Classify the audios. In this repo, I train a model on UrbanSound8K dataset, and achieve about 80% accuracy on test dataset.

There is a pre-trained model in urban_sound_train, trained epoch is 1000

• audio_train.py: Train audio model from scratch or restore from checkpoint.
• audio_params.py: Configuration for training a model.
• audio_inference_demo.py: Demo for test the trained model.
• ./audio/*: Dependencies of training, model and datasets.
• ./vggish/*: Dependencies of VGGish for feature extracting.

Conda are recommended, just need one line: conda env create -f conda.env.yml

1. Config parameters: audio_params.py.
2. Train the model: python audio_train.py. (It will create tfrecords automaticly if not exists)
3. Check the training process from tensorboard: tensorboard --logdir=./data/tensorboard
4. Test the model: python audio_inference_demo.py.

• TensorFlow: VGGish
• Google AudioSet
  • VGGish model checkpoint
  • Embedding PCA parameters
• pyAudioAnalysis(Ref.)

• urban sound dataset

• AudioSet: An ontology and human-labelled dataset for audio events
• CNN Architectures for Large-Scale Audio Classification