What’s your preferred model?

[treynortetik-creator]

Just starting to play around with this and don't have time to try out all the models.

Do you have a recommendation for a tts model for the best overall quality? I have the hardware to run larger ones, so that's not a concern. Just looking for the best quality with, ideally one I can add emotional tags to.

Also, have you had any luck getting an LLM tag a book for emotional tags for a tts model to read?

[zeropointnine]

Hi, MOSS-TTS is the best open-weights model at the moment when memory or inference time is not a concern, IMO. It's expressive without sounding awkward, has the best voice likeness IMO, and has a nice, fluid delivery. Its accuracy is very good, though most all models released in the past half-year are all good with accuracy.

I did play with LLM "auto-tagging" a little bit, but results so far have been... not that awesome, at least for the "bulk audio generation" use case.

I added an experimental stand-alone script in testx/tag_annotator.py for use with Higgs V3, which does respond to tags half-decently, but the output isn't consistently non-awkward enough to make it a net-positive (without reviewing each output and re-rolling with a new seed, etc).

The other model supporting a broad array of tags that I tried this with is Fish S2, but I found its tag behavior to be quite weak and inconsistent.

[treynortetik-creator]

Thank you for the info! This is exactly what I was looking to know. Much appreciated, and thanks for all your work on the app!