def main(argv):
fbank_mx = features.mel_fbank_mx(winlen_nfft=WINDOWSIZE / SOURCERATE,
fs=fs,
NUMCHANS=NUMCHANS,
LOFREQ=LOFREQ,
HIFREQ=HIFREQ)
scp_list = sys.argv[1]
vad_dir = sys.argv[2]
wav_dir = sys.argv[3]
ubm_file = sys.argv[4]
v_file = sys.argv[5]
out_dir = sys.argv[6]
print 'Loading UBM from', ubm_file
ubm_weights, ubm_means, ubm_covs = load_ubm(ubm_file)
GMM = gmm.gmm_eval_prep(ubm_weights, ubm_means, ubm_covs)
numG = ubm_means.shape[0]
dimF = ubm_means.shape[1]
# normalization of statistics - precomputing matrices
if ubm_covs.shape[1] == dimF:
ubm_norm = 1 / np.sqrt(ubm_covs);
print 'Loading T matrix from ', v_file, '...'
v = np.loadtxt(v_file, dtype=np.float32)
print 'Computing MVVT ...'
MVVT = iv.compute_VtV(v, numG)
print 'Loading list of files to process from ' + scp_list
seg_list = np.atleast_1d(np.loadtxt(scp_list, dtype=object))
# extract all sub-dir names
for dir in set(map(os.path.dirname, seg_list)):
mkdir_p(out_dir + '/' + dir)
# go over the scp and process the audio files
for ii, fn in enumerate(seg_list, 1):
try:
print 'Processing ', ii, '/', len(seg_list), fn
np.random.seed(777)
wav_file = wav_dir + '/' + fn + '.wav'
raw_file = wav_dir + '/' + fn + '.raw'
lab_file = vad_dir + '/' + fn + '.lab.gz'
ivec_out_file = out_dir + '/' + fn + '.ivec'
if os.path.isfile(wav_file):
print ' Reading wave file from ' + wav_file,
rate, sig = spiowav.read(wav_file)
if rate != 8000:
raise Exception(
'The input file ' + wav_file + ' is expected to be in 8000 Hz sampling rate, but ' + repr(
rate) + ' Hz detected')
else:
print ' Reading raw 8000Hz, 16bit-s, 1c, file from ' + raw_file,
sig = np.fromfile(raw_file, dtype='int16')
print '[t=' + repr(len(sig) / fs) + ' seconds, fs=' + repr(fs) + 'Hz, n=' + repr(len(sig)) + ' samples]'
if ADDDITHER > 0.0:
print ' Adding dither'
sig = features.add_dither(sig, ADDDITHER)
print ' Extracting features',
fea = features.mfcc_htk(sig,
window=WINDOWSIZE / SOURCERATE,
noverlap=(WINDOWSIZE - TARGETRATE) / SOURCERATE,
fbank_mx=fbank_mx,
_0='first',
NUMCEPS=NUMCEPS,
RAWENERGY=RAWENERGY,
PREEMCOEF=PREEMCOEF,
CEPLIFTER=CEPLIFTER,
ZMEANSOURCE=ZMEANSOURCE,
ENORMALISE=ENORMALISE,
ESCALE=0.1,
SILFLOOR=50.0,
USEHAMMING=True)
print '[n=' + repr(len(fea)) + ' frames]'
print ' Adding derivatives'
# [add_deriv] step
fea = features.add_deriv(fea, (deltawindow, accwindow))
print ' Reshaping to SFeaCat convention'
# [reshape] step
fea = fea.reshape(fea.shape[0], 3, -1).transpose((0, 2, 1)).reshape(fea.shape[0],
-1) # re-order coeffs like SFeaCut
if vad_dir == "auto":
print ' Computing VAD '
vad, n_regions, n_frames = compute_vad(sig, win_length=WINDOWSIZE / SOURCERATE,
win_overlap=(WINDOWSIZE - TARGETRATE) / SOURCERATE)[:len(fea)]
else:
print ' Loading VAD definition from ' + lab_file
vad, n_regions, n_frames = load_vad_lab_as_bool_vec(lab_file)[:len(fea)]
print ' Applying VAD [#frames=' + repr(n_frames) + ', #regions=' + repr(n_regions) + ']'
fea = fea[vad, ...]
