def wav2mgcf0(x, order=34, frame_window=512, zerofill_width=1024, shift_window=64, pass_const=0.4, min_pitch=20, max_pitch=500, mgcep_gamma=2):
# Convert from int to float32, but keep numbers as integers
x = x.astype('float32')
# Compute pitch and voicing strength
f0 = pysptk.rapt(x.astype(np.float32), 16000, shift_window, otype="pitch", min=min_pitch, max=max_pitch)
voicing_str = np.empty((f0.shape[0], 5), dtype=np.float32)
for i in range(5):
h = h_filters[i]
xf = lfilter(h, 1, x)
voicing_str[:, i] = pysptk.rapt(xf.astype(np.float32), 16000, shift_window, otype="mixed", min=20, max=500)[:,3]
# Compute MGC coefficients
mgc_cmd = 'frame -l {} -p {} | window -l {} -L {} | mgcep -m {} -a {} -c {} -l {} -e 0.0012'.format(frame_window, shift_window, frame_window, zerofill_width, order, pass_const, mgcep_gamma, zerofill_width)
p = subprocess.Popen(mgc_cmd, stdout=subprocess.PIPE, stdin=subprocess.PIPE, shell=True)
stdout, stderr = p.communicate(x.tobytes())
mgc = np.fromstring(stdout, dtype='float32').reshape((len(f0), order+1))
try:
idx = np.where( abs(voicing_str).sum(axis= 1) > 10)[0][0]
except IndexError:
idx = len(f0)
return mgc[:idx], f0[:idx], voicing_str[:idx]
def test_rapt_regression():
# Grund truth data is generated by:
#
# $ wav2raw pysptk/example_audio_data/arctic_a0007.wav
#
# $ x2x +sf ./pysptk/example_audio_data/arctic_a0007.raw | \
# pitch -a 0 -s 16 -p 80 -L 60 -H 240 -o 0 > \
# arctic_a007_p16_L60_H240_o0_rapt.pitch
#
# $ dmp +f arctic_a007_p16_L60_H240_o0_rapt.pitch | awk '{print $2}' >\
# arctic_a007_p16_L60_H240_o0_rapt.txt
#
# $ pitch -h
# ...
#
# SPTK: version 3.8
# CVS Info: $Id: pitch.c,v 1.46 2014/12/11 08:30:43 uratec Exp $
ground_truth_path = join(dirname(__file__), "data",
"arctic_a007_p16_L60_H240_o0_rapt.txt")
with open(ground_truth_path) as f:
ground_truth = np.asarray([float(s)
for s in [l for l in f.readlines()]])
ground_truth = ground_truth.astype(np.float32)
fs, x = wavfile.read(pysptk.util.example_audio_file())
assert fs == 16000
# Since SPTK might have memory corruption bug and the result might be
# non-deterministic, test it with multiple time...
for _ in range(5):
f0 = pysptk.rapt(x.astype(np.float32), fs=fs, hopsize=80,
min=60, max=240, voice_bias=0.0, otype=0)
assert np.allclose(ground_truth, f0)
def pysptk_featurize(audiofile):
labels = list()
features = list()
fs, x = wavfile.read(audiofile)
f0_swipe = pysptk.swipe(x.astype(np.float64),
fs=fs,
hopsize=80,
min=60,
max=200,
otype="f0")
features = features + stats(f0_swipe)
labels = stats_labels('f0_swipe', labels)
f0_rapt = pysptk.rapt(x.astype(np.float32),
fs=fs,
hopsize=80,
min=60,
max=200,
otype="f0")
features = features + stats(f0_rapt)
labels = stats_labels('f0_rapt', labels)
mgc = pysptk.mgcep(xw, 20, 0.0, 0.0)
features = features + stats(mgc)
labels = stats_labels('mel-spectrum envelope', labels)
return features, labels
def test_rapt_regression():
