A Novel Approach to Mimic Linguistic Features of Transcripts of Atypical Speech

It is often infeasible and costly to obtain speech data from individuals who exhibit atypical linguistic patterns. Although there have been considerable advances in the field of text generation, concerns remain about the potential leakage of identifiable information when generating data from patient speech transcripts. In response to this challenge, we present a novel approach to mimic the linguistic patterns found in transcripts of atypical speech. Our method estimates the probability distributions of linguistic features, devoid of sensitive information and unique to atypical typical speech. With these probability distributions, we create a mimicked atypical transcript by modifying selected content of a transcript of typical speech. To evaluate this mimicry strategy, three testing scenarios were considered: baseline test, where a classifier was trained and tested on real data; cross-data test, where a classifier was trained on real data and tested on mimicked data; and synthetic-boost test, where a classifier was trained on mimicked data and tested on real data. The real data consisted of non-parallel annotated spoken transcripts of cognitively intact individuals (typical speech) and those with probable Alzheimer’s disease (AD; atypical speech). In the synthetic-boost test, a Bidirectional Encoder Representations from Transformers (BERT) model classified the real data with an accuracy of 80.25%, exceeding the 75.86% accuracy achieved in the baseline test. This finding indicates that the mimicked data significantly enhanced the model’s ability to detect atypical Alzheimer’s speech. Accuracy significantly decreased when excluding shortening (63%, F1 score = 42%) and communicator (61%, F1 score= 41%) features from the mimicry process, underscoring their essential role in AD speech detection. We additionally compared our mimicry strategy with a GPT-4-o prompt-engineering baseline. The LLM-generated data yielded 73.9 % classification accuracy, 6.3 percentage points lower than mimicry, while exhibiting occasional hallucinated names/locations.