[Question] Do you take contributions?

[blazingbhavneek]

I recently tested this model at work, and it worked very well, we might use it further, the Delay model works very well but is slow, can I try to use SGLang/vLLM as backend and maybe add PR for that? That will increase its speed by a lot, while at it I might try quantization too so it works with smaller GPUs.

[gaoyang07]

Thanks for your interest in our model! We are actually planning to release the SGLang implementation of our MOSS-TTS model right after the Chinese New Year break. It will be available here: https://github.com/OpenMOSS/sglang

[blazingbhavneek]

Great!
Although is sglang willing to add a TTS model? vllm has vllm-omni project for multimodal generations but sglang seems to sticking to text/image only. I see some open issues for TTS models though

I will try to Quantize the model then for concurrency, or do you plan to release official Quantized variants?

[richservo]

https://github.com/richservo/comfyui-moss-tts

I made a comfy node set. Feel free to use it. Great model!!

[gaoyang07]

Great! Although is sglang willing to add a TTS model? vllm has vllm-omni project for multimodal generations but sglang seems to sticking to text/image only. I see some open issues for TTS models though

I will try to Quantize the model then for concurrency, or do you plan to release official Quantized variants?

It would be awesome if you were willing to do it 🎉. Any contribution is welcome.

[blazingbhavneek]

@gaoyang07 Can you upload/send some evaluation scripts? I tried the ones in seed-tts-eval on my work GPU and on the original model it's giving WER around 6 in English dataset. I have a Quantized version I want to evaluate it for performance loss

[YWMditto]

We currently use the widely accepted seed-tts-eval and cv3-eval for evaluation:
https://github.com/BytedanceSpeech/seed-tts-eval
https://github.com/FunAudioLLM/CV3-Eval/tree/main

[YWMditto]

The official quantitative version and llama.cpp are both open source. Please check the latest readme for usage instructions.

[YWMditto]

If there are no further issues, we will close this issue for now. We are very pleased that you are interested in and using our model. We will increase our open-source efforts in the future, so please stay tuned!

[gaoyang07]

@blazingbhavneek SGLang backend is ready now. Feel free to try it 🫡