Analyzing time-varying spectral characteristics of speech with function-on-scalar regression

The acoustic characteristics of noise from fricatives and stop releases are difficult to analyze. The spectral characteristics of such noise are multi-dimensional, and popular methods for analyzing them typically rely on reducing this complex information to one or a few discrete numbers, such as spectral moments or coefficients of discrete cosine transformations. In this paper, I propose using function-on-scalar regression models as a method for analyzing and mass-comparing spectra with minimal reduction of the complexity in the signal. The method is further useful for analyzing how spectra change as a function of time. The usefulness of this method is demonstrated with a corpus analysis of Danish aspirated stop releases, using the DanPASS corpus. The analysis finds that /t/ releases are invariably affricated; /k/ releases are highly affected by coarticulatory context; and /p/ releases are almost always dominated by aspiration in the latter half of the release, but are affricated in the first half in certain contexts.