Live translation in 20 languages

Ten years ago, live translation meant hiring a simultaneous interpreter. Five years ago, auto-captions started to work for English but sucked everywhere else. Today, live translation to 20+ languages runs on any Chrome sidebar for the price of a small API call. Here's what that unlocks.

What "20 languages" actually means

Most AI speech-to-text models handle a long tail of languages, but only well for the top dozen. TA pilot's practical list, ordered by transcription quality:

• English, Spanish, French, German, Italian, Portuguese — near-perfect
• Japanese, Korean, Traditional and Simplified Chinese — very good
• Russian, Polish, Dutch, Turkish, Ukrainian — good
• Arabic, Hindi, Thai, Vietnamese, Indonesian — good enough for casual,

not for legal/medical

The model behind it is Gemini 2.5 Flash, proxied through OpenRouter. Quality improves quarter-over-quarter as the underlying models do.

Where it helps most

International conferences

The conference is in English but 30% of the audience is non-native. Live captions in their language on their phone (different from the projected language on the big screen) means they can follow at full speed.

University lectures with international students

Same pattern, higher stakes. A student who misses 20% of the lecture because of language fails courses. Live captions in their L1 close the gap.

Recorded talk reach

A talk captioned live is a talk ready for YouTube upload the same day. Captioned videos get indexed, indexed videos get discovered, and discovered videos get watched. You're not just helping the live audience — you're helping future viewers.

Where it doesn't help (yet)

• Legal proceedings. Nuance matters. Use a human interpreter.
• Medical consultations. Liability matters. Use a human interpreter.
• Technical terminology domains. Cryptography, biology, specialized

law — the models get 85% right, which means 15% wrong in crucial ways.

For general business and educational talks, AI live translation is already better than nothing and often better than no interpreter at all.

Cost

The per-minute cost of live transcription + translation is a fraction of a penny per attendee. It's effectively free. The bottleneck is no longer cost — it's awareness that the feature exists and setup simplicity.

The setup step that matters

Pick the language at session creation. TA pilot asks for the language when you create a new session and uses it as the transcription hint for the rest of the talk. That single setting dramatically improves transcription accuracy vs. "auto-detect" which confuses proper nouns with French words and breaks mid-phrase.

If you're running a talk with international attendees, create the session in the primary language of the speaker, and let attendees pick their own caption language on their participant page.