Enhancing Schizophrenia Diagnosis with Multi-Spectral Speech Analysis and Attention Mechanisms

Schizophrenia is a complex and chronic psychiatric disorder that significantly affects cognition, perception, and communication. Among its many symptoms, speech impairments serve as a critical biomarker for early diagnosis, making speech-based assessments an effective tool for automated detection. This study introduces WNMSA-Net (Wideband-Narrowband-Mel Spectrogram Axial-Attention Network), an Axial-Attention-Based Deep Learning Model designed for high-accuracy schizophrenia classification and early detection. The model leverages Wideband Spectrograms (WBS), Narrowband Spectrograms (NBS), and Mel Spectrograms (MS) to extract essential speech features while utilizing Axial Attention Mechanisms to enhance time-frequency dependency modeling. Additionally, a Gated Multi-Channel Fusion (GMC-Fusion) module is incorporated to dynamically integrate spectral representations, ensuring optimal feature selection for classification. Experimental results demonstrate that WNMSA-Net effectively captures schizophrenia-related speech impairments, including reduced fluency, monotonic pitch variations, and abnormal intensity shifts. The model achieves an accuracy of 98.85%, significantly improving classification performance while enhancing early detection rates to 92.40%. Computational efficiency is also optimized, reducing training time by 21.6% and inference time by 17.7%, ensuring feasibility for real-time clinical applications. By combining multi-spectral speech analysis, attention-based feature extraction, and adaptive fusion mechanisms, WNMSA-Net establishes a new benchmark for AI-driven schizophrenia detection. This research highlights the potential of speech-driven artificial intelligence in psychiatric diagnostics, paving the way for automated, objective, and scalable mental health assessment tools that can support clinical decision-making and improve early intervention strategies.