Let Me Read You a Story in Your Mother Tongue! Kids Story Reader in Sorani Kurdish (Central Kurdish)
Bala Farhad, Hossein Hassani · Qeios · 2024
Text-to-speech (TTS) synthesis is the technique of generating synthetic speech from input text. Developing a TTS system for Sorani (Central) Kurdish is a challenge due to the lack of resources for the language. In this research, we assess the development of a storytelling TTS system in Sorani Kurdish for children aged five to ten by comparing two different TTS methodologies and technologies. We select proper children’s storybooks to build a Sorani storytelling TTS system. We used two female narrators to narrate the stories, based on which we created the necessary datasets that the two chosen TTS frameworks use. The collected records are nearly seven hours long and are pre-processed, segmented, and aligned with the transcribed texts. The final dataset includes approximately five hours of speech consisting of 4149 speech segments and 34,523 words. We use Tacotron2 and Variational Inference with adversarial learning for end-to-end Textto-Speech (VITS) frameworks. We evaluated the results objectively and subjectively. The results indicate that the sound quality of the VITS-based model and its understandability outperforms the Tacotron2 model by a mean opinion score of 3.41 versus 1.91 for Tacotron2. We attribute that to two factors: the amount of training data and the training period.