Re-train classifiers for Nattalia and BB data using current code, and put in examples/models/

[msonderegger]

[jkeshet]

@msonderegger Is this issue should be closed?

[msonderegger]

No, still needs to be done, though I don't think it's essential for the beta.

[jkeshet]

@thealk would you like to write a python script for training this data using the AutoVOT?

[thealk]

I mentioned this on another comment but yes, I could help with this however 1) it probably wouldn't be until next week 2) I would need some more information on how to retrain a classifier (both because I have only ever trained new classifiers using the scripts here and because my python skills are a bit rusty at this point). If you're(pl) okay with that, then we can assign this todo to me.

[msonderegger]

So, assuming the point of this issue was to post example classifiers that can be used by users (instead of training their own), I think we might actually want to do so differently. Re-training the nattalia and BB classifiers on the same data as in the JASA paper would take some time at this point, because they use our old "input file" format, and the input files would need to be converted into textgrids containing vot and window tiers.

On the other hand, what would be easy for me or Thea to do is to train classifiers for voiced and voiceless stops, for the full BB dataset I am now working with (a few thousand of each of voiced and voiceless) and for the full Glasgow dataset (similar numbers). We could then post those as examples trained on conversational (British) speech, and the classifiers from the tutorial would serve as examples trained on lab speech.

Does this alternative sound OK?

Yossi -- does this need to be done by this weekend to release v. 1.0? (What's your timeframe for the windows installer, assuming that's still todo before releasing 1.0?) If not, I think it's better if Thea trains these classifiers (though I don't think a python script is needed), to make sure it is doable for someone besides you and I to do so :) Thea -- I don't think this will actually take long; it's just a matter of swapping in pieces from the tutorial. But feel free to say if you don't have time.

[thealk]

Training a new classifier would be doable by this coming weekend for me, it would just be a matter of accessing the full set of bb files and generating a classifier with auto_vot_train.py in that case, right? If we decide to do something that would take a bit longer that's okay with me too, I just wouldn't be able to do it until next week (exams and deadlines).
Yossi, let me know what your ideal time frame would be and we can decide what to do.

On Oct 13, 2014, at 12:30 PM, msonderegger notifications@github.com wrote:

So, assuming the point of this issue was to post example classifiers that can be used by users (instead of training their own), I think we might actually want to do so differently. Re-training the nattalia and BB classifiers on the same data as in the JASA paper would take some time at this point, because they use our old "input file" format, and the input files would need to be converted into textgrids containing vot and window tiers.

On the other hand, what would be easy for me or Thea to do is to train classifiers for voiced and voiceless stops, for the full BB dataset I am now working with (a few thousand of each of voiced and voiceless) and for the full Glasgow dataset (similar numbers). We could then post those as examples trained on conversational (British) speech, and the classifiers from the tutorial would serve as examples trained on lab speech.

Does this alternative sound OK?

Yossi -- does this need to be done by this weekend to release v. 1.0? (What's your timeframe for the windows installer, assuming that's still todo before releasing 1.0?) If not, I think it's better if Thea trains these classifiers (though I don't think a python script is needed), to make sure it is doable for someone besides you and I to do so :) Thea -- I don't think this will actually take long; it's just a matter of swapping in pieces from the tutorial. But feel free to say if you don't have time.

—
Reply to this email directly or view it on GitHub.

[msonderegger]

@jkeshet what do you think? Is my note above clear?