def generateData(inFile, outFile, vuvInFile, vuvOutFile, featDim, resolution, thre=0.5):
if not os.path.isfile(inFile):
print inFile
return
if not os.path.isfile(vuvInFile):
print vuvInFile
return
labData = py_rw.read_raw_mat(inFile, featDim)
vuvData = py_rw.read_raw_mat(vuvInFile, 1)
if labData.shape[0] > vuvData.shape[0]:
labData = labData[0:vuvData.shape[0],:]
else:
vuvData = vuvData[0:labData.shape[0]]
assert labData.shape[0] == vuvData.shape[0], 'Unequal length vuv and lab'
maxTime = labData.shape[0] * resolution
labIdxBuf = np.zeros([int(maxTime)])
vuvBinBuf = np.zeros([int(maxTime)])
for idx in np.arange(labData.shape[0]):
st = idx * resolution
et = (idx + 1) * resolution
labIdxBuf[st:et] = idx
vuvBinBuf[st:et] = vuvData[idx]
randU = np.random.rand(et-st)
temp = vuvBinBuf[st:et]
temp[randU < thre] = 0
vuvBinBuf[st:et] = temp
py_rw.write_raw_mat(labIdxBuf, outFile)
py_rw.write_raw_mat(vuvBinBuf, vuvOutFile)
def generate(inFile1, inFile2, outfile):
data1 = np.asarray(py_rw.read_raw_mat(inFile1, 1), dtype=np.int32)
data2 = py_rw.read_raw_mat(inFile2, 1)
temp, _ = discreteF0.f0Conversion(data2.copy(), F0Max, F0Min, F0Inter,
'c2d', F0Conti)
data3 = np.zeros(data1.shape)
data3[data2[data1] > 0] = 1
py_rw.write_raw_mat(data3, outfile)
def tempWarpper(fileIn, outFile, outWav):
sr, wavdata = scipy.io.wavfile.read(fileIn)
if wavdata.dtype == np.int16:
#print outFile
transData = np.array(wavdata, dtype=np.float32) / np.power(2.0, 16-1)
py_rw.write_raw_mat(transData, outFile)
else:
print "Unsupported data type"
def tempWarpper(fileIn, outFile, outWav):
sr, wavdata = scipy.io.wavfile.read(fileIn)
if wavdata.dtype == np.int16:
#print outFile
transData = np.array(wavdata, dtype=np.float32) / np.power(2.0, 16 - 1)
py_rw.write_raw_mat(transData, outFile)
else:
print("Unsupported data type")
def f0ip2f0(f0_ip_file, vuv_file, output_file, vu_threshold=0.5, unvoiced_value=-1.0e+10):
#print("%s %s %s %f" % (f0_ip_file, vuv_file, output_file, vu_threshold))
f0_ip_data = py_rw.read_raw_lf0(f0_ip_file, 1)
vuv_data = py_rw.read_raw_mat(vuv_file, 1)
assert f0_ip_data.shape[0]==vuv_data.shape[0], "Error: unequal length %s %s" % (f0_ip_file, vuv_file)
#print(str(np.where(vuv_data<vu_threshold)))
f0_ip_data[vuv_data<vu_threshold] = unvoiced_value
py_rw.write_raw_mat(f0_ip_data, output_file)
def generateLabIndex(labfile, outfile, resolution = 80, res=50000):
sTime, eTime, labEntry=htslab.read_full_lab(labfile, res)
maxTime = eTime[-1] * resolution
outBuf = np.zeros([int(maxTime)])
for idx in np.arange(int(eTime[-1])):
st = idx * resolution
et = (idx + 1) * resolution
outBuf[st:et] = idx
py_rw.write_raw_mat(outBuf, outfile)
def generateLabIndex(labfile, outfile, resolution=80, res=50000):
sTime, eTime, labEntry = htslab.read_full_lab(labfile, res)
maxTime = eTime[-1] * resolution
outBuf = np.zeros([int(maxTime)])
for idx in np.arange(int(eTime[-1])):
st = idx * resolution
et = (idx + 1) * resolution
outBuf[st:et] = idx
py_rw.write_raw_mat(outBuf, outfile)
def CreateTimeMatrix(DataFile):
"""
"""
phofile = PhoAlign + os.path.sep + DataFile + '.lab'
sylfile = SylAlign + os.path.sep + DataFile + '.lab'
worfile = WorAlign + os.path.sep + DataFile + '.lab'
phodata = htslab.read_full_lab(phofile)
syldata = htslab.read_full_lab(sylfile)
wordata = htslab.read_full_lab(worfile)
if len(wordata[1]) != len(syldata[1]) or wordata[1][-1] != syldata[1][-1]:
print("\t Unequal Length %s" % (DataFile))
return 0
