Exaggeration and cfg_weight not supported in turbo model
#22
by
Jmurata
- opened
Hi, I saw in the model yaml file at line 299 this: supports_cfg: false . What about exaggeration? In one implementation when passing these parameters to the tts engine there is a warning: "CFG, min_p and exaggeration are not supported by Turbo version and will be ignored". Can you clarify a little this issue? Also, does the multilanguage model ignore also exaggeration and cfg_weight? Thanks
Hi
@Jmurata
Yeah they were removed for the following reasons
- CFG: this requires a batch size of 2 for a single generation. Since this model is optimised for speed we decided to remove it. There are literally 2x less compute because of this.
- Exaggeration: We did train with this enabled - however the resulting naturalness was not as good as without. I suspect this is as a result of the smaller param size in this model.
So generally I would recommend using the original model if CFG/exaggeration is preferred.
Hi, thanks for the info, is the multilingual model also lacking CFG/exaggeration? Thanks