def writeSeparatedSignals(self, suffix='.wav'):
"""Writes the separated signals to the files in self.files.
If suffix contains 'VUIMM', then this method will take
the WF0 and HF0 that contain the estimated unvoiced elements.
"""
if 'VUIMM' in suffix:
WF0 = self.SIMMParams['WUF0']
HF0 = self.SIMMParams['HUF0']
else:
WF0 = self.SIMMParams['WF0']
HF0 = self.SIMMParams['HF0']
WGAMMA = self.SIMMParams['WGAMMA']
HGAMMA = self.SIMMParams['HGAMMA']
HPHI = self.SIMMParams['HPHI']
HM = self.SIMMParams['HM']
WM = self.SIMMParams['WM']
alphaR = self.SIMMParams['alphaR']
alphaL = self.SIMMParams['alphaL']
betaR = self.SIMMParams['betaR']
betaL = self.SIMMParams['betaL']
windowSizeInSamples = self.stftParams['windowSizeInSamples']
WPHI = np.dot(WGAMMA, HGAMMA)
SPHI = np.dot(WPHI, HPHI)
SF0 = np.dot(WF0, HF0)
hatSXR = (alphaR**2) * SF0 * SPHI + np.dot(np.dot(WM, betaR**2),HM)
hatSXL = (alphaL**2) * SF0 * SPHI + np.dot(np.dot(WM, betaL**2),HM)
hatSXR = np.maximum(hatSXR, eps)
hatSXL = np.maximum(hatSXL, eps)
hatVR = (alphaR**2) * SPHI * SF0 / hatSXR * self.XR
vestR = slf.istft(hatVR,
hopsize=self.stftParams['hopsize'],
nfft=self.stftParams['NFT'],
window=slf.sinebell(windowSizeInSamples)) / 4.0
hatVR = (alphaL**2) * SPHI * SF0 / hatSXL * self.XL
vestL = slf.istft(hatVR,
hopsize=self.stftParams['hopsize'],
nfft=self.stftParams['NFT'],
window=slf.sinebell(windowSizeInSamples)) / 4.0
vestR = np.array(np.round(vestR*self.scaleData), dtype=self.dataType)
vestL = np.array(np.round(vestL*self.scaleData), dtype=self.dataType)
wav.write(self.files['voc_output_file'][:-4] + suffix,
self.fs,
np.array([vestR,vestL]).T)
hatMR = (np.dot(np.dot(WM,betaR ** 2), HM)) / hatSXR * self.XR
mestR = slf.istft(hatMR,
hopsize=self.stftParams['hopsize'],
nfft=self.stftParams['NFT'],
window=slf.sinebell(windowSizeInSamples)) / 4.0
hatMR = (np.dot(np.dot(WM,betaL ** 2), HM)) / hatSXL * self.XL
mestL = slf.istft(hatMR,
hopsize=self.stftParams['hopsize'],
nfft=self.stftParams['NFT'],
window=slf.sinebell(windowSizeInSamples)) / 4.0
mestR = np.array(np.round(mestR*self.scaleData), dtype=self.dataType)
mestL = np.array(np.round(mestL*self.scaleData), dtype=self.dataType)
wav.write(self.files['mus_output_file'][:-4] + suffix,
self.fs,
np.array([mestR,mestL]).T)
def writeSeparatedSignals(self, suffix='.wav'):
"""Writes the separated signals to the files in self.files.
If suffix contains 'VUIMM', then this method will take
the WF0 and HF0 that contain the estimated unvoiced elements.
