Kadir Nar

I am developing the VoiceHub library to run popular TTS models in a single library. Currently, it supports the Orpheus, Vui, and Dia models. I will add other models(llasa,kokoro, styletts, melotts, f5tts...)

@blu0190 @oguzergin neden yeni domain alsın basit bir şey için?

Tantuni endeksi sayfasından bazı görüntüleri paylaşıyorum. tantuni.oguzergin.net

Github: github.com/Vyvo-Labs/Vyvo…

The Orpheus and VyvoTTS models now support sglang and vllm libraries. Additionally, the snac model now runs much faster. TTFT: VLLM: 6 ms SgLang: 10 ms

@ChristophSchuh6 Aratako had 30k hours of emotion data in their dataset. There are very few open-source emotion datasets for English. So making a voice design model can be very difficult. Maybe I can do this by producing synthetic data with echo and qwen3-tts, but it might not sound natural.

@kadirnardev En voice design :)

I made more Triton-based optimizations to the Snac codec model to further speed up the Orpheus-TTS and VyvoTTS models. Don't forget to star my GitHub repo for more optimizations!

@ChristophSchuh6 Do you want the English version of the irodori-tts model?

@kadirnardev I think it would be interesting to train an Echo-like diffusion transformer that ingests SNAC artifact-corrupted audio and outputs high-quality 48 kHz DAC VAE without artifacts. Maybe a 400-million-parameter model or something like that could do it, just fixing artifacts. 🙂

@kint0kur Önemli olan VLLM ve SgLang desteği. Bunları eklersen iyi sonuç alırsın. Flow matching tabanlı da model tasarladım ve TTFT değeri 10ms felandı.

@kadirnardev Vyvo veya orpheus farketmez aslında. Single GPU üzerinde paralelde latency çok yükselmeden ne kadar request karşılayabiliyor onu merak ediyorum. Flow matching tabanlı bir model üzerinde çalıştım scale etmek çok zor autoregressive de çalışmıyorlar ttft çok yükseliyor paralelde.

@kint0kur Bu codec modeli. Önemli olan hangi TTS modeli kullandığınız. VyvoTTS için daha önce 100 kullanıcı için TTFT değeri 70-80 ms değerindeydi. Bu optimizasyon ile daha hızlı olacak.

@kadirnardev Paralel istek durumunda latencyler ne durumda? Single H100 üzerinde kaç tane paralel requesti karşılayabilir?

@PuterOnX Yes, and it's not open source

@kadirnardev Qwen omni just came out right?

The Qwen team is no longer releasing their models as open source, and this is a big problem for us. We need small models to train many models like TTS, STT, Omni, and others. Previously there was LLaMA, but they're no longer releasing either. The Qwen team won't be releasing anymore either. Our only hope is the LFM models. Minimax, Kimi, and GLM teams are releasing great models for open source, but none of them release small models. And if these companies also stop releasing open source, it's going to be really bad :(

@foreignsplat Yes, I saw your new models and I'm very happy about it.

@kadirnardev Liquid.ai

@yukiarimo Should we still use Llama models?

These models perform great since they're newly released, but in 3-4 months we'll need better models. For example, if the LFM team doesn't release new small models, would it be okay for you to use old Gemma models? New models should be released constantly. When Gemma was first released, they were great models, but now they're not up to date.

It says 39 message requests in the chat section, but when I click on it, there are no messages. If I didn't reply to your message, send another message.

@overlordayn They didn't release the TTS model. Why aren't we training multiple LFM-based models?

@kadirnardev They have an audio model that does tts and stt both (lfm2.5 audio)

Should I train TTS with the LFM2.5-350M model or the Omni model?

@billyG881 The Neucodec model quality isn't good. I'm thinking of using a better codec.

@kadirnardev Hopefully with NEUCODEC codec as its SOTA and has been trained on plenty of multi-lingual data 🤓🤓🤓

@jesujopi3D Data quality is bad.

@kadirnardev Would this one work? datacollective.mozillafoundation.org/datasets/cmn4z…

@jesujopi3D If you have an open source dataset, I can train it.

@kadirnardev Are you considering training any TTS model with multilingual datasets or Spanish data? I’ve been using Queen 3 TTS, but the VRAM consumption I get even with quantized models is a bit high for a 12GB card like mine…

@WaelShaikh I had trained LFM models before, training them again could be good. x.com/kadirnardev/st…

@kadirnardev Definitely on the 350M model. Would love to see how it performs. LFM makes some of the fastest LLMs, I wonder if the speedup would even benefit the TTS.

@liquidai Thank you for publishing it as open source❤️

Today, we release LFM2.5-350M. Agentic loops at 350M parameters. A 350M model trained for reliable data extraction and tool use, where models at this scale typically struggle. <500MB when quantized, built for environments where compute, memory, and latency are constrained. 🧵