what have you modify in this repo to support cross language ?

[emmacirl]

@Kyubyong Thanks for your contribution! I found that the ideas that net1 for recognition and net2 for synthesis is similar with 'deep voice conversion'. So I am wondering what have you modify to support cross language? Looking forward to your reply

[Kyubyong]

Actually nothing but one: replacing single phonemes with triphones. I know it's not enough for this task, so I'm looking for a better architecture/model.

[emmacirl]

@Kyubyong Thanks a lot for your fast reply! So in order to support cross language different language database is applied to train net2, is it ?

[Kyubyong]

No for training always English speech samples are uswd. For inference different languages are to be tested.

[emmacirl]

The TIMIT is labeled with triphones, and Net2 is trained with English. Finally voice with any language can be synthesized?

[Kyubyong]

Exactly.

[emmacirl]

Got it !

[lightwithshadow]

Thank you for your contribution. How long did it take you to train Net1? thank U Looking forward to your reply！