YAMNet TFLite Tested on 2,000 ESC-50 Audio Clips to Evaluate Sound Tagging Accuracy

A developer evaluated Google's YAMNet TFLite audio tagging model against the ESC-50 environmental sound dataset, which contains 2,000 five-second recordings spanning 50 sound categories. YAMNet natively predicts 521 AudioSet sound classes, so the researcher manually mapped those labels to ESC-50's 50 categories and five coarse groupings such as animals, human non-speech, and urban sounds. Each audio clip was converted to 16 kHz mono PCM using FFmpeg before being passed to the TFLite model, with frame-level scores averaged into a single clip-level prediction. The experiment measured fine-grained accuracy at ranks 1 and 3, as well as coarse category accuracy, to determine whether broader groupings yield more reliable results than specific labels. The full code and reproducible environment have been published openly, requiring only standard tools and a few hundred megabytes of storage to replicate.