Towards Better Recognition of Spontaneous Children's Speech: Speaker-Clustering Fine-Tuning of Whisper
Year: 2024
Type: conference-paper
Abstract: End-to-end based models have recently shown promising improvements in the area of child speech recognition. How-ever, the challenge of spontaneous child speech recognition remains unsolved because of the large variability in children's acoustic and linguistic features. In this study, we examine the effects of two different speaker grouping methods to im-prove the fine-tuning performance of Whisper model. Our re-sults indicate that the inter-speaker variability may not be per-fectly represented by age. Moreover, we examine the effects of partial parameter fine-tuning of Whisper models demon-strating better results on spontaneous child speech compared with state-of-the-art fine-tuning methods.
Institution: Norwegian University of Science and Technology
Language: English
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W4404037595