def extract_features_file(self,
audio,
static=True,
plots=False,
fmt="npy",
kaldi_file=""):
"""Extract the glottal 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.
>>> glottal=Glottal()
>>> file_audio="../audios/001_a1_PCGITA.wav"
>>> features1=glottal.extract_features_file(file_audio, static=True, plots=True, fmt="npy")
>>> features2=glottal.extract_features_file(file_audio, static=True, plots=True, fmt="dataframe")
>>> features3=glottal.extract_features_file(file_audio, static=False, plots=True, fmt="torch")
>>> glottal.extract_features_file(file_audio, static=False, plots=False, fmt="kaldi", kaldi_file="./test.ark")
"""
if audio.find('.wav') == -1 and audio.find('.WAV') == -1:
raise ValueError(audio + " is not a valid wav file")
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
nF = int((len(data_audio) / size_frameS / overlap)) - 1
data_audiof = np.asarray(data_audio * (2**15), dtype=np.float32)
f0 = pysptk.sptk.rapt(data_audiof,
fs,
int(0.01 * fs),
min=20,
max=500,
voice_bias=-0.2,
otype='f0')
sizef0 = int(self.size_frame / 0.01)
stepf0 = int(self.size_step / 0.01)
startf0 = 0
stopf0 = sizef0
avgGCIt = np.zeros(nF)
varGCIt = np.zeros(nF)
avgNAQt = np.zeros(nF)
varNAQt = np.zeros(nF)
avgQOQt = np.zeros(nF)
varQOQt = np.zeros(nF)
avgH1H2t = np.zeros(nF)
varH1H2t = np.zeros(nF)
avgHRFt = np.zeros(nF)
varHRFt = np.zeros(nF)
rmwin = []
for l in range(nF):
data_frame = data_audio[int(l * size_stepS):int(l * size_stepS +
size_frameS)]
f0_frame = f0[startf0:stopf0]
pf0framez = np.where(f0_frame != 0)[0]
f0nzframe = f0_frame[pf0framez]
if len(f0nzframe) < 10:
startf0 = startf0 + stepf0
stopf0 = stopf0 + stepf0
rmwin.append(l)
continue
GCI = SE_VQ_varF0(data_frame, fs, f0=f0_frame)
g_iaif = IAIF(data_frame, fs, GCI)
g_iaif = g_iaif - np.mean(g_iaif)
g_iaif = g_iaif / max(abs(g_iaif))
glottal = cumtrapz(g_iaif)
glottal = glottal - np.mean(glottal)
glottal = glottal / max(abs(glottal))
startf0 = startf0 + stepf0
stopf0 = stopf0 + stepf0
gci_s = GCI[:]
GCId = np.diff(gci_s)
avgGCIt[l] = np.mean(GCId / fs)
varGCIt[l] = np.std(GCId / fs)
NAQ, QOQ, T1, T2, H1H2, HRF = get_vq_params(
glottal, g_iaif, fs, GCI)
avgNAQt[l] = np.mean(NAQ)
varNAQt[l] = np.std(NAQ)
avgQOQt[l] = np.mean(QOQ)
varQOQt[l] = np.std(QOQ)
avgH1H2t[l] = np.mean(H1H2)
varH1H2t[l] = np.std(H1H2)
avgHRFt[l] = np.mean(HRF)
varHRFt[l] = np.std(HRF)
if plots:
self.plot_glottal(data_frame, fs, GCI, g_iaif, glottal,
avgGCIt[l], varGCIt[l])
if len(rmwin) > 0:
varGCIt = np.delete(varGCIt, rmwin)
avgNAQt = np.delete(avgNAQt, rmwin)
varNAQt = np.delete(varNAQt, rmwin)
avgQOQt = np.delete(avgQOQt, rmwin)
varQOQt = np.delete(varQOQt, rmwin)
avgH1H2t = np.delete(avgH1H2t, rmwin)
varH1H2t = np.delete(varH1H2t, rmwin)
avgHRFt = np.delete(avgHRFt, rmwin)
varHRFt = np.delete(varHRFt, rmwin)
feat = np.stack((varGCIt, avgNAQt, varNAQt, avgQOQt, varQOQt, avgH1H2t,
varH1H2t, avgHRFt, varHRFt),
axis=1)
if fmt == "npy" or fmt == "txt":
if static:
return dynamic2static(feat)
else:
return feat
elif fmt == "dataframe" or fmt == "csv":
if static:
feat_st = dynamic2static(feat)
