def get_modspec(args, srate=16000, window=np.hanning):
wavs = args.scp
add_reverb = args.add_reverb
set_unity_gain = args.set_unity_gain
nmodulations = args.nmodulations
order = args.order
fduration = args.fduration_modspec
frate = args.frate
nfilters = args.nfilters
fbank = createFbank(nfilters, int(2 * fduration * srate), srate)
if add_reverb:
if add_reverb == 'small_room':
sr_r, rir = read('./RIR/RIR_SmallRoom1_near_AnglA.wav')
rir = rir[:, 1]
rir = rir / np.power(2, 15)
elif add_reverb == 'large_room':
sr_r, rir = read('./RIR/RIR_LargeRoom1_far_AnglA.wav')
rir = rir[:, 1]
rir = rir / np.power(2, 15)
elif add_reverb == 'clean':
print('%s: No reverberation added!' % sys.argv[0])
else:
raise ValueError('Invalid type of reverberation!')
with open(wavs, 'r') as fid:
all_feats = {}
for line in fid:
tokens = line.strip().split()
uttid, inwav = tokens[0], ' '.join(tokens[1:])
if inwav[-1] == '|':
proc = subprocess.run(inwav[:-1],
shell=True,
stdout=subprocess.PIPE)
sr, signal = read(io.BytesIO(proc.stdout))
else:
sr, signal = read(inwav)
assert sr == srate, 'Input file has different sampling rate.'
# I want to work with numbers from 0 to 1 so....
signal = signal / np.power(2, 15)
if add_reverb:
if not add_reverb == 'clean':
signal = addReverb(signal, rir)
time_frames = np.array([
frame
for frame in getFrames(signal, srate, frate, fduration, window)
])
cos_trans = freqAnalysis.dct(time_frames) / np.sqrt(
2 * int(srate * fduration))
[frame_num, ndct] = np.shape(cos_trans)
if set_unity_gain:
feats = np.zeros((frame_num, nfilters * (nmodulations - 1)))
else:
feats = np.zeros((frame_num, nfilters * nmodulations))
print('%s: Computing Features for file: %s' % (sys.argv[0], uttid))
sys.stdout.flush()
for i in range(frame_num):
if set_unity_gain:
each_feat = np.zeros([nfilters, nmodulations - 1])
else:
each_feat = np.zeros([nfilters, nmodulations])
for j in range(nfilters):
filt = fbank[j, 0:-1]
band_dct = filt * cos_trans[i, :]
xlpc, gg = computeLpcFast(
band_dct, order) # Compute LPC coefficients
if set_unity_gain:
gg = 1
mod_spec = computeModSpecFromLpc(gg, xlpc, nmodulations)
if set_unity_gain:
mod_spec = mod_spec[1:]
each_feat[j, :] = mod_spec
if set_unity_gain:
each_feat = np.reshape(each_feat,
(1, nfilters * (nmodulations - 1)))
else:
each_feat = np.reshape(each_feat,
(1, nfilters * nmodulations))
feats[i, :] = each_feat
all_feats[uttid] = feats
return all_feats
def getFeats(args, srate=16000, window=np.hanning):
wavs = args.scp
scp_type = args.scp_type
outfile = args.outfile
add_reverb = args.add_reverb
coeff_0 = args.coeff_0
coeff_n = args.coeff_n
order = args.order
fduration = args.fduration
frate = args.frate
nfilters = args.nfilters
kaldi_cmd = args.kaldi_cmd
# Set up mel-filterbank
fbank_type = args.fbank_type.strip().split(',')
if args.complex_modulation:
dur = int(fduration * srate)
else:
dur = int(2 * fduration * srate)
if fbank_type[0] == "mel":
if len(fbank_type) < 2:
raise ValueError('Mel filter bank not configured properly....')
fbank = createFbank(nfilters,
dur,
srate,
warp_fact=float(fbank_type[1]))
elif fbank_type[0] == "cochlear":
if len(fbank_type) < 6:
raise ValueError(
'Cochlear filter bank not configured properly....')
if int(fbank_type[3]) == 1:
print(
'%s: Alpha is fixed and will not change as a function of the center frequency...'
% sys.argv[0])
fbank = createFbankCochlear(nfilters,
dur,
srate,
om_w=float(fbank_type[1]),
alp=float(fbank_type[2]),
fixed=int(fbank_type[3]),
bet=float(fbank_type[4]),
warp_fact=float(fbank_type[5]))
else:
raise ValueError(
'Invalid type of filter bank, use mel or cochlear with proper configuration'
)
coeff_num = coeff_n - coeff_0 + 1
if args.keep_even:
temp = np.arange(0, coeff_num)
if coeff_0 % 2 == 0:
# It starts from odd coefficients
feat_len = temp[1::2].shape[0]
else:
feat_len = temp[0::2].shape[0]
elif args.complex_modulation:
if args.absolute_value:
feat_len = coeff_num
else:
feat_len = 2 * coeff_num
else:
feat_len = coeff_num
if args.compensate_noise:
if args.complex_modulation:
fmax = coeff_num / (fduration)
faxis = np.linspace(0, fmax, coeff_n)
else:
fmax = coeff_num / (2 * fduration)
faxis = np.linspace(0, fmax, coeff_n)
if args.no_window:
print('%s: Using square windows' % sys.argv[0])
window = sq_wind
if add_reverb:
if add_reverb == 'small_room':
sr_r, rir = read('./RIR/RIR_SmallRoom1_near_AnglA.wav')
rir = rir[:, 1]
rir = rir / np.power(2, 15)
elif add_reverb == 'large_room':
sr_r, rir = read('./RIR/RIR_LargeRoom1_far_AnglA.wav')
rir = rir[:, 1]
rir = rir / np.power(2, 15)
elif add_reverb == 'clean':
print('%s: No reverberation added!' % sys.argv[0])
else:
raise ValueError('Invalid type of reverberation!')
with open(wavs, 'r') as fid:
all_feats = {}
for line in fid:
tokens = line.strip().split()
uttid, inwav = tokens[0], ' '.join(tokens[1:])
if scp_type == 'wav':
if inwav[-1] == '|':
try:
proc = subprocess.run(inwav[:-1],
shell=True,
stdout=subprocess.PIPE)
sr, signal = read(io.BytesIO(proc.stdout))
skip_rest = False
except Exception:
skip_rest = True
else:
try:
sr, signal = read(inwav)
skip_rest = False
except Exception:
skip_rest = True
assert sr == srate, 'Input file has different sampling rate.'
elif scp_type == 'segment':
try:
cmd = 'wav-copy ' + inwav + ' - '
proc = subprocess.run(cmd,
shell=True,
stdout=subprocess.PIPE)
sr, signal = read(io.BytesIO(proc.stdout))
skip_rest = False
except Exception:
skip_rest = True
else:
raise ValueError(
'Invalid type of scp type, it should be either wav or segment'
)
if not skip_rest:
