def mfcc_sad(debugging):
recipe = pp.make_pipeline(steps=[
pp.speech.AudioReader(sr=16000, sr_new=8000),
pp.speech.PreEmphasis(coeff=0.97),
pp.base.Converter(converter=WAV_FILES,
input_name='path', output_name='name'),
# ====== STFT ====== #
pp.speech.STFTExtractor(frame_length=0.025, step_length=0.01,
n_fft=512, energy=False),
# ====== spectrogram ====== #
pp.speech.PowerSpecExtractor(power=2.0, output_name='spec'),
pp.speech.MelsSpecExtractor(n_mels=24, fmin=20, fmax=3700,
output_name='mspec'),
pp.speech.MFCCsExtractor(n_ceps=24, remove_first_coef=True,
first_coef_energy=True,
output_name='mfcc'),
# ====== SAD ====== #
pp.speech.SADthreshold(energy_threshold=0.5, smooth_window=5,
output_name='sad'),
pp.speech.SADgmm(nb_mixture=3, smooth_window=3,
input_name='energy', output_name='sad'),
pp.speech.ApplyingSAD(input_name=('mspec',)),
pp.speech.AcousticNorm(input_name=('mspec',), mean_var_norm=True,
windowed_mean_var_norm=True, win_length=121),
# ====== cleaning ====== #
pp.base.DeleteFeatures(input_name=('stft', 'raw', 'spec', 'sad',
'sad_threshold', 'energy')),
pp.base.AsType(dtype='float16')
], debug=debugging)
return recipe
def mfcc(augmentation=None):
delete_list = ['stft', 'spec', 'raw', 'mfcc_energy', 'sad_threshold']
if augmentation is not None:
delete_list.append('sad')
extractors = pp.make_pipeline(steps=[
SREAugmentor(augmentation) if isinstance(augmentation, string_types
) else SREAudioReader(),
pp.speech.PreEmphasis(coeff=0.97, input_name='raw'),
# ====== STFT ====== #
pp.speech.STFTExtractor(frame_length=Config.FRAME_LENGTH,
step_length=Config.STEP_LENGTH,
n_fft=Config.NFFT,
window=Config.WINDOW,
padding=False,
energy=False),
# ====== for x-vector ====== #
pp.speech.PowerSpecExtractor(
power=2.0, input_name='stft', output_name='spec'),
pp.speech.MelsSpecExtractor(n_mels=24,
fmin=20,
fmax=3700,
input_name=('spec', 'sr'),
output_name='mspec'),
pp.speech.MFCCsExtractor(n_ceps=24,
remove_first_coef=True,
first_coef_energy=True,
input_name='mspec',
output_name='mfcc'),
# ====== extract SAD ====== #
pp.speech.SADthreshold(energy_threshold=0.5,
energy_mean_scale=0.5,
frame_context=2,
proportion_threshold=0.12,
smooth_window=5,
input_name='mfcc_energy',
output_name='sad') if augmentation is None else
SADreader(ds_path=os.path.join(PATH_ACOUSTIC_FEATURES, 'mfcc')),
pp.speech.ApplyingSAD(input_name=('mspec', 'mfcc'),
sad_name='sad',
keep_unvoiced=False if CURRENT_STATE ==
SystemStates.EXTRACT_FEATURES else True),
# ====== normalization ====== #
pp.speech.AcousticNorm(mean_var_norm=True,
windowed_mean_var_norm=True,
win_length=301,
input_name=('mspec', 'mfcc')),
# ====== post processing ====== #
pp.base.DeleteFeatures(input_name=delete_list),
pp.base.AsType(dtype='float16'),
])
return extractors
def bnf(augmentation=None):
raise NotImplementedError
bnf_network = N.models.BNF_2048_MFCC40()
recipe = pp.make_pipeline(steps=[
SREAudioReader(),
pp.speech.PreEmphasis(coeff=0.97),
# ====== STFT ====== #
pp.speech.STFTExtractor(frame_length=Config.FRAME_LENGTH,
step_length=Config.STEP_LENGTH,
n_fft=Config.NFFT,
window=Config.WINDOW),
# ====== SAD ====== #
pp.base.RenameFeatures(input_name='stft_energy', output_name='energy'),
