def __init__(self, datapath, framelen, overlap, start=0, stop=None, spkrid=None):
"""
datapath: path to TIMIT raw data (using WAV format)
framelen: length of the acoustic frames
overlap: amount of acoustic samples to overlap
start: index of first frame to use
end: index of last frame to use
FIXME: start and end here are kind of hackish if used with
spkrid...
"""
data = TimitFullCorpusReader(datapath)
print start, stop
# Some list comprehension/zip magic here (but it works!)
if spkrid:
utterances = data.utteranceids(spkrid=spkrid)
else:
utterances = data.utteranceids()
if stop is not None:
utterances = utterances[start:stop]
else:
utterances = utterances[start:]
print "# of utterances: ", len(utterances)
uttfr = [data.frames(z, framelen, overlap) for z in
utterances]
fr, ph = zip(*[(x[0], x[1]) for x in uttfr])
fr = np.vstack(fr)*2**-15
ph = list(itertools.chain(*ph))
X = fr[:,0:framelen-1]
y = np.asmatrix([fr[:,framelen-1]]).T # y.ndim has to be 2
if stop is None:
stop = len(y)
super(TimitFrameData,self).__init__(X=X, y=y)
def __init__(self, datapath, framelen, overlap, start=0, stop=None):
"""
datapath: path to TIMIT raw data (using WAV format)
framelen: length of the acoustic frames
overlap: amount of acoustic samples to overlap
start: index of first TIMIT file to be used
end: index of last TIMIT file to be used
"""
data = TimitFullCorpusReader(datapath)
# Some list comprehension/zip magic here (but it works!)
if stop is None:
utterances = data.utteranceids()[start:]
else:
utterances = data.utteranceids()[start:stop]
spkrfr = [data.frames(z, framelen, overlap) for z in
utterances]
fr, ph = zip(*[(x[0], x[1]) for x in spkrfr])
framedata = np.vstack(fr)*2**-15
ph = list(itertools.chain(*ph))
# making y a one-hot output
one_hot = np.zeros((len(ph),len(data.phonelist)),dtype='float32')
idx = [data.phonelist.index(p) for p in ph]
for i in xrange(len(ph)):
one_hot[i,idx[i]] = 1.
X = np.hstack([framedata[:,0:framelen-1], one_hot])
y = np.asmatrix(framedata[:,framelen-1]).T
super(TimitFramePhoneData,self).__init__(X=X, y=y)
def __init__(self, spkrid, phone, framelen, overlap, start, stop):
data = TimitFullCorpusReader('/home/jfsantos/data/TIMIT/')
# Some list comprehension/zip magic here (but it works!)
spkrfr = [data.frames(z, 160, 159) for z in
data.utteranceids(spkrid=spkrid)]
fr, ph = zip(*[(x[0], x[1]) for x in spkrfr])
fr = np.vstack(fr)*2**-15
ph = list(itertools.chain(*ph))
# Get all elements for which the phone is 'iy'
iy_idx = [i for i,x in enumerate(ph) if x == 'iy']
fr_iy = fr[iy_idx]
X = fr_iy[:,0:159]
y = np.array([fr_iy[:,159]]).T # y.ndim has to be 2
super(TimitPhoneData,self).__init__(X=X[start:stop], y=y[start:stop])
