whatsound - a Python ML toolkit for audio classification

whatsound is a toolkit for training, and testing audio classification using a neural network.

• Music
• Speech
• Ambient/Noise
• Silence

This toolkit uses Essentia for audio feature extraction, and PyBrain for the use of a backpropagation neural network for the training and testing of classification.

How it works

WS_classify.py

This is the toolkit for classification, which exposes the main classification functionality.

Source modules

The project is split into modules fit for different purposes.

/core

These modules are needed for audio training and classification.

WS_extractor.py

Extracts audio features from a stream. The Essentia library is used for audio analysis. The features which are used for extraction are:

• MFCC
• Zero crossing rate
• Key strength
• Spectral Flux
• Pitch strength
• LPC

WS_utils.py

Utility functions

WS_global_data.py

These are global parameters - settings for the neural net, training parameters, audio settings and classifier types.

WS_network.py

This module allows training and testing of a data set, with optional the following parameters:

• weights : the path to a PyBrain weights XML file
• dataset: the path to a directory containing audio samples split by class
• split: the ratio with which to split the data set between training/testing