM) based classifier to obtain more accurate phoneme strings. From the experiments on Bangla speech corpus prepared by us, it is observed that the proposed method provides higher phoneme recognition performance than the existing method. Moreover, it requires a fewer mixture components in the HMMs.