I am looking for a way to take a .wav file and extract features from it that could be used for further classification (the features could be MFCC statistics, an autoencoder representation, etc.). I know it is likely to be task dependent, but I am looking for some standard set of features that make sense. I am trying to categorize an audio signal (mostly music) into a set of categories such as "quiet music", "thrilling music", etc.

I was wondering if there is existing software (preferably for Linux) that takes such audio signal (such as in a .wav format) and outputs a feature fingerprint.

  • $\begingroup$ Welcome to SE.SP! This question is way too broad: we'd need to know what features you're after because it can be very application-specific. What are you trying to achieve? $\endgroup$ – Peter K. May 19 '17 at 13:32
  • $\begingroup$ I made an edit to explain that. $\endgroup$ – signaler May 19 '17 at 14:18
  • $\begingroup$ OK! Let's see what that brings. I've re-opened. $\endgroup$ – Peter K. May 19 '17 at 15:09

The openSMILE audio feature extraction toolkit may be able to provide the functionality you desire, where the input is a .wav file and the output extracted audio features. See: http://audeering.com/technology/opensmile/

openSMILE provides a command line executable that is coupled with a configuration file that defines the features to be extracted. The executable has binaries for Linux and Windows (32 bit), and I've also built the executable from source for use on macOS. It provides its own objects that calculate features such as MFCCs, energy, pitch, but also "chroma" musical features (which may help with music classification). It comes with many prebuilt feature configurations that have been used for audio classification tasks, such as the Audio Visual Emotion Challenge (http://sspnet.eu/avec2017/).

Below is an example of how openSMILE can be used to extract features from a single .wav file and a chosen configuration. The output.arff is a format used by the Weka machine learning library.

SMILExtract -C config/emobase.conf -I input.wav -O output.arff

Custom configurations can also be written using the openSMILE configuration language and also the extensive PDF documentation included with openSMILE's download.

This has the desired functionality as described in the original question:

looking for a way to take a .wav file and extract features from it that could be used for further classification


A relatively easy way to do that would be to write MATLAB/GNU Octave/Python scripts for data extraction (the latter two being open-source). A major advantage is that a lot of the work can be done automatically, provided you script it. Depending on what you would like to do, a lot of features may exist in form of built-in functions.

While GUI software for this application may exist as well, I think that writing your own scripts is worth your while because of repeatability and you being in control of every step in the extraction chain. This makes it easy for you slightly alter features, for example changing the standard bark-band definition so that the two lowest bands are split into two bands each, which I've seen in a couple of audio classification publications.

Another advantage is that for further classification algorithms written by yourself, you can easily connect them to your code.


Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.