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 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
global ti
now = time.clock()
elapsed = now-ti
ti = now
print(func, elapsed)
import ssp
import numpy as np
import matplotlib.pyplot as plt
lap("Import")
# Load and do basic AR to reconstruct the spectrum
pcm = ssp.PulseCodeModulation()
wav = pcm.WavSource(file)
print("File:", file, "rate:", pcm.rate, "size:", wav.size)
if ssp.parameter("ZF", 0) == 1:
wav = ssp.ZeroFilter(wav)
f = ssp.Frame(wav, size=256, period=128)
f = ssp.Window(f, np.hanning(256))
print("frame:", f.shape[0], "x", f.shape[1])
lap("Frame")
e = ssp.Energy(f)
p = ssp.Periodogram(f)
lap("Periodogram")
order = pcm.speech_ar_order()
a = ssp.Autocorrelation(f)
a, g = ssp.ARLevinson(a, order)
lap("Levinson")
ls = ssp.ARSpectrum(a, g, nSpec=128)
lap("Spectrum")
global ti
now = time.clock()
elapsed = now-ti
ti = now
print func, elapsed
import ssp
import numpy as np
import matplotlib.pyplot as plt
lap("Import")
# Load and do basic AR to reconstruct the spectrum
pcm = ssp.PulseCodeModulation()
wav = pcm.WavSource(file)
print "File:", file, "rate:", pcm.rate, "size:", wav.size
if ssp.parameter("ZF", 0) == 1:
wav = ssp.ZeroFilter(wav)
f = ssp.Frame(wav, size=256, period=128)
f = ssp.Window(f, np.hanning(256))
print "frame:", f.shape[0], "x", f.shape[1]
lap("Frame")
e = ssp.Energy(f)
p = ssp.Periodogram(f)
lap("Periodogram")
order = pcm.speech_ar_order()
a = ssp.Autocorrelation(f)
a, g = ssp.ARLevinson(a, order)
lap("Levinson")
ls = ssp.ARSpectrum(a, g, nSpec=128)
lap("Spectrum")
def decode(tuple):
"""
Decode a speech waveform.
"""
(ark, g, pitch, hnr) = tuple
print("Frame padding:", opt.padding)
nFrames = len(ark)
assert(len(g) == nFrames)
assert(len(pitch) == nFrames)
assert(len(hnr) == nFrames)
# The original framer padded the ends so the number of samples to
# synthesise is a bit less than you might think
if opt.ola:
frameSize = framePeriod * 2
nSamples = framePeriod * (nFrames-1)
else:
frameSize = framePeriod
nSamples = frameSize * (nFrames-1)
ex = opt.glottal
if opt.glottal == 'cepgm' and (opt.encode or opt.decode or opt.pitch):
order = ark.shape[-1] - 2
ar = ark[:,0:order]
theta = ark[:,-2]
magni = np.exp(ark[:,-1])
else:
ar = ark
# Use the original AR residual; it should be a very good reconstruction.
if ex == 'ar':
e = ssp.ARExcitation(f, ar, g)
# Just noise. This is effectively a whisper synthesis.
elif ex == 'noise':
e = np.random.normal(size=(nFrames, frameSize))
# Just harmonics, and with a fixed F0. This is the classic robot
# synthesis.
elif ex == 'robot':
ew = np.zeros(nSamples)
period = int(1.0 / 200 * r)
for i in range(0, len(ew), period):
ew[i] = period
e = ssp.Frame(ew, size=frameSize, period=framePeriod)
# Synthesise harmonics plus noise in the ratio suggested by the HNR.
elif ex == 'synth':
# Harmonic part
mperiod = int(1.0 / np.mean(pitch) * r)
gm = ssp.GlottalModel(ssp.parameter('Pulse', 'impulse'))
pr, pg = ssp.pulse_response(gm, pcm, period=mperiod, order=lpOrder[r])
h = np.zeros(nSamples)
i = 0
frame = 0
while i < nSamples and frame < len(pitch):
period = int(1.0 / pitch[frame] * r)
if i + period > nSamples:
break
weight = np.sqrt(hnr[frame] / (hnr[frame] + 1))
h[i:i+period] = gm.pulse(period, pcm) * weight
i += period
frame = i // framePeriod
h = ssp.ARExcitation(h, pr, 1.0)
fh = ssp.Frame(h, size=frameSize, period=framePeriod, pad=opt.padding)
# Noise part
n = np.random.normal(size=nSamples)
n = ssp.ZeroFilter(n, 1.0) # Include the radiation impedance
fn = ssp.Frame(n, size=frameSize, period=framePeriod, pad=opt.padding)
for i in range(len(fn)):
fn[i] *= np.sqrt(1.0 / (hnr[i] + 1))
hgain = ssp.parameter("HGain", 1.0)