# I want to work with numbers from 0 to 1 so....
# signal = signal / np.power(2, 15)
if add_reverb:
if not add_reverb == 'clean':
signal = addReverb(signal, rir)
time_frames = np.array([
frame for frame in getFrames(signal, srate, frate,
fduration, window)
])
if args.complex_modulation:
cos_trans = freqAnalysis.ifft(time_frames)
cos_trans = cos_trans[:, :int(fduration * srate / 2)]
else:
cos_trans = freqAnalysis.dct(time_frames) / np.sqrt(
2 * int(srate * fduration))
[frame_num, ndct] = np.shape(cos_trans)
feats = np.zeros((frame_num, nfilters * feat_len))
print('%s: Computing Features for file: %s, also %d' %
(sys.argv[0], uttid, time_frames.shape[0]))
sys.stdout.flush()
for i in range(frame_num):
each_feat = np.zeros([nfilters, feat_len])
for j in range(nfilters):
filt = fbank[j, 0:-1]
band_dct = filt * cos_trans[i, :]
if args.complex_modulation:
xlpc, gg = computeLpcFast(
band_dct, order,
keepreal=False) # Compute LPC coefficients
mod_spec = computeModSpecFromLpc(gg, xlpc, coeff_n)
if args.compensate_noise:
mod_spec = mod_spec * faxis
if args.absolute_value:
temp2 = np.abs(mod_spec[coeff_0 - 1:coeff_n])
else:
temp2 = np.append(
np.real(mod_spec[coeff_0 - 1:coeff_n]),
np.imag(mod_spec[coeff_0 - 1:coeff_n]))
else:
xlpc, gg = computeLpcFast(band_dct, order)
mod_spec = np.real(
computeModSpecFromLpc(gg, xlpc, coeff_n))
if args.compensate_noise:
mod_spec = mod_spec * faxis
if args.absolute_value:
temp2 = np.abs(mod_spec[coeff_0 - 1:coeff_n])
else:
temp2 = mod_spec[coeff_0 - 1:coeff_n]
if args.keep_even:
if coeff_0 % 2 == 0:
each_feat[j, :] = temp2[1::2]
else:
each_feat[j, :] = temp2[0::2]
else:
each_feat[j, :] = temp2
each_feat = np.reshape(each_feat, (1, nfilters * feat_len))
feats[i, :] = each_feat
all_feats[uttid] = feats
dict2Ark(all_feats, outfile, kaldi_cmd)
def getFeats(args, srate=16000, window=np.hanning):
wavs=args.scp
segment=args.segment
outfile=args.outfile
add_reverb=args.add_reverb
set_unity_gain=args.set_unity_gain
nmodulations=args.nmodulations
order=args.order
fduration=args.fduration
frate=args.frate
nfilters=args.nfilters
kaldi_cmd=args.kaldi_cmd
fbank = createFbank(nfilters, int(2*fduration*srate), srate)
if add_reverb:
if add_reverb=='small_room':
sr_r, rir=read('./RIR/RIR_SmallRoom1_near_AnglA.wav')
rir=rir[:,1]
rir=rir/np.power(2,15)
elif add_reverb=='large_room':
sr_r, rir=read('./RIR/RIR_LargeRoom1_far_AnglA.wav')
rir=rir[:,1]
rir=rir/np.power(2,15)
elif add_reverb=='clean':
print('%s: No reverberation added!' % sys.argv[0])
else:
raise ValueError('Invalid type of reverberation!')
wav_in_buffer='' # Wav file that is currently in RAM
# Load Location and Ids of all wav files
wav_ids=[]; wav_locs=[]
with open(wavs, 'r') as fid:
for line in fid:
tokens = line.strip().split()
uttid, inwav = tokens[0], ' '.join(tokens[1:])
wav_ids.append(uttid)
wav_locs.append(inwav)
# Compute features for all the segments
with open(segment, 'r') as fid_s:
all_feats={}
for line_s in fid_s:
token_s = line_s.strip().split()
seg_id=token_s[0]; wav_id=token_s[1]
# Load wav file it is already not in RAM unload
if wav_in_buffer!=wav_id:
wav_in_buffer=wav_id
inwav=wav_locs[wav_ids.index(wav_id)]
if inwav[-1] == '|':
proc = subprocess.run(inwav[:-1], shell=True,
stdout=subprocess.PIPE)
sr, signal_big = read(io.BytesIO(proc.stdout))
else:
sr, signal_big = read(inwav)
assert sr == srate, 'Input file has different sampling rate.'
t_beg=int(float(token_s[2])*sr); t_end=int(float(token_s[3])*sr)
signal=signal_big[t_beg:t_end]
signal=signal/np.power(2,15)
if add_reverb:
if not add_reverb=='clean':
signal=addReverb(signal,rir)
time_frames = np.array([frame for frame in
getFrames(signal, srate, frate, fduration, window)])
cos_trans=freqAnalysis.dct(time_frames)/np.sqrt(2*int(srate * fduration))
[frame_num, ndct]=np.shape(cos_trans)
feats=np.zeros((frame_num,nfilters*nmodulations))
print('%s: Computing Features for file: %s and segment: %s' % (sys.argv[0],wav_id,seg_id))
sys.stdout.flush()
for i in range(frame_num):
each_feat=np.zeros([nfilters,nmodulations])
for j in range(nfilters):
filt=fbank[j,0:-1]
band_dct=filt*cos_trans[i,:]
xlpc, gg=computeLpcFast(band_dct,order) # Compute LPC coefficients
if set_unity_gain:
gg=1
mod_spec=computeModSpecFromLpc(gg,xlpc,nmodulations)
each_feat[j,:]=mod_spec
each_feat=np.reshape(each_feat,(1,nfilters*nmodulations))
feats[i,:]=each_feat
all_feats[seg_id]=feats
dict2Ark(all_feats,outfile,kaldi_cmd)
def get_mfcc(args, srate=16000, window=np.hamming):
wavs = args.scp
add_reverb = args.add_reverb
nfft = args.nfft
context = args.context
fduration = args.fduration_mfcc
frate = args.frate
nfilters = args.nfilters
fbank = createFbank(nfilters, nfft, srate)
if add_reverb:
if add_reverb == 'small_room':
sr_r, rir = read('./RIR/RIR_SmallRoom1_near_AnglA.wav')
rir = rir[:, 1]
rir = rir / np.power(2, 15)
elif add_reverb == 'large_room':
sr_r, rir = read('./RIR/RIR_LargeRoom1_far_AnglA.wav')
rir = rir[:, 1]
rir = rir / np.power(2, 15)
elif add_reverb == 'clean':
print('%s: No reverberation added!' % sys.argv[0])
else:
raise ValueError('Invalid type of reverberation!')
sys.stdout.flush()
with open(wavs, 'r') as fid:
all_feats = {}
for line in fid:
tokens = line.strip().split()
uttid, inwav = tokens[0], ' '.join(tokens[1:])
print('%s: Computing Features for file: %s' % (sys.argv[0], uttid))
sys.stdout.flush()
if inwav[-1] == '|':
proc = subprocess.run(inwav[:-1],
shell=True,
stdout=subprocess.PIPE)
sr, signal = read(io.BytesIO(proc.stdout))
else:
sr, signal = read(inwav)
assert sr == srate, 'Input file has different sampling rate.'
signal = signal / np.power(2, 15)
if add_reverb:
if not add_reverb == 'clean':
signal = addReverb(signal, rir)
time_frames = np.array([
frame
for frame in getFrames(signal, srate, frate, fduration, window)
])
melEnergy_frames = np.log10(
np.matmul(np.abs(fft(time_frames, int(nfft / 2 + 1), axis=1)),
np.transpose(fbank)))
mfcc_feats = dct(melEnergy_frames, axis=1)
mfcc_feats = mfcc_feats[:, 0:13]
if args.context:
mfcc_feats = spliceFeats(mfcc_feats, context)
all_feats[uttid] = mfcc_feats
return all_feats
def extractModSpecFeatures(wavs,
outdir,
phone_map,
phn_file_dir,
get_phone_labels=True,
only_center=True,
around_center=1,
ignore_edge=False,
nmodulations=12,
order=50,
fduration=0.5,
frate=100,
nfft=512,
nfilters=15,
srate=16000,
window=np.hanning):
'''Extract the Modulation Spectral Features.
Args:
wavs (list): List of (uttid, 'filename or pipe-command').
outdir (string): Output of an existing directory.
phone_map(string): Map of the phonemes from Kaldi
get_phone_labels(bool): Set True if you want to get the phoneme labels
fduration (float): Frame duration in seconds.
frate (int): Frame rate in Hertz.
hz2scale (function): Hz -> 'scale' conversion.
nfft (int): Number of points to compute the FFT.
nfilters (int): Number of filters.
postproc (function): User defined post-processing function.
srate (int): Expected sampling rate of the audio.
scale2hz (function): 'scale' -> Hz conversion.
srate (int): Expected sampling rate.
window (function): Windowing function.
