Hybrid DNN-HMM-Based Approach for Telugu Language Speech Recognition

In this paper, we discuss our research work in Telugu Indian language speech data for building a large language vocabulary to build Telugu speech recognition system. We have collected speech research data around 2628 utterances in Telugu language among 14 Indian regional languages as part of Speech Quality Control Project under Bhashini-NLTM (National Language Translation Mission) in collaboration with IIT Madras. In this research work, we discuss our methodology for the collection of Telugu data and developing Telugu speech database. The proposed Hybrid DNN-HMM acoustic model implemented for building Telugu ASR in Kaldi achieved WER score of 0.7% when the inverse_language_model weight is 17 on the test dataset of Telugu speech database. In our paper, we present the preliminary results that are obtained by building speech recognition using the acoustic models created on Telugu language database by implementing Kaldi toolkit for speech in building ASR in Telugu language. In the process of our work, some of the ASR models, methods and speech processing performance metrics that we used in ASR in order to maintain quality corpus are discussed.