Ulan Abdurazakov (@defoemark) header
Ulan Abdurazakov (@defoemark) profile picture

Crafting realtime voice AI that feels human @nine_nine_six

Oakland, CA
Joined April 2013
Photo by Ulan Abdurazakov (@defoemark)Photo by Ulan Abdurazakov (@defoemark)Photo by Ulan Abdurazakov (@defoemark)Photo by Ulan Abdurazakov (@defoemark)Photo by Ulan Abdurazakov (@defoemark)Photo by Ulan Abdurazakov (@defoemark)Photo by Ulan Abdurazakov (@defoemark)Photo by Ulan Abdurazakov (@defoemark)Photo by Ulan Abdurazakov (@defoemark)Photo by Ulan Abdurazakov (@defoemark)
Congrats! 🔥
Replying to @dialt_ai
@dialt_ai is now on the VAmos frontier for conversational voice models! HMU if you’re training s2s models, would be glad to chat and see how we can help.
1
1
250
Have been using @MOSS_Global’s open-source models, checking out their approaches, and reading the papers for quite some time. Big fan of their work. Congrats on MOSS-Live and good luck with the commercial product!
MOSS-Live is here. Our full-duplex, real-time voice AI can listen, speak, and get things done at the same time. 🎙️ While tasks run in the background, it can keep listening, seeing what you see, and talking with you. • Natural conversation: It distinguishes real interruptions from “mm-hmm,” waits when you pause to think, stays out of side conversations, and gives brief acknowledgments while listening. • Speech awareness & live translation: It tracks where it is in its spoken response, picks up after interruptions, and translates while you speak. • Proactive help: Set rules for when it should correct mistakes, keep count, or remind you. • Tools & background tasks: Check weather or stock prices, or let longer tasks run in the background without putting the conversation on hold. • Multimodal perception: Use video calls or screen sharing to give it visual context. It can also recognize environmental sounds and describe voice qualities. Try MOSS-Live ➡️ shorturl.at/96uVt
1
1
16
996
Ulan Abdurazakov retweeted
Introducing Audio8-TTS-Realistic, @Edge0_AI’s new open-weight speech model: realistic, zero-shot voice cloning in just 0.6B parameters. • Human-like prosody and expressive delivery • Clones a voice from one short clip, in 23 languages • 601M params: small enough to fine-tune and serve affordably Currently #1 on the Hugging Face TTS Arena, where people vote blind on which voice sounds more human. Open weights are available now on @huggingface. Try it free.
63
164
16
1,568
104,940
A new highly efficient audio codec focused primarily on speech
新たな可変フレームレート型ニューラル音声コーデック“RubatoCodec”を開発し、OSSとして公開しました! 重みや学習・推論コード、論文など一通り公開しています。 主にTTS用途での実用性を考えて開発しており、特に英語・日本語ではかなり高いパフォーマンスを実現しています。
1
12
788
Ulan Abdurazakov retweeted
Today we're coming out of stealth, introducing Neywa Labs 🚀 We built the infrastructure to build and deploy AI that runs on the devices people already own. Private, offline, unlimited. One mission: make AI available to everyone in the world. neywalabs.ai/
269
274
107
2,182
332,254
Dessa is dogwalking gpt
5
23
1,323
full duplex is the only way for AI-human interaction
We believe voice is the most natural way to interact with intelligence, so we created Dessa, the most authentic full-duplex speech model ever built. Talking with it feels like talking with a person…full of emotion, nuance, pauses, and a natural cadence. block.xyz/dessa
2
4
231
How does TYGER know if "yeah" means "go on" or "my turn"? It doesn't read transcripts. It listens to both your mic and the agent's own voice through a frozen Nemotron ASR encoder (FP32), and decides every 80 ms: listen, speak, yield, or "mm-hm". No VAD. Works with any LLM + TTS.
1
2
2
28
1,508
Ulan Abdurazakov retweeted
Replying to @defoemark
@defoemark at @nine_nine_six used Open Yap 1K in their new TYGER model. Go try it out on nineninesix.ai/
Meet TYGER: a decision model for speech that opens the way to full-duplex voice agents. It listens to the mic and your TTS at once, and every 80ms decides whether to speak, yield, or chime in with an “mm-hm”. No need for VAD anymore. Works with any LLM and TTS. huggingface.co/nineninesix/t…
3
1
13
468
Meet TYGER: a decision model for speech that opens the way to full-duplex voice agents. It listens to the mic and your TTS at once, and every 80ms decides whether to speak, yield, or chime in with an “mm-hm”. No need for VAD anymore. Works with any LLM and TTS. huggingface.co/nineninesix/t…
35
57
11
839
34,087
Thanks to @matiasdrejer and @agenticdataco for Yap-1K speech dataset that was used to train the model.
1
4
1,287
Most voice agents work like walkie-talkies: you talk, they wait for silence, they talk. People don't: we interrupt, pause to think, and ask question while others talk. Full-duplex means listening while talking. Tyger transcribes and decides when to speak, yield, or chime in.
10
4
2
95
6,068
The hard part of full-duplex isn't generating speech, it's knowing when. A pause can mean "I'm done" or "I'm thinking." Speech over the agent can be an interruption or just "yeah."
1
5
611
14M-param decision head on top of @NVIDIAAI Nemotron ASR. Here is is an example of local full-duplex voice agent (Apache-2.0): github.com/nineninesix-ai/ty… Hosted Demo (English only): nineninesix.ai Feedback on where turn-taking still feels off is very welcome.
1
3
30
2,538
6 weeks of nonstop work, and TYGER is out 🐯 An open-source full-duplex voice agent (and model that enables it) that listens you while you talk or while it talks, chimes in, calls tools and runs on a single GPU. Try it and tell me where it breaks.
Today we're releasing TYGER 🐯 an open full-duplex voice model, and TYGER S2S, an open-source agent built on it. Interrupt it. It listens while you talk and calls your tools. 🎙 Try it: nineninesix.ai/ 🤗 Model: huggingface.co/nineninesix/t… 💻 Code: github.com/nineninesix-ai/ty…
3
1
7
353
Voice acting bench from @laion_ai
AI voices can speak. Can they act? With TTSArena we built a Voice Acting Arena: two models, same scene, you pick the more believable performance. 🎭 We would love to hear what you think! 🙂🦁 huggingface.co/blog/TTS-AGI/…
208