Measured, not claimed.
Every number on this site comes from this page. We use a public meeting recording, so anyone can repeat the test and check our work.
Last run 25 October 2026 · Jiva 0.1.16 · MacBook Pro, M4 Pro, 24 GBResults
| RIGHT SPEAKER ↑ | WORDS WRONG ↓ | SPEAKER LABELS | WORDS IN TRANSCRIPT | WHERE IT RUNS | |
|---|---|---|---|---|---|
| The real meeting (human transcript) | 100% | 0% | 4 people | 2,503 | — |
| Jiva | 90% | 21% | 6 | 2,129 | On your Mac |
| A popular AI notetaker | 43% | 17% | 1 ("Them") | 2,213 | Cloud |
One meeting so far (AMI ES2004a, 17.5 minutes, 4 people). "Right speaker" is the share of words given to the right person. The other app labels everyone on the call as "Them", so only the person who talks most can count as right. Jiva showed 6 labels for 4 people: it split two voices into extra labels.
How we test
- 1A public meetingES2004a from the AMI corpus (CC BY 4.0): a real 17.5-minute, four-person design meeting with a human-written transcript of 2,503 words and speaker labels.
- 2Same audio, same timeThe recording plays on one Mac while Jiva and a popular AI notetaker both record it, so both hear exactly the same thing.
- 3Scored against the human transcriptWords wrong, words given to the right person, and how many speaker labels each app shows. Same rules for both: lower case, no punctuation, filler words like "um" removed.
- 4Repeat it yourselfThe meeting and its human transcript are free to download, so anyone can repeat the test with any app.
Speed on your Mac
| MEETING LENGTH | MACBOOK PRO, M4 PRO, 24 GB |
|---|---|
| 17.5 min | Transcript and summary ready about 1 minute after the meeting ends |
Where Jiva is still behind
We list these because they're true, and because a page that only shows wins isn't worth trusting.
- Words: Jiva gets 21% of words wrong to the other app's 17%. On the words both apps heard, Jiva made fewer mistakes; it loses on quick back-and-forth, where it misses more short phrases.
- Extra labels: Jiva showed 6 speaker labels for 4 people, so two voices were split in two.
- So far this is one meeting. We'll add more meetings and report the average.
Try Jiva free.
Download for Maccurl -fsSL https://jiva.works/install.sh | bash