|
danstowell@5
|
1
|
|
danstowell@5
|
2 smacpy - simple-minded audio classifier in python
|
|
danstowell@5
|
3 =================================================
|
|
danstowell@5
|
4
|
|
danstowell@5
|
5 Copyright (c) 2012 Dan Stowell and Queen Mary University of London
|
|
danstowell@5
|
6 (incorporating code Copyright (c) 2009 Gyorgy Fazekas and Queen Mary University of London)
|
|
danstowell@6
|
7 - for licence information see the file named COPYING.
|
|
danstowell@5
|
8
|
|
danstowell@5
|
9 This is a classifier that you can train on a set of labelled audio files, and then it predicts a label for further audio files.
|
|
danstowell@5
|
10 It is designed with two main aims:
|
|
danstowell@6
|
11
|
|
danstowell@6
|
12 1. to provide a baseline against which to test more advanced audio classifiers;
|
|
danstowell@6
|
13 2. to provide a simple code example of a classifier which people are free to build on.
|
|
danstowell@5
|
14
|
|
danstowell@33
|
15 It uses a workflow which was very common before the age of deep learning, and might still be useful for low-complexity audio tasks: take an audio clip as input, converting it frame-by-frame into MFCCs, and modelling the MFCC "bag of frames" with a GMM.
|
|
danstowell@5
|
16
|
|
danstowell@5
|
17 Requirements
|
|
danstowell@5
|
18 ------------
|
|
danstowell@33
|
19 * Python 2.7 or later, or Python 3
|
|
danstowell@5
|
20 * Python modules:
|
|
danstowell@5
|
21 * numpy
|
|
danstowell@33
|
22 * [librosa](http://librosa.org/)
|
|
danstowell@20
|
23 * [sckikit-learn](http://scikit-learn.sourceforge.net/)
|
|
danstowell@5
|
24
|
|
danstowell@35
|
25 It has been tested on python 2.7 and 3.8 (on Ubuntu).
|
|
danstowell@7
|
26
|
|
danstowell@5
|
27
|
|
danstowell@5
|
28 Usage example 1: commandline
|
|
danstowell@5
|
29 -------------
|
|
danstowell@5
|
30 If you invoke the script from the commandline (e.g. "python smacpy.py") it will assume there is a folder called "wavs"
|
|
danstowell@5
|
31 and inside that folder are multiple WAV files, each of which has an underscore in the filename,
|
|
danstowell@5
|
32 and the class label is the text BEFORE the underscore.
|
|
danstowell@19
|
33 It will train a model using the wavs, and then test it on the same wavs (dividing the collection up so it can do a "crossvalidated" test).
|
|
danstowell@5
|
34
|
|
danstowell@7
|
35 To train and test on different folders, you can run it like this:
|
|
danstowell@7
|
36
|
|
danstowell@7
|
37 python smacpy.py -t trainwavs -T testwavs
|
|
danstowell@7
|
38
|
|
danstowell@7
|
39
|
|
danstowell@5
|
40 Usage example 2: from your own code
|
|
danstowell@5
|
41 -------------
|
|
danstowell@5
|
42 In this hypothetical example we train on four audio files, labelled as either 'usa' or 'uk', and then test on a separate audio file of someone called hubert:
|
|
danstowell@5
|
43
|
|
danstowell@5
|
44 from smacpy import Smacpy
|
|
danstowell@5
|
45 model = Smacpy("wavs/training", {'karen01.wav':'usa', 'john01.wav':'uk', 'steve02.wav':'usa', 'joe03.wav':'uk'})
|
|
danstowell@5
|
46 model.classify('wavs/testing/hubert01.wav')
|
|
danstowell@5
|
47
|