DataTime = np.int(syldata[1][-1])
# default, update at every frame level
dataMat = np.bitwise_or(np.zeros([DataTime], dtype=np.int8),
bitInfo['fra'])
preSyl = ''
preWor = ''
for idx1 in range(len(syldata[0])):
frameStart = syldata[0][idx1]
frameEnd = syldata[1][idx1]
# update the phoneme state
dataMat[frameStart] = np.bitwise_or(dataMat[frameStart],
bitInfo['pho'])
syllabel = syldata[2][idx1]
worlabel = wordata[2][idx1]
if syllabel != preSyl:
dataMat[frameStart] = np.bitwise_or(dataMat[frameStart],
bitInfo['syl'])
if worlabel != preWor:
dataMat[frameStart] = np.bitwise_or(dataMat[frameStart],
bitInfo['wor'])
if len(preWor) == 0 or preWor == phraseSym or worlabel == phraseSym:
dataMat[frameStart] = np.bitwise_or(dataMat[frameStart],
bitInfo['phr'])
preSyl = syllabel
preWor = worlabel
if CheckBinary:
pholabel = phodata[2][idx1]
for t in range(np.int(frameStart), np.int(frameEnd)):
print("%d, %s [%s %s %s]" %
(t, np.binary_repr(dataMat[t], len(bitInfo)),
pholabel[0:6], syllabel[0:6], worlabel[0:6]))
py_rw.write_raw_mat(dataMat, DataDir + os.path.sep + DataFile + '.bin',
'u1')
return DataTime
def generateLabIndex(labfile, outfile, labDim, resolution):
data = py_rw.read_raw_mat(labfile, labDim)
maxTime = data.shape[0] * resolution
outBuf = np.zeros([int(maxTime)])
for idx in np.arange(int(data.shape[0])):
st = idx * resolution
et = (idx + 1) * resolution
outBuf[st:et] = idx
py_rw.write_raw_mat(outBuf, outfile)
def generateLabIndex(labfile, outfile, labDim, resolution):
data = py_rw.read_raw_mat(labfile, labDim)
maxTime = data.shape[0] * resolution
outBuf = np.zeros([int(maxTime)])
for idx in np.arange(int(data.shape[0])):
st = idx * resolution
et = (idx + 1) * resolution
outBuf[st:et] = idx
py_rw.write_raw_mat(outBuf, outfile)
def split(fileName, inDir, uvDir, uvT, f0Ext, vuExt):
conlf0Name = inDir + os.path.sep + fileName + f0Ext
vuName = uvDir + os.path.sep + fileName + vuExt
print(fileName, end=' ')
if os.path.isfile(conlf0Name) and os.path.isfile(vuName):
conlf0 = py_rw.read_raw_mat(conlf0Name, 1)
vu = py_rw.read_raw_mat(vuName, 1)
assert conlf0.shape[0] == vu.shape[0], ': lf0 uv unequal length'
conlf0[vu < uvT] = -1.0e+10
py_rw.write_raw_mat(conlf0, conlf0Name)
print(': done')
else:
print(': not found')
def split(fileName, inDir, uvDir, uvT, f0Ext, vuExt):
conlf0Name = inDir + os.path.sep + fileName + f0Ext
vuName = uvDir + os.path.sep + fileName + vuExt
print fileName,
if os.path.isfile(conlf0Name) and os.path.isfile(vuName):
conlf0 = py_rw.read_raw_mat(conlf0Name, 1)
vu = py_rw.read_raw_mat(vuName, 1)
assert conlf0.shape[0] == vu.shape[0], ': lf0 uv unequal length'
conlf0[vu < uvT] = -1.0e+10
py_rw.write_raw_mat(conlf0, conlf0Name)
print ': done'
else:
print ': not found'
def f0convert(f0File, qF0Output, vuvOutputFile, f0Zero, f0Max, f0Min, f0Inter, f0Conti, f0Type):
if f0Type == 0:
data = py_rw.read_raw_mat(f0File, 1)
idx = data > 0
data[idx] = 1127.0 * np.log(data[idx]/700.0 + 1)
elif f0Type == 1:
data = py_rw.read_raw_lf0(f0File, 1)
idx = data > 0
F0Idx = data>f0Zero
dataClas, vuv = discreteF0.f0Conversion(data.copy(), f0Max, f0Min, f0Inter, 'c2d', f0Conti)
dataClas[vuv<1] = 0.0
py_rw.write_raw_mat(dataClas, qF0Output)
py_rw.write_raw_mat(vuv, vuvOutputFile)
def tempWarpper(fileIn, outFile, outWav, quanLevel):
sr, wavdata = scipy.io.wavfile.read(inFile)
if wavdata.dtype == np.int16:
transData = wavTool.wavformConvert(wavdata, 16, True, quanLevel)
elif wavdata.dtype == np.int32:
transData = wavTool.wavformConvert(wavdata, 32, True, quanLevel)
else:
print "Unsupported data type"
py_rw.write_raw_mat(transData, outFile)
if outWav is not None:
recoData = wavTool.wavformDeconvert(transData, quanLevel)
wavTool.waveSaveFromFloat(recoData, outWav, sr=sr)
def tempWarpper(fileIn, outFile, outWav, quanLevel):
sr, wavdata = scipy.io.wavfile.read(inFile)
if wavdata.dtype == np.int16:
transData = wavTool.wavformConvert(wavdata, 16, True, quanLevel)
elif wavdata.dtype == np.int32:
transData = wavTool.wavformConvert(wavdata, 32, True, quanLevel)
else:
print "Unsupported data type"
py_rw.write_raw_mat(transData, outFile)
if outWav is not None:
recoData = wavTool.wavformDeconvert(transData, quanLevel)
wavTool.waveSaveFromFloat(recoData, outWav, sr=sr)
def createMdnConfig(mdnConfigFile,
MDNType,
MDNTargetDim,
ARDynamic=None,
tieVariance=0):
"""Create the mdn.config for MDN CURRENNT
"""
if ARDynamic is None:
# default, no AR dynamic
ARDynamic = np.ones([len(MDNType)]) * -1.0
bias = 0
MDNNNOutDim = []
for idx, mdnConfig in enumerate(MDNType):
temp = distParaNum(MDNTargetDim[idx][1] - MDNTargetDim[idx][0],
mdnConfig, tieVariance, ARDynamic[idx])
MDNNNOutDim.append([bias, bias + temp])
bias = temp + bias
#print MDNNNOutDim
# check and generating the MDN configuration
mdnconfigdata = np.zeros([1 + len(MDNType) * 5], dtype=np.float32)
mdnconfigdata[0] = len(MDNType)
tmp = 0
for idx, mdnConfig in enumerate(MDNType):
mdntarDim = MDNTargetDim[idx]
mdnoutDim = MDNNNOutDim[idx]
tmp1 = distParaNum(mdntarDim[1] - mdntarDim[0], mdnConfig, tieVariance,
ARDynamic[idx])
tmp2 = (mdnoutDim[1] - mdnoutDim[0])
if mdnConfig > 0:
assert tmp1 == tmp2, "Error in MDN mixture configuraiton"
tmp = tmp + tmp2
elif mdnConfig < 0:
assert mdntarDim[1] - mdntarDim[
0] == 1, "Softmax to 1 dimension targert"
#tmp = tmp + 1
tmp = tmp + tmp2
mdnConfig = -1 # change it back to -1
else:
tmp = tmp + tmp2
mdnconfigdata[(idx * 5 + 1):((idx + 1) * 5 + 1)] = [
mdnoutDim[0], mdnoutDim[1], mdntarDim[0], mdntarDim[1], mdnConfig
]
#print "Dimension of output of NN should be %d" % (tmp)
py_rw.write_raw_mat(mdnconfigdata, mdnConfigFile)
return tmp
def data_normalize(dataFile, dim, mask=None, ran=False):
data = drh.read_raw_mat(dataFile, dim)
if ran==True:
ranfile = os.path.dirname(dataFile)+os.path.sep+'random_index'
if os.path.isfile(ranfile):
print "Found random index %s" % (ranfile)
randidx = np.asarray(drh.read_raw_mat(ranfile,1), dtype=np.int32)
if randidx.shape[0]!=data.shape[0]:
print "But it unmatches the data. New random_index will be generated"
randidx = np.array(range(data.shape[0]))
random.shuffle(randidx)
drh.write_raw_mat(randidx, ranfile)
else:
pass
else:
randidx = np.array(range(data.shape[0]))
random.shuffle(randidx)
drh.write_raw_mat(randidx, ranfile)
data = data[randidx,:]
meanData = data.mean(axis=0)
stdData = data.std(axis=0)
if mask is None:
pass
else:
stdData[mask>0] = 0.0 # set to zero
idx = stdData>0.000001 # a threshold to normalize
data[:,idx] = (data[:,idx]-meanData[idx])/stdData[idx]
drh.write_raw_mat(data,dataFile+'.norm')
drh.write_raw_mat(np.concatenate((meanData, stdData)), dataFile+'.mv')
def CreateTimeMatrix(DataFile):
"""
"""
phofile = PhoAlign + os.path.sep + DataFile + '.lab'
sylfile = SylAlign + os.path.sep + DataFile + '.lab'
worfile = WorAlign + os.path.sep + DataFile + '.lab'
phodata = htslab.read_full_lab(phofile)
syldata = htslab.read_full_lab(sylfile)
wordata = htslab.read_full_lab(worfile)
