def extract_spec(filename,
samp_freq,
frame_length_ms=25,
frame_shift_ms=10,
round_to_power_of_two=True,
snip_edges=True):
'''
extract spectrogram using kaldi
args:
filename: wav file path
samp_freq: sample frequence
return:
spectrogram: (frame, fre)
'''
# get rspec and wspec
with open('wav.scp', 'w') as f:
f.write('test1 ' + filename + '\n')
rspec = 'scp,p:' + 'wav.scp'
wspec = 'ark,t:' + 'spec.ark'
# set po
usage = """Extract MFCC features.Usage: example.py [opts...] <rspec> <wspec>"""
po = ParseOptions(usage)
po.register_float("min-duration", 0.0, "minimum segment duration")
opts = po.parse_args()
# set options
spec_opts = SpectrogramOptions()
spec_opts.frame_opts.samp_freq = samp_freq
spec_opts.frame_opts.frame_length_ms = frame_length_ms
spec_opts.frame_opts.frame_shift_ms = frame_shift_ms
spec_opts.frame_opts.round_to_power_of_two = round_to_power_of_two
spec_opts.frame_opts.snip_edges = snip_edges
spec_opts.register(po)
spec = Spectrogram(spec_opts)
sf = spec_opts.frame_opts.samp_freq
with SequentialWaveReader(rspec) as reader, MatrixWriter(wspec) as writer:
for key, wav in reader:
if wav.duration < opts.min_duration:
continue
assert (wav.samp_freq >= sf)
assert (wav.samp_freq % sf == 0)
s = wav.data()
s = s[:, ::int(wav.samp_freq / sf)]
m = SubVector(mean(s, axis=0))
f = spec.compute_features(m, sf, 1.0)
f_array = np.array(f)
writer[key] = f
return f_array
return num_success != 0
if __name__ == "__main__":
usage = """Compute VAD.
Usage: compute-vad [options...] <wav-rspecifier> <feats-wspecifier>
"""
po = ParseOptions(usage)
po.register_float(
"min-duration",
0.0,
"Minimum duration of segments to process in seconds (default: 0.0).",
)
po.register_int(
"channel",
-1,
"Channel to extract (-1 -> mono (default), 0 -> left, 1 -> right)",
)
po.register_int("frame-window", 25,
"Length of frame window in ms (default: 25)")
po.register_int("frame-shift", 10,
"Length of frame shift in ms (default: 10)")
po.register_int("nfft", 512, "Number of DFT points (default: 256)")
po.register_int(
"arma-order",
5,
file=sys.stderr)
return num_success != 0
if __name__ == "__main__":
usage = """Compute VAD.
Usage: compute-vad [options...] <wav-rspecifier> <feats-wspecifier>
"""
po = ParseOptions(usage)
po.register_float(
"min-duration", 0.0, "Minimum duration of segments "
"to process (in seconds).")
po.register_int(
"channel", -1, "Channel to extract (-1 -> expect mono, "
"0 -> left, 1 -> right)")
po.register_int("frame-window", 25, "Length of frame window in ms "
"default is 25ms")
po.register_int("frame-shift", 10, "Length of frame shift in ms "
"default is 10ms")
po.register_int("nfft", 256, "Number of DFT points " "default is 256")
po.register_int(
"arma-order", 5, "Length of ARMA window that will be applied "
"to the spectrogram")
po.register_int("ltsv-ctx-window", 50,
"Context window for LTSV computation "
"default is 50")
logging.basicConfig(format="%(levelname)s (%(module)s[{}]:%(funcName)s():"
"%(filename)s:%(lineno)s) %(message)s"
.format(__version__), level=logging.INFO)
usage = """Decode features using GMM-based model.
Usage: gmm-decode-faster.py [options] model-in fst-in features-rspecifier
words-wspecifier [alignments-wspecifier [lattice-wspecifier]]
Note: lattices, if output, will just be linear sequences;
use gmm-latgen-faster if you want "real" lattices.
"""
po = ParseOptions(usage)
decoder_opts = FasterDecoderOptions()
decoder_opts.register(po, True)
po.register_float("acoustic-scale", 0.1,
"Scaling factor for acoustic likelihoods")
po.register_bool("allow-partial", True,
"Produce output even when final state was not reached")
po.register_str("word-symbol-table", "",
"Symbol table for words [for debug output]");
opts = po.parse_args()
if po.num_args() < 4 or po.num_args() > 6:
po.print_usage()
sys.exit()
model_rxfilename = po.get_arg(1)
fst_rxfilename = po.get_arg(2)
feature_rspecifier = po.get_arg(3)
words_wspecifier = po.get_arg(4)
alignment_wspecifier = po.get_opt_arg(5)
if __name__ == '__main__':
usage = """Create MFCC feature files.
