def get_fromants(sound, fs=11025):
# Read from file.
# spf = sound
x = sound
# Get Hamming window.
N = len(x)
w = np.hamming(N)
# Apply window and high pass filter.
x1 = x * w
x1 = lfilter([1.], [1., 0.63], x1)
# Get LPC.
ncoeff = 2 + fs / 1000
A, e, k = lpc(x1, ncoeff)
# Get roots.
rts = np.roots(A)
rts = [r for r in rts if np.imag(r) >= 0]
# Get angles.
angz = np.arctan2(np.imag(rts), np.real(rts))
# Fs = spf.getframerate() #Gregory comment
frqs = sorted(angz * (fs / (2 * math.pi)))
return frqs
def compute_formants(audio_buffer):
N = len(audio_buffer)
Fs = 8000 # sampling frequency
hamming_window = np.hamming(N)
window = audio_buffer * hamming_window
# Apply a pre-emphasis filter; this amplifies high-frequency components and attenuates low-frequency components.
# The purpose in voice processing is to remove noise.
filtered_buffer = lfilter([1], [1., 0.63], window)
ncoeff = 2 + Fs / 1000
A, e, k = lpc(filtered_buffer, ncoeff)
roots = np.roots(A)
roots = [r for r in roots if np.imag(r) >= 0]
angz = np.arctan2(np.imag(roots), np.real(roots))
unsorted_freqs = angz * (Fs / (2 * math.pi))
freqs = sorted(unsorted_freqs)
# also get the indices so that we can get the bandwidths in the same order
indices = np.argsort(unsorted_freqs)
sorted_roots = np.asarray(roots)[indices]
#compute the bandwidths of each formant
bandwidths = -1/2. * (Fs/(2*math.pi))*np.log(np.abs(sorted_roots))
return freqs, bandwidths
def formant(arr):
# entered array is the recorded array using the data acquisation
#duration = 2;
#fs = 11025;
#sd.default.samplerate = fs;
#print("Speak now")
#arr = sd.rec(duration*fs, samplerate=fs, channels=1, blocking=True)
#from scikits.talkbox import lpc
# applying hamming window
Fs = 11025
#rr = [1,2,3,4,5,6,7,7,88]
N = len(arr)
window = numpy.hamming(N)
arr1 = arr * window
arr1 = lfilter([1], [1., 0.63], arr1)
n_coeff = int(2 + Fs / 1000) #no of coefficients
A, e, k = (lpc(arr1, n_coeff)) #applying lpc
#A.numpolyz(A)
rts = numpy.roots(A)
rts = [r for r in rts if numpy.imag(r) >= 0] #only positive roots
angz = numpy.arctan2(numpy.imag(rts), numpy.real(rts)) # taking angles
for_freq = sorted(angz * (Fs / (2 * math.pi)))
return (for_freq)
def get_formants(file_path):
# Read from file.
spf = wave.open(file_path, 'r') # http://www.linguistics.ucla.edu/people/hayes/103/Charts/VChart/ae.wav
# Get file as numpy array.
x = spf.readframes(-1)
x = numpy.fromstring(x, 'Int16')
# Get Hamming window.
N = len(x)
w = numpy.hamming(N)
# Apply window and high pass filter.
x1 = x * w
x1 = lfilter([1], [1., 0.63], x1)
# Get LPC.
Fs = spf.getframerate()
ncoeff = 2 + Fs / 1000
A, e, k = lpc(x1, ncoeff)
# Get roots.
rts = numpy.roots(A)
rts = [r for r in rts if numpy.imag(r) >= 0]
# Get angles.
angz = numpy.arctan2(numpy.imag(rts), numpy.real(rts))
# Get frequencies.
Fs = spf.getframerate()
frqs = sorted(angz * (Fs / (2 * math.pi)))
return frqs
def feature_extraction_lp_group_delay(y, fs=44100, statistics=True, lpgd_params=None, win_params=None):
eps = numpy.spacing(1)
nfft = lpgd_params['nfft']
lp_order = lpgd_params['lp_order']
y = y + eps
frames = segment_axis(y, win_params['win_length'], win_params['hop_length']);
print 'frames : ' + str(frames.shape)
a,e,k = lpc(frames, lp_order)
print 'a : ' + str(a.shape)
A = fft(a, nfft)
A = 1/A
phaseA = numpy.unwrap(numpy.angle(A))
print 'phaseA: ' + str(phaseA.shape)
phaseA = phaseA[:,0:nfft/2]
print 'phaseA: ' + str(phaseA.shape)
tauA = -1 * numpy.diff(phaseA)
print 'tauA' + str(tauA.shape)
# tau = numpy.concatenate((tauA, tauA[-1]))
# tau = tau
feature_matrix = tauA
feature_matrix = dct(feature_matrix, n=20)
print 'fm: ' + str(feature_matrix.shape)
# Collect into data structure
if statistics:
return {
'feat': feature_matrix,
'stat': {
'mean': numpy.mean(feature_matrix, axis=0),
'std': numpy.std(feature_matrix, axis=0),
'N': feature_matrix.shape[0],
'S1': numpy.sum(feature_matrix, axis=0),
'S2': numpy.sum(feature_matrix ** 2, axis=0),
}
}
else:
return {
'feat': feature_matrix}
def do_lpc(spec, order, error_normal=False):
coeff, error, k = lpc(spec, order, axis=0)
if error_normal:
error = np.reshape(error, (1, len(error)))
error = np.repeat(error, order + 1, axis=0)
return coeff / error
else:
return coeff[1:, :]
def feature_extraction_lp_group_delay(y, fs=44100, statistics=True, lpgd_params=None, win_params=None):
eps = numpy.spacing(1)
nfft = lpgd_params['nfft']
lp_order = lpgd_params['lp_order']
y = y + eps
frames = segment_axis(y, win_params['win_length'], win_params['hop_length']);
print 'frames : ' + str(frames.shape)
a,e,k = lpc(frames, lp_order)
print 'a : ' + str(a.shape)
A = fft(a, nfft)
A = 1/A
phaseA = numpy.unwrap(numpy.angle(A))
print 'phaseA: ' + str(phaseA.shape)
phaseA = phaseA[:,0:nfft/2]
print 'phaseA: ' + str(phaseA.shape)
tauA = -1 * numpy.diff(phaseA)
print 'tauA' + str(tauA.shape)
# tau = numpy.concatenate((tauA, tauA[-1]))
# tau = tau
feature_matrix = tauA
feature_matrix = dct(feature_matrix, n=20)
print 'fm: ' + str(feature_matrix.shape)
# Collect into data structure
if statistics:
return {
'feat': feature_matrix,
'stat': {
'mean': numpy.mean(feature_matrix, axis=0),
'std': numpy.std(feature_matrix, axis=0),
'N': feature_matrix.shape[0],
'S1': numpy.sum(feature_matrix, axis=0),
'S2': numpy.sum(feature_matrix ** 2, axis=0),
}
}
else:
return {
'feat': feature_matrix}
def lpc_filter(x, lpc_mem):
# linear prediction coefficients
a = lpc(x, len(lpc_mem))
coeff = np.asarray(a[0])
est_frames, lpc_mem = lfilter(0 - coeff, 1, x, -1, lpc_mem)
res_frames = x - est_frames
return coeff, lpc_mem, res_frames
def LSF(arr):
Fs = 11025
#rr = [1,2,3,4,5,6,7,7,88]
N = len(arr)
window = np.hamming(N)
arr1 = arr * window
arr1 = sp.signal.lfilter([1], [1., 0.63], arr1)
n_coeff = int(2 + Fs / 1000) #no of coefficients
A, e, k = (lpc(arr1, n_coeff)) #applying lpc
lsfs = poly2lsf(A)
return lsfs
def get_formants(audio): """ Calculate the formant frequencies of the audio segment. This method was taken from http://stackoverflow.com/questions/25107806/estimate-formants-using-lpc-in-python and should be confirmed with prof Niesler. Things I wonder about: - Peak picking? Where is this happening? - Why is the HPF necessary? - How do we confirm that this works? INPUTS: ======= audio: List containing audio data OUTPUTS: ======== freq: Formant frequencies F1-F5 """ N = len(audio) w = np.hamming(N) # Apply hamming window and High Pass filter audio = lfilter([1],[1,0.63], (audio*w) ) # ncoeff = 2 + fs/1000 """paper used 14th order LPC""" A, e, k = lpc(audio,1) roots = np.roots(A) roots = [r for r in roots if np.imag(r) >= 0] # Get angles angles = np.arctan(np.imag(roots), np.real(roots)) # Get Frequencies freq = sorted(angles * (fs / (2 * math.pi))) return freq[0:5]
def compare_to_talkbox(self):
try:
import scikits.talkbox as tbox
except ImportError:
return
order = 1
y = generate_recursive_noise()
a_pyp = sp.analysis.lpc(y,order=order)
a_tbx = tbox.lpc(y,order=order)
for ap, at in zip(a_pyp, a_tbx):
self.assertAlmostEqual(ap, at, delta=0.01)
def fft(fs, data):
max_freq = 5000
ncoeff = 2 + fs / 1000
a, e, k = lpc(data, ncoeff)
w, h = scipy.signal.freqz(1, a, worN=512)
freqs = fs * w / (2 * np.pi)
ans = 20 * np.log10(abs(h))
freqs = [freq for freq in freqs if freq < max_freq]
ans = ans[:len(freqs)]
labels = {'xlabel': u'Częstotliwość [Hz]', 'ylabel': 'Wzmocnienie [dB]'}
return {'y_vector': ans, 'x_vector': freqs, 'labels': labels}
def get_formants(x, Fs):
#for e in x:
# print >> sys.stderr, e
# Get Hamming window.
N = len(x)
w = np.hamming(N)
# Apply window and high pass (pre-emphasis) filter.
x1 = x * w
x1 = lfilter([1.], [1., 0.63], x1)
# Resample to make estimates better??
new_Fs = 22050
new_N = np.floor((float(N) * float(new_Fs)) / Fs)
#print new_N
x1 = resample(x1, new_N, window=None)
Fs = int(new_Fs)
# Get LPC.
ncoeff = 0 + Fs / 1000
A, e, k = lpc(x1, ncoeff)
try:
# Get roots.
rts = np.roots(A)
rts = [r for r in rts if np.imag(r) >= 0]
# Get angles.
angz = np.arctan2(np.imag(rts), np.real(rts))
