Has emotion support only existed for coqui API users?

[matbee-eth]

What would it take to get emotion control out of this?

[eginhard]

Yes, the emotion arguments in

coqui-ai-TTS/TTS/api.py

Line 243 in a682fa8

emotion: str = None,

could only be used with the proprietary Coqui Studio models, so they are no longer useful. If someone adds a model that supports generating emotional speech, they could be hooked up to that.

[IIEleven11]

Theoretically XTTSv2 should be able to do this because it is a child of the tortoise model and this is where that emotion control comes from. I've been trying to decipher this over the last few days and I found my way here to see if anyone else has any information. Along with some other things it appears intentionally obfuscated/omitted. The tokenizer/vocab show little to help. I am maybe jumping to conclusions though, if anyone has any insight please let me know.

[agonzalezd]

An option could be adding another embedder for tags, similar to a speaker embedding, only for emotions. This embedder would simply generate an embedding for a label instead of using another neural model for computing this feature, using this, for example. A small embedding could do it, since no more than 10 emotions will be used in a dataset (hopefully). The embedding dimesion could be updated in the model configuration.