Summary

Copiste may become (one day) a radio ripper featuring metadata extraction and audio recognition using learning algorithms. Currently, it allows music / speech detection on audio files as well as live radio streams. The core software relies on a large number of XML files and is therefore easy to hack on the fly.

The computing part of the project relies on libvlc, which provides a large set of decoders and a powerful streaming pipeline. The other dependencies are Qt and boost. The project can be built either with SCons or with CodeBlocks.

Tools

nnat : Neural Network Analysis Tool

This tool takes a neural network described in an XML format (see the networks directory for examples) and trains it to fit a given corpus, described in another XML format (see the corpus directory for examples). It can also evaluate how good a network fits a corpus, and provides a small graphical interface that can be used to visualize 2D networks.

featdraw : Feature Drawer

This tool takes an audio file and draws the variations of some features in this file. The features are described by a pipeline (XML file, see the pipeline directory), allowing the user to change on the fly how features are computed.

mkcorpus : Make Corpus

This tool uses a pipeline and a set of audio files to create a corpus (as in nnat) where the features are written. It uses another XML file, which defines the different audio classes.

specdraw : Draw Spectrum

This tool just takes as input an audio file and draws the variations of the frequency spectrum during the time.

classify : Classify an audio file

This program takes an audio file, a pipeline and a network, and says what class the file belongs to.

live : Live stream classification

This program reads a live stream (e.g. a radio), and shows the variations of the features and the classification of the stream.