We conduct Automatic speech recognition (ASR) experiments on the Common Voice Spontaneous Speech dataset by Mozilla Data Collective, consisting of 21 low-resource languages across four continents of the world. We fine-tune popular multilingual speech models on all languages of this dataset, and observe that while a single-best-model solution doesn’t exist, the Massively Multilingual Speech model and Whisper achieve superior performance on certain languages. Through n-gram language modeling decoding experiments, we observe a significant improvement in error rate over greedy decoding by up to 27.3%. We follow our experiments with a close linguistic error analysis of the best performing models on Scots (sco) and Nubi (kcn) - two of the languages in our dataset, with very little prior audio and text modeling research. We highlight the morphosyntactic errors induced during speech recognition and perform a holistic analysis of these languages. We finally advocate for the importance of building efficient and accurate ASR tools for modeling speech in endangered languages with scarce resources, and their applications to language revitalization, language learning assistance, and accessibility.

Online social networks are ubiquitous and user-friendly. Nevertheless, it is vital to detect and moderate offensive content to maintain decency and empathy. However, mining social media texts is a complex task since users don’t adhere to any fixed patterns. Comments can be written in any combination of languages and many of them may be low-resource. In this paper, we present our system for the LT-EDI shared task on detecting homophobia and transphobia in social media comments. We experiment with a number of monolingual and multilingual transformer based models such as mBERT along with a data augmentation technique for tackling class imbalance. Such pretrained large models have recently shown tremendous success on a variety of benchmark tasks in natural language processing. We observe their performance on a carefully annotated, real life dataset of YouTube comments in English as well as Tamil. Our submission achieved ranks 9, 6 and 3 with a macro-averaged F1-score of 0.42, 0.64 and 0.58 in the English, Tamil and Tamil-English subtasks respectively. The code for the system has been open sourced.