def extract_features(fname, bdir, sox, htk_mfc, mfc_extension, stereo_wav,
gammatones, spectrograms, filterbanks):
#def extract_features(fname, bdir):
if fname[-4:] != '.wav':
return
rawfname = bdir+'/'+fname[:-4]+'.rawaudio'
wavfname = bdir+'/'+fname
tempfname = bdir+'/'+fname[:-4]+'_temp.wav'
# temp fname with .wav for sox
mfccfname = bdir+'/'+fname[:-4]+mfc_extension
if sox:
shutil.move(wavfname, tempfname)
call(['sox', tempfname, wavfname])
#call(['sox', '-G', tempfname, '-r 16k', wavfname])
# w/o headers, sox uses extension
shutil.move(tempfname, rawfname)
if htk_mfc:
call(['HCopy', '-C', 'wav_config', wavfname, mfccfname])
srate = 16000
#srate, sound = wavfile.read(wavfname)
sound, srate = readwav(wavfname)
if stereo_wav and len(sound.shape) == 2: # in mono sound is a list
sound = 0.5 * (sound[:, 0] + sound[:, 1])
# for stereo wav, sum both channels
if gammatones:
gammatonefname = bdir+'/'+fname[:-4]+'_gamma.npy'
tmp_snd = loadsound(wavfname)
gamma_cf = erbspace(20*Hz, 20*kHz, N_GAMMATONES_FILTERS)
gamma_fb = Gammatone(tmp_snd, gamma_cf)
with open(gammatonefname, 'w') as o_f:
npsave(o_f, gamma_fb.process())
if spectrograms:
powerspec, _, _, _ = specgram(sound, NFFT=int(srate
* SPECGRAM_WINDOW), Fs=srate, noverlap=int(srate
* SPECGRAM_OVERLAP)) # TODO
specgramfname = bdir+'/'+fname[:-4]+'_specgram.npy'
with open(specgramfname, 'w') as o_f:
npsave(o_f, powerspec.T)
if filterbanks:
# convert to Mel filterbanks
fbanks = Spectral(nfilt=N_FBANKS, # nb of filters in mel bank
alpha=0.97, # pre-emphasis
do_dct=False, # we do not want MFCCs
compression='log',
fs=srate, # sampling rate
lowerf=50, # lower frequency
frate=FBANKS_RATE, # frame rate
wlen=FBANKS_WINDOW, # window length
nfft=1024, # length of dft
do_deltas=False, # speed
do_deltasdeltas=False # acceleration
)
sound /= np.abs(sound).max(axis=0) # TODO put that as option
fbank = fbanks.transform(sound)
fbanksfname = bdir+'/'+fname[:-4]+'_fbanks.npy'
with open(fbanksfname, 'w') as o_f:
npsave(o_f, fbank)
# TODO wavelets scattergrams / scalograms
print "dealt with file", wavfname
def main(args, samplegraph):
t1 = time.time()
g = Graph()
print("Reading...")
g.read_edgelist(samplegraph)
model = Spectral(graph=g, dim=args.representation_size)
t2 = time.time()
print(t2 - t1)
print("Saving embeddings...")
model.save_embeddings(args.output)
def extract_features(df, label2ix, spec_kwargs, vad_kwargs,
stacksize=1, frate=100, return_y=False):
if return_y:
return_y = 'label' in df
X = {}
if return_y:
y = {}
spectrum_encoder = Spectral(**spec_kwargs)
vad_encoder = VAD(**vad_kwargs)
for ix, fname in enumerate(df.filename.unique()):
sig, fs = wavread(fname)
if fs != spec_kwargs['fs']:
raise ValueError('expected samplerate {}, got {}'.format(
spec_kwargs['fs'], fs)
)
spec = spectrum_encoder.transform(sig)
spec = (spec - spec.mean(0)) / spec.std(0)
if stacksize > 1:
spec = roll_array(spec, stacksize)
vad = vad_encoder.activations(sig)
vad = vad.reshape(vad.shape[0], -1)
if stacksize > 1:
vad = roll_array(vad, stacksize)
X_curr = []
if return_y:
y_curr = []
rows_iter = df[df.filename == fname].iterrows()
for _, row in rows_iter:
start = row.start
end = row.end
start_fr = int(start * frate)
end_fr = int(end * frate)
feat = np.hstack(
(spec[start_fr: end_fr],
vad[start_fr: end_fr])
)
X_curr.append(
feat.astype(np.float32)
)
if return_y:
y_curr.append(
np.ones(feat.shape[0], dtype=np.uint8) * \
label2ix[row.label]
)
X[fname] = np.vstack(X_curr)
if return_y:
y[fname] = np.hstack(y_curr)
if return_y:
return X, y
else:
return X
def extract_features(df,
label2ix,
spec_kwargs,
vad_kwargs,
stacksize=1,
frate=100,
return_y=False):
if return_y:
return_y = 'label' in df
X = {}
if return_y:
y = {}
spectrum_encoder = Spectral(**spec_kwargs)
vad_encoder = VAD(**vad_kwargs)
for ix, fname in enumerate(df.filename.unique()):
sig, fs = wavread(fname)
if fs != spec_kwargs['fs']:
raise ValueError('expected samplerate {}, got {}'.format(
spec_kwargs['fs'], fs))
spec = spectrum_encoder.transform(sig)
spec = (spec - spec.mean(0)) / spec.std(0)
if stacksize > 1:
spec = roll_array(spec, stacksize)
vad = vad_encoder.activations(sig)
vad = vad.reshape(vad.shape[0], -1)
if stacksize > 1:
vad = roll_array(vad, stacksize)
X_curr = []
if return_y:
y_curr = []
rows_iter = df[df.filename == fname].iterrows()
for _, row in rows_iter:
start = row.start
end = row.end
start_fr = int(start * frate)
end_fr = int(end * frate)
feat = np.hstack((spec[start_fr:end_fr], vad[start_fr:end_fr]))
