Closed Ryu1845 closed 9 months ago
Thanks for the interest. I plan to release some of the codes after the peer review process. If you plan to try it earlier, you can simply refer to a VALL-E's code, add a loop over all relative positions and train it with transducer loss.
The transducer loss has been implemented in pytorch: from torchaudio.functional import rnnt_loss
.
Nice, thank you very much
Any ETA on the code release?