Attention-Based Deep Neural Networks for Speech and Music Discrimination

Abstract

Conventional techniques for speech and music discrimination often rely on handcrafted features and shallow models, leading to limitations in capturing complex temporal dependencies and handling noisy or degraded audio signals. To address these challenges, this paper proposes a novel approach that employs attention-based deep neural networks (DNNs) for improved speech and music classification. The proposed method utilizes a combination of LSTM and attention mechanisms to effectively model the temporal dynamics of audio signals and focus on the most discriminative parts. By incorporating attention, the model can learn to weigh different segments of the audio signal based on their relevance for classification. Experiments are conducted on two publicly available standard speech/music datasets: SS and GTZAN, to evaluate the performance of the proposed method. Initially, we have found that a combination of Mel-frequency cepstral coefficients (MFCCs), spectral roll-off deviation, spectral flux deviation, and zero-crossing rate (ZCR) deviation provides the best performance for speech and music discrimination using a Gaussian mixture model (GMM) classifier, achieving an accuracy of 92.68%. By incorporating attention mechanisms on the hybrid features (simple, LSTM and Hierarchical), the model can focus on the most discriminative parts of the audio signal, leading to improved classification accuracy (test accuracy of 99.58% for SS and 97.58% for GTZAN) and generalization over the previous work. The results demonstrate significant improvements over conventional techniques in terms of accuracy, precision, recall, and F1-score.