A two-stage transfer-learning pipeline that predicts a patient's Mini-Mental State Examination (MMSE) score from spontaneous-speech audio: OpenAI's Whisper transcribes each recording, RoBERTa encodes the transcript into a [CLS] embedding, and a regression head maps it to a normalized MMSE score. Chosen for Whisper's robustness to accents and pathological speech and RoBERTa's sensitivity to subtle linguistic markers of cognitive decline, it reaches RMSE 2.6911 on the PROCESS Grand Challenge 2025 PROCESS V1 dataset, about 10% below the challenge baseline.
A short primary-research paper predicting dementia severity from spontaneous-speech audio via a two-stage transfer-learning pipeline. Each recording (from the PROCESS Grand Challenge 2025 PROCESS V1 …