Word-finding difficulties occur with aging and neurological disease, but they are hard to localize and measure. Standard tests are labor-intensive, yield little data, and are insensitive to mild impairments in people with high cognitive reserve. Automated analysis of natural speech promises a quantitative and scalable approach. This work sits alongside studies in which neural networks read medical histories or assess pronunciation, but here, for the first time, lexical and acoustic features are combined to evaluate memory.
Participants performed the Famous Faces task: they looked at photographs of well-known people and told everything they remembered. Recordings were transcribed with the WhisperX neural network, and participant utterances were separated from interviewer utterances using the pyannote algorithm. Then lexical features were extracted from the text: total word count and the proportion of rare words (Zipf frequency), while audio yielded utterance and pause durations. Pauses shorter than 250 milliseconds were counted as speech, and speech segments shorter than 150 milliseconds as noise.
The groups did not differ in lexical measures: total word count, proportion of rare words, and other metrics were the same. However, patients with temporal lobe epilepsy spoke in shorter phrases (on average 5.54 seconds versus 6.59 in healthy controls), made shorter pauses (0.61 versus 0.67 seconds), and fell silent more often (10.81 pauses per minute versus 9.29). These differences were statistically significant. The lexical index correlated with MoCA scores (r=0.35) and the Boston Naming Test (r=0.62) in patients, but not with memory or verbal fluency tests. Pause frequency negatively correlated with the accuracy of autobiographical memories (r=-0.31).
The work shows that lexical and temporal speech features reflect different cognitive functions: vocabulary is linked to naming, while speech rate is linked to the efficiency of recall. Unlike neighboring studies, where speech was analyzed to diagnose dementia or assess pronunciation, here a single pipeline is applied to spontaneous narration in epilepsy. This opens the way to sensitive digital markers that do not hit the ceiling of standard tests.
The method can be extended to other tasks that require oral retelling, such as describing pictures or retelling stories. Speech markers can be tracked over time to catch deterioration before it becomes obvious. It is promising to combine speech analysis with brain activity data to link pauses to epileptiform discharges.
The results may be useful in neurology for monitoring patients with epilepsy and early stages of Alzheimer's disease, as well as in telemedicine for remote assessment of cognitive function.
The next step is to test how speech measures change after temporal lobe surgery and whether they correlate with neuroimaging data.
The work addresses the unsolved problem of sensitive assessment of cognitive function in people with high reserve, where standard tests see no changes. It also links subjective complaints about word finding to objective speech parameters.
Word frequency was measured on the Zipf scale: a word with frequency 3 occurs about once per million words, and one with frequency 4 about once per ten thousand. Rare words turned out to be a marker of vocabulary.
Not published on arXiv. Here are its page, the PDF and the formalization: you can check it yourself, not only read our account.