# Get frequencies.
frqs = angz * (Fs / (2 * math.pi))
frq_indices = np.argsort(frqs)
frqs = [frqs[i] for i in frq_indices]
bws = [
-1 / 2 * (Fs / (2 * np.pi)) * np.log(np.abs(rts[i]))
for i in frq_indices
]
frqs = [
freq for freq, bw in itertools.izip(frqs, bws)
if freq > 90 and bw < 400
]
except np.linalg.LinAlgError:
frqs = []
return frqs
def lpcc(arr):
Fs = 11025
#rr = [1,2,3,4,5,6,7,7,88]
N = len(arr)
window = numpy.hamming(N)
arr1 = (arr * window)
arr1 = lfilter([1], [1., 0.63], arr1)
n_coeff = int(2 + Fs / 1000) #no of coefficients
A, e, k = (lpc(arr1, n_coeff)) #applying lpc
fft_var = numpy.fft.fft(A, 1024)
fft_var = abs(fft_var) #taking abs
squared_array = numpy.square(fft_var)
ar = [] #power spectrum
for i in squared_array:
ar.append(e / i)
log_signal = numpy.log(ar)
ifft_signal = numpy.fft.ifft(numpy.transpose(log_signal), 1024)
return ifft_signal[:14].tolist()
def getFormants(frames, sr):
# calculate number of LPC coefficients to use
ncoeff = 2 + sr/1000
# calculate LPC coefficients
c = lpc(frames, ncoeff)[0]
# obtain roots of LPC
A = np.diag(np.ones((c.shape[1]-2,), float), -1)
cs = -c[:,1:]
Z = np.array([np.vstack((cp, A[1:])) for cp in cs])
# root calculation using eigen method: VERY SLOW
eig = np.linalg.eigvals(Z)
arc = np.arctan2(np.imag(eig), np.real(eig))
# convert to Hz and sort ascending
formant = []
pi2 = 0.05*sr/np.pi
[formant.append(sorted(pi2*a[a>0])[:4]) for a in arc]
return np.array(formant)
def getFormants(frames, sr):
# calculate number of LPC coefficients to use
ncoeff = 2 + sr / 1000
# calculate LPC coefficients
c = lpc(frames, ncoeff)[0]
# obtain roots of LPC
A = np.diag(np.ones((c.shape[1] - 2, ), float), -1)
cs = -c[:, 1:]
Z = np.array([np.vstack((cp, A[1:])) for cp in cs])
# root calculation using eigen method: VERY SLOW
eig = np.linalg.eigvals(Z)
arc = np.arctan2(np.imag(eig), np.real(eig))
# convert to Hz and sort ascending
formant = []
pi2 = 0.05 * sr / np.pi
[formant.append(sorted(pi2 * a[a > 0])[:4]) for a in arc]
return np.array(formant)
def formant_freqs(fs, data):
max_freq = 5000
ncoeff = 2 + fs / 1000
a, e, k = lpc(data, ncoeff)
w, h = scipy.signal.freqz(1, a, worN=512)
freqs = fs * w / (2 * np.pi)
ans = 20 * np.log10(abs(h))
rts = np.roots(a)
rts = [r for r in rts if np.imag(r) >= 0]
angs = np.arctan2(np.imag(rts), np.real(rts))
formants = sorted(angs * (fs / (2 * math.pi)))
formants = filter(lambda formant: formant != 0 and formant < max_freq, formants)
freqs = [freq for freq in freqs if freq < max_freq]
ans = ans[:len(freqs)]
labels = {'xlabel': u'Częstotliwość [Hz]', 'ylabel': 'Wzmocnienie [dB]'}
return {'y_vector': ans, 'x_vector': freqs, 'labels': labels, 'cursors': formants}
def get_formants(x, fs):
"""
Estimate formants using LPC.
See:
http://www.mathworks.com/help/signal/ug/formant-estimation-with-lpc-coefficients.html
http://www.phon.ucl.ac.uk/courses/spsci/matlab/lect10.html
"""
# b, a = scipy.signal.butter(5, 1.0, 'low', analog=True)
# x = scipy.signal.filtfilt(b, a, x)
if not np.any(x):
# All zeroes
return []
x1 = lfilter([1.], [1., 0.63], x)
# Get Hamming window.
# N = len(x)
# w = np.hamming(N)
# # Apply window.
# x1 = x * w
# Get LPC.
ncoeff = 2 + fs / 1000
A, e, k = lpc(x1, ncoeff)
# Get roots.
rts = np.roots(A)
rts = [r for r in rts if np.imag(r) > 0]
# Get angles.
angz = np.arctan2(np.imag(rts), np.real(rts))
# Get frequencies.
frqs = sorted(angz * (fs / (2 * math.pi)))
return frqs
def get_formants(dummy):
# Read from file.
spf = wave.open("/home/ponco/devel/mel_cepstral_coeff_neural/vowels/EMartin.wav", 'r') # http://www.linguistics.ucla.edu/people/hayes/103/Charts/VChart/ae.wav
# Get file as numpy array.
x = spf.readframes(-1)
x = np.fromstring(x, 'Int16')
# Get Hamming window.
N = len(x)
w = np.hamming(N)
# Apply window and high pass filter.
x1 = x * w
x1 = lfilter([1], [1., 0.63], x1)
Fs = spf.getframerate()
#Fs = 44100
ncoeff = 2 + Fs / 1000
A, e, k = lpc(x1, ncoeff)
# Get LPC.
#A, e, k = lpc(x1, 8)
# Get roots.
rts = np.roots(A)
rts = [r for r in rts if np.imag(r) >= 0]
# Get angles.
angz = np.arctan2(np.imag(rts), np.real(rts))
# Get frequencies.
#Fs = spf.getframerate()
frqs = sorted(angz * (Fs / (2 * math.pi)))
return frqs
def stlpc(longSignal, order=10, windowLength=1024, hopsize=512, axis=-1):
"""Compute "Short Time LPC":
Cut the input signal in frames
Compute the LPC on each of the frames
"""
lengthSignal = longSignal.size
currentWindow = np.zeros([windowLength,])
Nb_windows = np.ceil((lengthSignal - windowLength) / (np.double(hopsize)) + 1.0)
STLpc = np.ones([order + 1, Nb_windows])
rootLpc = np.zeros([order, Nb_windows], dtype=np.complex)
freqLpc = np.ones([(order - 2.0)/2.0, Nb_windows])
specFromLpc = np.zeros([windowLength / 2.0 + 1, Nb_windows])
sigmaS = np.zeros([Nb_windows, ])
b_preamp=np.array([1.0,-0.99])
a_preamp=np.array([1.0])
longSignalPreamp = scipy.signal.lfilter(b_preamp,a_preamp,longSignal)
for n in np.arange(Nb_windows):
beginFrame = n * hopsize
endFrame = np.minimum(n * hopsize + windowLength, lengthSignal)
currentWindow[:endFrame-beginFrame] = longSignalPreamp[beginFrame: endFrame]
currentWindow *= np.hamming(windowLength)
STLpc[:,n], sigmaS[n], trash = tb.lpc(currentWindow, order)
specFromLpc[:,n] = lpc2spec(STLpc[:,n], sigmaS[n], fs, windowLength)
rootLpc[:,n] = np.roots(STLpc[:,n])
freqLpcTmp = np.angle(rootLpc[:,n]) / (2.0 * np.pi) * fs
freqLpcTmp = freqLpcTmp[freqLpcTmp>0.0]
freqLpcTmp.sort()
nbMinPositiveRoots = freqLpcTmp[0:(order - 2.0)/2.0].size
freqLpc[0:nbMinPositiveRoots,n] = freqLpcTmp[0:(order - 2.0)/2.0]
return STLpc, rootLpc, freqLpc, specFromLpc, sigmaS
def get_formants(x, fs):
"""
Estimate formants using LPC.
See:
http://www.mathworks.com/help/signal/ug/formant-estimation-with-lpc-coefficients.html
http://www.phon.ucl.ac.uk/courses/spsci/matlab/lect10.html
"""
# Get Hamming window.
N = len(x)
w = np.hamming(N)
# Apply window.
x1 = x * w
# Apply pre-emphasis filter.
# x1 = lfilter([1.0], [1.0, -0.63], x1)
# Get LPC.
ncoeff = 2 + fs / 1000
A, e, k = lpc(x1, ncoeff)
# Get roots.
rts = np.roots(A)
rts = [r for r in rts if np.imag(r) > 0]
# Get angles.
angz = np.arctan2(np.imag(rts), np.real(rts))
# Get frequencies.
frqs = sorted(angz * (fs / (2 * math.pi)))
return frqs
def phormants(x, Fs):
N = len(x)
w = numpy.hamming(N)
# Apply window and high pass filter.
x1 = x * w
x1 = lfilter([1], [1., 0.63], x1)
# Get LPC.
ncoeff = 2 + Fs / 1000
A, e, k = lpc(x1, ncoeff)
#A, e, k = lpc(x1, 8)
# Get roots.
rts = numpy.roots(A)
rts = [r for r in rts if numpy.imag(r) >= 0]
# Get angles.
angz = numpy.arctan2(numpy.imag(rts), numpy.real(rts))
# Get frequencies.
frqs = sorted(angz * (Fs / (2 * math.pi)))
return frqs
def phormants(x, Fs):
N = len(x)
w = numpy.hamming(N)
# Apply window and high pass filter.
x1 = x * w
x1 = lfilter([1], [1., 0.63], x1)
# Get LPC.
ncoeff = 2 + Fs / 1000
A, e, k = lpc(x1, ncoeff)
# A, e, k = lpc(x1, 8)
# Get roots.
rts = numpy.roots(A)
rts = [r for r in rts if numpy.imag(r) >= 0]
# Get angles.
angz = numpy.arctan2(numpy.imag(rts), numpy.real(rts))
# Get frequencies.
frqs = sorted(angz * (Fs / (2 * math.pi)))
return frqs
def get_formants(file_path):
# Read from file.
spf = wave.open(
file_path, 'r'
) # http://www.linguistics.ucla.edu/people/hayes/103/Charts/VChart/ae.wav
# Get file as numpy array.
x = spf.readframes(-1)
x = numpy.fromstring(x, 'Int16')
# Get Hamming window.
N = len(x)
w = numpy.hamming(N)
# Apply window and high pass filter.
x1 = x * w
x1 = lfilter([1], [1., 0.63], x1)
# Get LPC.
Fs = spf.getframerate()
ncoeff = 2 + Fs / 1000
A, e, k = lpc(x1, ncoeff)
# Get roots.
rts = numpy.roots(A)
rts = [r for r in rts if numpy.imag(r) >= 0]
# Get angles.
angz = numpy.arctan2(numpy.imag(rts), numpy.real(rts))
# Get frequencies.
Fs = spf.getframerate()
frqs = sorted(angz * (Fs / (2 * math.pi)))
return frqs
def find_start_and_end_of_hit(vec_x, fs, vec_idx_transients):
idx_start = []
idx_end = []
L_x = len(vec_x)
# ___ some user parameters___
T_window = 5 # sec
L_window = int(math.floor(T_window * fs))
if sp.remainder(L_window, 2) == 1:
L_window = L_window + 1
L_half_window = L_window / 2
for cur_idx_transient in vec_idx_transients:
idx_start = int(cur_idx_transient)- L_half_window
idx_end = int(cur_idx_transient) + L_half_window
if idx_start < 0: idx_start = 0
if idx_end > L_x: idx_end = L_x
print("idx_start: {}".format(idx_start) )
print("idx_end: {}".format(idx_end))
#pdb.set_trace()
cur_vec_x = vec_x[idx_start:idx_end+1]
vec_envelope = np.abs(sgn.hilbert(vec_x, axis=0))
plt.figure(2)
plt.plot(cur_vec_x)
plt.hold(True)
plt.plot(vec_envelope)
plt.legend(['input signal', 'hilbert envelope'])
plt.show()
# calculate a threshold
#if b_debug
# tempfig('transient start search');
# plot([x vec_envelope]);
#end
# whiten the signal via lpc
# vec_a = [1, 2, 3] # lpc(x,32)
(vec_a, vec_prediction_error, vec_k) = talkbox.lpc(vec_x, 32, axis=0)
vec_prediction_error = sgn.lfilter(vec_a, 1, vec_x, axis=0)
# print(vec_prediction_error)
plt.figure(3)
plt.plot(vec_prediction_error)
plt.title('prediction error')
plt.show()
#vec_signal_whitened = scipy.filter([0 -a(2:end)], 1, x);
#vec_signal_diff = x- vec_signal_whitened;
L_vec_x = len(vec_x)
idx_mid = (L_vec_x-1) / 2;
if sp.mod(idx_mid, 1) > 0:
idx_mid = sp.floor(idx_mid)
vec_prediction_error_power = vec_prediction_error**2
threshold = 0.9 * np.max(vec_prediction_error_power)
#idx_temp_to_the_left = idx_mid - find(vec_prediction_error_power(idx_mid:-1:1) > threshold, 1, 'first') + 1;
idx_temp_to_the_left = idx_mid - sp.nonzero(vec_prediction_error_power[idx_mid:0:-1] > threshold)[0] + 1
#idx_temp_to_the_right = find(vec_prediction_error_power(idx_mid:end) > threshold, 1, 'first') + idx_mid - 1;
idx_temp_to_the_right = idx_mid+ sp.nonzero(vec_prediction_error_power[idx_mid-1:L_vec_x] > threshold)[0]
if len(idx_temp_to_the_left) == 0:
idx_start = idx_temp_to_the_right[0]
elif len(idx_temp_to_the_right) == 0:
idx_start = idx_temp_to_the_left[0]
else:
# which one is closer?
if abs(idx_temp_to_the_left[0] - idx_mid) > abs(idx_temp_to_the_right[0] - idx_mid):
idx_start = idx_temp_to_the_right[0]
else:
idx_start = idx_temp_to_the_left[0]
#todo: think about this
idx_start_decay = idx_start
idx_start_transient = idx_start
#
#if b_debug
# tempfig('transient start search');
# hold on; plot(idx_start, vec_envelope(idx_start), 'Marker', 'o', 'MarkerFaceColor', 'r');
# plot(vec_signal_whitened, 'g');
# hold off;
# tempfig('prediction error');
# plot(vec_prediction_error_power);
# hold on;
# line([1 length(vec_prediction_error_power)], repmat(threshold, 1, 2), 'Color', 'r', 'LineWidth', 2); hold off;
#end
# filter that thing
order_filter_smooth = 300.