Note:
It is possible to use a Kaldi like style to read the audio
using a "pipe-command" e.g.: "sph2pipe -f wav /path/file.wav |"
'''
if not only_center:
fbank = createFbank(nfilters, int(2 * fduration * srate), srate)
# Get list of phonemes
phn_list = []
with open(phone_map, 'r') as fid2:
for line2 in fid2:
line2 = line2.strip().split()
if len(line2) == 2:
if 'sil' not in line2 and 'SIL' not in line2:
phn_list.append(line2[1])
phn_list = list(set(phn_list))
phn_list.sort()
with open(wavs, 'r') as fid:
# Initialize matrix for all features
if get_phone_labels:
all_feats = np.empty(nmodulations * nfilters + 1)
else:
all_feats = np.empty(nmodulations * nfilters)
for line in fid:
tokens = line.strip().split()
uttid, inwav = tokens[0], ' '.join(tokens[1:])
fname_phn = uttid + '.PHN'
if get_phone_labels:
# Get first line of phone file in the beginning
phn_file = open(fname_phn)
phn_line = phn_file.readline()
phn_locs = phn_line.strip().split()
# Get phoneme information
phone_now = phn_locs[2]
phone_end = int(int(phn_locs[1]) / 160)
beg_frame = int(int(phn_locs[0]) / 160)
if inwav[-1] == '|':
proc = subprocess.run(inwav[:-1],
shell=True,
stdout=subprocess.PIPE)
sr, signal = read(io.BytesIO(proc.stdout))
else:
sr, signal = read(inwav)
assert sr == srate, 'Input file has different sampling rate.'
# I want to work with numbers from 0 to 1 so....
signal = signal / np.power(2, 15)
time_frames = np.array([
frame for frame in getFrames(signal, srate, frate,
fduration, window)
])
cos_trans = freqAnalysis.dct(time_frames) / np.sqrt(
2 * int(srate * fduration))
[frame_num, ndct] = np.shape(cos_trans)
if get_phone_labels:
feats = np.zeros([frame_num, nmodulations * nfilters + 1])
else:
feats = np.zeros([frame_num, nmodulations * nfilters])
print('Computing Features for file: %s' % uttid)
for i in range(beg_frame, frame_num):
each_feat = np.zeros([nfilters, nmodulations])
for j in range(nfilters):
filt = fbank[j, 0:-1]
band_dct = filt * cos_trans[i, :]
xlpc, gg = computeLpcFast(
band_dct, order) # Compute LPC coefficients
mod_spec = computeModSpecFromLpc(
gg, xlpc, nmodulations)
each_feat[j, :] = mod_spec
each_feat = np.reshape(each_feat,
(1, nfilters * nmodulations))
if get_phone_labels:
# Udates to current phoneme
if i > phone_end:
# Get new phone label
phn_line = phn_file.readline()
if phn_line:
phn_locs = phn_line.strip().split()
phone_now = phn_locs[2]
phone_end = int(int(phn_locs[1]) / 160)
else:
break # Break if no more phones are remaining
ind = phn_list.index(phone_now)
each_feat = np.append(each_feat, ind)
feats[i, :] = each_feat
all_feats = np.vstack([all_feats, feats])
all_feats = all_feats[1:, :]
# Save the final BIG feature file
np.save(os.path.join(outdir), all_feats)
np.save(os.path.join(os.path.dirname(outdir), 'phone_list'),
phn_list)
else:
fbank = createFbank(nfilters, int(2 * fduration * srate), srate)
# Get list of phonemes
phn_list = []
with open(phone_map, 'r') as fid2:
for line2 in fid2:
line2 = line2.strip().split()
if len(line2) == 2:
if 'sil' not in line2 and 'SIL' not in line2:
phn_list.append(line2[1])
phn_list = list(set(phn_list))
phn_list.sort()
with open(wavs, 'r') as fid:
# Initialize matrix for all features
if get_phone_labels:
all_feats = np.empty(nmodulations * nfilters + 1)
else:
all_feats = np.empty(nmodulations * nfilters)
for line in fid:
tokens = line.strip().split()
uttid, inwav = tokens[0], ' '.join(tokens[1:])
if inwav[-1] == '|':
proc = subprocess.run(inwav[:-1],
shell=True,
stdout=subprocess.PIPE)
sr, signal = read(io.BytesIO(proc.stdout))
else:
sr, signal = read(inwav)
assert sr == srate, 'Input file has different sampling rate.'
# I want to work with numbers from 0 to 1 so....
signal = signal / np.power(2, 15)
fname_phn = uttid + '.PHN'
# Get all phones and their center
if os.path.isfile(os.path.join(phn_file_dir, fname_phn)):
phn_file = open(os.path.join(phn_file_dir, fname_phn))
phone_mid = np.empty(0)
phone_now = np.empty(0)
for line2 in phn_file:
phn_locs = line2.strip().split()
if phn_locs[2] in phn_list:
ind = phn_list.index(phn_locs[2])
phone_now = np.append(phone_now, ind)
phone_mid = np.append(
phone_mid,
int(int(phn_locs[0]) + int(phn_locs[1])) / 2)
time_frames = np.array([
frame for frame in getFrames(signal, srate, frate,
fduration, window)
])
cos_trans = freqAnalysis.dct(time_frames) / np.sqrt(
2 * int(srate * fduration))
[frame_num, ndct] = np.shape(cos_trans)
if ignore_edge:
phone_mid = phone_mid[1:-1]
phone_now = phone_now[1:-1]
only_compute = len(phone_mid)
if get_phone_labels:
feats = np.zeros([
only_compute * around_center,
nmodulations * nfilters + 1
])
else:
feats = np.zeros([
only_compute * around_center,
nmodulations * nfilters
])
print('Computing Features for file: %s' % uttid)
for kk in range(only_compute):
i_mid = int(np.floor((phone_mid[kk])))
for cont in range(around_center):
i = i_mid + cont - int((around_center - 1) / 2)
each_feat = np.zeros([nfilters, nmodulations])
for j in range(nfilters):
filt = fbank[j, 0:-1]
band_dct = filt * cos_trans[i, :]
xlpc, gg = computeLpcFast(
band_dct,
order) # Compute LPC coefficients
mod_spec = computeModSpecFromLpc(
gg, xlpc, nmodulations)
each_feat[j, :] = mod_spec
each_feat = np.reshape(
each_feat, (1, nfilters * nmodulations))
if get_phone_labels:
feats[around_center * kk +
cont, :] = np.append(
each_feat, phone_now[kk])
else:
feats[around_center * kk + cont, :] = each_feat
#if get_phone_labels:
# feats=np.append(feats,np.reshape(phone_now,(len(phone_now),1)),axis=1)
all_feats = np.vstack([all_feats, feats])
all_feats = all_feats[1:, :]
# Save the final BIG feature file
np.save(os.path.join(outdir), all_feats)
np.save(os.path.join(os.path.dirname(outdir), 'phone_list'),
phn_list)
def extractFDLPTfPattern(wavs,
outdir,
phone_map,
phn_file_dir,
get_phone_labels=False,
only_center=False,
ignore_edge=False,
order=50,
fduration=0.5,
frate=100,
nfft=20,
nfilters=15,
srate=16000,
window=np.hanning):
if not only_center:
fbank = createFbank(nfilters, int(2 * fduration * srate), srate)
# Get list of phonemes
phn_list = []
with open(phone_map, 'r') as fid2:
for line2 in fid2:
line2 = line2.strip().split()
if len(line2) == 2:
if 'sil' not in line2 and 'SIL' not in line2:
phn_list.append(line2[1])
phn_list = list(set(phn_list))
phn_list.sort()
with open(wavs, 'r') as fid:
# Initialize matrix for all features
if get_phone_labels:
all_feats = np.empty(nfilters + 1)
else:
all_feats = np.empty(nfilters)
for line in fid:
tokens = line.strip().split()
uttid, inwav = tokens[0], ' '.join(tokens[1:])
fname_phn = uttid + '.PHN'
if get_phone_labels:
if os.path.isfile(os.path.join(phn_file_dir, fname_phn)):
# Get first line of phone file in the beginning
phn_file = open(os.path.join(phn_file_dir, fname_phn))
phn_line = phn_file.readline()
phn_locs = phn_line.strip().split()
# Get phoneme information
phone_now = phn_locs[2]
phone_end = int(int(phn_locs[1]))
beg_frame = int(int(phn_locs[0]))
else:
break
else:
beg_frame = 0
if inwav[-1] == '|':
proc = subprocess.run(inwav[:-1],
shell=True,
stdout=subprocess.PIPE)
sr, signal = read(io.BytesIO(proc.stdout))
else:
sr, signal = read(inwav)
assert sr == srate, 'Input file has different sampling rate.'