pp.speech.SADgmm(nb_mixture=3,
smooth_window=3,
input_name='energy',
output_name='sad'),
# ====== BNF ====== #
pp.speech.PowerSpecExtractor(power=2.0, output_name='spec'),
pp.speech.MelsSpecExtractor(n_mels=Config.NCEPS,
fmin=Config.FMIN,
fmax=Config.FMAX,
input_name='spec',
output_name='mspec'),
pp.speech.MFCCsExtractor(n_ceps=Config.NCEPS,
remove_first_coef=False,
input_name='mspec',
output_name='mfcc'),
pp.base.AsType(dtype='float32', input_name='mfcc'),
pp.speech.BNFExtractor(input_name='mfcc',
output_name='bnf',
sad_name='sad',
network=bnf_network,
remove_non_speech=True,
stack_context=10,
pre_mvn=True,
batch_size=5218),
# ====== normalization ====== #
pp.speech.AcousticNorm(input_name=('bnf', ),
mean_var_norm=True,
windowed_mean_var_norm=True,
win_length=301),
# ====== cleaning ====== #
pp.base.DeleteFeatures(input_name=('stft', 'raw', 'energy',
'sad_threshold', 'spec', 'mspec',
'mfcc')),
pp.base.AsType(dtype='float16')
])
return recipe
def bnf_sad(debugging):
bnf_network = N.models.BNF_2048_MFCC40()
recipe = pp.make_pipeline(steps=[
pp.speech.AudioReader(sr=16000, sr_new=8000,
best_resample=True, remove_dc=True),
pp.speech.PreEmphasis(coeff=0.97),
pp.base.Converter(converter=WAV_FILES,
input_name='path', output_name='name'),
# ====== STFT ====== #
pp.speech.STFTExtractor(frame_length=0.025, step_length=0.010,
window='hamm', n_fft=512),
# ====== SAD ====== #
pp.base.RenameFeatures(input_name='stft_energy', output_name='energy'),
pp.speech.SADgmm(nb_mixture=3, smooth_window=3,
input_name='energy', output_name='sad'),
# ====== spectrogram ====== #
pp.speech.PowerSpecExtractor(power=2.0, output_name='spec'),
pp.speech.MelsSpecExtractor(n_mels=24, fmin=100, fmax=4000,
input_name='spec', output_name='mspec'),
pp.speech.MFCCsExtractor(n_ceps=20, remove_first_coef=True,
input_name='mspec', output_name='mfcc'),
pp.base.DeltaExtractor(input_name='mfcc', order=(0, 1, 2)),
pp.speech.ApplyingSAD(input_name=('mspec', 'mfcc')),
# ====== BNF ====== #
pp.speech.MelsSpecExtractor(n_mels=40, fmin=100, fmax=4000,
input_name='spec', output_name='mspec_bnf'),
pp.speech.MFCCsExtractor(n_ceps=40, remove_first_coef=False,
input_name='mspec_bnf', output_name='mfcc_bnf'),
pp.base.AsType(dtype='float32', input_name='mfcc_bnf'),
pp.speech.BNFExtractor(input_name='mfcc_bnf', output_name='bnf',
sad_name='sad',
network=bnf_network,
remove_non_speech=True,
stack_context=10, pre_mvn=True,
batch_size=5218),
# ====== normalization ====== #
pp.speech.AcousticNorm(input_name=('mspec', 'bnf', 'mfcc'),
mean_var_norm=True,
windowed_mean_var_norm=True,
win_length=301),
# ====== cleaning ====== #
pp.base.DeleteFeatures(input_name=('stft', 'raw', 'energy',
'sad', 'sad_threshold',
'spec', 'mspec_bnf', 'mfcc_bnf')),
pp.base.AsType(dtype='float16')
], debug=debugging)
return recipe
def mfcc_sad(debugging):
recipe = pp.make_pipeline(
steps=[
pp.speech.AudioReader(sr=16000, sr_new=8000),
pp.speech.PreEmphasis(coeff=0.97),
pp.base.Converter(converter=WAV_FILES,
input_name='path',
output_name='name'),
# ====== STFT ====== #
pp.speech.STFTExtractor(frame_length=0.025,
step_length=0.01,
n_fft=512,
energy=False),
# ====== spectrogram ====== #
pp.speech.PowerSpecExtractor(power=2.0, output_name='spec'),
