def prosody_dynamic(self, audio):
"""Extract the dynamic prosody features from an audio file
:param audio: .wav audio file.
:returns: array (N,13) with the prosody features extracted from an audio file. N= number of voiced segments
>>> prosody=Prosody()
>>> file_audio="../audios/001_ddk1_PCGITA.wav"
>>> features=prosody.prosody_dynamic(file_audio)
"""
fs, data_audio = read(audio)
data_audio = data_audio - np.mean(data_audio)
data_audio = data_audio / float(np.max(np.abs(data_audio)))
size_frameS = self.size_frame * float(fs)
size_stepS = self.step * float(fs)
overlap = size_stepS / size_frameS
if self.pitch_method == 'praat':
name_audio = audio.split('/')
temp_uuid = 'prosody' + name_audio[-1][0:-4]
if not os.path.exists(self.PATH + '/../tempfiles/'):
os.makedirs(self.PATH + '/../tempfiles/')
temp_filename_f0 = self.PATH + '/../tempfiles/tempF0' + temp_uuid + '.txt'
temp_filename_vuv = self.PATH + '/../tempfiles/tempVUV' + temp_uuid + '.txt'
praat_functions.praat_vuv(audio,
temp_filename_f0,
temp_filename_vuv,
time_stepF0=self.step,
minf0=self.minf0,
maxf0=self.maxf0)
F0, _ = praat_functions.decodeF0(temp_filename_f0,
len(data_audio) / float(fs),
self.step)
os.remove(temp_filename_f0)
os.remove(temp_filename_vuv)
elif self.pitch_method == 'rapt':
data_audiof = np.asarray(data_audio * (2**15), dtype=np.float32)
F0 = pysptk.sptk.rapt(data_audiof,
fs,
int(size_stepS),
min=self.minf0,
max=self.maxf0,
voice_bias=self.voice_bias,
otype='f0')
#Find pitch contour of EACH voiced segment
pitchON = np.where(F0 != 0)[0]
dchange = np.diff(pitchON)
change = np.where(dchange > 1)[0]
iniV = pitchON[0]
featvec = []
iniVoiced = (pitchON[0] * size_stepS) + size_stepS #To compute energy
seg_voiced = []
f0v = []
Ev = []
for indx in change:
finV = pitchON[indx] + 1
finVoiced = (pitchON[indx] *
size_stepS) + size_stepS #To compute energy
VoicedSeg = data_audio[int(iniVoiced):int(
finVoiced)] #To compute energy
temp = F0[iniV:finV]
tempvec = []
if len(VoicedSeg) > int(
size_frameS): #Take only segments greater than frame size
seg_voiced.append(VoicedSeg)
#Compute duration
dur = len(VoicedSeg) / float(fs)
#Pitch coefficients
x = np.arange(0, len(temp))
z = np.poly1d(np.polyfit(x, temp, self.P))
f0v.append(temp)
tempvec.extend(z.coeffs)
#Energy coefficients
temp = E_cont(VoicedSeg, size_frameS, size_stepS, overlap)
Ev.append(temp)
x = np.arange(0, len(temp))
z = np.poly1d(np.polyfit(x, temp, self.P))
tempvec.extend(z.coeffs)
tempvec.append(dur)
featvec.append(tempvec)
iniV = pitchON[indx + 1]
iniVoiced = (pitchON[indx + 1] *
size_stepS) + size_stepS #To compute energy
#Add the last voiced segment
finV = (pitchON[len(pitchON) - 1])
finVoiced = (pitchON[len(pitchON) - 1] *
size_stepS) + size_stepS #To compute energy
VoicedSeg = data_audio[int(iniVoiced):int(
finVoiced)] #To compute energy
temp = F0[iniV:finV]
tempvec = []
if len(VoicedSeg) > int(
size_frameS): #Take only segments greater than frame size
#Compute duration
dur = len(VoicedSeg) / float(fs)
tempvec.append(dur)
x = np.arange(0, len(temp))
z = np.poly1d(np.polyfit(x, temp, self.P))
tempvec.extend(z.coeffs)
#Energy coefficients
temp = E_cont(VoicedSeg, size_frameS, size_stepS, overlap)
x = np.arange(0, len(temp))
z = np.poly1d(np.polyfit(x, temp, self.P))
tempvec.extend(z.coeffs)
#Compute duration