# I want to work with numbers from 0 to 1 so....
signal = signal / np.power(2, 15)
time_frames = np.array([
frame for frame in getFrames(signal, srate, frate,
fduration, window)
])
cos_trans = freqAnalysis.dct(time_frames) / np.sqrt(
2 * int(srate * fduration))
[frame_num, ndct] = np.shape(cos_trans)
if get_phone_labels:
feats = np.zeros([frame_num, nfilters + 1])
else:
feats = np.zeros([frame_num, nfilters])
print('Computing Features for file: %s' % uttid)
for i in range(beg_frame, frame_num):
each_feat = np.zeros(nfilters)
for j in range(nfilters):
filt = fbank[j, 0:-1]
band_dct = filt * cos_trans[i, :]
xlpc, gg = computeLpcFast(
band_dct, order) # Compute LPC coefficients
w, h = freqz(np.sqrt(gg), xlpc, ndct)
h_mid = np.log10(
np.mean(
np.abs(h[int(ndct / 2 - 160):int(ndct / 2 +
160)])))
each_feat[j] = h_mid
if get_phone_labels:
# Updates to current phoneme
if i > phone_end:
# Get new phone label
phn_line = phn_file.readline()
if phn_line:
phn_locs = phn_line.strip().split()
phone_now = phn_locs[2]
phone_end = int(phn_locs[1])
phone_end = int(int(phn_locs[1]) / 160)
else:
break # Break if no more phones are remaining
ind = phn_list.index(phone_now)
each_feat = np.append(each_feat, ind)
feats[i, :] = each_feat
all_feats = np.vstack([all_feats, feats])
all_feats = all_feats[1:, :]
# Save the final BIG feature file
np.save(os.path.join(outdir), all_feats)
np.save(os.path.join(os.path.dirname(outdir), 'phone_list'),
phn_list)
else:
fbank = createFbank(nfilters, int(2 * fduration * srate), srate)
# Get list of phonemes
phn_list = []
with open(phone_map, 'r') as fid2:
for line2 in fid2:
line2 = line2.strip().split()
if len(line2) == 2:
if 'sil' not in line2 and 'SIL' not in line2:
phn_list.append(line2[1])
phn_list = list(set(phn_list))
phn_list.sort()
with open(wavs, 'r') as fid:
# Initialize matrix for all features
if get_phone_labels:
all_feats = np.empty(nfilters * nfft + 1)
else:
all_feats = np.empty(nfilters * nfft)
for line in fid:
tokens = line.strip().split()
uttid, inwav = tokens[0], ' '.join(tokens[1:])
if inwav[-1] == '|':
proc = subprocess.run(inwav[:-1],
shell=True,
stdout=subprocess.PIPE)
sr, signal = read(io.BytesIO(proc.stdout))
else:
sr, signal = read(inwav)
assert sr == srate, 'Input file has different sampling rate.'
# I want to work with numbers from 0 to 1 so....
signal = signal / np.power(2, 15)
fname_phn = uttid + '.PHN'
# Get all phones and their center
if os.path.isfile(os.path.join(phn_file_dir, fname_phn)):
phn_file = open(os.path.join(phn_file_dir, fname_phn))
phone_mid = np.empty(0)
phone_now = np.empty(0)
for line2 in phn_file:
phn_locs = line2.strip().split()
if phn_locs[2] in phn_list:
ind = phn_list.index(phn_locs[2])
phone_now = np.append(phone_now, ind)
phone_mid = np.append(
phone_mid,
int(int(phn_locs[0]) + int(phn_locs[1])) / 2)
time_frames = np.array([
frame for frame in getFrames(signal, srate, frate,
fduration, window)
])
cos_trans = freqAnalysis.dct(time_frames) / np.sqrt(
2 * int(srate * fduration))
[frame_num, ndct] = np.shape(cos_trans)
if ignore_edge:
phone_mid = phone_mid[1:-1]
phone_now = phone_now[1:-1]
only_compute = len(phone_mid)
feats = np.zeros([only_compute, nfilters * nfft])
print('Computing Features for file: %s' % uttid)
for kk in range(only_compute):
i = int(np.floor((phone_mid[kk])))
each_feat = each_feat = np.zeros([nfilters, nfft])
for j in range(nfilters):
filt = fbank[j, 0:-1]
band_dct = filt * cos_trans[i, :]
xlpc, gg = computeLpcFast(
band_dct, order) # Compute LPC coefficients
w, h = freqz(np.sqrt(gg), xlpc, nfft)
each_feat[j, :] = np.log10(np.abs(h))
each_feat = np.reshape(each_feat, (1, nfilters * nfft))
feats[kk, :] = each_feat
if get_phone_labels:
feats = np.append(feats,
np.reshape(phone_now,
(len(phone_now), 1)),
axis=1)
all_feats = np.vstack([all_feats, feats])
all_feats = all_feats[1:, :]
# Save the final BIG feature file
np.save(os.path.join(outdir), all_feats)
np.save(os.path.join(os.path.dirname(outdir), 'phone_list'),
phn_list)
def extractMelEnergyFeats(args, srate=16000, window=np.hamming):
""" Extract the mel scale filter-bank energy features
"""
wavs = args.scp
outfile = args.outfile
add_reverb = args.add_reverb
nfft = args.nfft
context = args.context
fduration = args.fduration
frate = args.frate
nfilters = args.nfilters
kaldi_cmd = args.kaldi_cmd
fbank = createFbank(nfilters, nfft, srate)
if add_reverb:
if add_reverb == 'small_room':
sr_r, rir = read('./RIR/RIR_SmallRoom1_near_AnglA.wav')
rir = rir[:, 1]
rir = rir / np.power(2, 15)
elif add_reverb == 'large_room':
sr_r, rir = read('./RIR/RIR_LargeRoom1_far_AnglA.wav')
rir = rir[:, 1]
rir = rir / np.power(2, 15)
elif add_reverb == 'clean':
print('%s: No reverberation added!' % sys.argv[0])
else:
raise ValueError('Invalid type of reverberation!')
