def mcep2spec(MCEP, alpha, dftlen):
misc.check_executable(
'mgc2sp',
'You can find it in SPTK (https://sourceforge.net/projects/sp-tk/)')
order = MCEP.shape[1] - 1
tmpspecfile = sp.gentmpfile('mcep2spec.mcep')
outspecfile = sp.gentmpfile('mcep2spec.spec')
try:
MCEP.astype(np.float32).tofile(tmpspecfile)
cmd = os.path.join(
sp.BINPATH, 'mgc2sp') + ' -a ' + str(alpha) + ' -g 0 -m ' + str(
int(order)) + ' -l ' + str(
dftlen) + ' -o 2 ' + tmpspecfile + ' > ' + outspecfile
ret = os.system(cmd)
if ret > 0: raise ValueError('ERROR during execution of mgc2sp')
SPEC = np.fromfile(outspecfile, dtype=np.float32)
SPEC = SPEC.reshape((-1, int(dftlen / 2) + 1))
except:
if os.path.exists(tmpspecfile): os.remove(tmpspecfile)
if os.path.exists(outspecfile): os.remove(outspecfile)
raise
if os.path.exists(tmpspecfile): os.remove(tmpspecfile)
if os.path.exists(outspecfile): os.remove(outspecfile)
return SPEC
def analysis(wav, fs, f0s, shift, dftlen):
'''
Compute features using the STRAIGHT vocoder.
(STRAIGHT's Analysis stage)
fs : [Hz]
f0s : [s,Hz]
shift : [s]
dftlen :
'''
#f0s = np.interp(np.arange(f0s[0,0], f0s[-1,0], shift), f0s[:,0], np.log(f0s[:,1])
f0shift = np.median(np.diff(f0s[:,0]))
tmprawwavfile = sigproc.gentmpfile('straight-analysis.raw')
tmpf0file = sigproc.gentmpfile('straight-analysis.f0')
tmpspecfile = sigproc.gentmpfile('straight-analysis.spec')
tmpaperfile = sigproc.gentmpfile('straight-analysis.aper')
try:
((wav.copy()*np.iinfo(np.int16).max).astype(np.int16)).tofile(tmprawwavfile)
np.savetxt(tmpf0file, f0s[:,1], fmt='%f')
#print('STRAIGHTPATH='+STRAIGHTPATH)
# Aperiodicity estimation
cmd = STRAIGHTPATH+'straight_bndap -float -nmsg -f '+str(fs)+' -shift '+str(shift*1000)+' -f0shift '+str(f0shift*1000)+' -f0file '+tmpf0file+' -fftl '+str(dftlen)+' -apord '+str(dftlen/2+1)+' -raw '+tmprawwavfile+' '+tmpaperfile # -nmsg
os.system(cmd)
# Spectral envelope estimation
cmd = STRAIGHTPATH+'straight_mcep -float -nmsg -f '+str(fs)+' -fftl '+str(dftlen)+' -apord '+str(dftlen/2+1)+' -shift '+str(shift*1000)+' -f0shift '+str(f0shift*1000)+' -f0file '+tmpf0file+' -pow -raw '+tmprawwavfile+' '+tmpspecfile # -nmsg
os.system(cmd)
SPEC = np.fromfile(tmpspecfile, dtype='float32')
SPEC = SPEC.reshape((-1, dftlen/2+1))
APER = np.fromfile(tmpaperfile, dtype='float32')
APER = APER.reshape((-1, dftlen/2+1))
except:
if os.path.exists(tmprawwavfile): os.remove(tmprawwavfile)
if os.path.exists(tmpf0file): os.remove(tmpf0file)
if os.path.exists(tmpspecfile): os.remove(tmpspecfile)
if os.path.exists(tmpaperfile): os.remove(tmpaperfile)
raise
if os.path.exists(tmprawwavfile): os.remove(tmprawwavfile)
if os.path.exists(tmpf0file): os.remove(tmpf0file)
if os.path.exists(tmpspecfile): os.remove(tmpspecfile)
if os.path.exists(tmpaperfile): os.remove(tmpaperfile)
if 0:
import matplotlib.pyplot as plt
plt.ion()
f, axs = plt.subplots(2, 1, sharex=True, sharey=True)
