|
wolffd@0
|
1 function demo8classification
|
|
wolffd@0
|
2 % To get familiar with different approaches of classification using
|
|
wolffd@0
|
3 % MIRtoolbox, and to assess their performances.
|
|
wolffd@0
|
4
|
|
wolffd@0
|
5 % Part 1. The aim of this experiment is to categorize a set of very short
|
|
wolffd@0
|
6 % musical excerpts according to their genres, through a supervised learning.
|
|
wolffd@0
|
7
|
|
wolffd@0
|
8 % 1.3. Select the training set for current directory.
|
|
wolffd@0
|
9 try
|
|
wolffd@0
|
10 cd train_set
|
|
wolffd@0
|
11 catch
|
|
wolffd@0
|
12 error('Please change current directory to ''MIRtoolboxDemos'' directory')
|
|
wolffd@0
|
13 end
|
|
wolffd@0
|
14
|
|
wolffd@0
|
15 % Load all the files of the folder into one audio structure (called for
|
|
wolffd@0
|
16 % instance training), and associate for each folder a label defined by the
|
|
wolffd@0
|
17 % first two letters of the respective file name.
|
|
wolffd@0
|
18 train = miraudio('Folder','Label',1:2);
|
|
wolffd@0
|
19 cd ..
|
|
wolffd@0
|
20
|
|
wolffd@0
|
21 % In the same way, select the testing set for current directory, and load
|
|
wolffd@0
|
22 % all the files including their labels:
|
|
wolffd@0
|
23 cd test_set
|
|
wolffd@0
|
24 test = miraudio('Folder','Label',1:2);
|
|
wolffd@0
|
25 cd ..
|
|
wolffd@0
|
26
|
|
wolffd@0
|
27 % 1.4. Compute the mel-frequency cepstrum coefficient for each different
|
|
wolffd@0
|
28 % audio file of both sets:
|
|
wolffd@0
|
29 mfcc_train = mirmfcc(train);
|
|
wolffd@0
|
30 mfcc_test = mirmfcc(test);
|
|
wolffd@0
|
31
|
|
wolffd@0
|
32 % 1.5. Estimate the label (i.e., genre) of each file from the testing set,
|
|
wolffd@0
|
33 % based on a prior learning using the training set. Use for this purpose
|
|
wolffd@0
|
34 % the classify function.
|
|
wolffd@0
|
35 help mirclassify
|
|
wolffd@0
|
36
|
|
wolffd@0
|
37 % Let's first try a classification based of mfcc, for instance, using the
|
|
wolffd@0
|
38 % minimum distance strategy:
|
|
wolffd@0
|
39 mirclassify(test,mfcc_test,train,mfcc_train)
|
|
wolffd@0
|
40
|
|
wolffd@0
|
41 % The results indicates the outcomes and the total correct classification
|
|
wolffd@0
|
42 % rate (CCR).
|
|
wolffd@0
|
43
|
|
wolffd@0
|
44 % 1.6. Let's try a k-nearest-neighbour strategy. For instance, for k = 5:
|
|
wolffd@0
|
45 mirclassify(test,mfcc_test,train,mfcc_train,5)
|
|
wolffd@0
|
46
|
|
wolffd@0
|
47 % 1.7. Use a Gaussian mixture modelling with one gaussian per class:
|
|
wolffd@0
|
48 mirclassify(test,mfcc_test,train,mfcc_train,'GMM',1)
|
|
wolffd@0
|
49
|
|
wolffd@0
|
50 % try also with three Gaussians per class.
|
|
wolffd@0
|
51 mirclassify(test,mfcc_test,train,mfcc_train,'GMM',3)
|
|
wolffd@0
|
52
|
|
wolffd@0
|
53 % As this strategy is stochastic, the results vary for every trial.
