BembaSpeech: A Speech Recognition Corpus for the Bemba Language

Proceedings of the 13th Conference on Language Resources and Evaluation (LREC 2022), pages 7277–7283 Marseille, 20-25 June 2022 © European Language Resources Association (ELRA), licensed under CC-BY-NC-4.0 7277 BembaSpeech: A Speech Recognition Corpus for the Bemba Language Claytone Sikasote1 and Antonios Anastasopoulos2 1Department of Computer Science, University of Zambia, Zambia 2Department of Computer Science, George Mason University, USA claytone.sikasote@cs.unza.zm, antonis@gmu.edu Abstract We present a preprocessed, ready-to-use automatic speech recognition corpus, BembaSpeech, consisting over 24 hours of read speech in the Bemba language, a written but low-resourced language spoken by over 30% of the population in Zambia. To assess its usefulness for training and testing ASR systems for Bemba, we explored different approaches; supervised pre-training (training from scratch), cross-lingual transfer learning from a monolingual English pre-trained model using DeepSpeech on the portion of the dataset and fine-tuning large scale self-supervised Wav2Vec2.0 based multilingual pre-trained models on the complete Bem- baSpeech corpus. From our experiments, the 1 billion XLS-R parameter model gives the best results. The model achieves a word error rate (WER) of 32.91%, results demonstrating that model capacity significantly improves performance and that multilingual pre-trained models transfers cross-lingual acoustic representation better than monolingual pre-trained English model on the Bem- baSpeech for the Bemba ASR. Lastly, results also show that the corpus can be used for building ASR systems for Bemba language. Keywords: Automatic Speech Recognition, ASR corpus, Low-resource language, Bemba language, Zambia 1. Introduction Speech-to-Text, also known as Automatic Speech Recognition(ASR) or simply just Speech Recognition (SP), is the task of recognising and transcribing spoken utterances into text. In recent years, there has been a tremendous growth in popularity of speech-enabled ap- plications. This can be attributed to their usability and integration across wide domain applications, such as voice over control systems. However, building well- performing ASR systems typically requires massive amounts of transcribed speech, as well as large text cor- pora. This is generally not an issue for well-resourced languages such as English and Chinese, where ASR ap- plications have been successfully built with remarkable results (Amodei et al., 2016, et alia). Unfortunately, this is not the case for Africa and its over 2000 languages (Heine and Nurse, 2000). The preva- lence of speech recognition applications for African languages is very low. This can at least partially be attributed to the lack or unavailability of linguis- tic resources (speech and text) for most African lan- guages (Martinus and Abbott, 2019). This is particu- larly the case with Zambian languages. There exist no general speech or textual datasets curated for building natural language processing systems, including ASR systems. In this paper we present a speech corpus, BembaSpeech, consisting of over 24 hours of read speech in Bemba, a written but under-resourced language spoken by over 30% of the population in Zambia. We also present an end-to-end speech recognition model obtained by fine- tuning a large scale Wav2Vec2.0 based multilingual pre-trained model, XLS-R on BembaSpeech corpus. To our knowledge this is the first work carried out towards building ASR systems for …

Content truncated to 500 words