sys.stdout.flush()
with open(wavs, 'r') as fid:
all_feats = {}
for line in fid:
tokens = line.strip().split()
uttid, inwav = tokens[0], ' '.join(tokens[1:])
print('%s: Computing Features for file: %s' % (sys.argv[0], uttid))
sys.stdout.flush()
if inwav[-1] == '|':
proc = subprocess.run(inwav[:-1],
shell=True,
stdout=subprocess.PIPE)
if inwav[0:6] == 'ffmpeg':
riff_chunk_size = len(proc.stdout) - 8
q = riff_chunk_size
b = []
for i in range(4):
q, r = divmod(q, 256)
b.append(r)
riff = proc.stdout[:4] + bytes(b) + proc.stdout[8:]
sr, signal = read(io.BytesIO(riff))
else:
sr, signal = read(io.BytesIO(proc.stdout))
else:
sr, signal = read(inwav)
assert sr == srate, 'Input file has different sampling rate.'
signal = signal / np.power(2, 15)
if add_reverb:
if not add_reverb == 'clean':
signal = addReverb(signal, rir)
time_frames = np.array([
frame
for frame in getFrames(signal, srate, frate, fduration, window)
])
melEnergy_frames = np.log10(
np.matmul(np.abs(fft(time_frames, int(nfft / 2 + 1), axis=1)),
np.transpose(fbank)))
mfcc_feats = dct(melEnergy_frames, axis=1)
mfcc_feats = mfcc_feats[:, 0:13]
if args.context:
mfcc_feats = spliceFeats(mfcc_feats, context)
all_feats[uttid] = mfcc_feats
getKaldiArk(all_feats, outfile, kaldi_cmd)
def extractModSpecFeatures(args, srate=16000, window=np.hanning):
wavs = args.scp
outfile = args.outfile
phone_map = args.phn_file
phn_file_dir = args.phn_file_dir
get_phone_labels = args.get_phone_labels
add_reverb = args.add_reverb
set_unity_gain = args.set_unity_gain
nmodulations = args.nmodulations
order = args.order
fduration = args.fduration
frate = args.frate
nfilters = args.nfilters
'''Extract the Modulation Spectral Features.
Args:
wavs (list): List of (uttid, 'filename or pipe-command').
outdir (string): Output of an existing directory.
phone_map(string): Map of the phonemes from Kaldi
get_phone_labels(bool): Set True if you want to get the phoneme labels
fduration (float): Frame duration in seconds.
frate (int): Frame rate in Hertz.
hz2scale (function): Hz -> 'scale' conversion.
nfft (int): Number of points to compute the FFT.
nfilters (int): Number of filters.
postproc (function): User defined post-processing function.
srate (int): Expected sampling rate of the audio.
scale2hz (function): 'scale' -> Hz conversion.
srate (int): Expected sampling rate.
window (function): Windowing function.
Note:
It is possible to use a Kaldi like style to read the audio
using a "pipe-command" e.g.: "sph2pipe -f wav /path/file.wav |"
'''
fbank = createFbank(nfilters, int(2 * fduration * srate), srate)
if add_reverb:
if add_reverb == 'small_room':
sr_r, rir = read('./RIR/RIR_SmallRoom1_near_AnglA.wav')
rir = rir[:, 1]
rir = rir / np.power(2, 15)
elif add_reverb == 'large_room':
sr_r, rir = read('./RIR/RIR_LargeRoom1_far_AnglA.wav')
rir = rir[:, 1]
rir = rir / np.power(2, 15)
elif add_reverb == 'clean':
print('%s: No reverberation added!' % sys.argv[0])
else:
raise ValueError('Invalid type of reverberation!')
# Get list of phonemes
phn_list = []
with open(phone_map, 'r') as fid2:
for line2 in fid2:
line2 = line2.strip().split()
if len(line2) == 2:
if 'sil' not in line2 and 'SIL' not in line2:
phn_list.append(line2[1])
phn_list = list(set(phn_list))
phn_list.sort()
with open(wavs, 'r') as fid:
all_feats = {}
for line in fid:
tokens = line.strip().split()
uttid, inwav = tokens[0], ' '.join(tokens[1:])
if inwav[-1] == '|':
proc = subprocess.run(inwav[:-1],
shell=True,
stdout=subprocess.PIPE)
sr, signal = read(io.BytesIO(proc.stdout))
else:
sr, signal = read(inwav)
assert sr == srate, 'Input file has different sampling rate.'
# I want to work with numbers from 0 to 1 so....
signal = signal / np.power(2, 15)
if add_reverb:
if not add_reverb == 'clean':
signal = addReverb(signal, rir)
fname_phn = uttid + '.PHN'
# Get all phones and their center
if os.path.isfile(os.path.join(phn_file_dir, fname_phn)):
phn_file = open(os.path.join(phn_file_dir, fname_phn))
phone_mid = np.empty(0)
phone_now = np.empty(0)
for line2 in phn_file:
phn_locs = line2.strip().split()
if phn_locs[2] in phn_list:
ind = phn_list.index(phn_locs[2])
phone_now = np.append(phone_now, ind)
phone_mid = np.append(
phone_mid,
int(int(phn_locs[0]) + int(phn_locs[1])) / 2)
time_frames = np.array([
frame for frame in getFrames(signal, srate, frate,
fduration, window)
])
cos_trans = freqAnalysis.dct(time_frames) / np.sqrt(
2 * int(srate * fduration))
[frame_num, ndct] = np.shape(cos_trans)
only_compute = len(phone_mid)
if get_phone_labels:
feats = np.zeros(
[only_compute, nmodulations * nfilters + 1])
else:
feats = np.zeros([only_compute, nmodulations * nfilters])
print('Computing Features for file: %s' % uttid)
sys.stdout.flush()
for kk in range(only_compute):
i = int(np.floor((phone_mid[kk])))
each_feat = np.zeros([nfilters, nmodulations])
for j in range(nfilters):
filt = fbank[j, 0:-1]
band_dct = filt * cos_trans[i, :]
#band_dct=band_dct[band_dct>0]
xlpc, gg = computeLpcFast(
band_dct, order) # Compute LPC coefficients
if set_unity_gain:
gg = 1
mod_spec = computeModSpecFromLpc(
gg, xlpc, nmodulations)
each_feat[j, :] = mod_spec
each_feat = np.reshape(each_feat,
(1, nfilters * nmodulations))
if get_phone_labels:
feats[kk, :] = np.append(each_feat, phone_now[kk])
else:
feats[kk, :] = each_feat
all_feats[uttid] = feats
# Save the final BIG feature file
pickle.dump(all_feats, open(outfile, 'wb'))
np.save(os.path.join(os.path.dirname(outfile), 'phone_list'), phn_list)
def getFeats(args, srate=16000, window=np.hamming):
wavs = args.scp
scp_type = args.scp_type
outfile = args.outfile
coeff_num = args.coeff_num
coeff_range = args.coeff_range
order = args.order
fduration = args.fduration
frate = args.frate
nfilters = args.nfilters
kaldi_cmd = args.kaldi_cmd
add_noise = args.add_noise
add_reverb = args.add_reverb
if args.lifter_config:
fid = open(args.lifter_config, 'r')
lifter_config = fid.readline().strip().split(',')
lifter_config = np.asarray([float(x) for x in lifter_config])
# Set up mel-filterbank
fbank_type = args.fbank_type.strip().split(',')
if fbank_type[0] == "mel":
if len(fbank_type) < 2:
raise ValueError('Mel filter bank not configured properly....')
fbank = createFbank(nfilters, int(2 * fduration * srate), srate, warp_fact=float(fbank_type[1]))
elif fbank_type[0] == "cochlear":
if len(fbank_type) < 6:
raise ValueError('Cochlear filter bank not configured properly....')
if int(fbank_type[3]) == 1:
print('%s: Alpha is fixed and will not change as a function of the center frequency...' % sys.argv[0])
fbank = createFbankCochlear(nfilters, int(2 * fduration * srate), srate, om_w=float(fbank_type[1]),
alp=float(fbank_type[2]), fixed=int(fbank_type[3]), bet=float(fbank_type[4]),
warp_fact=float(fbank_type[5]))
else:
raise ValueError('Invalid type of filter bank, use mel or cochlear with proper configuration')
# Ignore odd modulations
if args.odd_mod_zero:
print('%s: Ignoring odd modulations... ' % sys.argv[0])
if add_noise:
if add_noise == "clean" or add_noise == "diff":
print('%s: No noise added!' % sys.argv[0])
else:
noise_info = add_noise.strip().split(',')
noise = load_noise(noise_info[0])
if add_reverb:
if add_reverb == 'small_room':
sr_r, rir = read('./RIR/RIR_SmallRoom1_near_AnglA.wav')
rir = rir[:, 1]
rir = rir / np.power(2, 15)
elif add_reverb == 'medium_room':
sr_r, rir = read('./RIR/RIR_MediumRoom1_far_AnglA.wav')
rir = rir[:, 1]
rir = rir / np.power(2, 15)
elif add_reverb == 'large_room':
sr_r, rir = read('./RIR/RIR_LargeRoom1_far_AnglA.wav')
rir = rir[:, 1]
rir = rir / np.power(2, 15)
elif add_reverb == 'clean':
print('%s: No reverberation added!' % sys.argv[0])
else:
raise ValueError('Invalid type of reverberation!')