X_curr.append(feat.astype(np.float32))
if return_y:
y_curr.append(
np.ones(feat.shape[0], dtype=np.uint8) * \
label2ix[row.label]
)
X[fname] = np.vstack(X_curr)
if return_y:
y[fname] = np.hstack(y_curr)
if return_y:
return X, y
else:
return X
def do_fbank(fname):
srate, sound = wavfile.read(fname)
fbanks = Spectral(nfilt=N_FBANKS, # nb of filters in mel bank
alpha=0.97, # pre-emphasis
do_dct=False, # we do not want MFCCs
fs=srate, # sampling rate
frate=FBANKS_RATE, # frame rate
wlen=FBANKS_WINDOW, # window length
nfft=1024, # length of dft
do_deltas=False, # speed
do_deltasdeltas=False # acceleration
)
fb = fbanks.transform(sound)
return fb
def do_fbank(fname):
srate, sound = wavfile.read(fname)
fbanks = Spectral(
nfilt=40, # nb of filters in mel bank
alpha=0.97, # pre-emphasis
do_dct=False, # we do not want MFCCs
fs=srate, # sampling rate
frate=100, # frame rate
wlen=0.025, # window length
nfft=1024, # length of dft
do_deltas=False, # speed
do_deltasdeltas=False # acceleration
)
fb = np.array(fbanks.transform(sound), dtype='float32')
return fb
def do_fbank(fname):
srate, sound = wavfile.read(fname)
fbanks = Spectral(
nfilt=N_FBANKS, # nb of filters in mel bank
alpha=0.97, # pre-emphasis
do_dct=False, # we do not want MFCCs
fs=srate, # sampling rate
frate=FBANKS_RATE, # frame rate
wlen=FBANKS_WINDOW, # window length
nfft=1024, # length of dft
do_deltas=False, # speed
do_deltasdeltas=False # acceleration
)
fb = fbanks.transform(sound)
print "did:", fname
#print fbnk.shape
return fb
def do_mfccs(fname):
"""Compute standard mfccs from a wav file"""
srate, sound = wavfile.read(fname)
fbanks = Spectral(
nfilt=40, # nb of filters in mel bank
alpha=0.97, # pre-emphasis
fs=srate, # sampling rate
frate=100, # frame rate
wlen=0.025, # window length
nfft=512, # length of dft
ncep=13, # nb of cepstral coefficients
lowerf=100,
upperf=6855.4976,
do_deltas=False, # speed
do_deltasdeltas=False # acceleration
)
fb = np.array(fbanks.transform(sound), dtype='float32')
return fb
def do_mfccs(fname):
"""Compute standard mfccs from a wav file"""
srate, sound = wavfile.read(fname)
fbanks = Spectral(
nfilt=40, # nb of filters in mel bank
alpha=0.97, # pre-emphasis
fs=srate, # sampling rate
frate=100, # frame rate
wlen=0.025, # window length
nfft=512, # length of dft
ncep=13, # nb of cepstral coefficients
lowerf=100,
upperf=6855.4976,
do_deltas=False, # speed
do_deltasdeltas=False # acceleration
)
fb = np.array(fbanks.transform(sound), dtype='float32')
return fb
def do_fbank(fname):
"""Compute standard filterbanks from a wav file"""
sound, srate = sf.read(fname)
#f = Sndfile(fname,'r')
#srate = f.samplerate
#nf = f.nframes
#sound = f.read_frames(nf)
fbanks = Spectral(
nfilt=40, # nb of filters in mel bank
alpha=0.97, # pre-emphasis
do_dct=False, # we do not want MFCCs
fs=srate, # sampling rate
frate=100, # frame rate
wlen=0.025, # window length
nfft=1024, # length of dft
do_deltas=False, # speed
do_deltasdeltas=False # acceleration
)
fb = np.array(fbanks.transform(sound), dtype='float32')
return fb
def do_fbank(fname):
fn = bdir + fname + '.wav'
try:
with open(fn[:-3] + 'npy', 'rb') as rfb:
fb = np.load(rfb)
except IOError:
srate, sound = wavfile.read(fn)
fbanks = Spectral(nfilt=N_FBANKS, # nb of filters in mel bank
alpha=0.97, # pre-emphasis
do_dct=False, # we do not want MFCCs
fs=srate, # sampling rate
frate=FBANKS_RATE, # frame rate
wlen=FBANKS_WINDOW, # window length
nfft=1024, # length of dft
do_deltas=False, # speed
do_deltasdeltas=False # acceleration
)
fb = np.array(fbanks.transform(sound), dtype='float32')
print "did:", fn
#print fb.shape
return fb
def run_all_spectral_clustering(
rad,
date_range,
boxcox=False,
norm=True,
params=["bmnum", "v", "p_l", "w_l", "slist", "elv", "time_index"],
methods=["spc", "spcb", "spcc"],
m_params={
"spc": {},
"spcb": {},
"spcc": {}
},
n_clusters=20):
"""
Invoke all spectral clustering algorithm
rad: Radar code
date_range: Date range
"""
fd = FetchData(rad, date_range)
beams, _ = fd.fetch_data(
v_params=["elv", "v", "w_l", "gflg", "p_l", "slist", "v_e"])
rec = fd.convert_to_pandas(beams)
rec["time_index"] = utils.time_days_to_index(
[x.to_pydatetime() for x in rec["time"].tolist()])
if boxcox: rec = utils.boxcox_tx(rec)
if norm: rec = utils.normalize(rec, params)
print("\n", rec.head())
for method in methods:
print("\n >> Running {c} clustering".format(c=method))
model = Spectral(method, rec[params].values, n_clusters)
model.setup(m_params[method])
model.run()
print("\n Estimating model skills.")
