Tacotron2: How to make multispeaker tacotron2?
I wanna try to train it on multiple datasets with multiple speakers, what changes should I make? Any guides/examples?
hadaev8
All 6 comments
Try to use Tacotron GST
asmodaay
on 4 Aug 2019
Add a speaker embedding layer and append the speaker embedding to the encoder outputs, replicating it over the time dimension. Note that you'll have to change the input dimension of your attention rnn.
We're going to release Multispeaker Tacotron GST: stay tuned!
rafaelvalle
on 13 Aug 2019
@rafaelvalle
What mean memory_lengths in decoder? Should I add +1 for speaker embedding?
Also, any eta for this realize? Like week or month?
hadaev8
on 17 Aug 2019
Maybe I need to train more, or should not to start from single speaker checkpoint, but it produces the same voice with every speaker id.
hadaev8
on 19 Aug 2019
Add a speaker embedding layer and append the speaker embedding to the encoder outputs, replicating it over the time dimension. Note that you'll have to change the input dimension of your attention rnn.
We're going to release Multispeaker Tacotron GST: stay tuned!
Hi Rafael, @rafaelvalle
Do you have projected timeline when the Tacotron GST will be released? Looking forward to testing it out!
yxt132
on 19 Aug 2019
https://github.com/NVIDIA/mellotron
Our mellotron repo uses a Multispeaker Tacotron with GST.
rafaelvalle
on 22 Nov 2019
Related issues
MXGray
路
5Comments
yliess86
路
6Comments
mahdeto
路
4Comments
kannadaraj
路
6Comments
deli666
路
5Comments
Most helpful comment
Add a speaker embedding layer and append the speaker embedding to the encoder outputs, replicating it over the time dimension. Note that you'll have to change the input dimension of your attention rnn.
We're going to release Multispeaker Tacotron GST: stay tuned!