L'accent cleaning par l'intelligence artificielle: Enjeux éthiques, responsabilité épistémique et reproduction des hiérarchies linguistiques dans les technologies vocales

Counts differ because each database indexes a different set of publications. We treat OpenAlex as the canonical count; Google Scholar is not shown (no API, and crawling it violates its ToS).

The widespread deployment of artificial intelligence–based speech technologies is profoundly reshaping contemporary language practices. These systems, trained predominantly on corpora drawn from standardized or geopolitically dominant varieties, tend to normalize phonetic features that diverge from the reference norm, a phenomenon known as accent cleaning. This contribution examines this process through two questions: to what extent do automatic speech recognition (ASR) systems contribute to the reproduction of linguistic and social hierarchies, and what responsibility do researchers in the language sciences bear in relation to these biases? The central hypothesis is that these models privilege certain phonetic norms at the expense of diversity, producing forms of symbolic exclusion for speakers with marginalized accents. An exploratory comparative analysis examines three ASR systems (Google Speech-to-Text, OpenAI Whisper, and Amazon Transcribe) using a speech corpus from speakers of three contrasting varieties of French (standard French, Moroccan French, and sub-Saharan French). The anticipated results suggest significant asymmetries in word error rates (WER) and mechanisms of active normalization. The study aims to document these biases empirically and examine the researcher’s role from a digital ethics perspective.