if len(fea) < 3:
raise NoVadException('Too few frames left: ' + str(len(fea)))
print ' Applying floating CMVN'
fea = features.cmvn_floating(fea, cmvn_lc, cmvn_rc, unbiased=True)
n_data, d_data = fea.shape
l = 0;
lc = 0
n = np.zeros((numG), dtype=np.float32)
f = np.zeros((numG, dimF), dtype=np.float32)
print ' Computing stats ...',
# Note that we compute the stats in in sub-chunks due to memory optimization
#
seq_data = split_seq(range(n_data), 1000)
for i in range(len(seq_data)):
dd = fea[seq_data[i], :]
l1, n1, f1 = gmm.gmm_eval(dd, GMM, return_accums=1)
l = l + l1.sum()
lc = lc + l1.shape[0]
n = n + n1;
f = f + f1;
print '[avg llh=' + repr(l / lc) + ', #frames=' + repr(n_data) + ']'
n, f = normalize_stats(n, f, ubm_means, ubm_norm)
f = row(f.astype(v.dtype))
n = row(n.astype(v.dtype))
print ' Computing i-vector'
w = iv.estimate_i(n, f, v, MVVT).T
# write it to the disk
print ' Saving ivec to:', ivec_out_file
# np.savetxt(ivec_out_file, w.ravel(), newline=' ', fmt='%f')
ivio.write_binary_ivector(ivec_out_file, w.ravel(), n_data / 100.0)
except NoVadException as e:
print e
print "Warning: No features generated for segment: " + fn
except:
raise
raise Exception(
'The input file ' + wav_file +
' is expected to be in 8000 Hz sampling rate, but ' +
repr(rate) + ' Hz detected')
else:
print ' Reading raw 8000Hz, 16bit-s, 1c, file from ' + raw_file,
sig = np.fromfile(raw_file, dtype='int16')
print '[t=' + repr(
len(sig) / fs) + ' seconds, fs=' + repr(fs) + 'Hz, n=' + repr(
len(sig)) + ' samples]'
if ADDDITHER > 0.0:
print ' Adding dither'
sig = features.add_dither(sig, ADDDITHER)
print ' Extracting features',
fea = features.mfcc_htk(sig,
window=WINDOWSIZE / SOURCERATE,
noverlap=(WINDOWSIZE - TARGETRATE) /
SOURCERATE,
fbank_mx=fbank_mx,
_0='first',
NUMCEPS=NUMCEPS,
RAWENERGY=RAWENERGY,
PREEMCOEF=PREEMCOEF,
CEPLIFTER=CEPLIFTER,
ZMEANSOURCE=ZMEANSOURCE,
ENORMALISE=ENORMALISE,
ESCALE=0.1,
fbank_mx = features.mel_fbank_mx(winlen,
samplerate,
NUMCHANS=64,
LOFREQ=20.0,
HIFREQ=7600,
htk_bug=False)
else:
raise ValueError(
f'Only 8kHz and 16kHz are supported. Got {samplerate} instead.'
)
LC = 150
RC = 149
np.random.seed(3) # for reproducibility
signal = features.add_dither(
(signal * 2**15).astype(int))
for segnum in range(len(labs)):
seg = signal[labs[segnum, 0]:labs[segnum, 1]]
if seg.shape[
0] > 0.01 * samplerate: # process segment only if longer than 0.01s
# Mirror noverlap//2 initial and final samples
seg = np.r_[seg[noverlap // 2 - 1::-1], seg,
seg[-1:-winlen // 2 - 1:-1]]
fea = features.fbank_htk(seg,
window,
noverlap,
fbank_mx,
USEPOWER=True,
ZMEANSOURCE=True)
fea = features.cmvn_floating_kaldi(
def process_wav(self, wav_file, mode="ivector", vad_dir="auto"):
if mode not in ["ivector", "statistics", "mfcc"]:
return False
else:
# all constans are initialized in __init__() method
# READ WAVE AND COMPUTE IVECTOR
sig, rate = librosa.load(wav_file)
#print(librosa.get_duration(sig, rate))
# wav conversion
sig, rate = self.wav_conversion(sig, rate)
#import sounddevice as sd
#sd.play(sig, rate)
if rate != 8000:
raise Exception(
'The input file ' + wav_file +
