Google's Live Translate gambit raises the stakes for India's voice AI

The News

Google has switched on Gemini 3.5 Live Translate, an audio model that turns spoken words into spoken words across languages with barely a beat of delay. The launch was detailed on 9 June 2026 by product manager Anuda Weerasinghe and senior staff software engineer Tony Lu, and it lands inside several products at once.

The model automatically detects more than 70 languages and handles over 2,000 language combinations. Crucially, it does not just convert text; it preserves a speaker's intonation, pacing and pitch, so the translated voice sounds like the original person rather than a flat synthetic read-out. Instead of waiting for a sentence to finish, it generates translated speech continuously, staying in sync with only a few seconds of lag.

Developers get it first through Google AI Studio and the Gemini Live API in public preview. Google Meet receives a private preview for select business customers this month, with a wider rollout planned later in 2026, while the consumer Google Translate app is rolling out globally on Android and iOS, including a new "listening mode" that pipes translations into an earpiece. Every clip carries a SynthID watermark to flag it as machine-made. Ride-hailing group Grab is already testing the system across its 10-plus million monthly voice calls, with broadcaster CJ ENM and infrastructure firm LiveKit reporting strong accuracy and low latency.

Why It Matters

Live, voice-preserving translation has been one of the most over-promised features in consumer technology. Meta shipped its SeamlessM4T speech model in August 2023 with similar ambitions, but real conversations exposed awkward pauses and robotic output. Google's pitch here is that the lag is now short enough, and the voice natural enough, for the technology to fade into the background of an actual conversation.

That matters because translation is shifting from a standalone app you open to a layer baked into calls, meetings and earbuds. Whoever owns that layer owns a chunk of how the next billion internet users communicate. By folding the model into Translate, Meet and a developer API on the same day, Google is trying to set the default before rivals at OpenAI and Apple can respond.

Indian Angle

No market tests this promise harder than India. The country recognises 22 scheduled languages and hundreds of dialects, and the central government has poured resources into Bhashini, the MeitY-backed translation mission designed to deliver public services in local tongues. A polished, low-latency Google model raises the bar that home-grown efforts such as Sarvam AI and Krutrim are measured against, and it sharpens the question of how well 70-plus languages actually cover Hindi, Tamil, Telugu, Bengali and Marathi in practice.

For Indian businesses, the Grab use case is the tell. Domestic gig platforms, call centres and quick-commerce operators field millions of cross-language voice interactions daily, and an API that handles them in near real time could trim the cost of multilingual support staff. Indian developers can begin building on the Gemini Live API now, though they will be watching usage pricing closely, since voice tokens billed in dollars add up fast at Indian scale.

There is a regulatory wrinkle too. Routing live voice through Google's cloud means personal speech data leaves the device, which puts the feature squarely inside the Digital Personal Data Protection Act's consent and localisation debates. Enterprises in banking and healthcare, where conversations are sensitive, will want clarity on where that audio is processed before they switch it on.

FAQ

When can I use it?

Developers can access it today via Google AI Studio and the Gemini Live API in public preview. The Google Translate app rollout is underway on Android and iOS, while Google Meet support starts as a private preview for select business customers this month, expanding later in 2026.

How does it compare to Bhashini?

Bhashini is India's public, government-built translation stack aimed at citizen services and Indian languages specifically. Gemini 3.5 Live Translate is a global commercial model covering 70-plus languages. They will overlap on Indian-language use cases, but Bhashini's edge is local-language depth and public-sector integration.

What does it mean for Indian developers?

They get a production-grade speech-to-speech API without training their own model, useful for support bots, ride-hailing and meetings. The trade-off is dependency on Google's pricing and cloud, plus data-residency questions under Indian law.

Where can I read the original announcement?

Google published full details on its official blog, linked in the attribution below.

This story was reported by Google DeepMind. Read the full original coverage at Google DeepMind.