Google details efforts to expand AI language coverage to 1,000 languages

Google described a shift from text translation to direct audio understanding, arguing that traditional pipelines strip away tone, pacing, emotion, and context. The company says people communicate in messy, real-world ways—such as code-switching or using slang—and that current systems often fail to capture these nuances.

The company highlighted two new models: Gemini 3.5 Live Translate, which supports real-time spoken translation across 70 languages and more than 2,000 language pairs, and Gemini 3.5 Transcribe, described as the company’s most precise speech-to-text model to date. The latter is designed to handle noisy environments and complex jargon, and powers features like Rambler on Android Gboard, which removes filler words, fixes grammar and punctuation, and allows voice-based editing and language switching.

Google reiterated its goal to support the world’s 1,000 most-spoken languages, up from the 300 languages currently covered by its products. To reach this goal, the company is leveraging the Universal Speech Model, which was trained on 12 million hours of audio and uses cross-lingual transfer learning to apply patterns learned from data-rich languages to under-resourced ones.

The company emphasized the role of local partnerships in building datasets for underrepresented languages. One example is WAXAL, an open speech dataset for 27 Sub-Saharan African languages spoken by more than 100 million people across 26 countries. WAXAL was developed with Makerere University and Digital Umuganda and aims to capture tonal variation and conversational speech patterns.