Generation of target speech with speaker individuality based on accent conversion for English pronunciation learning

This paper first provides an overview of the English pronunciation learning support tool. The tool aims to use "accent-modified speech that retains the learner’s voice quality" as the "target speech." We propose a new conversion model based on conventional methods for this accent conversion. Specifically, we improve the conventional LSTM-based DNN model for accent conversion by adopting a transformer-based model. Our experiments investigated the model’s ability to handle the unique katakana pronunciation characteristic of Japanese speakers. The results confirmed the effectiveness of the proposed conversion method, although challenges remain, such as the scarcity of Japanese speech data and the need to improve the accuracy of speaker identity retention.