A Grab driver who speaks mostly Malay and a tourist who speaks only Mandarin have always needed a shared third language, or a lot of pointing. Google's newest AI is trying to erase that gap, and one of the first companies testing it runs millions of rides a month in Malaysia.
On 9 June 2026, Google introduced Gemini 3.5 Live Translate, an audio model that converts speech to speech in more than 70 languages in near real time. According to Google, it detects the spoken language automatically and keeps the speaker's tone, pitch and pacing, instead of the flat, robotic voice older tools produced. Ars Technica reports that the model runs continuously rather than waiting for each person to finish a sentence, so it trails a live conversation by only a few seconds.
Editor
Kai T chevron_right
Table of Contents
Where you can actually use it
Google is putting the model in three places. Anyone can get it through the Google Translate app on Android and iOS, where connecting a pair of headphones gives you a running translation that mirrors the other person's voice. Android phones add a "listening mode" that pipes the translated audio through the earpiece, so you can hold the phone to your ear like an ordinary call without headphones. Developers can build on it through the Gemini Live API and Google AI Studio in public preview, while Google Meet gets it first for enterprise customers, jumping from 5 supported languages to more than 70.
Why this lands differently in Malaysia
Malaysia is one of the most linguistically dense countries on earth, with more than 130 living languages and a population that routinely folds Malay, Mandarin, English and a home dialect into a single sentence. Government primary schools teach in Malay, Mandarin and Tamil, and Tamil alone has well over a million speakers here. A translator that switches between languages without being told which one is being spoken fits the way Malaysians already talk.
The clearest signal for local readers is who Google named as an early tester. Grab, the Southeast Asian super-app that dominates e-hailing in Malaysia, is trialling the model so drivers and passengers can understand each other at pickup. Google says Grab users already make more than 10 million voice calls a month. If that test becomes a shipping feature, the daily friction between a driver and a tourist, or a driver and a customer who speak different first languages, could shrink to a single phone held between them.
The part worth watching
Two caveats matter. First, how useful this is for Malaysia depends on which languages and accents the voice model handles well. Google has named a 70+ language count but not how cleanly it manages Malaysian code-switching or regional tongues such as Hokkien and Cantonese, which even text translation has handled unevenly. Second, every clip the model generates carries a SynthID watermark, Google's invisible marker for AI-made audio. That matters as voice-cloning scams climb, a threat Malaysian authorities have flagged repeatedly.
The timing also draws a contrast. Apple spent its WWDC keynote this week on a rebuilt, Gemini-assisted Siri that is launching English-first, with Bahasa Melayu still unsupported. Google, in the same week, dropped a translation model straight into a free app that anyone with an Android phone can open today.
Live translation has been promised for years and mostly underwhelmed. The test that counts now is not the demo on a Google stage. It is whether your next Grab driver can understand you without either of you having to switch languages first.
Images courtesy of Utsman Media and Vitaly Gariev on Unsplash.