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,
"Length of ARMA window that will be applied to the spectrogram",
)
po.register_int(
"ltsv-ctx-window",
50,
"%(filename)s:%(lineno)s) %(message)s".format(__version__),
level=logging.INFO)
usage = """Use Principal component analysis for dimension reduction.
For the details, Please refer to website:
https://scikit-learn.org/stable/modules/generated/sklearn.decomposition.PCA.html
Usage: pca-vector.py [options] <vector-rspecifier> <vector-wspecifier
e.g.
pca-vector.py scp:data/train/ivector.scp ark:data/train/low_dim_vector.ark
see also: two-dim-vector-visual.py
"""
po = ParseOptions(usage)
po.register_int(
"output-dim", 2, "dimension of the output vectors."
" For visualization, only 2 is allowed in this program. (2 by default)"
)
opts = po.parse_args()
if (po.num_args() != 2):
po.print_usage()
sys.exit()
vector_rspecifier = po.get_arg(1)
vector_wspecifier = po.get_arg(2)
isSuccess = pca_vector(vector_rspecifier,
vector_wspecifier,
output_dim=opts.output_dim)
if not isSuccess:
sys.exit()
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 (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)
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")
po.register_float(
"threshold", 0.01, "Parameter for sigmoid scaling in LTSV "
"default is 0.01")
Posterior-formatted posterior:
<uttid> [[(0,0.1), (1,0.89), (5,0.01)],
[(1,0,9), (5,0.1)],
...
[(0,0.8), (1,0.2)]]
...
Usage: feat-to-post.py [options] feature_rspecifier posteriors_wspecifier
e.g.
feat-to-post scp:feats.scp ark:post.ark
"""
po = ParseOptions(usage)
po.register_int("top-n", 10,
"only keep highest N posteriors per frame, 10 by default")
po.register_bool("rescale", False,
"rescale top N posteriors to let summation equals to 1, false by default")
opts = po.parse_args()
if (po.num_args() != 2):
po.print_usage()
sys.exit()
feature_rspecifier = po.get_arg(1)
posterior_wspecifier = po.get_arg(2)
isSuccess = feat_to_post(feature_rspecifier, posterior_wspecifier,
opts.top_n, opts.rescale)
if not isSuccess:
sys.exit()
logging.basicConfig(
format="%(levelname)s (%(module)s[{}]:%(funcName)s():"
"%(filename)s:%(lineno)s) %(message)s".format(__version__),
level=logging.INFO)
usage = """Use MiniBatchKMeans for vector clustering. It outputs cluster assignments
For the details, Please refer to website:
https://scikit-learn.org/stable/modules/generated/sklearn.cluster.MiniBatchKMeans.html#sklearn.cluster.MiniBatchKMeans
Usage: kmeans-vector.py [options] <vector-rspecifier> <utt2clusterid-rxfilename>
e.g.
kmeans-vector.py scp:data/train/ivector.scp data/train/utt2clusterid
"""
po = ParseOptions(usage)
po.register_int(
"n-clusters", 8,
"The number of clusters to form as well as the number of centroids to generate. default=8"
)
po.register_int(
"random-state", 0,
"Determines random number generation for centroid initialization and random reassignment. "
"Use an int to make the randomness deterministic. ")
po.register_int("batch-size", 6, "Size of the mini batches.")
po.register_int(
"max-iter", 100,
"Maximum number of iterations over the complete dataset before stopping independently of "
"any early stopping criterion heuristics.")
opts = po.parse_args()
if (po.num_args() != 2):
po.print_usage()
sys.exit()
logging.basicConfig(
format="%(levelname)s (%(module)s[{}]:%(funcName)s():"
"%(filename)s:%(lineno)s) %(message)s".format(__version__),
level=logging.INFO)
usage = """Use t-sne (t-distributed Stochastic Neighbor Emedding) for dimension reduction.
For the details, Please refer to website:
https://scikit-learn.org/stable/modules/generated/sklearn.manifold.TSNE.html#sklearn.manifold.TSNE
Usage: tsne-vector.py [options] <vector-rspecifier> <vector-wspecifier
e.g.
tsne-vector.py scp:data/train/ivector.scp ark:data/train/low_dim_vector.ark
"""
po = ParseOptions(usage)
po.register_int(
"output-dim", 2, "dimension of the output vectors."
" For visualization, only 2 is allowed in this program. (2 by default)"
)
po.register_double(
"perplexity", 30,
"The perplexity is related to the number of nearest neighbors that is used"
" in other mainfold learning algorithms. Large datasets usually require a"
" large perplexity. Consider selecting a value between 5 and 50. Different"
" values can result in significantly different results. (30 by default)"
)
po.register_double(
"learning-rate", 200.0,
"The learning rate for t-sne is usually in the range [10.0, 1000.0]. If the"
" learning rate is too high, the data may look like a \'ball\' with any point"
" approximately equidistant from its nearest neighbors. If the learning rate"
" is too low, most points may look compressed in a dense cloud with few outliers."
" If the cost function gets stuck in a bad local minimum increasing the learning"
# 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 = """Convert features into posterior format, which is the generic format
of NN training target in Karel's nnet1 tools.
(spped is not an issue for reasonably low NN-output dimensions)
Usage: feat-to-post.py [options] feature_rspecifier posteriors_wspecifier
e.g.
feat-to-post scp:feats.scp ark:post.ark
"""
po = ParseOptions(usage)
po.register_int("top-n", 10, "N posteriors per frame, 10 by default")
opts = po.parse_args()
if (po.num_args() != 2):
po.print_usage()
sys.exit()
feature_rspecifier = po.get_arg(1)
posterior_wspecifier = po.get_arg(2)
isSuccess = feat_to_post(feature_rspecifier, posterior_wspecifier,
opts.top_n)
if not isSuccess:
sys.exit()
file=sys.stderr)
return num_success != 0
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_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 "
