def test_extract_features(utterances_index, features):
config = pipeline.get_default_config(features,
with_cmvn=False,
with_pitch=False)
feats = pipeline.extract_features(config, utterances_index)
feat1 = feats[utterances_index[0][0]]
assert feat1.is_valid()
assert feat1.shape[0] == 140
assert feat1.dtype == np.float32
config = pipeline.get_default_config(features,
with_cmvn=False,
with_pitch=True)
feats = pipeline.extract_features(config, utterances_index)
feat2 = feats[utterances_index[0][0]]
assert feat2.is_valid()
assert feat2.shape[0] == 140
assert feat2.shape[1] == feat1.shape[1] + 3
utterances_index = [('u1', utterances_index[0][1], 0, 1)]
config = pipeline.get_default_config(features,
with_cmvn=False,
with_pitch=False)
feats = pipeline.extract_features(config, utterances_index)
feat3 = feats[utterances_index[0][0]]
assert feat3.is_valid()
assert feat3.shape[0] == 98
assert feat3.shape[1] == feat1.shape[1]
def test_cmvn(utterances_index, by_speaker, with_vad):
config = pipeline.get_default_config(
'mfcc', with_cmvn=True, with_pitch=False, with_delta=False)
config['cmvn']['by_speaker'] = by_speaker
config['cmvn']['with_vad'] = with_vad
feats = pipeline.extract_features(config, utterances_index)
feat2 = feats[utterances_index[0][0]]
assert feat2.is_valid()
assert feat2.shape[0] == 140
assert feat2.shape[1] == 13
def command_extract(args):
# setup the logger (level given by -q/-v arguments)
if args.quiet:
log = utils.null_logger()
else:
if args.verbose == 0:
level = 'warning'
elif args.verbose == 1:
level = 'info'
else: # verbose >= 2
level = 'debug'
log = utils.get_logger(name='speech-features', level=level)
# forward the initialized log to shennong
utils._logger = log
# make sure the output file is not already existing and have a
# valid extension
output_file = args.output_file
if os.path.exists(output_file):
log.error('output file already exist: %s', output_file)
return
output_ext = os.path.splitext(output_file)[1]
if output_ext not in supported_extensions().keys():
log.error(
'output file has an unsupported extension "%s", must be in %s',
output_ext, ", ".join(supported_extensions().keys()))
return
# make sure the input config and wavs_index exists
for filename in (args.config, args.utts_index):
if not os.path.exists(filename):
log.error('input file not found: %s', filename)
# read the utterances file as a list of lists, ignore empty lines
# in the file
utterances = [
utt.split(' ') for utt in
(utt.strip() for utt in open(args.utts_index, 'r'))
if utt]
# run the pipeline
features = pipeline.extract_features(
args.config, utterances, njobs=args.njobs, log=log)
# save the features
log.info('saving the features to %s', output_file)
features.save(output_file)
def main():
# parse input arguments
parser = argparse.ArgumentParser()
parser.add_argument('data_directory', help='input/output data directory')
parser.add_argument('config_file', help='YAML configuration file')
parser.add_argument(
'corpus', choices=['english', 'xitsonga'], help='corpus to process')
parser.add_argument(
'-j', '--njobs', type=int, default=4, metavar='<int>',
help='number of parallel jobs (default to %(default)s)')
parser.add_argument(
'-v', '--verbose', action='store_true', help='increase log level')
args = parser.parse_args()
# check and setup arguments
data_directory = args.data_directory
if not os.path.isdir(data_directory):
raise ValueError(f'directory not found: {data_directory}')
config = args.config_file
if not os.path.isfile(config):
raise ValueError(f'file not found: {config}')
try:
os.makedirs(os.path.join(data_directory, 'features'))
except FileExistsError:
pass
log = get_logger(level='debug' if args.verbose else 'info')
