Automatic Speech Recognition and Query By Example for Creole Languages Documentation

Findings of the Association for Computational Linguistics: ACL 2022, pages 2512 - 2520 May 22-27, 2022 c⃝2022 Association for Computational Linguistics Automatic Speech Recognition and Query By Example for Creole Languages Documentation Cécile Macaire1, Didier Schwab1, Benjamin Lecouteux1, Emmanuel Schang2 1Univ. Grenoble Alpes, CNRS, Grenoble INP, LIG, 38000 Grenoble, France 2LLL, UMR 7270, Univ. Orléans & CNRS 1first.last@univ-grenoble-alpes.fr 2emmanuel.schang@univ-orleans.fr Abstract We investigate the exploitation of self- supervised models for two Creole languages with few resources: Gwadloupéyen and Morisien. Automatic language processing tools are almost non-existent for these two lan- guages. We propose to use about one hour of annotated data to design an automatic speech recognition system for each language. We evaluate how much data is needed to obtain a query-by-example system that is usable by linguists. Moreover, our experiments show that multilingual self-supervised models are not necessarily the most efficient for Creole languages. 1 Introduction There is a long tradition of description of creole languages since, at least, the pioneering work of Hugo Schuchardt (1842-1927). Creole languages have sometimes been assigned a special role in lin- guistics: as a type of ‘mixed languages’, they are often considered as illustrating a break in language transmission and do not fit the generally assumed historical/genetic tree model1. Meanwhile they remain, for many of them, under-resourced lan- guages. Gwadloupéyen, spoken mainly on Guade- loupe Island (France) by around 700.000 speakers, is a vigorous but largely under-equipped and under- resourced language. Morisien, (Mauritius Island) is spoken by approximately a million speakers. These two languages still suffer from a low social status and remain mainly spoken languages (rather than written). The CREAM project aims at providing linguists with new methods for computational language doc- umentation: Automatic Speech Recognition and keyword-spotting (Query-by-Example, Ram et al., 2020). The CREAM project teams up field lin- 1But see DeGraff (2004) or Corcoran (2001) for a severe criticism of the "creation myth". guists with computer scientists in order to address this low resource challenge. Our contribution is twofold: 1) we introduce new methods for creole language documentation based on a combination of automatic speech recogni- tion and keyword-spotting (in particular Query-by- example QbE) ; 2) our experiments shed new light on some key assumptions about creole languages, i.e. its distance from the lexifier language2. Gwad- loupéyen (gcf) and Morisien (mfe) are two French- based creole languages. This category groups lan- guages that share French as the most important part of their lexicon, but have a significantly different grammar. We illustrate this phenomenon in (1-a.) and (1-b.) for gcf and mfe respectively, where most of the words (if not all) are clearly identifiable by a French speaker despite the difference in orthogra- phy: (1) a. fo need ou 2SG desann go.down Gwadloup Guadeloupe ‘you’ve got to come to Guadeloupe" b. Zan John kontign continue reste. stay ‘John continues to stay.’ (from Henri and Kihm, 2015) While these languages are well studied and vig- orous, they are mostly spoken languages used in context of a dominant language: French for Gwad- loupéyen (see …

Content truncated to 500 words