One more thing from me:
This is the proposal I was most pleased to find, and I think there’s an obvious collaboration here if you’re open to it.
One line jumped out: “live-translate one main-stage talk into six Indian languages”. I’ve spent the last few months building something that does exactly that, live, and I’d like to offer it to you rather than propose anything competing.
What it is
A live captioning system. Speech from a talk is transcribed as it’s spoken, translated into several languages at once, and each subtitle carries the timestamps of the audio it describes — so several languages sit on screen together, in step with the speaker, or each person reads one language on their own phone.
It runs today, on ordinary hardware: sixteen languages from an 8-watt ARM board on my home internet connection, six of them simultaneously in a single session. The software is public domain.
The reason I think it’s interesting for your hub specifically is scale. Live human interpretation is better than anything a machine does — but it doesn’t scale to every talk on every day, and it doesn’t carry over to the recordings afterwards. This does: every session, all four days, plus subtitled archives that are searchable in each language. It doesn’t replace what you’re planning. It could cover the ninety talks nobody has an interpreter for.
The honest part, which matters more
I don’t speak any of your eight languages, and machine translation into Indic languages is genuinely the weak point rather than a detail I’m glossing over.
I’ve benchmarked the freely available general-purpose models, and the English→Hindi output is not good enough to put on a screen. It garbles technical sentences and occasionally invents them outright. The promising direction is IndicTrans2 from AI4Bharat, built specifically for Indic languages and covering all eight of yours — but there’s real work between now and November to get it right.
That’s exactly why I’d rather work with you than at you. The people who should decide whether the output is good enough are the people who can read it. If a caption is wrong, or clumsy, or unintentionally condescending in Marathi, I have no way of knowing. Machine translation without native-speaker review is how you end up with confident nonsense on a screen in front of the very people it was supposed to serve. Your Language Operations function is precisely the thing that would make this safe to attempt.
A smaller, very tractable job
Ethereum vocabulary breaks these models in funny and consistent ways. The word orchestrator comes back as a musical orchestra in five of the languages I tested. “Stream” becomes electricity in German, and a brook in Czech.
The fix is a glossary of terms that shouldn’t be translated, or should be translated one specific way. It’s a one-off piece of work that improves every language at once — and it’s the kind of thing your community would do far better than I can.
What I’m proposing
That we submit this jointly, or at least as two proposals that reference each other. I’ve written the idea up here: Live auto-generated multilingual subtitles for Devcon 8, on infrastructure paid in ETH
You bring the languages, the reviewers, and the standing in those communities. I bring the machine, the infrastructure, and the operators to run it.
Very happy to be told this isn’t useful, or that you’d rather keep it human-only — that’s a legitimate call and I’d understand it. But if the goal is that someone in the audience who thinks in Gujarati can follow a talk given in fast technical English, I have a thing that might help, and I’d rather it were shaped by you than by me.