Hey everyone — Kai here, one of the builders on Subanana. Adding a bit of the "why" behind a few decisions Aric mentioned:
Why we route per language. Our first product taught us that no single speech model is best at everything. Great English accuracy often turns into mangled Cantonese or messy code-switching. So Live Captions uses the same routing as our subtitle tool: each language goes to the model that handles it best, not whichever vendor we happen to be locked into.
Why the captions are a moment late. This one was a debate for us. Raw real-time output flickers, rewrites itself, and breaks mid-sentence, which is tiring to read on a phone. We chose a short delay so we can clean up the text before anyone sees it. A slightly later caption you can trust beats an instant one you have to re-read.
Why no meeting bot. It listens to your mic or your computer's audio, so it works the same on a conference stage, in a church, a classroom, or next to a Zoom call. You don't need to invite anything into the meeting.
Built for the production desk too. Beyond the QR code on phones, you can run up to 5 languages at once, put 2 on the big screen, float a subtitle box over your slides, or send a transparent caption layer into OBS, vMix, or ProPresenter.
Every plan, including free, gets a 5-minute live session, so you can try it on a real meeting today.
I'm hanging around all day. Happy to answer technical questions, and I'd love to hear which language pairs you'd want us to stress-test next. 🙏
PH 用户
@lakshya_singh @aric_fung @connect_kai @kevin_wong21 Good launch, congrats!
We have a constant problem with Armenian-language transcription =)
I’ll definitely give it a try!
PH 用户
Congrats on the launch, been on the beta and it clicked at a demo meeting last week:
Hong Kong client, local team + two people over from the Shenzhen office, and the client went "don't mind them, carry on in Cantonese", which is always awkward. Put the QR up, they read it in Simplified on their phones, done. I kept the Cantonese line up next to the translation and it actually follows how people talk here, half Cantonese half English, instead of rewriting it into written Chinese or mangling the English words like every other tool I tried.
Having worked on AI products myself I know how little Cantonese data there is to build on, so genuinely impressed you pulled this off.
PH 用户
Wow Aric! It's truly awesome cause online events usually gets viewers from different countries and it makes them more accesible and friendly. Wish you all the best here
PH 用户
This is exactly the product that I've been dreaming for. This is especially going to be useful for Georgian language, as transcription for Georgian content is usually problematic.
PH 用户
从没想过自己会需要这种东西，但现在 Live captions 感觉已经是任何活动都必不可少的配置了。恭喜 Subanana 团队发布！
Why we route per language. Our first product taught us that no single speech model is best at everything. Great English accuracy often turns into mangled Cantonese or messy code-switching. So Live Captions uses the same routing as our subtitle tool: each language goes to the model that handles it best, not whichever vendor we happen to be locked into.
Why the captions are a moment late. This one was a debate for us. Raw real-time output flickers, rewrites itself, and breaks mid-sentence, which is tiring to read on a phone. We chose a short delay so we can clean up the text before anyone sees it. A slightly later caption you can trust beats an instant one you have to re-read.
Why no meeting bot. It listens to your mic or your computer's audio, so it works the same on a conference stage, in a church, a classroom, or next to a Zoom call. You don't need to invite anything into the meeting.
Built for the production desk too. Beyond the QR code on phones, you can run up to 5 languages at once, put 2 on the big screen, float a subtitle box over your slides, or send a transparent caption layer into OBS, vMix, or ProPresenter.
Every plan, including free, gets a 5-minute live session, so you can try it on a real meeting today.
I'm hanging around all day. Happy to answer technical questions, and I'd love to hear which language pairs you'd want us to stress-test next. 🙏
We have a constant problem with Armenian-language transcription =)
I’ll definitely give it a try!
Hong Kong client, local team + two people over from the Shenzhen office, and the client went "don't mind them, carry on in Cantonese", which is always awkward.
Put the QR up, they read it in Simplified on their phones, done.
I kept the Cantonese line up next to the translation and it actually follows how people talk here, half Cantonese half English, instead of rewriting it into written Chinese or mangling the English words like every other tool I tried.
Having worked on AI products myself I know how little Cantonese data there is to build on, so genuinely impressed you pulled this off.