def main():
Ns = 44100*400 #44100*60*3 #3000000
print("Loading data")
# load data
#fname = 'sandbox/test_cpea_20120602_part3.wav' # delay = 45.2 ms = 1993 samples (from Friture cross-correlation measurements)
fname = 'sandbox/gary_moore_separate_chambre.wav'
#fname = 'sandbox/gary_moore_loner_chambre.wav'
#fname = 'sandbox/test_cpea_20120602.wav'
#data, fs, enc = wavread(fname)
SAMPLING_RATE, data = wavfile.read(fname)
print(data.shape, data.dtype)
data = data.astype(np.float64) # convert from integer to floats for computations
ht = None
#data = data[:Ns,:]
data /= 2**16
#data = data[::-1,:]
# original signal
x = data[:Ns,0]
# measured signal
m = data[:Ns,1]
print("Data loaded")
# We will decimate several times
# no decimation => 1/fs = 23 µs resolution
# 1 ms resolution => fs = 1000 Hz is enough => can divide the sampling rate by 44 !
# if I decimate 2 times (2**2 = 4 => 0.092 ms (3 cm) resolution)!
# if I decimate 3 times (2**3 = 8 => 0.184 ms (6 cm) resolution)!
# if I decimate 4 times (2**4 = 16 => 0.368 ms (12 cm) resolution)!
# if I decimate 5 times (2**5 = 32 => 0.7 ms (24 cm) resolution)!
# (actually, I could fit a gaussian on the cross-correlation peak to get
# higher resolution even at low sample rates)
Ndec = 2
subsampled_sampling_rate = SAMPLING_RATE/2**(Ndec)
[bdec, adec] = generated_filters.params['dec']
zfs0 = subsampler_filtic(Ndec, bdec, adec)
zfs1 = subsampler_filtic(Ndec, bdec, adec)
delayrange_s = DEFAULT_DELAYRANGE # confidence range
# separate the channels
x0 = x
x1 = m
#subsample them
print("subsampling x0")
x0_dec, zfs0 = subsampler(Ndec, bdec, adec, x0, zfs0)
print("subsampling x1")
x1_dec, zfs1 = subsampler(Ndec, bdec, adec, x1, zfs1)
time = 2*delayrange_s
length = time*subsampled_sampling_rate
Ns_dec = Ns/2**(Ndec)
overlap = 0.5
Nb = int((1./overlap)*(Ns_dec/length - 1))
old_Xcorr = np.zeros((length), np.complex128)
print("time =", time)
print("length =", length)
print("Ns_dec = ", Ns_dec)
print("Nb =", Nb)
# Hann window to mitigate non-periodicity effects
window = numpy.hanning(length)
#l2 = length/2 + 1
# Hann window to mitigate non-periodicity effects
#window2 = numpy.hanning(l2)
#f1 = plt.figure()
#plt1 = f1.add_subplot(1,1,1)
#f2 = plt.figure()
#plt2 = f2.add_subplot(1,1,1)
for i in range(Nb):
print(i)
# retrieve last one-second of data
d0 = x0_dec[i*length*overlap:i*length*overlap + length]
d1 = x1_dec[i*length*overlap:i*length*overlap + length]
std0 = numpy.std(d0)
std1 = numpy.std(d1)
if std0>0. and std1>0.:
Xcorr = generalized_cross_correlation(d0, d1)
#plt1.plot(Xcorr)
#plt.figure()
#plt.plot(Xcorr)
#plt.draw()
if old_Xcorr != None and old_Xcorr.shape == Xcorr.shape:
# smoothing
alpha = 0.3
smoothed_Xcorr = alpha*Xcorr + (1. - alpha)*old_Xcorr
else:
smoothed_Xcorr = Xcorr
#plt2.plot(Xcorr)
#plt.draw()
absXcorr = numpy.abs(smoothed_Xcorr)
ind = argmax(absXcorr)
#h2_ = fft(absXcorr**2)
#N = len(h2_)
#h2_[N/50:-N/50] = 0.