skill = Skills(model.data, model.obj.labels_)
return
def __init__(self,
stacksize=40,
normalize='mvn',
n_noise_fr=0,
fs=16000,
window_length=0.050,
window_shift=0.010,
nfft=1024,
scale='mel',
lowerf=120,
upperf=7000,
nfilt=40,
taper_filt=True,
compression='log',
dct=False,
nceps=13,
log_e=True,
lifter=22,
deltas=False,
remove_dc=False,
medfilt_t=0,
medfilt_s=(0, 0),
noise_fr=0,
pre_emph=0.97,
feat_cache=None,
noise_cache=None,
wav_cache=None,
n_jobs=1,
verbose=False):
self.stacksize = stacksize
self.normalize = normalize
if self.normalize == 'mvn':
self.normalizer = StandardScaler()
elif self.normalize == 'zca':
self.normalizer = ZCA()
elif self.normalize == 'minmax':
self.normalizer = MinMaxScaler()
else:
self.normalizer = IdentityTransform()
self.n_noise_fr = n_noise_fr
self.fs = fs
self.window_length = window_length
self.window_shift = window_shift
self.nfft = nfft
self.scale = scale
self.lowerf = lowerf
self.upperf = upperf
self.nfilt = nfilt
self.taper_filt = taper_filt
self.compression = compression
self.dct = dct
self.nceps = nceps
self.log_e = log_e
self.lifter = lifter
self.deltas = deltas
self.remove_dc = remove_dc
self.medfilt_t = medfilt_t
self.medfilt_s = medfilt_s
self.noise_fr = noise_fr
self.pre_emph = pre_emph
self.n_jobs = n_jobs
self.verbose = verbose
self.encoder = Spectral(fs=fs,
window_length=window_length,
window_shift=window_shift,
nfft=nfft,
scale=scale,
lowerf=lowerf,
upperf=upperf,
nfilt=nfilt,
taper_filt=taper_filt,
compression=compression,
dct=dct,
nceps=nceps,
log_e=log_e,
lifter=lifter,
deltas=deltas,
remove_dc=remove_dc,
medfilt_t=medfilt_t,
medfilt_s=medfilt_s,
noise_fr=noise_fr,
pre_emph=pre_emph)
self.D = self.encoder.n_features * self.stacksize
self.wav_cache = wav_cache if wav_cache else {}
self.noise_cache = noise_cache if noise_cache else {}
self.feat_cache = feat_cache if feat_cache else {}
class FeatureLoader(TransformerMixin, BaseEstimator):
def __init__(self,
stacksize=40,
normalize='mvn',
n_noise_fr=0,
fs=16000,
window_length=0.050,
window_shift=0.010,
nfft=1024,
scale='mel',
lowerf=120,
upperf=7000,
nfilt=40,
taper_filt=True,
compression='log',
dct=False,
nceps=13,
log_e=True,
lifter=22,
deltas=False,
remove_dc=False,
medfilt_t=0,
medfilt_s=(0, 0),
noise_fr=0,
pre_emph=0.97,
feat_cache=None,
noise_cache=None,
wav_cache=None,
n_jobs=1,
verbose=False):
self.stacksize = stacksize
self.normalize = normalize
if self.normalize == 'mvn':
self.normalizer = StandardScaler()
elif self.normalize == 'zca':
self.normalizer = ZCA()
elif self.normalize == 'minmax':
self.normalizer = MinMaxScaler()
else:
self.normalizer = IdentityTransform()
self.n_noise_fr = n_noise_fr
self.fs = fs
self.window_length = window_length
self.window_shift = window_shift
self.nfft = nfft
self.scale = scale
self.lowerf = lowerf
self.upperf = upperf
self.nfilt = nfilt
self.taper_filt = taper_filt
self.compression = compression
self.dct = dct
self.nceps = nceps
self.log_e = log_e
self.lifter = lifter
self.deltas = deltas
self.remove_dc = remove_dc
self.medfilt_t = medfilt_t
self.medfilt_s = medfilt_s
self.noise_fr = noise_fr
self.pre_emph = pre_emph
self.n_jobs = n_jobs
self.verbose = verbose
self.encoder = Spectral(fs=fs,
window_length=window_length,
window_shift=window_shift,
nfft=nfft,
scale=scale,
lowerf=lowerf,
upperf=upperf,
nfilt=nfilt,
taper_filt=taper_filt,
compression=compression,
dct=dct,
nceps=nceps,
log_e=log_e,
lifter=lifter,
deltas=deltas,
remove_dc=remove_dc,
medfilt_t=medfilt_t,
medfilt_s=medfilt_s,
noise_fr=noise_fr,
pre_emph=pre_emph)
self.D = self.encoder.n_features * self.stacksize
self.wav_cache = wav_cache if wav_cache else {}
self.noise_cache = noise_cache if noise_cache else {}
self.feat_cache = feat_cache if feat_cache else {}
def clear_cache(self):
self.wav_cache = {}
self.noise_cache = {}
self.feat_cache = {}
def get_params(self, deep=True):
p = super(FeatureLoader, self).get_params()
del p['n_jobs']
del p['verbose']
return p
def get_key(self):
"""'Frozen' dictionary representation of this object's parameters.
