Live testing not good

[songpu2015617]

Dear Authors:

We trained your model actually (topcoderCRNN) which is good at training and even regression test(which means unseen data from training). But the live test is really bad. Live means we use microphone recorded data to test. Could you let us know what's the reason causing this? Is it the feature issue or the micro recorded data are much different from training data? Thanks.

[Bartzi]

Hmm, I'd say it is because the data you record via microphone is quite different to the training data and thus not on the trained data manifold. You should be able to solve this problem by incorporating some training data that is similar to your recordings.