A Traditional Machine Learning Framework for Kannada Text-to-Speech Synthesis
Main Article Content
Abstract
Speech synthesis plays a pivotal role in assistive technology, particularly for individuals with visual impairments and literacy challenges. This research presents a machine learning-based text-to-speech (TTS) system designed specifically for the Kannada language, aiming to generate clear, intelligible, and natural-sounding speech without relying on complex deep learning architectures. The proposed system leverages classical machine learning models, including Decision Tree Regression and K-Nearest Neighbors, to predict prosody features such as pitch and duration for each phoneme extracted from Kannada text. These predicted parameters guide a concatenative synthesis engine, which assembles prerecorded phoneme audio segments to produce fluent speech. The system architecture emphasizes interpretability, efficiency, and low computational overhead, making it suitable for deployment in lowresource environments such as embedded devices or regional mobile applications. Experimental evaluations confirm that the model can generate understandable Kannada speech while maintaining flexibility and ease of training, demonstrating the potential of traditional machine learning in regional language speech generation.