MASRI-HEADSET: A Maltese Corpus for Speech Recognition

Proceedings of the 12th Conference on Language Resources and Evaluation (LREC 2020), pages 6381–6388 Marseille, 11–16 May 2020 c⃝ European Language Resources Association (ELRA), licensed under CC-BY-NC 6381 MASRI-HEADSET: A Maltese Corpus for Speech Recognition Carlos Mena, Albert Gatt, Andrea DeMarco, Claudia Borg, Lonneke van der Plas, Amanda Muscat, Ian Padovani University of Malta {carlos.hernandez, albert.gatt, andrea.demarco, claudia.borg, lonneke.vanderplas amanda.muscat.11, ian.padovani.16}@um.edu.mt Abstract Maltese, the national language of Malta, is spoken by approximately 500,000 people. Speech processing for Maltese is still in its early stages of development. In this paper, we present the first spoken Maltese corpus designed purposely for Automatic Speech Recognition (ASR). The MASRI-HEADSET corpus was developed by the MASRI project at the University of Malta. It consists of 8 hours of speech paired with text, recorded by using short text snippets in a laboratory environment. The speakers were recruited from different geographical locations all over the Maltese islands, and were roughly evenly distributed by gender. This paper also presents some initial results achieved in baseline experiments for Maltese ASR using Sphinx and Kaldi. The MASRI-HEADSET Corpus is publicly available for research/academic purposes. Keywords: Maltese, Speech Corpora, Automatic Speech Recognition 1. Introduction As digital resources and tools for Natural Language Pro- cessing and language-enabled interfaces become ever more common in daily use and commercial applications, it has become increasingly important to ensure that all languages are adequately represented in the digital sphere. This con- cern is evident, for example, in a recent resolution passed by the European Parliament to safeguard language equality in the digital age (European Parliament, 2018). In this regard, so-called ‘under-resourced’ or ‘low- resourced’ languages have been an ongoing concern within some sectors of the NLP community. For example, in a series of white papers published in 2011-12, reviewing the level of digital support for 31 of the languages of the Eu- ropean Union, the METANET Initiative1 concluded that for a significant number of these languages, support in most areas of NLP was at best fragmentary and in some cases, weak or non-existent. The METANET papers took the step of determining the level of support for different languages with reference to specific tasks or domains, namely: machine translation, text anal- ysis, speech/language resources and speech processing. In principle, a language could have strong support in a sub- set of these, with weak or fragmentary support in others. This is reminiscent of the strategy used by Krauwer (2003), who discussed the notion of an ‘under-resourced’ language in terms of a minimal set of language resources (a Basic Language Resource Kit, or BLARK) necessary to undertake further precompetitive research and education. A some- what more wide-ranging definition was more recently of- fered by Besacier et al. (2014), in their review of ASR for low-resourced languages, where the criteria included the following: • Lack of a unique writing system or stable orthography. • Lack of linguistic expertise. • Limited presence on the web. 1http://www.meta-net.eu/ • Lack of electronic resources for speech and language processing. 1.1. The case of Maltese …

Content truncated to 500 words