# Grund truth data is generated by:
#
# $ wav2raw pysptk/example_audio_data/arctic_a0007.wav
#
# $ x2x +sf ./pysptk/example_audio_data/arctic_a0007.raw | \
# pitch -a 0 -s 16 -p 80 -L 60 -H 240 -o 0 > \
# arctic_a007_p16_L60_H240_o0_rapt.pitch
#
# $ dmp +f arctic_a007_p16_L60_H240_o0_rapt.pitch | awk '{print $2}' >\
# arctic_a007_p16_L60_H240_o0_rapt.txt
#
# $ pitch -h
# ...
#
# SPTK: version 3.8
# CVS Info: $Id: pitch.c,v 1.46 2014/12/11 08:30:43 uratec Exp $
ground_truth_path = join(dirname(__file__), "data",
"arctic_a007_p16_L60_H240_o0_rapt.txt")
with open(ground_truth_path) as f:
ground_truth = np.asarray([float(s)
for s in [l for l in f.readlines()]])
ground_truth = ground_truth.astype(np.float32)
fs, x = wavfile.read(pysptk.util.example_audio_file())
assert fs == 16000
# Since SPTK might have memory corruption bug and the result might be
# non-deterministic, test it with multiple time...
for i in range(5):
f0 = pysptk.rapt(x.astype(np.float32), fs=fs, hopsize=80,
min=60, max=240, voice_bias=0.0, otype=0)
assert np.allclose(ground_truth, f0)
def f0gram(filename,hparams,index):
fs, x = wavfile.read(filename)
if x.ndim > 1:
x = np.mean(x,axis=1)
x = librosa.core.resample(x.astype(np.float32),fs,hparams.sample_rate,'kaiser_best')
x = x[index[0]:index[1]]
f0_rapt = pysptk.rapt(x, fs=hparams.sample_rate, hopsize=256, min=10, max=7600, otype="f0")
return f0_rapt
def wav2f0(self, wav_file, f0_dir):
sr, x = io.wavfile.read(wav_file)
bn = os.path.basename(wav_file)
f0_file_path = os.path.join(f0_dir, bn.split('.')[0] + ".f0")
f0 = pysptk.rapt(x.astype(np.float32),
fs=sr,
hopsize=self.hop_length,
min=self.pitch_floor,
max=self.pitch_ceiling,
otype="f0").astype(np.float32)
f0.tofile(f0_file_path)
return f0_file_path
def process(filename):
'''
The function decomposes a wav file into F0, mel-cepstral coefficients, and aperiodicity
:param filename: path to wav file
:return: .lf0, .mgc and .bap files
'''
# pdb.set_trace()
file_id = os.path.basename(filename).split(".")[0]
print('\n' + file_id)
### WORLD ANALYSIS -- extract vocoder parameters ###
# x, fs = librosa.core.load(filename, sr=16000)
fs, x = wavfile.read(filename)
# warnning this parameter is important
alpha = pysptk.util.mcepalpha(fs)
hopesize = int(0.005 * fs)
# pdb.set_trace()
f0 = pysptk.rapt(x.astype(np.float32),
fs=fs,
hopsize=hopesize,
min=60,
max=600,
voice_bias=0.0,
otype=1)
f0 = f0.astype(np.float64)
x = x.astype(np.float64) / (2**15)
_, timeaxis = pyworld.harvest(x,
fs,
frame_period=5,
f0_floor=60.0,
f0_ceil=600)
spectrogram = pyworld.cheaptrick(x, f0, timeaxis, fs)
aperiodicity = pyworld.d4c(x, f0, timeaxis, fs)
f0 = f0[:, None]
lf0 = f0.copy()
lf0 = lf0.astype(np.float32)
nonzero_indices = np.where(f0 != 0)
lf0[nonzero_indices] = np.log(f0[nonzero_indices])
zero_indices = np.where(f0 == 0)
lf0[zero_indices] = -1.0E+10
write_binfile(lf0,
os.path.join(lf0_dir, file_id + '.lf0'),
dtype=np.float32)
mc = pysptk.sp2mc(spectrogram, mcsize, alpha=alpha)
mc = mc.astype(np.float32)
write_binfile(mc,
os.path.join(mgc_dir, file_id + '.mgc'),
dtype=np.float32)
