At Google I/O 2025, the company unveiled a major upgrade to Google Meet: real-time speech translation, powered by a new large language audio model developed by Google DeepMind. The feature allows users to carry on seamless conversations in different languages, preserving not just words, but also voice tone and emotional nuance.
Now, instead of relying on text captions or waiting for translations, Meet users can hear instant spoken translations layered naturally over the original audio. You’ll still faintly hear the speaker’s voice beneath the overlay, maintaining a sense of authenticity and presence that Google says is crucial for real conversations.
This marks a meaningful evolution in AI-powered communication, with potential use cases ranging from family conversations across generations—like English-speaking grandchildren talking with Spanish-speaking grandparents—to multinational corporate teams collaborating in real time across continents.
Low Latency, High Fidelity: Breaking Language Barriers in Real Time
According to Google, what sets this system apart is its exceptionally low latency, allowing multiple users to speak and hear translated responses almost instantly. This level of speed and accuracy makes true multilingual group conversations possible in a way that earlier systems simply couldn’t support.
The initial rollout begins in beta this week for consumer AI subscribers, starting with English and Spanish. More languages, including Italian, German, and Portuguese, are expected in the coming weeks. Google is also preparing the feature for its business customers, with Workspace early testing planned for later this year.
By tapping DeepMind’s audio language models, Google is aiming not just to break down language barriers—but to make cross-cultural conversations feel natural, emotional, and human again. Whether it’s a team brainstorm or a family reunion, this new Meet experience brings a bold vision of AI-powered inclusivity closer to reality.
