KazTTS

This work presents a cross-domain system for emotion prediction and speech synthesis for Kazakh, a low-resource language in AI research.

By fine-tuning the ai- forever/mGPT- 1.3B-Kazakh model for emotion prediction from text inputs and integrating it with the facebook/mms-tts-kaz and BigVGAN models, we generate emotionally expressive speech. Leveraging the KazEmoTTS dataset, ourapproach demonstrates high emotion prediction accuracy and enhanced naturalness in synthesized speech, bridging a critical gap in Kazakh language resources.