Book a call
← All work
otobeanWeb app and iOS app built · pre-launch
Meeting minutes from audio · 音豆

Stop spending an hour on the minutes after every meeting.

For managers, teachers and anyone who writes up meetings or lessons

Record the meeting and otobean writes it up — speakers labelled, Cantonese, Mandarin and English understood, ready to send as a PDF.

00:12:48

ALet’s confirm the Q3 targets first.

BOps wants handling time down.

AProposal due next Friday.

4A4 templates
2document styles
9mascot states
The problem

Transcription tools stop at a wall of text. People still spend an hour turning a meeting, lecture or interview into minutes, notes or a report someone else can read.

The product
otobean transcript view with speaker labels and audio player
otobean A4 meeting minutes with review checklist
otobean home: drop audio to transcribe
Drop audio in; recent recordings and documents stay one click away.
otobean Ask across recordings
Ask across recordings — answers link back to the moment it was said.

Screens from the otobean web prototype, with sample data.

Inside the pipeline

A recording goes in. A document you can send comes out.

Five stages, each with an output you can check. Nothing is invented: anything the recording doesn’t say is marked “Not specified”.

Before
  • An hour writing minutes after every meeting
  • Mixed-language speech that tools mis-transcribe
  • Action items buried in a wall of text
With otobean
  • You review and send instead of writing
  • Speakers and languages handled in one transcript
  • Decisions and owners listed, each linked to the moment it was said
01 · CaptureOutput
Weekly operations meeting · 42:16
What you get

Start from any recording — a meeting, a lecture or a phone call.

Audio is normalised and split into segments so long recordings are processed in parallel.

normalisationsegmentationparallel processing
You stay in the loop

Correct the transcript and review the document before anything is shared.

Nothing invented

Unknown owners, dates or venues stay blank — never guessed.

Private by design

PDFs are produced on your own device, not in a cloud browser.

What we built
  • Speaker-labelled, timestamped transcripts you can correct before anything is generated
  • Approved A4 templates — meeting minutes, lecture notes, interview summary, formal report — in a Professional and a Friendly style
  • A structuring layer that fills a validated schema; documents are rendered from approved HTML, never from model output
  • PDF export on the device, so recordings and reports do not pass through a cloud browser
  • Bean, the mascot, appears only at key moments: recording, processing, done, needs attention
Built with

We build the AI layer ourselves, on proven platforms that scale.

AI layer
IN-HOUSEProprietary ASR pipeline · code-switching, diarisationIN-HOUSELLM structuring · schema-constrained output, validate-and-repairIN-HOUSETemplate renderer · approved HTML only
Front end
Nuxt 4VueTypeScript
Mobile
FlutteriOSAndroid (next)
Back end & infrastructure
Cloudflare WorkersD1 · QueuesDrizzle ORM

Developed by P. Chan & Co.

Next case study
mirubean→
Your screenshots, finally sorted — and fact-checked.