Deep-layer cortical tracking abruptly collapses in the absence of language comprehension

The brain processes speech through several stages, converting acoustic input into linguistic comprehension. However, the precise point at which comprehension diverges from general acoustic processing has been hard to pinpoint, as both are closely linked during natural speech.

In this study, researchers aligned the activity of approximately 145,000 artificial neurons from an audio large language model with high-resolution magnetoencephalography. By employing single-neuron interpretability, they were able to track, layer by layer, the computations the cortex performed during natural listening. Comparing native listeners with individuals hearing an unfamiliar language, while keeping the acoustic input physically identical, revealed that comprehension sustained brain-model alignment throughout the model's deep layers.

In the absence of comprehension, alignment persisted in the acoustic encoder, but deep-layer tracking abruptly collapsed. The sparse alignments that remained were found to correspond only to physical acoustic features. This finding highlights the specific stage where comprehension separates from perception, providing an interpretable, non-invasive method to determine whether speech is understood.