The French-Algerian Code-Switching Triggered audio corpus (FACST)

The French-Algerian Code-Switching Triggered audio corpus (FACST) Djegdjiga Amazouz1, Martine Adda-Decker1,2, Lori Lamel2 1LPP-CNRS Université Sorbonne Nouvelle Paris-III 2LIMSI, CNRS, Paris Saclay University, Orsay, France djegdjiga.amazouz@univ-paris3.fr, {madda, lamel}@limsi.fr Abstract The French Algerian Code-Switching Triggered corpus (FACST) was created in order to support a variety of studies in phonetics, prosody and natural language processing. The first aim of the FACST corpus is to collect a spontaneous Code-switching speech (CS) corpus. In order to obtain a large quantity of spontaneous CS utterances in natural conversations experiments were carried out on how to elicit CS. Applying a triggering protocol by means of code-switched questions was found to be effective in eliciting CS in the responses. To ensure good audio quality, all recordings were made in a soundproof room or in a very calm room. This paper describes FACST corpus, along with the principal steps to build a CS speech corpus in French-Algerian languages and data collection steps. We also explain the selection criteria for the CS speakers and the recording protocols used. We present the methods used for data segmentation and annotation, and propose a conventional transcription of this type of speech in each language with the aim of being well-suited for both computational linguistic and acoustic-phonetic studies. We provide an a quantitative description of the FACST corpus along with results of linguistic studies, and discuss some of the challenges we faced in collecting CS data. Keywords: Code-switching, bilingual speakers, oral speech data, French, Arabic 1. Introduction When multilingual speakers communicate using two shared languages in a conversation, they may switch between these languages in a same sentence or utterance. Switching may become a common practice in social groups with strong language contact. This simultaneous use of two language codes can take several forms at the lexical, syntactic and communication levels, and is commonly referred to as code-switching (CS) (Muysken, 2000; Grosjean, 1995). CS may present complex structures in particular at mor- phological, syntactical and phonetic levels. Hence, CS has interested several research fields such as sociolinguis- tics and interactional linguistics (Gumperz, 1982; Auer, 2010; Bullock, 2012; Gardner-Chloros, 2009; Sebba, 2012; Tabouret-Keller and Page, 1970; Ziamari, 2008; Piccinini, 2012; Gullberg et al., 2012) with numerous studies de- scribing CS phenomena, explaining the process of its re- alization and analyzing its productions. Computational lin- guistics enable the study of large textual and spoken cor- pora (Çetino˘glu et al., 2016; Schultz and Kirchhoff, 2006). Consequently, experiments have been carried out on lan- guage identification (LID) for both oral and written CS as well as automatic identification of CS languages pairs (Lyu et al., 2015; Lyu and Lyu, 2008; Modipa et al., 2013). Fur- thermore, quantifying CS in large corpora at the word and oral segment levels gives a precise idea about the use of CS in bilingual communities (Amazouz et al., 2016). It also allows the determination of which language is dominant in the CS speech and in the daily CS practices. Applying lan- guage processing tools can help to compare CS in differ- ent …

Content truncated to 500 words