def test_kaldiserializer_badfile(tmpdir, mfcc_col, missing):
filename = str(tmpdir.join('foo.ark'))
mfcc_col.save(filename)
os.remove(str(tmpdir.join(missing)))
with pytest.raises(IOError) as err:
FeaturesCollection.load(filename)
assert 'file not found: {}'.format(str(tmpdir.join(missing))) in str(err)
def test_kaldiserializer_baditems(tmpdir, mfcc_col):
mfcc_col2 = FeaturesCollection(one=mfcc_col['mfcc'], two=mfcc_col['mfcc'])
mfcc_col.save(str(tmpdir.join('one.ark')))
mfcc_col2.save(str(tmpdir.join('two.ark')))
os.remove(str(tmpdir.join('two.times.ark')))
shutil.copyfile(str(tmpdir.join('one.times.ark')),
str(tmpdir.join('two.times.ark')))
with pytest.raises(ValueError) as err:
FeaturesCollection.load(str(tmpdir.join('two.ark')))
assert 'items differ in data and times' in str(err.value)
os.remove(str(tmpdir.join('one.properties.json')))
shutil.copyfile(str(tmpdir.join('two.properties.json')),
str(tmpdir.join('one.properties.json')))
with pytest.raises(ValueError) as err:
FeaturesCollection.load(str(tmpdir.join('one.ark')))
assert 'items differ in data and properties' in str(err.value)
def test_kaldiserializer(mfcc_col, tmpdir, scp):
mfcc_col.save(str(tmpdir.join('foo.ark')), scp=scp)
assert os.path.isfile(str(tmpdir.join('foo.ark')))
assert os.path.isfile(str(tmpdir.join('foo.times.ark')))
assert os.path.isfile(str(tmpdir.join('foo.properties.json')))
if scp:
assert os.path.isfile(str(tmpdir.join('foo.scp')))
assert os.path.isfile(str(tmpdir.join('foo.times.scp')))
mfcc_col2 = FeaturesCollection.load(str(tmpdir.join('foo.ark')))
assert mfcc_col2 == mfcc_col
def analyze_serializer(features, ext, output_dir):
with tempfile.TemporaryDirectory(dir=output_dir) as tmpdir:
filename = os.path.join(tmpdir, 'features' + ext)
print('writing {}...'.format(filename))
t1 = datetime.datetime.now()
features.save(filename)
t2 = datetime.datetime.now()
t_write = t2 - t1
print('took {}'.format(t_write))
f_size = os.path.getsize(filename)
print('filesize: {}'.format(sizeof_fmt(f_size)))
print('reading {}...'.format(filename))
t1 = datetime.datetime.now()
features2 = FeaturesCollection.load(filename)
t2 = datetime.datetime.now()
t_read = t2 - t1
print('took {}'.format(t_read))
print('rw equality: {}'.format(features2 == features))
return (t_write, t_read, f_size)
def test_extract_features_full(ext, wav_file, wav_file_8k, wav_file_float32,
capsys, tmpdir):
# difficult case with parallel jobs, different sampling rates,
# speakers and segments
index = [('u1', wav_file, 's1', 0, 1),
('u2', wav_file_float32, 's2', 1, 1.2),
('u3', wav_file_8k, 's1', 1, 3)]
config = pipeline.get_default_config('mfcc')
# disable VAD because it can alter the cmvn result (far from (0,
# 1) when the signal includes non-voiced frames)
config['cmvn']['with_vad'] = False
feats = pipeline.extract_features(config,
index,
njobs=2,
log=utils.get_logger())
# ensure we have the expected log messages
messages = capsys.readouterr().err
assert 'INFO - get 3 utterances from 2 speakers in 3 wavs' in messages
assert 'WARNING - several sample rates found in wav files' in messages
for utt in ('u1', 'u2', 'u3'):
assert utt in feats
assert feats[utt].dtype == np.float32
# check properies
p1 = feats['u1'].properties
p2 = feats['u2'].properties
p3 = feats['u3'].properties
assert p1['audio']['file'] == wav_file
assert p1['audio']['duration'] == 1.0
assert p2['audio']['file'] == wav_file_float32
assert p2['audio']['duration'] == pytest.approx(0.2)
assert p3['audio']['file'] == wav_file_8k
assert p3['audio']['duration'] < 0.5 # ask 3s but get duration-tstart
assert p1['mfcc'] == p2['mfcc']
assert p1['mfcc']['sample_rate'] != p3['mfcc']['sample_rate']
assert p1.keys() == {
'audio', 'mfcc', 'cmvn', 'pitch', 'delta', 'speaker', 'pipeline'
}
assert p1.keys() == p2.keys() == p3.keys()
assert p1['pipeline'] == p2['pipeline'] == p3['pipeline']
# check shape. mfcc*delta + pitch = 13 * 3 + 3 = 42
assert feats['u1'].shape == (98, 42)
assert feats['u2'].shape == (18, 42)
assert feats['u3'].shape == (40, 42)
# check cmvn
assert feats['u2'].data[:, :13].mean() == pytest.approx(0.0, abs=1e-6)
assert feats['u2'].data[:, :13].std() == pytest.approx(1.0, abs=1e-6)
data = np.vstack((feats['u1'].data[:, :13], feats['u3'].data[:, :13]))
assert data.mean() == pytest.approx(0.0, abs=1e-6)
assert data.std() == pytest.approx(1.0, abs=1e-6)
assert np.abs(data.mean()) <= np.abs(feats['u1'].data[:, :13].mean())
assert np.abs(data.std() - 1.0) <= np.abs(feats['u1'].data[:, :13].std() -
1.0)
assert np.abs(data.mean()) <= np.abs(feats['u3'].data[:, :13].mean())
assert np.abs(data.std() - 1.0) <= np.abs(feats['u3'].data[:, :13].std() -
1.0)
# save / load the features
filename = str(tmpdir.join('feats' + ext))
feats.save(filename)
feats2 = FeaturesCollection.load(filename)
assert feats2 == feats