#h2 = ifft(h2_)
#ind = argmax(h2)
#plt.figure()
#plt.plot(Xcorr)
#plt.plot(absXcorr)
#plt.draw()
# normalize
#Xcorr_max_norm = Xcorr_unweighted[ind]/(d0.size*std0*std1)
Xcorr_extremum = smoothed_Xcorr[ind]
Xcorr_max_norm = abs(smoothed_Xcorr[ind])/(3*numpy.std(smoothed_Xcorr))
delay_ms = 1e3*float(ind)/subsampled_sampling_rate
# delays larger than the half of the window most likely are actually negative
if delay_ms > 1e3*time/2.:
delay_ms -= 1e3*time
# store for smoothing
old_Xcorr = smoothed_Xcorr
else:
delay_ms = 0.
Xcorr_max_norm = 0.
Xcorr_extremum = 0.
# debug wrong phase detection
#if Xcorr_extremum < 0.:
# numpy.save("Xcorr.npy", Xcorr)
# numpy.save("smoothed_Xcorr.npy", smoothed_Xcorr)
c = 340. # speed of sound, in meters per second (approximate)
distance_m = delay_ms*1e-3*c
# home-made measure of the significance
slope = 0.12
p = 3
x = (Xcorr_max_norm>1.)*(Xcorr_max_norm-1.)
x = (slope*x)**p
correlation = int((x/(1. + x))*100)
delay_message = "%.1f ms = %.2f m" %(delay_ms, distance_m)
correlation_message = "%d%%" %(correlation)
if Xcorr_extremum >= 0:
polarity_message = "In-phase"
else:
polarity_message = "Reversed phase"
print(ind, delay_message, correlation_message, polarity_message)
plt.show()
from friture import generated_filters
from friture.audiobackend import SAMPLING_RATE
from friture.delay_estimator import subsampler, subsampler_filtic
Ns = int(1e4)
#y = np.random.rand(Ns)
f = 2e1
t = np.linspace(0, float(Ns)/SAMPLING_RATE, Ns)
y = np.cos(2.*np.pi*f*t)
Ndec = 2
subsampled_sampling_rate = SAMPLING_RATE/2**(Ndec)
[bdec, adec] = generated_filters.PARAMS['dec']
zfs0 = subsampler_filtic(Ndec, bdec, adec)
Nb = 10
l = int(Ns/Nb)
Nb0 = Nb/2
print("Nb0 =", Nb0, "Nb1 =", Nb - Nb0)
print("subsample first parts")
for i in range(Nb0):
ydec, zfs0 = subsampler(Ndec, bdec, adec, y[i*l:(i+1)*l], zfs0)
if i == 0:
y_dec = ydec
else:
y_dec = np.append(y_dec, ydec)
def main():
Ns = 44100 * 400 #44100*60*3 #3000000
print("Loading data")
# load data
#fname = 'sandbox/test_cpea_20120602_part3.wav' # delay = 45.2 ms = 1993 samples (from Friture cross-correlation measurements)
fname = 'sandbox/gary_moore_separate_chambre.wav'
#fname = 'sandbox/gary_moore_loner_chambre.wav'
#fname = 'sandbox/test_cpea_20120602.wav'
#data, fs, enc = wavread(fname)
SAMPLING_RATE, data = wavfile.read(fname)
print(data.shape, data.dtype)
data = data.astype(
np.float64) # convert from integer to floats for computations
ht = None
#data = data[:Ns,:]
data /= 2**16
#data = data[::-1,:]
# original signal
x = data[:Ns, 0]
# measured signal
m = data[:Ns, 1]
print("Data loaded")
