Google Says TranslateGemma Translates 55 Languages On Device
Google published a post about its language AI work, and every figure in it is Google's own. Read it as a company describing its own models and products, not as independent reporting or third party evaluation.
Transcript
Google says its new translation models run on phones without internet. It posted the details today.
Google's post says TranslateGemma is a lightweight open model built from Gemini, trained across fifty five languages and running on device.
Google says its sign language model powers sign to text dictation in Gboard and Live Transcribe on Pixel eleven, first for American Sign Language.
Google says Gemini three point five Live Translate handles real time speech across seventy languages and over two thousand language pairs.
Google published all of this itself, and cites no independent check of the numbers.
Inboxsmith builds AI phone and email agents for small business. Please like and subscribe for more news.
Sources
Every claim in this video comes from the top ranking coverage of this topic. The claims and where each one came from:
- Every narrated claim is drawn from this post. Hook: the post's "Overcoming real-world constraints" section says TranslateGemma runs efficiently on-device, so high quality translation no longer requires a connection to the cloud or the internet; the narration compresses that to running on phones without internet, which is a simplification, since the post says on-device rather than naming phones specifically. Scene 1: the same section says TranslateGemma is a family of lightweight open translation models built from Gemini and trained across 55 languages; the narration says lightweight open model in the singular rather than the post's family of models, and says running on device for the post's runs efficiently on-device. Scene 2: the "Designing for accessibility" section says Sign Language-to-Text (SL2T), trained across 50+ sign languages, powers sign-to-text dictation in Gboard and Live Transcribe on Pixel 11, starting with American Sign Language (ASL) to English; the narration says its sign language model instead of naming SL2T, says first for American Sign Language for the post's starting with American Sign Language to English, and omits both the 50+ sign languages figure and the post's framing of this as a first step toward the 70 million people worldwide who rely on sign language. Scene 3: the "Going from text to true understanding" section says Gemini 3.5 Live Translate powers real-time spoken translation across 70 languages and 2,000+ language pairs, naturally capturing code-switching and emotional cues; the narration says handles real time speech for the post's powers real-time spoken translation, says over two thousand for the post's 2,000+, and omits the code-switching and emotional cues detail. Scene 4 is an editorial honesty note, not an article claim. The post's opening milestone figures (more than 300 languages, more than 7 billion people, 86% of the global population) are deliberately not narrated here, because our channel already published a Short earlier the same day built on that stat from a sibling Manyika post. Honesty note: this is Google writing about Google, a self published description of its own models and products, and the post cites no independent verification of any of these figures; its only external citation is a Forbes article used to support an unrelated statistic about global internet access.(Google's own corporate blog (blog.google), first party post titled "AI for everyone in every language", bylined James Manyika, SVP Research, Labs, Technology and Society, dated Sep 15 2026, read from a local watcher snapshot captured 2026-09-15T16:02:25Z and confirmed against a live fetch on the same day. No YouTube transcripts were available for this story (0 of 5 returned), so this post is the sole source.)