BDConvSGNet: A Lightweight Neural Network for Isolated Bengali Digit Sound Classification

Speech recognition systems in low-resourced languages are gaining popularity, as command-based systems in such languages enable users to interact with technology without requiring extensive technical knowledge. This paper proposes a lightweight neural network architecture, BDConvSGNet, to classify Bengali isolated digit sounds. The proposed model combines a Convolutional Neural Network (CNN), a Bidirectional Gated Recurrent Unit (Bi-GRU), and a Squeeze-and-Excite (SE) mechanism. The model uses Mel-Frequency Cepstral Coefficients (MFCCs) as features from the isolated digit sounds to achieve efficient classification. With a total of only 97,408 parameters, BDConvSGNet is compact and consumes less memory, making it suitable for resource-constrained environments. The model was evaluated on three different datasets, consisting of 984, 2,252, and 13,554 samples, respectively. It achieved high classification accuracy with mean scores of 94.00%, 98.76%, and 99.20% in 10- fold cross-validation on these datasets. The results demonstrate that BDConvSGNet is a robust, efficient, and reliable solution for Bengali digit sound classification.