the timestamp doesn't match? about pocketsphinx-python HOT 9 CLOSED

This is because for some reason -remove_silence and -remove_noise are enabled by default in the pocketsphinx configuration. This makes it drop silence and noise frames, so the frame indices only correspond to actual speech. It speeds up decoding but confuses everybody! You can fix it like this:

cfg = pocketsphinx.Decoder.get_default_config()
cfg.set_boolean('-remove_silence', False)
cfg.set_boolean('-remove_noise', False)
decoder = pocketsphinx.Decoder(cfg)

etc.

from pocketsphinx-python.

Hi @dhdaines

remove_noise should be almost always enabled because models are trained with noise removal feature.

remove_silence is also a very good thing because it allows to properly compute CMN estimate, thus much better accuracy.

The current proposal is to use continuous processing, then timing will be correct. The implementation of continuous processing is not great though.

from pocketsphinx-python.

If you use continuous processing, doesn't this prevent you from getting whole-utterance CMN, though?

from pocketsphinx-python.

I suppose that is only relevant if you're doing offline recognition, of course...

from pocketsphinx-python.

Whole utterance cmn is also ok, and still CMN estimation is best without silence frames which could be really long in real life (2 seconds of silence around short command) unlike in common ASR databases. Thats what remove_silence is doing and this is why it is enabled by default.

Timing and whole-utterance processing are not very easy, it is correct.

from pocketsphinx-python.

Hi, is there a way to get the starting and ending time-step of each phoneme. I am still a little bit confused by that.

from pocketsphinx-python.

from pocketsphinx-python.

Hi, There isn't a particularly easy way to do that at the moment. The search is word-based (otherwise it would be horribly slow) - if you use allphone decoding you will get phone segmentations but the phoneme accuracy isn't very good. The intention is that state align search can be used as a second pass to get phone alignments. (in fact, it will give them to you, but this involves writing code) I am trying to find some time to implement this.
…

Thank you for your reply! I still wondering does continuous processing approximate the time steps?

from pocketsphinx-python.

This module is obsolete; python bindings are now in pocketsphinx

from pocketsphinx-python.