Why does my transcript mix up speakers?

The details

• 01Reline captures two things separately: your microphone, and the audio your computer is playing. Those are different signals on your own machine, so the boundary between you and the rest of the call is measured rather than inferred, and your words do not get attributed to the far side by accident.
• 02Inside the far side it genuinely is a mix. Three remote people arrive as one stream, and telling them apart depends on their voices being distinguishable and on them not speaking over each other. That is where transcripts of lively calls degrade, in every product on the market.
• 03The hardest case for all of them is a room of people sharing one laptop microphone. Every voice is roughly the same distance from the same mic in the same acoustics, so speaker separation has very little to hold on to. A microphone in the middle of the table changes the result more than any software setting.
• 04Generic labels appear for a structural reason: audio carries no names. A tool only knows a voice belongs to a particular person if the meeting platform told it, which requires joining the call, or if a human said so. That is why transcripts arrive with numbered speakers until someone puts names to them.
• 05In order of effect, the fixes are: give remote people their own device instead of one shared speakerphone, put a decent microphone in the middle of an in-person table, avoid a Bluetooth headset microphone, and let people finish sentences in the first minutes while the voices are still being established.