Machine Learning Approaches to Analyze Speech-Evoked Neurophysiological Responses

Area of Science:

• Neuroscience
• Computational Linguistics
• Machine Learning

Background:

• Speech-evoked neurophysiological responses are crucial for understanding speech and language processing.
• Traditional analysis methods have limitations in capturing the complexity of neural responses to natural speech.
• Machine learning (ML) offers novel approaches to analyze these complex neural signals.

Purpose of the Study:

• To highlight the practical application of ML-based approaches for analyzing speech-evoked neurophysiological responses.
• To introduce decoding and encoding ML models for speech-related neural data.
• To demonstrate the utility of ML in understanding speech and language processing.

Main Methods:

• Reviewed two categories of ML models: decoding (classifying neural responses) and encoding (predicting neural responses from stimuli).
• Focused on decoding classification for phonological categories and encoding temporal response functions.
• Applied these methods to electroencephalography (EEG) data from Mandarin lexical tones and English phonemes in continuous speech.

Main Results:

• Demonstrated ML classification of EEG responses to Mandarin lexical tones and English phonemes.
• Showcased the ability of temporal response functions to predict EEG responses from acoustic features in natural speech.
• Observed statistically significant neural effects at the individual level across all examples.

Conclusions:

• ML-based approaches significantly complement traditional methods for analyzing neurophysiological responses to speech.
• These ML techniques provide a deeper understanding of natural speech and language processing.
• The application of ML is valuable in both typical and clinical populations using ecologically valid paradigms.