"""
if 'VUIMM' in suffix:
WF0 = self.SIMMParams['WUF0']
HF0 = self.SIMMParams['HUF0']
else:
WF0 = self.SIMMParams['WF0']
HF0 = self.SIMMParams['HF0']
WGAMMA = self.SIMMParams['WGAMMA']
HGAMMA = self.SIMMParams['HGAMMA']
HPHI = self.SIMMParams['HPHI']
HM = self.SIMMParams['HM']
WM = self.SIMMParams['WM']
alphaR = self.SIMMParams['alphaR']
alphaL = self.SIMMParams['alphaL']
betaR = self.SIMMParams['betaR']
betaL = self.SIMMParams['betaL']
windowSizeInSamples = self.stftParams['windowSizeInSamples']
WPHI = np.dot(WGAMMA, HGAMMA)
SPHI = np.dot(WPHI, HPHI)
SF0 = np.dot(WF0, HF0)
hatSXR = (alphaR**2) * SF0 * SPHI + np.dot(np.dot(WM, betaR**2), HM)
hatSXL = (alphaL**2) * SF0 * SPHI + np.dot(np.dot(WM, betaL**2), HM)
hatSXR = np.maximum(hatSXR, eps)
hatSXL = np.maximum(hatSXL, eps)
hatVR = (alphaR**2) * SPHI * SF0 / hatSXR * self.XR
vestR = slf.istft(hatVR,
hopsize=self.stftParams['hopsize'],
nfft=self.stftParams['NFT'],
window=slf.sinebell(windowSizeInSamples)) / 4.0
hatVR = (alphaL**2) * SPHI * SF0 / hatSXL * self.XL
vestL = slf.istft(hatVR,
hopsize=self.stftParams['hopsize'],
nfft=self.stftParams['NFT'],
window=slf.sinebell(windowSizeInSamples)) / 4.0
vestR = np.array(np.round(vestR * self.scaleData), dtype=self.dataType)
vestL = np.array(np.round(vestL * self.scaleData), dtype=self.dataType)
wav.write(self.files['voc_output_file'][:-4] + suffix, self.fs,
np.array([vestR, vestL]).T)
hatMR = (np.dot(np.dot(WM, betaR**2), HM)) / hatSXR * self.XR
mestR = slf.istft(hatMR,
hopsize=self.stftParams['hopsize'],
nfft=self.stftParams['NFT'],
window=slf.sinebell(windowSizeInSamples)) / 4.0
hatMR = (np.dot(np.dot(WM, betaL**2), HM)) / hatSXL * self.XL
mestL = slf.istft(hatMR,
hopsize=self.stftParams['hopsize'],
nfft=self.stftParams['NFT'],
window=slf.sinebell(windowSizeInSamples)) / 4.0
mestR = np.array(np.round(mestR * self.scaleData), dtype=self.dataType)
mestL = np.array(np.round(mestL * self.scaleData), dtype=self.dataType)
wav.write(self.files['mus_output_file'][:-4] + suffix, self.fs,
np.array([mestR, mestL]).T)
def __init__(self, inputAudioFilename,
windowSize=0.0464, nbIter=10,
numCompAccomp=40,
minF0=60, maxF0=2000, stepNotes=16,
K_numFilters=4,
P_numAtomFilters=30,
imageCanvas=None, wavCanvas=None,
progressBar=None,
verbose=True,
outputDirSuffix='/'):
"""During init, process is initiated, STFTs are computed,
and the parameters are stored.
Parameters
----------
inputAudioFilename : string
filename of the input audio file
windowSize : double, optional
analysis frame ('windows') size, in s. By default, 0.0464s
nbIter : integer, optional
number of iterations for the estimation algorithm. By default, 10
numCompAccomp : integer, optional
number of components for the accompaniment, default = 40
minF0 : double/integer, optional
lowest F0 candidate (in Hz), default=60Hz
maxF0 : double/integer, optional
highest F0 candidate (in Hz), default=2000Hz
stepNotes : integer, optional
number of F0 candidates in one semitone, default=16 F0s/semitone
K_numFilters : integer, optional
number of filter spectral shapes, default=4
P_numAtomFilters : integer, optional
number of atomic filter smooth spectral shapes, default=30
imageCanvas : MplCanvas/MplCanvas3Axes, optional
an instance of the MplCanvas/MplCanvas3Axes, giving access to the
axes where to draw the HF0 image. By default=None
wavCanvas : MplCanvas/MplCanvas3Axes, optional
an instance of the MplCanvas/MplCanvas3Axes, giving access to the
axes to draw the waveform of the input signal.
progressBar : boolean, optional ???