pp.speech.MelsSpecExtractor(n_mels=24,
fmin=20,
fmax=3700,
output_name='mspec'),
pp.speech.MFCCsExtractor(n_ceps=24,
remove_first_coef=True,
first_coef_energy=True,
output_name='mfcc'),
# ====== SAD ====== #
pp.speech.SADthreshold(energy_threshold=0.5,
smooth_window=5,
output_name='sad'),
pp.speech.SADgmm(nb_mixture=3,
smooth_window=3,
input_name='energy',
output_name='sad'),
pp.speech.ApplyingSAD(input_name=('mspec', )),
pp.speech.AcousticNorm(input_name=('mspec', ),
mean_var_norm=True,
windowed_mean_var_norm=True,
win_length=121),
# ====== cleaning ====== #
pp.base.DeleteFeatures(input_name=('stft', 'raw', 'spec', 'sad',
'sad_threshold', 'energy')),
pp.base.AsType(dtype='float16')
],
debug=debugging)
return recipe
def mfcc(augmentation=None):
delete_list = ['stft', 'spec', 'raw',
'mfcc_energy', 'sad_threshold']
if augmentation is not None:
delete_list.append('sad')
extractors = pp.make_pipeline(steps=[
SREAugmentor(augmentation)
if isinstance(augmentation, string_types) else
SREAudioReader(),
pp.speech.PreEmphasis(coeff=0.97, input_name='raw'),
# ====== STFT ====== #
pp.speech.STFTExtractor(frame_length=Config.FRAME_LENGTH,
step_length=Config.STEP_LENGTH,
n_fft=Config.NFFT, window=Config.WINDOW,
padding=False, energy=False),
# ====== for x-vector ====== #
pp.speech.PowerSpecExtractor(power=2.0,
input_name='stft', output_name='spec'),
pp.speech.MelsSpecExtractor(n_mels=24, fmin=20, fmax=3700,
input_name=('spec', 'sr'), output_name='mspec'),
pp.speech.MFCCsExtractor(n_ceps=24,
remove_first_coef=True, first_coef_energy=True,
input_name='mspec', output_name='mfcc'),
# ====== extract SAD ====== #
pp.speech.SADthreshold(energy_threshold=0.5, energy_mean_scale=0.5,
frame_context=2, proportion_threshold=0.12,
smooth_window=5,
input_name='mfcc_energy', output_name='sad')
if augmentation is None else
SADreader(ds_path=os.path.join(PATH_ACOUSTIC_FEATURES, 'mfcc')),
pp.speech.ApplyingSAD(input_name=('mspec', 'mfcc'), sad_name='sad',
keep_unvoiced=False if CURRENT_STATE == SystemStates.EXTRACT_FEATURES else True),
# ====== normalization ====== #
pp.speech.AcousticNorm(mean_var_norm=True, windowed_mean_var_norm=True,
win_length=301, input_name=('mspec', 'mfcc')),
# ====== post processing ====== #
pp.base.DeleteFeatures(input_name=delete_list),
pp.base.AsType(dtype='float16'),
])
return extractors
def bnf(augmentation=None):
raise NotImplementedError
bnf_network = N.models.BNF_2048_MFCC40()
recipe = pp.make_pipeline(steps=[
SREAudioReader(),
pp.speech.PreEmphasis(coeff=0.97),
# ====== STFT ====== #
pp.speech.STFTExtractor(frame_length=Config.FRAME_LENGTH,
step_length=Config.STEP_LENGTH,
n_fft=Config.NFFT,
window=Config.WINDOW),
# ====== SAD ====== #
pp.base.RenameFeatures(input_name='stft_energy', output_name='energy'),
pp.speech.SADgmm(nb_mixture=3, smooth_window=3,
input_name='energy', output_name='sad'),
# ====== BNF ====== #
pp.speech.PowerSpecExtractor(power=2.0, output_name='spec'),
pp.speech.MelsSpecExtractor(n_mels=Config.NCEPS,
fmin=Config.FMIN, fmax=Config.FMAX,
input_name='spec', output_name='mspec'),
pp.speech.MFCCsExtractor(n_ceps=Config.NCEPS, remove_first_coef=False,
input_name='mspec', output_name='mfcc'),