|
|
wolffd@0
|
54 mirclassify(test,mfcc_test,train,mfcc_train,'GMM',1)
|
|
wolffd@0
|
55 mirclassify(test,mfcc_test,train,mfcc_train,'GMM',1)
|
|
wolffd@0
|
56 mirclassify(test,mfcc_test,train,mfcc_train,'GMM',3)
|
|
wolffd@0
|
57 mirclassify(test,mfcc_test,train,mfcc_train,'GMM',3)
|
|
wolffd@0
|
58
|
|
wolffd@0
|
59 % 1.8. Carry out the classification using other features such as spectral
|
|
wolffd@0
|
60 % centroid:
|
|
wolffd@0
|
61 spectrum_train = mirspectrum(train);
|
|
wolffd@0
|
62 spectrum_test = mirspectrum(test);
|
|
wolffd@0
|
63 centroid_train = mircentroid(spectrum_train);
|
|
wolffd@0
|
64 centroid_test = mircentroid(spectrum_test);
|
|
wolffd@0
|
65 mirclassify(test,centroid_test,train,centroid_train,'GMM',1)
|
|
wolffd@0
|
66 mirclassify(test,centroid_test,train,centroid_train,'GMM',1)
|
|
wolffd@0
|
67 mirclassify(test,centroid_test,train,centroid_train,'GMM',3)
|
|
wolffd@0
|
68 mirclassify(test,centroid_test,train,centroid_train,'GMM',3)
|
|
wolffd@0
|
69
|
|
wolffd@0
|
70 % try also spectral entropy and spectral irregularity.
|
|
wolffd@0
|
71 entropy_train = mirentropy(spectrum_train);
|
|
wolffd@0
|
72 entropy_test = mirentropy(spectrum_test);
|
|
wolffd@0
|
73 mirclassify(test,entropy_test,train,entropy_train,'GMM',1)
|
|
wolffd@0
|
74 mirclassify(test,entropy_test,train,entropy_train,'GMM',1)
|
|
wolffd@0
|
75 mirclassify(test,entropy_test,train,entropy_train,'GMM',3)
|
|
wolffd@0
|
76 mirclassify(test,entropy_test,train,entropy_train,'GMM',3)
|
|
wolffd@0
|
77
|
|
wolffd@0
|
78 irregularity_train = mirregularity(spectrum_train,'Contrast',.1);
|
|
wolffd@0
|
79 irregularity_test = mirregularity(spectrum_test,'Contrast',.1);
|
|
wolffd@0
|
80 mirclassify(test,irregularity_test,train,irregularity_train,'GMM',1)
|
|
wolffd@0
|
81 mirclassify(test,irregularity_test,train,irregularity_train,'GMM',1)
|
|
wolffd@0
|
82 mirclassify(test,irregularity_test,train,irregularity_train,'GMM',3)
|
|
wolffd@0
|
83 mirclassify(test,irregularity_test,train,irregularity_train,'GMM',3)
|
|
wolffd@0
|
84
|
|
wolffd@0
|
85 % Try classification based on a set of features such as:
|
|
wolffd@0
|
86 mirclassify(test,{entropy_test,centroid_test},...
|
|
wolffd@0
|
87 train,{entropy_train,centroid_train},'GMM',1)
|
|
wolffd@0
|
88 mirclassify(test,{entropy_test,centroid_test},...
|
|
wolffd@0
|
89 train,{entropy_train,centroid_train},'GMM',1)
|
|
wolffd@0
|
90 mirclassify(test,{entropy_test,centroid_test},...
|
|
wolffd@0
|
91 train,{entropy_train,centroid_train},'GMM',3)
|
|
wolffd@0
|
92 mirclassify(test,{entropy_test,centroid_test},...
|
|
wolffd@0
|
93 train,{entropy_train,centroid_train},'GMM',3)
|
|
wolffd@0
|
94
|
|
wolffd@0
|
95 % 1.9. By varying the features used for classification, the strategies and
|
|
wolffd@0
|
96 % their parameters, try to find an optimal strategy that give best correct
|
|
wolffd@0
|
97 % classification rate.
|
|
wolffd@0
|
98 bright_train = mirbrightness(spectrum_train);
|
|
wolffd@0
|
99 bright_test = mirbrightness(spectrum_test);
|
|
wolffd@0
|
100 rolloff_train = mirbrightness(spectrum_train);
|
|
wolffd@0
|
101 rolloff_test = mirbrightness(spectrum_test);
|
|
wolffd@0
|
102 spread_train = mirspread(spectrum_train);
|
|
wolffd@0
|
103 spread_test = mirspread(spectrum_test);
|
|
wolffd@0
|
104 mirclassify(test,{bright_test,rolloff_test,spread_test},...
