Balinese text-to-speech dataset as digital cultural heritage

Area of Science:

• Computational Linguistics
• Speech Technology
• Sociolinguistics

Background:

• Balinese language lacks representation in current speech technologies.
• Limited digital resources hinder Natural Language Processing (NLP) research and voice technology applications for Balinese.
• A significant challenge is the scarcity of comprehensive Balinese voice datasets.

Purpose of the Study:

• To create a diverse Balinese speech dataset featuring native speakers and various politeness levels.
• To support the development of Text-to-Speech (TTS), speech recognition, and voice-to-text technologies for Balinese.
• To facilitate NLP research and digital preservation of the Balinese language.

Main Methods:

• Audio data acquisition from native Balinese speakers using the Badung dialect.
• Application of denoising techniques to enhance audio quality.
• Categorization of recordings based on Balinese politeness levels (Alus Singgih, Alus Sor, Alus Mider, Mider, Andap), phrases, and alphabets.

Main Results:

• A dataset comprising 1187 audio recordings was successfully developed.
• The dataset captures a broad spectrum of social variations within the Balinese language.
• The resource includes diverse linguistic elements, enhancing its utility for various applications.

Conclusions:

• The created dataset is a valuable resource for advancing speech technologies for the Balinese language.
• This initiative contributes to the digital preservation of Balinese and fosters NLP research for low-resource languages.
• The dataset opens avenues for future research in computational linguistics and speech processing for regional languages.