Modeling regional accents of French for inclusive speech recognition

Abstract

Automatic speech recognition (ASR) systems do not perform equally well for all speakers, which can reinforce existing feelings of social exclusion. Regional accents, or differences in pronunciation, are believed to be an important source of gaps in performance for multiple languages. In this work, I explore the relationship between regional accent and the performance of a French-language DNN-HMM speech recognition system. To characterize performance, I first calculated the Word error rate (WER) of 133 speakers from across France, Switzerland, Canada, Burkina Faso, Côte d’Ivoire, and Cameroon for the same written text. Then, to operationalize regional accent, I extracted phonetic alignments and posterior probabilities for the same speech samples. Following that, I explored how changing the level of phonetic representation in the acoustic model, language rescoring, and modifying the pronunciation lexicon impacted WER across different regions. Overall, there was considerable variation within regions with respect to both pronunciation and WER, which made it difficult to establish direct relationship between accent and performance. Furthermore, both region-specific and language-wide pronunciation modifications were able to lower WER for speakers with non-normative accents. For future work, I suggest feature-level simulation of the acoustic model to establish a relationship between specific accent features and ASR accuracy.