Bilingual Speech Recognition based on Deep Neural Networks and Directed Acyclic Word Graphs

This study explores the use of Bilingual speech recognition system for Indian languages along with Kaldi toolkit and Directed Acyclic Word Graphs as an innovative idea. When using a Large Vocabulary Continuous Speech Recognition (LVCSR), the crucial task is to establish a relationship between sub-word acoustic units across the particular languages. It forms the core for building automatic speech recognition system for multiple languages. Deep neural networks were employed with 2 hidden layers for acoustic modeling to create a Bidirectional Long Short-Term Memory Networks (BLSTM) model. For bilingual speech recognition, standard Mel-Frequency Cepstral Coefficients (MFCC) generated on audio along with Gaussian Mixture Model/Hidden Markov Model (GMM/HMM) were used to align the reference text. The final language models implemented were statistically pruned trigram models. The study aimed at building a refined Telugu-English bilingual speech recognition system by using a Directed Acyclic Word Graph (DAWG) to map phonetically similar words in both the languages.