vec_b_filter_smooth = 1 / order_filter_smooth * sp.ones(order_filter_smooth,)
vec_a_filter_smooth = [1]
#disp(vec_b_filter_smooth)
#vec_b_smooth = 1/order_smoothing_filter * ones(order_smoothing_filter, 1);
#vec_a_smooth = 1;
#x_envelope_smoothed = filtfilt(vec_b_smooth, vec_a_smooth, x_envelope);
vec_envelope_smoothed = sgn.filtfilt(vec_b_filter_smooth, vec_a_filter_smooth, vec_envelope, padtype=None)
#tempfig('selected waveform');
#% subplot(211);
#hold on, plot((0:length(x)-1) / fs, x_envelope_smoothed, 'r'); hold off;
#
# try to find the beginning of the decay phase
temp_max = np.max(vec_envelope_smoothed, axis=0)
# disp(temp_max)
#temp_max = max(x_envelope_smoothed);
#tempfig('selected waveform'); hold on;
#% subplot(211);
#plot((idx_start_decay-1) / fs, temp_max, 'Marker', 'o', 'MarkerSize', 15, 'Color', 'g');
#
#% tempfig('selected waveform'); hold on;
#% subplot(212);
#% plot((0:length(x)-2) / fs, diff((x_envelope_smoothed)), 'r'); hold off;
#
#tempfig('envelope histogram');
#hist(x_envelope_smoothed, 100);
#
# find the end of the decay phase
threshold = np.percentile(vec_envelope_smoothed[idx_start_decay:L_vec_x], 30)
#threshold = quantile(x_envelope_smoothed(idx_start_decay:end), 0.3);
#idx_end_decay = find(x_envelope_smoothed(idx_start_decay:end) <= threshold, 1, 'first') + idx_start_decay-1 ;
idx_end_decay = sp.nonzero(vec_envelope_smoothed[idx_start_decay:L_vec_x] <= threshold)[0][0] + idx_start_decay
#tempfig('selected waveform'); hold on;
#% subplot(211);
#plot((idx_end_decay-1) / fs, x_envelope_smoothed(idx_end_decay), 'Marker', 'o', 'MarkerSize', 15, 'Color', 'g');
#
#
#% now model the decay
#% (to estimate the decay time)
#if true
# val_start = (x_envelope_smoothed(idx_start_decay));
# val_end = (x_envelope_smoothed(idx_end_decay));
# decay_constant = 1*( log(val_end) - log(val_start) ) / (idx_end_decay - idx_start_decay);
#else
# val_start = x_envelope_smoothed(idx_start_decay);
#
#end
#
# % plot the model
# x_model = (val_start) * exp(1*decay_constant * (0:(idx_end_decay - idx_start_decay))');
# tempfig('selected waveform'); hold on;
# plot((idx_start_decay:idx_end_decay) / fs, x_model, 'k');
#
# tau_decay_ms = -1 / (decay_constant * fs) * 1000
#
idx_end_transient = idx_end_decay
print('idx_start_transient: {}'.format(idx_start_transient))
print('idx_end_transient: {}'.format(idx_end_transient))
return (idx_start_transient, idx_end_transient)
def get_formant_locations_from_raw_long_frame(v_sig, v_pm, nx, fft_len):
'''
nx: frame index
'''
#v_sig, fs = la.read_audio_file(wavfile)
# Epoch detection:
#v_pm_sec, v_voi = la.reaper_epoch_detection(wavfile)
#v_pm = lu.round_to_int(v_pm_sec * fs)
# Raw-long Frame extraction:
v_frm_long = v_sig[v_pm[nx-2]:v_pm[nx+2]+1]
# Win:
left_len = v_pm[nx] - v_pm[nx-2]
right_len = v_pm[nx+2] - v_pm[nx]
v_win = la.gen_non_symmetric_win(left_len, right_len, np.hanning, b_norm=False)
v_frm_long_win = v_frm_long * v_win
# Spectrum:
v_mag = la.remove_hermitian_half(np.absolute(np.fft.fft(v_frm_long_win, n=fft_len)[None,:]))[0]
v_mag_db = la.db(v_mag)
# Mel warping:
alpha = 0.50 # 0.55 - 0.60
#ncoeffs = 2048 # must be even
v_mag_mel = la.sp_mel_warp(v_mag[None,:], fft_len/2, alpha=alpha, in_type=3)[0]
v_sp_cmplx = la.build_min_phase_from_mag_spec(v_mag_mel[None,:])[0]
v_sp_cmplx_ext = la.add_hermitian_half(v_sp_cmplx[None,:], data_type='complex')[0]
v_frm_long_win_mel = np.fft.ifft(v_sp_cmplx_ext).real
if False:
plt.close('all')
pl(la.db(v_mag))
pl(la.db(np.absolute(v_sp_cmplx)))
pl(v_frm_long_win_mel)
# Formant extraction -LPC method:--------------------------------------------------
n_lpc_coeffs = 30 # 40
v_lpc_mel, v_e, v_refl = lpc(v_frm_long_win_mel, n_lpc_coeffs)
v_lpc_mag_mel = lpc_to_mag(v_lpc_mel, fft_len=fft_len)
v_lpc_mag_mel_db = la.db(v_lpc_mag_mel)
v_lpc_mag_mel_db = v_lpc_mag_mel_db - np.mean(v_lpc_mag_mel_db) + np.mean(la.db(v_mag_mel))
v_frmnts_bins_mel, v_frmnts_gains_db = get_formant_locations_from_spec_env(v_lpc_mag_mel_db)
# Getting bandwidth:
fft_len_half = 1 + fft_len / 2
v_vall_bins = get_formant_locations_from_spec_env(-v_lpc_mag_mel_db)[0]
v_vall_bins = np.r_[0, v_vall_bins, fft_len_half-1]
nfrmnts = v_frmnts_bins_mel.size
v_frmnts_bw_mel = np.zeros(nfrmnts) - 1.0
for nx_f in xrange(nfrmnts):
#Left slope:
curr_frmnt_bin = v_frmnts_bins_mel[nx_f]
curr_vall_l_bin = v_vall_bins[nx_f]
curr_vall_r_bin = v_vall_bins[nx_f+1]
curr_midp_l = int((curr_frmnt_bin + curr_vall_l_bin) / 2.0)
curr_midp_r = int((curr_frmnt_bin + curr_vall_r_bin) / 2.0)
# Protection:
if curr_midp_l==curr_frmnt_bin:
curr_midp_l = curr_vall_l_bin
if curr_midp_r==curr_frmnt_bin:
curr_midp_r = curr_vall_r_bin
#print(nx_f)
# 27 y 32
slope_l = (v_frmnts_gains_db[nx_f] - v_lpc_mag_mel_db[curr_midp_l]) / (curr_frmnt_bin - curr_midp_l).astype(float)
slope_r = (v_frmnts_gains_db[nx_f] - v_lpc_mag_mel_db[curr_midp_r]) / (curr_frmnt_bin - curr_midp_r).astype(float)
slope_ave = (slope_l - slope_r) / 2.0
v_frmnts_bw_mel[nx_f] = 1.0 / slope_ave
# Filtering by bandwidth:
# bw_thress = 7.0
# v_frmnts_bins_mel = v_frmnts_bins_mel[v_frmnts_bw_mel<bw_thress]
# v_frmnts_gains_db = v_frmnts_gains_db[v_frmnts_bw_mel<bw_thress]
# v_frmnts_bw_mel = v_frmnts_bw_mel[v_frmnts_bw_mel<bw_thress]
# Computing frame short:--------------------------------
# Win:
left_len_short = v_pm[nx] - v_pm[nx-1]
right_len_short = v_pm[nx+1] - v_pm[nx]
v_win_short = la.gen_non_symmetric_win(left_len_short, right_len_short, np.hanning, b_norm=False)
v_frm_short = v_sig[v_pm[nx-1]:v_pm[nx+1]+1]
v_frm_short_win = v_frm_short * v_win_short
shift = v_pm[nx] - v_pm[nx-1]
# Formant extraction - True envelope method:----------------------------------------
# Not finished.
#v_true_env_db = la.true_envelope(v_mag_db[None,:], in_type='db', ncoeffs=400, thres_db=0.1)[0]
if True:
plt.figure(); plt.plot(la.db(v_mag_mel)); plt.plot(v_lpc_mag_mel_db); plt.grid(); plt.show()
#pl(v_mag_db)
if True: import ipdb; ipdb.set_trace(context=8) # breakpoint 906d26d6 //
return v_mag_db, v_lpc_mag_mel_db, v_frmnts_bins_mel, v_frmnts_gains_db, v_frmnts_bw_mel, v_frm_short_win, shift
def get_formant_locations_from_raw_long_frame(v_sig, v_pm, nx, fft_len):
'''
nx: frame index
'''
#v_sig, fs = la.read_audio_file(wavfile)
# Epoch detection:
#v_pm_sec, v_voi = la.reaper_epoch_detection(wavfile)
#v_pm = lu.round_to_int(v_pm_sec * fs)
# Raw-long Frame extraction:
v_frm_long = v_sig[v_pm[nx-2]:v_pm[nx+2]+1]
# Win:
left_len = v_pm[nx] - v_pm[nx-2]
right_len = v_pm[nx+2] - v_pm[nx]
v_win = la.gen_non_symmetric_win(left_len, right_len, np.hanning, b_norm=False)
v_frm_long_win = v_frm_long * v_win
# Spectrum:
v_mag = np.absolute(np.fft.fft(v_frm_long_win, n=fft_len))
v_mag_db = la.db(la.remove_hermitian_half(v_mag[None,:])[0])
# Formant extraction -LPC method:--------------------------------------------------
v_lpc, v_e, v_refl = lpc(v_frm_long_win, 120)
b_use_lpc_roots = False
if b_use_lpc_roots:
v_lpc_roots = np.roots(v_lpc)
v_lpc_angles = np.angle(v_lpc_roots)
v_lpc_angles = v_lpc_angles[v_lpc_angles>=0]
v_lpc_angles = np.sort(v_lpc_angles)
fft_len_half = 1 + fft_len / 2
v_lpc_roots_bins = v_lpc_angles * fft_len_half / np.pi
v_lpc_mag = lpc_to_mag(v_lpc, fft_len=fft_len)
v_lpc_mag_db = la.db(v_lpc_mag)
v_lpc_mag_db = v_lpc_mag_db - np.mean(v_lpc_mag_db) + np.mean(v_mag_db)
v_frmnts_bins, v_frmnts_gains_db = get_formant_locations_from_spec_env(v_lpc_mag_db)
# Getting bandwidth:
fft_len_half = 1 + fft_len / 2
v_vall_bins = get_formant_locations_from_spec_env(-v_lpc_mag_db)[0]
v_vall_bins = np.r_[0, v_vall_bins, fft_len_half-1]
nfrmnts = v_frmnts_bins.size
v_frmnts_bw = np.zeros(nfrmnts) - 1.0
for nx_f in xrange(nfrmnts):
#Left slope:
curr_frmnt_bin = v_frmnts_bins[nx_f]
curr_vall_l_bin = v_vall_bins[nx_f]
curr_vall_r_bin = v_vall_bins[nx_f+1]
curr_midp_l = int((curr_frmnt_bin + curr_vall_l_bin) / 2.0)
curr_midp_r = int((curr_frmnt_bin + curr_vall_r_bin) / 2.0)
# Protection:
if curr_midp_l==curr_frmnt_bin:
curr_midp_l = curr_vall_l_bin
if curr_midp_r==curr_frmnt_bin:
curr_midp_r = curr_vall_r_bin
#print(nx_f)
# 27 y 32
#if ((nx==73) and (nx_f==27)): import ipdb; ipdb.set_trace(context=8) # breakpoint c4f78f1e //
slope_l = (v_frmnts_gains_db[nx_f] - v_lpc_mag_db[curr_midp_l]) / (curr_frmnt_bin - curr_midp_l).astype(float)
slope_r = (v_frmnts_gains_db[nx_f] - v_lpc_mag_db[curr_midp_r]) / (curr_frmnt_bin - curr_midp_r).astype(float)
slope_ave = (slope_l - slope_r) / 2.0
v_frmnts_bw[nx_f] = 1.0 / slope_ave
# Filtering by bandwidth:
bw_thress = 7.0
v_frmnts_bins = v_frmnts_bins[v_frmnts_bw<bw_thress]
v_frmnts_gains_db = v_frmnts_gains_db[v_frmnts_bw<bw_thress]
v_frmnts_bw = v_frmnts_bw[v_frmnts_bw<bw_thress]
# Computing frame short:--------------------------------
# Win:
left_len_short = v_pm[nx] - v_pm[nx-1]
right_len_short = v_pm[nx+1] - v_pm[nx]
v_win_short = la.gen_non_symmetric_win(left_len_short, right_len_short, np.hanning, b_norm=False)
v_frm_short = v_sig[v_pm[nx-1]:v_pm[nx+1]+1]
v_frm_short_win = v_frm_short * v_win_short
shift = v_pm[nx] - v_pm[nx-1]
