The research on the Parkinson′s disease (PD) speech recognition algorithm is important for timely diagnosis and treatment. However, the existing public PD speech datasets are characterized by small sample sizes, which is one of the main challenges faced by existing PD speech recognition methods. To address this issue, a novel dual-side two-stage means clustering envelope and convolution sparse transfer learning model is proposed. First, for the dataset side, multiple groups ofconvolution kernels are trained, which is based on the source domain dataset. Then, the optimal convolution kernels are filtered by the encoded intermediate dataset. Finally, the target domain dataset is encoded by the optimalkernels. In regard to deep instances clustering envelope,an iterative mean clustering algorithm is designed to construct the deep instance space. Secondly, various classifiers are developed after sample / feature parallel selection. Finally, the classification results of different instance layers are fused. In the experiment,the representative PD speech datasets are selected for verification. Experimental results show that the main innovative parts of the proposed algorithm are effective. Compared with more than ten classical algorithms,the obvious improvements in terms of classification accuracy are achieved 97. 8% . In addition, the proposed algorithmhas potential in clinicalapplications for acceptable time complexity.