Ultravox Import + Creating training bouts about deepsqueak

[DrCoffey]

hmm, I'll take a look at what we updated. It looks like some of the detections are outside the range of the audio file. It is trying to make a box, but there isn't audio there to box.

The fastest way for us to deal with this is if you could send one of the offending files over google drive or something.

[DrCoffey]

Hey Bavo,

I think the first error was because you have a call at the very end of the file. I think I fixed it, let me know if that worked!

-Kevin

[BavoKempen]

Hey Kevin,

Unfortunately it did not work.

Kind regards,

Bavo

[MxMarx]

I think I fixed the ultravox issue now.

Regarding network training, that issue might be the result of training images that are smaller than the neural network input layer. I think I've seen a similar issue before which went away after expanding the boxes around the shortest calls.

We're also in the process of hand-reviewing more files from different recording conditions in an effort to train networks that are more generalizable. If you have any human reviewed data, we'd be more than willing to add them to our training set.

-Russell

[DrCoffey]

Hey Bavo,

The error had to do with extremely short calls being too small for the network training. I think we fixed it, so let us know if that fixes your issue!

-Kevin

[DrCoffey]

Hey Bavo,

I'm going to close this particular issue, but if you run into any more issues, feel free to open another.

-Kevin

[BavoKempen]

Dear Kevin and Russell,

Firstly, my apologies for the very delayed answer, I did not find the time to test and run the aforementioned solutions to my problems.
Secondly, the import issue of Ultravox is solved indeed. Nonetheless, some error still occur with both creating image bouts and training which I will post in a new thread.

Kind regards,

Bavo