Used as key in caching.
"""
p = self.get_params()
del p['wav_cache']
del p['noise_cache']
del p['feat_cache']
return tuple(sorted(p.items()))
def _load_wav(self, fname):
"""
Memoized audio loader.
"""
key = fname
if not key in self.wav_cache:
sig, fs_ = wavread(fname)
if self.fs != fs_:
raise ValueError('sampling rate should be {0}, not {1}. '
'please resample.'.format(self.fs, fs_))
if len(sig.shape) > 1:
warnings.warn('stereo audio: merging channels')
sig = (sig[:, 0] + sig[:, 1]) / 2
self.wav_cache[key] = sig
return self.wav_cache[key]
def _fill_noise_cache(self, X):
for fname in X[:, 0]:
self._extract_noise(fname)
def _extract_noise(self, fname):
cfg = (('fs', self.fs), ('window_length', self.window_length),
('window_shift', self.window_shift), ('nfft', self.nfft),
('remove_dc', self.remove_dc), ('medfilt_t', self.medfilt_t),
('medfilt_s', self.medfilt_s), ('pre_emph', self.pre_emph))
key = (fname, cfg)
if not key in self.noise_cache:
if self.n_noise_fr == 0:
self.noise_cache[key] = None
else:
sig = self._load_wav(fname)
nsamples = (self.n_noise_fr + 2) * self.encoder.fshift
spec = self.encoder.get_spectrogram(sig[:nsamples])[2:, :]
noise = spec.mean(axis=0)
noise = np.clip(noise, 1e-4, np.inf)
self.noise_cache[key] = noise
return self.noise_cache[key]
def _fill_feat_cache(self, X_keys):
sigs = [self._load_wav(fname) for fname, _ in X_keys]
noises = [self._extract_noise(fname) for fname, _ in X_keys]
p = Parallel(n_jobs=self.n_jobs, verbose=0)(
delayed(extract_features_at)(sig, noise, start, self.stacksize,
self.encoder)
for (fname, start), sig, noise in izip(X_keys, sigs, noises))
r = {x_key: feat for x_key, feat in izip(X_keys, p)}
key = self.get_key()
self.feat_cache[key].update(r)
def get_specs(self, X):
key = self.get_key()
# list of [(filename, start)]
X_keys = [(X[ix, 0], X[ix, 1]) for ix in xrange(X.shape[0])]
if key in self.feat_cache:
# check for missing keys
missing_X_keys = [
x_key for x_key in X_keys if not x_key in self.feat_cache[key]
]
self._fill_feat_cache(missing_X_keys)
else:
self.feat_cache[key] = {}
self._fill_feat_cache(X_keys)
return np.vstack((self.feat_cache[key][x_key] for x_key in X_keys))
def fit(self, X, y=None):
"""Load audio and optionally estimate mean and covar
Parameters
----------
X : ndarray with columns
filename, start, end
y :
"""
r = self.get_specs(X)
self.normalizer.fit(r)
return self
def transform(self, X, y=None):
"""Load audio and perform feature extraction.
