Skip to content
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

Problem with very short and noisy audio during inference when providing xvector embeddings #80

Open
Phlayne opened this issue Jul 3, 2024 · 1 comment

Comments

@Phlayne
Copy link

Phlayne commented Jul 3, 2024

Hello,

I am pretty new to ESPnet and I am attempting to perform inference using the vctk_tts_train_xvector_transformer_raw_phn_tacotron_g2p_en_no_space_train.loss.ave pretrained model.

Steps Taken:

  • I used speechbrain/spkrec-xvect-voxceleb to create speaker embeddings for specific voices.
  • I provided one of these embeddings to the pretrained TTS model.

The problem is that the generated audios are extremely short (0.125 or 0.013 seconds) and sound noisy.

I am using the Python API. I only provided text and spembs fields when calling the Text2Speech class. I also have successfully used the Python API with other pretrained models that do not require speaker embeddings. I am unsure if there are additional arguments or steps required when using this specific model with speaker embeddings.

If more information is needed, I am happy to provide it. Has anyone experienced a similar issue or can provide guidance on how to resolve this?

Thank you for your assistance,

@sw005320
Copy link
Collaborator

sw005320 commented Jul 3, 2024

maybe, I'm wrong, but the model might use the Kaldi xvector, not speechbrain one.

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment
Labels
None yet
Projects
None yet
Development

No branches or pull requests

2 participants