transcripts
Why a shared conference-room microphone can't tell who spoke — and what actually fixes it
A hybrid meeting: four people in a conference room around one speakerphone, three people remote. Afterwards the transcript credits half of what was said to "Conference Room B", and the AI summary assigns an action item to someone who wasn't even in the building.
The tool wasn't careless. It was asked to do something it can't reliably do.
One microphone, many voices
A transcript names speakers in one of two ways:
- By connection. Each person's audio arrives on its own device, so the system knows whose words they are because of where they came from. This is reliable, and it's how remote participants get the right name.
- By voice. Several people share one microphone, so the system has to guess who's talking from what their voices sound like. This is called diarization, and it's a guess.
A conference room is case 2 for everyone in it. To the meeting software, the room is one participant. Everything said near that microphone belongs to the room until something tries to split it up by voice, and voice-splitting breaks down in exactly the conditions rooms create: people talking over each other, similar voices, someone far from the mic, someone who only says "yeah, agreed".
What voice enrollment fixes, and what it doesn't
Some room systems can identify in-room speakers by name if each person enrolls a voice profile first. Microsoft documents this for Teams Rooms with Intelligent Speakers. In-room people are identified in the live transcript, but only if they've enrolled and opted in, and were invited to the meeting individually. Anyone without a profile shows up as "Speaker 1, 2 (Room name)", and identification stops working when more than 50 invited participants have enrolled voices.
That's a real improvement for one company's own rooms and regular staff. It doesn't help with the visiting client, the contractor, the new hire who hasn't enrolled yet, or a room without that hardware. Those are exactly the people whose words you'd least like to misattribute.
The fix: give every voice its own microphone
The reliable answer is case 1 for everyone. Each person in the room joins on their own device, so their words arrive on their own connection.
- Everyone joins from their laptop, with laptop speakers muted, so only one device in the room plays sound. Otherwise every laptop picks up every other laptop's audio and you get echo.
- Phones work too. A phone on the table in front of each person is a good, close microphone. Mute its speaker the same way.
- Keep the room's speakerphone for listening only, or turn its microphone off, so the same words don't arrive twice.
The transcript then names each line by the device it came from, not by a guess about a voice.
How Meetful handles a room
Meetful is built on case 1. Every person's microphone is transcribed separately, in their own browser, so the name on a line is the connection it arrived on. When people sit in the same room and a laptop's microphone picks up a neighbour, Meetful drops a line when someone else was clearly louder at that moment, or when the line is mostly a copy of what another person in the call had just said. One person's speech doesn't get credited twice, once to them and once to whoever was sitting next to them.
So the action item lands on the person who said "I'll do it", and the recap can quote them. More on why that matters: wrong speaker attribution in meeting transcripts.
Free, no time limit on calls, nothing to install. Start a meeting.