if len(wordata[1]) != len(syldata[1]) or wordata[1][-1] != syldata[1][-1]:
print "\t Unequal Length %s" % (DataFile)
return 0
DataTime = np.int(syldata[1][-1])
# default, update at every frame level
dataMat = np.bitwise_or(np.zeros([DataTime], dtype=np.int8), bitInfo['fra'])
preSyl = ''
preWor = ''
for idx1 in xrange(len(syldata[0])):
frameStart = syldata[0][idx1]
frameEnd = syldata[1][idx1]
# update the phoneme state
dataMat[frameStart] = np.bitwise_or(dataMat[frameStart], bitInfo['pho'])
syllabel = syldata[2][idx1]
worlabel = wordata[2][idx1]
if syllabel != preSyl:
dataMat[frameStart] = np.bitwise_or(dataMat[frameStart], bitInfo['syl'])
if worlabel != preWor:
dataMat[frameStart] = np.bitwise_or(dataMat[frameStart], bitInfo['wor'])
if len(preWor)==0 or preWor == phraseSym or worlabel == phraseSym:
dataMat[frameStart] = np.bitwise_or(dataMat[frameStart], bitInfo['phr'])
preSyl = syllabel
preWor = worlabel
if CheckBinary:
pholabel = phodata[2][idx1]
for t in range(np.int(frameStart), np.int(frameEnd)):
print "%d, %s [%s %s %s]" % (t,np.binary_repr(dataMat[t], len(bitInfo)),
pholabel[0:6], syllabel[0:6], worlabel[0:6])
py_rw.write_raw_mat(dataMat, DataDir+os.path.sep+DataFile+'.bin', 'u1')
return DataTime
def meanStdNormMask(fileScps,
fileDims,
fileNormMask,
meanStdOutPath,
f0Dim=-1):
"""
"""
assert len(fileDims) == len(fileNormMask), \
"Unequal length feature dim & norm mask"
# calcualte the mean/std
stats.getMeanStd_merge(fileScps, fileDims, meanStdOutPath + '.unmasked',
f0Dim)
meanStdData = py_rw.read_raw_mat(meanStdOutPath + '.unmasked', 1)
assert meanStdData.shape[0] == sum(fileDims) * 2, \
"%s dimension not %d" % (meanStdOutPath + '.unmasked', sum(fileDims) * 2)
featDims = []
startDim = 0
for dim in fileDims:
featDims.append([startDim, startDim + dim])
startDim = startDim + dim
for dimRange, normMask in zip(featDims, fileNormMask):
if len(normMask) == 0:
pass
elif len(normMask) == 1 and (normMask[0] == 0
or normMask[0] == 'not_norm'):
meanStdData[dimRange[0]:dimRange[1]] = 0.0
meanStdData[dimRange[0] + sum(fileDims):dimRange[1] +
sum(fileDims)] = 1.0
elif len(normMask) == 2:
assert dimRange[0] <= normMask[0], 'normMask range error' % (
str(normMask))
assert dimRange[1] >= normMask[1], 'normMask range error' % (
str(normMask))
meanStdData[normMask[0]:normMask[1]] = 0.0
meanStdData[normMask[0] + sum(fileDims):normMask[1] +
sum(fileDims)] = 1.0
else:
print("Wrong format of NormMask %s" % (str(normMask)))
print('normmask %s' % (str(normMask)))
py_rw.write_raw_mat(meanStdData, meanStdOutPath)
def createMdnConfig(mdnConfigFile, MDNType, MDNTargetDim, ARDynamic=None, tieVariance=0):
"""Create the mdn.config for MDN CURRENNT
"""
if ARDynamic is None:
# default, no AR dynamic
ARDynamic = np.ones([len(MDNType)]) * -1.0
bias = 0
MDNNNOutDim = []
for idx, mdnConfig in enumerate(MDNType):
temp = distParaNum(MDNTargetDim[idx][1] - MDNTargetDim[idx][0], mdnConfig,
tieVariance, ARDynamic[idx])
MDNNNOutDim.append([bias, bias+temp])
bias = temp+bias
#print MDNNNOutDim
# check and generating the MDN configuration
mdnconfigdata = np.zeros([1+len(MDNType)*5], dtype = np.float32)
mdnconfigdata[0] = len(MDNType)
tmp = 0
for idx, mdnConfig in enumerate(MDNType):
mdntarDim = MDNTargetDim[idx]
mdnoutDim = MDNNNOutDim[idx]
tmp1 = distParaNum(mdntarDim[1]-mdntarDim[0], mdnConfig, tieVariance, ARDynamic[idx])
tmp2 = (mdnoutDim[1]-mdnoutDim[0])
if mdnConfig > 0:
assert tmp1 == tmp2, "Error in MDN mixture configuraiton"