e = fn + fh * hgain
# Like harmonics plus noise, but with explicit sinusoids instead of time
# domain impulses.
elif ex == 'sine':
order = 20
sine = ssp.Harmonics(r, order)
h = np.zeros(nSamples)
for i in range(0, len(h)-framePeriod, framePeriod):
frame = i // framePeriod
period = int(1.0 / pitch[frame] * r)
weight = np.sqrt(hnr[frame] / (hnr[frame] + 1))
h[i:i+framePeriod] = ( sine.sample(pitch[frame], framePeriod)
* weight )
fh = ssp.Frame(h, size=frameSize, period=framePeriod, pad=opt.padding)
n = np.random.normal(size=nSamples)
fn = ssp.Frame(n, size=frameSize, period=framePeriod, pad=opt.padding)
for i in range(len(fn)):
fn[i] *= np.sqrt(1.0 / (hnr[i] + 1))
e = fn + fh*10
# High order linear prediction. Synthesise the harmonics using noise to
# excite a high order polynomial with roots resembling harmonics.
elif ex == 'holp':
# Some noise
n = np.random.normal(size=nSamples)
fn = ssp.Frame(n, size=frameSize, period=framePeriod)
# Use the noise to excite a high order AR model
fh = np.ndarray(fn.shape)
for i in range(len(fn)):
hoar = ssp.ARHarmonicPoly(pitch[i], r, 0.7)
fh[i] = ssp.ARResynthesis(fn[i], hoar, 1.0 / linalg.norm(hoar)**2)
print(i, pitch[i], linalg.norm(hoar), np.min(fh[i]), np.max(fh[i]))
print(' ', np.min(hoar), np.max(hoar))
# fh[i] *= np.sqrt(r / pitch[i]) / linalg.norm(fh[i])
# fh[i] *= np.sqrt(hnr[i] / (hnr[i] + 1))
# Weight the noise as for the other methods
for i in range(len(fn)):
fn[i] *= np.sqrt(1.0 / (hnr[i] + 1))
e = fh # fn + fh*30
# Shaped excitation. The pulses are shaped by a filter to have a
# rolloff, then added to the noise. The resulting signal is
# flattened using AR.
elif ex == 'shaped':
# Harmonic part
gm = ssp.GlottalModel(ssp.parameter('Pulse', 'impulse'))
gm.angle = pcm.hertz_to_radians(np.mean(pitch)*0.5)
h = np.zeros(nSamples)
i = 0
frame = 0
while i < nSamples and frame < len(pitch):
period = int(1.0 / pitch[frame] * r)
if i + period > nSamples:
break
weight = np.sqrt(hnr[frame] / (hnr[frame] + 1))
h[i:i+period] = gm.pulse(period, pcm) * weight
i += period
frame = i // framePeriod
# Filter to mimic the glottal pulse
hfilt = ssp.parameter("HFilt", None)
hpole1 = ssp.parameter("HPole1", 0.98)
hpole2 = ssp.parameter("HPole2", 0.8)
angle = pcm.hertz_to_radians(np.mean(pitch)) * ssp.parameter("Angle", 1.0)
if hfilt == 'pp':
h = ssp.ZeroFilter(h, 1.0)
h = ssp.PolePairFilter(h, hpole1, angle)
fh = ssp.Frame(h, size=frameSize, period=framePeriod)
# Noise part
n = np.random.normal(size=nSamples)
zero = ssp.parameter("NoiseZero", 1.0)
n = ssp.ZeroFilter(n, zero) # Include the radiation impedance
npole = ssp.parameter("NPole", None)
nf = ssp.parameter("NoiseFreq", 4000)
if npole is not None:
n = ssp.PolePairFilter(n, npole, pcm.hertz_to_radians(nf))
fn = ssp.Frame(n, size=frameSize, period=framePeriod, pad=opt.padding)
for i in range(len(fn)):
fn[i] *= np.sqrt(1.0 / (hnr[i] + 1))
# Combination
assert(len(fh) == len(fn))
hgain = ssp.parameter("HGain", 1.0)
e = fn + fh * hgain
hnw = np.hanning(frameSize)
for i in range(len(e)):
ep = ssp.Window(e[i], hnw)
#ep = e[i]
eac = ssp.Autocorrelation(ep)
ea, eg = ssp.ARLevinson(eac, order=lpOrder[r])
e[i] = ssp.ARExcitation(e[i], ea, eg)
elif ex == 'ceplf':
omega, alpha = ssp.glottal_pole_lf(
f, pcm, pitch, hnr, visual=(opt.graphic == "ceplf"))
epsilon = ssp.parameter("Epsilon", 5000.0)
h = np.zeros(nSamples)
i = 0
frame = 0
while i < nSamples and frame < len(pitch):
period = int(1.0 / pitch[frame] * r)
if i + period > nSamples:
break
weight = np.sqrt(hnr[frame] / (hnr[frame] + 1))
pu = np.zeros((period))
T0 = pcm.period_to_seconds(period)
print(T0,)
Te = ssp.lf_te(T0, alpha[frame], omega[frame], epsilon)
if Te:
pu = ssp.pulse_lf(pu, T0, Te, alpha[frame], omega[frame], epsilon)
h[i:i+period] = pu * weight
i += period
frame = i // framePeriod
fh = ssp.Frame(h, size=frameSize, period=framePeriod, pad=opt.padding)
# Noise part
n = np.random.normal(size=nSamples)
zero = ssp.parameter("NoiseZero", 1.0)
n = ssp.ZeroFilter(n, zero) # Include the radiation impedance
fn = ssp.Frame(n, size=frameSize, period=framePeriod, pad=opt.padding)
for i in range(len(fn)):
fn[i] *= np.sqrt(1.0 / (hnr[i] + 1))
# Combination
assert(len(fh) == len(fn))
hgain = ssp.parameter("HGain", 1.0)
e = fn + fh * hgain
hnw = np.hanning(frameSize)
for i in range(len(e)):
ep = ssp.Window(e[i], hnw)
#ep = e[i]
eac = ssp.Autocorrelation(ep)
ea, eg = ssp.ARLevinson(eac, order=lpOrder[r])
e[i] = ssp.ARExcitation(e[i], ea, eg)
elif ex == 'cepgm':
