<p>Machine learning procedures have become efficient at using acoustic information to distinguish modal speech, but their application on other speech modalities remains largely overlooked. Operating under the hypothesis that non-spoken speech classification requires a more robust acoustic analysis in which to train models, this study applies a spectrotemporal modulation (STM) analysis to examine its effectiveness at distinguishing modal and whistled Spanish speech. A linear mixed-effects model revealed that in comparison to modal speech, whistled speech had significantly increased spectrotemporal modulations generally above 1 cycles per octave. Features identified as significantly relevant for distinguishing speech modalities were extracted, and an automatic speech modality classification task was developed. STM- and MFCC-feature models performed similarly, boasting weighted accuracy performances of 92% and 94%, respectively. Although an STM analytical approach had not previously been applied to whistled speech, our results support existing evidence on its acoustic characteristics. Moreover, our findings have key implications for using STM-features for automatic speech modality classification tasks, as it reduced the feature space at little cost to performance.</p>