Parameters
----------
X : ndarray
"""
r = self.get_specs(X)
r = self.normalizer.transform(r)
return as_strided(r,
shape=(r.shape[0] // self.stacksize,
r.shape[1] * self.stacksize),
strides=(r.strides[0] * self.stacksize,
r.strides[1]))
def fit_transform(self, X, y=None):
r = self.get_specs(X)
r = self.normalizer.fit_transform(r)
return as_strided(r,
shape=(r.shape[0] // self.stacksize,
r.shape[1] * self.stacksize),
strides=(r.strides[0] * self.stacksize,
r.strides[1]))
def process(folder,
debug=False,
htk_mfc=False,
forcemfcext=False,
stereo_wav=False,
gammatones=False,
spectrograms=False,
filterbanks=False,
sox=True):
""" applies to all *.wav in folder """
# first find if we produce normalized MFCC, otherwise note it in the ext
# because we can then normalize on the whole corpus with another py script
mfc_extension = '.mfc_unnorm'
wcfg = open('wav_config', 'r')
for line in wcfg:
if "ENORMALISE" in line:
mfc_extension = '.mfc'
if forcemfcext:
mfc_extension = '.mfc'
print "MFC extension:", mfc_extension
if gammatones:
try:
from brian import Hz, kHz
from brian.hears import loadsound, erbspace, Gammatone
except ImportError:
print >> sys.stderr, "You need Brian Hears"
print >> sys.stderr, "http://www.briansimulator.org/docs/\
hears.html"
sys.exit(-1)
if spectrograms:
try:
from pylab import specgram
except ImportError:
print >> sys.stderr, "You need Pylab"
sys.exit(-1)
fbanks = None
if filterbanks:
try:
sys.path.append('../spectral')
from spectral import Spectral
except ImportError:
print >> sys.stderr, "You need spectral (in the parent folder)"
print >> sys.stderr, "https://github.com/mwv/spectral"
sys.exit(-1)
# run through all the folders and files in the path "folder"
# and put a header to the waves, save the originals as .rawaudio
# use HCopy to produce MFCC files according to "wav_config" file
for bdir, _, files in os.walk(folder):
for fname in files:
if fname[-4:] != '.wav':
continue
rawfname = bdir+'/'+fname[:-4]+'.rawaudio'
wavfname = bdir+'/'+fname
tempfname = bdir+'/'+fname[:-4]+'_temp.wav'
# temp fname with .wav for sox
mfccfname = bdir+'/'+fname[:-4]+mfc_extension
if sox:
shutil.move(wavfname, tempfname)
call(['sox', tempfname, wavfname])
# w/o headers, sox uses extension
shutil.move(tempfname, rawfname)
if htk_mfc:
call(['HCopy', '-C', 'wav_config', wavfname, mfccfname])
srate = 16000
srate, sound = wavfile.read(wavfname)
if stereo_wav and len(sound.shape) == 2: # in mono sound is a list
sound = sound[:, 0] + sound[:, 1]
# for stereo wav, sum both channels
if gammatones:
gammatonefname = bdir+'/'+fname[:-4]+'_gamma.npy'
tmp_snd = loadsound(wavfname)
gamma_cf = erbspace(20*Hz, 20*kHz, N_GAMMATONES_FILTERS)
gamma_fb = Gammatone(tmp_snd, gamma_cf)
with open(gammatonefname, 'w') as o_f:
npsave(o_f, gamma_fb.process())
if spectrograms:
powerspec, _, _, _ = specgram(sound, NFFT=int(srate
* SPECGRAM_WINDOW), Fs=srate, noverlap=int(srate
* SPECGRAM_OVERLAP)) # TODO
specgramfname = bdir+'/'+fname[:-4]+'_specgram.npy'
with open(specgramfname, 'w') as o_f:
npsave(o_f, powerspec.T)
if filterbanks:
# convert to Mel filterbanks
if fbanks == None: # assume parameters are fixed
fbanks = Spectral(nfilt=N_FBANKS, # nb of filters in mel bank
alpha=0.97, # pre-emphasis
do_dct=False, # we do not want MFCCs
fs=srate, # sampling rate
frate=FBANKS_RATE, # frame rate
wlen=FBANKS_WINDOW, # window length
nfft=1024, # length of dft
do_deltas=False, # speed
do_deltasdeltas=False # acceleration
)
fbank = fbanks.transform(sound)[0] # first dimension is for
# deltas & deltasdeltas
fbanksfname = bdir+'/'+fname[:-4]+'_fbanks.npy'
with open(fbanksfname, 'w') as o_f:
npsave(o_f, fbank)
# TODO wavelets scattergrams / scalograms
print "dealt with file", wavfname
_, _, fs, nframes, _, _ = fid.getparams()
sig = np.array(struct.unpack_from("%dh" % nframes,
fid.readframes(nframes)))
fid.close()
return sig, fs
FBANKS_WINDOW = 0.025 # 25ms
FBANKS_RATE = 100 # 10ms
N_FBANKS = 40
for wavfname in sys.argv[1:]:
sound, srate = readwav(wavfname)
fbanks = Spectral(
nfilt=N_FBANKS, # nb of filters in mel bank
alpha=0.97, # pre-emphasis
do_dct=False, # we do not want MFCCs
compression='log',
fs=srate, # sampling rate
#lowerf=50, # lower frequency
frate=FBANKS_RATE, # frame rate
wlen=FBANKS_WINDOW, # window length
nfft=1024, # length of dft
do_deltas=False, # speed
do_deltasdeltas=False # acceleration
)
fbank = fbanks.transform(sound)
fbanksfname = wavfname[:-4] + '_fbanks.npy'
with open(fbanksfname, 'w') as o_f:
np.save(o_f, fbank)
def process(folder,
debug=False,
htk_mfc=False,
forcemfcext=False,
stereo_wav=False,
gammatones=False,
spectrograms=False,
filterbanks=False,
sox=True):
""" applies to all *.wav in folder """
# first find if we produce normalized MFCC, otherwise note it in the ext
# because we can then normalize on the whole corpus with another py script
mfc_extension = '.mfc_unnorm'
wcfg = open('wav_config', 'r')