# Infer the unstable poles via complex cepstrum, then build an explicit
# glottal model.
if not (opt.encode or opt.decode or opt.pitch):
theta, magni = ssp.glottal_pole_gm(
f, pcm, pitch, hnr, visual=(opt.graphic == "cepgm"))
h = np.zeros(nSamples)
i = 0
frame = 0
while i < nSamples and frame < len(pitch):
period = int(1.0 / pitch[frame] * r)
if i + period > nSamples:
break
h[i] = 1 # np.random.normal() ** 2
i += period
frame = i // framePeriod
fh = ssp.Frame(h, size=frameSize, period=framePeriod, pad=opt.padding)
gl = ssp.MinPhaseGlottis()
for i in range(len(fh)):
# This is minimum phase; the glotter will invert if required
gl.setpolepair(np.abs(magni[frame]), theta[frame])
fh[i] = gl.glotter(fh[i])
if linalg.norm(fh[i]) > 1e-6:
fh[i] *= np.sqrt(len(fh[i])) / linalg.norm(fh[i])
weight = np.sqrt(hnr[i] / (hnr[i] + 1))
fh[i] *= weight
if (opt.graphic == "h"):
fig = ssp.Figure(1, 1)
hPlot = fig.subplot()
hPlot.plot(h, 'r')
fig.show()
# Noise part
n = np.random.normal(size=nSamples)
zero = ssp.parameter("NoiseZero", 1.0)
n = ssp.ZeroFilter(n, zero) # Include the radiation impedance
fn = ssp.Frame(n, size=frameSize, period=framePeriod, pad=opt.padding)
for i in range(len(fn)):
fn[i] *= np.sqrt(1.0 / (hnr[i] + 1))
# Combination
assert(len(fh) == len(fn))
hgain = ssp.parameter("HGain", 1.0)
e = fn + fh * hgain
hnw = np.hanning(frameSize)
for i in range(len(e)):
ep = ssp.Window(e[i], hnw)
#ep = e[i]
eac = ssp.Autocorrelation(ep)
ea, eg = ssp.ARLevinson(eac, order=lpOrder[r])
e[i] = ssp.ARExcitation(e[i], ea, eg)
else:
print("Unknown synthesis method")
exit
if opt.excitation:
s = e.flatten('C')/frameSize
else:
s = ssp.ARResynthesis(e, ar, g)
if opt.ola:
# Asymmetric window for OLA
sw = np.hanning(frameSize+1)
sw = np.delete(sw, -1)
s = ssp.Window(s, sw)
s = ssp.OverlapAdd(s)
else:
s = s.flatten('C')
gain = ssp.parameter("Gain", 1.0)
return s * gain
def decode(tuple):
"""
Decode a speech waveform.
"""
(ark, g, pitch, hnr) = tuple
print("Frame padding:", opt.padding)
nFrames = len(ark)
assert (len(g) == nFrames)
assert (len(pitch) == nFrames)
assert (len(hnr) == nFrames)
# The original framer padded the ends so the number of samples to
# synthesise is a bit less than you might think
if opt.ola:
frameSize = framePeriod * 2
nSamples = framePeriod * (nFrames - 1)
else:
frameSize = framePeriod
nSamples = frameSize * (nFrames - 1)
ex = opt.glottal
if opt.glottal == 'cepgm' and (opt.encode or opt.decode or opt.pitch):
order = ark.shape[-1] - 2
ar = ark[:, 0:order]
theta = ark[:, -2]
magni = np.exp(ark[:, -1])
else:
ar = ark
# Use the original AR residual; it should be a very good reconstruction.
if ex == 'ar':
e = ssp.ARExcitation(f, ar, g)
# Just noise. This is effectively a whisper synthesis.
elif ex == 'noise':
e = np.random.normal(size=(nFrames, frameSize))
# Just harmonics, and with a fixed F0. This is the classic robot
# synthesis.
elif ex == 'robot':
ew = np.zeros(nSamples)
period = int(1.0 / 200 * r)
for i in range(0, len(ew), period):
ew[i] = period
e = ssp.Frame(ew, size=frameSize, period=framePeriod)
