LEVERAGING DEEP LEARNING FOR NATURAL MARATHI TEXT-TO-SPEECH SYNTHESIS
Authors/Creators
Description
Abstract
This research paper describes a deep learning-based method that uses an end-to-end architecture to generate high-quality Marathi TTS. Marathi is a morphologically rich Indo-Aryan language and the generation of speech in this language is challenging due to the shortage of annotated datasets, variations in dialectal patterns, and complex prosody. The authors have proposed a VITS-based TTS system that combines text normalization, phoneme conversion, and prosody-informed acoustic modeling to overcome these difficulties. Naturalness and intelligibility have been improved by training the system on a Marathi speech corpus comprising multiple speakers and dialects. The work also includes a comparative evaluation of the study outcomes with traditional concatenative methods and neural baselines such as Tacotron2. In order to evaluate the systems' precision from the point of view of sound reproduction, the authors resort to objective metrics such as Mel-Cepstral Distortion (MCD), F0 Root Mean Square Error (RMSE), and Character Error Rate (CER). The subjective Mean Opinion Score (MOS) tests are used to evaluate perceptual quality. The findings demonstrate that, compared to the baseline systems, significant enhancements in naturalness, clarity, and prosody modeling have been achieved. The current research serves as the basis for producing large-scale and linguistically complex TTS systems in low-resource Indian languages. The authors propose that their system be used in the creation of audiobooks, accessibility tools, virtual assistants, e-learning platforms, and government information services. Their subsequent work will involve broadening the dataset, enabling emotional TTS, and making the model suitable for on-device, real-time inference.
Keywords: Marathi TTS, Deep Learning, VITS, AI Speech Synthesis, NLP, Prosody Modeling.
Files
JAMRSD 05-01-2026_1.pdf
Additional details
Software
- Repository URL
- https://ijamrsd.com/