Hierarchical multi-branch system for heterogeneous audio classification

Heterogeneous audio classification faces challenges such as the variability of sound sources and the need to maintain taxonomic consistency across hierarchical levels. Modern systems, such as those based on CLAP representations, integrate specific branches of acoustic features and post-processing with KNN to achieve hierarchical F1 scores above 80%. This multi-branch approach combines complementary models —for example, using log-STFT and diverse classification heads— to improve accuracy in seconds and overall consistency. At Q2BSTUDIO we apply similar techniques of artificial intelligence for businesses, adapting hierarchical architectures to real-world audio analysis scenarios, from surveillance to multimedia indexing. Our custom software services allow us to implement training pipelines augmented with quality data, such as the filtered BSD35k subsets, and deploy them in cloud environments (AWS, Azure) to scale inferences. Additionally, we integrate AI agents that refine predictions in real time and business intelligence services with Power BI to visualize classification metrics. Cybersecurity is also key when processing sensitive acoustic data, and we offer perimeter protection solutions. Thus, the multi-branch approach not only optimizes sound taxonomy but also becomes an enabler for custom applications in sectors such as industry or smart cities.