Text-to-Voice Conversion for Indic Languages

The demand for sophisticated Text-to-Speech (TTS) systems is rising due to linguistic diversity, with Indic languages posing particular difficulties. Conventional text-to-speech (TTS) synthesis algorithms frequently fall short in capturing the rich expressiveness and subtle cultural nuances of Indic languages, resulting in an artificial voice output devoid of authentic prosody. By integrating the HiFi-GAN for high-fidelity voice cloning and the Tortoise TTS framework into a hybrid TTS architecture, this solution suggests a novel way to get beyond these restrictions. Allowing the system to pick up on and replicate the complex patterns and nuanced subtleties of Indic speech. High-fidelity waveforms are effectively synthesized from mel-spectrograms by the HiFi-GAN vocoder, while extremely expressive and natural-sounding speech waveforms are produced by the Tortoise TTS model, which is based on Denoising Diffusion Restoration Models. High-quality training data is ensured via a complex data processing pipeline that includes silence reduction, audio segmentation, and transcription filtering. The suggested system performs much better than current TTS systems on both Mean Opinion Score (MOS) and Comparative Mean Opinion Score (CMOS) according to extensive subjective evaluations.