Text-to-Speech (TTS) technology has advanced considerably, enabling natural and intelligible speech synthesis directly from text. This work presents the development of a high-quality Sinhala TTS system based on the VITS architecture, an end-to-end model that combines variational inference with adversarial training. To address the low-resource nature of Sinhala, the system was trained on Sinhala-script input derived from a curated subset of the Pathnirvana speech corpus. Three configurations were investigated: single-speaker male, single-speaker female, and multi-speaker models. Experimental evaluations, using both subjective measures—Mean Opinion Score (MOS) and Semantically Unpredictable Sentences (SUS)—and the objective Mel Cepstral Distortion (MCD) metric, indicate that the proposed system outperforms existing Sinhala TTS solutions. It achieved an MOS of 4.62 for intelligibility, 4.18 for naturalness, an SUS intelligibility score of 85.83%, and an average MCD of 13.27 dB, establishing a new benchmark for Sinhala speech synthesis.