The early maximum likelihood estimation model of audiovisual integration in speech perception

Area of Science:

• Cognitive Science
• Auditory Neuroscience
• Speech Processing

Background:

• Speech perception relies on integrating auditory and visual cues.
• The McGurk-MacDonald illusion highlights audiovisual integration.
• Existing models like the fuzzy logical model of perception (FLMP) have limitations in parsimony and interpretability.

Purpose of the Study:

• Introduce and evaluate the early maximum likelihood estimation (MLE) model for audiovisual speech perception.
• Compare the performance of early MLE with existing models using cross-validation.
• Address the lack of comprehensive computational accounts for perceptual audiovisual integration.

Main Methods:

• Developed an early maximum likelihood estimation (MLE) model with three variations.
• Applied cross-validation techniques to assess model goodness-of-fit and flexibility.
• Tested models on a published dataset previously used for the FLMP.

Main Results:

• Cross-validation favored the early MLE model over more complex models.
• Conventional error measures favored more complex models, indicating potential over-fitting.
• The early MLE model demonstrated greater parsimony by imposing constraints reflecting experimental designs.

Conclusions:

• The early MLE model provides a more constrained and parsimonious account of audiovisual speech integration.
• Cross-validation is crucial for evaluating model performance, detecting over-fitting in complex models like FLMP.
• This study offers a new computational framework for understanding how visual speech information enhances auditory perception.