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():
parser = argparse.ArgumentParser(
description=__doc__,
formatter_class=argparse.RawDescriptionHelpFormatter,
)
parser.add_argument('data_dir', help='input directory with wavs')
parser.add_argument(
'output_dir',
default='/tmp',
nargs='?',
help='output directory (created files are deleted at exit)')
args = parser.parse_args()
# load audio data and compute total duration
audio_data = {
os.path.basename(f): Audio.load(f)
for f in list_files_with_extension(args.data_dir, '.wav')
}
total_duration = datetime.timedelta(
seconds=int(sum(a.duration for a in audio_data.values())))
print('found {} wav files, total duration of {}'.format(
len(audio_data), str(total_duration)))
# compute the features (default MFCC)
print('computing MFCC features...')
t1 = datetime.datetime.now()
processor = MfccProcessor()
features = FeaturesCollection(
**{k: processor.process(v)
for k, v in audio_data.items()})
t2 = datetime.datetime.now()
print('took {}'.format(t2 - t1))
# save the features in all the supported formats
data = {
'duration': total_duration,
'data': {
ext: analyze_serializer(features, ext, args.output_dir)
for ext in supported_extensions().keys()
}
}
print_results(data)
for v in (True, False)])
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
@pytest.mark.parametrize('ext', supported_extensions().keys())
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,
def test_get_serializer_byext(ext):
h = serializers.get_serializer(FeaturesCollection, 'foo' + ext, None)
assert not os.path.isfile('foo' + ext)
assert isinstance(h, serializers.supported_extensions()[ext])