Researchers trained a single deep neural network to turn stereo sound directly into high‑resolution electroencephalography (EEG) signals. EEG is a noninvasive measure of brain electrical activity. The model was trained on about 250 hours of EEG from 92 people listening to tonebursts, speech, and music. It produced continuous predictions that matched known patterns from both early brainstem and later cortical responses to sound.

At a high level the network is a causal encoder–decoder: it sees the audio waveform and produces an EEG waveform aligned in time. Audio was preprocessed with a 24‑band filter per ear (100 Hz–16 kHz) and resampled to 40 kHz. The encoder used a WaveNet‑style stack of dilated convolutions and the model projected its internal representation into 16 latent neural components that were mapped to EEG channels. The designers also included a parallel pathway to model fast, stimulus‑locked recording artifacts during training, but that artifact pathway was turned off for the evaluations reported here. Training used segments about a minute long, a loss that emphasizes timing and high frequencies (a log short‑time Fourier transform loss), and a 10% held‑out test set.

The team tested whether the model reproduced three standard kinds of responses. First, auditory brainstem responses (ABRs) are fast waves that reflect early, subcortical stages of hearing. Second, temporal response functions (TRFs) capture how the brain tracks continuous speech at subcortical and cortical timescales. Third, the binaural interaction component (BIC) reflects how sounds to both ears combine in the brainstem. Some evaluation stimuli were new or of types not seen during training (for example, clicks and serial presentation were not part of training). Despite that, predicted ABRs and TRFs showed established waveform shapes and stimulus‑dependent changes. When compared to human data, many model predictions had correlation values that fell within the range seen across individual subjects. For example, model correlations for cortical TRFs reached the 93–100th percentile in several test conditions, while pABR (parallel ABR) correlations ranged from about the 21st percentile at 500 Hz up to the 97th percentile at 2 kHz for different frequencies.