BULBasaa: A Bilingual Basaa-French Speech Corpus for the Evaluation of Language Documentation Tools

BULBasaa: A Bilingual B`as`a´a-French Speech Corpus for the Evaluation of Language Documentation Tools Fatima Hamlaoui⋄‡, Emmanuel-Moselly Makasso‡, Markus M¨uller∗, Jonas Engelmann‡, Gilles Adda†, Alex Waibel∗⋆, Sebastian St¨uker∗ ⋄University of Toronto, Toronto, Canada ‡Leibniz-Zentrum Allgemeine Sprachwissenschaft, Berlin, Germany ∗Karlsruhe Institute of Technology, Karlsruhe, Germany †LIMSI, CNRS, Paris, France ⋆LTI, CMU, Pittsburgh, USA contact: f.hamlaoui@utoronto.ca, m.mueller@kit.edu, sebastian.stueker@kit.edu Abstract B`as`a´a is one of the three Bantu languages of BULB (Breaking the Unwritten Language Barrier), a project whose aim is to provide NLP-based tools to support linguists in documenting under-resourced and unwritten languages. To develop technologies such as automatic phone transcription or machine translation, a massive amount of speech data is needed. Approximately 50 hours of B`as`a´a speech were thus collected and then carefully re-spoken and orally translated into French in a controlled environment by a few bilingual speakers. For a subset of ≈10 hours of the corpus, each utterance was additionally phonetically transcribed to establish a golden standard for the output of our NLP tools. The experiments described in this paper are meant to provide an automatic phonetic transcription using a set of derived phone-like units. As every language features a specific set of idiosyncrasies, automating the process of phonetic unit discovery in its entirety is a challenging task. Within BULB, we envision a workflow where linguists are able to refine the set of automatically discovered units and the system is then able to re-iterate on the data, providing a better approximation of the actual phone set. Keywords: B`as`a´a, Northwest Bantu, Computational linguistics, unsupervised phone discovery, under-resourced languages 1. Introduction About two-thirds of the approximately 7,000 languages spoken today count less than 1,000 speakers and are in dan- ger of disappearing (Simons and Fennig, 2017). In a world that is rapidly changing, some estimate that 70% to 90% of the languages spoken today will fall out of use by the end of this century. To provide communities with mod- ern tools that will increase the vitality of their language and support its use in a variety of contexts, as well as to help lin- guists in their efforts to learn about human cognition though the study of language diversity, the BULB (Breaking the Unwritten Language Barrier) project has been developing Natural Language Processing-based tools to help and ac- celerate language documentation. To achieve this goal, a critical mass of speech data is necessary. In the present pa- per, we describe the corpus created for B`as`a´a (Cameroon), one of the three under-resourced Bantu languages on which the project concentrates and outline a method for the un- supervised evaluation of the derived phone set. The other two languages part the of the project are Myene (Gabon) and Mboshi (Republic of Congo) (Godard et al., 2018). We first provide basic information about B`as`a´a in Section 2. In Section 3., we describe the data collection process and pro- vide a detailed overview of the corpus. In the remainder of the paper, we concentrate on the subset of our data that was used for evaluation. Data preparation steps …

Content truncated to 500 words