Skip to content

LongCat-AudioDit #11

Description

@honestabelink

https://github.com/meituan-longcat/LongCat-AudioDiT

This is the only model that "actually" surpassed Index-TTS in term of speaker similarity for me.
Many of the voices I care about are near out of distribution on most models due to lack of dataset coverage.
LongCat-AudioDiT and Index-TTSv2 so far are the only ones to meaningfully get close.

The only issue is it does have a slightly higher WER than I would like.

Would love to see a formal implementation. Though honestly we're one release away from doing everything on the cpu + faster then realtime, kinda like how moonshine is good enough for most ASR applications.

PS.
https://github.com/neuphonic/neutts - cpu
https://github.com/ekwek1/soprano - cpu
https://github.com/kdrkdrkdr/supertonic.embed - only because they have a great model + dumb position on what monetization in the tts space looks like

Thanks for your work. I plan to benchmark audio.cpp against my optimizations as well. If I have anything to contribute I will open a PR in the future.

Metadata

Metadata

Assignees

No one assigned

    Labels

    new modelRequest for new model support

    Projects

    No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions