I stopped missing half my standups — notes on live bilingual captions without a meeting bot

I work in a company where the meeting language is not my native one.

Standups are fine on a good day. Design reviews are worse: three people talking over each other, acronyms I do not know yet, and someone saying “ship it” while I am still parsing the previous sentence. Native meeting captions help sometimes. They also fail in the ways that matter when you are already behind:

1. They are often host-gated — if the organizer never turns them on, you get nothing.
2. They are usually one language at a time — I want the original and a translation side by side when a term is ambiguous.
3. They do not help me take a note without leaving the conversation — by the time I open a doc, the thread has moved on.
4. The “AI notetaker” path often means a bot joins the call — extra participant, recording notice, IT policy questions, and a little social tax every time.

So I stopped optimizing for “better post-call transcript” and started optimizing for staying present while the call is still happening.

The constraint that changed the product shape

Offline meeting tools optimize for a clean recording and a polished summary afterward.

Live tools optimize for a different set:

1. Work on Google Meet, Zoom, and Microsoft Teams without waiting for each platform’s feature flag.
2. Latency low enough that captions feel attached to the voice — for me, roughly a second is the line between aid and distraction.
3. Original + translation available together when I want learning mode, not only a single target language.
4. Notes tied to the same timeline as the captions, so “that sentence about Redis” is clickable later.
5. Prefer a participant-side capture over inviting another identity into the room.

That last one is the fork in the road. Bot-join tools are easy to demo. They are also the reason some teams never install them.

Why "no bot" is not just a privacy slogan

When a bot joins, three things happen at once:

• Everyone sees an extra participant (or a recording banner).
• Policy and consent become a meeting-level decision, not a personal tool choice.
• You inherit every platform’s bot permissions model.

Participant-side captions flip that: you install something that listens on your machine or in your browser tab. Other people do not get a new invitee. You still need to be honest about what is transcribed for you, but you avoid the social and IT blast radius of a room-wide bot.

That choice has engineering consequences. Browser tab capture covers Meet and web Zoom/Teams well. The Zoom or Teams desktop client does not live in a tab — so a Windows desktop listener became the twin path for “every meeting on this PC,” same account, same minutes pool.

Instant vs accurate translation is a real UX split

Early on I treated “translation quality” as one knob.

In a live meeting it is two jobs:

• Instant: keep latency ultra-low for back-and-forth. You accept that a half-finished clause might look rough.
• Accurate: wait for a fuller sentence, spend a bit more model time, get better technical terms and phrasing.

I switch between them in the same call. Standup banter wants Instant. A design review full of service names wants Accurate. Shipping only one mode means half your meetings feel wrong.

Captions that become notes beat a second app

The pause-loop for meetings is not scrubbing a video. It is:

hear something important → panic → open Notion → lose the next two sentences → never find the quote later.

Tapping a caption line into a note, with the transcript kept after the call, closed that loop better than “remember to write it down.” Asking the meeting a question afterward (“what did we decide about caching?”) only works if the live capture was good enough to build a timeline on.

Summaries are useful. They are not a substitute for being able to read the line that was said at 00:17.

What I would tell myself before building

1. Solve missing the live conversation, not “AI meeting notes” as a category.
2. Treat bot vs participant-side as a product decision, not an implementation detail.
3. Ship bilingual as a first-class mode — original-only and translation-only are modes, not the whole product.
4. Accept that browser + desktop are two listeners for one account if you care about Zoom/Teams native apps.
5. Auto-stop when the room goes quiet. Forgetting a tab open should not silently burn your monthly minutes.

The thing I actually use

I packaged this workflow as SonicMeet — Chrome/Edge extension, optional Windows desktop app, freemium meeting hours.

I am not doing a pricing pitch here. The interesting part for me was treating live meeting understanding as a latency + bilingual layout + no-bot problem, not only an ASR accuracy race.

Open questions

• When do you keep Instant mode on by default vs nudge people toward Accurate for technical calls?
• How do you explain desktop audio capture without sounding like spyware?
• For bilingual captions on a laptop: original on top, translation on top, or a user preference you change mid-call?

If you have built realtime speech tools for meetings — or you have your own ritual for surviving standups in a second language — I would love to hear what you landed on in the comments.

Built this because I needed it. Feedback: [e-mail]