Chris Bartley, a School of Computer Science PhD student here at the University of Sheffield, has developed a speech and language AI text-to-speech system for endangered languages, with a particular focus upon Manx, the heritage language of the Isle of Man, where he originates from.
The developed technologies by Chris allow Artificial Intelligence systems to speak Manx for the first time, where, alongside other similar languages, data and resources are limited, if not scarce. Chris’ research tackles this problem head-on, investigating how to design more data-efficient approaches and make better use of what already exists rather than simply demanding more. The language itself, according to the 2021 census, found that around 2,200 people still speak the language today, and it is hoped that with Chris’ developed technologies and tools, it can help ‘lighten the load’ for some native Manx speakers.
The project required months of research, collecting and documenting Manx resources from a variety of sources, including YouTube, Manx Radio, and Learn Manx. In total, 350 hours of speech and eight million Manx words in order to lay the groundwork for the developed technologies.
Speaking about the project, Chris, who has a keen passion for languages and studied French, Spanish, and Mandarin at University, said: “We often hear about the negatives of AI, but text-to-speech technology has been shown to have clear long-term value. It makes language content more accessible for learners and teachers, and can enable assistive tools like screen reading for the visually impaired.
Until now, Manx has largely been excluded from modern speech and language technologies, mainly due to a lack of training data. Developing these systems with little to no data is a challenge, and addressing this is the focus of my PhD research.”
And having recently authored his first paper, “How I Built ASR for Endangered Languages with a Spoken Dictionary”, the project has resulted in an automatic speech recognition (ASR) system for the Manx language.
In the near future, Chris aims to release the tool publicly as well as automatic speech recognition models for community use, alongside the possibility of translation.