tmp = tmp + tmp2
elif mdnConfig < 0:
assert mdntarDim[1]-mdntarDim[0]==1, "Softmax to 1 dimension targert"
tmp = tmp + 1
mdnConfig = -1 # change it back to -1
else:
tmp = tmp + tmp2
mdnconfigdata[(idx*5+1):((idx+1)*5+1)] = [mdnoutDim[0],mdnoutDim[1],
mdntarDim[0],mdntarDim[1],
mdnConfig]
#print "Dimension of output of NN should be %d" % (tmp)
py_rw.write_raw_mat(mdnconfigdata, mdnConfigFile)
return tmp
def meanStdNormMask(fileScps, fileDims, fileNormMask, meanStdOutPath, f0Dim=-1):
"""
"""
assert len(fileDims) == len(fileNormMask), \
"Unequal length feature dim & norm mask"
# calcualte the mean/std
stats.getMeanStd_merge(fileScps, fileDims, meanStdOutPath + '.unmasked', f0Dim)
meanStdData = py_rw.read_raw_mat(meanStdOutPath + '.unmasked', 1)
assert meanStdData.shape[0] == sum(fileDims) * 2, \
"%s dimension not %d" % (meanStdOutPath + '.unmasked', sum(fileDims) * 2)
featDims = []
startDim = 0
for dim in fileDims:
featDims.append([startDim, startDim + dim])
startDim = startDim + dim
for dimRange, normMask in zip(featDims, fileNormMask):
if len(normMask) == 0:
pass
elif len(normMask) == 1 and normMask[0] == 0:
meanStdData[dimRange[0]:dimRange[1]] = 0.0
meanStdData[dimRange[0]+sum(fileDims):dimRange[1]+sum(fileDims)] = 1.0
elif len(normMask) == 2:
assert dimRange[0] <= normMask[0], 'normMask range error' % (str(normMask))
assert dimRange[1] >= normMask[1], 'normMask range error' % (str(normMask))
meanStdData[normMask[0]:normMask[1]] = 0.0
meanStdData[normMask[0]+sum(fileDims):normMask[1]+sum(fileDims)] = 1.0
else:
print "Wrong format of NormMask %s" % (str(normMask))
print 'normmask %s' % (str(normMask))
py_rw.write_raw_mat(meanStdData, meanStdOutPath)
#!/usr/bin/python
from speechTools import wavTool
from scipy.io import wavfile
from ioTools import readwrite as py_rw
import os
import sys
import numpy as np
dirPath = sys.argv[1]
quantiBitNum = int(sys.argv[2])
samplingRate = int(sys.argv[3])
fileList = py_rw.read_txt_list(dirPath + '/gen.scp')
for fileName in fileList:
fileName = fileName.rstrip('\n')
nameHtk = dirPath + os.path.sep + os.path.basename(fileName).rstrip('.htk') + '.htk'
nameRaw = dirPath + os.path.sep + os.path.basename(fileName).rstrip('.htk') + '.raw'
nameWav = dirPath + os.path.sep + os.path.basename(fileName).rstrip('.htk') + '.wav'
print nameRaw, nameWav
data = py_rw.read_htk(nameHtk, 'f4', 'b')
if quantiBitNum > 0:
quantiLevel = np.power(2, quantiBitNum)-1
py_rw.write_raw_mat(data, nameRaw)
wavTool.raw2wav(nameRaw, nameWav, quantiLevel, samplingRate=samplingRate)
else:
wavfile.write(nameWav, samplingRate, data)
from ioTools import readwrite
import kaldi_io
import numpy as np
args = sys.argv
data_dir = args[1]
out_dir = args[2]
dataname = basename(data_dir)
xvector_file = "exp/0007_voxceleb_v2_1a/exp/xvector_nnet_1a/am_nsf/xvectors_" + dataname + "/xvector.scp"
xvec_out_dir = join(out_dir, "xvector")
pitch_out_dir = join(out_dir, "f0")
# Write pitch features
pitch_file = join(data_dir, 'pitch.scp')
pitch2shape = {}
for key, mat in kaldi_io.read_mat_scp(pitch_file):
#pitch2shape[key] = mat.shape
readwrite.write_raw_mat(mat[:, 1], join(pitch_out_dir, key + '.f0'))
'''
# Write xvector features
with open(xvector_file) as f:
for key, mat in kaldi_io.read_vec_flt_scp(f):
#print key, mat.shape
plen = pitch2shape[key][0]
mat = mat[np.newaxis]
xvec = np.repeat(mat, plen, axis=0)
readwrite.write_raw_mat(xvec, join(xvec_out_dir, key+'.xvector'))
'''
