Dynamic spectral features stabilize Japanese vowel recognition across whisper and normal voice

Whispered speech strips away the vocal-fold vibration that anchors most acoustic vowel analysis, and a new study of Japanese vowels shows why static measurements struggle under that shift. Researchers compared conventional formant-based descriptors with dynamic spectral features drawn from Vowel Inherent Spectral Change (VISC), which tracks how formant trajectories evolve continuously over the course of a vowel rather than sampling it at fixed points.

The evaluation covered vowels produced in both normal and whispered phonation. Static formant measures, which summarize the spectrum at discrete time points, proved less consistent when the phonation mode changed. The dynamic spectral representations, by contrast, preserved more stable vowel discrimination across the two conditions, suggesting that the time-varying structure of the spectrum carries information that survives the loss of voicing.

Rather than pursuing algorithmic optimization, the team used Support Vector Machine (SVM) and Random Forest (RF) classifiers as analytical probes to test how robust each representation was. Similar performance trends appeared across both classification frameworks, indicating that differences in outcomes were driven primarily by the properties of the acoustic representations themselves rather than by the choice of model.

The authors argue that this points to representation-level design as a key lever for robust speech signal analysis under phonation-induced variability. Low-dimensional dynamic features, they conclude, offer a practical foundation for vowel modeling in conditions where phonation varies, including whispered speech, where traditional static descriptors often fail to capture the dynamic spectral behavior of vowels.

Subject of Research: Robust acoustic vowel representation in Japanese under normal and whispered phonation using dynamic spectral analysis

Article Title: Robust acoustic representation of Japanese vowels under phonation variability using dynamic spectral analysis

Article References: Zhang, Y., Cao, T., & Liu, Y. (2026). Robust acoustic representation of Japanese vowels under phonation variability using dynamic spectral analysis. PLOS One, 21(10), e0361125. https://doi.org/10.1371/journal.pone.0361125

Image Credits: AI Generated

DOI: 10.1371/journal.pone.0361125

Keywords: phonetics, Japanese vowels, whispered speech, formants, Vowel Inherent Spectral Change, acoustic features, Support Vector Machine, Random Forest, phonation variability, speech signal analysis, Robust, acoustic