Hardware-Efficient Neural Network for Voice Disorder Classification from Multi-Source Datasets

This paper presents a hardware-efficient neural network-based approach for classification of healthy voices and pathological conditions including functional dysphonia, hyperkinetic dysphonia, hypokinetic dysphonia, and reflux laryngitis. Using two widely-used datasets—the Saarbruecken Voice Database (SVD) database and the Voiced database—we extracted clinically relevant acoustic features such as jitter, shimmer, harmonics-to-noise ratio (HNR), fundamental frequency (f₀), and formant frequencies to train a compact convolutional neural network (CNN) optimized for deployment on Field Programmable Gate Array (FPGA) platforms. The proposed model achieved an overall classification accuracy of 91.4%, with consistently high sensitivity and specificity across all five categories. Post-training quantization was applied to convert the model into an 8-bit fixed-point representation, significantly reducing memory usage and computational overhead. The CNN was synthesized using Vivado High-Level Synthesis (HLS) and implemented on a Zynq-7000 FPGA, where it achieved real-time inference with an average latency of 80 ms per sample. These results confirm the feasibility of integrating deep learning-based voice diagnostics into portable, low-power clinical devices. The system's robustness, efficiency, and accuracy make it highly suitable for early detection and continuous monitoring of voice disorders in both clinical and telemedicine environments.