An end-to-end interpolated Automatic speech recognition system with punctuated transcripts for the Hindi language

Abstract

The Automatic Speech Recognition System (ASR) produces transcripts that often are misinterpreted and confuses the reader due to lack of context and punctuations. The presence of punctuation in the text improves readability and helps in better cognitive understanding. A wide variety of work has been done on English but Hindi which is the third-largest spoken language in the world, after English and Mandarin, still remains in the shadows. This paper aims to extend the technology to a wider section and introduces an end-to-end system, interpolating an automatic speech recognition system and natural language processing to produce high-quality punctuated transcriptions for the Hindi language. An ASR is implemented using the Kaldi toolkit leveraging the hybrid deep neural networks and the punctuation restoration is done with Bidirectional RNNs with an attention network.