# Synthesise harmonics plus noise in the ratio suggested by the HNR.
elif ex == 'synth':
# Harmonic part
mperiod = int(1.0 / np.mean(pitch) * r)
gm = ssp.GlottalModel(ssp.parameter('Pulse', 'impulse'))
pr, pg = ssp.pulse_response(gm, pcm, period=mperiod, order=lpOrder[r])
h = np.zeros(nSamples)
i = 0
frame = 0
while i < nSamples and frame < len(pitch):
period = int(1.0 / pitch[frame] * r)
if i + period > nSamples:
break
weight = np.sqrt(hnr[frame] / (hnr[frame] + 1))
h[i:i + period] = gm.pulse(period, pcm) * weight
i += period
frame = i // framePeriod
h = ssp.ARExcitation(h, pr, 1.0)
fh = ssp.Frame(h, size=frameSize, period=framePeriod, pad=opt.padding)
# Noise part
n = np.random.normal(size=nSamples)
n = ssp.ZeroFilter(n, 1.0) # Include the radiation impedance
fn = ssp.Frame(n, size=frameSize, period=framePeriod, pad=opt.padding)
for i in range(len(fn)):
fn[i] *= np.sqrt(1.0 / (hnr[i] + 1))
hgain = ssp.parameter("HGain", 1.0)
e = fn + fh * hgain
# Like harmonics plus noise, but with explicit sinusoids instead of time
# domain impulses.
elif ex == 'sine':
order = 20
sine = ssp.Harmonics(r, order)
h = np.zeros(nSamples)
for i in range(0, len(h) - framePeriod, framePeriod):
frame = i // framePeriod
period = int(1.0 / pitch[frame] * r)
weight = np.sqrt(hnr[frame] / (hnr[frame] + 1))
h[i:i + framePeriod] = (sine.sample(pitch[frame], framePeriod) *
weight)
fh = ssp.Frame(h, size=frameSize, period=framePeriod, pad=opt.padding)
n = np.random.normal(size=nSamples)
fn = ssp.Frame(n, size=frameSize, period=framePeriod, pad=opt.padding)
for i in range(len(fn)):
fn[i] *= np.sqrt(1.0 / (hnr[i] + 1))
e = fn + fh * 10
# High order linear prediction. Synthesise the harmonics using noise to
# excite a high order polynomial with roots resembling harmonics.
elif ex == 'holp':
# Some noise
n = np.random.normal(size=nSamples)
fn = ssp.Frame(n, size=frameSize, period=framePeriod)
# Use the noise to excite a high order AR model
fh = np.ndarray(fn.shape)
for i in range(len(fn)):
hoar = ssp.ARHarmonicPoly(pitch[i], r, 0.7)
fh[i] = ssp.ARResynthesis(fn[i], hoar, 1.0 / linalg.norm(hoar)**2)
print(i, pitch[i], linalg.norm(hoar), np.min(fh[i]), np.max(fh[i]))
print(' ', np.min(hoar), np.max(hoar))
# fh[i] *= np.sqrt(r / pitch[i]) / linalg.norm(fh[i])
# fh[i] *= np.sqrt(hnr[i] / (hnr[i] + 1))
# Weight the noise as for the other methods
for i in range(len(fn)):
fn[i] *= np.sqrt(1.0 / (hnr[i] + 1))
e = fh # fn + fh*30
# Shaped excitation. The pulses are shaped by a filter to have a
# rolloff, then added to the noise. The resulting signal is
# flattened using AR.
elif ex == 'shaped':
# Harmonic part
gm = ssp.GlottalModel(ssp.parameter('Pulse', 'impulse'))
gm.angle = pcm.hertz_to_radians(np.mean(pitch) * 0.5)
h = np.zeros(nSamples)
i = 0
frame = 0
while i < nSamples and frame < len(pitch):
period = int(1.0 / pitch[frame] * r)
if i + period > nSamples:
break
weight = np.sqrt(hnr[frame] / (hnr[frame] + 1))
h[i:i + period] = gm.pulse(period, pcm) * weight
i += period
frame = i // framePeriod
# Filter to mimic the glottal pulse
hfilt = ssp.parameter("HFilt", None)
hpole1 = ssp.parameter("HPole1", 0.98)
hpole2 = ssp.parameter("HPole2", 0.8)
angle = pcm.hertz_to_radians(np.mean(pitch)) * ssp.parameter(
"Angle", 1.0)
if hfilt == 'pp':
h = ssp.ZeroFilter(h, 1.0)
h = ssp.PolePairFilter(h, hpole1, angle)
fh = ssp.Frame(h, size=frameSize, period=framePeriod)
# Noise part
n = np.random.normal(size=nSamples)
zero = ssp.parameter("NoiseZero", 1.0)
n = ssp.ZeroFilter(n, zero) # Include the radiation impedance
npole = ssp.parameter("NPole", None)
nf = ssp.parameter("NoiseFreq", 4000)
if npole is not None:
n = ssp.PolePairFilter(n, npole, pcm.hertz_to_radians(nf))
fn = ssp.Frame(n, size=frameSize, period=framePeriod, pad=opt.padding)
for i in range(len(fn)):
fn[i] *= np.sqrt(1.0 / (hnr[i] + 1))
