Cloning/extracting voices on Base talker, injecting it into CustomVoice for instruct support

[hat-commits]

This project (https://github.com/gabriele-mastrapasqua/qwen3-tts) has a useful feature where you create a cloned voice to a .qvoice file once, then reuse it even on the CustomVoice model so you get --instruct style control (without having to do a whole training/fine-tuning session)

Base and CustomVoice models share 99.98% identical transformer weights. The only meaningful differences are in the codec embedding table and a few special tokens. The Delta format exploits this by storing only the per-weight differences (int16 deltas, LZ4 compressed). At load time, these deltas patch the CustomVoice weights to exactly match the Base model — producing PCM-level bit-identical output.

https://github.com/gabriele-mastrapasqua/qwen3-tts/blob/main/docs/custom-voices.md

I am not sure how feasible this is, but it would be really neat.