Speaker Independent Accent Based Speech Recognition for Malayalam Isolated Words: An LSTM-RNN Approach

Automatic speech recognition (ASR) has been a very active area of research for the past few decades. Though there are great advancements in ASR in many languages accent-based speech recognition is an area that is yet to be explored in many languages. Speech recognition by humans is an intuitive process and so is a tough process to make the computers automatically recognize human speech. Although speech recognition has achieved promising achievements for many languages; speech recognition for the Malayalam language is still in infancy. The scarcity of the datasets makes it researchers difficult to do the experiments. Here in this paper, we have experimented with Long Short-Term Memory (LSTM) a Recurrent Neural Network (RNN), for recognizing the accent-based isolated words in Malayalam. The datasets we used here have been constructed manually under a natural recording environment. We used Mel Frequency Cepstral Coefficient (MFCC) methods to extract the features from the audio signals. LSTM with RNN is used to train and build the model since this technology significantly outperforms all other feed-forward deep neural networks and other statistical methodologies.