# Where VoiceLK Goes Next

Canonical HTML: https://voicelk.dev/blog/where-voicelk-goes-next

Reflecting on VoiceLK’s achievements, current limitations, and future opportunities to improve Sinhala–English text-to-speech.

Author: T.H. Mihiran Thilakarathna

Published: 2026-10-21T02:00:05.656Z

Updated: 2026-10-21T02:00:05.867Z

## Where VoiceLK Goes Next

*Reflecting on what we have built and exploring the future of Sinhala–English text-to-speech.*

VoiceLK began with a practical challenge: existing Sinhala text-to-speech systems can struggle when English technical terms appear within Sinhala sentences. Our goal was to explore how a combination of language-processing rules and neural speech synthesis could help address this problem.

Throughout this journey, we developed a hybrid pipeline combining text normalisation, language routing, custom grapheme-to-phoneme conversion, and the VITS neural speech synthesis architecture. We also prepared speech datasets, built a technical-term pronunciation lexicon, trained our model, and integrated the components into a software application.

Our experiments produced encouraging results, including **100% language-routing accuracy on 22 test tokens**, **98.9% English lexicon fidelity**, and **94.8% Sinhala lexicon fidelity** in the reported evaluations. These results represent important milestones in our development, while also reminding us that individual metrics do not fully describe the quality of a generated voice.

There is still work to do. Future improvements could focus on increasing speech intelligibility, strengthening text normalisation, improving robustness across different input sentences, and evaluating generated speech with human listeners. Additional data preparation and testing would also help us better understand how the system performs in practical situations.

We see VoiceLK as more than a model-training exercise. It is an opportunity to explore how language technology can support Sinhala speakers when learning, reading, or working with English technical terminology.

The next stage is about moving beyond promising experiments towards a more thoroughly evaluated, accessible, and dependable application.

**VoiceLK is a continuing journey—built through research, experimentation, and the belief that language technology should work better for the people who use it.**