Decoding speech information from EEG data with 4-, 7- and 11-month-old infants: Using convolutional neural network,
Mahmoud Keshavarzi1, Áine Ní Choisdealbha1, Adam Attaheri1
1Centre for Neuroscience in Education, Department of Psychology, University of Cambridge, Downing Street, Cambridge CB2 3EB, UK.
Journal of Neuroscience Methods
|December 21, 2023
Summary
Computational models can decode infant neural activity related to speech. The choice of model impacts developmental findings, highlighting the need to understand model strengths for language acquisition research.
Area of Science:
- Neuroscience
- Computational Linguistics
- Developmental Psychology
Background:
- Computational models for decoding neural activity into speech are common in adults but underutilized in infant research.
- Existing models like convolutional neural networks (CNNs), backward linear models, and mutual information (MI) models offer potential for analyzing infant neural data.
Purpose of the Study:
- To apply and compare three computational models (CNN, MI, backward linear) for decoding speech envelope information from infant neural activity.
- To investigate the performance of these models in different frequency bands (delta and theta) and across different infant ages (4, 7, and 11 months).
Main Methods:
- EEG recordings were collected from 50 infants (4, 7, 11 months) passively listening to nursery rhymes.
- Backward linear, CNN, and MI models were used to decode speech envelope information from neural activity.
- Model performance was evaluated in delta and theta frequency bands.
Main Results:
- All three models successfully decoded delta-band neural activity from 4 months of age.
- Two models also showed significant performance for theta-band activity.
- Models performed better with delta-band responses, and no age-related developmental effects were observed.
Conclusions:
- The selection of a decoding algorithm significantly influences conclusions drawn about infant speech processing development.
- Understanding the strengths and limitations of different computational models is crucial for advancing research on how the brain develops language capabilities.
