Fix conditional drop for CFG when conditioning on semantic/phoneme tokens

lucidrains / voicebox-pytorch

Implementation of Voicebox, new SOTA Text-to-speech network from MetaAI, in Pytorch

MIT License

589 stars 49 forks source link

Fix conditional drop for CFG when conditioning on semantic/phoneme tokens #36

Closed lucasnewman closed 9 months ago

lucasnewman commented 9 months ago

I noticed a minor issue where the semantic/phoneme conditioning wouldn't have the null token ids populated in the input to the embedding. (The code to handle it was already present but wasn't used.)

lucidrains commented 9 months ago

🤦 thank you Lucas!