Mercurial > hg > chourdakisreiss2016

annotate experiment-reverb/code/selfsim.py @ 2:c87a9505f294 tip

Find changesets by keywords (author, files, the commit message), revision number or hash, or revset expression.
Added LICENSE for code, removed .wav files
author Emmanouil Theofanis Chourdakis <e.t.chourdakis@qmul.ac.uk>
date Sat, 30 Sep 2017 13:25:50 +0100
parents 246d5546657c
children
rev   line source
e@0 1 # -*- coding: utf-8 -*-
e@0 2 """
e@0 3 Created on Mon Jun 8 11:19:15 2015
e@0 4
e@0 5 @author: mmxgn
e@0 6 """
e@0 7 # Codes taken from: https://github.com/urinieto/msaf/blob/master/msaf/algorithms/foote/segmenter.py
e@0 8
e@0 9
e@0 10
e@0 11 if __name__=="__main__":
e@0 12 from sys import argv
e@0 13 if len(argv) != 3:
e@0 14 print("Incorrect number of arguments:")
e@0 15 print("Usage: ")
e@0 16 print("%s <input>")
e@0 17 print("")
e@0 18 print("Arguments:")
e@0 19 print("<input>\tThe input filename. Can be .wav, .mp3, etc...")
e@0 20 print("<output_folder>\tThe output folders. Segments will be stored under names 'name_segN'")
e@0 21 sys.exit(-1)
e@0 22 else:
e@0 23 print("[II] Applying the method found in: ")
e@0 24 print("[II] Automatic Audio Segmentation using a measure of Audio Novelty")
e@0 25 print("[II] - Jonathar Foote ")
e@0 26 print("[II] Loading libraries")
e@0 27
e@0 28 import essentia
e@0 29 from essentia import Pool
e@0 30 from essentia.standard import *
e@0 31 import csv
e@0 32 import yaml
e@0 33
e@0 34 # reqyures matplotlib
e@0 35 from pylab import *
e@0 36
e@0 37 #requires numpy
e@0 38 from numpy import *
e@0 39
e@0 40 import wave
e@0 41
e@0 42
e@0 43 from scipy.spatial import distance
e@0 44 from scipy.ndimage import filters
e@0 45 d = {}
e@0 46 v = {}
e@0 47
e@0 48 fname = argv[1]
e@0 49 outfdir = argv[2]
e@0 50
e@0 51 print "[II] Using filename: %s" % fname
e@0 52 print "[II] Using output folder: %s" % outfdir
e@0 53
e@0 54 name = fname.split('.')[-2].split('/')[-1]
e@0 55
e@0 56 print "[II] Segments will be saved in the form '%s/%s_segN.mp3'" % (outfdir, name)
e@0 57
e@0 58
e@0 59 trackname = fname.split('.')[0].split('/')[-1]
e@0 60
e@0 61
e@0 62 # if outfname.partition('.')[-1].lower() not in ['json', 'yaml']:
e@0 63 # print("Please choose a .json or .yaml as an output file.")
e@0 64 # sys.exit(-1)
e@0 65 # else:
e@0 66 # if outfname.partition('.')[-1].lower() == 'json':
e@0 67 # output = YamlOutput(filename = outfname, format='json')
e@0 68 # else:
e@0 69 # output = YamlOutput(filename = outfname, format='yaml')
e@0 70
e@0 71 print("Feature extraction of `%s\'" % fname)
e@0 72
e@0 73 # Sampling Rate
e@0 74 SR = 21000.0
e@0 75
e@0 76
e@0 77 # Audio Loader
e@0 78 loader = MonoLoader(filename = fname, sampleRate=SR)
e@0 79
e@0 80 # Lowpass audio
e@0 81 lp = LowPass(cutoffFrequency=SR/4, sampleRate=SR)
e@0 82
e@0 83 # Audio
e@0 84 audio = lp(loader())
e@0 85
e@0 86
e@0 87
e@0 88 # For MFCCs
e@0 89
e@0 90 w_hanning = Windowing(type = "hann")
e@0 91 spectrum = Spectrum()
e@0 92 mfcc = MFCC()
e@0 93
e@0 94
e@0 95 frameSize = int(0.2 * SR) # Change this depending whether it's music or sound
e@0 96
e@0 97 pool = essentia.Pool()
e@0 98
e@0 99
e@0 100
e@0 101 for frame in FrameGenerator(audio, frameSize = frameSize, hopSize = frameSize/2):
e@0 102 mfcc_bands, mfcc_coeffs = mfcc(spectrum(w_hanning(frame)))
e@0 103 pool.add("lowlevel.mfcc_selfsim", mfcc_coeffs)
e@0 104
e@0 105 mfcc_coeffs = pool['lowlevel.mfcc_selfsim']
e@0 106
e@0 107 # selfsim = 1 - pairwise_distances(mfcc_coeffs)#, metric = "cosine")