# Combination
assert (len(fh) == len(fn))
hgain = ssp.parameter("HGain", 1.0)
e = fn + fh * hgain
hnw = np.hanning(frameSize)
for i in range(len(e)):
ep = ssp.Window(e[i], hnw)
#ep = e[i]
eac = ssp.Autocorrelation(ep)
ea, eg = ssp.ARLevinson(eac, order=lpOrder[r])
e[i] = ssp.ARExcitation(e[i], ea, eg)
elif ex == 'ceplf':
omega, alpha = ssp.glottal_pole_lf(f,
pcm,
pitch,
hnr,
visual=(opt.graphic == "ceplf"))
epsilon = ssp.parameter("Epsilon", 5000.0)
h = np.zeros(nSamples)
i = 0
frame = 0
while i < nSamples and frame < len(pitch):
period = int(1.0 / pitch[frame] * r)
if i + period > nSamples:
break
weight = np.sqrt(hnr[frame] / (hnr[frame] + 1))
pu = np.zeros((period))
T0 = pcm.period_to_seconds(period)
print(T0, )
Te = ssp.lf_te(T0, alpha[frame], omega[frame], epsilon)
if Te:
pu = ssp.pulse_lf(pu, T0, Te, alpha[frame], omega[frame],
epsilon)
h[i:i + period] = pu * weight
i += period
frame = i // framePeriod
fh = ssp.Frame(h, size=frameSize, period=framePeriod, pad=opt.padding)
# Noise part
n = np.random.normal(size=nSamples)
zero = ssp.parameter("NoiseZero", 1.0)
n = ssp.ZeroFilter(n, zero) # Include the radiation impedance
fn = ssp.Frame(n, size=frameSize, period=framePeriod, pad=opt.padding)
for i in range(len(fn)):
fn[i] *= np.sqrt(1.0 / (hnr[i] + 1))
# Combination
assert (len(fh) == len(fn))
hgain = ssp.parameter("HGain", 1.0)
e = fn + fh * hgain
hnw = np.hanning(frameSize)
for i in range(len(e)):
ep = ssp.Window(e[i], hnw)
#ep = e[i]
eac = ssp.Autocorrelation(ep)
ea, eg = ssp.ARLevinson(eac, order=lpOrder[r])
e[i] = ssp.ARExcitation(e[i], ea, eg)
elif ex == 'cepgm':
# Infer the unstable poles via complex cepstrum, then build an explicit
# glottal model.
if not (opt.encode or opt.decode or opt.pitch):
theta, magni = ssp.glottal_pole_gm(f,
pcm,
pitch,
hnr,
visual=(opt.graphic == "cepgm"))
h = np.zeros(nSamples)
i = 0
frame = 0
while i < nSamples and frame < len(pitch):
period = int(1.0 / pitch[frame] * r)
if i + period > nSamples:
break
h[i] = 1 # np.random.normal() ** 2
i += period
frame = i // framePeriod
fh = ssp.Frame(h, size=frameSize, period=framePeriod, pad=opt.padding)
gl = ssp.MinPhaseGlottis()
for i in range(len(fh)):
# This is minimum phase; the glotter will invert if required
gl.setpolepair(np.abs(magni[frame]), theta[frame])
fh[i] = gl.glotter(fh[i])
if linalg.norm(fh[i]) > 1e-6:
fh[i] *= np.sqrt(len(fh[i])) / linalg.norm(fh[i])
weight = np.sqrt(hnr[i] / (hnr[i] + 1))
fh[i] *= weight
if (opt.graphic == "h"):
fig = ssp.Figure(1, 1)
hPlot = fig.subplot()
hPlot.plot(h, 'r')
fig.show()
# Noise part
n = np.random.normal(size=nSamples)
zero = ssp.parameter("NoiseZero", 1.0)
n = ssp.ZeroFilter(n, zero) # Include the radiation impedance
fn = ssp.Frame(n, size=frameSize, period=framePeriod, pad=opt.padding)
for i in range(len(fn)):
fn[i] *= np.sqrt(1.0 / (hnr[i] + 1))
# Combination
assert (len(fh) == len(fn))
hgain = ssp.parameter("HGain", 1.0)
e = fn + fh * hgain
hnw = np.hanning(frameSize)
for i in range(len(e)):
ep = ssp.Window(e[i], hnw)
#ep = e[i]
eac = ssp.Autocorrelation(ep)
ea, eg = ssp.ARLevinson(eac, order=lpOrder[r])
e[i] = ssp.ARExcitation(e[i], ea, eg)
else:
print("Unknown synthesis method")
exit
if opt.excitation:
s = e.flatten('C') / frameSize
else:
s = ssp.ARResynthesis(e, ar, g)
if opt.ola:
# Asymmetric window for OLA
sw = np.hanning(frameSize + 1)
sw = np.delete(sw, -1)
s = ssp.Window(s, sw)
s = ssp.OverlapAdd(s)
else:
s = s.flatten('C')
gain = ssp.parameter("Gain", 1.0)
return s * gain
#
from optparse import OptionParser
op = OptionParser()
(option, arg) = op.parse_args()
if (len(arg) < 1):
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)
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)
a = ssp.Autocorrelation(a, input='psd')
a, g = ssp.ARLevinson(a, lpOrder)
def decode((ar, g, pitch, hnr)):
"""
Decode a speech waveform.