pp.base.AsType(dtype='float32', input_name='mfcc'),
pp.speech.BNFExtractor(input_name='mfcc', output_name='bnf', sad_name='sad',
network=bnf_network,
remove_non_speech=True,
stack_context=10, pre_mvn=True,
batch_size=5218),
# ====== normalization ====== #
pp.speech.AcousticNorm(input_name=('bnf',),
mean_var_norm=True,
windowed_mean_var_norm=True,
win_length=301),
# ====== cleaning ====== #
pp.base.DeleteFeatures(input_name=('stft', 'raw', 'energy',
'sad_threshold',
'spec', 'mspec', 'mfcc')),
pp.base.AsType(dtype='float16')
])
return recipe
extractors = pp.make_pipeline(
steps=[
pp.speech.AudioReader(sr_new=8000,
best_resample=True,
remove_dc=True),
pp.speech.PreEmphasis(coeff=0.97),
pp.base.Converter(
converter=lambda x: os.path.basename(x).split('.')[0],
input_name='path',
output_name='name'),
# ====== STFT ====== #
pp.speech.STFTExtractor(frame_length=0.025,
step_length=0.005,
n_fft=512,
window='hamm',
energy=False),
# ====== spectrogram ====== #
pp.speech.PowerSpecExtractor(power=2.0, output_name='spec'),
pp.speech.MelsSpecExtractor(n_mels=24,
fmin=64,
fmax=4000,
input_name=('spec', 'sr'),
output_name='mspec'),
pp.speech.MFCCsExtractor(n_ceps=20,
remove_first_coef=True,
first_coef_energy=True,
input_name='mspec',
output_name='mfcc'),
pp.base.DeltaExtractor(input_name='mfcc', order=(0, 1, 2)),
# ====== SAD ====== #
pp.base.RenameFeatures(input_name='mfcc_energy',
output_name='energy'),
pp.speech.SADthreshold(energy_threshold=0.55,
smooth_window=5,
input_name='energy',
output_name='sad'),
# ====== normalization ====== #
pp.base.DeleteFeatures(input_name=('stft', 'spec', 'sad_threshold')
),
pp.speech.AcousticNorm(mean_var_norm=True,
windowed_mean_var_norm=True,
input_name=('mspec', 'mfcc')),
# ====== post processing ====== #
pp.base.AsType(dtype='float16'),
],
debug=False)
mpi = MPI(jobs=cmds, func=mpi_fn, ncpu=cpu_count() - 1, batch=12)
for i in mpi:
prog.add(i)
# ===========================================================================
# Extract Acoustic features
# ===========================================================================
jobs = get_all_files(wav_path, filter_func=lambda x: '.wav' == x[-4:])
assert len(jobs) == TOTAL_FILES
# ====== configuration ====== #
if not os.path.exists(outpath) or args.ds:
extractors = pp.make_pipeline(steps=[
pp.speech.AudioReader(sr=None,
sr_new=8000,
best_resample=True,
remove_dc=True),
pp.base.Converter(
converter=lambda x: os.path.basename(x).split('.')[0],
input_name='path',
output_name='name'),
pp.base.AsType(dtype='float16', input_name='raw')
],
debug=False)
processor = pp.FeatureProcessor(jobs=jobs,
path=outpath,
extractor=extractors,
n_cache=0.08,
ncpu=None,
override=True)
processor.run()
pp.validate_features(processor,
path='/tmp/tidigits',
nb_samples=12,
extractors = pp.make_pipeline(steps=[
pp.speech.AudioReader(sr=FeatureConfigs.sr, dataset=audio),
pp.speech.PreEmphasis(coeff=0.97),
pp.speech.Dithering(),
# ====== STFT ====== #
pp.speech.STFTExtractor(frame_length=FeatureConfigs.frame_length,
step_length=FeatureConfigs.step_length,
n_fft=FeatureConfigs.n_fft,
window=FeatureConfigs.window),
# ====== SAD ====== #
pp.base.RenameFeatures(input_name='stft_energy', output_name='energy'),
pp.speech.SADgmm(nb_mixture=3, nb_train_it=25,
input_name='energy', output_name='sad'),
# ====== for x-vector ====== #
pp.speech.PowerSpecExtractor(power=2.0, output_name='spec'),
pp.speech.MelsSpecExtractor(n_mels=24,
fmin=FeatureConfigs.fmin, fmax=FeatureConfigs.fmax,
input_name=('spec', 'sr'), output_name='mspec'),
# ====== BNF ====== #
pp.speech.MelsSpecExtractor(n_mels=FeatureConfigs.n_mels,
fmin=FeatureConfigs.fmin, fmax=FeatureConfigs.fmax,
input_name=('spec', 'sr'), output_name='mspec_bnf'),
pp.speech.MFCCsExtractor(n_ceps=FeatureConfigs.n_ceps, remove_first_coef=False,
input_name='mspec', output_name='mfcc_bnf'),
pp.base.AsType(dtype='float32', input_name='mfcc_bnf'),
pp.speech.BNFExtractor(input_name='mfcc_bnf', output_name='bnf',
stack_context=10, pre_mvn=True,
sad_name='sad', remove_non_speech=False,
network=bnf_network,
batch_size=2048),
# ====== MFCCs with deltas ====== #
pp.speech.MFCCsExtractor(n_ceps=20, remove_first_coef=True,
input_name='mspec', output_name='mfcc'),
pp.base.DeltaExtractor(input_name='mfcc', order=(0, 1, 2)),
# ====== SDC ====== #
pp.speech.MFCCsExtractor(n_ceps=7, remove_first_coef=True,
input_name='mspec', output_name='sdc'),
pp.speech.RASTAfilter(rasta=True, input_name='sdc', output_name='sdc'),
# ====== normalization ====== #
pp.base.DeleteFeatures(input_name=('stft', 'spec',
'mspec_bnf', 'mfcc_bnf',
'sad_threshold')),
pp.speech.AcousticNorm(mean_var_norm=True, windowed_mean_var_norm=True,
sad_name=None, ignore_sad_error=True,
input_name=('mspec', 'mfcc', 'sdc', 'bnf')),
# ====== post processing ====== #
pp.base.EqualizeShape0(input_name=('mspec', 'mfcc', 'sdc', 'bnf',
'energy', 'sad')),
pp.base.AsType(dtype='float16'),
], debug=args.debug)
def bnf_all(debugging):
bnf_network = N.models.BNF_2048_MFCC40()
recipe = pp.make_pipeline(
steps=[
pp.speech.AudioReader(sr=16000,
sr_new=8000,
best_resample=True,
remove_dc=True),
pp.speech.PreEmphasis(coeff=0.97),
pp.base.Converter(converter=WAV_FILES,
input_name='path',
output_name='name'),
# ====== STFT ====== #
pp.speech.STFTExtractor(frame_length=0.025,
step_length=0.010,
window='hamm',
n_fft=512),
# ====== SAD ====== #
pp.base.RenameFeatures(input_name='stft_energy',
output_name='energy'),
pp.speech.SADgmm(nb_mixture=3,
smooth_window=3,
input_name='energy',
output_name='sad'),
# ====== spectrogram ====== #
pp.speech.PowerSpecExtractor(power=2.0, output_name='spec'),
pp.speech.MelsSpecExtractor(n_mels=24,
fmin=100,
fmax=4000,
input_name='spec',
output_name='mspec'),
pp.speech.MFCCsExtractor(n_ceps=20,
remove_first_coef=True,
input_name='mspec',
output_name='mfcc'),
pp.base.DeltaExtractor(input_name='mfcc', order=(0, 1, 2)),
# ====== BNF ====== #
pp.speech.MelsSpecExtractor(n_mels=40,
fmin=100,
fmax=4000,
input_name='spec',
output_name='mspec_bnf'),
pp.speech.MFCCsExtractor(n_ceps=40,
remove_first_coef=False,
input_name='mspec_bnf',
output_name='mfcc_bnf'),
pp.base.AsType(dtype='float32', input_name='mfcc_bnf'),
pp.speech.BNFExtractor(input_name='mfcc_bnf',
output_name='bnf',
sad_name='sad',
network=bnf_network,
remove_non_speech=False,
stack_context=10,
pre_mvn=True,
batch_size=1234),
# ====== normalization ====== #
pp.speech.AcousticNorm(input_name=('mspec', 'bnf', 'mfcc'),
