How we measure

Public recordings with human-made transcripts. We do not use customer recordings.

• Read speech, in many languages
• Meeting recordings, with close and room microphones
• Mixed-language conversations

Live results use pieces of audio of up to 12 seconds, as in a live meeting. In every comparison, KaigiFlow and the comparison are given exactly the same audio.

The error rate is the share of the spoken words (the spoken characters, for Cantonese) that came out wrong. Lower is better. Missed words are words that were spoken but never appeared on screen. Times fewer is the error rate of standard cloud speech recognition divided by ours.

The numbers describe those tests, not every meeting. Where we are only level with standard cloud speech recognition, we make no claim.

Standard cloud speech recognition is a widely used general-purpose cloud service, tested by us on the same audio pieces.

The only comparison with other products is Cantonese, on a public Cantonese benchmark that the other services also publish results on: our live result against their published offline result. We show those products as Peer A to D and do not name them. For other languages, published figures are not comparable with ours, so we make no comparison.

Our own tests ran between 15 and 18 October 2026. Published peer figures were read on 19 October 2026.