def encode(a, pcm):
"""
Encode a speech waveform. The encoding framers (frames and pitch)
pad the frames so that the first frame is centered on sample zero.
This is consistent with STRAIGHT and SPTK (I hope!). At least, it
means the pitch can have longer frame lengths and still align with
the OLA'd frames.
"""
if opt.ola:
frameSize = pcm.seconds_to_period(0.025, 'atleast') # 25ms frame size
else:
frameSize = framePeriod
pitchSize = pcm.seconds_to_period(0.1, 'atmost')
print "Encoding with period", framePeriod, "size", frameSize, \
"and pitch window", pitchSize
# First the pitch as it's on the unaltered waveform. The frame
# should be long with no window. 1024 at 16 kHz is 64 ms.
pf = ssp.Frame(a, size=pitchSize, period=framePeriod)
pitch, hnr = ssp.ACPitch(pf, pcm)
# Pre-emphasis
pre = ssp.parameter("Pre", None)
if pre is not None:
a = ssp.PoleFilter(a, pre) / 5
# Keep f around after the function so the decoder can do a
# reference decoding on the real excitaton.
global f
f = ssp.Frame(a, size=frameSize, period=framePeriod)
#aw = np.hanning(frameSize+1)
aw = ssp.nuttall(frameSize+1)
aw = np.delete(aw, -1)
w = ssp.Window(f, aw)
ac = ssp.Autocorrelation(w)
lp = ssp.parameter('AR', 'levinson')
if lp == 'levinson':
ar, g = ssp.ARLevinson(ac, lpOrder[r])
elif lp == 'ridge':
ar, g = ssp.ARRidge(ac, lpOrder[r], 0.03)
elif lp == 'lasso':
ar, g = ssp.ARLasso(ac, lpOrder[r], 5)
elif lp == 'sparse':
ar, g = ssp.ARSparse(w, lpOrder[r], ssp.parameter('Gamma', 1.414))
elif lp == 'student':
ar, g = ssp.ARStudent(w, lpOrder[r], ssp.parameter('DoF', 50.0))
if False:
fig = ssp.Figure(5, 1)
#stddev = np.sqrt(kVar)
sPlot = fig.subplot()
sPlot.plot(pitch, 'c')
#sPlot.plot(kPitch + stddev, 'b')
#sPlot.plot(kPitch - stddev, 'b')
sPlot.set_xlim(0, len(pitch))
sPlot.set_ylim(0, 500)
plt.show()
return (ar, g, pitch, hnr)
def encode(a, pcm):
"""
Encode a speech waveform. The encoding framers (frames and pitch)
pad the frames so that the first frame is centered on sample zero.
This is consistent with STRAIGHT and SPTK (I hope!). At least, it
means the pitch can have longer frame lengths and still align with
the OLA'd frames.
"""
if opt.ola:
frameSize = pcm.seconds_to_period(0.025, 'atleast') # 25ms frame size
else:
frameSize = framePeriod
pitchSize = pcm.seconds_to_period(0.1, 'atmost')
print "Encoding with period", framePeriod, "size", frameSize, \
"and pitch window", pitchSize
# First the pitch as it's on the unaltered waveform. The frame
# should be long with no window. 1024 at 16 kHz is 64 ms.
pf = ssp.Frame(a, size=pitchSize, period=framePeriod)
pitch, hnr = ssp.ACPitch(pf, pcm)
# Pre-emphasis
pre = ssp.parameter("Pre", None)
if pre is not None:
a = ssp.PoleFilter(a, pre) / 5
# Keep f around after the function so the decoder can do a
# reference decoding on the real excitaton.
global f
f = ssp.Frame(a, size=frameSize, period=framePeriod)
#aw = np.hanning(frameSize+1)
aw = ssp.nuttall(frameSize+1)
aw = np.delete(aw, -1)
w = ssp.Window(f, aw)
ac = ssp.Autocorrelation(w)
lp = ssp.parameter('AR', 'levinson')
if lp == 'levinson':
ar, g = ssp.ARLevinson(ac, lpOrder[r])
elif lp == 'ridge':
ar, g = ssp.ARRidge(ac, lpOrder[r], 0.03)
elif lp == 'lasso':
ar, g = ssp.ARLasso(ac, lpOrder[r], 5)
elif lp == 'sparse':
ar, g = ssp.ARSparse(w, lpOrder[r], ssp.parameter('Gamma', 1.414))
elif lp == 'student':
ar, g = ssp.ARStudent(w, lpOrder[r], ssp.parameter('DoF', 50.0))
if False:
fig = ssp.Figure(5, 1)
#stddev = np.sqrt(kVar)
sPlot = fig.subplot()
sPlot.plot(pitch, 'c')
#sPlot.plot(kPitch + stddev, 'b')
#sPlot.plot(kPitch - stddev, 'b')
sPlot.set_xlim(0, len(pitch))
sPlot.set_ylim(0, 500)
plt.show()
return (ar, g, pitch, hnr)
def setUp(self):
"""
Generate one (short) frame of a 1 kHz sinusoid at a sampling
rate of 16 kHz. It doesn't matter too much what it is, just
that it is representative of some natural signal.
