ResBiA-FusionNET: A Robust Deep Learning Framework with Harmonic-Contrast Mel Spectrogram for Audio Sound Classification

Revolutionizing audio classification hinges on sophisticated Deep Learning (DL) models that adeptly identify subtle spectral variations and temporal dynamics. These capabilities are essential for harnessing the full potential of complex sound signals across various applications. However, conventional approaches such as standard Convolutional Neural Networks (CNNs) and Recurrent Neural Networks (RNNs) struggle with short-term transients, long-term dependencies, and the relative importance of sound components, often leading to incomplete feature representation and reduced classification accuracy. To overcome these limitations, this study introduces the ResBiA-FusionNET framework, an integration approach designed to enhance classification robustness and computational efficiency across diverse acoustic environments. The framework begins with a Dynamic Audio Pre-processing Framework (DAPF), to decompose audio into multiple frequency bands, capturing fine-grained acoustic patterns while suppressing background noise. Feature extraction is further enriched using the Harmonic-Contrast Mel Spectrogram (HCMS), a hybrid technique to capture perceptually significant, dynamic, and harmonic properties of the audio signal. This multi-dimensional feature fusion enables more discriminative and robust sound representations, especially for acoustically similar classes. For classification, the study introduces ResBiAttentionNet, an advanced architecture that integrates Residual Networks with Bidirectional Attention Mechanisms, allowing the model to focus on crucial temporal segments in both directions while enabling efficient DL through residual connections. The framework is rigorously trained and validated on two comprehensive benchmark datasets, ESC-50 and UrbanSound8K. The proposed model attained an accuracy of 98.67%, demonstrating improvements over state-of-the-art. Furthermore, the framework demonstrates strong generalizability across diverse audio scenarios, highlighting its potential for real-world deployments. This robust architecture sets a new benchmark for audio signal analysis, offers robust precision and adaptability for real-world audio recognition applications.