Provide speaker embeddings to produce mel

[savinay]

Can we provide our own speaker embeddings to produce mel spectrograms using flowtron rather than use the speaker embeddings generated by flowtron? If yes, how should we normalize those embeddings?

[rafaelvalle]

yes, re-train the pre-trained Flowtron LibriTTS2K using your own speaker embeddings, preferably at least on LibriTTS2K.

[savinay]

Thanks for your reply! what are the settings to provide my own speaker embeddings? I have trained another model which generates speaker embeddings for the LibriTTS dataset. I would like to use those embeddings to train flowtron model.

[rafaelvalle]

there's no pre-written code for loading speaker embeddings externally. you'll need to change flowtron.py.

[savinay]

Thank you very much! that helps.