"""
self.pcm = ssp.PulseCodeModulation(16000)
self.seq = np.zeros(64)
p = self.pcm.seconds_to_period(1.0/1000);
for s in range(len(self.seq)):
self.seq[s] = np.sin(2*np.pi * s/p)
w = ssp.nuttall(len(self.seq)+1)
w = np.delete(w, -1)
self.seq = ssp.Window(self.seq, w)
def get_pitch(gen_path, basefilename):
(Fs, x) = io_wav.read(gen_path + basefilename + '.wav')
assert Fs == 16000
pcm = ssp.PulseCodeModulation(Fs)
frameSize = pcm.seconds_to_period(0.025, 'atleast') # 25ms Frame size
pitchSize = pcm.seconds_to_period(0.1, 'atmost') # 100ms Pitch size
pf = ssp.Frame(x, size=pitchSize, period=framePeriod)
pitch, ac = ssp.ACPitch(pf, pcm, loPitch,
hiPitch) # Initially pitch estimated
# Pre-emphasis
pre = ssp.parameter("Pre", None)
if pre is not None:
x = ssp.PoleFilter(x, pre) / 5
# Frame Splitting
f = ssp.Frame(x, size=frameSize, period=framePeriod)
# Windowing
aw = ssp.nuttall(frameSize + 1)
aw = np.delete(aw, -1)
w = ssp.Window(f, aw)
# Autocorrelation
ac = ssp.Autocorrelation(w)
if (len(ac) > len(pitch)):
d = len(ac) - len(pitch)
addon = np.ones(d) * pitch[-1]
pitch = np.hstack((pitch, addon))
# Save pitch as binary
lf0 = np.log(pitch)
lf0.astype('float32').tofile(gen_path + basefilename + '.lf0')
return pitch
print "Need one arg"
exit(1)
file = arg[0]
import ssp
import numpy as np
# Load and process
pcm = ssp.PulseCodeModulation()
a = pcm.WavSource(file)
if (ssp.parameter('Pre', None)):
a = ssp.ZeroFilter(a)
framePeriod = pcm.seconds_to_period(0.01)
frameSize = pcm.seconds_to_period(0.02, 'atleast')
f = ssp.Frame(a, size=frameSize, period=framePeriod)
w = ssp.nuttall(frameSize+1)
w = np.delete(w, -1)
wf = ssp.Window(f, w)
type = ssp.parameter('Type', 'psd')
if type == 'psd':
p = ssp.Periodogram(wf)
p = p[:,:p.shape[1]/2+1]
elif type == 'ar':
a = ssp.Autocorrelation(wf)
a, g = ssp.ARLevinson(a, pcm.speech_ar_order())
p = ssp.ARSpectrum(a, g, nSpec=128)
elif type == 'snr':
p = ssp.Periodogram(wf)
n = ssp.Noise(p)
p = ssp.SNRSpectrum(p, n)
p = p[:,:p.shape[1]/2+1]
# Defaults for 8 kHz
frameSize = 256
framePeriod = 80
lpOrder = 10
if pcm.rate == 16000:
frameSize = 400
framePeriod = 160
lpOrder = 12
# Basic preprocessing
g = np.ndarray((0))
a = ssp.ZeroFilter(a)
f = ssp.Frame(a, size=frameSize, period=framePeriod, pad=False)
f = ssp.Window(f, ssp.nuttall(frameSize))
# Next part depends on user
frontend = ssp.parameter("FrontEnd", "ar")
if frontend == "ar":
a = ssp.Autocorrelation(f)
a = ssp.AutocorrelationAllPassWarp(a,
alpha=ssp.mel[pcm.rate],
size=lpOrder + 1)
a, g = ssp.ARLevinson(a, lpOrder)
# ridge = Parameter('Ridge', 0.1)
# a, g = ARRidge(a, lpOrder, ridge)
# a, g = ARLasso(a, lpOrder, ridge)
elif frontend == "snr":
a = ssp.Periodogram(f)
n = ssp.Noise(a)
# Defaults for 8 kHz
frameSize = 256
framePeriod = 80
lpOrder = 10
if pcm.rate == 16000:
frameSize = 400
framePeriod = 160
lpOrder = 12
# Basic preprocessing
g = np.ndarray((0))
a = ssp.ZeroFilter(a)
f = ssp.Frame(a, size=frameSize, period=framePeriod, pad=False)
f = ssp.Window(f, ssp.nuttall(frameSize))
# Next part depends on user
frontend = ssp.parameter("FrontEnd", "ar")
if frontend == "ar":
a = ssp.Autocorrelation(f)
a = ssp.AutocorrelationAllPassWarp(a, alpha=ssp.mel[pcm.rate],
size=lpOrder+1)
a, g = ssp.ARLevinson(a, lpOrder)
# ridge = Parameter('Ridge', 0.1)
# a, g = ARRidge(a, lpOrder, ridge)
# a, g = ARLasso(a, lpOrder, ridge)
elif frontend == "snr":
a = ssp.Periodogram(f)
n = ssp.Noise(a)
a = ssp.SNRSpectrum(a, n * 0.1)