Frisian Text-to-Speech Using Cross-Lingual Transfer Learning: An Experiment with 5 Source Languages

Sep 14, 2022·

Do, P.

Matt Coler

Dijkstra, J.

Klabbers, E.

· 1 min read

Cite

Abstract

Speech synthesis, also called text-to-speech (TTS), is the generation of artificial speech, with wide-ranging applications such as speech-enabled assistants and translators, language learning tools, and speech-based accessibility devices. Since the 2010s, neural-network based TTS (neural TTS) has been the dominant architecture, producing much higher speech quality than its previous paradigms (Tan et al., 2021). However, neural TTS usually requires a large amount of high-quality training data. Accordingly, it is not obviously applicable to low-resource languages (LRLs), ie, languages that lack sufficient training data to facilitate research in speech technology and thus may fall behind in the age of far-reaching speech applications. To overcome this, one approach is to train neural TTS models utilizing training data from other languages (called ‘source languages’) to make up for the lack of data in the language of interest.

Type

Conference Paper

Publication

BOOK OF ABSTRACTS

This conference abstract presents research on developing text-to-speech systems for Frisian, a low-resource language, using cross-lingual transfer learning techniques. The study examines how different source languages can be leveraged to improve speech synthesis quality when limited training data is available for the target language.

The research addresses the critical challenge of developing speech technology for under-resourced languages, which often lack the extensive data resources needed for modern neural text-to-speech systems. By exploring cross-lingual approaches, this work contributes to methods that can help bridge the technological gap for minority languages.

Last updated on Sep 14, 2022