mean_var_norm=True,
windowed_mean_var_norm=True,
win_length=301),
# ====== cleaning ====== #
pp.base.DeleteFeatures(input_name=('stft', 'raw', 'energy', 'sad',
'sad_threshold', 'spec',
'mspec_bnf', 'mfcc_bnf')),
pp.base.AsType(dtype='float16')
],
debug=debugging)
return recipe
print(fmt % tuple(line))
# ===========================================================================
# More detail pipeline
# ===========================================================================
pp1 = make_pipeline(steps=[
speech.AudioReader(),
speech.STFTExtractor(frame_length=0.025, padding=False),
# spectra analysis
speech.PowerSpecExtractor(output_name='spec', power=1.0),
speech.PowerSpecExtractor(output_name='pspec', power=2.0),
speech.Power2Db(input_name='pspec', output_name='db'),
# Cepstra analysis
speech.MelsSpecExtractor(n_mels=40, input_name=('pspec', 'sr')),
speech.MFCCsExtractor(n_ceps=13, input_name='mspec'),
# others
speech.PitchExtractor(frame_length=0.025, f0=True),
speech.SADgmm(input_name='stft_energy'),
speech.RASTAfilter(input_name='mfcc', output_name='rasta'),
base.EqualizeShape0(input_name=None),
speech.AcousticNorm(input_name=('mfcc', 'mspec', 'spec'),
output_name=('mfcc_norm', 'mspec_norm', 'spec_norm')),
speech.ApplyingSAD(input_name='mfcc', output_name='mfcc_sad'),
base.StackFeatures(n_context=4, input_name='mfcc')
])
formatted_printer(feats=pp1.transform(AUDIO_PATH))
print("///////////////////////////")
# ===========================================================================
# Fast pipeline
# ===========================================================================
pp2 = make_pipeline(steps=[
extractors = pp.make_pipeline(
steps=[
pp.speech.AudioReader(sr=FeatureConfigs.sr, dataset=audio),
pp.speech.PreEmphasis(coeff=0.97),
pp.speech.Dithering(),
# ====== STFT ====== #
pp.speech.STFTExtractor(frame_length=FeatureConfigs.frame_length,
step_length=FeatureConfigs.step_length,
n_fft=FeatureConfigs.n_fft,
window=FeatureConfigs.window),
# ====== SAD ====== #
pp.base.RenameFeatures(input_name='stft_energy', output_name='energy'),
pp.speech.SADgmm(nb_mixture=3,
nb_train_it=25,
input_name='energy',
output_name='sad'),
# ====== for x-vector ====== #
pp.speech.PowerSpecExtractor(power=2.0, output_name='spec'),
pp.speech.MelsSpecExtractor(n_mels=24,
fmin=FeatureConfigs.fmin,
fmax=FeatureConfigs.fmax,
input_name=('spec', 'sr'),
output_name='mspec'),
# ====== BNF ====== #
pp.speech.MelsSpecExtractor(n_mels=FeatureConfigs.n_mels,
fmin=FeatureConfigs.fmin,
fmax=FeatureConfigs.fmax,
input_name=('spec', 'sr'),
output_name='mspec_bnf'),
pp.speech.MFCCsExtractor(n_ceps=FeatureConfigs.n_ceps,
remove_first_coef=False,
input_name='mspec',
output_name='mfcc_bnf'),
pp.base.AsType(dtype='float32', input_name='mfcc_bnf'),
pp.speech.BNFExtractor(input_name='mfcc_bnf',
output_name='bnf',
stack_context=10,
pre_mvn=True,
sad_name='sad',
remove_non_speech=False,
network=bnf_network,
batch_size=2048),
# ====== MFCCs with deltas ====== #
pp.speech.MFCCsExtractor(n_ceps=20,
remove_first_coef=True,
input_name='mspec',
output_name='mfcc'),
pp.base.DeltaExtractor(input_name='mfcc', order=(0, 1, 2)),
# ====== SDC ====== #
pp.speech.MFCCsExtractor(n_ceps=7,