# Set up mask
coeff_range = coeff_range.split(',')
lowpass = int(coeff_range[0])
highpass = int(coeff_range[1])
mask = []
for i in range(coeff_num):
if i >= lowpass and i <= highpass:
mask.append(1)
else:
mask.append(0)
mask = np.asarray(mask)
args.overlap_fraction = 1 - args.overlap_fraction
# Setup modulation weights
args.gamma_weight = args.gamma_weight.strip().split(',')
if not args.gamma_weight[0] == "None":
print('%s: Adding gamma filter on modulation frequencies...' % sys.argv[0])
x = np.linspace(0, order - 1, order)
scale = float(args.gamma_weight[0])
shape = float(args.gamma_weight[1])
pk_required = float(args.gamma_weight[2])
res = 2 * fduration
pk_required = pk_required * res
pk = (shape - 1) * scale
loc = -pk + pk_required
mod_wts = stats.gamma.pdf(x, a=shape, loc=loc, scale=scale) * 3 * scale
with open(wavs, 'r') as fid:
all_feats = {}
if args.write_utt2num_frames:
all_lens = {}
for line in fid:
tokens = line.strip().split()
uttid, inwav = tokens[0], ' '.join(tokens[1:])
if scp_type == 'wav':
if inwav[-1] == '|':
try:
proc = subprocess.run(inwav[:-1], shell=True, stdout=subprocess.PIPE)
sr, signal = read(io.BytesIO(proc.stdout))
skip_rest=False
except Exception:
skip_rest=True
else:
try:
sr, signal = read(inwav)
skip_rest = False
except Exception:
skip_rest = True
assert sr == srate, 'Input file has different sampling rate.'
elif scp_type == 'segment':
try:
cmd = 'wav-copy ' + inwav + ' - '
proc = subprocess.run(cmd, shell=True, stdout=subprocess.PIPE)
sr, signal = read(io.BytesIO(proc.stdout))
skip_rest = False
except Exception:
skip_rest = True
else:
raise ValueError('Invalid type of scp type, it should be either wav or segment')
# I want to work with numbers from 0 to 1 so....
# signal = signal / np.power(2, 15)
if not skip_rest:
if add_noise:
if not add_noise == "clean":
if add_noise == "diff":
a = [1, 2, 3, 2, 0, -2, -5, -2, 0, 2, 3, 2, 1]
signal = convolve(signal, a, mode='same')
else:
signal = add_noise_to_wav(signal, noise, float(noise_info[1]))
if add_reverb:
if not add_reverb == 'clean':
signal = addReverb(signal, rir)
tframes = signal.shape[0] # Number of samples in the signal
lfr = 1 / (args.overlap_fraction * fduration)
time_frames = np.array([frame for frame in
getFrames(signal, srate, lfr, fduration, window)])
cos_trans = freqAnalysis.dct(time_frames) / np.sqrt(2 * int(srate * fduration))
[frame_num, ndct] = np.shape(cos_trans)
feats = np.zeros((nfilters, int(np.ceil(tframes * frate / srate))))
ptr = int(0)
print('%s: Computing Features for file: %s' % (sys.argv[0], uttid))
sys.stdout.flush()
for i in range(0, frame_num):
for j in range(nfilters):
filt = fbank[j, 0:-1]
band_dct = filt * cos_trans[i, :]
xlpc, gg = computeLpcFast(band_dct, order) # Compute LPC coefficients
ms = computeModSpecFromLpc(gg, xlpc, coeff_num)
ms = ms * mask
if args.lifter_config:
ms = ms * lifter_config
if not args.gamma_weight[0] == "None":
ms = ms * mod_wts
if args.odd_mod_zero:
ms[1::2] = 0
ms = fft(ms, 2 * int(fduration * frate))
ms = np.abs(np.exp(ms))
kk = int(np.round(fduration * frate))
kkb2 = int(np.round(fduration * frate / 2))
ms = ms[0:kk] * np.hanning(kk) / window(kk)
if i == 0:
if feats.shape[1] < kkb2:
feats[j, :] += ms[kkb2:kkb2 + feats.shape[1]]
else:
feats[j, ptr:ptr + kkb2] += ms[kkb2:]
elif i == frame_num - 1 or i == frame_num - 2:
if ms.shape[0] >= feats.shape[1] - ptr:
feats[j, ptr:] += ms[:feats.shape[1] - ptr]
else:
feats[j, ptr:ptr + kk] += ms
else:
feats[j, ptr:ptr + kk] += ms
kk = int(np.round(fduration * frate * args.overlap_fraction))
kkb2 = int(np.round(fduration * frate / 2))
if i == 0:
ptr = int(ptr + kk - kkb2)
else:
ptr = int(ptr + kk + randrange(2))
all_feats[uttid] = np.log(np.clip(feats.T, a_max=None, a_min=0.00000000000001))
if args.write_utt2num_frames:
all_lens[uttid] = feats.shape[1]
dict2Ark(all_feats, outfile, kaldi_cmd)
if args.write_utt2num_frames:
with open(outfile + '.len', 'w+') as file:
for key, lens in all_lens.items():
p = "{:s} {:d}".format(key, lens)
file.write(p)
file.write("\n")
def compute_mel_spectrum(args, srate=16000, window=np.hamming):
wavs = args.scp
scp_type = args.scp_type
outfile = args.outfile
add_noise = args.add_noise
nfft = args.nfft
fduration = args.fduration
frate = args.frate
nfilters = args.nfilters
add_reverb = args.add_reverb
# Set up mel-filterbank
fbank_type = args.fbank_type.strip().split(',')
if fbank_type[0] == "mel":
if len(fbank_type) < 2:
raise ValueError('Mel filter bank not configured properly....')
fbank = createFbank(nfilters,
nfft,
srate,
warp_fact=float(fbank_type[1]))
elif fbank_type[0] == "cochlear":
if len(fbank_type) < 6:
raise ValueError(
'Cochlear filter bank not configured properly....')
if int(fbank_type[3]) == 1:
print(
'%s: Alpha is fixed and will not change as a function of the center frequency...'