e@0 108 selfsim = distance.pdist(mfcc_coeffs, metric='seuclidean')
e@0 109 selfsim = distance.squareform(selfsim)
e@0 110 selfsim /= selfsim.max()
e@0 111 selfsim = 1 - selfsim
e@0 112 # Calculating cosine distances as a better metric
e@0 113
e@0 114 C = array([[1,-1],[-1,1]])
e@0 115
e@0 116 def Novelty(S, C = array([[1, -1],[-1, 1]])):
e@0 117 L = C.shape[0]
e@0 118
e@0 119 horconcat = concatenate((S[:, 0:L/2], S, S[:,-L/2:]), axis=1)
e@0 120 verconcat = concatenate((horconcat[0:L/2,:], horconcat, horconcat[-L/2:,:]), axis=0)
e@0 121
e@0 122
e@0 123 N = zeros((S.shape[0],))
e@0 124
e@0 125 for i in range(0, len(N)):
e@0 126 S_ = 0
e@0 127 for m in range(-L/2, L/2):
e@0 128 for n in range(-L/2, L/2):
e@0 129 # print (m,n), (L/2+m, L/2+n)
e@0 130 S_ += C[L/2+m, L/2+n]*verconcat[i+m+L/2, i+n-L/2]
e@0 131 # S_ += verconcat[i+m+L/2, i+m-L/2]
e@0 132
e@0 133 # print S_
e@0 134 N[i] = S_
e@0 135
e@0 136 return N
e@0 137
e@0 138 def novel(S, C = array([[1, -1], [-1, 1]])):
e@0 139 N = S.shape[0]
e@0 140 M = C.shape[0]
e@0 141
e@0 142 novelty = zeros(N)
e@0 143
e@0 144 for i in xrange(M/2, N-M/2+1):
e@0 145 novelty[i] = sum(S[i-M/2:i+M/2,i-M/2:i+M/2] * C)
e@0 146
e@0 147 novelty += novelty.min()
e@0 148 novelty /= novelty.max()
e@0 149
e@0 150 return novelty
e@0 151
e@0 152
e@0 153
e@0 154 def pick_peaks(nc, L=32):
e@0 155 # Codes taken from: https://github.com/urinieto/msaf/blob/master/msaf/algorithms/foote/segmenter.py
e@0 156
e@0 157 """Obtain peaks from a novelty curve using an adaptive threshold."""
e@0 158 offset = nc.mean() / 20.
e@0 159
e@0 160 nc = filters.gaussian_filter1d(nc, sigma=4) # Smooth out nc
e@0 161
e@0 162 th = filters.median_filter(nc, size=L) + offset
e@0 163 #th = filters.gaussian_filter(nc, sigma=L/2., mode="nearest") + offset
e@0 164
e@0 165 peaks = []
e@0 166 for i in xrange(1, nc.shape[0] - 1):
e@0 167 # is it a peak?
e@0 168 if nc[i - 1] < nc[i] and nc[i] > nc[i + 1]:
e@0 169 # is it above the threshold?
e@0 170 if nc[i] > th[i]:
e@0 171 peaks.append(i)
e@0 172 #plt.plot(nc)
e@0 173 #plt.plot(th)
e@0 174 #for peak in peaks:
e@0 175 #plt.axvline(peak)
e@0 176 #plt.show()
e@0 177
e@0 178 return peaks
e@0 179
e@0 180 from scipy import signal
e@0 181 def compute_gaussian_krnl(M):
e@0 182 """Creates a gaussian kernel following Foote's paper."""
e@0 183 g = signal.gaussian(M, M / 3., sym=True)
e@0 184 G = np.dot(g.reshape(-1, 1), g.reshape(1, -1))
e@0 185 G[M / 2:, :M / 2] = -G[M / 2:, :M / 2]
e@0 186 G[:M / 2, M / 2:] = -G[:M / 2, M / 2:]
e@0 187 return G
e@0 188
e@0 189 K = compute_gaussian_krnl(96)
e@0 190 def kernelMatrix(L):
e@0 191 k1 = concatenate((ones((L/2,L/2)), -1*ones((L/2,L/2))))
e@0 192 k1 = concatenate((k1,-k1),axis=1)
e@0 193 return k1
e@0 194
e@0 195 N = novel(selfsim, K)
e@0 196 peaks = pick_peaks(N)
e@0 197
e@0 198 boundaries = array(peaks)*frameSize/2
e@0 199
e@0 200 sampleRate = SR
e@0 201
e@0 202 audio = MonoLoader(filename=fname, sampleRate = sampleRate)()
e@0 203
e@0 204 from scipy.io.wavfile import write as wavwrite
e@0 205
e@0 206 for b in range(1, len(boundaries)):
e@0 207 outname = '%s/%s_seg%d.wav' % (outfdir, name, b)
e@0 208 segment = audio[boundaries[b-1]:boundaries[b]]
e@0 209 if len(segment) >= 5*SR:
e@0 210 #audioout = MonoWriter(sampleRate = SR, filename=outname)
e@0 211 #audioout(segment)
e@0 212
e@0 213 wavwrite(outname, SR, segment)
e@0 214 print "[II] Saving %s" % outname
e@0 215
e@0 216
e@0 217
e@0 218
e@0 219