remove_first_coef=True,
input_name='mspec',
output_name='sdc'),
pp.speech.RASTAfilter(rasta=True, input_name='sdc', output_name='sdc'),
# ====== normalization ====== #
pp.base.DeleteFeatures(input_name=('stft', 'spec', 'mspec_bnf',
'mfcc_bnf', 'sad_threshold')),
pp.speech.AcousticNorm(mean_var_norm=True,
windowed_mean_var_norm=True,
sad_name=None,
ignore_sad_error=True,
input_name=('mspec', 'mfcc', 'sdc', 'bnf')),
# ====== post processing ====== #
pp.base.EqualizeShape0(input_name=('mspec', 'mfcc', 'sdc', 'bnf',
'energy', 'sad')),
pp.base.AsType(dtype='float16'),
],
debug=args.debug)
mpi = MPI(jobs=cmds, func=mpi_fn,
ncpu=cpu_count() - 1, batch=12)
for i in mpi:
prog.add(i)
# ===========================================================================
# Extract Acoustic features
# ===========================================================================
jobs = get_all_files(wav_path,
filter_func=lambda x: '.wav' == x[-4:])
assert len(jobs) == TOTAL_FILES
# ====== configuration ====== #
if not os.path.exists(outpath) or args.ds:
extractors = pp.make_pipeline(steps=[
pp.speech.AudioReader(sr=None, sr_new=8000, best_resample=True,
remove_dc=True),
pp.base.Converter(converter=lambda x: os.path.basename(x).split('.')[0],
input_name='path', output_name='name'),
pp.base.AsType(dtype='float16', input_name='raw')
], debug=False)
processor = pp.FeatureProcessor(jobs=jobs, path=outpath, extractor=extractors,
n_cache=0.08, ncpu=None, override=True)
processor.run()
pp.validate_features(processor, path='/tmp/tidigits', nb_samples=12,
override=True)
with open(os.path.join(outpath, 'README'), 'w') as f:
f.write(README)
# ====== check the preprocessed dataset ====== #
ds = F.Dataset(outpath, read_only=True)
print(ds)
print(ctext(ds.md5, 'yellow'))
ds.close()
len(os.listdir(PATH_ACOUSTIC_FEATURES)) != 14 or \
bool(args.acous):
extractors = pp.make_pipeline(steps=[
pp.speech.AudioReader(sr_new=8000, best_resample=True, remove_dc=True),
pp.speech.PreEmphasis(coeff=0.97),
pp.base.Converter(converter=lambda x: os.path.basename(x).split('.')[0],
input_name='path', output_name='name'),
# ====== STFT ====== #
pp.speech.STFTExtractor(frame_length=0.025, step_length=0.005,
n_fft=512, window='hamm', energy=False),
# ====== spectrogram ====== #
pp.speech.PowerSpecExtractor(power=2.0, output_name='spec'),
pp.speech.MelsSpecExtractor(n_mels=24, fmin=64, fmax=4000,
input_name=('spec', 'sr'), output_name='mspec'),
pp.speech.MFCCsExtractor(n_ceps=20,
remove_first_coef=True, first_coef_energy=True,
input_name='mspec', output_name='mfcc'),
pp.base.DeltaExtractor(input_name='mfcc', order=(0, 1, 2)),
# ====== SAD ====== #
pp.base.RenameFeatures(input_name='mfcc_energy', output_name='energy'),
pp.speech.SADthreshold(energy_threshold=0.55, smooth_window=5,
input_name='energy', output_name='sad'),
# ====== normalization ====== #
pp.base.DeleteFeatures(input_name=('stft', 'spec', 'sad_threshold')),
pp.speech.AcousticNorm(mean_var_norm=True, windowed_mean_var_norm=True,
input_name=('mspec', 'mfcc')),
# ====== post processing ====== #
pp.base.AsType(dtype='float16'),
], debug=False)
with np.warnings.catch_warnings():
np.warnings.filterwarnings('ignore')
