Open AI research lab Kyutai has released a groundbreaking streaming voice-to-voice (TTS) model with approximately 2 billion parameters. Designed for real-time responsiveness, the model provides ultra-low latent audio production (220 milliseconds) while maintaining high fidelity. It has undergone an unprecedented 2.5 million hours of audio training and licensed under the permitted CC-By-4.0, strengthening Kyutai’s commitment to openness and repeatability. This advancement redefined the efficiency and accessibility of large-scale speech generation models, especially for edge deployment and proxy AI.
Unlock performance: Latency of 32 concurrent users on a single L40 GPU is less than 350ms
The flow function of this model is its most unique feature. On a single NVIDIA L40 GPU, the system can provide up to 32 concurrent users while maintaining latency below 350ms. For personal use, the model remains low as 220ms for a generation incubation period, enabling almost real-time applications such as conversation agents, voice assistants and real-time narrative systems. This performance is enabled by Kyutai’s novel delay flow modeling method, which allows the model to gradually generate speech as text arrives.
Key technical indicators:
- Model size: ~2B parameters
- Training data: 2.5 million hours of speech
- Incubation period: 220ms single user,
- Language Support: English and French
- license: CC-BY-4.0 (open source)
Delay flow modeling: architecture real-time responsiveness
Kyutai’s innovation anchors in delayed flow modeling, a technology that allows speech synthesis to begin before available full input text. This approach is specially designed to balance the predicted quality with response speed, thereby enabling high-throughput flow tts. Unlike conventional self-cyclotron models with response lag, the architecture maintains time coherence while achieving faster synthesis than real-time.
The architecture’s code base and training recipes are available in Kyutai’s GitHub repository to support full repeatability and community contribution.
Model availability and open research commitment
Kyutai published model weights and reasoning scripts on Embrace Faces, making it available to researchers, developers and business teams. The permitted CC-BY-4.0 license encourages unrestricted adaptation and integration into the application if appropriate attribution is maintained.
This version supports batch processing and stream inference, making it a versatile foundation for voice cloning, live chat bots, accessibility tools and more. Kyutai lays the foundation for the multilingual TTS pipeline with its read-out model in English and French.
The enlightenment of real-time AI applications
By lurking speech production into the 200ms range, Kyutai’s model narrows the perceived delay between intent and speech, making it feasible:
- Session AI: Similar to human voice interface, low transition
- Assistive Technology: Faster screen readers and voice feedback systems
- Media production: Dubbing with fast iteration cycles
- Edge devices: Optimize inferences for low power or device environments
The ability to serve 32-bit users on a single L40 GPU without degradation also makes it attractive for efficiently scaling voice services in a cloud environment.
Conclusion: Open, fast and ready to deploy
Kyutai’s streaming TTS release is a milestone in speech AI. With high-quality integration, real-time latency and generous licensing, it addresses the critical needs of researchers and real-world product teams. The reproducibility, multilingual support and scalable performance of the model make it an outstanding alternative to proprietary solutions.
For more details, you can explore the official model card for hugging faces, the technical notes on the Kyutai website, and the implementation details on Github.
Sana Hassan, a consulting intern at Marktechpost and a dual-degree student at IIT Madras, is passionate about applying technology and AI to address real-world challenges. He is very interested in solving practical problems, and he brings a new perspective to the intersection of AI and real-life solutions.
