Researchers are strengthening Scottish Gaelic resources – using automatic speech and handwriting recognition to advance Gaelic language technology
EFI supports interdisciplinary and data-driven research which focusses on navigating an increasingly complex future. Lying at the intersection between ethnography, linguistics and data-driven innovation, Dr Will Lamb and his team are creating and refining the world’s first Scottish Gaelic Speech Recognition System.
Automatic Speech Recognition
Automatic Speech Recognition (ASR) technology can be used to translate spoken language into written text. It is used for many purposes in the lives of majority language speakers, for example in subtitling, voice assistant software and dictation services. For minority languages however, these services are often unavailable or inaccurate.
ASR and minority languages
ASR technology needs to be ‘trained’ with real language input to become more accurate. ASR systems analyse spoken language data, for example, audio recordings from native speakers, and learn about its patterns and structures. The more natural language data there is available, the more accurate the ASR technology can become.
For majority languages such as English there is a wealth of ‘real world’ language data available, for example, from literature, television, news and radio. This data can be used for training an increasingly accurate ASR system. But for minority languages such as Scottish Gaelic there is usually less of this data available. This means that ASR systems for minority languages struggle to reach the high levels of accuracy that are possible for majority languages.
ASR and Scottish Gaelic
In comparison with other minority languages, Scottish Gaelic has a surprising level of language technology available. Over the past 10 years, researchers have developed a Scottish Gaelic handwriting recogniser to help transcribe handwritten manuscripts into digital text and a speech recogniser to transcribe audio files and an ‘aligner’ to create timestamps for the audio represented by the transcription.
Through the process of building the Scottish Gaelic handwriting recogniser, Dr Will Lamb and his team were able to create a database of natural language data which could be used to train ASR technology for Scottish Gaelic. One use of this has been for applying Gaelic subtitles to programmes for MG ALBA, a Gaelic media outlet.
The future of ASR and Scottish Gaelic
Improving the accuracy of ASR systems for Scottish Gaelic is important for multiple reasons. Accurate automatic subtitling in Scottish Gaelic will enable native speakers and language learners to use Gaelic media to maintain and develop their language skills. It will also make Gaelic programming more accessible to d/Deaf audiences.
Improved ASR technology could also be used to improve dictation services, which allow spoken language to be converted into text automatically by a web tool. This could be used to support children with learning difficulties in school.
Conclusion
Building accurate ASR technology for Scottish Gaelic is important for supporting native speakers and language learners. It can also help document and revitalise Scottish Gaelic as a minority language, support its use and encourage the development of useful resources for language learners.
“One of the aims of EFI is to address the challenges and opportunities posed by data driven innovation in the Arts, Humanities and Social Sciences. Will’s project is an excellent example of how computer-based technology and digital methods can be applied in to the study of human culture”
Professor Melissa Terras, Research Director (EFI)
Researcher profiles
Professor William Lamb is an EFI Research Affiliate and Personal Chair in Gaelic Ethnology and Linguistics. His research interests lie within Scottish oral tradition, Gaelic linguistics, and Natural Language Processing. His work is linked with EFI’s Creative Industries theme which supports work which links data-driven methods with creative and heritage industries.
Dr Beatrice Alex is Senior Lecturer and Chancellor’s Fellow at EFI and LLC. She leads the Edinburgh Language Technology Group and the Clinical NLP Group. Her research focuses on text mining and natural language processing to extract information from raw text.
Acknowledgements
The investigators would like to thank Michael Bauer, Lucy Evans and Dr Mark Sinclair for their contributions to this work.
This work has been supported by the Arts and Humanities Research Council [grant number AH/W001934/1], the Soillse Research Fund and DDI/SFC. Collaborators and partners include: Am Faclair Beag, Ceòlas Uibhist Ltd, European Ethnological Research Centre, Grace Note Publications, Guthan nan Eilean / Island Voices, LearnGaelic (MG Alba), National Folklore Collection (University College Dublin), Ruairidh MacIllEathain, Sabhal Mòr Ostaig, The National Library of Scotland, The School of Scottish Studies Archives, Tobar an Dualchais / Kist o Riches, University of the Highlands and Islands and Prof Wilson McLeod.
The work has also been supported by the Data Driven Innovation network, which supports research innovation at the intersection between academic disciplines and real-world challenges.
