Gemini 3.5 Live Translate: New AI translates language in real time

With Gemini 3.5 Live Translate, Google is launching a new real-time translation for over 70 languages. The AI ​​no longer waits for the end of the sentence, but instead reproduces what is spoken fluently. It even adapts to the original voice pitch.

AI translator for over 70 languages

Google has announced the release of Gemini 3.5 Live Translate. The new model for speech-to-speech translation is now available for the in-house translation app on Android and iOS. At the same time, a test phase begins for the video conferencing service Google Meet and for software developers. In contrast to previous systems, artificial intelligence does not wait for the end of a sentence, but rather translates continuously. This is intended to produce a smooth output without long pauses that is only a few seconds behind the original speaker. The software automatically recognizes over 70 languages ​​and adapts the output to the accent and pitch of the person speaking. One disadvantage, however, is that during continuous processing the context may occasionally be missing, which often only becomes clear at the end of the sentence. Voice translation in Google Meet with Gemini 3.5 Live Translate

New listening mode for smartphones

Like the company in one Blog post writes, there is no need for special hardware. Until now, live translation often only worked in combination with headphones. The update for the translation app now brings a new listening mode. Users can hold the smartphone to their ear like a regular phone call to play the translated audio track directly via the earpiece. This makes it much easier to use when traveling. For now, however, the practical update remains exclusively reserved for the Android mobile operating system, while iOS users still have to connect headphones.

Extension for Google Meet

In the business environment, Google Meet benefits from the new technology. The video conferencing software previously only supported five languages ​​for language translation. The update to the new model increases the number to over 70. This enables more than 2,000 language combinations within one meeting. However, the rollout is currently only taking place as a private preview version for selected Workspace customers. To prevent misuse of the technology, the company integrates an invisible watermark called SynthID into the generated audio signals. The process permanently marks the output as machine-generated and is intended to curb the spread of false information. The watermark cannot currently be removed. This is an advantage for authenticity testing, but a limitation for certain purposes.