# Formant extraction - True envelope method:----------------------------------------
# Not finished.
#v_true_env_db = la.true_envelope(v_mag_db[None,:], in_type='db', ncoeffs=400, thres_db=0.1)[0]
if False:
plt.figure(); plt.plot(v_mag_db); plt.plot(v_lpc_mag_db); plt.grid(); plt.show()
return v_mag_db, v_lpc_mag_db, v_frmnts_bins, v_frmnts_gains_db, v_frmnts_bw, v_frm_short_win, shift
import numpy as np
from matplotlib import pyplot as plt
from math import pi
from scipy.io import wavfile
from scipy import signal as sig
import das
from scikits.talkbox import lpc
#%% load wav
fs, wav = wavfile.read('audio/glas_aaa.wav')
wav = wav / 2**15
# wav = wav * sig.hamming(wav.size)
f, wav_spec = das.get_spectrum(wav, fs)
#%% get LPC parameters
a_lp, e, k = lpc(wav, 25)
b_inv = np.concatenate(([0], -a_lp[1:]))
wav_est = sig.lfilter(b_inv, 1, wav)
wav_err = wav - wav_est
G = e
f, err_spec = das.get_spectrum(wav_err, fs)
#%% plot
#plt.figure()
#plt.plot(wav)
#plt.plot(est_wav)
#plt.figure()
#plt.plot(err)
#%% LP filter impulse response and transfer function
x = np.zeros(.02 * fs)
def lpc_coeff(sig):
""" return ndarray"""
signal = reshapeSignal(sig)
A, e, k = lpc(signal, 12) # 12 coefficients (n/sizeof window) *k coefficients
return k
def fundEstimator(soundIn, fs, t=None, debugFig = 0, maxFund = 1500, minFund = 300, lowFc = 200, highFc = 6000, minSaliency = 0.5):
"""
Estimates the fundamental frequency of a complex sound.
soundIn is the sound pressure waveformlog spectrogram.
fs is the sampling rate
t is a vector of time values in s at which the fundamental will be estimated.
The sound must include at least 1024 sample points
The optional parameter with defaults are
Some user parameters (should be part of the function at some time)
debugFig = 0 Set to zero to eliminate figures.
maxFund = 1500 Maximum fundamental frequency
minFund = 300 Minimum fundamental frequency
lowFc = 200 Low frequency cut-off for band-passing the signal prior to auto-correlation.
highFc = 6000 High frequency cut-off
minSaliency = 0.5 Threshold in the auto-correlation for minimum saliency - returns NaN for pitch values is saliency is below this number
Returns
sal - the time varying pitch saliency - a number between 0 and 1 corresponding to relative size of the first auto-correlation peak
fund - the time-varying fundamental in Hz at the same resolution as the spectrogram.
fund2 - a second peak in the spectrum - not a multiple of the fundamental a sign of a second voice
form1 - the first formant, if it exists
form2 - the second formant, if it exists
form3 - the third formant, if it exists
soundLen - length of sal, fund, fund2, form1, form2, form3
"""
# Band-pass filtering signal prior to auto-correlation
soundLen = len(soundIn)
nfilt = 1024
if soundLen < 1024:
print 'Error in fundEstimator: sound too short for bandpass filtering, len(soundIn)=%d\n' % soundLen
return (0, 0, 0, 0, 0, 0, 0)
# high pass filter the signal
highpassFilter = firwin(nfilt-1, 2*lowFc/fs, pass_zero=False)
padlen = min(soundLen-10, 3*len(highpassFilter))
soundIn = filtfilt(highpassFilter, [1.0], soundIn, padlen=padlen)
# low pass filter the signal
lowpassFilter = firwin(nfilt, 2*highFc/fs)
padlen = min(soundLen-10, 3*len(lowpassFilter))
soundIn = filtfilt(lowpassFilter, [1.0], soundIn, padlen=padlen)
# Plot a spectrogram?
if debugFig:
plt.figure(9)
(tDebug ,freqDebug ,specDebug , rms) = spectrogram(soundIn, fs, 1000.0, 50, min_freq=0, max_freq=10000, nstd=6, log=True, noise_level_db=50, rectify=True)
plot_spectrogram(tDebug, freqDebug, specDebug)
# Initializations and useful variables
if t is None:
# initialize t to be spaced by 500us increments
sound_dur = len(soundIn) / fs
_si = 1e-3
npts = int(sound_dur / _si)
t = np.arange(npts) * _si
nt=len(t)
soundRMS = np.zeros(nt)
fund = np.zeros(nt)
fund2 = np.zeros(nt)
sal = np.zeros(nt)
form1 = np.zeros(nt)
form2 = np.zeros(nt)
form3 = np.zeros(nt)
# Calculate the size of the window for the auto-correlation
alpha = 5 # Number of sd in the Gaussian window
winLen = int(np.fix((2.0*alpha/minFund)*fs)) # Length of Gaussian window based on minFund
if (winLen%2 == 0): # Make a symmetric window
winLen += 1
winLen2 = 2**12+1 # This looks like a good size for LPC - 4097 points
gt, w = gaussian_window(winLen, alpha)
gt2, w2 = gaussian_window(winLen2, alpha)
maxlags = int(2*ceil((float(fs)/minFund)))
# First calculate the rms in each window
for it in range(nt):
tval = t[it] # Center of window in time
tind = int(np.fix(tval*fs)) # Center of window in ind
tstart = tind - (winLen-1)/2
tend = tind + (winLen-1)/2
if tstart < 0:
winstart = - tstart
tstart = 0
else:
winstart = 0
if tend >= soundLen:
windend = winLen - (tend-soundLen+1) - 1
tend = soundLen-1
else:
windend = winLen-1
soundWin = soundIn[tstart:tend]*w[winstart:windend]
soundRMS[it] = np.std(soundWin)
soundRMSMax = max(soundRMS)
# Calculate the auto-correlation in windowed segments and obtain 4 guess values of the fundamental
# fundCorrGuess - guess from the auto-correlation function
# fundCorrAmpGuess - guess form the amplitude of the auto-correlation function
# fundCepGuess - guess from the cepstrum
# fundStackGuess - guess taken from a fit of the power spectrum with a harmonic stack, using the fundCepGuess as a starting point
# Current version use fundStackGuess as the best estimate...
soundlen = 0
for it in range(nt):
fund[it] = float('nan')
sal[it] = float('nan')
fund2[it] = float('nan')
form1[it] = float('nan')
form2[it] = float('nan')
form3[it] = float('nan')
if (soundRMS[it] < soundRMSMax*0.1):
continue
soundlen += 1
tval = t[it] # Center of window in time
tind = int(np.fix(tval*fs)) # Center of window in ind
tstart = tind - (winLen-1)/2
tend = tind + (winLen-1)/2
if tstart < 0:
winstart = - tstart
tstart = 0
else:
winstart = 0
if tend >= soundLen:
windend = winLen - (tend-soundLen+1) - 1
tend = soundLen-1
else:
windend = winLen-1
tstart2 = tind - (winLen2-1)/2
tend2 = tind + (winLen2-1)/2
if tstart2 < 0:
winstart2 = - tstart2
tstart2 = 0
else:
winstart2 = 0
if tend2 >= soundLen:
windend2 = winLen2 - (tend2-soundLen+1) - 1
tend2 = soundLen-1
else:
windend2 = winLen2-1
soundWin = soundIn[tstart:tend]*w[winstart:windend]
soundWin2 = soundIn[tstart2:tend2]*w2[winstart2:windend2]
# Apply LPC to get time-varying formants and one additional guess for the fundamental frequency
A, E, K = talkbox.lpc(soundWin2, 8) # 8 degree polynomial
rts = np.roots(A) # Find the roots of A
rts = rts[np.imag(rts)>=0] # Keep only half of them
angz = np.arctan2(np.imag(rts),np.real(rts))
# Calculate the frequencies and the bandwidth of the formants
frqsFormants = angz*(fs/(2*np.pi))
indices = np.argsort(frqsFormants)
bw = -1/2*(fs/(2*np.pi))*np.log(np.abs(rts))
# Keep formants above 1000 Hz and with bandwidth < 1000
formants = []
for kk in indices:
if ( frqsFormants[kk]>1000 and bw[kk] < 1000):
formants.append(frqsFormants[kk])
formants = np.array(formants)
if len(formants) > 0 :
form1[it] = formants[0]
if len(formants) > 1 :
form2[it] = formants[1]
if len(formants) > 2 :
form3[it] = formants[2]
# Calculate the auto-correlation
lags = np.arange(-maxlags, maxlags+1, 1)
autoCorr = correlation_function(soundWin, soundWin, lags)
ind0 = int(mlab.find(lags == 0)) # need to find lag zero index
# find peaks
indPeaksCorr = detect_peaks(autoCorr, mph=max(autoCorr)/10)
# Eliminate center peak and all peaks too close to middle
indPeaksCorr = np.delete(indPeaksCorr,mlab.find( (indPeaksCorr-ind0) < fs/maxFund))
pksCorr = autoCorr[indPeaksCorr]
# Find max peak
if len(pksCorr)==0:
pitchSaliency = 0.1 # 0.1 goes with the detection of peaks greater than max/10
else:
indIndMax = mlab.find(pksCorr == max(pksCorr))[0]
indMax = indPeaksCorr[indIndMax]
fundCorrGuess = fs/abs(lags[indMax])
pitchSaliency = autoCorr[indMax]/autoCorr[ind0]
sal[it] = pitchSaliency
if sal[it] < minSaliency:
continue
# Calculate the envelope of the auto-correlation after rectification
envCorr = temporal_envelope(autoCorr, fs, cutoff_freq=maxFund, resample_rate=None)
locsEnvCorr = detect_peaks(envCorr, mph=max(envCorr)/10)
pksEnvCorr = envCorr[locsEnvCorr]
# The max peak should be around zero
indIndEnvMax = mlab.find(pksEnvCorr == max(pksEnvCorr))
# Take the first peak not in the middle
if indIndEnvMax+2 > len(locsEnvCorr):
fundCorrAmpGuess = fundCorrGuess
indEnvMax = indMax
else:
indEnvMax = locsEnvCorr[indIndEnvMax+1]
fundCorrAmpGuess = fs/lags[indEnvMax]
# Calculate power spectrum and cepstrum
Y = fft(soundWin, n=winLen+1)
f = (fs/2.0)*(np.array(range((winLen+1)/2+1), dtype=float)/float((winLen+1)/2))
fhigh = mlab.find(f >= highFc)[0]
powSound = 20.0*np.log10(np.abs(Y[0:(winLen+1)/2+1])) # This is the power spectrum
powSoundGood = powSound[0:fhigh]
maxPow = max(powSoundGood)
powSoundGood = powSoundGood - maxPow # Set zero as the peak amplitude
powSoundGood[powSoundGood < - 60] = -60
# Calculate coarse spectral enveloppe
p = np.polyfit(f[0:fhigh], powSoundGood, 3)
powAmp = np.polyval(p, f[0:fhigh])
# Cepstrum
CY = dct(powSoundGood-powAmp, norm = 'ortho')
tCY = 2000.0*np.array(range(len(CY)))/fs # Units of Cepstrum in ms
fCY = 1000.0/tCY # Corresponding fundamental frequency in Hz.