# std threshold
# if std < threshold, set std = 1.0 (not use std for normalization)
stdThresh = 0.0000001
# If mean and std are calculated over multiple types of features,
# please specify fileDirs, fileExts, dataDims for each type of feature.
# The output contains a mean/std vector concatenated from all types of features
dataMvBuffer = []
for idx, fileDir in enumerate(fileDirs):
fileExt = fileExts[idx]
dataDim = dataDims[idx]
dataMv = np.zeros([2, dataDim], dtype=np.float64)
dataCounter = 0
with open(fileList, 'r') as filePtr:
for fileName in filePtr:
fileName = fileName.rstrip('\n')
print(fileName)
fileName = fileDir + '/' + fileName + fileExt
dataMv, dataCounter = temprapper(fileName, dataDim, dataMv,
dataCounter)
dataMv[1] = np.sqrt(dataMv[1] / (dataCounter - 1))
dataMv = np.asarray(dataMv, dtype=np.float32)
dataMvBuffer.append(dataMv)
dataMvBuffer = np.concatenate(dataMvBuffer, axis=1)
dataMvBuffer[1, np.where(dataMvBuffer[1, :] < stdThresh)] = 1.0
py_rw.write_raw_mat(dataMvBuffer, dataOut)
def lf02f0_file(data_file, out_file, thres=10.0):
data = py_rw.read_raw_lf0(data_file, 1)
data[data>thres] = lf02f0(data[data>thres])
data[data<=thres] = 0.0
py_rw.write_raw_mat(data, out_file)
frameSum = frameSum + data.shape[0]
F0Idx = data>10
dataClas, vuv = discreteF0.f0Conversion(data, F0Max, F0Min, F0Inter, 'c2d', F0Conti)
dataClas[vuv<1] = 0.0
#dataClas = np.zeros([data.shape[0]])
#if F0Conti:
# # Continous F0
# pass
# dataClas[F0Idx] = np.round((data[F0Idx] - F0Min)/(F0Max - F0Min) * (F0Inter - 1))
#else:
# # Discontinuous F0, leave one dimension for unvoiced
# dataClas[F0Idx] = np.round((data[F0Idx] - F0Min)/(F0Max - F0Min) * (F0Inter - 2)) + 1
tmpmax = np.max(data[F0Idx])
tmpmin = np.min(data[F0Idx])
tmpmax2 = np.max(dataClas[F0Idx])
tmpmin2 = np.min(dataClas[F0Idx])
print tmpmax, tmpmin, tmpmax2, tmpmin2
frameMax = np.max([frameMax, tmpmax])
frameMin = np.min([frameMin, tmpmin])
filePath = fileOut + os.path.sep + fileName + fileExt + '_class'
py_rw.write_raw_mat(dataClas, filePath)
print "\nmax F0 event, min F0 event, #frame"
print frameMax, frameMin, frameSum
fileDir = sys.argv[2]
fileList = sys.argv[3]
fileExt = sys.argv[4]
fileDim = int(sys.argv[5])
fileOut = sys.argv[6]
cnt = 0
with open(fileList, 'r') as filePtr:
for idx, fileName in enumerate(filePtr):
cnt = cnt + 1
gvData = np.zeros([cnt, fileDim])
cnt = 0
with open(fileList, 'r') as filePtr:
for idx, fileName in enumerate(filePtr):
fileName = fileName.rstrip('\n')
data = py_rw.read_raw_mat(fileDir + os.path.sep + fileName + fileExt, fileDim)
if (fileExt == '.lf0' or fileExt =='.f0') and fileDim == 1:
data = data[data>0]
gvData[cnt, :] = gv(data)
cnt = cnt + 1
#print fileName
py_rw.write_raw_mat(gvData, fileOut + os.path.sep + 'gv.data.bin')
print fileOut, '\t', np.median(gvData, axis=0)
# the netTopo should be
netTopo = [382, 512, 256, 259]
# the layerType should be
# [0]: input layer
# [1]: feedforward layer
# [2]: blstm layer
layerType = [0, 1, 1, 1]
# Now, specify the connection between the 1st and 2nd, 2nd and 3rd, 3rd and 4th layer
# [input_start_dimension, input_end_dimension, output_start_dimension, output_end_dimension]
#
# I want to connect the 1-256th components of 2nd layer to the 1-128th elements of 3rd layer
# it's [0, 256, 0, 128]
# Then 257-512th components of 2nd layer to the 129-256th of 3rd layer
# [256, 512, 128, 256]
# NOTE: python style of index [0, 5] => [0, 1, 2, 3, 4]
netconfig= [[], # void for the 1st-2nd layer
[[0, 256, 0, 128], [256, 512, 128, 256]], # 2nd-3rd layer
[]] # void for the 3rd-4th layer
#
netTopo = np.array(netTopo)
netTopoNoInput = netTopo[1:]
layerNM, TotalWNum = getNetStruct(netTopo, netTopoNoInput, layerType)
weight = genWeightMast(TotalWNum, netconfig, layerNM, netTopoNoInput,
netTopo, layerType, 1)
py_rw.write_raw_mat(weight, './weightMask')
#os.system("cat ./" + __file__ + " > ./log")
if __name__ == "__main__":
if sys.argv[1] == 'gv':
fileDir = sys.argv[2]
fileList = sys.argv[3]
fileExt = sys.argv[4]
fileDim = int(sys.argv[5])
fileOut = sys.argv[6]
cnt = 0
with open(fileList, 'r') as filePtr:
