Dysarthria is one of the early symptoms of stroke, and the traditional diagnosis method relies on the doctors listening to the patient's pronunciation, which suffers from low accuracy. We constructed a multi-feature fusion feature engineering and combined it with machine learning methods for early recognition of stroke. According to the specificity of stroke pathological speech, different features will imply different information, which will lead to different recognition results. Therefore, we extracted features with different classes of information of the stroke pathological speech and combined them with different machine learning models for comparison experiments. Finally, the random forest model with a combination of prosodic features and spectral features achieved an accuracy of 94.33%, a precision of 91.89%, a recall of 99.14%, and an f1-score of 95.17%. The model has been deployed into an application running on smart devices. Stroke patients can use this for early recognition at home to save valuable time for early warning and intervention treatment.