featvec.append(tempvec)
return np.asarray(featvec)
def prosody_dynamic(audio,
size_frame=0.03,
size_step=0.01,
minf0=60,
maxf0=350,
voice_bias=-0.2,
energy_thr_percent=0.025,
P=5,
pitch_method='praat'):
"""
Based on:
Najim Dehak, "Modeling Prosodic Features With Joint Factor Analysis for Speaker Verification", 2007
"""
fs, data_audio = read(audio)
data_audio = data_audio - np.mean(data_audio)
data_audio = data_audio / float(np.max(np.abs(data_audio)))
size_frameS = size_frame * float(fs)
size_stepS = size_step * float(fs)
overlap = size_stepS / size_frameS
nF = int((len(data_audio) / size_frameS / overlap)) - 1
data_audiof = np.asarray(data_audio * (2**15), dtype=np.float32)
if pitch_method == 'praat':
name_audio = audio.split('/')
temp_uuid = 'pros' + name_audio[-1][0:-4]
temp_filename_vuv = path_app + '/../tempfiles/tempVUV' + temp_uuid + '.txt'
temp_filename_f0 = path_app + '/../tempfiles/tempF0' + temp_uuid + '.txt'
praat_functions.praat_vuv(audio,
temp_filename_f0,
temp_filename_vuv,
time_stepF0=size_step,
minf0=minf0,
maxf0=maxf0)
F0, _ = praat_functions.decodeF0(temp_filename_f0,
len(data_audio) / float(fs),
size_step)
os.remove(temp_filename_vuv)
os.remove(temp_filename_f0)
elif pitch_method == 'rapt':
F0 = pysptk.sptk.rapt(data_audiof,
fs,
int(size_stepS),
min=minf0,
max=maxf0,
voice_bias=voice_bias,
otype='f0')
#Find pitch contour of EACH voiced segment
pitchON = np.where(F0 != 0)[0]
dchange = np.diff(pitchON)
change = np.where(dchange > 1)[0]
iniV = pitchON[0]
featvec = []
iniVoiced = (pitchON[0] * size_stepS) + size_stepS #To compute energy
seg_voiced = []
f0v = []
Ev = []
for indx in change:
finV = pitchON[indx] + 1
finVoiced = (pitchON[indx] *
size_stepS) + size_stepS #To compute energy
VoicedSeg = data_audio[int(iniVoiced):int(
finVoiced)] #To compute energy
temp = F0[iniV:finV]
tempvec = []
if len(VoicedSeg) > int(
size_frameS): #Take only segments greater than frame size
seg_voiced.append(VoicedSeg)
#Compute duration
dur = len(VoicedSeg) / float(fs)
tempvec.append(dur)
#Pitch coefficients
x = np.arange(0, len(temp))
z = np.poly1d(np.polyfit(x, temp, P))
f0v.append(temp)
#fitCoeff.append(z.coeffs)
tempvec.extend(z.coeffs)
#Energy coefficients
temp = E_cont(VoicedSeg, size_frameS, size_stepS, overlap)
Ev.append(temp)
x = np.arange(0, len(temp))
z = np.poly1d(np.polyfit(x, temp, P))
tempvec.extend(z.coeffs)
featvec.append(tempvec)
iniV = pitchON[indx + 1]
iniVoiced = (pitchON[indx + 1] *
size_stepS) + size_stepS #To compute energy
#Add the last voiced segment
finV = (pitchON[len(pitchON) - 1])
finVoiced = (pitchON[len(pitchON) - 1] *
size_stepS) + size_stepS #To compute energy
VoicedSeg = data_audio[int(iniVoiced):int(finVoiced)] #To compute energy
temp = F0[iniV:finV]
tempvec = []
if len(VoicedSeg) > int(
size_frameS): #Take only segments greater than frame size
#Compute duration
dur = len(VoicedSeg) / float(fs)
tempvec.append(dur)
x = np.arange(0, len(temp))
z = np.poly1d(np.polyfit(x, temp, P))
tempvec.extend(z.coeffs)
#Energy coefficients
temp = E_cont(VoicedSeg, size_frameS, size_stepS, overlap)
x = np.arange(0, len(temp))
z = np.poly1d(np.polyfit(x, temp, P))
tempvec.extend(z.coeffs)
#Compute duration
featvec.append(tempvec)
if flag_plots:
plot_pros(data_audio, fs, F0, seg_voiced, Ev, featvec, f0v)
return np.asarray(featvec)