Automatic Speech Recognition in Malayalam Using DNN-based Acoustic Modelling

In the paper, we describe Malayalam automatic speech recognition using DNN-based acoustic modelling. Mel-frequency cepstral coefficient features (MFCC) are extracted, and acoustic modelling is done using GMM-HMM, SGMM and DNN. The dependency of the performance on different factors like the type of modelling used and alignment algorithms employed are studied. Training and testing of the system were performed using the open-source Kaldi toolkit. We investigated the performance by varying number of hidden layers and hidden units. Testing of the models was carried out on very large vocabulary continuous Malayalam speech recognition task. We obtained word rate (WER) of 3.01% and the sentence error rate (SER) of 14.59% for the proposed system. DNN modelling shows significant improvement over GMM-HMM modelling framework in the design.