Dataset size considerations for robust acoustic and phonetic speech encoding models in EEG

Area of Science:

• Neuroscience
• Auditory Neuroscience
• Cognitive Science

Background:

• Electroencephalography (EEG) studies of auditory and speech processing often use lengthy, tedious experimental paradigms to ensure data robustness.
• Naturalistic stimuli and system identification techniques like multivariate temporal receptive field (mTRF) analyses are increasingly used to study brain responses to speech features.
• The amount of neural data required for stable and generalizable mTRF model fitting remains unclear, impacting experimental design.

Purpose of the Study:

• To determine the minimum amount of electroencephalography (EEG) data needed for stable multivariate temporal receptive field (mTRF) model fitting in naturalistic speech processing experiments.
• To investigate how different types of naturalistic stimuli (sentences, movies, audiobooks) and feature representations influence data requirements for robust mTRF analyses.
• To provide practical guidelines for researchers to optimize data collection, minimizing participant fatigue while maintaining signal quality in auditory and speech processing studies.

Main Methods:

• Utilized previously collected EEG data from sentence and movie stimuli, alongside an open-source audiobook dataset.
• Applied multivariate temporal receptive field (mTRF) analyses to assess the stability of EEG receptive field structures across varying training dataset sizes.
• Investigated the impact of different naturalistic stimuli on the data duration required for reliable mTRF model fitting.

Main Results:

• EEG receptive field structure stabilization was observed after approximately 200 s of TIMIT sentence data, 600 s of movie trailer data, and 460 s of audiobook data for training.
• These findings suggest specific minimum data requirements for fitting mTRFs from diverse naturalistic listening paradigms.
• The type of stimulus significantly influences the amount of data necessary for robust model parameter estimation.

Conclusions:

• Established data collection benchmarks for fitting mTRFs in naturalistic speech processing research using EEG.
• Recommendations are provided to aid researchers in designing efficient experiments, particularly for populations with limited tolerance for long sessions (e.g., children, patients).
• This work facilitates future studies on auditory and speech processing in both healthy and clinical populations, balancing signal quality with participant comfort.