Automatic Speech Recognition and Translation for Low Resource Languages

10CoRePooL—Corpus for Resource-Poor Languages: Badaga Speech Corpus

Barathi Ganesh H.B.^1,2, Jyothish Lal G.¹*, Jairam R.^1,2, Soman K.P.¹, Kamal N.S.² and Sharmila B.³

¹Center for Computational Engineering and Networking (CEN), Amrita Vishwa Vidyapeetham, Coimbatore, India

²RBG.AI, Resilience Business Grids LLP, SREC Incubation Center, Coimbatore, Tamil Nadu, India

³Sri Ramakrishna Engineering College Coimbatore, Tamil Nadu, India

Abstract

This chapter presents a corpus named CoRePooL that stands for Corpus for Resource-Poor Languages. As voice-specific human-machine interaction applications are accelerated by deep learning algorithms, the lack of resources constrains the scalability in applying to resource-poor languages. In CoRePooL version 0.1.0, we released 420 min of monolingual supervised corpus and 968 minutes of multilingual unsupervised corpus for the Badaga language from the Dravidian language family. The annotation of supervised corpus helps in performing speech-to-text, text-to-speech, translation, gender, and speaker identification. The unsupervised corpus would help self-supervised algorithms which compute latent representations. We also provided the baseline for all the tasks by fine-tuning the foundation models on the released corpus. The code, models, and data are made publicly available at https://github.com/rbg-research/CoRePooL.

Keywords: CoRePooL, Badaga language, speech-to-text, text-to-speech, translation, gender identification, speaker identification ...