# Finding the Gap in Sinhala Voice Technology

Canonical HTML: https://voicelk.dev/blog/finding-the-gap-in-sinhala-voice-technology

Exploring the evolution of Sinhala text-to-speech systems, their limitations, and the research gap that inspired VoiceLK.

Author: T.H. Mihiran Thilakarathna

Published: 2026-10-20T23:25:56.411Z

Updated: 2026-10-20T23:25:56.614Z

## Finding the Gap in Sinhala Voice Technology

*Sinhala speech technology has come a long way. But could existing systems handle the way Sri Lankans actually speak about technology?*

In our previous post, we explored the challenge that inspired VoiceLK: making Sinhala–English mixed-language speech sound natural. To understand how we could address this problem, we first needed to look at what had already been achieved in Sinhala Text-to-Speech (TTS) research.

We discovered that Sinhala speech synthesis was not a new idea. Researchers had spent years developing systems that could transform Sinhala text into spoken audio. Each generation brought new techniques, new possibilities, and improvements in speech quality.

However, one important question remained: **Could these systems naturally pronounce English technical terms when they appeared in the middle of a Sinhala sentence?**

## How Sinhala Text-to-Speech Has Evolved

The development of Sinhala TTS reflects a gradual shift from traditional speech synthesis techniques towards neural approaches.

Some of the systems examined in our literature review included:

- **Festival-si (2007):** An early Sinhala TTS system based on diphone concatenation.
- **MaryTTS (2018):** Explored unit-selection techniques for Sinhala speech synthesis.
- **Deep Voice 3 (2019):** Represented the use of neural approaches in speech generation.
- **TacoSi (2023):** Applied the Tacotron architecture to Sinhala TTS.
- **VAENAR-TTS (2024):** Explored another neural approach, although the reported system experienced humming-like output.
- **Sinhala VITS (2025):** Demonstrated further progress using an end-to-end neural architecture.

These developments showed us that Sinhala speech synthesis was progressing. Neural models, in particular, offered new opportunities to improve how synthetic speech sounded.

Yet, better speech synthesis did not automatically solve the problem of mixed-language pronunciation.

## The Missing Piece: Sinhala–English Code-Switching

Think about a student learning ICT through a Sinhala explanation. The sentence may contain Sinhala grammar alongside English words such as *database*, *algorithm*, *CPU*, or *network*.

For a human speaker, switching between these languages is natural. We recognise the words, understand their context, and pronounce them accordingly.

A speech synthesis system faces a different challenge.

A system designed primarily for pure Sinhala text may struggle when English terms, acronyms, technical symbols, or mixed-language expressions appear within the same sentence. Pronunciation errors can interrupt the flow of speech and make educational explanations less effective.

Our literature review highlighted several connected limitations:

**1. Limited code-switching support**

Existing Sinhala TTS research did not provide the integrated solution we wanted for naturally speaking Sinhala and English technical content together.

**2. A lack of mixed-language training data**

The reviewed systems frequently relied on Sinhala-focused datasets. This left an important challenge in preparing speech data that represented natural Sinhala–English code-switching.

**3. Difficult pronunciation rules**

Sinhala spelling does not always map directly to pronunciation in a simple, one-to-one way. English technical terms, abbreviations, and acronyms introduce additional complexity.

**4. Incomplete text normalisation**

Numbers, dates, URLs, technical symbols, units, and version numbers need to be interpreted correctly before they are converted into speech. If the text is processed incorrectly, even a good speech model may produce an unnatural result.

**5. Naturalness and deployment challenges**

Producing intelligible speech is only part of the problem. A useful system must also maintain a consistent voice, handle longer inputs, and operate efficiently enough for practical use.

## Turning the Research Gap Into an Opportunity

These findings helped us define the direction of VoiceLK.

Instead of treating pronunciation and speech generation as a single problem, we explored a hybrid architecture that combines two complementary ideas.

The first is a **rule-based language-processing component** that handles text normalisation, language routing, pronunciation rules, acronyms, and technical vocabulary.

The second is a **neural speech model** that converts phoneme sequences into spoken audio.

By combining these components, our goal was to make pronunciation more predictable while using neural speech synthesis to produce a more natural voice.

This approach also shaped our data strategy. We needed speech recordings and transcripts that represented the way Sinhala and English are actually combined in technical communication—not simply a collection of isolated Sinhala sentences.

## From Literature Review to Research Direction

Identifying the research gap was one of the most important steps in the VoiceLK journey. It helped us move beyond asking whether we could build a Sinhala voice and towards a more specific research question:

*How can we develop a practical text-to-speech framework that naturally handles Sinhala–English code-switched technical content?*

That question became the foundation for our hybrid approach, pronunciation lexicon, dataset preparation, neural voice training, and evaluation process.

The existing research gave us a starting point. The limitations we identified gave us a direction.

**Our next challenge was to turn that direction into a system design.** In the next post, we will explore how we designed the VoiceLK approach and why we chose to combine rule-based pronunciation handling with neural speech synthesis.

*VoiceLK — Building a more natural voice for Sinhala–English digital learning.*