SARA: SYNTHETIC AZERBAIJANI RECORDING ALIGNMENT – A UNIFIED TTS FRAMEWORK TRAINED ON SYNTHETIC AND REAL SPEECH

Speech synthesis is crucial part of the modern assistive AI (Artificial Intelligence). While the development in this field has great burst, solutions for low-resource languages stays in background due to lack of the data. Our paper demonstrates the novel solution to overcome this problem for Azerbaijani Language and proposes a model which can generate fluent voice in Azerbaijani Language. Our solution is based on self-supervised learning with multi-lingual Text-to-Speech (TTS) model to pretrain the VITS (Variational Inference for Text to Speech) architecture from scratch. Additional Fine-tuning dataset has been obtained by segmenting voiced news articles along with text of the news by using whisper. Moreover, obtained datasets for fine-tuning and pre-training has been made available for further development and use-cases as an open-source. Model has been evaluated according to the LLM (Large Language Models) and Human Opinion by using Mean Opinion Scores (MOS) which proved efficiency of the approach and usability of the proposed model for further applications. Model is available at: https://huggingface.co/BHOSAI/SARA_TTS