# AI transcription on a meeting with five non-native speakers: what the accuracy looked like **URL:** \ **Category:** Experiments & Case Studies\ **Created:** [September 16, 2026, 7:59pm UTC](https://humanizeaiforum.com/t/ai-transcription-on-a-meeting-with-five-non-native-speakers-what-the-accuracy-looked-like/214 "2026-09-16T19:59:07Z")\ **Posts on this page:** 6\ **Page:** 1
**Author:** ![thomas.seeker](https://avatars.discourse-cdn.com/v4/letter/t/b4bc9f/32.png) [@thomas.seeker](https://humanizeaiforum.com/u/thomas.seeker)\ **Post date:** [September 16, 2026, 7:59pm UTC](https://humanizeaiforum.com/t/ai-transcription-on-a-meeting-with-five-non-native-speakers-what-the-accuracy-looked-like/214/1 "2026-09-16T19:59:07Z")
AI transcription on a meeting with five non-native speakers: what the accuracy looked like. ran an experiment on our last all-hands. five out of eight people speaking had English as a second language. different first languages, different accent profiles. one tool, same settings throughout. sharing the results because the accuracy gap was significant and I don’t think most teams doing AI transcription are aware of it. ---
**Author:** ![stormhaze71](https://avatars.discourse-cdn.com/v4/letter/s/e47774/32.png) [@stormhaze71](https://humanizeaiforum.com/u/stormhaze71)\ **Post date:** [September 17, 2026, 11:49pm UTC](https://humanizeaiforum.com/t/ai-transcription-on-a-meeting-with-five-non-native-speakers-what-the-accuracy-looked-like/214/2 "2026-09-17T23:49:07Z")
the accent accuracy gap is documented in the research and the gap is not small. models trained primarily on North American and British English perform measurably worse on other accent profiles. the performance gap compounds with technical vocabulary. ---
**Author:** ![tallest.tower.2016](https://avatars.discourse-cdn.com/v4/letter/t/c89c15/32.png) [@tallest.tower.2016](https://humanizeaiforum.com/u/tallest.tower.2016)\ **Post date:** [September 18, 2026, 7:45pm UTC](https://humanizeaiforum.com/t/ai-transcription-on-a-meeting-with-five-non-native-speakers-what-the-accuracy-looked-like/214/3 "2026-09-18T19:45:09Z")
the fairness implication is the one that should be getting more attention. if the meeting record is less accurate for non-native speakers, their contributions are being underrepresented in the official record. that’s not a neutral technical limitation. ---
**Author:** ![codewrath58](https://avatars.discourse-cdn.com/v4/letter/c/ecccb3/32.png) [@codewrath58](https://humanizeaiforum.com/u/codewrath58)\ **Post date:** [September 19, 2026, 10:42am UTC](https://humanizeaiforum.com/t/ai-transcription-on-a-meeting-with-five-non-native-speakers-what-the-accuracy-looked-like/214/4 "2026-09-19T10:42:07Z")
share the tool and the accuracy rates if you can. this kind of specific data is hard to find. ---
**Author:** ![stormhaze71](https://avatars.discourse-cdn.com/v4/letter/s/e47774/32.png) [@stormhaze71](https://humanizeaiforum.com/u/stormhaze71)\ **Post date:** [September 20, 2026, 3:51pm UTC](https://humanizeaiforum.com/t/ai-transcription-on-a-meeting-with-five-non-native-speakers-what-the-accuracy-looked-like/214/5 "2026-09-20T15:51:30Z")
the tool was a common meeting transcription service, not naming it specifically because I’m not sure the results are representative of their current model. accuracy by speaker (rough estimate based on manual comparison with video): native English speakers: around 94-96% word accuracy. non-native speakers with lower accent variation (European first languages): 85-90%. non-native speakers with higher accent variation (Asian first languages, Arabic): 70-78%. one speaker with a regional accent and technical vocabulary: 65%. the technical vocabulary problem was visible across all speakers but worse for the lower-accuracy group. product names, methodology terms, and proper nouns were the highest-error category. the consent issue came up separately. three people had not been told the meeting was being transcribed. we’ve since changed the process to explicit opt-in notification at meeting start. ---
**Author:** ![thomas.seeker](https://avatars.discourse-cdn.com/v4/letter/t/b4bc9f/32.png) [@thomas.seeker](https://humanizeaiforum.com/u/thomas.seeker)\ **Post date:** [September 21, 2026, 5:56am UTC](https://humanizeaiforum.com/t/ai-transcription-on-a-meeting-with-five-non-native-speakers-what-the-accuracy-looked-like/214/6 "2026-09-21T05:56:29Z")
the consent point is the one most teams skip because they announce the transcription policy in a general email and then assume everyone has read and consented. explicit notification at each meeting is a different standard and a higher one.