Automatic Speech Recognition in Sanskrit: A New Speech Corpus and Modelling Insights

Automatic speech recognition (ASR) in Sanskrit is interesting, owing to the\nvarious linguistic peculiarities present in the language. The Sanskrit language\nis lexically productive, undergoes euphonic assimilation of phones at the word\nboundaries and exhibits variations in spelling conventions and in\npronunciations. In this work, we propose the first large scale study of\nautomatic speech recognition (ASR) in Sanskrit, with an emphasis on the impact\nof unit selection in Sanskrit ASR. In this work, we release a 78 hour ASR\ndataset for Sanskrit, which faithfully captures several of the linguistic\ncharacteristics expressed by the language. We investigate the role of different\nacoustic model and language model units in ASR systems for Sanskrit. We also\npropose a new modelling unit, inspired by the syllable level unit selection,\nthat captures character sequences from one vowel in the word to the next vowel.\nWe also highlight the importance of choosing graphemic representations for\nSanskrit and show the impact of this choice on word error rates (WER). Finally,\nwe extend these insights from Sanskrit ASR for building ASR systems in two\nother Indic languages, Gujarati and Telugu. For both these languages, our\nexperimental results show that the use of phonetic based graphemic\nrepresentations in ASR results in performance improvements as compared to ASR\nsystems that use native scripts.\n