axs[0].imshow(sp.mag2db(SPEC.T), origin='lower', aspect='auto', interpolation='none', extent=(0, SPEC.shape[0]*shift, 0, fs/2))
axs[1].imshow(APER.T, origin='lower', aspect='auto', interpolation='none', extent=(0, SPEC.shape[0]*shift, 0, fs/2))
from IPython.core.debugger import Pdb; Pdb().set_trace()
return SPEC, APER
def spec2mcep(SPEC, alpha, order):
misc.check_executable(
'mcep',
'You can find it in SPTK (https://sourceforge.net/projects/sp-tk/)')
dftlen = 2 * (SPEC.shape[1] - 1)
tmpspecfile = sp.gentmpfile('spec2mcep.spec')
outspecfile = sp.gentmpfile('spec2mcep.mcep')
try:
SPEC.astype(np.float32).tofile(tmpspecfile)
cmd = os.path.join(
sp.BINPATH, 'mcep'
) + ' -a ' + str(alpha) + ' -m ' + str(int(order)) + ' -l ' + str(
dftlen
) + ' -e 1.0E-8 -j 0 -f 0.0 -q 3 ' + tmpspecfile + ' > ' + outspecfile
ret = os.system(cmd)
if ret > 0: raise ValueError('ERROR during execution of mcep')
MCEP = np.fromfile(outspecfile, dtype=np.float32)
MCEP = MCEP.reshape((-1, 1 + int(order)))
except:
if os.path.exists(tmpspecfile): os.remove(tmpspecfile)
if os.path.exists(outspecfile): os.remove(outspecfile)
raise
if os.path.exists(tmpspecfile): os.remove(tmpspecfile)
if os.path.exists(outspecfile): os.remove(outspecfile)
return MCEP
def analysis_aper(wav, fs, f0s, shift, dftlen):
'''
Compute features using the STRAIGHT vocoder.
(STRAIGHT's Analysis stage)
fs : [Hz]
f0s : [s,Hz]
shift : [s]
dftlen :
'''
#f0s = np.interp(np.arange(f0s[0,0], f0s[-1,0], shift), f0s[:,0], np.log(f0s[:,1])
f0shift = np.median(np.diff(f0s[:,0]))
tmprawwavfile = sigproc.gentmpfile('straight-analysis-aper.raw')
tmpf0file = sigproc.gentmpfile('straight-analysis-aper.f0')
tmpaperfile = sigproc.gentmpfile('straight-analysis-aper.aper')
try:
((wav.copy()*np.iinfo(np.int16).max).astype(np.int16)).tofile(tmprawwavfile)
np.savetxt(tmpf0file, f0s[:,1], fmt='%f')
# Aperiodicity estimation
cmd = STRAIGHTPATH+'straight_bndap -float -nmsg -f '+str(fs)+' -shift '+str(shift*1000)+' -f0shift '+str(f0shift*1000)+' -f0file '+tmpf0file+' -fftl '+str(dftlen)+' -apord '+str(dftlen/2+1)+' -raw '+tmprawwavfile+' '+tmpaperfile # -nmsg
ret = os.system(cmd)
if ret>0: raise ValueError('ERROR during execution of straight_bndap')
APER = np.fromfile(tmpaperfile, dtype='float32')
APER = APER.reshape((-1, dftlen/2+1))
except:
if os.path.exists(tmprawwavfile): os.remove(tmprawwavfile)
if os.path.exists(tmpf0file): os.remove(tmpf0file)
if os.path.exists(tmpaperfile): os.remove(tmpaperfile)
raise
if os.path.exists(tmprawwavfile): os.remove(tmprawwavfile)
if os.path.exists(tmpf0file): os.remove(tmpf0file)
if os.path.exists(tmpaperfile): os.remove(tmpaperfile)
if 0:
import matplotlib.pyplot as plt
plt.ion()
plt.imshow(APER.T, origin='lower', aspect='auto', interpolation='none', extent=(0, APER.shape[0]*shift, 0, fs/2))
from IPython.core.debugger import Pdb; Pdb().set_trace()
return APER
def sv56demo(wav, fs, level=-26):
'''
Normalize the waveform in amplitude
'''
misc.check_executable('sv56demo', 'source?')
