This repository will contain notebooks and code snippets for working with Google AudioSet, a large-scale dataset of manually annotated audio events.

tutorial_reading_audioset_tfrecords.ipynb explains how to open a TFRecord that contains AudioSet feature vectors (embeddings), video IDs and labels.

AudioSet consists of an expanding ontology of 632 audio event classes and a collection of 2,084,320 human-labeled 10-second sound clips drawn from YouTube videos. The ontology is specified as a hierarchical graph of event categories, covering a wide range of human and animal sounds, musical instruments and genres, and common everyday environmental sounds.

Instead of raw audio, every example is made available as a sequence of 128-dimension embeddings (one embedding for every second of audio). The embeddings were obtained after passing the log-scaled mel-spectrograms of the audio waveform through a deep convolutional neural network trained on millions of audio clips (VGGish).

Creators/users of the dataset can be found at audioset-users.