Menu

Best choice for feature extraction tool

Help
bat10
2010-07-14
2012-09-22
  • bat10

    bat10 - 2010-07-14

    Hi!

    The perl script that come with SphinxTrain use wave2feat as the feature
    extraction tool. It seems to me, however, that you cannot tell wave2feat which
    features to extract, i.e. 1s_c_d_dd or others. The -feat options does not
    appear to accept anything other than "sphinx", whatever that is supposed to
    mean. On the other hand, the feature stream can be specified in
    sphinx_train.cfg. How does wave2feat know which features to extract, i.e. what
    feature stream the user specified in the config file?

    Aside from wave2feat, there are at least two other tools for feature
    extraction: sphinx_fe and agg_seg. sphinx_fe takes very similar arguments as
    wave2feat, but not exactly the same. It doesn't seem to allow telling which
    features to extract either. This is in contrast to the (deprecated?) agg_seg,
    which appears to work rather differently.

    All this leaves me with the question: What tool should I use for feature
    extraction, if I want to train with, and decode, MS WAV files with SphinxTrain
    and the Sphinx 3 decoder according to some user-defined feature set? And why
    are there so many?

    Advice is appreciated!

     
  • Nickolay V. Shmyrev

    It seems to me, however, that you cannot tell wave2feat which features to
    extract, i.e. 1s_c_d_dd or others.

    In CMUSphinx features with derivatives are computed by all applications on the
    fly. In file only cepstral vector is stored which is wave2feat task. There is
    no need to specify 1s_c_d_dd to wave2feat. 1s_c_d_dd feature format is
    specified to training programs and decoder.

    This is in contrast to the (deprecated?) agg_seg,

    agg_seg i snot for feature extraction. agg_seg accumulate feature vectors and
    use the accumulated values to quantize the vector space. This functionality is
    useful in semi-continuous training initialization.

    What tool should I use for feature extraction, if I want to train with, and
    decode, MS WAV files with SphinxTrain and the Sphinx 3 decoder according to
    some user-defined feature set?

    Extract cepstral feature vector with wave2feat or write your own code to
    extract feature vectors and store them into CMUSPhinx binary feat file.
    Specify lengh of feature vector in sphinx_train.cfg. After that specify
    feature format (1s_c, 1s_c_d_dd or some other) in sphinx_train.cfg to be used
    across training and decoding code.

    And why are there so many?

    There is only tool in Sphinxtrain to extract features - wave2feat.

     
  • bat10

    bat10 - 2010-07-14

    Thanks for clarifying, this makes sense.

     

Log in to post a comment.

Want the latest updates on software, tech news, and AI?
Get latest updates about software, tech news, and AI from SourceForge directly in your inbox once a month.