Multi-Source Transfer Learning Method Based on BP Fine-Tuning for Spontaneous Speech Recognition in Parkinson's Disease

Speech-based diagnosis of Parkinson’s disease (PD) has emerged as a prominent research focus. However, existing speech-based PD diagnostic methods suffer from insufficient utilization of speech data and the limitation that speech acquisition requires subjects to read a fixed text. To address these limitations, this paper proposes a Multi-Source Transfer Learning Method Based on BP Fine-Tuning (MTL-BPFT) for continuous speech recognition in PD. Specifically, the method pre-trains a BP neural network model using multi-source fixed-text speech data, freezes lower-layer parameters to preserve domain-invariant feature extraction capabilities, and introduces dynamic pruning mechanisms alongside elastic regularization constraints to fine-tune trainable layer parameters with limited labeled target-domain data, thereby enhancing recognition accuracy in data-scarce scenarios. Experimental results on two tasks demonstrate that the proposed algorithm achieves accuracies of 95.22% and 80.31%, respectively, significantly outperforming state-of-the-art methods and baseline approaches. These results illustrate that the proposed method provides highly generalizable technical support for spontaneous speech recognition in Parkinson’s disease.