From Annotation to Automation: Preparing Connacht Irish for Automatic Speech
Recognition
What does it take to teach a computer to recognize a language? While recent advances
in AI can make language technology appear almost effortless, effective automated
systems depend on a less visible foundation: carefully prepared linguistic data.
This talk presents our work on Connacht Irish using Besra, a tool we have developed for
rapid audio and text processing and model training. We will demonstrate how
recordings are transcribed, annotated, standardized, and prepared for training, and how
this careful human work subsequently enables automatic phonetic transcription.
Rather than focusing on large language models, the talk highlights how important the
pre-AI stages of creating computational linguistic resources are.
While this workflow focuses on language documentation for low-resource languages,
the talk also aims to demonstrate how such a workflow may be useful within a
classroom for language teaching. By training a model in this way, Besra may be useful
for pronunciation practice, identifying phonetic di:erences between speakers of a
language, and the creation of searchable spoken-language corpora.
By demonstrating the methodology we use to transform audio recordings for Connacht
Irish into an automated system, the talk illustrates how computational methods can
build on the careful work of linguists and language teachers, and how this combination
can create new possibilities for language teaching and research.
The guest lecture will be held in English. No prior registration is necessary.
