-
Notifications
You must be signed in to change notification settings - Fork 2.6k
New issue
Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.
By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.
Already on GitHub? Sign in to your account
[TTS] Add Mixed Representation Training #3473
Conversation
blisc
commented
Jan 20, 2022
•
edited
Loading
edited
- Adds phoneme probability to our current g2p module.
- Updates FastPitch yaml to use this functionality.
Signed-off-by: Jocelyn Huang <jocelynh@nvidia.com>
Signed-off-by: Jason <jasoli@nvidia.com>
Signed-off-by: Jason <jasoli@nvidia.com>
Signed-off-by: Jason <jasoli@nvidia.com>
Signed-off-by: Jason <jasoli@nvidia.com>
Signed-off-by: Jason <jasoli@nvidia.com>
Signed-off-by: Jason <jasoli@nvidia.com>
Signed-off-by: Jason <jasoli@nvidia.com>
Signed-off-by: Jason <jasoli@nvidia.com>
See this bug: #3593 |
Signed-off-by: Jason <jasoli@nvidia.com>
Signed-off-by: Jason <jasoli@nvidia.com>
Signed-off-by: Jason <jasoli@nvidia.com>
Signed-off-by: Jason <jasoli@nvidia.com>
Signed-off-by: Jason <jasoli@nvidia.com>
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
Great PR, thanks for this feature. Need to fix comments.
@@ -10,8 +10,8 @@ sup_data_path: ??? | |||
sup_data_types: [ "align_prior_matrix", "pitch" ] |
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
Some important notes about FastPitch + Aligner in mixed representation training overall:
- Did you try to train this model? According to FastPitch paper, "However, the mixed approach of training on phonemes and graphemes introduced unpleasant artifacts", have you checked that the artifacts do not appear? If so, it is very-very big result!
- Despite the fact that this config is very cool, I would prefer to create separate config for mixer representation training, because it is not ordinary case. Moreover,
beta_binomial_interpolator
is not so accurate. What do you think about it? I suggest to keep it unchanged and createconf/fastpitch_align_mixed_representation_v1.05.yaml
(we will rename it toconf/fastpitch_align_mixed_representation.yaml
in 1.8.0).
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
I think this will be the default fastpitch config moving forward as it is what will be used in riva.
Signed-off-by: Jason <jasoli@nvidia.com>
Signed-off-by: Jason <jasoli@nvidia.com>
This pull request introduces 2 alerts when merging 4f09e27 into 8313b01 - view on LGTM.com new alerts:
|
Signed-off-by: Jason <jasoli@nvidia.com>
Signed-off-by: Jason <jasoli@nvidia.com>
Signed-off-by: Jason <jasoli@nvidia.com>
Signed-off-by: Jason <jasoli@nvidia.com>
Signed-off-by: Jason <jasoli@nvidia.com>
Signed-off-by: Jason <jasoli@nvidia.com>
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
Great PR, thanks.
We need to think more about accuracy of beta binomial interpolator and also think about potential bug in TTSDataset
* Update CMUdict with ADLR version pronunciations Signed-off-by: Jocelyn Huang <jocelynh@nvidia.com> * minor updates for finetuning Signed-off-by: Jason <jasoli@nvidia.com> * update conf Signed-off-by: Jason <jasoli@nvidia.com> * merge Signed-off-by: Jason <jasoli@nvidia.com> * update Signed-off-by: Jason <jasoli@nvidia.com> * update Signed-off-by: Jason <jasoli@nvidia.com> * bug fixes Signed-off-by: Jason <jasoli@nvidia.com> * update config Signed-off-by: Jason <jasoli@nvidia.com> * bf16 support Signed-off-by: Jason <jasoli@nvidia.com> * bf16 support Signed-off-by: Jason <jasoli@nvidia.com> * bugfix Signed-off-by: Jason <jasoli@nvidia.com> * update Signed-off-by: Jason <jasoli@nvidia.com> * finalize changes Signed-off-by: Jason <jasoli@nvidia.com> * undo notebook 1.6.0 pins Signed-off-by: Jason <jasoli@nvidia.com> * more 1.6.0 undos Signed-off-by: Jason <jasoli@nvidia.com> * wip Signed-off-by: Jason <jasoli@nvidia.com> * update num_workers Signed-off-by: Jason <jasoli@nvidia.com> * update hypers Signed-off-by: Jason <jasoli@nvidia.com> * revert to main _align yamls Signed-off-by: Jason <jasoli@nvidia.com> * update yamls Signed-off-by: Jason <jasoli@nvidia.com> * cleanup Signed-off-by: Jason <jasoli@nvidia.com> * remove unnecessary line Signed-off-by: Jason <jasoli@nvidia.com> * address comments Signed-off-by: Jason <jasoli@nvidia.com> * update vocoder mel uploading; add contextmanager to mixed g2p Signed-off-by: Jason <jasoli@nvidia.com> * update comments; make prob required argument Signed-off-by: Jason <jasoli@nvidia.com> * added val check Signed-off-by: Jason <jasoli@nvidia.com> * update message Signed-off-by: Jason <jasoli@nvidia.com> * update Signed-off-by: Jason <jasoli@nvidia.com> * revert num workers Signed-off-by: Jason <jasoli@nvidia.com> Co-authored-by: Jocelyn Huang <jocelynh@nvidia.com>