Discrete representation of perceptual structure underlying consonant confusions

Area of Science:

• Speech Perception
• Acoustic Phonetics
• Psychometrics

Background:

• Speech perception is often modeled as discrete, necessitating analytic models for sound similarities.
• Existing models may not fully capture the complexity of perceptual similarities across conditions.

Purpose of the Study:

• To apply the INDCLUS (INdividual Differences CLUStering) model to discrete analytic modeling of speech sound perception.
• To represent perceptual similarities in consonant sounds using a discrete, cluster-based approach.
• To investigate how stimulus degradation affects the perceptual attributes of consonants.

Main Methods:

• Utilized the INDCLUS model to analyze three-way similarity data (stimuli x stimuli x conditions).
• Applied INDCLUS to the classic Miller and Nicely consonant confusion data.
• Analyzed a 14-cluster solution explaining 82.9% of the variance across 17 listening conditions.

Main Results:

• The INDCLUS model successfully generated a discrete representation of consonant perceptual structure.
• Cluster weights varied with stimulus degradation, indicating differential impact on common and unique perceptual attributes.
• Low-pass filtering and noise primarily degraded unique attributes (e.g., place of articulation cues), while high-pass filtering affected both unique and common attributes.

Conclusions:

• Perceptual similarities among consonants can be accurately modeled by additive combinations of discrete acoustic attributes.
• The weights of these attributes are modulated by the type of stimulus degradation.
• This approach provides a robust framework for understanding speech sound perception and its relationship to acoustic properties.