for idx, fileName in enumerate(filePtr):
cnt = cnt + 1
gvData = np.zeros([cnt, fileDim])
cnt = 0
with open(fileList, 'r') as filePtr:
for idx, fileName in enumerate(filePtr):
fileName = fileName.rstrip('\n')
data = py_rw.read_raw_mat(
fileDir + os.path.sep + fileName + fileExt, fileDim)
if (fileExt == '.lf0' or fileExt == '.f0') and fileDim == 1:
data = data[data > 0]
gvData[cnt, :] = gv(data)
cnt = cnt + 1
#print fileName
py_rw.write_raw_mat(gvData, fileOut + os.path.sep + 'gv.data.bin')
print(fileOut, '\t', np.median(gvData, axis=0))
def f02lf0_file(data_file, out_file, thres=10.0, unvoiced_value=-1.0e+10):
data = py_rw.read_raw_lf0(data_file, 1)
data[data>thres] = f02lf0(data[data>thres])
data[data<=thres] = unvoiced_value
py_rw.write_raw_mat(data, out_file)
from ioTools import readwrite as py_rw
import os
import sys
import numpy as np
dirPath = sys.argv[1]
quantiBitNum = int(sys.argv[2])
samplingRate = int(sys.argv[3])
fileList = py_rw.read_txt_list(dirPath + '/gen.scp')
for fileName in fileList:
fileName = fileName.rstrip('\n')
nameHtk = dirPath + os.path.sep + os.path.basename(fileName).rstrip(
'.htk') + '.htk'
nameRaw = dirPath + os.path.sep + os.path.basename(fileName).rstrip(
'.htk') + '.raw'
nameWav = dirPath + os.path.sep + os.path.basename(fileName).rstrip(
'.htk') + '.wav'
print nameRaw, nameWav
data = py_rw.read_htk(nameHtk, 'f4', 'b')
if quantiBitNum > 0:
quantiLevel = np.power(2, quantiBitNum) - 1
py_rw.write_raw_mat(data, nameRaw)
wavTool.raw2wav(nameRaw,
nameWav,
quantiLevel,
samplingRate=samplingRate)
else:
wavfile.write(nameWav, samplingRate, data)
bias = temp+bias
print MDNNNOutDim
# check and generating the MDN configuration
mdnconfigdata = np.zeros([1+len(MDNType)*5], dtype = np.float32)
mdnconfigdata[0] = len(MDNType)
tmp = 0
for idx, mdntype in enumerate(MDNType):
mdntarDim = MDNTargetDim[idx]
mdnoutDim = MDNNNOutDim[idx]
tmp1 = kmixPara(mdntarDim[1]-mdntarDim[0], mdntype, tieVariance)
tmp2 = (mdnoutDim[1]-mdnoutDim[0])
if mdntype > 0:
assert tmp1 == tmp2, "Error in MDN mixture configuraiton"
tmp = tmp + tmp2
elif mdntype < 0:
assert mdntarDim[1]-mdntarDim[0]==1, "Softmax to 1 dimension targert"
tmp = tmp + (mdnoutDim[1]-mdnoutDim[0])
mdntype = -1 # change it back to -1
else:
tmp = tmp + tmp2
mdnconfigdata[(idx*5+1):((idx+1)*5+1)] = [mdnoutDim[0],mdnoutDim[1],
mdntarDim[0],mdntarDim[1],
mdntype]
print "Dimension of output of NN should be %d" % (tmp)
py_rw.write_raw_mat(mdnconfigdata, mdnconfig)
import sys
from os.path import join, basename
from ioTools import readwrite
from kaldiio import ReadHelper
args = sys.argv
ppg_file = args[1]
out_dir = args[2]
ppg_out_dir = join(out_dir, "ppg")
print("Writing PPG feats.....")
# Write ppg features
with ReadHelper('scp:'+ppg_file) as reader:
for key, mat in reader:
readwrite.write_raw_mat(mat, join(ppg_out_dir, key+'.ppg'))
print("Finished writing PPG feats.")
source_stats = {}
with open(statsdir + dataname + "/" +
key.split("-")[0].split("_")[0]) as f:
source_stats = json.load(f)
selected_target_speaker_list = [target_spk]
pseudo_speaker_f0_stats = {"mu_s": 0, "var_s": 0, "std_s": 0}
for selected_target_speaker in selected_target_speaker_list:
target_speaker_stats = {}
with open(statsdir + dataset_of_target + "/" +
selected_target_speaker) as f:
target_speaker_stats = json.load(f)
mu = target_speaker_stats["mu_s"]
var = target_speaker_stats["var_s"]
pseudo_speaker_f0_stats["mu_s"] += mu
pseudo_speaker_f0_stats["var_s"] += var
pseudo_speaker_f0_stats["var_s"] /= len(selected_target_speaker_list)
pseudo_speaker_f0_stats["mu_s"] /= len(selected_target_speaker_list)
pseudo_speaker_f0_stats["std_s"] = math.sqrt(
pseudo_speaker_f0_stats["var_s"])
transfomation = {
**source_stats, "mu_t": pseudo_speaker_f0_stats["mu_s"],
"std_t": pseudo_speaker_f0_stats["std_s"]
}
f0t = log_linear_transformation(f0.copy(), transfomation)
readwrite.write_raw_mat(f0t, join(pitch_out_dir, key + '.f0'))
if os.path.isdir(dataOut):
pass
else:
os.mkdir(dataOut)
for dataFile in dataList:
if re.search(dataPattern, dataFile):