lowInd = mlab.find(fCY<lowFc)
if lowInd.size > 0:
flowCY = mlab.find(fCY < lowFc)[0]
else:
flowCY = fCY.size
fhighCY = mlab.find(fCY < highFc)[0]
# Find peak of Cepstrum
indPk = mlab.find(CY[fhighCY:flowCY] == max(CY[fhighCY:flowCY]))[-1]
indPk = fhighCY + indPk
fmass = 0
mass = 0
indTry = indPk
while (CY[indTry] > 0):
fmass = fmass + fCY[indTry]*CY[indTry]
mass = mass + CY[indTry]
indTry = indTry + 1
if indTry >= len(CY):
break
indTry = indPk - 1
if (indTry >= 0 ):
while (CY[indTry] > 0):
fmass = fmass + fCY[indTry]*CY[indTry]
mass = mass + CY[indTry]
indTry = indTry - 1
if indTry < 0:
break
fGuess = fmass/mass
if (fGuess == 0 or np.isnan(fGuess) or np.isinf(fGuess) ): # Failure of cepstral method
fGuess = fundCorrGuess
fundCepGuess = fGuess
# Force fundamendal to be bounded
if (fundCepGuess > maxFund ):
i = 2
while(fundCepGuess > maxFund):
fundCepGuess = fGuess/i
i += 1
elif (fundCepGuess < minFund):
i = 2
while(fundCepGuess < minFund):
fundCepGuess = fGuess*i
i += 1
# Fit Gaussian harmonic stack
maxPow = max(powSoundGood-powAmp)
# This is the matlab code...
# fundFitCep = NonLinearModel.fit(f(1:fhigh)', powSoundGood'-powAmp, @synSpect, [fundCepGuess ones(1,9).*log(maxPow)])
# modelPowCep = synSpect(double(fundFitCep.Coefficients(:,1)), f(1:fhigh))
vars = np.concatenate(([fundCepGuess], np.ones(9)*np.log(maxPow)))
bout = leastsq(residualSyn, vars, args = (f[0:fhigh], powSoundGood-powAmp))
modelPowCep = synSpect(bout[0], f[0:fhigh])
errCep = sum((powSoundGood - powAmp - modelPowCep)**2)
vars = np.concatenate(([fundCepGuess*2], np.ones(9)*np.log(maxPow)))
bout2 = leastsq(residualSyn, vars, args = (f[0:fhigh], powSoundGood-powAmp))
modelPowCep2 = synSpect(bout2[0], f[0:fhigh])
errCep2 = sum((powSoundGood - powAmp - modelPowCep2)**2)
if errCep2 < errCep:
bout = bout2
modelPowCep = modelPowCep2
fundStackGuess = bout[0][0]
if (fundStackGuess > maxFund) or (fundStackGuess < minFund ):
fundStackGuess = float('nan')
# A second cepstrum for the second voice
# CY2 = dct(powSoundGood-powAmp'- modelPowCep)
fund[it] = fundStackGuess
if not np.isnan(fundStackGuess):
powLeft = powSoundGood- powAmp - modelPowCep
maxPow2 = max(powLeft)
f2 = 0
if ( maxPow2 > maxPow*0.5): # Possible second peak in central area as indicator of second voice.
f2 = f[mlab.find(powLeft == maxPow2)]
if ( f2 > 1000 and f2 < 4000):
if (pitchSaliency > minSaliency):
fund2[it] = f2
#% modelPowCorrAmp = synSpect(double(fundFitCorrAmp.Coefficients(:,1)), f(1:fhigh))
#%
#% errCorr = sum((powSoundGood - powAmp' - modelPowCorr).^2)
#% errCorrAmp = sum((powSoundGood - powAmp' - modelPowCorrAmp).^2)
#% errCorrSum = sum((powSoundGood - powAmp' - (modelPowCorr+modelPowCorrAmp) ).^2)
#%
#% f1 = double(fundFitCorr.Coefficients(1,1))
#% f2 = double(fundFitCorrAmp.Coefficients(1,1))
#%
#% if (pitchSaliency > minSaliency)
#% if (errCorr < errCorrAmp)
#% fund(it) = f1
#% if errCorrSum < errCorr
#% fund2(it) = f2
#% end
#% else
#% fund(it) = f2
#% if errCorrSum < errCorrAmp
#% fund2(it) = f1
#% end
#% end
#%
#% end
if (debugFig ):
plt.figure(10)
plt.subplot(4,1,1)
plt.cla()
plt.plot(soundWin)
# f1 = double(fundFitCorr.Coefficients(1,1))
# f2 = double(fundFitCorrAmp.Coefficients(1,1))
titleStr = 'Saliency = %.2f Pitch AC = %.2f (Hz) Pitch ACA = %.2f Pitch C %.2f (Hz)' % (pitchSaliency, fundCorrGuess, fundCorrAmpGuess, fundStackGuess)
plt.title(titleStr)
plt.subplot(4,1,2)
plt.cla()
plt.plot(1000*(lags/fs), autoCorr)
plt.plot([1000.*lags[indMax]/fs, 1000*lags[indMax]/fs], [0, autoCorr[ind0]], 'k')
plt.plot(1000.*lags/fs, envCorr, 'r', linewidth= 2)
plt.plot([1000*lags[indEnvMax]/fs, 1000*lags[indEnvMax]/fs], [0, autoCorr[ind0]], 'g')
plt.xlabel('Time (ms)')
plt.subplot(4,1,3)
plt.cla()
plt.plot(f[0:fhigh],powSoundGood)
plt.axis([0, highFc, -60, 0])
plt.plot(f[0:fhigh], powAmp, 'b--')
plt.plot(f[0:fhigh], modelPowCep + powAmp, 'k')
# plt.plot(f(1:fhigh), modelPowCorrAmp + powAmp', 'g')
for ih in range(1,6):
plt.plot([fundCorrGuess*ih, fundCorrGuess*ih], [-60, 0], 'r')
plt.plot([fundStackGuess*ih, fundStackGuess*ih], [-60, 0], 'k')
if f2 != 0:
plt.plot([f2, f2], [-60, 0], 'g')
plt.xlabel('Frequency (Hz)')
# title(sprintf('Err1 = %.1f Err2 = %.1f', errCorr, errCorrAmp))
plt.subplot(4,1,4)
plt.cla()
plt.plot(tCY, CY)
# plot(tCY, CY2, 'k--')
plt.plot([1000/fundCorrGuess, 1000/fundCorrGuess], [0, max(CY)], 'r')
plt.plot([1000/fundStackGuess, 1000/fundStackGuess], [0, max(CY)], 'k')
#% plot([(pkClosest-1)/fs (pkClosest-1)/fs], [0 max(CY)], 'g')
#% if ~isempty(ipk2)
#% plot([(pk2-1)/fs (pk2-1)/fs], [0 max(CY)], 'b')
#% end
#% for ip=1:length(pks)
#% plot([(locs(ip)-1)/fs (locs(ip)-1)/fs], [0 pks(ip)/4], 'r')
#% end
plt.axis([0, 1000*np.size(CY)/(2*fs), 0, max(CY)])
plt.xlabel('Time (ms)')
plt.pause(1)
# Fix formants.
meanf1 = np.mean(form1[~np.isnan(form1)])
meanf2 = np.mean(form2[~np.isnan(form2)])
meanf3 = np.mean(form3[~np.isnan(form3)])
for it in range(nt):
if ~np.isnan(form1[it]):
df11 = np.abs(form1[it]-meanf1)
df12 = np.abs(form1[it]-meanf2)
df13 = np.abs(form1[it]-meanf3)
if df12 < df11:
if df13 < df12:
if ~np.isnan(form3[it]):
df33 = np.abs(form3[it]-meanf3)
if df13 < df33:
form3[it] = form1[it]
else:
form3[it] = form1[it]
else:
if ~np.isnan(form2[it]):
df22 = np.abs(form2[it]-meanf2)
if df12 < df22:
form2[it] = form1[it]
else:
form2[it] = form1[it]
form1[it] = float('nan')
if ~np.isnan(form2[it]):
df21 = np.abs(form2[it]-meanf1)
df22 = np.abs(form2[it]-meanf2)
df23 = np.abs(form2[it]-meanf3)
if df21 < df22 :
if ~np.isnan(form1[it]):
df11 = np.abs(form1[it]-meanf1)
if df21 < df11:
form1[it] = form2[it]
else:
form1[it] = form2[it]
form2[it] = float('nan')
elif df23 < df22:
if ~np.isnan(form3[it]):
df33 = np.abs(form3[it]-meanf3)
if df23 < df33:
form3[it] = form2[it]
else:
form3[it] = form2[it]
form2[it] = float('nan')
if ~np.isnan(form3[it]):
df31 = np.abs(form3[it]-meanf1)
df32 = np.abs(form3[it]-meanf2)
df33 = np.abs(form3[it]-meanf3)
if df32 < df33:
if df31 < df32:
if ~np.isnan(form1[it]):
df11 = np.abs(form1[it]-meanf1)
if df31 < df11:
form1[it] = form3[it]
else:
form1[it] = form3[it]
else:
if ~np.isnan(form2[it]):
df22 = np.abs(form2[it]-meanf2)
if df32 < df22:
form2[it] = form3[it]
else:
form2[it] = form3[it]
form3[it] = float('nan')
return (sal, fund, fund2, form1, form2, form3, soundlen)
def targets_map_fn(indexes):
rval = []
for sequence_index, example_index in self._fetch_index(indexes):
rval.append(lpc(self.samples_sequences[sequence_index][example_index].T,self.lpc_order)[0][1:].ravel())
return rval
def calFormants(frame):
formants = []
Fs = 7418
preemph = [1.0, 0.63]
frame = lfilter(preemph, 1, frame)
A, e, k = lpc(frame, 8)
A = numpy.nan_to_num(A)
rts = numpy.roots(A)
rts = rts[numpy.imag(rts) >= 0]
angz = []
for a in range(0, len(rts)):
ang = math.atan2(numpy.imag(rts[a]), numpy.real(rts[a]))
angz.insert(a, ang)
# print("angz", angz)
freqs = numpy.multiply(angz, (Fs / (2 * math.pi)))
freqs = sorted(freqs, reverse=True)
indices = numpy.argsort(freqs)
# print("freq and indices", freqs, indices)
bw = []
for a in range(0, len(indices)):
b = (-1 / 2) * (Fs /
(2 * math.pi)) * math.log(abs(rts[indices[a]]), 10)
bw.insert(a, b)
# print("bw", bw)
nn = 0
formants = []
for kk in range(0, len(freqs)):
if (freqs[kk] > 90 and bw[kk] < 400):
formants.insert(nn, freqs[kk])
nn = nn + 1
if (nn < 5):
if nn == 3: # indexing from zero -1 to matlab
formants.insert(3, 3500)
formants.insert(4, 3700)
# print ("formants")
if nn == 4: # indexing from zero so -1 to matlab
formants.insert(4, 3700)
if nn == 2: # indexing from zero so -1 to matlab
formants.insert(2, 3700)
formants.insert(3, 3700)
formants.insert(4, 3700)
if nn == 1: # indexing from zero so -1 to matlab
formants.insert(1, 3700)
formants.insert(2, 3700)
formants.insert(2, 3700)
formants.insert(4, 3700)
if nn == 0: # indexing from zero so -1 to matlab
formants.insert(0, 3700)
formants.insert(1, 3700)
formants.insert(2, 3700)
formants.insert(2, 3700)
formants.insert(4, 3700)
formants_5 = formants[:]
form = numpy.array(formants_5)
form.shape = (5, )
return form
def stlpc(longSignal,
order=10,
windowLength=1024,
hopsize=512,
samplingrate=16000,
axis=-1):
"""Compute 'Short Term LPC':
Cut the input signal in frames
Compute the LPC on each of the frames (through talkbox)
"""
fs = samplingrate
# adding zeros to have the first frame centered on 0:
data = np.concatenate((np.zeros(windowLength / 2), longSignal))
lengthSignal = data.size
# number of windows, and resizing the data,
# in accordance with stft from sffhmm.py:
nbWindows = np.ceil((lengthSignal - windowLength) /
(np.double(hopsize)) + 1.0) + 1
newLengthSignal = (nbWindows - 1) * hopsize + windowLength
data = np.concatenate([data, np.zeros([newLengthSignal - lengthSignal])])
currentWindow = np.zeros([
windowLength,
])
# number of coefficients for the LPC decomposition is `order+1`
STLpc = np.ones([order + 1, nbWindows])
