How much AI meeting captions mishear people with your gender
When a video-call transcript becomes the record of what you said, your voice may not be treated like everyone else's.
By Maria Lee·
In 2024, AI speech recognition put female speakers at 5.16 word error rate and male speakers at 13.3. The figures come from an ISCA/Interspeech 2024 paper on a fairness dataset of U.S. voice recordings. But gender was only one split. In calls, interviews, reviews and pay conversations, accent, language status and region can also shape what the transcript says you said.
Male speakers hit 13.3 on Whisper-small
The Interspeech 2024 paper said its dataset included approximately 26.5K utterances in recorded speech by 593 people in the United States. Those people were paid to record and submit audio of themselves saying voice commands, with self-reported age, gender, ethnicity, geographic location, native-English status and first language.
Its category table listed 321 female speakers with 14,422 utterances and 272 male speakers with 12,049 utterances. In the evaluation table, the female line ran 6.76, 3.82, 5.16 and 3.86 across the reported systems, while the male line ran 12.06, 5.75, 13.3 and 7.91.
A clean transcript can outlast the voice
A transcript looks official because it is neat, time-stamped and searchable. In a performance review, a hiring panel debrief or a pay discussion, the clean page can outlive the actual voice. A missed word can change how direct, prepared or confident someone sounded.
If your job depends on being understood when you speak, the basics of why public speaking is important in the workplace now include how your words survive captions, summaries and searchable transcripts.
Second-language English was penalized in another test
A separate ACM/FAccT 2022 conference paper examined the Speech Accent Archive, which it described as a database of short English language recordings. Its Table 1 said word error rate was significantly higher, with p<0.05, for second-language speakers of English than first-language speakers for both ASR systems tested.
For Amazon Transcribe, the paper reported a reference first-language English baseline of 19.58 and an English L2 coefficient of 5.14. For Google Speech-to-Text, it reported a baseline of 26.55 and an English L2 coefficient of 7.91.
Belfast was singled out against London
The ACM/FAccT 2022 paper also looked at geography inside British English. It said that, compared with speakers from London, Google performed significantly worse for speakers from Belfast, again at p<0.05. That is a narrower finding than the L1 and L2 result, but it points to the same practical problem.
The voice on a work call is not just words in a vacuum. It carries accent, rhythm and local pronunciation. A person can speak clearly to coworkers and still be mistranscribed by a system trained and tested unevenly across speaker groups.
Why the answer changes by model
The two papers do not make one simple ranking of who is always misheard most. The Interspeech 2024 fairness table has male speakers higher than female speakers on the Whisper-small line. Its native-English cut, however, lists L1 at 9.54 and L2 at 5.63 for the same column.
Geography and income also moved the rate in that 2024 table. In the listed column, low income, shown as US <= $50k, was 7.69. Medium income, shown as US = $50k-$100k, was 10.94. Affluent, shown as US = $100k+, was 5.55.
Every group the caption model split out
The people-like-you cut is limited to the groups reported in the public table. The gender cut lists female and male speakers only. The linguistic cut uses the source labels L1 and L2, while the geography-related cut is reported through income bands.
| Cut | Group | Word error rate (WER) |
|---|---|---|
| Gender | Female | 5.16 |
| Gender | Male | 13.3 |
| Geographic variation | Low (US <= $50k) | 7.69 |
| Geographic variation | Medium (US = $50k-$100k) | 10.94 |
| Geographic variation | Affluent (US = $100k+) | 5.55 |
| Linguistic variation | L1 | 9.54 |
| Linguistic variation | L2 | 5.63 |
Word error rate in the Interspeech 2024 fairness table, using the column that includes 5.16 for female speakers and 13.3 for male speakers
A worked example makes the risk easier to picture. Take a pay-review answer where the important words are your role, the project, the amount of responsibility and the ask. A transcript following a 13.3 error rate gives more room for damage than one following 5.16, even when both speakers were clear to human listeners.
What this means for you
If captions mangle your words, do not treat the transcript as neutral proof that you spoke badly. The public numbers show group-level differences across gender, accent, language status and other cuts. The machine may be adding noise to a sentence you said clearly.
Before a high-stakes call, slow down on names, numbers and the exact sentence that carries your ask. After the call, scan the transcript for the few lines that matter. If it got them wrong, correct the record in writing while the conversation is still fresh.
Sources