-
Notifications
You must be signed in to change notification settings - Fork 0
/
Synchronize.py
336 lines (311 loc) · 12.4 KB
/
Synchronize.py
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
"""
Purpose: To synchronize cover songs using my similarity fusion technique to do
an alignment, and pyrubberband to appropriately time shift
"""
import numpy as np
import os
import glob
import scipy.io as sio
import scipy.misc
import time
import matplotlib.pyplot as plt
from GeometricCoverSongs.CSMSSMTools import *
from GeometricCoverSongs.BlockWindowFeatures import *
from GeometricCoverSongs.pyMIRBasic.AudioIO import *
from GeometricCoverSongs.pyMIRBasic.Onsets import *
import json
import pyrubberband as pyrb
import subprocess
def getGreedyPerm(D):
"""
Purpose: Naive O(N^2) algorithm to do the greedy permutation
param: D (NxN distance matrix for points)
return: (permutation (N-length array of indices),
lambdas (N-length array of insertion radii))
"""
N = D.shape[0]
#By default, takes the first point in the list to be the
#first point in the permutation, but could be random
perm = np.zeros(N, dtype=np.int64)
lambdas = np.zeros(N)
ds = D[0, :]
for i in range(1, N):
idx = np.argmax(ds)
perm[i] = idx
lambdas[i] = ds[idx]
ds = np.minimum(ds, D[idx, :])
return (perm, lambdas)
def syncBlocks(path, CSM, beats1, beats2, Fs, hopSize, XAudio1, XAudio2, BeatsPerBlock, fileprefix = ""):
"""
:param path: Px2 array representing a partial warping path to align two songs
:param CSM: The cross similarity matrix between two songs
:param beats1: An array of beat onsets for song 1 in increments of hopSize
:param beats2: An array of beat onsets for song 2 in increments of hopSize
:param XAudio1: The raw audio samples for song 1
:param XAudio2: The raw audio samples for song 2
:param BeatsPerBlock: The number of beats per block for each pixel in the CSM
:param fileprefix: Prefix of each stretched block to save. By default, blank,\
so no debugging info saved
:returns (XFinal: An NSamples x 2 array with the first song along the first column\
and the second synchronized song along the second column,\
beatsFinal: An array of the locations in samples of the beat onsets in XFinal \
scoresFinal: An array of matching scores for each beat)
"""
XFinal = np.array([[0, 0]])
beatsFinal = [] #The final beat locations based on hop size
scoresFinal = []
for i in range(path.shape[0]):
[j, k] = [path[i, 0], path[i, 1]]
if j >= CSM.shape[0] or k >= CSM.shape[1]:
break
scoresFinal.append(CSM[j, k])
t1 = beats1[j]*hopSize
t2 = beats1[j+BeatsPerBlock]*hopSize
s1 = beats2[k]*hopSize
s2 = beats2[k+BeatsPerBlock]*hopSize
x1 = XAudio1[t1:t2]
x2 = XAudio2[s1:s2]
#Figure out the time factor by which to stretch x2 so it aligns
#with x1
fac = float(len(x1))/len(x2)
print("fac = ", fac)
x2 = pyrb.time_stretch(x2, Fs, 1.0/fac)
print("len(x1) = %i, len(x2) = %i"%(len(x1), len(x2)))
N = min(len(x1), len(x2))
x1 = x1[0:N]
x2 = x2[0:N]
X = np.zeros((N, 2))
X[:, 0] = x1
X[:, 1] = x2
if len(fileprefix) > 0:
filename = "%s_%i.mp3"%(fileprefix, i)
sio.wavfile.write("temp.wav", Fs, X)
subprocess.call(["avconv", "-i", "temp.wav", filename])
beat1 = beats1[j+1]*hopSize-t1
beatsFinal.append(XFinal.shape[0])
XFinal = np.concatenate((XFinal, X[0:beat1, :]))
return (XFinal, beatsFinal, scoresFinal)
def expandBeats(beats, bSub):
if bSub == 1:
return beats
import scipy.interpolate as interp
idx = np.arange(beats.size)
idxx = (np.arange(bSub*beats.size)/float(bSub))[0:-bSub+1]
y = interp.spline(idx, beats, idxx)
y = np.round(y)
return np.array(y, dtype = np.int64)
def synchronize(filename1, filename2, hopSize, TempoBiases, bSub, FeatureParams, CSMTypes,\
Kappa, fileprefix = "", doPlot = False, outputSnippets = True, doNegative = False):
"""
:param filename1: First song path
:param filename2: Second song path
:param hopSize: Hop size (in samples) to be used between feature windows