"""
nFrames = len(ar)
assert(len(g) == nFrames)
assert(len(pitch) == nFrames)
assert(len(hnr) == nFrames)
# The original framer padded the ends so the number of samples to
# synthesise is a bit less than you might think
if opt.ola:
frameSize = framePeriod * 2
nSamples = framePeriod * (nFrames-1)
else:
frameSize = framePeriod
nSamples = frameSize * (nFrames-1)
ex = ssp.parameter('Excitation', 'synth')
# Use the original AR residual; it should be a very good
# reconstruction.
if ex == 'ar':
e = ssp.ARExcitation(f, ar, g)
# Just noise. This is effectively a whisper synthesis.
elif ex == 'noise':
e = np.random.normal(size=f.shape)
# Just harmonics, and with a fixed F0. This is the classic robot
# syntheisis.
elif ex == 'robot':
ew = np.zeros(nSamples)
period = int(1.0 / 200 * r)
for i in range(0, len(ew), period):
ew[i] = period
e = ssp.Frame(ew, size=frameSize, period=framePeriod)
# Synthesise harmonics plus noise in the ratio suggested by the
# HNR.
elif ex == 'synth':
# Harmonic part
mperiod = int(1.0 / np.mean(pitch) * r)
gm = ssp.GlottalModel(ssp.parameter('Pulse', 'impulse'))
pr, pg = ssp.pulse_response(gm, pcm, period=mperiod, order=lpOrder[r])
h = np.zeros(nSamples)
i = 0
frame = 0
while i < nSamples and frame < len(pitch):
period = int(1.0 / pitch[frame] * r)
if i + period > nSamples:
break
weight = np.sqrt(hnr[frame] / (hnr[frame] + 1))
h[i:i+period] = gm.pulse(period, pcm) * weight
i += period
frame = i // framePeriod
h = ssp.ARExcitation(h, pr, 1.0)
fh = ssp.Frame(h, size=frameSize, period=framePeriod)
# Noise part
n = np.random.normal(size=nSamples)
n = ssp.ZeroFilter(n, 1.0) # Include the radiation impedance
fn = ssp.Frame(n, size=frameSize, period=framePeriod)
for i in range(len(fn)):
fn[i] *= np.sqrt(1.0 / (hnr[i] + 1))
hgain = ssp.parameter("HGain", 1.0)
e = fn + fh * hgain
# Like harmonics plus noise, but with explicit sinusoids instead
# of time domain impulses.
elif ex == 'sine':
order = 20
sine = ssp.Harmonics(r, order)
h = np.zeros(nSamples)
for i in range(0, len(h)-framePeriod, framePeriod):
frame = i // framePeriod
period = int(1.0 / pitch[frame] * r)
weight = np.sqrt(hnr[frame] / (hnr[frame] + 1))
h[i:i+framePeriod] = ( sine.sample(pitch[frame], framePeriod)
* weight )
fh = ssp.Frame(h, size=frameSize, period=framePeriod)
n = np.random.normal(size=nSamples)
fn = ssp.Frame(n, size=frameSize, period=framePeriod)
for i in range(len(fn)):
fn[i] *= np.sqrt(1.0 / (hnr[i] + 1))
e = fn + fh*10
# High order linear prediction. Synthesise the harmonics using
# noise to excite a high order polynomial with roots resembling
# harmonics.
elif ex == 'holp':
# Some noise
n = np.random.normal(size=nSamples)
fn = ssp.Frame(n, size=frameSize, period=framePeriod)
# Use the noise to excite a high order AR model
fh = np.ndarray(fn.shape)
for i in range(len(fn)):
hoar = ssp.ARHarmonicPoly(pitch[i], r, 0.7)
fh[i] = ssp.ARResynthesis(fn[i], hoar, 1.0 / linalg.norm(hoar)**2)
print i, pitch[i], linalg.norm(hoar), np.min(fh[i]), np.max(fh[i])
print ' ', np.min(hoar), np.max(hoar)
# fh[i] *= np.sqrt(r / pitch[i]) / linalg.norm(fh[i])
# fh[i] *= np.sqrt(hnr[i] / (hnr[i] + 1))
# Weight the noise as for the other methods
for i in range(len(fn)):
fn[i] *= np.sqrt(1.0 / (hnr[i] + 1))