tmpinfname = sigproc.gentmpfile('interface-sv56demo-in.raw')
tmpoutfname = sigproc.gentmpfile('interface-sv56demo-out.raw')
try:
(wav.copy() * np.iinfo(np.int16).max).astype(
np.int16).tofile(tmpinfname)
levelstr = ''
if level != None:
levelstr = ' -lev ' + str(level)
cmd = os.path.join(
sigproc.BINPATH, 'sv56demo') + ' ' + levelstr + ' -sf ' + str(
fs) + ' ' + tmpinfname + ' ' + tmpoutfname + ' 640'
out = subprocess.check_output(cmd,
stderr=subprocess.STDOUT,
shell=True)
#print(out)
# Put outputs in a dictionary
m = re.findall(r'\n?\s*(.*): .* ([-+]?[0-9]+\.?[0-9]*).*', out)
meta = dict(m)
meta.pop('Input file', None)
for key in meta:
meta[key] = float(meta[key])
wav = np.fromfile(tmpoutfname, dtype=np.int16)
wav = wav / float(np.iinfo(np.int16).max)
except:
if os.path.exists(tmpinfname): os.remove(tmpinfname)
if os.path.exists(tmpoutfname): os.remove(tmpoutfname)
raise
if os.path.exists(tmpinfname): os.remove(tmpinfname)
if os.path.exists(tmpoutfname): os.remove(tmpoutfname)
return wav, meta
def avread(filename):
'''
Read any audio
'''
tmpoutfname = sp.gentmpfile('avread.wav')
try:
os.system('avconv -i ' + filename + ' ' + tmpoutfname +
' >/dev/null 2>&1')
#os.system('avconv -i '+filename+' '+tmpoutfname)
wav, fs, enc = wavread(tmpoutfname)
except:
if os.path.exists(tmpoutfname): os.remove(tmpoutfname)
raise
if os.path.exists(tmpoutfname): os.remove(tmpoutfname)
return wav, fs, enc
def reaper(wav,
fs,
shift,
f0min,
f0max,
rmsteps=False,
outpm=False,
prehp_cutoff=None):
'''
Estimate the f0
prehp_cutoff: High-pass the signal before estimating f0 (def. None). Value is [Hz].
Source:
https://github.com/google/REAPER
'''
misc.check_executable(
'reaper',
'You can download the source from https://github.com/google/REAPER')
if not prehp_cutoff is None:
b, a = scipy.signal.butter(8,
prehp_cutoff / (0.5 / shift),
btype='high')
wav = scipy.signal.filtfilt(b, a, wav)
tmpwavfile = sigproc.gentmpfile('interface-reaper-in.wav')
tmpf0file = sigproc.gentmpfile('interface-reaper-out.f0')
tmppmfile = sigproc.gentmpfile('interface-reaper-out.pm')
try:
fileio.wavwrite(tmpwavfile, wav, fs, np.int16)
cmd = os.path.join(
sigproc.BINPATH, 'reaper') + ' -i ' + tmpwavfile + ' -m ' + str(
f0min) + ' -x ' + str(f0max) + ' -e ' + str(
shift) + ' -a -f ' + tmpf0file + ' -p ' + tmppmfile
os.system(cmd + ' 2>&1 >/dev/null') # Run it silent
# Read the f0 values and the corresponding analysis times
f0s = list()
with open(tmpf0file) as f0file:
n = 1
for line in f0file:
if n > 7: # Skip the EST-file header TODO Neater way to do this!
