An Automatic Pipeline For Building Emotional Speech Dataset

Emotional speech synthesis has gained attention for enhancing the naturalness and expressiveness of synthesized speech. However, developing emotional speech synthesis presents challenges due to the lack of high-quality emotional speech datasets, especially for low-resource languages. This paper proposes an automated process for building datasets of emotional speech that reduces dataset generation time by 92%. The pipeline combines trained models and leverages publicly available data. The effectiveness of the proposed pipeline is examined and demonstrated using a Vietnamese case study.