# We will decimate several times
# no decimation => 1/fs = 23 µs resolution
# 1 ms resolution => fs = 1000 Hz is enough => can divide the sampling rate by 44 !
# if I decimate 2 times (2**2 = 4 => 0.092 ms (3 cm) resolution)!
# if I decimate 3 times (2**3 = 8 => 0.184 ms (6 cm) resolution)!
# if I decimate 4 times (2**4 = 16 => 0.368 ms (12 cm) resolution)!
# if I decimate 5 times (2**5 = 32 => 0.7 ms (24 cm) resolution)!
# (actually, I could fit a gaussian on the cross-correlation peak to get
# higher resolution even at low sample rates)
Ndec = 2
subsampled_sampling_rate = SAMPLING_RATE / 2**(Ndec)
[bdec, adec] = generated_filters.params['dec']
zfs0 = subsampler_filtic(Ndec, bdec, adec)
zfs1 = subsampler_filtic(Ndec, bdec, adec)
delayrange_s = DEFAULT_DELAYRANGE # confidence range
# separate the channels
x0 = x
x1 = m
#subsample them
print("subsampling x0")
x0_dec, zfs0 = subsampler(Ndec, bdec, adec, x0, zfs0)
print("subsampling x1")
x1_dec, zfs1 = subsampler(Ndec, bdec, adec, x1, zfs1)
time = 2 * delayrange_s
length = time * subsampled_sampling_rate
Ns_dec = Ns / 2**(Ndec)
overlap = 0.5
Nb = int((1. / overlap) * (Ns_dec / length - 1))
old_Xcorr = np.zeros((length), np.complex128)
print("time =", time)
print("length =", length)
print("Ns_dec = ", Ns_dec)
print("Nb =", Nb)
# Hann window to mitigate non-periodicity effects
window = numpy.hanning(length)
#l2 = length/2 + 1
# Hann window to mitigate non-periodicity effects
#window2 = numpy.hanning(l2)
#f1 = plt.figure()
#plt1 = f1.add_subplot(1,1,1)
#f2 = plt.figure()
#plt2 = f2.add_subplot(1,1,1)
for i in range(Nb):
print(i)
# retrieve last one-second of data
d0 = x0_dec[i * length * overlap:i * length * overlap + length]
d1 = x1_dec[i * length * overlap:i * length * overlap + length]
std0 = numpy.std(d0)
std1 = numpy.std(d1)
if std0 > 0. and std1 > 0.:
Xcorr = generalized_cross_correlation(d0, d1)
#plt1.plot(Xcorr)
#plt.figure()
#plt.plot(Xcorr)
#plt.draw()
if old_Xcorr != None and old_Xcorr.shape == Xcorr.shape:
# smoothing
alpha = 0.3
smoothed_Xcorr = alpha * Xcorr + (1. - alpha) * old_Xcorr
else:
smoothed_Xcorr = Xcorr
#plt2.plot(Xcorr)
#plt.draw()
absXcorr = numpy.abs(smoothed_Xcorr)
ind = argmax(absXcorr)
#h2_ = fft(absXcorr**2)
#N = len(h2_)
#h2_[N/50:-N/50] = 0.
#h2 = ifft(h2_)
#ind = argmax(h2)
#plt.figure()
#plt.plot(Xcorr)
#plt.plot(absXcorr)
#plt.draw()
# normalize
#Xcorr_max_norm = Xcorr_unweighted[ind]/(d0.size*std0*std1)
Xcorr_extremum = smoothed_Xcorr[ind]
Xcorr_max_norm = abs(
smoothed_Xcorr[ind]) / (3 * numpy.std(smoothed_Xcorr))
delay_ms = 1e3 * float(ind) / subsampled_sampling_rate
# delays larger than the half of the window most likely are actually negative
if delay_ms > 1e3 * time / 2.:
delay_ms -= 1e3 * time
# store for smoothing
old_Xcorr = smoothed_Xcorr
else:
delay_ms = 0.
Xcorr_max_norm = 0.
Xcorr_extremum = 0.
# debug wrong phase detection
#if Xcorr_extremum < 0.:
# numpy.save("Xcorr.npy", Xcorr)
# numpy.save("smoothed_Xcorr.npy", smoothed_Xcorr)
c = 340. # speed of sound, in meters per second (approximate)
distance_m = delay_ms * 1e-3 * c
# home-made measure of the significance
slope = 0.12
p = 3
x = (Xcorr_max_norm > 1.) * (Xcorr_max_norm - 1.)
x = (slope * x)**p
correlation = int((x / (1. + x)) * 100)
delay_message = "%.1f ms = %.2f m" % (delay_ms, distance_m)
correlation_message = "%d%%" % (correlation)
if Xcorr_extremum >= 0:
polarity_message = "In-phase"
else:
polarity_message = "Reversed phase"
print(ind, delay_message, correlation_message, polarity_message)
plt.show()