Audio Chime

Paper Implementation for:

A convolutional neural network approach for acoustic scene classification [Paper]

We have worked on the CHIME Audio Dataset for Audio tagging. We train on 48KHz and test on 16KHz Audio.

About the Dataset

The annotations are based on a set of 7 label classes. For each chunk, multi-label annotations were first obtained for each of 3 annotators. There are 1946 such 'strong agreement' chunks is the development dataset, and 816 such 'strong agreement' chunks in the evaluation dataset.

Cloning the repo

Go ahead and clone this repository using

$ git clone https://github.com/DeepLearn-lab/audio_CHIME.git

Quick Run

If you are looking for a quick running version go inside single_file folder and run

$ python mainfile.py

Detailed Task

The process involves three steps:

1. Feature Extraction
2. Training on Development Dataset
3. Testing on Evaluation Dataset

Feature Extraction

We are going to extract mel frequencies on raw audio waveforms. Go ahead and uncomment
feature_extraction function which would extract these features and save it in the .f pickle.

Training

We train our model on these extracted featuers. We use a convolution neural network for training and testing purpose. Alteration in model can be done in model.py file. All hyper-parameters can be set in util.py. Once you have made all the required changes or want to run on the pre-set ones, run

$ python mainfile.py 

This will run the model which we test and use EER for rating our model.