# Kyutai — AURA: Audio-Driven Streaming Avatars in Real Time

- Company: Kyutai (kyutai.org)
- Announced: 2026-11-20
- Category: new-model
- Coverage: not counted
- Announcement: yes
- Group: models
- Source: https://kyutai.org/blog/2026-11-20-aura
- Record: https://forck.live/items/39258-aura-audio-driven-streaming-avatars-in-real-time
- Subject: Moshi / Pocket TTS
- Open weights: yes
- Models affected: AURA 1.3B, AURA 5B

Kyutai built AURA, a family of causal, few-step speech-to-video models that generate talking-head video chunk by chunk without audio look-ahead, for streaming uses such as voice assistants and virtual avatars. AURA distills 50-step bidirectional teachers into 4-step causal students that stream with a rolling KV cache, so only the current chunk's audio is needed to start generating. AURA 1.3B streams 576p video in real time on a single GPU at 20.7 FPS for 16 Hz playback, 23× faster than its teacher, while AURA 5B streams 720p at 9.7 FPS, matching its TalkVerse teacher at 17× the speed. Kyutai released the paper, the code, the weights of both models and a dataset of 30K synthetic videos with their precomputed teacher trajectories, plus a live demo. AURA 1.3B gives a face to Unmute, Kyutai's voice AI, generating the avatar live from Unmute's speech in real time on a single H100 GPU. With 4 denoising steps instead of 50, AURA matches or improves on its TalkVerse teacher across most automatic metrics, and both models have the best identity preservation of all methods on both test sets. In an anonymous human evaluation of 400 pairwise comparisons from 27 raters, both AURA models were preferred over Hallo3 on all four axes.

## Evidence

Verbatim from https://kyutai.org/blog/2026-11-20-aura:

> Audio-driven video diffusion models can now synthesize realistic talking heads, but they generally need dozens of denoising steps over multi-second video chunks, and the audio of the whole chunk before they can start. That rules them out for streaming applications such as voice assistants and virtual avatars, where the video must follow the voice with low latency.
> 
> We built AURA ( AU dio-d R iven streaming A vatar), a family of causal, few-step speech-to-video models. AURA distills 50-step bidirectional teachers into 4-step causal students that generate talking-head video chunk by chunk with a rolling KV cache, with no audio look-ahead: only the audio of the current chunk is needed to start generating. It comes in two sizes. AURA 1.3B streams 576p video in real time on a single GPU (20.7 FPS for 16 Hz playback, 23× faster than its teacher), and AURA 5B streams 720p video at 9.7 FPS, matching its TalkVerse teacher at 17× the speed. As part of Kyutai’s open-science mission, we release the paper , the code , the weights of both models and a dataset of 30K synthetic videos with their precomputed teacher trajectories, and a live demo .
> 
> Talk to AURA

---

Record: https://forck.live/items/39258-aura-audio-driven-streaming-avatars-in-real-time
Catalogue: https://forck.live/llms.txt
Current issue: https://forck.live/feed.md
