Intelligent Analysis of Speech Data from Speech Rehabilitation Devices: Grading and Recovery Prediction Algorithms for Articulation Dysfunction

Speech disorders impair patients' social and quality of life; speech rehabilitation devices have accumulated massive speech data, but traditional manual analysis is inefficient and subjective, while existing methods focus on onedimensional features with insufficient accuracy and no reliable rehabilitation progress prediction. This paper constructs multimodal speech features (acoustics, rhythm, pronunciation movements), proposes an attention-based CNN-BiLSTM classification model for damage degree, and develops an LSTMGRU prediction model combining patient observation data. Experiments show cerebral palsy patients' actual scores decreased from$\sim \mathbf{4.2}$to$\sim \mathbf{2}$, with LSTM-GRU predictions closely matching (3.5 actual vs 3.55 predicted in week 4), while single LSTM had obvious deviations, providing a reliable basis for clinical rehabilitation plans.