???
verbose : boolean, optional
Whether to write out or not information about the evolution of the
algorithm. By default=False.
outputDirSuffix : string, optional
the subfolder name (to be appended to the full path to the audio
signal), where the output files are going to be written. By default
='/'
"""
self.files['inputAudioFilename'] = str(inputAudioFilename)
self.imageCanvas = imageCanvas
self.wavCanvas = wavCanvas
self.displayEvolution = True
self.verbose=verbose
if self.imageCanvas is None:
self.displayEvolution = False
if inputAudioFilename[-4:] != ".wav":
raise ValueError("File not WAV file? Only WAV format support, "+\
"for now...")
self.files['outputDirSuffix'] = outputDirSuffix
self.files['outputDir'] = str('/').join(\
self.files['inputAudioFilename'].split('/')[:-1])+\
'/'+self.files['outputDirSuffix'] +'/'
if os.path.isdir(self.files['outputDir']):
print "Output directory already existing - "+\
"NB: overwriting files in:"
print self.files['outputDir']
else:
print "Creating output directory"
print self.files['outputDir']
os.mkdir(self.files['outputDir'])
self.files['pathBaseName'] = self.files['outputDir'] + \
self.files['inputAudioFilename'\
].split('/')[-1][:-4]
self.files['mus_output_file'] = str(self.files['pathBaseName']+\
'_acc.wav')
self.files['voc_output_file'] = str(self.files['pathBaseName']+\
'_lead.wav')
self.files['pitch_output_file'] = str(self.files['pathBaseName']+\
'_pitches.txt')
print "Writing the different following output files:"
print " separated lead in", \
self.files['voc_output_file']
print " separated accompaniment in", \
self.files['mus_output_file']
print " separated lead + unvoc in", \
self.files['voc_output_file'][:-4] + '_VUIMM.wav'
print " separated acc - unvoc in", \
self.files['mus_output_file'][:-4] + '_VUIMM.wav'
print " estimated pitches in", \
self.files['pitch_output_file']
# read the WAV file and store the STFT
self.fs, data = wav.read(self.files['inputAudioFilename'])
# for some bad format wav files, data is a str?
# cf. files from beat/tempo evaluation campaign of MIREX
## print self.fs, data
self.scaleData = 1.2 * np.abs(data).max() # to rescale the data.
self.dataType = data.dtype
data = np.double(data) / self.scaleData # makes data vary from -1 to 1
if data.shape[0] == data.size: # data is multi-channel
print "The audio file is not stereo. Making stereo out of mono."
print "(You could also try the older separateLead.py...)"
data = np.vstack([data,data]).T
if data.shape[1] != 2:
print "The data is multichannel, but not stereo... \n"
print "Unfortunately this program does not scale well. Data is \n"
print "reduced to its 2 first channels.\n"
data = data[:,0:2]
# parameters for the STFT:
self.stftParams['windowSizeInSamples'] = \
slf.nextpow2(np.round(windowSize * self.fs))
self.stftParams['hopsize'] = self.stftParams['windowSizeInSamples']/8
self.stftParams['NFT'] = self.stftParams['windowSizeInSamples']
self.SIMMParams['niter'] = nbIter
self.SIMMParams['R'] = numCompAccomp
print "Some parameter settings:"
print " Size of analysis windows: ", \
self.stftParams['windowSizeInSamples']
print " Hopsize: ", self.stftParams['hopsize']
print " Size of Fourier transforms: ", self.stftParams['NFT']
print " Number of iterations to be done: ",self.SIMMParams['niter']
print " Number of elements in WM: ", self.SIMMParams['R']
self.XR, F, N = slf.stft(data[:,0], fs=self.fs,
hopsize=self.stftParams['hopsize'] ,
window=slf.sinebell(\
self.stftParams['windowSizeInSamples']),
nfft=self.stftParams['NFT'] )
self.XL, F, N = slf.stft(data[:,1], fs=self.fs,
hopsize=self.stftParams['hopsize'] ,
window=slf.sinebell(\
self.stftParams['windowSizeInSamples']),
nfft=self.stftParams['NFT'] )
