transducer

[BestSonny]

@awni In the infer code, are you still using ground truth labels for testing phase? This confused me since we do not have a ground truth when applying to an unseen data. Or do you just forward a fake input (such as batch_size x 1 with all zero label) when in the testing phase?

Also, will you maintain this excellent project in the future?

Thank you very much.

[awni]

Yes great point. In the committed code the ground truth labels are being used for the transducer. I have some code to do this properly using the inferred output but I haven't had a chance to clean it up and commit it yet. That's why I haven't posted "transducer" results in the README yet.

[awni]

And yes, I will continue to support this project though any collaborators would be very welcome!!

[BestSonny]

@awni Have you made any progress on cleaning up the decoder?

[yjchun]

Just for reference, here is my inference code.
https://gist.github.com/yjchun/aae0b86f89ffb46a54c28e7747124ee4

May not be useful as it is because it uses my model class's internal logic.
The problem is inference is awfully slow. Takes 3 seconds for 10 seconds of audio with 400 hidden layers. 161 chars = 161 prediction+joint network calls at minimum.
Had to read Grave's paper in and out for a few days for this piece of code.

And thank you @awni for sharing your transducer loss implementation which was very helpful.

[Duum]

Here is my implement for greed decoder.
you may need some change for your lstm decoder network.