Statistical, Multi-scale and Attention-based Layer Pooling of Wav2Vec-2 Speech Embeddings for Parkinson’s Disease Detection

Self-supervised pre-trained speech models such as wav2vec 2.0 provide rich frame-level embeddings that are increasingly used for clinical voice screening, including Parkinson’s disease (PD). Optimally aggregating the frame-level embeddings is a significant task, with the underexplored question of how to best aggregate the frame-level embeddings into fixed-length utterance descriptors for the downstream task of binary classification (healthy controls (HC) vs. PD). To address this, our study compared three wav2vec 2.0 variants, the base model and two fine-tuned variants (one adapted for dysarthric corpora), across multiple layer depths, ten statistical aggregation functions (e.g., mean, median, quantiles), and two proposed advanced schemes (attention-based and multi-scale average pooling). We used the MDVR-KCL as a read-speech corpus (16 PD, 21 HC). Contrary to the expectation that sophisticated pooling would help, statistical aggregations such as mean pooling consistently provided better performance and robustness. Early representations (pre-Transformer and 1st Transformer block) were often the most informative, and mean aggregation produced relatively high, low-variance scores across models and depths. Attention and multi-scale pooling did not yield consistent gains. Moreover, wav2vec-based embeddings outperformed traditional acoustic baselines. Applying the supervised feature selection (ANOVA F-value) further improved performance, i.e. conservative selection (12 features) achieved a mean balanced accuracy of 0.87 and precision of 0.92, with top configurations exceeding a 0.93 balanced accuracy and had a precision of 1.0. The findings empirically support the continued use of mean pooling as a viable strategy for temporal aggregation of latent features for PD wav2vec-based detection.