[Feature Request] Kokoro-style Voice Mixing

[OwenTyme]

In the past, I've used a CLI fork of Kokoro that allows blending voices together, to produce unique mixtures. I always found that to be a fascinating and very useful thing, which I miss since I've started using Pocket in its place.

Here's the fork I'm talking about, linking to the section on command-line options, where that's spelled out:
https://github.com/nazdridoy/kokoro-tts?tab=readme-ov-file#options

It's based on percentages, like so: 'voice1:30,voice2:70'.

I really don't entirely understand how it works and have no idea if the same thing is possible with Pocket, but I'd love having that as a feature, because it allows toning down a harshly accented sample into something less so, or tweaking a voice with a little hint of another.

My understanding is that the af_heart from Kokoro is a synthetic voice produced in this fashion. If I remember right, it's a blending of af_bella and one other.

So, is there any way to do this with Pocket? Again, I'd love this to be a feature.

Thank you.

[manukyutai]

Hello!

Fun project! I've never tried that, but here are two ways you may be able to achieve that off the top of my head:

• Generating a mixed sample with your Kokoro branch and using it as prompt
• Mixing CFG style: Every step, compute to latents out of the backbone transformer, one from voice 1 and one from voice 2. Perform a weighted average of those latents, and use this to condition the MLP sampler.

[OwenTyme]

Hello!

Fun project! I've never tried that, but here are two ways you may be able to achieve that off the top of my head:

* Generating a mixed sample with your Kokoro branch and using it as prompt

* Mixing CFG style: Every step, compute to latents out of the backbone transformer, one from voice 1 and one from voice 2. Perform a weighted average of those latents, and use this to condition the MLP sampler.

To be honest, I was trying to make a feature request and while I've made an attempt to look over the code, I don't understand enough of it to even begin to try your suggestion.

Python is not a programming language I've ever really been comfortable with, so I'm totally out of my depth on this one.