bap = pyworld.code_aperiodicity(aperiodicity, fs)
bap = bap.astype(np.float32)
write_binfile(bap,
os.path.join(bap_dir, file_id + '.bap'),
dtype=np.float32)
def get_f0(wav, sr, fmin=60, fmax=400, spec_len=None):
if wav.dtype == np.float32:
wav = wav * 32768.0
f0 = pysptk.rapt(wav,
fs=sr,
hopsize=args.hop_length,
min=fmin,
max=fmax,
otype='f0')
f0norm = normalize(f0, xmin=0, xmax=fmax)
if spec_len is not None and spec_len != f0.shape[0]:
n_pad = spec_len - f0.shape[0]
f0norm = np.pad(f0norm, [0, n_pad]) # pad into spec length
f0norm = padding_reduction(f0norm, r=args.r)
# f0norm = f0norm[::args.r]
return f0norm
def process_file(data, fs, window_len, log):
chunk_size = int(window_len / 1000 * fs)
pdqs = []
for chunk in [
data[idx:idx + chunk_size]
for idx in range(0, len(data), chunk_size)
]:
f0 = pysptk.rapt(chunk.astype(np.float32),
fs=fs,
hopsize=HOP_SIZE,
min=60,
max=600,
otype="f0")
f0_positive = f0[f0 > 0]
if f0_positive.size != 0:
std = np.std(f0_positive)
avg = np.average(f0_positive)
if avg == 0:
pdq = 0
else:
pdq = std / avg
else:
pdq = 0
pdqs.append(float(pdq))
avg_pdq = np.average(pdqs)
log(f"Avg pitch quotient: {avg_pdq}")
return {
'values': pdqs,
'frame_length': window_len, #ms
'avg': float(avg_pdq)
}
def extract_emphasis(self, index_chunk):
chunks_path = self.video_path + "chunks/"
if (index_chunk <= 9):
file = self.chunks_path + "chunk-0" + str(index_chunk) + ".wav"
else:
file = self.chunks_path + "chunk-" + str(index_chunk) + ".wav"
fs, x = wavfile.read(file)
assert fs == 16000
f0_swipe = pysptk.rapt(x.astype(np.float32),
fs=fs,
hopsize=20,
min=60,
otype="pitch")
a = []
f = []
X_Frequecies_Vector = []
for w in f0_swipe:
if w != 0:
f.append(w)
pitch = float(np.median(f))
if (np.isnan(pitch)):
pitch = 0
#dbs = 20*np.log10( np.sqrt(np.mean(x**2)) )
dbs = 2000 * np.log10(np.sqrt(np.mean(
x**2))) / 5 * (self.n_chunks - index_chunk)
if (np.isnan(dbs)):
dbs = 0
return pitch, dbs
def __test(x, fs, hopsize, min, max, otype):
f0 = pysptk.rapt(x, fs, hopsize, min=min, max=max, otype=otype)
assert np.all(np.isfinite(f0))
if otype == 1:
assert np.all(f0 >= 0)
def feature_extraction(x, fs, feats_df, lp_ord, ID, label):
#Extract features from signal x (identified as ID), and concatenate them to dataframe feats_df
#Features' reference: (see Appendix)
#[1]https://link.springer.com/article/10.1007/s10439-013-0741-6
#[2]https://espace.library.uq.edu.au/data/UQ_344963/s41943203_phd_submission.pdf?dsi_version=c5434db897ab74b192ca295a9eeca041&Expires=1585086202&Key-Pair-Id=APKAJKNBJ4MJBJNC6NLQ&Signature=c8k8DmG~KIxg0ToTO8rebm2MzHneCzJGkjSFRB7BYTEQ-MHXEr0ocHmISrldP3hFf9qmeiL11ezyefcNeRVeKIQ9PVjOl9pn7rXWcjA1o2voPn1VnDd8n7G2cT31apdj0LNMclhlXRPnCsGD66qDRqa3d-xaqqXhEqU73aw3ZgBgroO213MfJOqFhJxxXo2QEia0bSlDRTeX9KhSczFK-IFTPC6GwFL2L04por8pQRI3HF7E3f26O9zp9OhkwxSU9qfJah20WxZLA4PxREdv7JGoVBinR6T0mTcIaQi~B4IzYjSPSsTTADMNk5znVYIvSqgtMT~DY~qwlfq4SRdFjQ__
#do features in a frame-basis
x_frames = spe_feats.sigproc.framesig(
x, config.frame_len, config.frame_step,
config.win_func) #DOUBT: should I use window or not?
