Can speech foundation models effectively identify languages in low-resource multilingual aging populations?

Area of Science:

• Computational Linguistics
• Speech Technology
• Gerontology

Background:

• Speech foundation models (SFMs) excel in various tasks but their efficacy with elderly, multilingual speech is not well-understood.
• Spoken language identification is used as a proxy task to evaluate SFMs on this demographic.

Purpose of the Study:

• To investigate the performance of SFMs on multilingual speech from older adults.
• To identify key qualities for SFMs serving aging populations: robustness to duration, demographic invariance, and few-shot transferability.

Main Methods:

• Zero-shot evaluation of SFMs on spoken language identification using elderly, multilingual speech.
• Few-shot learning experiments to assess transferability in different model sizes.
• Analysis of performance variations based on input duration and speaker nativity.

Main Results:

• A significant performance decrease was observed for shorter speech inputs in zero-shot evaluations.
• Native speakers' speech yielded consistently better results than non-native speakers across all languages.
• Few-shot learning demonstrated improved transferability, particularly with larger SFM architectures.

Conclusions:

• SFMs require further development to effectively process the nuances of elderly, multilingual speech.
• Robustness to input duration and speaker nativity are critical areas for improvement.
• Larger models and few-shot learning strategies show promise for enhancing SFM performance in low-resource scenarios for aging populations.