Auditory EEG encoding model that predicts continuous 5 kHz scalp EEG from binaural sound, from brainstem responses through cortical speech tracking.
No providers recorded yet. Browse all providers
A clinician measuring an auditory brainstem response, a neuroscientist fitting a temporal response function to an audiobook, and a hearing scientist probing binaural integration are all recording the same scalp EEG as it follows sound, yet each typically works with a model built for one response, one stage of the pathway, or one stimulus class. This model, from Thomas Stoll and Ross Maddox at the University of Michigan's Kresge Hearing Research Institute, takes the opposite approach. It is a single neural network that turns a binaural acoustic waveform into the continuous, high-sample-rate EEG a listener would produce. It is described in the September 2026 arXiv preprint "A Deep Neural Network for Predicting Continuous Human EEG Across the Auditory Pathway in Response to Sound", and the paper does not give it a name.
The authors frame it as a foundation model of human auditory electrophysiology. ABRs, speech TRFs and the binaural interaction component all become readouts of one simulated signal, analyzed as if it had been recorded. Most EEG foundation models, such as EEGPT, encode recorded EEG into features for decoding. This one runs the other way, predicting EEG from sound, in the spirit of sequence-to-signal models like Enformer.
Stereo audio passes through a causal 24-band-per-ear filterbank covering 100 Hz–16 kHz. It is log-compressed and pooled from 40 kHz down to 5 kHz. A nine-layer causal WaveNet encoder follows (kernel size 8, dilations 2^d, gated activations, residual connections), and a four-layer 1×1-convolution bottleneck projects the result to 16 latent components. The model is also conditioned on subject identity and on demographic and audiometric metadata. Training used about 250 hours of EEG from 92 young adults (ages 18–38), drawn from the lab's earlier studies of tonebursts, speech and music. The recordings were acquired at 10 kHz or more, downsampled to 5 kHz and high-pass filtered at 0.1 Hz. Plain MSE loss gave poor results, so the loss is computed on the short-time Fourier transform of the prediction error. This penalizes timing as well as amplitude, and it keeps the high-power low frequencies from dominating.
Evaluation used newly generated or held-out stimuli. Clicks, monaural sounds, periodic presentation and serial pABR presentation did not appear in the training data at all. Pearson correlations between the model and human grand averages ranged from 0.60 to 0.94 for pABR waveforms across 0.5–8 kHz, and from 0.79 to 0.93 for cortical speech TRFs. All but one condition fell within the distribution of individual subjects' correlations with the grand average. The exception was one cortical TRF, where the model sat above that distribution. The predicted binaural interaction component had a latency of 6.40 ms, inside the published range of 5.58–6.90 ms. Its amplitude of 0.44 µV was above the reported range but not significantly different from it.
The authors see the model as a platform for in silico auditory neuroscience: piloting paradigms, stimulus designs and analysis methods on predicted EEG before committing to costly recording sessions. Because the neural pathway is causal, it could also sit inside streaming hearing-technology pipelines, where predicted neural responses serve to evaluate or tune hearing-aid processing, or act as a front end for automated feature extraction.
The work shows that one network trained on pooled EEG from many experiments can recover textbook auditory phenomena from the brainstem to the cortex. Those phenomena include rate and frequency effects on wave V, serial-versus-parallel differences, the stronger subcortical response to low-F0 speech, and binaural interaction. Its scope is limited so far. It has been evaluated only through the default subject, and the subject-specific conditioning has not been analyzed or validated. The training cohort is young adults and is not stratified by hearing ability. The model overestimates responses to serial presentation and underestimates TRF amplitudes. The study is a preprint, and no code or trained weights have been released with it.
Much of this page is generated or calculated automatically. Flag anything that looks off and we will re-run it.