|
|
wolffd@0
|
105 train,{bright_train,rolloff_train,spread_train},'GMM',3)
|
|
wolffd@0
|
106 skew_train = mirskewness(spectrum_train);
|
|
wolffd@0
|
107 skew_test = mirskewness(spectrum_test);
|
|
wolffd@0
|
108 kurtosis_train = mirkurtosis(spectrum_train);
|
|
wolffd@0
|
109 kurtosis_test = mirkurtosis(spectrum_test);
|
|
wolffd@0
|
110 flat_train = mirflatness(spectrum_train);
|
|
wolffd@0
|
111 flat_test = mirflatness(spectrum_test);
|
|
wolffd@0
|
112 mirclassify(test,{skew_test,kurtosis_test,flat_test},...
|
|
wolffd@0
|
113 train,{skew_train,kurtosis_train,flat_train},'GMM',3)
|
|
wolffd@0
|
114 for i = 1:3
|
|
wolffd@0
|
115 mirclassify(test,{mfcc_test,centroid_test,skew_test,kurtosis_test,...
|
|
wolffd@0
|
116 flat_test,entropy_test,irregularity_test,...
|
|
wolffd@0
|
117 bright_test,rolloff_test,spread_test},...
|
|
wolffd@0
|
118 train,{mfcc_train,centroid_train,skew_train,kurtosis_train,...
|
|
wolffd@0
|
119 flat_train,entropy_train,irregularity_train,...
|
|
wolffd@0
|
120 bright_train,rolloff_train,spread_train},'GMM',3)
|
|
wolffd@0
|
121 end
|
|
wolffd@0
|
122
|
|
wolffd@0
|
123 % You can also try to change the size of the training and testing sets (by
|
|
wolffd@0
|
124 % simply interverting them for instance).
|
|
wolffd@0
|
125 for i = 1:3
|
|
wolffd@0
|
126 mirclassify(train,{mfcc_train,centroid_train,skew_train,kurtosis_train,...
|
|
wolffd@0
|
127 flat_train,entropy_train,irregularity_train,...
|
|
wolffd@0
|
128 bright_train,rolloff_train,spread_train},...
|
|
wolffd@0
|
129 test,{mfcc_test,centroid_test,skew_test,kurtosis_test,...
|
|
wolffd@0
|
130 flat_test,entropy_test,irregularity_test,...
|
|
wolffd@0
|
131 bright_test,rolloff_test,spread_test},'GMM',3)
|
|
wolffd@0
|
132 end
|
|
wolffd@0
|
133
|
|
wolffd@0
|
134 %%
|
|
wolffd@0
|
135 % Part 2. In this second experiment, we will try to cluster the segments of
|
|
wolffd@0
|
136 % an audio file according to their mutual similarity.
|
|
wolffd@0
|
137
|
|
wolffd@0
|
138 % 2.1. To simplify the computation, downsample
|
|
wolffd@0
|
139 % the audio file to 11025 Hz.
|
|
wolffd@0
|
140 a = miraudio('czardas','Sampling',11025);
|
|
wolffd@0
|
141
|
|
wolffd@0
|
142 % 2.2. Decompose the file into successive frames of 2 seconds with half-
|
|
wolffd@0
|
143 % overlapping.
|
|
wolffd@0
|
144 f = mirframe(a,2,.1);
|
|
wolffd@0
|
145
|
|
wolffd@0
|
146 % 2.3. Segment the file based on the novelty of the key strengths.
|
|
wolffd@0
|
147 n = mirnovelty(mirkeystrength(f),'KernelSize',5)
|
|
wolffd@0
|
148 p = mirpeaks(n)
|
|
wolffd@0
|
149 s = mirsegment(a,p)
|
|
wolffd@0
|
150
|
|
wolffd@0
|
151 % 2.4. Compute the key strengths of each segment.
|
|
wolffd@0
|
152 ks = mirkeystrength(s)
|
|
wolffd@0
|
153
|
|
wolffd@0
|
154 % 2.5. Cluster the segments according to their key strengths.
|
|
wolffd@0
|
155 help mircluster
|
|
wolffd@0
|
156 mircluster(s,ks)
|
|
wolffd@0
|
157
|
|
wolffd@0
|
158 % The k means algorithm used in the clustering is stochastic, and its
|
|
wolffd@0
|
159 % results may vary at each run. By default, the algorithm is run 5 times
|
|
wolffd@0
|
160 % and the best result is selected. Try the analysis with a higher number of
|
|
wolffd@0
|
161 % runs:
|
|
wolffd@0
|
162 mircluster(s,ks,'Runs',10)
|