:param TempoBiases: The different tempo levels to be tried for beat tracking
:param bSub: The factor by which to subdivide the beats
:param FeatureParams: Dictionary of feature parameters
:param CSMTypes: Dictionary of CSM types for different features
:param Kappa: Nearest neighbor fraction for making binary CSM
:param fileprefix: File prefix for debugging plots and intermediate audio files
:param doPlot: Whether to plot alignment
:param outputSnippets: Whether to output aligned audio snippets block by block
:param doNegative: Whether to sample negative matches
"""
print("Loading %s..."%filename1)
(XAudio1, Fs) = getAudioLibrosa(filename1)
print("Loading %s..."%filename2)
(XAudio2, Fs) = getAudioLibrosa(filename2)
print("Fs = ", Fs)
maxScore = 0.0
maxRes = {}
for TempoBias1 in TempoBiases:
for TempoBias2 in TempoBiases:
print("Doing TempoBias1 = %i, TempoBias2 = %i..."%(TempoBias1, TempoBias2))
(tempo, beats1) = getBeats(XAudio1, Fs, TempoBias1, hopSize, filename1)
beats1 = expandBeats(beats1, bSub)
(Features1, O1) = getBlockWindowFeatures((XAudio1, Fs, tempo, beats1, hopSize, FeatureParams))
(tempo, beats2) = getBeats(XAudio2, Fs, TempoBias2, hopSize, filename2)
beats2 = expandBeats(beats2, bSub)
(Features2, O2) = getBlockWindowFeatures((XAudio2, Fs, tempo, beats2, hopSize, FeatureParams))
print("Doing similarity fusion")
K = 20
NIters = 3
res = getCSMSmithWatermanScoresEarlyFusionFull(Features1, O1, Features2, O2, Kappa, K, NIters, CSMTypes, doPlot = True, conservative = False)
sio.savemat("Synced.mat", res)
print("res.keys() = ", res.keys())
print("score = ", res['score'])
if res['score'] > maxScore:
print("New maximum score!")
maxScore = res['score']
maxRes = res
res['beats1'] = beats1
res['beats2'] = beats2
res['TempoBias1'] = TempoBias1
res['TempoBias2'] = TempoBias2
res = maxRes
print("TempoBias1 = %i, TempoBias2 = %i"%(res['TempoBias1'], res['TempoBias2']))
beats1 = res['beats1']
beats2 = res['beats2']
bs = hopSize*beats1/float(Fs)
print("Interval 1: %.3g"%np.mean(bs[1::]-bs[0:-1]))
bs = hopSize*beats2/float(Fs)
print("Interval 2: %.3g"%np.mean(bs[1::]-bs[0:-1]))
CSM = res['CSM']
CSM = CSM/np.max(CSM) #Normalize so highest score is 1
path = np.array(res['path'])
if doPlot:
plt.clf()
plt.figure(figsize=(20, 8))
plt.subplot(121)
plt.imshow(CSM, cmap = 'afmhot')
plt.hold(True)
plt.plot(path[:, 1], path[:, 0], '.')
plt.subplot(122)
plt.plot(path[:, 0], path[:, 1])
plt.savefig("%sBlocksAligned.svg"%fileprefix, bbox_inches = 'tight')
#Now extract signal snippets that are in correspondence, beat by beat
BeatsPerBlock = FeatureParams['MFCCBeatsPerBlock']
path = np.flipud(path)
(XFinal, beatsFinal, scoresFinal) = syncBlocks(path, CSM, beats1, beats2, Fs, hopSize, XAudio1, XAudio2, BeatsPerBlock, fileprefix = "")
#Write out true positives synced
if len(fileprefix) > 0:
sio.wavfile.write("temp.wav", Fs, XFinal)
subprocess.call(["avconv", "-i", "temp.wav", "%sTrue.mp3"%fileprefix])
#Write out true positives beat times and scores
[beatsFinal, scoresFinal] = [np.array(beatsFinal), np.array(scoresFinal)]
if len(fileprefix) > 0:
sio.savemat("%sTrue.mat"%fileprefix, {"beats1":beats1, "beats2":beats2, "path":path, "beats":beatsFinal, "scores":scoresFinal, "BeatsPerBlock":BeatsPerBlock, "hopSize":hopSize})
#Now save negative examples (same number as positive blocks)
if doNegative:
NBlocks = path.shape[0]
x = CSM.flatten()
idx = np.argsort(x)
idx = idx[0:5*CSM.shape[0]]
idxy = np.unravel_index(idx, CSM.shape)
idx = np.zeros((idx.size, 2), dtype = np.int64)
idx[:, 0] = idxy[0]
idx[:, 1] = idxy[1]
D = getCSM(idx, idx)
#Do furthest point sampling on negative locations
(perm, lambdas) = getGreedyPerm(D)
path = idx[perm[0:NBlocks], :]
if doPlot:
plt.clf()
plt.imshow(CSM, interpolation = 'nearest', cmap = 'afmhot')
plt.hold(True)
plt.plot(path[:, 1], path[:, 0], '.')
