Google Expands Its Multilingual AI Tools Online and Offline

Google highlights native audio models, open datasets, and local partnerships to cover hundreds of languages. The company claims a presence in over 300 languages, usage across nine platforms, and offline solutions, while aiming for 1,000 languages and better consideration of cultures and communities.

Mass Usage and Concrete Integrations

Google's language technologies are integrated into its ecosystem, connecting over five billion people through nine platforms, including Search, Android, Chrome, YouTube, and Google Play. The voice assistant Ask Viamo Anything, supported by Google to bring Gemini to basic phones, has already been tested in Rwanda and provided answers to over two million questions. Google is also rolling out sign language dictation to text in Gboard and Live Transcribe on Pixel 11, starting with ASL, in a context where around 70 million people rely on sign language. Additionally, Māori language experts have contributed to improving the pronunciation of place names in Google Maps, with the goal of making the voice synthesis reflect local heritage.

Covering Hundreds of Languages and Billions of Speakers

Google states that its products support daily interactions in over 300 languages, spoken by more than seven billion people, which is 86% of the global population. Launched in 2006, Google Translate has expanded from a small number of languages to over 250, driven by advancements in AI. The company claims to dedicate its research to systems that respect cultural nuances and the richness of expressions, beyond mere text translation.

Functioning Offline with Embedded Models

More than three billion people lack reliable Internet access. For these users, Google has designed TranslateGemma, a family of open and lightweight models derived from Gemini and trained on 55 languages, which operates directly on the device without a cloud connection. The company emphasizes that the hardware requirements of powerful models exclude hundreds of millions of users of basic phones, which motivates the collaboration with partners for solutions tailored to these devices.

Data from the Ground for Underrepresented Languages

To address the predominance of certain languages on the Internet, Google has modified its data collection strategy by relying on local partnerships. WAXAL, developed in collaboration with Makerere University and Digital Umuganda, gathers voice recordings covering 27 languages from Sub-Saharan Africa, used by over 100 million individuals across more than 26 countries, while taking into account tonal differences and conversational rhythms. In India, the Vaani Project, led with IISc and Bhashini, adopts a regional approach and has already collected over 30,000 hours of recordings in 109 languages from more than 155,000 people. The Amplify Initiative has brought together over 1,600 specialists and 20 universities across four continents, including UFMG, IIT Kharagpur, and Makerere, to gather 15,000 multimodal data points. Meanwhile, Language Explorer visualizes LinguaMeta, presented by Google as the largest open-source linguistic repository, mapping over 7,000 languages; the tool has been recognized by Fast Company. Programs supported by Google.org, such as the Center for Digital Language Inclusion and the AI Singapore Aquarium Project, aim for field applications in agriculture, health, and education.

From Text Pipeline to Native Audio with Gemini

Historical approaches to speech recognition, based on a transcription-processing-synthesis sequence, lost information about tone, rhythm, emotion, and context, while real speech includes laughter, hesitations, and code-switching. Google trains models like Gemini to process audio directly. Gemini 3.5 Live Translate offers real-time voice translation in 70 languages and over 2,000 pairs, taking into account code-switching and emotional cues. Google presents Gemini 3.5 Transcribe as its most accurate recognition model to date; it produces formatted text from noisy or specialized audio and serves functions like Rambler on Android Gboard to clean up fillers, correct grammar and punctuation, and rewrite in voice while changing languages.

Accessibility, Local Pronunciation, and Research Trajectory

Google states it designs its products for accessibility with SL2T, trained on over 50 sign languages and deployed for sign language dictation to text, initially on Pixel 11 with ASL, aimed at a community of around 70 million people. Collaboration with Māori experts in New Zealand has helped improve local pronunciations in Google Maps, with authentic pronunciations integrated into voice synthesis. Google aims to support the 1,000 most spoken languages, relying on a Universal Speech Model trained on 12 million hours and on cross-lingual transfer learning, after 25 years of open research and over 400 papers. The company emphasizes an ambition for depth and respect for cultures and plans to continue close work with local communities.