values = line.split()
f0s.append((float(values[0]), float(values[2])))
n += 1
f0s = np.array(f0s)
f0s[f0s[:, 1] < 0, 1] = 0
if rmsteps:
f0s = resampling.f0s_rmsteps(f0s)
# Read the f0 values and the corresponding analysis times
pms = list()
with open(tmppmfile) as pmfile:
n = 1
for line in pmfile:
if n > 7: # Skip the EST-file header TODO Neater way to do this!
values = line.split()
pms.append(float(values[0]))
n += 1
pms = np.array(pms)
except:
if os.path.exists(tmpwavfile): os.remove(tmpwavfile)
if os.path.exists(tmpf0file): os.remove(tmpf0file)
if os.path.exists(tmppmfile): os.remove(tmppmfile)
raise
if os.path.exists(tmpwavfile): os.remove(tmpwavfile)
if os.path.exists(tmpf0file): os.remove(tmpf0file)
if os.path.exists(tmppmfile): os.remove(tmppmfile)
if outpm:
return f0s, pms
else:
return f0s
def synthesis(fs, f0s, SPEC, shift, APER=None, f0_forcecont=False, verbose=0):
if f0s.shape[0]!=SPEC.shape[0] or (APER!=None and (SPEC.shape[0]!=APER.shape[0])):
raise ValueError('Features dont have the same length (F0={}, SPEC={}, APER={}).'.format(f0s.shape, SPEC.shape, APER.shape))
if APER!=None and (SPEC.shape[1]!=APER.shape[1]):
raise ValueError('Spectral features dont have the same width (SPEC={}, APER={}).'.format(SPEC.shape, APER.shape))
dftlen = (SPEC.shape[1]-1)*2
if APER==None:
APER = np.zeros(SPEC.shape)
APER[f0s[:,1]>0,:int((float(dftlen)/fs)*4000.0)] = -100.0
if f0_forcecont:
# Replace zero values by interpolated values
idx = f0s[:,1]>0
f0s[:,1] = np.exp(np.interp(f0s[:,0], f0s[idx,0], np.log(f0s[idx,1])))
tmpf0file = sigproc.gentmpfile('straight-synthesis.f0')
tmpaperfile = sigproc.gentmpfile('straight-synthesis.aper')
tmpspecfile = sigproc.gentmpfile('straight-synthesis.spec')
tmpwavfile = sigproc.gentmpfile('straight-synthesis.wav')
try:
np.savetxt(tmpf0file, f0s[:,1], fmt='%f')
APER.astype(np.float32).tofile(tmpaperfile)
SPEC.astype(np.float32).tofile(tmpspecfile)
cmd = STRAIGHTPATH+'synthesis_fft \
-f '+str(fs)+' \
-fftl '+str(dftlen)+' \
-spec \
-shift '+str(shift*1000)+' \
-sigp 1.2 \
-sd 0.5 \
-cornf 4000 \
-bw 70.0 \
-delfrac 0.2 \
-apfile '+tmpaperfile+' \
-float \
'+tmpf0file+' \
'+tmpspecfile+' \
'+tmpwavfile+' ' #>/dev/null 2>&1
if verbose==0:
cmd += ' > /dev/null 2>&1'
#print(cmd)
os.system(cmd)
wavfs, wav = scipy.io.wavfile.read(tmpwavfile)
wavdtype = wav.dtype
wav = wav / float(np.iinfo(wavdtype).max)
except:
if os.path.exists(tmpf0file): os.remove(tmpf0file)
if os.path.exists(tmpaperfile): os.remove(tmpaperfile)
if os.path.exists(tmpspecfile): os.remove(tmpspecfile)
if os.path.exists(tmpwavfile): os.remove(tmpwavfile)
raise
if os.path.exists(tmpf0file): os.remove(tmpf0file)
if os.path.exists(tmpaperfile): os.remove(tmpaperfile)
if os.path.exists(tmpspecfile): os.remove(tmpspecfile)
if os.path.exists(tmpwavfile): os.remove(tmpwavfile)
return wav