5 things to know about Kyutai, Xavier Niel's French lab which wants to revolutionize voice assistants using AI

© Kyutai

The French AI sector is booming. In just eight months of existence, the Kyutai laboratory has reached a major milestone by unveiling Moshi. This generative artificial intelligence (AI) voice assistant, entirely designed in France, already rivals OpenAI’s GPT-4o.

Based in the Marais in Paris, Kyutai has received support of 300 million euros from Xavier Niel, former Google CEO Eric Schmidt and Rodolphe Saadé, head of CMA-CGM. The laboratory has around fifteen researchers, who have been working tirelessly since November.

Moshi is the first prototype of Kyutai

During an event organized on July 3, they presented the first fruit of their work. The Moshi AI, which required 10 million euros to design, is capable of communicating vocally with its interlocutors. She can answer their questions as well as transmit information or advise them.

It is based on a large language model (LLM) called Helium. Comprising 7 billion parameters, it is multimodal, that is to say it is based on audio and textual data.

Kyutai used data synthetics to train Moshi

The Kyutai team claims to have used data created from scratch to train its AI, reports Les Échos. A cunning strategy when we observe the numerous conflicts with the law of artificial intelligence publishers. Often, they train their models on copyrighted content, exposing themselves to heavy financial penalties.

On the audio side, the AI ​​has been trained from recordings of calls made between 1994 and 2002 in the United States.

200% Deposit Bonus up to €3,000 180% First Deposit Bonus up to $20,000

Moshi « thinks while speaking »

Kyutai AI already outperforms GPT-4o on one point: its response time. According to the laboratory's researchers, Moshi records a latency ranging from 3 to 5 milliseconds. This is much better than the 160 milliseconds found on average on the market, they argued. Its prediction techniques are so advanced that the AI ​​does not hesitate to cut off its interlocutors.

For the moment, it only speaks in English , a limitation that Kyutai wishes to remedy in the future.

Kyutai called on an artist for Moshi's voice

Beyond the speed of his responses, it was especially Moshi's voice that impressed the audience. An artist named Alice collaborated with the laboratory, and performed various scenarios for 20 hours in order to produce as many intonations as possible.

As a result, Moshi can adapt to the user's manner of speaking, and is capable of imitating up to 70 different emotions. She is also able to whisper and recognize the emotions of her interlocutor.

Kyutai does not intend to stop there

In addition to learning other languages ​​and particularly French, Kyutai intends to run its model on small devices, such as a smartphone or a computer. This would allow it to become a voice assistant that directly competes with those of technological giants, such as Siri or Alexa. Currently, it is the Scaleway cloud that powers the AI.

The time is above all for research, says Kyutai, with monetization not being one of its priorities at the moment. In addition to voice AI, video generation could also enter into the group's plans.

📍 To not miss any news from Presse-citron, follow us on Google News and WhatsApp.

[ ]