for line in wcfg:
if "ENORMALISE" in line:
mfc_extension = '.mfc'
if forcemfcext:
mfc_extension = '.mfc'
print "MFC extension:", mfc_extension
if gammatones:
try:
from brian import Hz, kHz
from brian.hears import loadsound, erbspace, Gammatone
except ImportError:
print >> sys.stderr, "You need Brian Hears"
print >> sys.stderr, "http://www.briansimulator.org/docs/\
hears.html"
sys.exit(-1)
if spectrograms:
try:
from pylab import specgram
except ImportError:
print >> sys.stderr, "You need Pylab"
sys.exit(-1)
fbanks = None
if filterbanks:
try:
sys.path.append('../spectral')
from spectral import Spectral
except ImportError:
print >> sys.stderr, "You need spectral (in the parent folder)"
print >> sys.stderr, "https://github.com/mwv/spectral"
sys.exit(-1)
# run through all the folders and files in the path "folder"
# and put a header to the waves, save the originals as .rawaudio
# use HCopy to produce MFCC files according to "wav_config" file
for bdir, _, files in os.walk(folder):
for fname in files:
if fname[-4:] != '.wav':
continue
rawfname = bdir + '/' + fname[:-4] + '.rawaudio'
wavfname = bdir + '/' + fname
tempfname = bdir + '/' + fname[:-4] + '_temp.wav'
# temp fname with .wav for sox
mfccfname = bdir + '/' + fname[:-4] + mfc_extension
if sox:
shutil.move(wavfname, tempfname)
call(['sox', tempfname, wavfname])
# w/o headers, sox uses extension
shutil.move(tempfname, rawfname)
if htk_mfc:
call(['HCopy', '-C', 'wav_config', wavfname, mfccfname])
srate = 16000
srate, sound = wavfile.read(wavfname)
if stereo_wav and len(sound.shape) == 2: # in mono sound is a list
sound = sound[:, 0] + sound[:, 1]
# for stereo wav, sum both channels
if gammatones:
gammatonefname = bdir + '/' + fname[:-4] + '_gamma.npy'
tmp_snd = loadsound(wavfname)
gamma_cf = erbspace(20 * Hz, 20 * kHz, N_GAMMATONES_FILTERS)
gamma_fb = Gammatone(tmp_snd, gamma_cf)
with open(gammatonefname, 'w') as o_f:
npsave(o_f, gamma_fb.process())
if spectrograms:
powerspec, _, _, _ = specgram(
sound,
NFFT=int(srate * SPECGRAM_WINDOW),
Fs=srate,
noverlap=int(srate * SPECGRAM_OVERLAP)) # TODO
specgramfname = bdir + '/' + fname[:-4] + '_specgram.npy'
with open(specgramfname, 'w') as o_f:
npsave(o_f, powerspec.T)
if filterbanks:
# convert to Mel filterbanks
if fbanks == None: # assume parameters are fixed
fbanks = Spectral(
nfilt=N_FBANKS, # nb of filters in mel bank
alpha=0.97, # pre-emphasis
do_dct=False, # we do not want MFCCs
fs=srate, # sampling rate
frate=FBANKS_RATE, # frame rate
wlen=FBANKS_WINDOW, # window length
nfft=1024, # length of dft
do_deltas=False, # speed
do_deltasdeltas=False # acceleration
)
fbank = fbanks.transform(sound)[0] # first dimension is for
# deltas & deltasdeltas
fbanksfname = bdir + '/' + fname[:-4] + '_fbanks.npy'
with open(fbanksfname, 'w') as o_f:
npsave(o_f, fbank)
# TODO wavelets scattergrams / scalograms
print "dealt with file", wavfname
def process(folder,debug=False,htk_mfcc=False,forcemfcext=False,stereo_wave=False,gammatones=False,spectograms=False,filterbanks=False,sox=True):
mfc_extension = '.mfc_unnorm'
wcfg = open('wav_config','r')
for line in wcfg:
if "ENORMALISE" in line:
mfc_extension = '.mfc'
if forcemfcext:
mfc_extension = '.mfc'
print "MFC Extension is", mfc_extension
if gammatones:
try:
from brian import Hz, kHz
from brian.hears import loadsound, erbspace, Gammatone
except ImportError:
print >> sys.stderr, "You need Brian Hears"
sys.exit(-1)
if spectograms:
try:
from pylab import specgram
except ImportError:
print >> sys.stderr,'You need Pylab'
sys.exit(-1)
fbanks = None
if filterbanks:
try:
sys.path.append('../spectral')
from spectral import Spectral
except ImportError:
print >> sys.stderr, 'you need spectral (in the parent folder)'
for bdir, _ , files in os.walk(folder):
for fname in files:
if fname[-4:] != '.WAV':
continue
rawfname= bdir + '/' + fname[:-4]+'.rawaudio'
wavfname = bdir + '/'+ fname
tempfname = bdir + '/' + fname[:-4] + '_temp.wav'
mfccfname = bdir + '/' + fname[:-4] + '.txt'
if sox:
shutil.move(wavfname, tempfname)
call(['sox',tempfname,wavfname])
shutil.move(tempfname,wavfname)
if htk_mfcc:
call(['HCopy','-C','wav_config',wavfname,mfccfname])
srate = 16000
srate, sound = wavfile.read(wavfname)
if stereo_wave and len(sound.shape == 2):
sound = sound[:,0]+ sound[:,1]
if gammatones:
gammatonefname = bdir + '/' + fname[:-4] + '_gamma.npy'
tmp_snd = loadsound(wavfname)
gamma_cf = erbspace(20*Hz, 20*kHz, n_gmammatones_filters)
gamma_fb = Gammatone(tmp_snd, gamma_cf)
with open(gammatonefname,'w') as o_f:
npsave(o_f, gamma_fb.process())
if spectograms:
powersspec, _,_,_ = specgram(sound, NFFT=int(srate * specgram_window), Fs=srate,noverlap=int(srate*specgram_window))
specgramfname = bdir + '/' + fname[:-4]+'_specgram.npy'
with open(specgramfname,'w') as o_f:
npsave(o_f , powerspec.T)
if filterbanks:
if fbanks ==None:
fbanks = Spectral(nfilt = n_fbanks, alpha=0.97,do_dct=False, fs=srate, frate=fbanks_rate, wlen=fbanks_window,nfft=1024,do_deltas=False,do_deltasdeltas=False)
fbank = fbanks.transform(sound)[0]
fbanksfname = bdir + '/' + fname[:-4]+'_fbanks.npy'
with open(fbanksfname,'w') as o_f:
npsave(o_f, fbank)
print "Dealt with the file ", wavfname
class FeatureLoader(TransformerMixin, BaseEstimator):
def __init__(self,
stacksize=40,
normalize='mvn',
n_noise_fr=0,
fs=16000,
window_length=0.050,
window_shift=0.010,
nfft=1024,
scale='mel',
lowerf=120,
upperf=7000,
nfilt=40,
taper_filt=True,
compression='log',
dct=False,
nceps=13,
log_e=True,
lifter=22,
deltas=False,
remove_dc=False,
medfilt_t=0,
medfilt_s=(0, 0),
noise_fr=0,
pre_emph=0.97,
feat_cache=None, noise_cache=None, wav_cache=None,
n_jobs=1, verbose=False):
self.stacksize = stacksize
self.normalize = normalize
if self.normalize == 'mvn':
self.normalizer = StandardScaler()
elif self.normalize == 'zca':
self.normalizer = ZCA()
elif self.normalize == 'minmax':
self.normalizer = MinMaxScaler()
else:
self.normalizer = IdentityTransform()
self.n_noise_fr = n_noise_fr
self.fs = fs
self.window_length = window_length
self.window_shift = window_shift
self.nfft = nfft
self.scale = scale
self.lowerf = lowerf
self.upperf = upperf
self.nfilt = nfilt
self.taper_filt = taper_filt
self.compression = compression
self.dct = dct
self.nceps = nceps
self.log_e = log_e
self.lifter = lifter
self.deltas = deltas
self.remove_dc = remove_dc
self.medfilt_t = medfilt_t
self.medfilt_s = medfilt_s
self.noise_fr = noise_fr
self.pre_emph = pre_emph
self.n_jobs = n_jobs
self.verbose = verbose
self.encoder = Spectral(
fs=fs,
window_length=window_length,
window_shift=window_shift,
nfft=nfft,
scale=scale,
lowerf=lowerf,
upperf=upperf,
nfilt=nfilt,
taper_filt=taper_filt,
compression=compression,
dct=dct,
nceps=nceps,
log_e=log_e,
lifter=lifter,
deltas=deltas,
remove_dc=remove_dc,
medfilt_t=medfilt_t,
medfilt_s=medfilt_s,
noise_fr=noise_fr,
pre_emph=pre_emph
)
self.D = self.encoder.n_features * self.stacksize
self.wav_cache = wav_cache if wav_cache else {}
self.noise_cache = noise_cache if noise_cache else {}
self.feat_cache = feat_cache if feat_cache else {}
def clear_cache(self):
self.wav_cache = {}
self.noise_cache = {}
self.feat_cache = {}
def get_params(self, deep=True):
p = super(FeatureLoader, self).get_params()
del p['n_jobs']
del p['verbose']
return p
def get_key(self):
"""'Frozen' dictionary representation of this object's parameters.
Used as key in caching.
"""
p = self.get_params()
del p['wav_cache']
del p['noise_cache']
del p['feat_cache']
return tuple(sorted(p.items()))
def _load_wav(self, fname):
"""
Memoized audio loader.
"""
key = fname
if not key in self.wav_cache:
sig, fs_ = wavread(fname)
if self.fs != fs_:
raise ValueError('sampling rate should be {0}, not {1}. '
'please resample.'.format(self.fs, fs_))
if len(sig.shape) > 1:
warnings.warn('stereo audio: merging channels')
sig = (sig[:, 0] + sig[:, 1]) / 2
self.wav_cache[key] = sig
return self.wav_cache[key]
def _fill_noise_cache(self, X):
for fname in X[:, 0]:
self._extract_noise(fname)
def _extract_noise(self, fname):
cfg = (
('fs', self.fs),
('window_length', self.window_length),
('window_shift', self.window_shift),
('nfft', self.nfft),
('remove_dc', self.remove_dc),
('medfilt_t', self.medfilt_t),
('medfilt_s', self.medfilt_s),
('pre_emph', self.pre_emph)
)
key = (fname, cfg)
if not key in self.noise_cache:
if self.n_noise_fr == 0:
self.noise_cache[key] = None
else:
sig = self._load_wav(fname)
nsamples = (self.n_noise_fr + 2) * self.encoder.fshift
spec = self.encoder.get_spectrogram(sig[:nsamples])[2:, :]
noise = spec.mean(axis=0)
noise = np.clip(noise, 1e-4, np.inf)
self.noise_cache[key] = noise
return self.noise_cache[key]
def _fill_feat_cache(self, X_keys):
sigs = [self._load_wav(fname) for fname, _ in X_keys]
noises = [self._extract_noise(fname) for fname, _ in X_keys]
p = Parallel(n_jobs=self.n_jobs, verbose=0)(
delayed(extract_features_at)(
sig, noise, start, self.stacksize, self.encoder)
for (fname, start), sig, noise in izip(X_keys, sigs, noises)
)
r = {x_key: feat
for x_key, feat in izip(X_keys, p)}
key = self.get_key()
self.feat_cache[key].update(r)
def get_specs(self, X):
key = self.get_key()
# list of [(filename, start)]
X_keys = [(X[ix, 0], X[ix, 1]) for ix in xrange(X.shape[0])]
if key in self.feat_cache:
# check for missing keys
missing_X_keys = [
x_key
for x_key in X_keys
if not x_key in self.feat_cache[key]
]
self._fill_feat_cache(missing_X_keys)
else:
self.feat_cache[key] = {}
self._fill_feat_cache(X_keys)
return np.vstack((self.feat_cache[key][x_key] for x_key in X_keys))
def fit(self, X, y=None):
"""Load audio and optionally estimate mean and covar
Parameters
----------
X : ndarray with columns
filename, start, end
y :
"""
r = self.get_specs(X)
self.normalizer.fit(r)
return self
def transform(self, X, y=None):
"""Load audio and perform feature extraction.
