def extract_features_file(self,
audio,
static=True,
plots=False,
fmt="npy",
kaldi_file=""):
"""Extract the phonation features from an audio file
:param audio: .wav audio file.
:param static: whether to compute and return statistic functionals over the feature matrix, or return the feature matrix computed over frames
:param plots: timeshift to extract the features
:param fmt: format to return the features (npy, dataframe, torch, kaldi)
:param kaldi_file: file to store kaldi features, only valid when fmt=="kaldi"
:returns: features computed from the audio file.
>>> phonation=Phonation()
>>> file_audio="../audios/001_a1_PCGITA.wav"
>>> features1=phonation.extract_features_file(file_audio, static=True, plots=True, fmt="npy")
>>> features2=phonation.extract_features_file(file_audio, static=True, plots=True, fmt="dataframe")
>>> features3=phonation.extract_features_file(file_audio, static=False, plots=True, fmt="torch")
>>> phonation.extract_features_file(file_audio, static=False, plots=False, fmt="kaldi", kaldi_file="./test")
"""
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.size_step * float(fs)
overlap = size_stepS / size_frameS
if self.pitch_method == 'praat':
name_audio = audio.split('/')
temp_uuid = 'phon' + name_audio[-1][0:-4]
if not os.path.exists(self.PATH + '/../tempfiles/'):
os.makedirs(self.PATH + '/../tempfiles/')
temp_filename_vuv = self.PATH + '/../tempfiles/tempVUV' + temp_uuid + '.txt'
temp_filename_f0 = self.PATH + '/../tempfiles/tempF0' + temp_uuid + '.txt'
praat_functions.praat_vuv(audio,
temp_filename_f0,
temp_filename_vuv,
time_stepF0=self.size_step,
minf0=self.minf0,
maxf0=self.maxf0)
F0, _ = praat_functions.decodeF0(temp_filename_f0,
len(data_audio) / float(fs),
self.size_step)
os.remove(temp_filename_vuv)
os.remove(temp_filename_f0)
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')
F0nz = F0[F0 != 0]
Jitter = jitter_env(F0nz, len(F0nz))
nF = int((len(data_audio) / size_frameS / overlap)) - 1
Amp = []
logE = []
apq = []
ppq = []
DF0 = np.diff(F0nz, 1)
DDF0 = np.diff(DF0, 1)
F0z = F0[F0 == 0]
totaldurU = len(F0z)
thresholdE = 10 * logEnergy([self.energy_thr_percent])
degreeU = 100 * float(totaldurU) / len(F0)
lnz = 0
for l in range(nF):
data_frame = data_audio[int(l * size_stepS):int(l * size_stepS +
size_frameS)]
energy = 10 * logEnergy(data_frame)
if F0[l] != 0:
Amp.append(np.max(np.abs(data_frame)))
logE.append(energy)
if lnz >= 12: # TODO:
amp_arr = np.asarray(
[Amp[j] for j in range(lnz - 12, lnz)])
#print(amp_arr)
apq.append(APQ(amp_arr))
if lnz >= 6: # TODO:
f0arr = np.asarray([F0nz[j] for j in range(lnz - 6, lnz)])
ppq.append(PPQ(1 / f0arr))
lnz = lnz + 1
Shimmer = shimmer_env(Amp, len(Amp))
apq = np.asarray(apq)
ppq = np.asarray(ppq)
logE = np.asarray(logE)
if len(apq) == 0:
print(
"warning, there is not enough long voiced segments to compute the APQ, in this case APQ=shimmer"
)
apq = Shimmer
if plots:
self.plot_phon(data_audio, fs, F0, logE)
if len(Shimmer) == len(apq):
feat_mat = np.vstack((DF0[5:], DDF0[4:], Jitter[6:], Shimmer[6:],
