def extract_features_for_subset(language, subset, feat_type, output_fn):
"""
Extract features for the subset in this language.
The `feat_type` parameter can be "mfcc" or "fbank".
"""
# Get speakers for subset
speakers_fn = path.join("..", "data", subset + "_spk.list")
print("Reading:", speakers_fn)
speakers = read_speakers(speakers_fn, language)
# Convert shorten audio to wav
wav_dir = path.join("wav", language, subset)
if not path.isdir(wav_dir):
os.makedirs(wav_dir)
print("Converting shn audio to wav:")
shorten_to_wav(language, speakers, wav_dir)
# Extract raw features
print("Extracting features:")
if feat_type == "mfcc":
feat_dict = features.extract_mfcc_dir(wav_dir)
elif feat_type == "fbank":
feat_dict = features.extract_fbank_dir(wav_dir)
else:
assert False, "invalid feature type"
# Perform per speaker mean and variance normalisation
print("Per speaker mean and variance normalisation:")
feat_dict = features.speaker_mvn(feat_dict)
# Write output
print("Writing:", output_fn)
np.savez_compressed(output_fn, **feat_dict)
def extract_features(feat_type, output_fn):
"""
Extract specified features.
The `feat_type` parameter can be "mfcc" or "fbank".
"""
# Raw features
feat_dict = {}
if feat_type == "mfcc":
feat_dict_wavkey = features.extract_mfcc_dir(xitsonga_datadir)
elif feat_type == "fbank":
feat_dict_wavkey = features.extract_fbank_dir(xitsonga_datadir)
else:
assert False, "invalid feature type"
for wav_key in feat_dict_wavkey:
feat_key = utils.uttlabel_to_uttkey(wav_key)
feat_dict[feat_key] = feat_dict_wavkey[wav_key]
# Read voice activity regions
fa_fn = path.join("..", "data", "xitsonga.wrd")
print("Reading:", fa_fn)
vad_dict = utils.read_vad_from_fa(fa_fn)
# Only keep voice active regions
print("Extracting VAD regions:")
feat_dict = features.extract_vad(feat_dict, vad_dict)
# Perform per speaker mean and variance normalisation
print("Per speaker mean and variance normalisation:")
feat_dict = features.speaker_mvn(feat_dict)
# Write output
print("Writing:", output_fn)
np.savez_compressed(output_fn, **feat_dict)
def extract_features_for_subset(subset, feat_type, output_fn):
"""
Extract specified features for a subset.
The `feat_type` parameter can be "mfcc" or "fbank".
"""
# Speakers for subset
speaker_fn = path.join("..", "data",
"buckeye_" + subset + "_speakers.list")
print("Reading:", speaker_fn)
speakers = set()
with open(speaker_fn) as f:
for line in f:
speakers.add(line.strip())
print("Speakers:", ", ".join(sorted(speakers)))
# Raw features
feat_dict = {}
print("Extracting features per speaker:")
for speaker in sorted(speakers):
if feat_type == "mfcc":
speaker_feat_dict = features.extract_mfcc_dir(
path.join(buckeye_datadir, speaker))
elif feat_type == "fbank":
speaker_feat_dict = features.extract_fbank_dir(
path.join(buckeye_datadir, speaker))
else:
assert False, "invalid feature type"
for wav_key in speaker_feat_dict:
feat_dict[speaker + "_" + wav_key[3:]] = speaker_feat_dict[wav_key]
# Read voice activity regions
fa_fn = path.join("..", "data", "buckeye_english.wrd")
print("Reading:", fa_fn)
vad_dict = utils.read_vad_from_fa(fa_fn)
# Only keep voice active regions
print("Extracting VAD regions:")
feat_dict = features.extract_vad(feat_dict, vad_dict)
# Perform per speaker mean and variance normalisation
print("Per speaker mean and variance normalisation:")
feat_dict = features.speaker_mvn(feat_dict)
# Write output
print("Writing:", output_fn)
np.savez_compressed(output_fn, **feat_dict)
def main():
args = check_argv()
print(datetime.now())
# Speaker set for the indicated subset
speaker_fn = path.join("..", "data",
"buckeye_" + args.subset + "_speakers.list")
print("Reading:", speaker_fn)
speakers = set()
with open(speaker_fn) as f:
for line in f:
speakers.add(line.strip())
print("Speakers:", sorted(speakers))
# Raw filterbanks
feat_dict = {}
print("Extracting features per speaker:")
for speaker in sorted(speakers):
speaker_feat_dict = features.extract_mfcc_dir(
path.join(buckeye_datadir, speaker))
for wav_key in speaker_feat_dict:
feat_dict[speaker + "_" + wav_key[3:]] = speaker_feat_dict[wav_key]
# Read voice activity regions
fa_fn = path.join("..", "data", "buckeye_english.wrd")
print("Reading:", fa_fn)
vad_dict = read_vad_from_fa(fa_fn)
# Only keep voice active regions
print("Extracting VAD regions:")
feat_dict = features.extract_vad(feat_dict, vad_dict)
# Perform per speaker mean and variance normalisation
print("Per speaker mean and variance normalisation:")
feat_dict = features.speaker_mvn(feat_dict)
# Write output
output_dir = path.join("mfcc", "buckeye")
if not path.isdir(output_dir):
os.makedirs(output_dir)
output_fn = path.join(output_dir, args.subset + ".dd.npz")
print("Writing:", output_fn)
np.savez_compressed(output_fn, **feat_dict)
def main():
args = check_argv()
print(datetime.now())
# Raw filterbanks
data_dir = path.join(zerospeech2019_datadir, args.dataset, args.subset)
if args.subset == "train":
print("Extracting unit discovery features:")
feat_dict = features.extract_mfcc_dir(path.join(data_dir, "unit"))
print("Extracting target voice features:")
feat_dict.update(
features.extract_mfcc_dir(path.join(data_dir, "voice"))
)
else:
print("Extracting test features:")
feat_dict = features.extract_mfcc_dir(data_dir)
# Read voice activity regions
vad_fn = path.join(zerospeech2019_datadir, args.dataset, "vads.txt")
vad_dict = read_vad(vad_fn)
# Only keep voice active regions
print("Extracting VAD regions:")
feat_dict = features.extract_vad(feat_dict, vad_dict)
# Perform per speaker mean and variance normalisation
print("Per speaker mean and variance normalisation:")
feat_dict = features.speaker_mvn(feat_dict)
# Write output
output_dir = path.join("mfcc", args.dataset)
if not path.isdir(output_dir):
os.makedirs(output_dir)
output_fn = path.join(output_dir, args.subset + ".dd.npz")
print("Writing:", output_fn)
np.savez_compressed(output_fn, **feat_dict)
print(datetime.now())
