# From Raw Audio to Training-Ready Data

Canonical HTML: https://voicelk.dev/blog/from-raw-audio-to-training-ready-data

How VoiceLK’s speech recordings were organised and prepared into a consistent audio format for text-to-speech model training.

Author: T.H. Mihiran Thilakarathna

Published: 2026-10-21T00:01:57.345Z

Updated: 2026-10-21T00:05:06.212Z

## From Raw Audio to Training-Ready Data

*Preparing speech recordings for training a Sinhala–English text-to-speech model.*

A high-quality text-to-speech system depends not only on its model architecture but also on the quality and consistency of its training data.

For VoiceLK, preparing the speech dataset was an important step in building a system capable of generating Sinhala–English speech. The recordings needed to be organised and prepared in a consistent format before they could be used for model training.

Our dataset included recordings from two individual speakers, along with additional speech data from Pathnirvana. The Speaker 01 dataset contained **3,979 audio clips, approximately 6.2 hours of speech**, while the Speaker 02 dataset contained **3,087 clips, approximately 8.2 hours**. The Pathnirvana dataset contributed another **3,300 clips**.

To maintain consistency, the single-speaker recordings were prepared in **22.05 kHz, mono-channel, 16-bit PCM** audio format.

This preparation stage helped organise the recordings into a format suitable for the speech synthesis pipeline. Consistent audio data is important because differences in recording formats and quality can affect model training and the clarity of generated speech.

Preparing the data was more than collecting audio files. It was about building a structured foundation for teaching VoiceLK how written text corresponds to spoken language.

**A reliable voice begins with reliable data.** The quality of this preparation directly supports the next stage: training our first AI voices.