# non need to store this.
## self.SXR = np.abs(self.XR) ** 2
## self.SXL = np.abs(self.XL) ** 2
# drawing the waveform to wavCanvas:
if not(self.wavCanvas is None):
if self.wavCanvas==self.imageCanvas:
self.wavCanvas.ax2.clear()
self.wavCanvas.ax2.plot(np.arange(data.shape[0]) / \
np.double(self.stftParams['hopsize']), \
data)
#self.wavCanvas.ax2.plot(np.arange(data.shape[0]) / \
# np.double(self.fs), \
# data)
self.wavCanvas.ax2.axis('tight')
self.wavCanvas.draw()
else:
self.wavCanvas.ax.clear()
self.wavCanvas.ax.plot(np.arange(data.shape[0]) / \
np.double(self.fs), \
data)
self.wavCanvas.ax.axis('tight')
self.wavCanvas.draw()
del data, F, N
# TODO: also process these as options:
self.SIMMParams['minF0'] = minF0
self.SIMMParams['maxF0'] = maxF0
self.F, self.N = self.XR.shape
# this is the number of F0s within one semitone
self.SIMMParams['stepNotes'] = stepNotes
# number of spectral shapes for the filter part
self.SIMMParams['K'] = K_numFilters
# number of elements in dictionary of smooth filters
self.SIMMParams['P'] = P_numAtomFilters
# number of chirped spectral shapes between each F0
# this feature should be further studied before
# we find a good way of doing that.
self.SIMMParams['chirpPerF0'] = 1
self.scopeAllowedHF0 = 4.0 / 1.0
# Create the harmonic combs, for each F0 between minF0 and maxF0:
self.SIMMParams['F0Table'], WF0 = \
slf.generate_WF0_chirped(minF0=self.SIMMParams['minF0'],
maxF0=self.SIMMParams['maxF0'],
Fs=self.fs,
Nfft=self.stftParams['NFT'],\
stepNotes=self.SIMMParams['stepNotes'],\
lengthWindow=\
self.stftParams['windowSizeInSamples'],
Ot=0.25,\
perF0=self.SIMMParams['chirpPerF0'],\
depthChirpInSemiTone=.15,
loadWF0=True,\
analysisWindow='sinebell')
self.SIMMParams['WF0'] = WF0[:self.F, :] # ensure same size as SX
# number of harmonic combs
self.SIMMParams['NF0'] = self.SIMMParams['F0Table'].size
# Normalization:
# by max or by sum?
self.SIMMParams['WF0'] = self.SIMMParams['WF0'] / \
np.outer(np.ones(self.F), \
np.sum(self.SIMMParams['WF0'],
axis=0))
# for debug:
if False: #DEBUG
self.imageCanvas.ax.imshow(np.log(np.abs(self.XR)),
aspect='auto',origin='lower')
self.imageCanvas.draw()
raise KeyboardInterrupt("Check these matrices !")
if False: #DEBUG
from IPython.Shell import IPShellEmbed
ipshell = IPShellEmbed()
ipshell()
plt.figure()
plt.imshow(np.log(self.SIMMParams['WF0']),
aspect='auto',
origin='lower',)
plt.figure()
plt.imshow(np.log(np.abs(self.XR)),aspect='auto',origin='lower')
# Create the dictionary of smooth filters, for the filter part of
# the lead isntrument:
self.SIMMParams['WGAMMA'] = \
slf.generateHannBasis(numberFrequencyBins=self.F,
sizeOfFourier=self.stftParams['NFT'],
Fs=self.fs,
frequencyScale='linear',
numberOfBasis=self.SIMMParams['P'],
overlap=.75)
def __init__(self,
inputAudioFilename,
windowSize=0.0464,
nbIter=10,
numCompAccomp=40,
minF0=60,
maxF0=2000,
stepNotes=16,
K_numFilters=4,
P_numAtomFilters=30,
imageCanvas=None,
wavCanvas=None,
progressBar=None,
verbose=True,
outputDirSuffix='/'):
"""During init, process is initiated, STFTs are computed,
and the parameters are stored.
