AI for everyone in every language

What Happened

Google has expanded its language coverage to more than 300 languages, with Gemini 3.5 Live Translate supporting real‑time spoken translation across 70 languages and 2,000+ language pairs. The company is training a Universal Speech Model on 12 million hours of audio using cross‑lingual transfer learning to improve speech understanding in under‑resourced languages. Google is also deploying lightweight on‑device models (TranslateGemma) and a voice AI assistant for feature phones (AVA) to reach users with limited or intermittent connectivity.

Why It Matters

Enterprise architects can leverage Google’s multilingual AI to reduce language barriers in global applications, potentially lowering localization costs and improving user engagement. However, integrating these models requires careful consideration of on‑device resource constraints and data privacy, especially when deploying in regions with limited connectivity. Governance around cultural nuance and accessibility must be built into the architecture to ensure compliance and inclusivity.

The Limitation

The performance of these models may vary significantly for low‑resource languages and on older hardware, so extensive testing in target locales is essential before full deployment.

Source