Usage: compute-mfcc-feats [options...] <wav-rspecifier> <feats-wspecifier>
"""
po = ParseOptions(usage)
mfcc_opts = MfccOptions()
mfcc_opts.register(po)
po.register_bool(
"subtract-mean", False, "Subtract mean of each feature"
"file [CMS]; not recommended to do it this way.")
po.register_float(
"vtln-warp", 1.0, "Vtln warp factor (only applicable "
"if vtln-map not specified)")
po.register_str(
"vtln-map", "", "Map from utterance or speaker-id to "
"vtln warp factor (rspecifier)")
po.register_str(
"utt2spk", "", "Utterance to speaker-id map rspecifier"
"(if doing VTLN and you have warps per speaker)")
po.register_int(
"channel", -1, "Channel to extract (-1 -> expect mono, "
"0 -> left, 1 -> right)")
po.register_float(
"min-duration", 0.0, "Minimum duration of segments "
"to process (in seconds).")
opts = po.parse_args()
if __name__ == '__main__':
# Configure log messages to look like Kaldi messages
from kaldi import __version__
logging.addLevelName(20, 'LOG')
logging.basicConfig(
format='%(levelname)s (%(module)s[{}]:%(funcName)s():'
'%(filename)s:%(lineno)s) %(message)s'.format(__version__),
level=logging.INFO)
usage = """Extract segments from a large audio file in WAV format.
Usage:
extract-segments [options] <wav-rspecifier> <segments-file> <wav-wspecifier>
"""
po = ParseOptions(usage)
po.register_float(
"min-segment-length", 0.1, "Minimum segment length "
"in seconds (reject shorter segments)")
po.register_float(
"max_overshoot", 0.5, "End segments overshooting audio "
"by less than this (in seconds) are truncated, "
"else rejected.")
opts = po.parse_args()
if po.num_args() != 3:
po.print_usage()
sys.exit()
wav_rspecifier = po.get_arg(1)
segments_rxfilename = po.get_arg(2)
wav_wspecifier = po.get_arg(3)
Usage: copy-matrix [options] <matrix-in-rspecifier> <matrix-out-wspecifier>
or copy-matrix [options] <matrix-in-rxfilename> <matrix-out-wxfilename>
e.g.
copy-matrix --binary=false 1.mat -
copy-matrix ark:2.trans ark,t:-
"""
po = ParseOptions(usage)
po.register_bool(
"binary", True,
"Write in binary mode (only relevant if output is a wxfilename)")
po.register_float(
"scale", 1.0,
"This option can be used to scale the matrices being copied.")
po.register_bool(
"apply-log", False,
"This option can be used to apply log on the matrices. Must be avoided if matrix has negative quantities."
)
po.register_bool("apply-exp", False,
"This option can be used to apply exp on the matrices")
po.register_float(
"apply-power", 1.0,
"This option can be used to apply a power on the matrices")
po.register_bool(
"apply-softmax-per-row", False,
"This option can be used to apply softmax per row of the matrices")
opts = po.parse_args()
sw02005-A sw02005 A
sw02005-B sw02005 B
interpreted as <utterance-id> <call-id> <side> and for each <call-id>
that has two sides, does the 'only-the-louder' computation, else does
per-utterance stats in the normal way.
Note: loudness is judged by the first feature component, either energy or c0
only applicable to MFCCs or PLPs (this code could be modified to handle filterbanks).
Usage: compute-cmvn-stats-two-channel [options] <reco2file-and-channel> <feats-rspecifier> <stats-wspecifier>
e.g.: compute-cmvn-stats-two-channel data/train_unseg/reco2file_and_channel scp:data/train_unseg/feats.scp ark,t:-
"""
po = ParseOptions(usage)
po.register_float(
"quieter_channel_weight", 0.01, "For the quieter channel,"
" apply this weight to the stats, so that we still get "
"stats if one channel always dominates.")
opts = po.parse_args()
if po.num_args() != 3:
po.print_usage()
sys.exit(1)
reco2file_and_channel_rxfilename = po.get_arg(1)
feats_rspecifier = po.get_arg(2)
stats_wspecifier = po.get_arg(3)
compute_cmvn_stats_two_channel(reco2file_and_channel_rxfilename,
feats_rspecifier, stats_wspecifier, opts)
po.register_int("sampling-rate", 16000,
"Sampling rate of waveforms and labels.")
po.register_int(
"signal-window-length", 200,
"Window length in ms (what will be presented to the network).")
po.register_int("label-window-length", 25,
"Window length of alignments / labels in ms.")
po.register_int("label-window-shift", 10,
"Window shift of alignments / labels in ms.")
po.register_bool(
"subtract-mean", False, "Subtract mean of each feature"
"file [CMS]; not recommended to do it this way.")
po.register_int(
"channel", -1, "Channel to extract (-1 -> expect mono, "
"0 -> left, 1 -> right)")
po.register_float(
"min-duration", 0.0, "Minimum duration of segments "
"to process (in seconds).")
opts = po.parse_args()
if (po.num_args() != 2):
po.print_usage()
sys.exit()
wav_rspecifier = po.get_arg(1)
feats_wspecifier = po.get_arg(2)
compute_mfcc_feats(wav_rspecifier, feats_wspecifier, opts, mfcc_opts)