Vowel decoding from single-trial speech-evoked electrophysiological responses: A feature-based machine learning...

Area of Science:

• Neuroscience
• Auditory Neuroscience
• Speech Processing

Background:

• Frequency-following responses (FFRs) reflect neural processing of auditory signals.
• Traditional FFR analysis requires averaging thousands of trials due to low single-trial signal-to-noise ratio.
• This limitation prevents the study of trial-by-trial neural dynamics.

Purpose of the Study:

• To develop a novel method for analyzing single-trial frequency-following responses (FFRs).
• To decode speech signal information from individual FFRs using machine learning.
• To overcome the limitations of traditional FFR analysis for assessing auditory function.

Main Methods:

• Collected scalp-recorded electrophysiological responses (FFRs) from participants listening to vowels.
• Projected FFRs onto a low-dimensional spectral feature space derived from speech signals.
• Trained a supervised machine learning classifier to discriminate vowel tokens on single-trial FFRs.

Main Results:

• Demonstrated reliable decoding of speech signals from single-trial FFRs.
• Successfully decomposed raw FFRs based on information-bearing spectral features.
• Showcased the potential of data-driven analysis for extracting interpretable features.

Conclusions:

• The developed method enables the analysis of single-trial FFRs, overcoming previous limitations.
• This data-driven approach allows for the investigation of trial-by-trial neural dynamics.
• Offers new possibilities for noninvasive assessment of human auditory function.