plt.savefig("%sBlocksMisaligned.svg"%fileprefix, bbox_inches = 'tight')
#Output negative example audio synced
(XFinal, beatsFinal, scoresFinal) = syncBlocks(path, CSM, beats1, beats2, Fs, hopSize, XAudio1, XAudio2, BeatsPerBlock, fileprefix = "%sFalse"%fileprefix)
sio.savemat("%sFalse.mat"%fileprefix, {"scores":scoresFinal, "BeatsPerBlock":BeatsPerBlock, "hopSize":hopSize})
return {'X':XFinal, 'Fs':Fs, 'beatsFinal':beatsFinal, 'scoresFinal':scoresFinal}
if __name__ == '__main__':
Kappa = 0.1
hopSize = 512
bSub = 1
TempoBiases = [0]
fileprefix = ""
doPlot = False
"""
filename1 = "DespacitoOrig.mp3"
filename2 = "DespacitoMetal.mp3"
fileprefix = "Despacito" #Save a JSON file with this prefix
artist1 = "Luis Fonsi ft. Daddy Yankee"
artist2 = "Leo Moracchioli"
songName = "Despacito"
"""
"""
filename1 = "WakaNoHands.webm"
filename2 = "DannyVolaNoHands.m4a"
artist1 = "Waka Flocka Flame"
artist2 = "Danny Vola"
fileprefix = "nohands"
songName = "No Hands"
"""
"""
filename1 = "LaFolia1.mp3"
filename2 = "LaFolia2.mp3"
artist1 = "Vivaldi"
artist2 = "Vivaldi"
fileprefix = "LaFolia"
songName = "La Folia"
"""
filename1 = "music/SmoothCriminalMJ.mp3"
filename2 = "music/SmoothCriminalAAF.mp3"
artist1 = "Michael Jackson"
artist2 = "Alien Ant Farm"
fileprefix = "smoothcriminal"
songName = "Smooth Criminal"
TempoBiases = [180]
"""
filename1 = "music/Rednex/CottoneyeJoe.mp3"
filename2 = "music/Rednex/CottoneyeJoeCover.mp3"
artist1 = "Rednex"
artist2 = "Leo Moracchioli"
fileprefix = "cottoneyejoe"
songName = "Cottoneye Joe"
TempoBiases = [60, 120, 180]
TempoBiases = [0]
"""
"""
filename1 = "music/Aha/AhaTakeOnMe.mp3"
filename2 = "music/Aha/MXPXTakeOnMe.mp3"
artist1 = "Aha"
artist2 = "MXPX"
fileprefix = "takeonme"
songName = "Take On Me"
TempoBiases = [60, 120, 180]
TempoBiases = [0]
"""
"""
filename1 = "music/HersheyBar/StanGetzQuartet.mp3"
filename2 = "music/HersheyBar/KenichiroNishihara.mp3"
artist1 = "Stan Getz Quartet"
artist2 = "Kenichiro Nishihara"
fileprefix = "hersheybar"
songName = "Hershey Bar"
TempoBiases = [60, 120, 180]
TempoBiases = [0]
"""
"""
filename1 = "music/SweetDreams/Eurythmics.mp3"
filename2 = "music/SweetDreams/MarilynManson.mp3"
artist1 = "Eurythmics"
artist2 = "Marilyn Manson"
fileprefix = "sweetdreams"
songName = "Sweet Dreams"
TempoBiases = [0]
"""
"""
filename1 = "music/MIDIExample/StayinAliveMIDI.mp3"
filename2 = "music/MIDIExample/StayinAlive.mp3"
artist1 = "BeeGeesMIDI"
artist2 = "BeeGees"
fileprefix = "stayinalive"
songName = "Stayin Alive"
TempoBiases = [120]
fileprefix = "beegees"
doPlot = True
"""
"""
filename1 = "music/Coldplay/InMyPlaceColdplay.mp3"
filename2 = "music/Coldplay/InMyPlaceMetal.mp3"
artist1 = "Coldplay"
artist2 = "Leo Moriachielli"
fileprefix = "coldplay"
songName = "In My Place"
TempoBiases = [0]
"""
"""
filename1 = "music/Oasis/WonderwallOasis.mp3"
filename2 = "music/Oasis/WonderwallMetal.mp3"
artist1 = "Oasis"
artist2 = "Leo Moriachielli"
fileprefix = "oasis"
songName = "Wonderwall"
TempoBiases = [180]
"""
FeatureParams = {'MFCCBeatsPerBlock':20, 'MFCCSamplesPerBlock':200, 'DPixels':50, 'ChromaBeatsPerBlock':20, 'ChromasPerBlock':40}
CSMTypes = {'MFCCs':'Euclidean', 'SSMs':'Euclidean', 'Chromas':'CosineOTI'}
res = synchronize(filename1, filename2, hopSize, TempoBiases, bSub, FeatureParams, CSMTypes, Kappa, fileprefix=fileprefix, doPlot=doPlot)
sio.wavfile.write("temp.wav", res['Fs'], res['X'])
subprocess.call(["avconv", "-i", "temp.wav", "%sTrue.mp3"%fileprefix])