head_st = []
df = {}
for k in [
"global avg", "global std", "global skewness",
"global kurtosis"
]:
for h in self.head:
head_st.append(k + " " + h)
for e, k in enumerate(head_st):
df[k] = [feat_st[e]]
return pd.DataFrame(df)
else:
df = {}
for e, k in enumerate(self.head):
df[k] = feat[:, e]
return pd.DataFrame(df)
elif fmt == "torch":
if static:
feat_s = dynamic2static(feat)
feat_t = torch.from_numpy(feat_s)
return feat_t
else:
return torch.from_numpy(feat)
elif fmt == "kaldi":
if static:
raise ValueError(
"Kaldi is only supported for dynamic features")
else:
name_all = audio.split('/')
dictX = {name_all[-1]: feat}
save_dict_kaldimat(dictX, kaldi_file)
def glottal_features(audio, flag_plots, size_frame=0.2, size_step=0.1):
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)
f0 = pysptk.sptk.rapt(data_audiof,
fs,
int(0.01 * fs),
min=20,
max=500,
voice_bias=-0.2,
otype='f0')
sizef0 = int(size_frame / 0.01)
stepf0 = int(size_step / 0.01)
startf0 = 0
stopf0 = sizef0
avgGCIt = np.zeros(nF)
varGCIt = np.zeros(nF)
avgNAQt = np.zeros(nF)
varNAQt = np.zeros(nF)
avgQOQt = np.zeros(nF)
varQOQt = np.zeros(nF)
avgH1H2t = np.zeros(nF)
varH1H2t = np.zeros(nF)
avgHRFt = np.zeros(nF)
varHRFt = np.zeros(nF)
rmwin = []
for l in range(nF):
data_frame = data_audio[int(l * size_stepS):int(l * size_stepS +
size_frameS)]
f0_frame = f0[startf0:stopf0]
pf0framez = np.where(f0_frame != 0)[0]
f0nzframe = f0_frame[pf0framez]
if len(f0nzframe) < 10:
startf0 = startf0 + stepf0
stopf0 = stopf0 + stepf0
rmwin.append(l)
print("frame " + str(l) + " from " + str(nF) +
"-" * int(100 * l / nF) + ">" + str(int(100 *
(l + 1) / nF)) + "%",
sep=' ',
end='\r')
continue
GCI = SE_VQ_varF0(data_frame, fs, f0=f0_frame)
g_iaif = IAIF(data_frame, fs, GCI)
g_iaif = g_iaif - np.mean(g_iaif)
g_iaif = g_iaif / max(abs(g_iaif))
glottal = cumtrapz(g_iaif)
glottal = glottal - np.mean(glottal)
glottal = glottal / max(abs(glottal))
startf0 = startf0 + stepf0
stopf0 = stopf0 + stepf0
gci_s = GCI[:]
GCId = np.diff(gci_s)
avgGCIt[l] = np.mean(GCId / fs)
varGCIt[l] = np.std(GCId / fs)
NAQ, QOQ, T1, T2, H1H2, HRF = get_vq_params(glottal, g_iaif, fs, GCI)
avgNAQt[l] = np.mean(NAQ)
varNAQt[l] = np.std(NAQ)
avgQOQt[l] = np.mean(QOQ)
varQOQt[l] = np.std(QOQ)
avgH1H2t[l] = np.mean(H1H2)
varH1H2t[l] = np.std(H1H2)
avgHRFt[l] = np.mean(HRF)
varHRFt[l] = np.std(HRF)
print("frame " + str(l) + " from " + str(nF) +
"-" * int(100 * l / nF) + ">" + str(int(100 *
(l + 1) / nF)) + "%",
sep=' ',
end='\r')
if flag_plots:
plot_glottal(data_frame, fs, GCI, g_iaif, glottal, avgGCIt[l],
varGCIt[l])
if len(rmwin) > 0:
varGCI = np.delete(varGCIt, rmwin)
avgNAQ = np.delete(avgNAQt, rmwin)
varNAQ = np.delete(varNAQt, rmwin)
avgQOQ = np.delete(avgQOQt, rmwin)
varQOQ = np.delete(varQOQt, rmwin)
avgH1H2 = np.delete(avgH1H2t, rmwin)
varH1H2 = np.delete(varH1H2t, rmwin)
avgHRF = np.delete(avgHRFt, rmwin)
varHRF = np.delete(varHRFt, rmwin)
return varGCI, avgNAQ, varNAQ, avgQOQ, varQOQ, avgH1H2, varH1H2, avgHRF, varHRF
else:
return varGCIt, avgNAQt, varNAQt, avgQOQt, varQOQt, avgH1H2t, varH1H2t, avgHRFt, varHRFt