% sys.argv[0])
fbank = createFbankCochlear(nfilters,
nfft,
srate,
om_w=float(fbank_type[1]),
alp=float(fbank_type[2]),
fixed=int(fbank_type[3]),
bet=float(fbank_type[4]),
warp_fact=float(fbank_type[5]))
else:
raise ValueError(
'Invalid type of filter bank, use mel or cochlear with proper configuration'
)
if add_noise:
if not add_noise == "clean" and not add_noise == "diff":
noise_info = add_noise.strip().split(',')
noise = load_noise(noise_info[0])
if add_reverb:
if add_reverb == 'small_room':
sr_r, rir = read('./RIR/RIR_SmallRoom1_near_AnglA.wav')
rir = rir[:, 1]
rir = rir / np.power(2, 15)
elif add_reverb == 'medium_room':
sr_r, rir = read('./RIR/RIR_MediumRoom1_far_AnglA.wav')
rir = rir[:, 1]
rir = rir / np.power(2, 15)
elif add_reverb == 'large_room':
sr_r, rir = read('./RIR/RIR_LargeRoom1_far_AnglA.wav')
rir = rir[:, 1]
rir = rir / np.power(2, 15)
elif add_reverb == 'clean':
print('%s: No reverberation added!' % sys.argv[0])
else:
raise ValueError('Invalid type of reverberation!')
with open(wavs, 'r') as fid:
all_feats = {}
if args.write_utt2num_frames:
all_lens = {}
for line in fid:
tokens = line.strip().split()
uttid, inwav = tokens[0], ' '.join(tokens[1:])
print('%s: Computing Features for file: %s' % (sys.argv[0], uttid))
sys.stdout.flush()
if scp_type == 'wav':
if inwav[-1] == '|':
try:
proc = subprocess.run(inwav[:-1],
shell=True,
stdout=subprocess.PIPE)
sr, signal = read(io.BytesIO(proc.stdout))
skip_rest = False
except Exception:
skip_rest = True
else:
try:
sr, signal = read(inwav)
skip_rest = False
except Exception:
skip_rest = True
assert sr == srate, 'Input file has different sampling rate.'
elif scp_type == 'segment':
try:
cmd = 'wav-copy ' + inwav + ' - '
proc = subprocess.run(cmd,
shell=True,
stdout=subprocess.PIPE)
sr, signal = read(io.BytesIO(proc.stdout))
skip_rest = False
except Exception:
skip_rest = True
else:
raise ValueError(
'Invalid type of scp type, it should be either wav or segment'
)
# signal = signal / np.power(2, 15)
if not skip_rest:
if add_noise:
if not add_noise == "clean":
if add_noise == "diff":
# signal = np.diff(signal)
a = [1, 2, 3, 2, 0, -2, -5, -2, 0, 2, 3, 2, 1]
signal = convolve(signal, a, mode='same')
else:
signal = add_noise_to_wav(signal, noise,
float(noise_info[1]))
if add_reverb:
if not add_reverb == 'clean':
signal = addReverb(signal, rir)
time_frames = np.array([
frame for frame in getFrames(signal, srate, frate,
fduration, window)
])
if args.spectrum_type == "log":
melEnergy_frames = np.log10(
np.matmul(
np.abs(
fft(time_frames, nfft,
axis=1)[:, :int(nfft / 2 + 1)]),
np.transpose(fbank)))
elif args.spectrum_type == "power":
melEnergy_frames = np.power(
np.matmul(
np.abs(
fft(time_frames, nfft,
axis=1)[:, :int(nfft / 2 + 1)]),
np.transpose(fbank)), 2)
else:
print("Spectrum type not supported! ")
sys.exit(1)
all_feats[uttid] = melEnergy_frames
if args.write_utt2num_frames:
all_lens[uttid] = melEnergy_frames.shape[0]
get_kaldi_ark(all_feats, outfile)
if args.write_utt2num_frames:
with open(outfile + '.len', 'w+') as file:
for key, lens in all_lens.items():
p = "{:s} {:d}".format(key, lens)
file.write(p)
file.write("\n")
def extractMelEnergyFeats(args, srate=16000, window=np.hamming):
""" Extract the mel scale filter-bank energy features
"""
wavs = args.scp
outfile = args.outfile
phone_map = args.phn_file
phn_file_dir = args.phn_file_dir
get_phone_labels = args.get_phone_labels
add_reverb = args.add_reverb
nfft = args.nfft
context = args.context
fduration = args.fduration
frate = args.frate
nfilters = args.nfilters
fbank = createFbank(nfilters, nfft, srate)
if add_reverb:
if add_reverb == 'small_room':
sr_r, rir = read('./RIR/RIR_SmallRoom1_near_AnglA.wav')
rir = rir[:, 1]
rir = rir / np.power(2, 15)
elif add_reverb == 'large_room':
sr_r, rir = read('./RIR/RIR_LargeRoom1_far_AnglA.wav')
rir = rir[:, 1]
rir = rir / np.power(2, 15)
elif add_reverb == 'clean':
print('%s: No reverberation added!' % sys.argv[0])
else:
raise ValueError('Invalid type of reverberation!')
sys.stdout.flush()
# Get list of phonemes
phn_list = []
with open(phone_map, 'r') as fid2:
for line2 in fid2:
line2 = line2.strip().split()
if len(line2) == 2:
if 'sil' not in line2 and 'SIL' not in line2:
phn_list.append(line2[1])
phn_list = list(set(phn_list))
phn_list.sort()
np.save('phone_list', phn_list)
with open(wavs, 'r') as fid:
all_feats = {}
for line in fid:
tokens = line.strip().split()
uttid, inwav = tokens[0], ' '.join(tokens[1:])
fname_phn = uttid + '.PHN'
uttid_base = uttid.split('_')[0]
fname_phn_base = uttid_base + '.PHN'
if isfile(join(phn_file_dir, fname_phn_base)):
fname_phn = fname_phn_base
# Get all the locations of phonemes
phone_now = np.empty(0)
phone_end = np.empty(0)
phone_beg = np.empty(0)
if isfile(join(phn_file_dir, fname_phn)):
print('%s: Computing Features for file: %s' %
(sys.argv[0], uttid))
sys.stdout.flush()
with open(join(phn_file_dir, fname_phn)) as phn_file:
for phn_line in phn_file:
phn_locs = phn_line.strip().split()
# Get phoneme information
phone_now = np.append(phone_now, phn_locs[2])
phone_end = np.append(phone_end, phn_locs[1])
phone_beg = np.append(phone_beg, phn_locs[0])
phn_file.close()
if np.size(phone_end) == 0:
print('%s: Corrupted Phone file.. hence skipped...' %
sys.argv[0])
continue
if inwav[-1] == '|':
proc = subprocess.run(inwav[:-1],
shell=True,
stdout=subprocess.PIPE)
sr, signal = read(io.BytesIO(proc.stdout))
else:
sr, signal = read(inwav)
assert sr == srate, 'Input file has different sampling rate.'