e = fh # fn + fh*30
# Shaped excitation. The pulses are shaped by a filter to have a
# rolloff, then added to the noise. The resulting signal is
# flattened using AR.
elif ex == 'shaped':
# Harmonic part
gm = ssp.GlottalModel(ssp.parameter('Pulse', 'impulse'))
gm.angle = pcm.hertz_to_radians(np.mean(pitch)*0.5)
h = np.zeros(nSamples)
i = 0
frame = 0
while i < nSamples and frame < len(pitch):
period = int(1.0 / pitch[frame] * r)
if i + period > nSamples:
break
weight = np.sqrt(hnr[frame] / (hnr[frame] + 1))
h[i:i+period] = gm.pulse(period, pcm) * weight
i += period
frame = i // framePeriod
# Filter to mimic the glottal pulse
hfilt = ssp.parameter("HFilt", None)
hpole1 = ssp.parameter("HPole1", 0.98)
hpole2 = ssp.parameter("HPole2", 0.8)
angle = pcm.hertz_to_radians(np.mean(pitch)) * ssp.parameter("Angle", 1.0)
# Noise part
n = np.random.normal(size=nSamples)
zero = ssp.parameter("NoiseZero", 1.0)
n = ssp.ZeroFilter(n, zero) # Include the radiation impedance
npole = ssp.parameter("NPole", None)
nf = ssp.parameter("NoiseFreq", 4000)
if npole is not None:
n = ssp.PolePairFilter(n, npole, pcm.hertz_to_radians(nf))
fn = ssp.Frame(n, size=frameSize, period=framePeriod)
for i in range(len(fn)):
fn[i] *= np.sqrt(1.0 / (hnr[i] + 1))
# Combination
assert(len(fh) == len(fn))
hgain = ssp.parameter("HGain", 1.0)
e = fn + fh * hgain
hnw = np.hanning(frameSize)
for i in range(len(e)):
ep = ssp.Window(e[i], hnw)
#ep = e[i]
eac = ssp.Autocorrelation(ep)
ea, eg = ssp.ARLevinson(eac, order=lpOrder[r])
e[i] = ssp.ARExcitation(e[i], ea, eg)
else:
print "Unknown synthesis method"
exit
if opt.excitation:
s = e.flatten('C')/frameSize
else:
s = ssp.ARResynthesis(e, ar, g)
if opt.ola:
# Asymmetric window for OLA
sw = np.hanning(frameSize+1)
sw = np.delete(sw, -1)
s = ssp.Window(s, sw)
s = ssp.OverlapAdd(s)
else:
s = s.flatten('C')
gain = ssp.parameter("Gain", 1.0)
return s * gain
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)
a = ssp.Autocorrelation(a, input='psd')
a, g = ssp.ARLevinson(a, lpOrder)
a = ssp.AutocorrelationAllPassWarp(a, alpha=ssp.mel[pcm.rate],
def decode((ar, g, pitch, hnr)):
"""
Decode a speech waveform.
"""
nFrames = len(ar)
assert(len(g) == nFrames)
assert(len(pitch) == nFrames)
assert(len(hnr) == nFrames)
# The original framer padded the ends so the number of samples to
# synthesise is a bit less than you might think
if opt.ola:
frameSize = framePeriod * 2
nSamples = framePeriod * (nFrames-1)
else:
frameSize = framePeriod
nSamples = frameSize * (nFrames-1)
ex = ssp.parameter('Excitation', 'synth')
# Use the original AR residual; it should be a very good
# reconstruction.
if ex == 'ar':
e = ssp.ARExcitation(f, ar, g)
# Just noise. This is effectively a whisper synthesis.
elif ex == 'noise':
e = np.random.normal(size=f.shape)
# Just harmonics, and with a fixed F0. This is the classic robot
# syntheisis.
elif ex == 'robot':
ew = np.zeros(nSamples)
period = int(1.0 / 200 * r)
for i in range(0, len(ew), period):
ew[i] = period
e = ssp.Frame(ew, size=frameSize, period=framePeriod)
# Synthesise harmonics plus noise in the ratio suggested by the
# HNR.
elif ex == 'synth':
# Harmonic part
mperiod = int(1.0 / np.mean(pitch) * r)
gm = ssp.GlottalModel(ssp.parameter('Pulse', 'impulse'))
pr, pg = ssp.pulse_response(gm, pcm, period=mperiod, order=lpOrder[r])
h = np.zeros(nSamples)
i = 0
frame = 0
while i < nSamples and frame < len(pitch):
period = int(1.0 / pitch[frame] * r)
if i + period > nSamples:
break
weight = np.sqrt(hnr[frame] / (hnr[frame] + 1))
h[i:i+period] = gm.pulse(period, pcm) * weight
i += period
frame = i // framePeriod
h = ssp.ARExcitation(h, pr, 1.0)
fh = ssp.Frame(h, size=frameSize, period=framePeriod)
# Noise part
n = np.random.normal(size=nSamples)
n = ssp.ZeroFilter(n, 1.0) # Include the radiation impedance
fn = ssp.Frame(n, size=frameSize, period=framePeriod)
for i in range(len(fn)):
fn[i] *= np.sqrt(1.0 / (hnr[i] + 1))
hgain = ssp.parameter("HGain", 1.0)
