Tagging Environmental Sounds with YAMNet TFLite: An ESC-50 Evaluation over 2,000 Clips

Short summary

Evaluated Google's YAMNet TFLite audio tagging model on ESC-50's 2,000 environmental sound clips, achieving 60.45% fine-grained accuracy and 78.70% coarse accuracy at 0.029s per clip. The model excels on direct label matches (69% fine accuracy) but struggles with ambiguous categories like washing_machine and drinking_sipping. Complete reproducible code, confusion matrices, and per-category performance breakdown included.

• •YAMNet TFLite achieved 60.45% fine@1 accuracy on ESC-50 audio classification with 2,000 clips
• •Model processes 5-second clips in 0.029s—practical for lightweight audio tagging pipelines
• •Detailed confusion analysis shows common errors (hen/rooster, snoring/breathing) and category-specific accuracy breakdowns

Generated with AI, which can make mistakes.