L'accent cleaning par l'intelligence artificielle : enjeux éthiques, responsabilité épistémique et reproduction des hiérarchies linguistiques dans les technologies vocales AI-Powered Speech Recognition: Ethical Issues, Epistemic Responsibility, and the Reproduction of Linguistic Hierarchies in Voice Technologies

Abstract

The widespread deployment of AI based speech technologies is profoundly reshaping contemporary language practices. These systems, primarily trained on corpora from standardized or dominant varieties, tend to normalize phonetic features that deviate from the reference norm a phenomenon known as "accent cleaning." This paper investigates this process through a dual question: to what extent do automatic speech recognition (ASR) systems contribute to the reproduction of linguistic and social hierarchies, and what responsibility do language researchers bear in the face of these biases? Our central hypothesis is that these models prioritize certain phonetic norms at the expense of diversity, producing forms of symbolic exclusion for speakers of marginalized accents. An exploratory comparative analysis is conducted on three ASR systems (Google SpeechtoText, OpenAI Whisper, Amazon Transcribe), based on a speech corpus from French speaking speakers of three contrasting varieties (Standard French, Moroccan French, Sub-Saharan French). Anticipated results suggest significant asymmetries in word error rates (WER) and active normalization mechanisms. The study aims to empirically document these biases and to examine the researcher's role within a digital ethics framework.