' is expected to be in 8000 Hz sampling rate, but ' +
repr(rate) + ' Hz detected')
# info about singnal printed
print '[t=' + repr(
len(sig) / fs) + ' seconds, fs=' + repr(fs) + 'Hz, n=' + repr(
len(sig)) + ' samples]'
if ADDDITHER > 0.0:
print ' Adding dither'
sig = features.add_dither(sig, ADDDITHER)
print ' Extracting features',
fea = features.mfcc_htk(sig,
window=WINDOWSIZE / SOURCERATE,
noverlap=(WINDOWSIZE - TARGETRATE) /
SOURCERATE,
fbank_mx=fbank_mx,
_0='first',
NUMCEPS=NUMCEPS,
RAWENERGY=RAWENERGY,
PREEMCOEF=PREEMCOEF,
CEPLIFTER=CEPLIFTER,
ZMEANSOURCE=ZMEANSOURCE,
ENORMALISE=ENORMALISE,
ESCALE=0.1,
SILFLOOR=50.0,
USEHAMMING=True)
print '[n=' + repr(len(fea)) + ' frames]'
print ' Adding derivatives'
# [add_deriv] step
fea = features.add_deriv(fea, (deltawindow, accwindow))
print ' Reshaping to SFeaCat convention'
# [reshape] step
fea = fea.reshape(fea.shape[0], 3, -1).transpose(
(0, 2, 1)).reshape(fea.shape[0],
-1) # re-order coeffs like SFeaCut
if vad_dir == "auto":
print ' Computing VAD '
vad, n_regions, n_frames = self.compute_vad(
sig,
win_length=WINDOWSIZE / SOURCERATE,
win_overlap=(WINDOWSIZE - TARGETRATE) /
SOURCERATE)[:len(fea)]
print ' Applying VAD [#frames=' + repr(
n_frames) + ', #regions=' + repr(n_regions) + ']'
fea = fea[0:len(vad), ...]
fea = fea[vad, ...]
if len(fea) < 3:
raise NoVadException('Too few frames left: ' +
str(len(fea)))
print ' Applying floating CMVN'
fea = features.cmvn_floating(fea,
cmvn_lc,
cmvn_rc,
unbiased=True)
if mode == "mfcc":
return fea
n_data, d_data = fea.shape
l = 0
lc = 0
n = np.zeros((self.numG), dtype=np.float32)
f = np.zeros((self.numG, self.dimF), dtype=np.float32)
print ' Computing stats ...',
# Note that we compute the stats in in sub-chunks due to memory optimization
#
seq_data = self.split_seq(range(n_data), 1000)
for i in range(len(seq_data)):
dd = fea[seq_data[i], :]
l1, n1, f1 = gmm.gmm_eval(dd, self.GMM, return_accums=1)
l = l + l1.sum()
lc = lc + l1.shape[0]
n = n + n1
f = f + f1
print '[avg llh=' + repr(
l / lc) + ', #frames=' + repr(n_data) + ']'
n, f = self.normalize_stats(n, f, self.ubm_means,
self.ubm_norm)
f = self.row(f.astype(self.v.dtype))
n = self.row(n.astype(self.v.dtype))
if mode == "statistics":
return f, n
print ' Computing i-vector'
w = iv.estimate_i(n, f, self.v, self.MVVT).T
print "IVECTOR", w
if mode == "ivector":
return w
fs,
NUMCHANS=40,
LOFREQ=20.0,
HIFREQ=7600,
htk_bug=False)
LC = 150
RC = 149
with open(out_seg_fn, "w") as seg_file:
with open(out_ark_fn, "wb") as ark_file:
for fn in file_names:
labs = (
np.loadtxt(in_lab_dir + "/" + fn + ".lab", usecols=(0, 1)) *
16000).astype(int)
signal, samplerate = sf.read(in_flac_dir + "/" + fn + ".flac")
signal = features.add_dither(
(signal * 2**(samplerate / 1000 - 1)).astype(int))
for segnum in range(len(labs)):
seg = signal[labs[segnum, 0]:labs[segnum, 1]]
seg = np.r_[
seg[noverlap // 2 - 1::-1], seg,
seg[-1:-winlen // 2 -
1:-1]] # Mirror noverlap//2 initial and final samples
fea = features.fbank_htk(seg,
window,
noverlap,
fbank_mx,
USEPOWER=True,
ZMEANSOURCE=True)
fea = features.cmvn_floating_kaldi(fea,
LC,
RC,