# number of corresponding formants is
# `floor((order-1)/2)
# indeed, if `order` is odd, then it's `(order-1)/2`, that is to say all
# poles, except the isolated one (which is real)
# if `order` is even, then it's equal to `(order-2)/2`,
#
# 20130514 wait, why is it not order/2 again?
nbFormants = int(order / 2)
rootLpc = np.zeros([order, nbWindows], dtype=np.complex)
freqLpc = np.ones([nbFormants, nbWindows])
# specFromLpc = np.zeros([windowLength / 2.0 + 1, nbWindows])
sigmaS = np.zeros([
nbWindows,
])
# pre-processing the data, amplifying high frequencies:
b_preamp = np.array([1.0, -0.99])
a_preamp = np.array([1.0])
longSignalPreamp = scipy.signal.lfilter(b_preamp, a_preamp, data)
for n in np.arange(nbWindows):
# getting the desired frame
beginFrame = n * hopsize
endFrame = np.minimum(n * hopsize + windowLength, lengthSignal)
currentWindow[:endFrame -
beginFrame] = longSignalPreamp[beginFrame:endFrame]
# windowing the frame
currentWindow *= np.hamming(windowLength)
# computing the LPC coefficients
STLpc[:, n], sigmaS[n], _ = tb.lpc(currentWindow, order)
# compute the corresponding spectrum - not necessary here
# specFromLpc[:,n] = lpc2spec(STLpc[:,n], sigmaS[n], fs, windowLength)
# compute the roots of the polynomial:
rootLpc[:, n] = np.roots(STLpc[:, n])
# convert to frequencies
freqLpcTmp = np.angle(rootLpc[:, n]) / (2.0 * np.pi) * fs
freqLpcTmp = freqLpcTmp[freqLpcTmp > 0.0]
freqLpcTmp.sort()
nbMinPositiveRoots = freqLpcTmp[0:nbFormants].size
freqLpc[0:nbMinPositiveRoots, n] = freqLpcTmp[0:nbFormants]
return STLpc, rootLpc, freqLpc, sigmaS #, specFromLpc,
def get_lpc(self, data, order=44):
# Use talkbox to get the linear predictive coding
from scikits.talkbox import lpc
coefs = lpc(data, order)
return coefs[0]
def convert_to_lpc(filename, n_coeff):
wave, sr = lp.load(filename, mono=True, sr=16000)
lpc_signal = lpc(wave, n_coeff)
return np.hstack((lpc_signal[0], lpc_signal[1], lpc_signal[2]))
def feature_extraction_gd(y, fs=44100, statistics=True, include_delta=True,
include_acceleration=True, lpgd_params=None, win_params=None, delta_params=None, acceleration_params=None):
eps = numpy.spacing(1)
nfft = lpgd_params['nfft']
lp_order = lpgd_params['lp_order']
y = y + eps
frames = segment_axis(y, win_params['win_length'], win_params['hop_length']);
#print 'frames : ' + str(frames.shape)
a,e,k = lpc(frames, lp_order)
#print 'a : ' + str(a.shape)
A = fft(a, nfft)
A = 1/A
phaseA = numpy.unwrap(numpy.angle(A))
#print 'phaseA: ' + str(phaseA.shape)
phaseA = phaseA[:,0:nfft/2]
#print 'phaseA: ' + str(phaseA.shape)
tauA = -1 * numpy.diff(phaseA)
#print 'tauA' + str(tauA.shape)
tauA = numpy.vstack((tauA, tauA[-1]))
# tau = tau
# print 'tauA' + str(tauA.shape)
feature_matrix = tauA
feature_matrix = dct(feature_matrix, n=20)
feature_matrix = feature_matrix.T
print feature_matrix.shape
if include_delta:
# Delta coefficients
feature_delta = librosa.feature.delta(feature_matrix, **delta_params)
# Add Delta Coefficients to feature matrix
feature_matrix = numpy.vstack((feature_matrix, feature_delta))
# print 'fm: ' + str(feature_matrix.shape)
if include_acceleration:
# Acceleration coefficients (aka delta)
feature_delta2 = librosa.feature.delta(feature_delta, order=2, **acceleration_params)
# Add Acceleration Coefficients to feature matrix
feature_matrix = numpy.vstack((feature_matrix, feature_delta2))
feature_matrix = feature_matrix.T
print 'fm: ' + str(feature_matrix.shape)
# Collect into data structure
if statistics:
return {
'tauA' : tauA,
'feat': feature_matrix,
'stat': {
'mean': numpy.mean(feature_matrix, axis=0),
'std': numpy.std(feature_matrix, axis=0),
'N': feature_matrix.shape[0],
'S1': numpy.sum(feature_matrix, axis=0),
'S2': numpy.sum(feature_matrix ** 2, axis=0),
}
}
else:
return {
'feat': feature_matrix}
def get_lpc(self,data,order=44):
# Use talkbox to get the linear predictive coding
from scikits.talkbox import lpc
coefs = lpc(data,order)
return coefs[0]
def fundEstimator(soundIn, fs, t=None, debugFig = 0, maxFund = 1500, minFund = 300, lowFc = 200, highFc = 6000, minSaliency = 0.5):
"""
Estimates the fundamental frequency of a complex sound.
soundIn is the sound pressure waveformlog spectrogram.
fs is the sampling rate
t is a vector of time values in s at which the fundamental will be estimated.
The sound must include at least 1024 sample points
The optional parameter with defaults are
Some user parameters (should be part of the function at some time)
debugFig = 0 Set to zero to eliminate figures.
maxFund = 1500 Maximum fundamental frequency
minFund = 300 Minimum fundamental frequency
lowFc = 200 Low frequency cut-off for band-passing the signal prior to auto-correlation.
highFc = 6000 High frequency cut-off
minSaliency = 0.5 Threshold in the auto-correlation for minimum saliency - returns NaN for pitch values is saliency is below this number
Returns
sal - the time varying pitch saliency - a number between 0 and 1 corresponding to relative size of the first auto-correlation peak
fund - the time-varying fundamental in Hz at the same resolution as the spectrogram.
fund2 - a second peak in the spectrum - not a multiple of the fundamental a sign of a second voice
form1 - the first formant, if it exists
form2 - the second formant, if it exists
form3 - the third formant, if it exists
soundLen - length of sal, fund, fund2, form1, form2, form3
"""
# Band-pass filtering signal prior to auto-correlation
soundLen = len(soundIn)
nfilt = 1024
if soundLen < 1024:
print 'Error in fundEstimator: sound too short for bandpass filtering, len(soundIn)=%d\n' % soundLen
return (0, 0, 0, 0, 0, 0, 0)
# high pass filter the signal
highpassFilter = firwin(nfilt-1, 2*lowFc/fs, pass_zero=False)
padlen = min(soundLen-10, 3*len(highpassFilter))
soundIn = filtfilt(highpassFilter, [1.0], soundIn, padlen=padlen)
# low pass filter the signal
lowpassFilter = firwin(nfilt, 2*highFc/fs)
padlen = min(soundLen-10, 3*len(lowpassFilter))
soundIn = filtfilt(lowpassFilter, [1.0], soundIn, padlen=padlen)
# Plot a spectrogram?
if debugFig:
plt.figure(9)
(tDebug ,freqDebug ,specDebug , rms) = spectrogram(soundIn, fs, 1000.0, 50, min_freq=0, max_freq=10000, nstd=6, log=True, noise_level_db=50, rectify=True)
plot_spectrogram(tDebug, freqDebug, specDebug)
# Initializations and useful variables
if t is None:
# initialize t to be spaced by 500us increments
sound_dur = len(soundIn) / fs
_si = 1e-3
npts = int(sound_dur / _si)
t = np.arange(npts) * _si
nt=len(t)
soundRMS = np.zeros(nt)
fund = np.zeros(nt)
fund2 = np.zeros(nt)
sal = np.zeros(nt)
form1 = np.zeros(nt)
form2 = np.zeros(nt)
form3 = np.zeros(nt)
# Calculate the size of the window for the auto-correlation
alpha = 5 # Number of sd in the Gaussian window
winLen = int(np.fix((2.0*alpha/minFund)*fs)) # Length of Gaussian window based on minFund
if (winLen%2 == 0): # Make a symmetric window
winLen += 1
winLen2 = 2**12+1 # This looks like a good size for LPC - 4097 points
gt, w = gaussian_window(winLen, alpha)
gt2, w2 = gaussian_window(winLen2, alpha)
maxlags = int(2*ceil((float(fs)/minFund)))
# First calculate the rms in each window
for it in range(nt):
tval = t[it] # Center of window in time
tind = int(np.fix(tval*fs)) # Center of window in ind
tstart = tind - (winLen-1)/2
tend = tind + (winLen-1)/2
if tstart < 0:
winstart = - tstart
tstart = 0
else:
winstart = 0
if tend >= soundLen:
windend = winLen - (tend-soundLen+1) - 1
tend = soundLen-1
else:
windend = winLen-1
soundWin = soundIn[tstart:tend]*w[winstart:windend]
soundRMS[it] = np.std(soundWin)
soundRMSMax = max(soundRMS)
# Calculate the auto-correlation in windowed segments and obtain 4 guess values of the fundamental
# fundCorrGuess - guess from the auto-correlation function
# fundCorrAmpGuess - guess form the amplitude of the auto-correlation function
# fundCepGuess - guess from the cepstrum
# fundStackGuess - guess taken from a fit of the power spectrum with a harmonic stack, using the fundCepGuess as a starting point