Parameters
----------
inputAudioFilename : string
filename of the input audio file
windowSize : double, optional
analysis frame ('windows') size, in s. By default, 0.0464s
nbIter : integer, optional
number of iterations for the estimation algorithm. By default, 10
numCompAccomp : integer, optional
number of components for the accompaniment, default = 40
minF0 : double/integer, optional
lowest F0 candidate (in Hz), default=60Hz
maxF0 : double/integer, optional
highest F0 candidate (in Hz), default=2000Hz
stepNotes : integer, optional
number of F0 candidates in one semitone, default=16 F0s/semitone
K_numFilters : integer, optional
number of filter spectral shapes, default=4
P_numAtomFilters : integer, optional
number of atomic filter smooth spectral shapes, default=30
imageCanvas : MplCanvas/MplCanvas3Axes, optional
an instance of the MplCanvas/MplCanvas3Axes, giving access to the
axes where to draw the HF0 image. By default=None
wavCanvas : MplCanvas/MplCanvas3Axes, optional
an instance of the MplCanvas/MplCanvas3Axes, giving access to the
axes to draw the waveform of the input signal.
progressBar : boolean, optional ???
???
verbose : boolean, optional
Whether to write out or not information about the evolution of the
algorithm. By default=False.
outputDirSuffix : string, optional
the subfolder name (to be appended to the full path to the audio
signal), where the output files are going to be written. By default
='/'
"""
self.files['inputAudioFilename'] = str(inputAudioFilename)
self.imageCanvas = imageCanvas
self.wavCanvas = wavCanvas
self.displayEvolution = True
self.verbose = verbose
if self.imageCanvas is None:
self.displayEvolution = False
if inputAudioFilename[-4:] != ".wav":
raise ValueError("File not WAV file? Only WAV format support, "+\
"for now...")
self.files['outputDirSuffix'] = outputDirSuffix
self.files['outputDir'] = str('/').join(\
self.files['inputAudioFilename'].split('/')[:-1])+\
'/'+self.files['outputDirSuffix'] +'/'
if os.path.isdir(self.files['outputDir']):
print "Output directory already existing - "+\
"NB: overwriting files in:"
print self.files['outputDir']
else:
print "Creating output directory"
print self.files['outputDir']
os.mkdir(self.files['outputDir'])
self.files['pathBaseName'] = self.files['outputDir'] + \
self.files['inputAudioFilename'\
].split('/')[-1][:-4]
self.files['mus_output_file'] = str(self.files['pathBaseName']+\
'_acc.wav')
self.files['voc_output_file'] = str(self.files['pathBaseName']+\
'_lead.wav')
self.files['pitch_output_file'] = str(self.files['pathBaseName']+\
'_pitches.txt')
print "Writing the different following output files:"
print " separated lead in", \
self.files['voc_output_file']
print " separated accompaniment in", \
self.files['mus_output_file']
print " separated lead + unvoc in", \
self.files['voc_output_file'][:-4] + '_VUIMM.wav'
print " separated acc - unvoc in", \
self.files['mus_output_file'][:-4] + '_VUIMM.wav'
print " estimated pitches in", \
self.files['pitch_output_file']
# read the WAV file and store the STFT
self.fs, data = wav.read(self.files['inputAudioFilename'])
# for some bad format wav files, data is a str?
# cf. files from beat/tempo evaluation campaign of MIREX
## print self.fs, data
self.scaleData = 1.2 * np.abs(data).max() # to rescale the data.
self.dataType = data.dtype
data = np.double(data) / self.scaleData # makes data vary from -1 to 1
if data.shape[0] == data.size: # data is multi-channel
print "The audio file is not stereo. Making stereo out of mono."
print "(You could also try the older separateLead.py...)"
data = np.vstack([data, data]).T
if data.shape[1] != 2:
print "The data is multichannel, but not stereo... \n"
print "Unfortunately this program does not scale well. Data is \n"
print "reduced to its 2 first channels.\n"
data = data[:, 0:2]
# parameters for the STFT:
self.stftParams['windowSizeInSamples'] = \
slf.nextpow2(np.round(windowSize * self.fs))
self.stftParams['hopsize'] = self.stftParams['windowSizeInSamples'] / 8
self.stftParams['NFT'] = self.stftParams['windowSizeInSamples']
self.SIMMParams['niter'] = nbIter
self.SIMMParams['R'] = numCompAccomp
print "Some parameter settings:"
print " Size of analysis windows: ", \
self.stftParams['windowSizeInSamples']
print " Hopsize: ", self.stftParams['hopsize']
print " Size of Fourier transforms: ", self.stftParams['NFT']
print " Number of iterations to be done: ", self.SIMMParams['niter']
print " Number of elements in WM: ", self.SIMMParams['R']
self.XR, F, N = slf.stft(data[:,0], fs=self.fs,
hopsize=self.stftParams['hopsize'] ,
window=slf.sinebell(\
self.stftParams['windowSizeInSamples']),
nfft=self.stftParams['NFT'] )
self.XL, F, N = slf.stft(data[:,1], fs=self.fs,
hopsize=self.stftParams['hopsize'] ,
window=slf.sinebell(\
self.stftParams['windowSizeInSamples']),
nfft=self.stftParams['NFT'] )