signal = signal / np.power(2, 15)
if add_reverb:
if not add_reverb == 'clean':
signal = addReverb(signal, rir)
time_frames = np.array([
frame for frame in getFrames(signal, srate, frate,
fduration, window)
])
melEnergy_frames = np.log10(
np.matmul(
np.abs(fft(time_frames, int(nfft / 2 + 1), axis=1)),
np.transpose(fbank)))
if args.context:
melEnergy_frames = spliceFeats(melEnergy_frames, context)
if get_phone_labels:
if args.context:
feats = np.empty(nfilters * (2 * context + 1) + 1)
else:
feats = np.empty(nfilters + 1)
else:
if args.context:
feats = np.empty(nfilters)
else:
feats = np.empty(nfilters * (2 * context + 1))
for num, phn in enumerate(phone_now):
now_frames = melEnergy_frames[
int(phone_beg[num]):int(phone_end[num]), :]
if get_phone_labels:
ind = phn_list.index(phn)
fr_num = now_frames.shape[0]
now_frames = np.concatenate(
(now_frames, np.tile(ind, (fr_num, 1))), axis=1)
feats = np.vstack([feats, now_frames])
feats = feats[1:, :]
all_feats[uttid] = feats
outfile = abspath(outfile)
pkl_file = outfile + '.pkl'
pickle.dump(all_feats, open(pkl_file, 'wb'))
with open(outfile + '.scp', 'w+') as file:
for item in list(all_feats.keys()):
file.write("%s %s\n" % (item, pkl_file))
def extractMelEnergyFeats(wavs,
outfile,
phone_map,
phn_file_dir,
fduration,
context,
frate,
nfft,
nfilters,
get_phone_labels=False,
add_reverb=True,
srate=16000,
window=np.hamming):
""" Extract the mel scale filter-bank energy features
"""
fbank = createFbank(nfilters, nfft, srate)
if add_reverb:
sr_r, rir = read('./RIR/RIR_LargeRoom1_far_AnglA.wav')
rir = rir[:, 1]
rir = rir / np.power(2, 15)
# Get list of phonemes
phn_list = []
with open(phone_map, 'r') as fid2:
for line2 in fid2:
line2 = line2.strip().split()
if len(line2) == 2:
if 'sil' not in line2 and 'SIL' not in line2:
phn_list.append(line2[1])
phn_list = list(set(phn_list))
phn_list.sort()
with open(wavs, 'r') as fid:
all_feats = {}
for line in fid:
tokens = line.strip().split()
uttid, inwav = tokens[0], ' '.join(tokens[1:])
fname_phn = uttid + '.PHN'
# Get all the locations of phonemes
phone_now = np.empty(0, dtype=int)
phone_mid = np.empty(0, dtype=int)
if isfile(join(phn_file_dir, fname_phn)):
print('%s: Computing Features for file: %s' %
(sys.argv[0], uttid))
sys.stdout.flush()
with open(join(phn_file_dir, fname_phn)) as phn_file:
for phn_line in phn_file:
phn_locs = phn_line.strip().split()
# Get phoneme information
if phn_locs[2] in phn_list:
ind = phn_list.index(phn_locs[2])
phone_now = np.append(phone_now, ind)
phone_mid = np.append(
phone_mid,
int((int(phn_locs[0]) + int(phn_locs[1])) / 2))
phn_file.close()
if np.size(phone_mid) == 0:
print('%s: Corrupted Phone file.. hence skipped...' %
sys.argv[0])
sys.stdout.flush()
continue
if inwav[-1] == '|':
proc = subprocess.run(inwav[:-1],
shell=True,
stdout=subprocess.PIPE)
sr, signal = read(io.BytesIO(proc.stdout))
else:
sr, signal = read(inwav)
assert sr == srate, 'Input file has different sampling rate.'
signal = signal / np.power(2, 15)
if add_reverb:
signal = addReverb(signal, rir)
time_frames = np.array([
frame for frame in getFrames(signal, srate, frate,
fduration, window)
])
melEnergy_frames = np.log10(
np.matmul(
np.abs(fft(time_frames, int(nfft / 2 + 1), axis=1)),
np.transpose(fbank)))
if args.context:
melEnergy_frames = spliceFeats(melEnergy_frames, context)
feats = melEnergy_frames[phone_mid, :]
if get_phone_labels:
feats = np.append(feats,
phone_now.reshape(len(phone_now), 1),
axis=1)
all_feats[uttid] = feats
pickle.dump(all_feats, open(outfile, 'wb'))
def extractModSpecFeatures(args, srate=16000, window=np.hamming):
""" Extract the mel scale filter-bank energy features
"""
wavs = args.scp
outfile = args.outfile
add_reverb = args.add_reverb
set_unity_gain = args.set_unity_gain
nmodulations = args.nmodulations
order = args.order
fduration = args.fduration
frate = args.frate
nfilters = args.nfilters
phone_map = args.phn_file
phn_file_dir = args.phn_file_dir
get_phone_labels = args.get_phone_labels
fbank = createFbank(nfilters, int(2 * fduration * srate), srate)
if add_reverb:
if add_reverb == 'small_room':
sr_r, rir = read('./RIR/RIR_SmallRoom1_near_AnglA.wav')
rir = rir[:, 1]
rir = rir / np.power(2, 15)
elif add_reverb == 'large_room':
sr_r, rir = read('./RIR/RIR_LargeRoom1_far_AnglA.wav')
rir = rir[:, 1]
rir = rir / np.power(2, 15)
elif add_reverb == 'clean':
print('%s: No reverberation added!' % sys.argv[0])
else:
raise ValueError('Invalid type of reverberation!')
sys.stdout.flush()
# Get list of phonemes
phn_list = []
with open(phone_map, 'r') as fid2:
for line2 in fid2:
line2 = line2.strip().split()
if len(line2) == 2:
if 'sil' not in line2 and 'SIL' not in line2:
phn_list.append(line2[1])
phn_list = list(set(phn_list))
phn_list.sort()
np.save('phone_list', phn_list)
with open(wavs, 'r') as fid:
all_feats = {}
for line in fid:
tokens = line.strip().split()
uttid, inwav = tokens[0], ' '.join(tokens[1:])
fname_phn = uttid + '.PHN'
fname_phn_base = uttid[0:-2] + '.PHN'
if isfile(join(phn_file_dir, fname_phn_base)):
fname_phn = fname_phn_base
# Get all the locations of phonemes
phone_now = np.empty(0)
phone_end = np.empty(0)
phone_beg = np.empty(0)
if isfile(join(phn_file_dir, fname_phn)):
print('%s: Computing Features for file: %s' %
(sys.argv[0], uttid))
sys.stdout.flush()
with open(join(phn_file_dir, fname_phn)) as phn_file:
for phn_line in phn_file:
phn_locs = phn_line.strip().split()
# Get phoneme information
phone_now = np.append(phone_now, phn_locs[2])
phone_end = np.append(phone_end, phn_locs[1])
phone_beg = np.append(phone_beg, phn_locs[0])
phn_file.close()
if np.size(phone_end) == 0:
print('%s: Corrupted Phone file.. hence skipped...' %
sys.argv[0])
continue
if inwav[-1] == '|':
proc = subprocess.run(inwav[:-1],
shell=True,
stdout=subprocess.PIPE)
sr, signal = read(io.BytesIO(proc.stdout))
else:
sr, signal = read(inwav)
assert sr == srate, 'Input file has different sampling rate.'
signal = signal / np.power(2, 15)
if add_reverb:
if not add_reverb == 'clean':
signal = addReverb(signal, rir)
time_frames = np.array([
frame for frame in getFrames(signal, srate, frate,
fduration, window)
])
cos_trans = dct(time_frames) / np.sqrt(
2 * int(srate * fduration))
[frame_num, ndct] = np.shape(cos_trans)
# Main feature computation loop
feats = np.zeros((frame_num, nfilters * nmodulations))
sys.stdout.flush()
for i in range(frame_num):
each_feat = np.zeros([nfilters, nmodulations])
for j in range(nfilters):
filt = fbank[j, 0:-1]
band_dct = filt * cos_trans[i, :]
xlpc, gg = computeLpcFast(
band_dct, order) # Compute LPC coefficients
if set_unity_gain:
gg = 1
mod_spec = computeModSpecFromLpc(
gg, xlpc, nmodulations)
each_feat[j, :] = mod_spec
each_feat = np.reshape(each_feat,
(1, nfilters * nmodulations))
feats[i, :] = each_feat
if not get_phone_labels:
now_feats = np.empty(nfilters * nmodulations)
else:
now_feats = np.empty(nfilters * nmodulations + 1)
for num, phn in enumerate(phone_now):
now_frames = feats[
int(phone_beg[num]):int(phone_end[num]), :]
if get_phone_labels:
ind = phn_list.index(phn)
fr_num = now_frames.shape[0]
now_frames = np.concatenate(
(now_frames, np.tile(ind, (fr_num, 1))), axis=1)
now_feats = np.vstack([now_feats, now_frames])
now_feats = now_feats[1:, :]
all_feats[uttid] = now_feats
pickle.dump(all_feats, open(outfile, 'wb'))