data = io.netcdf_file(dataDir + os.path.sep + dataFile)
uttNum = data.dimensions['numSeqs']
seqLengths = data.variables['seqLengths'][:].copy()
seqLengths = np.concatenate((np.array([0]), seqLengths)).cumsum()
seqTags = data.variables['seqTags'][:]
if inOutData == 1:
dataAll = data.variables['inputs'][:]
else:
dataAll = data.variables['targetPatterns'][:]
for i in range(uttNum):
outName = dataOut + os.path.sep + ''.join(seqTags[i])
for j, suf in enumerate(name):
outFile = outName + suf
tmpdata = dataAll[seqLengths[i]:seqLengths[i + 1],
dim[j][0]:dim[j][1]].copy()
if dataMV is not None:
tmpdata = tmpdata * varVec[dim[j][0]:dim[j][1]] + meanVec[
dim[j][0]:dim[j][1]]
py_rw.write_raw_mat(tmpdata, outFile)
print("%s Utt %d" % (dataFile, i))
del dataAll, seqTags, seqLengths, uttNum
data.close()
if os.path.isdir(dataOut):
pass
else:
os.mkdir(dataOut)
for dataFile in dataList:
if re.search(dataPattern, dataFile):
data = io.netcdf_file(dataDir+os.path.sep+dataFile)
uttNum = data.dimensions['numSeqs']
seqLengths = data.variables['seqLengths'][:].copy()
seqLengths = np.concatenate((np.array([0]), seqLengths)).cumsum()
seqTags = data.variables['seqTags'][:]
if inOutData == 1:
dataAll = data.variables['inputs'][:]
else:
dataAll = data.variables['targetPatterns'][:]
for i in xrange(uttNum):
outName = dataOut+os.path.sep+''.join(seqTags[i])
for j, suf in enumerate(name):
outFile = outName + suf
tmpdata = dataAll[seqLengths[i]:seqLengths[i+1],dim[j][0]:dim[j][1]].copy()
if dataMV is not None:
tmpdata = tmpdata*varVec[dim[j][0]:dim[j][1]]+meanVec[dim[j][0]:dim[j][1]]
py_rw.write_raw_mat(tmpdata, outFile)
print "%s Utt %d" % (dataFile, i)
del dataAll, seqTags, seqLengths, uttNum
data.close()
yaap_pitch_dir = join(data_dir, 'yaapt_pitch')
xvec_out_dir = join(out_dir, "xvector")
pitch_out_dir = join(out_dir, "f0")
# Write pitch features
pitch_file = join(data_dir, 'pitch.scp')
pitch2shape = {}
with ReadHelper('scp:'+pitch_file) as reader:
for key, mat in reader:
pitch2shape[key] = mat.shape[0]
kaldi_f0 = mat[:, 1].squeeze().copy()
yaapt_f0 = readwrite.read_raw_mat(join(yaap_pitch_dir, key+'.f0'), 1)
#unvoiced = np.where(yaapt_f0 == 0)[0]
#kaldi_f0[unvoiced] = 0
#readwrite.write_raw_mat(kaldi_f0, join(pitch_out_dir, key+'.f0'))
f0 = np.zeros(kaldi_f0.shape)
f0[:yaapt_f0.shape[0]] = yaapt_f0
readwrite.write_raw_mat(f0, join(pitch_out_dir, key+'.f0'))
# Write xvector features
with ReadHelper('scp:'+xvector_file) as reader:
for key, mat in reader:
#print key, mat.shape
plen = pitch2shape[key]
mat = mat[np.newaxis]
xvec = np.repeat(mat, plen, axis=0)
readwrite.write_raw_mat(xvec, join(xvec_out_dir, key+'.xvector'))
#
# skippara, feedforward can be specified using [row_start, row_end, col_start, col_end] for each block
# blstm:
netconfig = [[], [[0, 5, 0, 4], [5, 10, 4, 8]], []]
netTopo = [382, 768, 768, 768, 768, 768, 768, 259]
layerType = dupVec(1, len(netTopo))
layerType[0] = 0
netconfig = dupVec([[0, 382, 0, 382], [382, 382 + 256, 382, 382 + 256],
[382 + 256, 768, 382 + 256, 768]],
len(netTopo) - 1)
netconfig[0] = []
netconfig[1] = []
netconfig[len(netTopo) - 2] = [[0, 382, 0, 180],
[382, 382 + 256, 180, 184],
[382 + 256, 768, 184, 259]]
netTopo = np.array(netTopo)
netTopoNoInput = netTopo[1:]
layerNM, TotalWNum = getNetStruct(netTopo, netTopoNoInput, layerType)
weight = genWeightMast(TotalWNum, netconfig, layerNM, netTopoNoInput,
netTopo, layerType, 1)
py_rw.write_raw_mat(weight, './weightMask')
os.system("cat ./" + __file__ + " > ./log")
fileptr = open('./log', 'a')
fileptr.write('Number of parameter: %d \n' % (TotalWNum))
fileptr.write('Number of non-zero parameter: %d \n' % (np.sum(weight)))
fileptr.close()
print np.sum(weight)
import sys
from os.path import join, basename
from ioTools import readwrite
import kaldi_io
args = sys.argv
mspec_file = args[1]
out_dir = args[2]
mspec_out_dir = join(out_dir, "mel")
print "Writing MEL feats....."
# Write mspec features
for key, mat in kaldi_io.read_mat_scp(mspec_file):
#print key, mat.shape
readwrite.write_raw_mat(mat, join(mspec_out_dir, key + '.mel'))
print "Finished writing MEL feats."