# Current version use fundStackGuess as the best estimate...
soundlen = 0
for it in range(nt):
fund[it] = float('nan')
sal[it] = float('nan')
fund2[it] = float('nan')
form1[it] = float('nan')
form2[it] = float('nan')
form3[it] = float('nan')
if (soundRMS[it] < soundRMSMax*0.1):
continue
soundlen += 1
tval = t[it] # Center of window in time
tind = int(np.fix(tval*fs)) # Center of window in ind
tstart = tind - (winLen-1)/2
tend = tind + (winLen-1)/2
if tstart < 0:
winstart = - tstart
tstart = 0
else:
winstart = 0
if tend >= soundLen:
windend = winLen - (tend-soundLen+1) - 1
tend = soundLen-1
else:
windend = winLen-1
tstart2 = tind - (winLen2-1)/2
tend2 = tind + (winLen2-1)/2
if tstart2 < 0:
winstart2 = - tstart2
tstart2 = 0
else:
winstart2 = 0
if tend2 >= soundLen:
windend2 = winLen2 - (tend2-soundLen+1) - 1
tend2 = soundLen-1
else:
windend2 = winLen2-1
soundWin = soundIn[tstart:tend]*w[winstart:windend]
soundWin2 = soundIn[tstart2:tend2]*w2[winstart2:windend2]
# Apply LPC to get time-varying formants and one additional guess for the fundamental frequency
A, E, K = talkbox.lpc(soundWin2, 8) # 8 degree polynomial
rts = np.roots(A) # Find the roots of A
rts = rts[np.imag(rts)>=0] # Keep only half of them
angz = np.arctan2(np.imag(rts),np.real(rts))
# Calculate the frequencies and the bandwidth of the formants
frqsFormants = angz*(fs/(2*np.pi))
indices = np.argsort(frqsFormants)
bw = -1/2*(fs/(2*np.pi))*np.log(np.abs(rts))
# Keep formants above 1000 Hz and with bandwidth < 1000
formants = []
for kk in indices:
if ( frqsFormants[kk]>1000 and bw[kk] < 1000):
formants.append(frqsFormants[kk])
formants = np.array(formants)
if len(formants) > 0 :
form1[it] = formants[0]
if len(formants) > 1 :
form2[it] = formants[1]
if len(formants) > 2 :
form3[it] = formants[2]
# Calculate the auto-correlation
lags = np.arange(-maxlags, maxlags+1, 1)
autoCorr = correlation_function(soundWin, soundWin, lags)
ind0 = int(mlab.find(lags == 0)) # need to find lag zero index
# find peaks
indPeaksCorr = detect_peaks(autoCorr, mph=max(autoCorr)/10)
# Eliminate center peak and all peaks too close to middle
indPeaksCorr = np.delete(indPeaksCorr,mlab.find( (indPeaksCorr-ind0) < fs/maxFund))
pksCorr = autoCorr[indPeaksCorr]
# Find max peak
if len(pksCorr)==0:
pitchSaliency = 0.1 # 0.1 goes with the detection of peaks greater than max/10
else:
indIndMax = mlab.find(pksCorr == max(pksCorr))[0]
indMax = indPeaksCorr[indIndMax]
fundCorrGuess = fs/abs(lags[indMax])
pitchSaliency = autoCorr[indMax]/autoCorr[ind0]
sal[it] = pitchSaliency
if sal[it] < minSaliency:
continue
# Calculate the envelope of the auto-correlation after rectification
envCorr = temporal_envelope(autoCorr, fs, cutoff_freq=maxFund, resample_rate=None)
locsEnvCorr = detect_peaks(envCorr, mph=max(envCorr)/10)
pksEnvCorr = envCorr[locsEnvCorr]
# The max peak should be around zero
indIndEnvMax = mlab.find(pksEnvCorr == max(pksEnvCorr))
# Take the first peak not in the middle
if indIndEnvMax+2 > len(locsEnvCorr):
fundCorrAmpGuess = fundCorrGuess
indEnvMax = indMax
else:
indEnvMax = locsEnvCorr[indIndEnvMax+1]
fundCorrAmpGuess = fs/lags[indEnvMax]
# Calculate power spectrum and cepstrum
Y = fft(soundWin, n=winLen+1)
f = (fs/2.0)*(np.array(range((winLen+1)/2+1), dtype=float)/float((winLen+1)/2))
fhigh = mlab.find(f >= highFc)[0]
powSound = 20.0*np.log10(np.abs(Y[0:(winLen+1)/2+1])) # This is the power spectrum
powSoundGood = powSound[0:fhigh]
maxPow = max(powSoundGood)
powSoundGood = powSoundGood - maxPow # Set zero as the peak amplitude
powSoundGood[powSoundGood < - 60] = -60
# Calculate coarse spectral enveloppe
p = np.polyfit(f[0:fhigh], powSoundGood, 3)
powAmp = np.polyval(p, f[0:fhigh])
# Cepstrum
CY = dct(powSoundGood-powAmp, norm = 'ortho')
tCY = 2000.0*np.array(range(len(CY)))/fs # Units of Cepstrum in ms
fCY = 1000.0/tCY # Corresponding fundamental frequency in Hz.
lowInd = mlab.find(fCY<lowFc)
if lowInd.size > 0:
flowCY = mlab.find(fCY < lowFc)[0]
else:
flowCY = fCY.size
fhighCY = mlab.find(fCY < highFc)[0]
# Find peak of Cepstrum
indPk = mlab.find(CY[fhighCY:flowCY] == max(CY[fhighCY:flowCY]))[-1]
indPk = fhighCY + indPk
fmass = 0
mass = 0
indTry = indPk
while (CY[indTry] > 0):
fmass = fmass + fCY[indTry]*CY[indTry]
mass = mass + CY[indTry]
indTry = indTry + 1
if indTry >= len(CY):
break
indTry = indPk - 1
if (indTry >= 0 ):
while (CY[indTry] > 0):
fmass = fmass + fCY[indTry]*CY[indTry]
mass = mass + CY[indTry]
indTry = indTry - 1
if indTry < 0:
break
fGuess = fmass/mass
if (fGuess == 0 or np.isnan(fGuess) or np.isinf(fGuess) ): # Failure of cepstral method
fGuess = fundCorrGuess
fundCepGuess = fGuess
# Force fundamendal to be bounded
if (fundCepGuess > maxFund ):
i = 2
while(fundCepGuess > maxFund):
fundCepGuess = fGuess/i
i += 1
elif (fundCepGuess < minFund):
i = 2
while(fundCepGuess < minFund):
fundCepGuess = fGuess*i
i += 1
# Fit Gaussian harmonic stack
maxPow = max(powSoundGood-powAmp)
# This is the matlab code...
# fundFitCep = NonLinearModel.fit(f(1:fhigh)', powSoundGood'-powAmp, @synSpect, [fundCepGuess ones(1,9).*log(maxPow)])
# modelPowCep = synSpect(double(fundFitCep.Coefficients(:,1)), f(1:fhigh))
vars = np.concatenate(([fundCepGuess], np.ones(9)*np.log(maxPow)))
bout = leastsq(residualSyn, vars, args = (f[0:fhigh], powSoundGood-powAmp))
modelPowCep = synSpect(bout[0], f[0:fhigh])
errCep = sum((powSoundGood - powAmp - modelPowCep)**2)
vars = np.concatenate(([fundCepGuess*2], np.ones(9)*np.log(maxPow)))
bout2 = leastsq(residualSyn, vars, args = (f[0:fhigh], powSoundGood-powAmp))
modelPowCep2 = synSpect(bout2[0], f[0:fhigh])
errCep2 = sum((powSoundGood - powAmp - modelPowCep2)**2)
if errCep2 < errCep:
bout = bout2
modelPowCep = modelPowCep2
fundStackGuess = bout[0][0]
if (fundStackGuess > maxFund) or (fundStackGuess < minFund ):
fundStackGuess = float('nan')
# A second cepstrum for the second voice
# CY2 = dct(powSoundGood-powAmp'- modelPowCep)
fund[it] = fundStackGuess
if not np.isnan(fundStackGuess):
powLeft = powSoundGood- powAmp - modelPowCep
maxPow2 = max(powLeft)
f2 = 0
if ( maxPow2 > maxPow*0.5): # Possible second peak in central area as indicator of second voice.
f2 = f[mlab.find(powLeft == maxPow2)]
if ( f2 > 1000 and f2 < 4000):
if (pitchSaliency > minSaliency):
fund2[it] = f2
#% modelPowCorrAmp = synSpect(double(fundFitCorrAmp.Coefficients(:,1)), f(1:fhigh))
#%
#% errCorr = sum((powSoundGood - powAmp' - modelPowCorr).^2)
#% errCorrAmp = sum((powSoundGood - powAmp' - modelPowCorrAmp).^2)
#% errCorrSum = sum((powSoundGood - powAmp' - (modelPowCorr+modelPowCorrAmp) ).^2)
#%
#% f1 = double(fundFitCorr.Coefficients(1,1))
#% f2 = double(fundFitCorrAmp.Coefficients(1,1))
#%
#% if (pitchSaliency > minSaliency)
#% if (errCorr < errCorrAmp)
#% fund(it) = f1
#% if errCorrSum < errCorr
#% fund2(it) = f2
#% end
#% else
#% fund(it) = f2
#% if errCorrSum < errCorrAmp
#% fund2(it) = f1
#% end
#% end
#%
#% end
if (debugFig ):
plt.figure(10)
plt.subplot(4,1,1)
plt.cla()
plt.plot(soundWin)
# f1 = double(fundFitCorr.Coefficients(1,1))
# f2 = double(fundFitCorrAmp.Coefficients(1,1))
titleStr = 'Saliency = %.2f Pitch AC = %.2f (Hz) Pitch ACA = %.2f Pitch C %.2f (Hz)' % (pitchSaliency, fundCorrGuess, fundCorrAmpGuess, fundStackGuess)
plt.title(titleStr)
plt.subplot(4,1,2)
plt.cla()
plt.plot(1000*(lags/fs), autoCorr)
plt.plot([1000.*lags[indMax]/fs, 1000*lags[indMax]/fs], [0, autoCorr[ind0]], 'k')
plt.plot(1000.*lags/fs, envCorr, 'r', linewidth= 2)
plt.plot([1000*lags[indEnvMax]/fs, 1000*lags[indEnvMax]/fs], [0, autoCorr[ind0]], 'g')
plt.xlabel('Time (ms)')
plt.subplot(4,1,3)
plt.cla()
plt.plot(f[0:fhigh],powSoundGood)
plt.axis([0, highFc, -60, 0])
plt.plot(f[0:fhigh], powAmp, 'b--')
plt.plot(f[0:fhigh], modelPowCep + powAmp, 'k')
# plt.plot(f(1:fhigh), modelPowCorrAmp + powAmp', 'g')
for ih in range(1,6):
plt.plot([fundCorrGuess*ih, fundCorrGuess*ih], [-60, 0], 'r')
plt.plot([fundStackGuess*ih, fundStackGuess*ih], [-60, 0], 'k')
if f2 != 0:
plt.plot([f2, f2], [-60, 0], 'g')
plt.xlabel('Frequency (Hz)')
# title(sprintf('Err1 = %.1f Err2 = %.1f', errCorr, errCorrAmp))
plt.subplot(4,1,4)
plt.cla()
plt.plot(tCY, CY)
# plot(tCY, CY2, 'k--')
plt.plot([1000/fundCorrGuess, 1000/fundCorrGuess], [0, max(CY)], 'r')
plt.plot([1000/fundStackGuess, 1000/fundStackGuess], [0, max(CY)], 'k')
#% plot([(pkClosest-1)/fs (pkClosest-1)/fs], [0 max(CY)], 'g')
#% if ~isempty(ipk2)
#% plot([(pk2-1)/fs (pk2-1)/fs], [0 max(CY)], 'b')
#% end
#% for ip=1:length(pks)
#% plot([(locs(ip)-1)/fs (locs(ip)-1)/fs], [0 pks(ip)/4], 'r')
#% end
plt.axis([0, 1000*np.size(CY)/(2*fs), 0, max(CY)])
plt.xlabel('Time (ms)')
plt.pause(1)
