A new publication from the BabyRhythm project funded by the European Research Council (ERC)
Decoding speech information from EEG data with 4-, 7- and 11-month-old infants: Using convolutional neural network, mutual information-based and backward linear models
Here, we applied three computational methods to decode speech information from neural activity in the infant brain. Particularly, we developed a backward linear model, a convolutional neural network (CNN) model, and a mutual information-based (MI) model, which has not been explored in infant literature before. EEG data were collected from fifty infants as they listened passively to natural speech (sung or chanted nursery rhymes) presented through video with a female singer. Each model estimated speech information for these nursery rhymes in two distinct low-frequency bands, namely delta and theta, thought to convey different types of linguistic information. Remarkably, all three models exhibited significant performance levels for delta-band neural activity starting from 4 months of age, with two out of three models also displaying significant performance for theta-band activity. Furthermore, all models demonstrated higher accuracy in the delta-band neural responses. Comparing the models, both the backward linear and CNN models exhibited more convergence than the MI model. Our findings suggest that computational choices play a crucial role in studying the development of the neural representation of speech. Understanding the strengths and weaknesses of each modeling approach is essential for advancing our comprehension of how the human brain constructs a language system.
Read the article here.