The French start-up Kyutai has just unveiled Hibiki, an AI capable of translating live from the oral conversations from French to English, as an interpreter. A tool still in development, but very promising.
While Paris welcomes the first summit dedicated to artificial intelligence, the announcements are increasing, especially from French companies. Thus, after the arrival of the mobile version of Le Chat, the Conversational Robot of Mistral AI, on iOS and Android, a research laboratory in AI has just unveiled an experimental model of real -time translation.
Called Hibiki (“echo” in Japanese), this tool makes it possible to translate the words of a person in real time from French to English, like a human interpreter. It is developed by Kyuta, a French lab created by Xavier Niel, owner of the Iliad group and founder of Free, Rodolphe Saadé, CEO of the maritime carrier CMA CGM, and Eric Schmidt, ex-boss of Google.
Hibiki: like a human interpreter
In a demonstration videowe see how Hibiki ensures the French-English translation. The extract features a French interlocutor who explains the functioning of the AI and the tool translated simultaneously in the language of Shakespeare. The translation is done with a few seconds of lag like a translation during an interview or a press conference for example.
Meet Hibiki, Our Simultaneous Speech-to-Speech Translation Model, Currently Supporting.
Hibiki Produces Spoken and Text Translations of the Input Speech in real-time, while Preserving the speakers Voice and Optimally Adapting Its Paced on the Semantic Content of the pic.twitter.com/jomccc5vji– Kyutai (@kyutai_labs) FEBRUARY 6, 2025
As specified in press release laboratory: “Hibiki allows simultaneous translation while preserving the speaker’s voice and optimally adapting its rhythm to the semantic content of source discourse. Hibiki thus produces in real time an oral translation, as well as written, in the target language”.
In view of the demonstration, technology already seems well advanced when it is still in development phase. This is why, the hibiki model is available in open-source. Its code is therefore made accessible to the public, so that researchers will grasp it and make their changes, moreover, in its current version, the tool can be used on a phone without the need for Internet connection.
In addition, the director of modeling at Kyuta underlines that the technology has been trained not to say anything: “You cannot control it by text to make him tell what you want”. Finally, he claims to want to make his innovation “A model that can be used when traveling or when videoconferences or in case of streaming”. And he specifies that Hibiki is already capable of providing live subtitles.