Meeting Transcription API for Zoom, Google Meet and Microsoft Teams

What is a meeting transcription API?

A meeting transcription API turns a live video call into text your code can use. It does two separate jobs that are easy to confuse.

MeetStream sends a bot into the call as a participant, captures the audio and returns the transcript, on the same meeting bot API.

How the meeting transcription API works

1. POST a meeting link to the API.
2. A bot joins the call. Scheduled bots join on time, and ad hoc bots join in under ten seconds on average.
3. Live transcript chunks are pushed to your webhook while the meeting runs, if you chose a streaming engine.
4. A webhook fires when the call ends and the finished transcript is available to GET.

curl -X POST https://api.meetstream.ai/api/v1/bots/create_bot \
  -H "Authorization: Token $MEETSTREAM_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "meeting_link": "https://meet.google.com/abc-defg-hij",
    "bot_name": "Notetaker",
    "recording_config": { "transcript": { "provider": { "meetstream": {} } } },
    "callback_url": "https://your-app.example.com/meetstream"
  }'

The response is a list of utterances, each tied to a named participant with start and end times. No separate speaker identification step.

API PLAYGROUND

Speaker labels without diarization

Most transcription products identify speakers with speaker diarization: one mixed track, and a model infers where the voice changed.

That holds up in a quiet two-person call. Cross-talk, similar voices and larger rooms all put more weight on the model.

Nylas publicly scopes its transcription accuracy to meetings with up to 10 participants. That is what a diarization ceiling looks like when it is written down.

MeetStream captures each participant's audio separately where the platform allows it, so an utterance is tied to a speaker by the channel it arrived on.

On Zoom that per-participant audio isolation is complete. On Google Meet and Microsoft Teams it follows the active speaker.

Real-time vs post-call meeting transcription

Both come off the same bot. The question is when you need the words.

The real-time transcripts page covers the streaming interface, and voice agents covers agents that speak back.

For a stored record as well, stream during the call, then re-run the recording through a post-call engine.

LIVE CAPTURE

Choose your transcription engine

MeetStream ships its own engine, and lets you route audio to a provider you choose. The figures below are each provider's own published numbers, linked to source.

TRANSCRIPTION PROVIDERS

Two things to plan for. Deepgram's real-time multilingual mode covers ten languages, and its full batch list is wider. JigsawStack streams in English, so pair it with post-call jobs for other languages.

Supported platforms: Zoom, Google Meet and Microsoft Teams

One integration covers all three, and the same code path serves each of them.

Platform detail lives on the Zoom bot API, Google Meet bots and Teams bots pages.

Each platform has a transcription page of its own: the Zoom transcription API, the Google Meet transcription API and the Teams transcription API.

Zoom is the one with a setup step: a Marketplace app once, plus a one-time OBF consent per user. MeetStream handles the tokens after that.

MEETSTREAM DASHBOARD

Why not use native Zoom, Meet or Teams transcripts?

Every platform has some transcript feature of its own, and developers usually try those first. They run into the same four walls.

A bot-based transcription API works on any account, on the plans your customers already have, and returns text while the meeting is still live.

Meeting transcription API pricing

Capture costs $0.35 per bot-hour and transcription $0.10 per hour, so a transcribed meeting is $0.45 per hour all-in, dropping toward $0.25 at volume.

Every account starts with $5 of free credit. Full detail on the pricing page.

Here is how that compares against the rates competitors publish. These are their numbers, from their own pricing pages, linked so you can check them.

How to evaluate meeting transcription accuracy

Most buyers reach for Word Error Rate. WER counts the words an engine got wrong, so lower is better. It treats every word as equally important.

AssemblyAI's Universal-2 release reported a WER gain of roughly 3%, while 73% of people preferred the new output in blind evaluation.

The gains they noticed were 24% on rare words, 21% on alphanumerics and 15% on formatting. Almost none of that shows up in a WER headline.

For meeting transcripts specifically, judge four things.

Run your own audio through more than one engine before choosing. That is the reason MeetStream lets you switch providers rather than locking you to one.

Live and final, on one bot

A streaming engine gives you chunks during the call. A post-call engine gives you the stored record.

Products that want both stream during the call, then re-run the recording through a post-call engine.

Vocabulary tuning passes straight through, so Deepgram keyword lists and AssemblyAI key terms travel with the request.

Chunk delivery rules are on the webhooks page, and the security page covers what applies to transcripts.

Frequently asked questions