Interspeech 2018 Low Resource Automatic Speech Recognition Challenge for Indian Languages

Abstract

India has more than 1500 1 languages, with 30 of them spoken by more than one million native speakers. Most of them are low-resource and could greatly benefit from speech and language technologies. Building speech recognition support for these low-resource languages requires innovation in handling constraints on data size, while also exploiting the unique properties and similarities among Indian languages. With this goal, we organized a low-resource Automatic Speech Recognition challenge for Indian languages as part of Interspeech 2018. We released 50 hours of speech data with transcriptions for Tamil, Telugu and Gujarati, amounting to a total of 150 hours. Participants were required to only use the data we released for the challenge to preserve the low-resource setting, however, they were not restricted to work on any particular aspect of the speech recognizer. We received 109 submissions from 18 research groups and evaluated the systems in terms of Word Error Rate on a blind test set. In this paper we summarize the data, approaches and results of the challenge.

A Survey of Multilingual Models for Automatic Speech Recognition

“…Building on (Conneau et al, 2020a), pre-training with a diverse set of languages improves performance on languages not present in the pre-training data, which indicates that multilingual models may be useful for building speech technologies for very low resource languages, provided that some data for finetuning and evaluation is collected. Although these experiments do not include comparisons with monolingual models, the models beat previous SOTA models in terms of performance, which in turn are better performing than monolingual models (Srivastava et al, 2018). (Wang et al, 2021) proposes to combine supervised learning with labeled data and the CTC loss and SSL with the WAV2VEC2.0 loss in a Multitask Learning (MTL) scheme which they call UniSpeech.…”

Multilingual and code-switching ASR challenges for low resource Indian languages

“…Further, the selected data is split into a train (Trn) and test (Tst) without sentence overlap and with out-of-vocabulary (OOV) rates at about 30% between train and test sets. The Tel, Tam and Guj data are taken from Interspeech 2018 low resource automatic speech recognition challenge for Indian languages, for which, the data was provided by SpeechOcean.com and Microsoft [17]. The train and test sets are considered as-is for this challenge, however, the blind test set is modified with speed perturbations randomly between 1.1 to 1.4 (with increments of 0.05), and/or adding one noise randomly from white, babble and three noises chosen in the Musan dataset [18] considering the signal-to-noise ratio randomly between 18dB to 30dB at step of 1dB.…”

A Dataset for Measuring Reading Levels In India At Scale

“…Tundra is a comprehensive dataset with more than 14 languages built from audiobooks [17]. Perhaps the closest dataset to the one presented in this paper is the Low Resource Automatic Speech Recognition Challenge for Indian Languages [18]. It has over 150 hours of Tamil, Telugu and Gujarati languages.…”