# Fix formants.
meanf1 = np.mean(form1[~np.isnan(form1)])
meanf2 = np.mean(form2[~np.isnan(form2)])
meanf3 = np.mean(form3[~np.isnan(form3)])
for it in range(nt):
if ~np.isnan(form1[it]):
df11 = np.abs(form1[it]-meanf1)
df12 = np.abs(form1[it]-meanf2)
df13 = np.abs(form1[it]-meanf3)
if df12 < df11:
if df13 < df12:
if ~np.isnan(form3[it]):
df33 = np.abs(form3[it]-meanf3)
if df13 < df33:
form3[it] = form1[it]
else:
form3[it] = form1[it]
else:
if ~np.isnan(form2[it]):
df22 = np.abs(form2[it]-meanf2)
if df12 < df22:
form2[it] = form1[it]
else:
form2[it] = form1[it]
form1[it] = float('nan')
if ~np.isnan(form2[it]):
df21 = np.abs(form2[it]-meanf1)
df22 = np.abs(form2[it]-meanf2)
df23 = np.abs(form2[it]-meanf3)
if df21 < df22 :
if ~np.isnan(form1[it]):
df11 = np.abs(form1[it]-meanf1)
if df21 < df11:
form1[it] = form2[it]
else:
form1[it] = form2[it]
form2[it] = float('nan')
elif df23 < df22:
if ~np.isnan(form3[it]):
df33 = np.abs(form3[it]-meanf3)
if df23 < df33:
form3[it] = form2[it]
else:
form3[it] = form2[it]
form2[it] = float('nan')
if ~np.isnan(form3[it]):
df31 = np.abs(form3[it]-meanf1)
df32 = np.abs(form3[it]-meanf2)
df33 = np.abs(form3[it]-meanf3)
if df32 < df33:
if df31 < df32:
if ~np.isnan(form1[it]):
df11 = np.abs(form1[it]-meanf1)
if df31 < df11:
form1[it] = form3[it]
else:
form1[it] = form3[it]
else:
if ~np.isnan(form2[it]):
df22 = np.abs(form2[it]-meanf2)
if df32 < df22:
form2[it] = form3[it]
else:
form2[it] = form3[it]
form3[it] = float('nan')
return (sal, fund, fund2, form1, form2, form3, soundlen)
def stlpc(longSignal,
order=10,
windowLength=1024,
hopsize=512,
samplingrate=16000,
axis=-1):
"""Compute 'Short Term LPC':
Cut the input signal in frames
Compute the LPC on each of the frames (through talkbox)
"""
fs = samplingrate
# adding zeros to have the first frame centered on 0:
data = np.concatenate((np.zeros(windowLength/2),
longSignal))
lengthSignal = data.size
# number of windows, and resizing the data,
# in accordance with stft from sffhmm.py:
nbWindows = np.ceil((lengthSignal - windowLength) /
(np.double(hopsize)) + 1.0) + 1
newLengthSignal = (nbWindows - 1) * hopsize + windowLength
data = np.concatenate([data,
np.zeros([newLengthSignal - lengthSignal])])
currentWindow = np.zeros([windowLength,])
# number of coefficients for the LPC decomposition is `order+1`
STLpc = np.ones([order + 1, nbWindows])
# number of corresponding formants is
# `floor((order-1)/2)
# indeed, if `order` is odd, then it's `(order-1)/2`, that is to say all
# poles, except the isolated one (which is real)
# if `order` is even, then it's equal to `(order-2)/2`,
#
# 20130514 wait, why is it not order/2 again?
nbFormants = int(order / 2)
rootLpc = np.zeros([order, nbWindows], dtype=np.complex)
freqLpc = np.ones([nbFormants, nbWindows])
# specFromLpc = np.zeros([windowLength / 2.0 + 1, nbWindows])
sigmaS = np.zeros([nbWindows, ])
# pre-processing the data, amplifying high frequencies:
b_preamp=np.array([1.0,-0.99])
a_preamp=np.array([1.0])
longSignalPreamp = scipy.signal.lfilter(b_preamp,a_preamp,data)
for n in np.arange(nbWindows):
# getting the desired frame
beginFrame = n * hopsize
endFrame = np.minimum(n * hopsize + windowLength, lengthSignal)
currentWindow[:endFrame-beginFrame] = longSignalPreamp[beginFrame:
endFrame]
# windowing the frame
currentWindow *= np.hamming(windowLength)
# computing the LPC coefficients
STLpc[:,n], sigmaS[n], _ = tb.lpc(currentWindow, order)
# compute the corresponding spectrum - not necessary here
# specFromLpc[:,n] = lpc2spec(STLpc[:,n], sigmaS[n], fs, windowLength)
# compute the roots of the polynomial:
rootLpc[:,n] = np.roots(STLpc[:,n])
# convert to frequencies
freqLpcTmp = np.angle(rootLpc[:,n]) / (2.0 * np.pi) * fs
freqLpcTmp = freqLpcTmp[freqLpcTmp>0.0]
freqLpcTmp.sort()
nbMinPositiveRoots = freqLpcTmp[0:nbFormants].size
freqLpc[0:nbMinPositiveRoots,n] = freqLpcTmp[0:nbFormants]
return STLpc, rootLpc, freqLpc, sigmaS #, specFromLpc,
def _compute_formants(self, audio_buffer):
"""
Computes the frequencies of formants of the window of audio data, along with their bandwidths.
A formant is a frequency band over which there is a concentration of energy.
They correspond to tones produced by the vocal tract and are therefore often
used to characterize vowels, which have distinct frequencies. In the task of
speaker identification, it can be used to characterize a person's speech
patterns.
This implementation is based on the Matlab tutorial on Estimating Formants using
LPC (Linear Predictive Coding) Coefficients:
http://www.mathworks.com/help/signal/ug/formant-estimation-with-lpc-coefficients.html.
Help porting this to Python was found here :
http://stackoverflow.com/questions/25107806/estimate-formants-using-lpc-in-python.
Why LPC? http://dsp.stackexchange.com/questions/2482/speech-compression-in-lpc-how-does-the-linear-predictive-filter-work-on-a-gene
Here are some more details on why linear predictive analysis is a generally powerful tool
in audio processing: http://iitg.vlab.co.in/?sub=59&brch=164&sim=616&cnt=1108.
"""
# Get Hamming window. More on window functions can be found at https://en.wikipedia.org/wiki/Window_function
# The idea of the Hamming window is to smooth out discontinuities at the edges of the window.
# Simply multiply to apply the window.
N = len(audio_buffer)
Fs = 8000 # sampling frequency
hamming_window = np.hamming(N)
window = audio_buffer * hamming_window
# Apply a pre-emphasis filter; this amplifies high-frequency components and attenuates low-frequency components.
# The purpose in voice processing is to remove noise.
filtered_buffer = lfilter([1], [1., 0.63], window)
# Speech can be broken down into (1) The raw sound emitted by the larynx and (2) Filtering that occurs when transmitted from the larynx, defined by, for instance, mouth shape and tongue position.
# The larynx emits a periodic function defined by its amplitude and frequency.
# The transmission is more complex to model but is in the form 1/(1-sum(a_k * z^-k)), where the coefficients
# a_k sufficiently encode the function (because we know it's of that form).
# Linear Predictive Coding is a method for estimating these coefficients given a pre-filtered audio signal.
# These value are called the roots, because the are the points at which the difference
# from the actual signal and the reconstructed signal (using that transmission function) is closest to 0.
# See http://dsp.stackexchange.com/questions/2482/speech-compression-in-lpc-how-does-the-linear-predictive-filter-work-on-a-gene.
# Get the roots using linear predictive coding.
# As a rule of thumb, the order of the LPC should be 2 more than the sampling frequency (in kHz).
ncoeff = 2 + Fs / 1000
A, e, k = lpc(filtered_buffer, ncoeff)
roots = np.roots(A)
roots = [r for r in roots if np.imag(r) >= 0]
# Get angles from the roots. Each root represents a complex number. The angle in the
# complex coordinate system (where x is the real part and y is the imaginary part)
# corresponds to the "frequency" of the formant (in rad/s, however, so we need to convert them).
# Note it really is a frequency band, not a single frequency, but this is a simplification that is acceptable.
angz = np.arctan2(np.imag(roots), np.real(roots))
# Convert the angular frequencies from rad/sample to Hz; then calculate the
# bandwidths of the formants. The distance of the roots from the unit circle
# gives the bandwidths of the formants (*Extra credit* if you can explain this!).
unsorted_freqs = angz * (Fs / (2 * math.pi))
# Let's sort the frequencies so that when we later compare them, we don't overestimate
# the difference due to ordering choices.
freqs = sorted(unsorted_freqs)
# also get the indices so that we can get the bandwidths in the same order
indices = np.argsort(unsorted_freqs)
sorted_roots = np.asarray(roots)[indices]
#compute the bandwidths of each formant
bandwidths = -1 / 2. * (Fs /
(2 * math.pi)) * np.log(np.abs(sorted_roots))
if self.debug:
print("Identified {} formants.".format(len(freqs)))
return freqs, bandwidths
def feature_extraction_gd(y,
fs=44100,
statistics=True,
include_delta=True,
include_acceleration=True,
lpgd_params=None,
win_params=None,
delta_params=None,
acceleration_params=None):
eps = numpy.spacing(1)
nfft = lpgd_params['nfft']
lp_order = lpgd_params['lp_order']
y = y + eps
frames = segment_axis(y, win_params['win_length'],
win_params['hop_length'])
#print 'frames : ' + str(frames.shape)
a, e, k = lpc(frames, lp_order)
#print 'a : ' + str(a.shape)
A = fft(a, nfft)
A = 1 / A
phaseA = numpy.unwrap(numpy.angle(A))
#print 'phaseA: ' + str(phaseA.shape)
phaseA = phaseA[:, 0:nfft / 2]
#print 'phaseA: ' + str(phaseA.shape)
tauA = -1 * numpy.diff(phaseA)
#print 'tauA' + str(tauA.shape)
tauA = numpy.vstack((tauA, tauA[-1]))
# tau = tau
# print 'tauA' + str(tauA.shape)
feature_matrix = tauA
feature_matrix = dct(feature_matrix, n=20)
feature_matrix = feature_matrix.T
print feature_matrix.shape
if include_delta:
# Delta coefficients
feature_delta = librosa.feature.delta(feature_matrix, **delta_params)
# Add Delta Coefficients to feature matrix
feature_matrix = numpy.vstack((feature_matrix, feature_delta))
# print 'fm: ' + str(feature_matrix.shape)
if include_acceleration:
# Acceleration coefficients (aka delta)
feature_delta2 = librosa.feature.delta(feature_delta,
order=2,
**acceleration_params)
# Add Acceleration Coefficients to feature matrix
feature_matrix = numpy.vstack((feature_matrix, feature_delta2))
feature_matrix = feature_matrix.T
print 'fm: ' + str(feature_matrix.shape)
# Collect into data structure
if statistics:
return {
'tauA': tauA,
'feat': feature_matrix,
'stat': {
'mean': numpy.mean(feature_matrix, axis=0),
'std': numpy.std(feature_matrix, axis=0),
'N': feature_matrix.shape[0],
'S1': numpy.sum(feature_matrix, axis=0),
'S2': numpy.sum(feature_matrix**2, axis=0),
}
}
else:
return {'feat': feature_matrix}