apq[6:], ppq, logE[6:])).T
else:
feat_mat = np.vstack((DF0[11:], DDF0[10:], Jitter[12:],
Shimmer[12:], apq, ppq[6:], logE[12:])).T
feat_v = dynamic2statict([DF0, DDF0, Jitter, Shimmer, apq, ppq, logE])
if fmt == "npy" or fmt == "txt":
if static:
return feat_v
else:
return feat_mat
elif fmt == "dataframe" or fmt == "csv":
if static:
head_st = []
df = {}
for k in ["avg", "std", "skewness", "kurtosis"]:
for h in self.head:
head_st.append(k + " " + h)
for e, k in enumerate(head_st):
df[k] = [feat_v[e]]
return pd.DataFrame(df)
else:
df = {}
for e, k in enumerate(self.head):
df[k] = feat_mat[:, e]
return pd.DataFrame(df)
elif fmt == "torch":
if static:
feat_t = torch.from_numpy(feat_v)
return feat_t
else:
return torch.from_numpy(feat_mat)
elif fmt == "kaldi":
if static:
raise ValueError(
"Kaldi is only supported for dynamic features")
else:
name_all = audio.split('/')
dictX = {name_all[-1]: feat_mat}
save_dict_kaldimat(dictX, kaldi_file)
else:
raise ValueError(fmt + " is not supported")
def phonationVowels(audio, flag_plots, size_frame=0.04,size_step=0.02,minf0=60,maxf0=350, voice_bias=-0.2,energy_thr_percent=0.025, pitch_method='praat'):
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
if pitch_method == 'praat':
name_audio=audio.split('/')
temp_uuid='phon'+name_audio[-1][0:-4]
if not os.path.exists('../tempfiles/'):
os.makedirs('../tempfiles/')
temp_filename_vuv='../tempfiles/tempVUV'+temp_uuid+'.txt'
temp_filename_f0='../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':
data_audiof=np.asarray(data_audio*(2**15), dtype=np.float32)
F0=pysptk.sptk.rapt(data_audiof, fs, int(size_stepS), min=minf0, max=maxf0, voice_bias=voice_bias, otype='f0')
F0nz=F0[F0!=0]
Jitter=jitter_env(F0nz, len(F0nz))
nF=int((len(data_audio)/size_frameS/overlap))-1
Amp=[]
logE=[]
apq=[]
ppq=[]
DF0=np.diff(F0nz, 1)
DDF0=np.diff(DF0,1)
F0z=F0[F0==0]
totaldurU=len(F0z)
thresholdE=10*logEnergy([energy_thr_percent])
degreeU=100*float(totaldurU)/len(F0)
lnz=0
for l in range(nF):
data_frame=data_audio[int(l*size_stepS):int(l*size_stepS+size_frameS)]
energy=10*logEnergy(data_frame)
if F0[l]!=0:
Amp.append(np.max(np.abs(data_frame)))
logE.append(10*logEnergy(data_frame))
if lnz>=12: # TODO:
amp_arr=np.asarray([Amp[j] for j in range(lnz-12, lnz)])
#print(amp_arr)
apq.append(APQ(amp_arr))
if lnz>=6: # TODO:
f0arr=np.asarray([F0nz[j] for j in range(lnz-6, lnz)])
ppq.append(PPQ(1/f0arr))
lnz=lnz+1
print("frame "+str(l) +" from "+str(nF)+"-"*int(100*l/nF)+">"+str(int(100*(l+1)/nF))+"%", sep=' ', end='\r', flush=True)
Shimmer=shimmer_env(Amp, len(Amp))
apq=np.asarray(apq)
ppq=np.asarray(ppq)
logE=np.asarray(logE)
F0semi=np.asarray([Hz2semitones(F0nz[l]) for l in range(len(F0nz))])
if flag_plots:
plot_phon(data_audio,fs,F0,logE)
print("Jitter=", len(Jitter))
print("Shimmer", len(Shimmer))
print("APQ", len(apq))
print("PPQ", len(ppq))
print("DF0", len(DF0))
print("DDF0", len(DDF0))
print("Energy", len(logE))
print("degree unvoiced",degreeU)
return F0, DF0, DDF0, F0semi, Jitter, Shimmer, apq, ppq, logE, degreeU
