Acoustic characteristics of whispered vowels: A dynamic feature exploration

Abstract

Whispered speech exhibits distinct acoustic characteristics compared to phonated speech due to the absence of vocal cord vibration. Previous research primarily focused on static features, neglecting the crucial role of dynamic spectral changes in vowel perception. This study investigates the effectiveness of dynamic features, specifically Vowel Inherent Spectral Change (VISC), in characterizing whispered vowels. VISC, a robust measure of dynamic spectral change, captures the continuous variation of formant frequencies throughout vowel articulation. It provides a richer representation than static features measured at a single point. This study analyzes formant frequencies and two key VISC metrics − vector length (VL) and spectral angle (α) in whispered and phonated Japanese vowels. VL represents the magnitude of formant change, while α reflects the direction of formant movement in acoustic space. Additionally, a vowel classification experiment using a Support Vector Machine (SVM) is conducted to directly compare the performance of models trained on static features versus those trained on a combination of static and dynamic features. Results reveal the effectiveness of VISC as a dynamic feature for characterizing whispered vowels. Through VISC analysis, a vowel classification experiment, and an exploration of the acoustic space of whispered vowels, dynamic features exhibit greater robustness across different phonation types in contrast to traditional static features. This study highlights the potential of VISC for enhancing whispered speech recognition. It further contributes to a more comprehensive understanding of whispered vowel acoustics and informs the development of more effective whisper-related speech technology.