Aaron Earned an Iron Urn: Speech-to-IPA Models Improve Diagnostic of Pronunciation

Abstract

Learning the proper pronunciation is one of the key aspects of foreign language acquisition. Assessment of the correctness of pronunciation requires the involvement of expert phoneticians and linguists, severely limiting the scalability of learning solutions. However, the recent adaptation of the Transformer architecture to the audio domain opens the way for automatic model-based assessment of pronunciation. In this paper, we present the pronunciation diagnostic tool developed at PUT and we experimentally evaluate the correlation between expert human assessment and automatic model assessment. By combining the Wav2Vec model and the IPA representation, we prove that pronunciation assessment can be performed automatically with high precision.