Ulan Abdurazakov

Building the fastest, cheapest realtime voice for AI

[mrfakename @realmrfakename]

Qwen3-ASR-Enhanced v0.1

An early checkpoint, consider it a beta release. Soon I will release a better version with more stable nonverbal tag support.

Huge thanks to @huggingface for sponsoring the compute used to train this model!

Thanks to LAION for some of the data.

[Ulan Abdurazakov @defoemark]

@Pendrokar Thanks 👍

[Ulan Abdurazakov @defoemark]

We just open-sourced KaniTTS2 — a 400M param text-to-speech model that runs in 3GB VRAM with voice cloning support.

And we’re releasing the full pretrain code so you can train your own TTS from scratch for any language.
https://t.co/jfLxIRyOlZ, https://t.co/wcv3L74pbK
Apache 2.0

[Ulan Abdurazakov @defoemark]

@ldenoue Got it!

[Ulan Abdurazakov @defoemark]

For those who uses Voice AI, how important is voice cloning feature for you

[Ulan Abdurazakov @defoemark]

@fahdmirza Yeah we need to work on voice cloning! And that “fuggedaboutit” thing

[Ulan Abdurazakov @defoemark]

We have released 2nd version recently

[Hugging Models @HuggingModels]

Meet kani-tts-370m: a multilingual text-to-speech model that's quietly becoming a community favorite. It turns written text into natural-sounding speech across multiple languages. Perfect for devs building accessible apps!

[Ulan Abdurazakov @defoemark]

@Prince_Canuma Ya good point, updated license. Thanks 👍

[Ulan Abdurazakov @defoemark]

Seems like we’ve made the most realistic Voice AI assistant. And completely useless. But he is funny

[Ulan Abdurazakov @defoemark]

@FabioAngela79 Start with 200 hrs and then up to 1000hrs. It should start talking Italian after 1 epoch of 200hrs

[Ulan Abdurazakov @defoemark]

What if you could build your own TTS model that speaks your language - your accent - for around $200? Then host it locally or in the cloud for a tiny fraction of what any proprietary voice AI service charges. Sound interesting?

[Ulan Abdurazakov @defoemark]

@gabrielstuff France is top 1 in voice AI

[Ulan Abdurazakov @defoemark]

@ysu_ChatData The main tradeoff is that you cannot put a lot of languages into onr model. Better to build seperate models for each language, or group of languages. We will ship streaming and batching example soon

[Ulan Abdurazakov @defoemark]

@kgrchz Yeah mix of synthetic and real world data is the best scenario imho

[Ulan Abdurazakov @defoemark]

@kgrchz Yes. And it will sound exactly like them. Those voices sometimes lack human likeness but generally are more stable

[Ulan Abdurazakov @defoemark]

@dnl23 @rodrimora Actually we wanted to try Greek, even found dataset on HF: https://t.co/1SZBhEMVv9. It needs a seperate model definitely. Just haven't gotten around to it yet

[Ulan Abdurazakov @defoemark]

@kgrchz The main issue - where to find enough data. Decent quality speech recordings

[Ulan Abdurazakov @defoemark]

@TyagiShailesh Thanks!

[Ulan Abdurazakov @defoemark]

@peregil We have an example of how to make dataset. A mix of real world speech recordings: 1-40 sec. You can do up to 1 min too. https://t.co/j3DxP8WTo7

[Ulan Abdurazakov @defoemark]

@untalpablodz Stay tuned. Will be at least 20x cheaper with about the same quality

[Ulan Abdurazakov @defoemark]

@prateekhh Drop me a line. We have only Hessian rn, would love to find more