Parameters
----------
X : ndarray
"""
r = self.get_specs(X)
r = self.normalizer.transform(r)
return as_strided(
r,
shape=(r.shape[0]//self.stacksize, r.shape[1]*self.stacksize),
strides=(r.strides[0]*self.stacksize, r.strides[1])
)
def fit_transform(self, X, y=None):
r = self.get_specs(X)
r = self.normalizer.fit_transform(r)
return as_strided(
r,
shape=(r.shape[0]//self.stacksize, r.shape[1]*self.stacksize),
strides=(r.strides[0]*self.stacksize, r.strides[1])
)
def __init__(self,
stacksize=40,
normalize='mvn',
n_noise_fr=0,
fs=16000,
window_length=0.050,
window_shift=0.010,
nfft=1024,
scale='mel',
lowerf=120,
upperf=7000,
nfilt=40,
taper_filt=True,
compression='log',
dct=False,
nceps=13,
log_e=True,
lifter=22,
deltas=False,
remove_dc=False,
medfilt_t=0,
medfilt_s=(0, 0),
noise_fr=0,
pre_emph=0.97,
feat_cache=None, noise_cache=None, wav_cache=None,
n_jobs=1, verbose=False):
self.stacksize = stacksize
self.normalize = normalize
if self.normalize == 'mvn':
self.normalizer = StandardScaler()
elif self.normalize == 'zca':
self.normalizer = ZCA()
elif self.normalize == 'minmax':
self.normalizer = MinMaxScaler()
else:
self.normalizer = IdentityTransform()
self.n_noise_fr = n_noise_fr
self.fs = fs
self.window_length = window_length
self.window_shift = window_shift
self.nfft = nfft
self.scale = scale
self.lowerf = lowerf
self.upperf = upperf
self.nfilt = nfilt
self.taper_filt = taper_filt
self.compression = compression
self.dct = dct
self.nceps = nceps
self.log_e = log_e
self.lifter = lifter
self.deltas = deltas
self.remove_dc = remove_dc
self.medfilt_t = medfilt_t
self.medfilt_s = medfilt_s
self.noise_fr = noise_fr
self.pre_emph = pre_emph
self.n_jobs = n_jobs
self.verbose = verbose
self.encoder = Spectral(
fs=fs,
window_length=window_length,
window_shift=window_shift,
nfft=nfft,
scale=scale,
lowerf=lowerf,
upperf=upperf,
nfilt=nfilt,
taper_filt=taper_filt,
compression=compression,
dct=dct,
nceps=nceps,
log_e=log_e,
lifter=lifter,
deltas=deltas,
remove_dc=remove_dc,
medfilt_t=medfilt_t,
medfilt_s=medfilt_s,
noise_fr=noise_fr,
pre_emph=pre_emph
)
self.D = self.encoder.n_features * self.stacksize
self.wav_cache = wav_cache if wav_cache else {}
self.noise_cache = noise_cache if noise_cache else {}
self.feat_cache = feat_cache if feat_cache else {}
fid = wave.open(fname, 'r')
_, _, fs, nframes, _, _ = fid.getparams()
sig = np.array(struct.unpack_from("%dh" % nframes,
fid.readframes(nframes)))
fid.close()
return sig, fs
FBANKS_WINDOW = 0.025 # 25ms
FBANKS_RATE = 100 # 10ms
N_FBANKS = 40
for wavfname in sys.argv[1:]:
sound, srate = readwav(wavfname)
fbanks = Spectral(nfilt=N_FBANKS, # nb of filters in mel bank
alpha=0.97, # pre-emphasis
do_dct=False, # we do not want MFCCs
compression='log',
fs=srate, # sampling rate
#lowerf=50, # lower frequency
frate=FBANKS_RATE, # frame rate
wlen=FBANKS_WINDOW, # window length
nfft=1024, # length of dft
do_deltas=False, # speed
do_deltasdeltas=False # acceleration
)
fbank = fbanks.transform(sound)
fbanksfname = wavfname[:-4]+'_fbanks.npy'
with open(fbanksfname, 'w') as o_f:
np.save(o_f, fbank)
