Gaussian-mixture-model-based spectral cue detection for modeling lexically guided learning

The acoustic cue patterns used to realize and distinguish among phonemes (and thus among target words) in a spoken utterance include the spectral characteristics of frication noise, which signal the difference between, e.g., /s/ and / ʃ/ in American English. As part of a larger project developing a system of modules for automatic detection and interpretation of such cues, this study focuses on enhancing an existing Gaussian-mixture-model-based module for detecting spectral bursts/frication (Tubbs et al. 2024), which is designed to detect the spectral characteristics of frication noise that distinguish between these two obstruent consonants. The model provides the posterior probability of the acoustic cue (place of articulation of spectral burst/frication) given the relevant speech-related measurements. These detection probabilities are used to decide the identity of the phoneme, given a sequence of speech stimuli containing ambiguous spectral bursts embedded within a lexical context. The distributions for /s/ and /ʃ/ are then updated to reflect these decisions. The transparent nature of the GMM-based framework enables explicit tracking of the mechanism that can be used to model lexically guided learning (Cummings and Theodore 2023).