IITG-INDIGO Submissions for Interspeech-2021 Multilingual and Code-Switching ASR Challenges for Low Resource Indian Languages

The submissions made by the INDIGO team at Indian Institute of Technology Guwahati (IITG) to the Interspeech-2021 multilingual and code-switching Automatic Speech Recognition (ASR) challenges for low resource Indian languages are described in this publication. For multilingual ASR case, the CTC-loss based end-to-end ASR systems are developed for 6 target languages, namely, Gujarati, Hindi, Odia, Marathi, Telugu and Tamil. For the code-switching case, separate ASR systems are developed for Hindi-English and Bengali-English code-switching data using the time delay neural network (TDNN) based acoustic model employing lattice-free, maximum mutual information (LF-MMI) training. The language wise results and the average of all the 6 languages and the other having 5 languages (all except Marathi) and their average was 73.24%. Without Marathi, the average %WER obtained was 65.60%. The TDNN system was found to give better results than the baseline on the basis of average Transliterated %WER. But the average %WER of TDNN system was slightly higher than the baseline. The raw average scored 30.73 %, whereas the transliterated average scored 27.49% (better than baseline)