Comparing LTASS across languages using natural and AI speech samples

Abstract

Hearing aid prescription protocols rely on estimates of speech spectra. One such estimate, the International Long-Term Average Speech Spectrum (ILTASS), is a common reference for the distribution of spectral energy within speech signals. However, recent research has suggested differences in speech spectra between some languages and the ILTASS standard. These disparities raise questions regarding the applicability of a language-general approach to hearing aid gain application. This study presents a comparative analysis of Long-Term Average Speech Spectra (LTASS) across various languages. Utilizing speech samples from both human and AI sources, LTASS was measured for speakers of different languages. Subsequently, these LTASS profiles were contrasted with the established ILTASS reference. Results reveal that while LTASS was very similar across languages, deviations were observed between LTASS derived from the human speech and AI speech data compared to ILTASS. Specifically, measured LTASS demonstrated a reduced 500 Hz peak in contrast to ILTASS. This research underscores the importance of accounting for spectral variations introduced during the recording process in the field of speech and hearing research. It also emphasizes the importance of standardized recording protocols to enhance the precision of hearing technologies.