Classifying Environmental Sounds with CLAP and ONNX Runtime: A 1-NN Evaluation on ESC-50

Short summary

Researchers evaluated CLAP audio embeddings on environmental sound classification using ONNX Runtime and the ESC-50 dataset. A 1-nearest-neighbor approach achieved 90% fine-grained accuracy and 100% top-5 accuracy on Apple Silicon CPU. The evaluation demonstrates that CLAP's embedding space preserves meaningful acoustic properties despite using only one reference recording per class.

• •CLAP audio encoder exported to ONNX achieved 90% fine-grained accuracy on ESC-50 environmental sound classification
• •1-nearest neighbor approach with single reference per class reached 100% top-5 accuracy
• •Reproducible pipeline runs 0.055 seconds per clip on Apple Silicon, with code and setup instructions available

Generated with AI, which can make mistakes.