# non need to store this.
## self.SXR = np.abs(self.XR) ** 2
## self.SXL = np.abs(self.XL) ** 2
# drawing the waveform to wavCanvas:
if not (self.wavCanvas is None):
if self.wavCanvas == self.imageCanvas:
self.wavCanvas.ax2.clear()
self.wavCanvas.ax2.plot(np.arange(data.shape[0]) / \
np.double(self.stftParams['hopsize']), \
data)
#self.wavCanvas.ax2.plot(np.arange(data.shape[0]) / \
# np.double(self.fs), \
# data)
self.wavCanvas.ax2.axis('tight')
self.wavCanvas.draw()
else:
self.wavCanvas.ax.clear()
self.wavCanvas.ax.plot(np.arange(data.shape[0]) / \
np.double(self.fs), \
data)
self.wavCanvas.ax.axis('tight')
self.wavCanvas.draw()
del data, F, N
# TODO: also process these as options:
self.SIMMParams['minF0'] = minF0
self.SIMMParams['maxF0'] = maxF0
self.F, self.N = self.XR.shape
# this is the number of F0s within one semitone
self.SIMMParams['stepNotes'] = stepNotes
# number of spectral shapes for the filter part
self.SIMMParams['K'] = K_numFilters
# number of elements in dictionary of smooth filters
self.SIMMParams['P'] = P_numAtomFilters
# number of chirped spectral shapes between each F0
# this feature should be further studied before
# we find a good way of doing that.
self.SIMMParams['chirpPerF0'] = 1
self.scopeAllowedHF0 = 4.0 / 1.0
# Create the harmonic combs, for each F0 between minF0 and maxF0:
self.SIMMParams['F0Table'], WF0 = \
slf.generate_WF0_chirped(minF0=self.SIMMParams['minF0'],
maxF0=self.SIMMParams['maxF0'],
Fs=self.fs,
Nfft=self.stftParams['NFT'],\
stepNotes=self.SIMMParams['stepNotes'],\
lengthWindow=\
self.stftParams['windowSizeInSamples'],
Ot=0.25,\
perF0=self.SIMMParams['chirpPerF0'],\
depthChirpInSemiTone=.15,
loadWF0=True,\
analysisWindow='sinebell')
self.SIMMParams['WF0'] = WF0[:self.F, :] # ensure same size as SX
# number of harmonic combs
self.SIMMParams['NF0'] = self.SIMMParams['F0Table'].size
# Normalization:
# by max or by sum?
self.SIMMParams['WF0'] = self.SIMMParams['WF0'] / \
np.outer(np.ones(self.F), \
np.sum(self.SIMMParams['WF0'],
axis=0))
# for debug:
if False: #DEBUG
self.imageCanvas.ax.imshow(np.log(np.abs(self.XR)),
aspect='auto',
origin='lower')
self.imageCanvas.draw()
raise KeyboardInterrupt("Check these matrices !")
if False: #DEBUG
from IPython.Shell import IPShellEmbed
ipshell = IPShellEmbed()
ipshell()
plt.figure()
plt.imshow(
np.log(self.SIMMParams['WF0']),
aspect='auto',
origin='lower',
)
plt.figure()
plt.imshow(np.log(np.abs(self.XR)), aspect='auto', origin='lower')
# Create the dictionary of smooth filters, for the filter part of
# the lead isntrument:
self.SIMMParams['WGAMMA'] = \
slf.generateHannBasis(numberFrequencyBins=self.F,
sizeOfFourier=self.stftParams['NFT'],
Fs=self.fs,
frequencyScale='linear',
numberOfBasis=self.SIMMParams['P'],
overlap=.75)