e = fn + fh * hgain
# Like harmonics plus noise, but with explicit sinusoids instead
# of time domain impulses.
elif ex == 'sine':
order = 20
sine = ssp.Harmonics(r, order)
h = np.zeros(nSamples)
for i in range(0, len(h)-framePeriod, framePeriod):
frame = i // framePeriod
period = int(1.0 / pitch[frame] * r)
weight = np.sqrt(hnr[frame] / (hnr[frame] + 1))
h[i:i+framePeriod] = ( sine.sample(pitch[frame], framePeriod)
* weight )
fh = ssp.Frame(h, size=frameSize, period=framePeriod)
n = np.random.normal(size=nSamples)
fn = ssp.Frame(n, size=frameSize, period=framePeriod)
for i in range(len(fn)):
fn[i] *= np.sqrt(1.0 / (hnr[i] + 1))
e = fn + fh*10
# High order linear prediction. Synthesise the harmonics using
# noise to excite a high order polynomial with roots resembling
# harmonics.
elif ex == 'holp':
# Some noise
n = np.random.normal(size=nSamples)
fn = ssp.Frame(n, size=frameSize, period=framePeriod)
# Use the noise to excite a high order AR model
fh = np.ndarray(fn.shape)
for i in range(len(fn)):
hoar = ssp.ARHarmonicPoly(pitch[i], r, 0.7)
fh[i] = ssp.ARResynthesis(fn[i], hoar, 1.0 / linalg.norm(hoar)**2)
print i, pitch[i], linalg.norm(hoar), np.min(fh[i]), np.max(fh[i])
print ' ', np.min(hoar), np.max(hoar)
# fh[i] *= np.sqrt(r / pitch[i]) / linalg.norm(fh[i])
# fh[i] *= np.sqrt(hnr[i] / (hnr[i] + 1))
# Weight the noise as for the other methods
for i in range(len(fn)):
fn[i] *= np.sqrt(1.0 / (hnr[i] + 1))
e = fh # fn + fh*30
# Shaped excitation. The pulses are shaped by a filter to have a
# rolloff, then added to the noise. The resulting signal is
# flattened using AR.
elif ex == 'shaped':
# Harmonic part
gm = ssp.GlottalModel(ssp.parameter('Pulse', 'impulse'))
gm.angle = pcm.hertz_to_radians(np.mean(pitch)*0.5)
h = np.zeros(nSamples)
i = 0
frame = 0
while i < nSamples and frame < len(pitch):
period = int(1.0 / pitch[frame] * r)
if i + period > nSamples:
break
weight = np.sqrt(hnr[frame] / (hnr[frame] + 1))
h[i:i+period] = gm.pulse(period, pcm) * weight
i += period
frame = i // framePeriod
# Filter to mimic the glottal pulse
hfilt = ssp.parameter("HFilt", None)
hpole1 = ssp.parameter("HPole1", 0.98)
hpole2 = ssp.parameter("HPole2", 0.8)
angle = pcm.hertz_to_radians(np.mean(pitch)) * ssp.parameter("Angle", 1.0)
if hfilt == 'pp':
h = ssp.ZeroFilter(h, 1.0)
h = ssp.PolePairFilter(h, hpole1, angle)
if hfilt == 'g':
h = ssp.GFilter(h, hpole1, angle, hpole2)
if hfilt == 'p':
h = ssp.PFilter(h, hpole1, angle, hpole2)
fh = ssp.Frame(h, size=frameSize, period=framePeriod)
# Noise part
n = np.random.normal(size=nSamples)
zero = ssp.parameter("NoiseZero", 1.0)
n = ssp.ZeroFilter(n, zero) # Include the radiation impedance
npole = ssp.parameter("NPole", None)
nf = ssp.parameter("NoiseFreq", 4000)
if npole is not None:
n = ssp.PolePairFilter(n, npole, pcm.hertz_to_radians(nf))
fn = ssp.Frame(n, size=frameSize, period=framePeriod)
for i in range(len(fn)):
fn[i] *= np.sqrt(1.0 / (hnr[i] + 1))
# Combination
assert(len(fh) == len(fn))
hgain = ssp.parameter("HGain", 1.0)
e = fn + fh * hgain
hnw = np.hanning(frameSize)
for i in range(len(e)):
ep = ssp.Window(e[i], hnw)
#ep = e[i]
eac = ssp.Autocorrelation(ep)
ea, eg = ssp.ARLevinson(eac, order=lpOrder[r])
e[i] = ssp.ARExcitation(e[i], ea, eg)
else:
print "Unknown synthesis method"
exit
if opt.excitation:
s = e.flatten('C')/frameSize
else:
s = ssp.ARResynthesis(e, ar, g)
if opt.ola:
# Asymmetric window for OLA
sw = np.hanning(frameSize+1)
sw = np.delete(sw, -1)
s = ssp.Window(s, sw)
s = ssp.OverlapAdd(s)
else:
s = s.flatten('C')
gain = ssp.parameter("Gain", 1.0)
return s * gain