SpeechBlender: Speech augmentation framework for mispronunciation data generation
Published in Speech and Language Technology in Education Workshop (SLaTE 2023), 2022
A fine-grained data augmentation pipeline that generates mispronunciation errors by masked, mix-factor blending of phonetic units. Achieves SOTA on Speechocean762 (+2.0 PCC) and +4.6 F1 on Arabic AraVoiceL2.