Status log

[Kyubyong]

22 Nov. 2017. Has completed the first draft. I've tested the current hyperparameters on only Nick dataset which is 8 hours long, but not on LJ which is 24 hours long. The results were not good, not terrible. As I tried with the same hyperparameters as the original paper with no success, I changed some of them. Amongst them are application of dilation and positional embedding instead of positional encoding. I found the attention plot of the last layer looks monotonic somewhat, but not clearly. I think the key signal that the network works is, of course, the attention plots.

[lsq357]

Thank you for your great work!
Can you show me the environment such as python2 or python3, tensorflow version and so on?

[Kyubyong]

Sure. Python 2 TF 1.3 linux

[lsq357]

Thanks!
I am using LJSpeech-1.0 dataset to train, Can you show me the alignment curve after convergence and how many steps I need to train only using Nick dataset.

[lsq357]

Any plan to add `JOINT REPRESENTATION OF CHARACTERS AND PHONEMES' as the (deepvoice3,part 3.2) saying

and the keithito tacotron experiments showed faster convergeｎｃｅ.

Also, as my experiments, JOINT REPRESENTATION OF CHARACTERS AND PHONEMES does coverge fasteｒ.

[arijit17]

Do you have the synthesized speech files somewhere?

[chenxf0619]

hello， Kyubyong, we have pull your code, we test your code with LJ-speech data. we found the synthesized wav files has nothing to do with the content of the "test_sents.txt". Do you have any guide for us?