# load input utterances
utterances = [line.split(' ') for line in open(os.path.join(
data_directory, f'{args.corpus}.utts'), 'r')]
# extract the features
features = pipeline.extract_features(
config, utterances, njobs=args.njobs, log=log)
# save them
h5f_file = os.path.join(
data_directory, 'features', f'{args.corpus}_{os.path.basename(config)}'
.replace('.yaml', '.h5f'))
features.save(h5f_file)
def test_check_speakers(utterances_index, capsys):
log = utils.get_logger(level='info')
config = pipeline.get_default_config('mfcc')
with pytest.raises(ValueError) as err:
pipeline.extract_features(config, [(utterances_index[0][1], )],
log=log)
assert 'no speaker information provided' in str(err)
capsys.readouterr() # clean the buffer
config = pipeline.get_default_config('mfcc', with_cmvn=False)
pipeline.extract_features(config, utterances_index, log=log)
log_out = capsys.readouterr()
assert 'cmvn' not in log_out.err
assert '(CMVN disabled)' in log_out.err
config = pipeline.get_default_config('mfcc', with_cmvn=True)
config['cmvn']['by_speaker'] = False
pipeline.extract_features(config, utterances_index, log=log)
log_out = capsys.readouterr().err
assert 'cmvn by utterance' in log_out
assert '(CMVN by speaker disabled)' in log_out
def test_config_bad(utterances_index):
with pytest.raises(ValueError) as err:
pipeline.get_default_config('bad')
assert 'invalid features "bad"' in str(err)
config = pipeline.get_default_config('mfcc')
del config['mfcc']
with pytest.raises(ValueError) as err:
pipeline.extract_features(config, utterances_index)
assert 'the configuration does not define any features' in str(err)
config = pipeline.get_default_config('mfcc')
config['plp'] = config['mfcc']
with pytest.raises(ValueError) as err:
pipeline.extract_features(config, utterances_index)
assert 'more than one features extraction processor' in str(err)
config = pipeline.get_default_config('mfcc')
config['invalid'] = config['mfcc']
with pytest.raises(ValueError) as err:
pipeline.extract_features(config, utterances_index)
assert 'invalid keys in configuration' in str(err)
config = pipeline.get_default_config('mfcc')
del config['cmvn']['with_vad']
parsed = pipeline._init_config(config)
assert 'cmvn' in parsed
assert parsed['cmvn']['with_vad']
config = pipeline.get_default_config('mfcc')
del config['cmvn']['by_speaker']
c = pipeline._init_config(config)
assert not c['cmvn']['by_speaker']
config = pipeline.get_default_config('mfcc')
del config['pitch']['postprocessing']
c = pipeline._init_config(config)
assert c['pitch']['postprocessing'] == {}
c for c in pairs.columns if len(c) > 2 and c[-2:] == "_1"
]]
items_1.columns = [c[:-2] for c in items_1.columns]
items_2 = pairs[[
c for c in pairs.columns if len(c) > 2 and c[-2:] == "_2"
]]
items_2.columns = [c[:-2] for c in items_2.columns]
if set(items_1.columns) != set(items_2.columns):
eprint("""Issue with pair file (<F>):
columns don't match""".replace("<F>", str(args.pair_file)))
sys.exit(1)
if not set(['file', 'onset', 'offset', 'speaker']).issubset(
items_1.columns):
eprint("""Issue with pair file (<F>): missing 'file', 'speaker',
'onset', or 'offset' column(s)""".replace("<F>", str(args.pair_file)))
sys.exit(1)
items = pd.concat([items_1, items_2], sort=True).drop_duplicates()
file_spk_ = items[['file', 'speaker']].drop_duplicates()
utterance_index = [(str(i), ) + tuple(x)
for (i, x) in enumerate(file_spk_.values)]
utterance_ids = dict(((f, s), uid) for (uid, f, s) in utterance_index)
features = snpipeline.extract_features(args.shennong_config_file,
utterance_index,
njobs=args.njobs)
pairs['distance'] = calculate_distances(pairs,
features,
utterance_ids,
njobs=args.njobs)
pairs.to_csv(args.output_file, index=False)
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