Automatic Recognition of Speech Sounds Using Convolutional Neural Networks

Abstract

—Kurdish language has gained significant importance across various domains, yet it remains underrepresented in computational linguistics due to the scarcity of annotated datasets. This study develops and validates a robust Convolutional Neural Network (CNN)-based system for automatic recognition of Kurdish speech commands. A dedicated data-collection website was designed and deployed to gather over 4000 voice samples spanning 10 Kurdish commands, recorded by participants aged 7–64 across diverse environments, genders, and dialect backgrounds. The collected data (totaling 4527 samples) was partitioned into training (70%), validation (15%), and testing (15%) sets and preprocessed through low-pass filtering, amplitude normalization, and auditory spectrogram computation in MATLAB R2023b using the Audio Toolbox. The CNN architecture comprised three convolutional layers (3×3 filters, Rectified Linear Unit (ReLU) activation, 2×2 max-pooling with 0.25 dropout), followed by fully connected layers of 128 and 64 neurons, and a 10-class SoftMax output. The Adam optimizer (learning rate 0.001, batch size 32, 25 epochs) was used for training. The model achieved a validation accuracy of 98.17% and an independent test-set accuracy of 97.4%, with per-command recognition rates ranging from 91.5% to 100%. These results demonstrate the effectiveness of the proposed approach for Kurdish speech recognition and lay a foundation for future advancements in low-resource language processing.