▲ Gemini 3.5 Live Translate
Google has released a translation model capable of real-time interpretation in over 70 languages.
Google announced on June 9 (local time) that it is fully integrating "Gemini 3.5 Live Translate," a translation model based on its latest artificial intelligence (AI), into the Google Translate app.
Google introduced the new translation model as shifting from a consecutive interpretation method to a "continuous real-time generation" approach that closely resembles simultaneous interpretation.
The model balances the need to translate speech immediately with the need to wait for context to ensure higher quality, according to the company.
As a result, the delay between the original speaker's voice and the translated audio is only a few seconds.
Furthermore, while users previously had to manually select the languages to be translated, the new model automatically detects the languages and translates them accordingly.
It can also be used in multilingual conversation environments where multiple languages are mixed.
The technology for voice generation has also been enhanced, preserving the original speaker's intonation, tone, and pitch to deliver a natural-sounding voice.
Previously, users had to connect earphones to use voice interpretation on the Google Translate app for iPhone and Android; however, they can now listen to the translated audio by holding the smartphone to their ear, just like a phone call, without the need for earphones.
Google plans to apply this feature to its video conferencing platform, "Google Meet," to support multilingual meetings.
Major service companies that have previewed the model, such as Grab and Agora, have also praised the feature.
Baek Hyun-jung, Chief AI Officer (CAIO) at CJ ENM, who participated in the initial testing, evaluated it by saying, "It demonstrates quality outstanding enough to provide a more vivid experience to viewers in Korea and around the world."
(Photo: Provided by Google, Yonhap News)
※ Please note: This article was translated by AI and may contain errors.