#at least for formant estimation i should
nr_frames = x_frames.shape[0]
#print(nr_frames)
#0)Wavelets #TODO
#DOUBT: if log-energy feature is included, should I also include the first mfcc coefficient (c0) ?
#1)mfcc
mfcc_feat = spe_feats.mfcc(x,
fs,
winlen=config.frame_len_s,
winstep=config.frame_step_s,
numcep=config.cep_num,
winfunc=config.win_func)
#Apply zscore normalization
mfcc_feat = zscore(mfcc_feat, axis=1, ddof=1)
#deltas to capture
mfcc_delta_feat = spe_feats.delta(
mfcc_feat,
1) #mfcc_delta_feat = np.subtract(mfcc_feat[:-1], mfcc_feat[1:]) #same
mfcc_deltadelta_feat = spe_feats.delta(mfcc_delta_feat, 1)
#2)zero-crossing rate
zcr_feat = np.apply_along_axis(get_zcr, 1, x_frames)
#3)Formant frequencies
#using LP-coeffcs-based method
#formant_feat = np.apply_along_axis(get_formants, 1, x_frames, lp_ord, nr_formants)
#Note: for the moment, it seems some frames are ill-conditioned for lp computing,
#current solution - we skip those and fill with NaN values
formants_feat = np.empty((nr_frames, 4))
formants_feat[:] = np.nan
for i_frame in range(0, nr_frames):
try:
formants_feat[i_frame] = get_formants(x_frames[i_frame],
config.lp_ord,
config.nr_formants)
except:
pass
#4)Log-energy
logEnergy_feat = np.apply_along_axis(get_logEnergy, 1, x_frames)
#5)Pitch (F0)
F0_feat = np.apply_along_axis(get_F0, 1, x_frames, fs)
#TODO: compute also F0 with pysptk (a python wrapper for SPTK library), it probably gives better results
#https://github.com/r9y9/pysptk/blob/master
#F0_feat = pysptk.rapt(x.astype(np.float32), fs=fs, hopsize=frame_step, min=50, max=500, ,voice_bias=0.0 ,otype=\"f0\")
#Compare the values between swipe and rapt
#F0_feat = pysptk.swipe(x.astype(np.float64), fs=fs,hopsize = config.frame_step, min=50, max=500, otype="f0")
F0_feat = pysptk.rapt(x.astype(np.float32),
fs=fs,
hopsize=config.frame_step,
min=50,
max=500,
otype="f0")
#right frame size???
#Change the window size from 450 to 40 to 100
# Keep swipe , change min to 50 and max - 500
#EXample pysptk.swipe(x.astype(np.float64), fs=fs, hopsize=80, min=60, max=200, otype="f0")
#6)Kurtosis
kurt_feat = np.apply_along_axis(kurtosis, 1, x_frames)
#7)Bispectrum Score (BGS)
#TODO: see PhD thesis for more info on this feature
#8)Non-Gaussianity Score (NGS)
#TODO: see PhD thesis for more info on this feature
#9) Adding skewness as measure of non-gaussianity (not in paper)
skew_feat = np.apply_along_axis(skew, 1, x_frames)
#DOUBT: 10) Shannon entropy GETTING -inf in all cases, WHY??? Don't include until fixed
entropy_feat = np.apply_along_axis(get_entropy, 1, x_frames)
#TODO: add small value in all entries, this may fix the problem
mfcc_cols = ['mfcc_%s' % s for s in range(0, config.cep_num)]
mfcc_delta_cols = ['mfcc_d%s' % s for s in range(0, config.cep_num)]
mfcc_deltadelta_cols = ['mfcc_dd%s' % s for s in range(0, config.cep_num)]
formants_cols = ['F%s' % s for s in range(1, config.nr_formants + 1)]
feats_segment = pd.concat([
pd.DataFrame({
'Id': ID,
'kurt': kurt_feat,
'logEnergy': logEnergy_feat,
'zcr': zcr_feat,
'F0': F0_feat,
'skewness': skew_feat,
'label': label,
'entropy': entropy_feat
}),
pd.DataFrame(mfcc_feat, columns=mfcc_cols),
pd.DataFrame(mfcc_delta_feat, columns=mfcc_delta_cols),
pd.DataFrame(mfcc_deltadelta_feat, columns=mfcc_deltadelta_cols),
