Low resource speech intent classification using MFCC features.

Abstract

Speech-based user interfaces have revolutionized digital interactions, yet developing them for low-resource languages remains a challenge due to limited labeled speech data. This research proposes a Convolutional Neural Network (CNN)-based approach utilizing Mel-Frequency Cepstral Coefficients (MFCC) along with delta and delta- delta features for effective speech intent classification in Sinhala and Tamil. The methodology incorporates audio preprocessing, MFCC feature extraction, and data augmentation techniques such as noise addition, pitch shifting, and time stretching. A stratified cross-validation framework is used to ensure fair and consistent evaluation. The proposed model achieves 96.92% accuracy on the Sinhala dataset (7,624 samples) and 93.81% on the Tamil dataset (400 samples, ~0.5 hours of speech), representing a substantial improvement over prior methods. These results demonstrate the effectiveness of the CNN-based approach in capturing meaningful acoustic patterns for intent recognition in low-resource settings. The study offers a scalable, efficient solution for speech intent classification and contributes to the advancement of inclusive voice-enabled technologies.