Benchmarking and Domain Adaptation of Automatic Speech Recognition (ASR) for Adolescent Health Communication in Ghanaian Languages

TL;DR

This paper presents a study on automatic speech recognition for adolescent health communication in three Ghanaian languages. The researchers benchmarked five ASR systems and performed supervised domain adaptation, resulting in significant improvements in word error rate. The work validates the importance of in-domain data for real-world applications.

Why It Matters

The research has practical significance for developing voice-first applications in low-resource languages, particularly for sensitive topics like adolescent health. It highlights the need for in-domain data and supervised domain adaptation to improve ASR performance. The findings can inform the development of more effective and culturally sensitive health communication systems.

Key Findings

• Fine-tuning reduced WER on every language
• Ewe language showed the most dramatic improvement with a 44.5 pp drop in WER
• In-domain data is the binding constraint for ASR performance in these languages