pd.DataFrame(formants_feat, columns=formants_cols)
],
axis=1)
#print(nr_frames)
feats_df = feats_df.append(feats_segment, ignore_index=True, sort=False)
return feats_df
def __test(x, fs, hopsize, min, max, otype):
f0 = pysptk.rapt(x, fs, hopsize, min=min, max=max, otype=otype)
assert np.all(np.isfinite(f0))
if otype == 1:
assert np.all(f0 >= 0)
order = 34
frame_window = 512
zerofill_width = 1024
shift_window = 64
pass_const = 0.4
min_pitch = 20
max_pitch = 500
mgcep_gamma = 2
e = 0.0012
#-a 0 -s 16
SPTK.rapt(raw.astype(np.float32),
fs=16000,
hopsize=shift_window,
min=min_pitch,
max=max_pitch,
otype="pitch")
f0[1200:1300]
import pipes, os, subprocess, tempfile
import numpy as np
frame_cmd = 'frame -l {} -p {}'.format(frame_window, shift_window)
raw2 = raw.astype('float32')
#raw2 = np.arange(5000).astype('float32')
p = subprocess.Popen(frame_cmd,
mgc2, f02 = wav2mgcf0(raw)
order=34
frame_window=512
zerofill_width=1024
shift_window=64
pass_const=0.4
min_pitch=20
max_pitch=500
mgcep_gamma=2
e = 0.0012
#-a 0 -s 16
SPTK.rapt(raw.astype(np.float32), fs=16000,
hopsize=shift_window, min=min_pitch, max=max_pitch, otype="pitch")
f0[1200:1300]
import pipes, os, subprocess, tempfile
import numpy as np
frame_cmd = 'frame -l {} -p {}'.format(frame_window, shift_window)
raw2 = raw.astype('float32')
#raw2 = np.arange(5000).astype('float32')
p = subprocess.Popen(frame_cmd, stdout=subprocess.PIPE, stdin=subprocess.PIPE, shell=True)
stdout, stderr = p.communicate(raw2.tobytes())
def get_f0(waveform,
sample_rate,
hop_length_seconds=0.01,
method='swipe',
f0_min=60,
f0_max=300):
"""Compute the F0 contour using PYSPTK: https://github.com/r9y9/pysptk/.
Args:
waveform (np.array, [T, ]): waveform over which to compute f0
sample_rate (int > 0): number of samples per second in waveform
hop_length (int): hop size argument in pysptk.swipe. Corresponds to hopsize
in the window sliding of the computation of f0.
method (str): is one of 'swipe' or 'rapt'. Define which method to use for f0
calculation. See https://github.com/r9y9/pysptk
Returns:
dict: Dictionary containing keys:
"contour" (np.array, [1, t1]): f0 contour of waveform. Contains unvoiced
frames.
"values" (np.array, [1, t2]): nonzero f0 values waveform. Note that this
discards all unvoiced frames. Use to compute mean, std, and other statistics.
"mean" (float): mean of the f0 contour.
"std" (float): standard deviation of the f0 contour.
"""
assert method in (
'swipe',
'rapt'), "The method argument should be one of 'swipe' or 'rapt'."
hop_length = numseconds_to_numsamples(hop_length_seconds, sample_rate)
if method == 'swipe':
f0_contour = swipe(
waveform.astype(np.float64),
fs=sample_rate,
hopsize=hop_length,
min=f0_min,
max=f0_max,
otype="f0",
)[np.newaxis, :]
elif method == 'rapt':
# For this estimation, waveform needs to be in the int PCM format.
f0_contour = rapt(
np.round(waveform * 32767).astype(np.float32),
fs=sample_rate,
hopsize=hop_length,
min=f0_min,
max=f0_max,
otype="f0",
)[np.newaxis, :]
# Remove unvoiced frames.
f0_values = f0_contour[:, np.where(f0_contour[0, :] != 0)][0]
f0_mean = np.mean(f0_values[0])
f0_std = np.std(f0_values[0])
return {
"contour": f0_contour,
"values": f0_values,
"mean": f0_mean,
"std": f0_std,
}
