First acestep.cpp, now this?

[theaerotoad]

Amazing start on this one, @ServeurpersoCom. I'm still enjoying acestep.cpp , and went looking a cpp / ggml based alternative to my current local TTS server. Can't believe you picked omnivoice up too. Thanks for sharing this.

You might consider adding a "buy me a beer/coffee" link at some point. I think I owe you several.

[ServeurpersoCom]

Thanks, that's great! Let me know if you find a bug :)
It's finished, however, I'll probably redo the GGUFs later to conform the tensor names to the llama.cpp standard (at least for the LLM architecture, and follow KoboldCpp for what isn't standardized in llama.cpp but which they have already standardized).

I'm trying to GGML-upstream the accelerated audio math operators (SNAKE and COL2IM_1D); they're used everywhere for lots of fun things, like Qwen TTS or Omni. These will also be available in minimal CPP, which will allow me to see how to merge the code into another project that supports multiple TTS models.

I've put it on a Telegram bot: you send a voice message with a voice taken from TV, and the AI ​​(also a local LLM) speaks like that! But the ultimate goal is for the llama.cpp WebUI to be able to speak, in real-time streaming (Qwen3 Omni and other), to do like the ChatGPT app..

[ServeurpersoCom]

You can also try this one: https://github.com/ServeurpersoCom/qwentts.cpp
it works almost the same! I still need to write the documentation, but the GGUFs are available here: https://huggingface.co/Serveurperso/Qwen3-TTS-GGUF

[theaerotoad]

Thanks! I'd struggled getting some other ggml based variants of qwentts going in Vulkan well earlier. Honestly, this omnivoice port is so good, I settled happily on this as my TTS backend for all my local homegrown tools. But I'll check out qwentts too!

I have a webserver version of this (back before the nicer C api fixes) this is OpenAI API compatible, preloads voices (shaves off a few ms) and has upload to add new voices. It was about 10 back and forths with Gemini--very happy to share or do a pull if it's of interest.