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

The human brain is a master of sound, capable of taking a chaotic stream of vibrations and turning it into a coherent story, a clear instruction, or a shared joke. For decades, scientists have known that this transformation happens in stages. First, the ear and the early parts of the brain catch the raw physical properties of the sound: its pitch, its volume, and its rhythm. Later, other parts of the brain take those sounds and stitch them together into words, sentences, and meaning. But exactly where the line is drawn between simply hearing a noise and understanding what it says has remained a mystery. In the middle of a flowing conversation, the physical sound and the meaning are so tightly woven together that it is nearly impossible to tell where one ends and the other begins. This gap matters because without knowing where comprehension starts, it is difficult to measure whether someone who cannot speak or move—perhaps due to a severe injury—is actually understanding the world around them.

A team of researchers has now mapped this boundary with unprecedented precision, using a new method that treats the brain like a complex machine and compares it directly to a sophisticated computer model. By listening to the brain's electrical signals while people listened to natural speech, and then matching those signals to the internal workings of an artificial intelligence trained on language, they discovered a sharp dividing line. When a listener understands the language, the brain's activity mirrors the deep, complex layers of the computer model that handle meaning. But when the listener hears the same sounds in a language they do not know, that deep connection vanishes instantly. The brain continues to track the basic sounds perfectly, but the part of the brain responsible for higher-level understanding simply stops following the model. This finding pinpoints the exact moment in the brain's processing chain where hearing turns into understanding, offering a clear, non-invasive way to see if speech is being comprehended, even when the listener cannot say so.

To find this boundary, the researchers turned to a powerful tool: a large language model designed to process audio, specifically a system called Qwen2.5-Omni. This model is built like a multi-layered factory. The first layers act as a microphone, breaking down the raw sound waves into basic components like frequency and loudness. As the data moves deeper into the model, through dozens of subsequent layers, it begins to recognize patterns, words, and eventually, the complex relationships between ideas. The researchers wanted to see if the human brain follows this same path. They recorded the brain activity of native English speakers listening to an English podcast and native Russian speakers listening to a Russian podcast. At the same time, they fed the exact same audio into the computer model and watched how the activity of its 145,000 individual artificial neurons changed over time.

The results were striking. In the native listeners, the brain's activity aligned with the computer model from the very first layer of sound processing all the way to the deepest layers where meaning is constructed. It was as if the brain and the computer were walking in perfect step, layer by layer. The researchers could see that as the computer model moved from simple sound detection to complex linguistic understanding, the human brain did the same, just with a slight delay. This confirmed that the brain processes natural speech in a hierarchical way, moving from the physical to the abstract, and that this process can be tracked by comparing it to a machine that does the same thing.

To find the exact point where understanding diverges from simple hearing, the researchers introduced a crucial twist. They took the Russian audio and played it for a group of native English speakers who did not understand Russian. The sound waves were physically identical to what the native Russian speakers heard, but for this group, the sounds were just noise. They could hear the rhythm and the volume, but they could not grasp the meaning. When the researchers compared the brain activity of these non-comprehending listeners to the computer model, a dramatic change occurred. The brain still matched the computer's early layers perfectly, tracking the basic sounds with the same precision as the native speakers. However, the moment the computer model moved into its deeper layers, where it began to construct meaning, the alignment with the human brain collapsed.

This collapse was abrupt and total. Beyond a specific layer in the computer model, the brain of the non-comprehending listener stopped following the computer's internal state. The few remaining connections that survived were not tracking words or ideas; they were tracking only the most basic physical features of the sound, such as loudness or the duration of a word. It was as if the brain, realizing it could not find meaning in the stream, retreated to the safety of the raw data, ignoring the complex layers where comprehension lives. The researchers identified this breaking point with high precision, locating it at a specific layer in the model's architecture. This suggests that the brain does not gradually lose its grip on meaning as a language becomes unfamiliar; rather, it maintains a perfect grip on the sounds until it hits a wall, at which point the higher-level tracking simply switches off.

The study also revealed that this boundary is not a fuzzy zone but a distinct threshold. By analyzing the data layer by layer, the researchers found that the transition from high tracking to no tracking happened so sharply that it could be pinpointed to a single layer in the model. This level of detail was only possible because they looked at individual units of the computer model one by one, rather than averaging them together. Previous studies that looked at the brain and computer models as a whole often missed this distinction because they blended the simple sound processing with the complex meaning processing. By isolating each step, the researchers showed that the brain's ability to track deep, abstract representations depends entirely on the listener's ability to understand the language.

This discovery provides a new way to look at how the brain builds meaning. It confirms that comprehension is not just a louder version of hearing; it is a distinct computational stage that requires the listener to have the necessary linguistic knowledge. When that knowledge is missing, the brain does not struggle to find meaning; it simply stops trying to track the complex patterns and focuses entirely on the physical properties of the sound. This finding has potential implications for understanding how the brain works in situations where communication is difficult, such as with patients who are unable to speak or respond. If the brain's deep-layer tracking collapses without comprehension, then measuring this tracking could serve as a reliable, non-invasive test to see if a person understands speech, even if they cannot tell us so. The researchers note that while their findings are based on specific languages and models, the method offers a clear path to testing whether this boundary exists universally across different languages and types of speech.

The work represents a significant step forward in bridging the gap between artificial intelligence and neuroscience. By using a computer model that processes language in a way that is transparent and interpretable, the researchers were able to read the brain's activity like a map. They did not just see that the brain was active; they saw exactly which part of the brain was active and what it was doing at every moment. This approach moves beyond asking whether the brain understands speech to asking how it builds that understanding, layer by layer. The result is a clear, detailed picture of the journey from sound to meaning, showing that the moment we stop understanding a language is the moment our brain stops following the path of meaning and returns to the safety of the sound itself.