Um sistema de avaliação de pronúncia via encoder-decoder transformer

Abstract

Mastery of accurate pronunciation remains a significant hurdle in both literacy acquisition and foreign language learning, as phonetic deviations (i.e., substitutions, omissions, or distortions of sounds) directly impact speech intelligibility, as well as academic and professional performance. Leveraging recent advancements in Automatic Speech Recognition (ASR) and phonetic processing techniques, this work describes the development of a system that converts voice recordings into International Phonetic Alphabet (IPA)-aligned phonetic transcriptions and assesses the accuracy of the expected pronunciation. The platform combines the Whisper-large-v3-turbo model for ASR with a forced alignment module, generating feedback that highlights both correctly produced phonemes and those requiring correction.