# Designing the VoiceLK Approach

Canonical HTML: https://voicelk.dev/blog/designing-the-voicelk-approach

Exploring the hybrid architecture that combines rule-based pronunciation handling with neural speech synthesis.

Author: T.H. Mihiran Thilakarathna

Published: 2026-10-20T23:45:31.943Z

Updated: 2026-10-20T23:45:32.145Z

## Designing the VoiceLK Approach

*Combining the precision of pronunciation rules with the naturalness of neural speech synthesis.*

After identifying the limitations of existing Sinhala Text-to-Speech systems, we needed an approach that could handle Sinhala–English code-switching more effectively.

VoiceLK was designed around a hybrid framework that combines **rule-based language processing** with **neural speech synthesis**.

The rule-based component handles text normalisation, language identification, pronunciation rules, acronyms, and technical terminology. The neural component uses the VITS architecture to transform phoneme sequences into spoken audio.

By combining these two components, our goal was to make mixed-language pronunciation more predictable while generating clear, natural-sounding speech.

This design became the foundation for VoiceLK's text-processing pipeline, pronunciation lexicon, speech dataset preparation, model training, and web-based application.

**The principle behind our approach was simple: reliable pronunciation handling, combined with neural voice generation, to support more natural Sinhala–English speech.**

In the next stage, we explore how we built the language intelligence behind VoiceLK.