SCRIBE RESULTS / 22 SEPTEMBER 2026
Lyrebird
Averages across 5 graded runs of 11 consultations, the same audio and instruction for every scribe. 10 cases (no French).
2m 6sTime to a signable note · ranked 3 of 7
2%Notes ready to sign as written · ranked 6 of 7
9.8Significant note errors per run · ranked 6 of 7
10.2Hallucination flags per run · ranked 7 of 7
16.7%Word error rate · ranked 4 of 7
Summary
Lyrebird ranked 3 of 7 for time to a signable note at 2m 6s and for review time at 1m 46s. Its prompted note was ready 20s after stop, ranked 2 of 6.
Lyrebird averaged 10.2 hallucination flags per run, ranked 7 of 7. 2% of its notes were ready to sign as written, ranked 6 of 7, and this ranged from 0% to 10% between runs. Its editing effort was 22%. Lyrebird was tested on 10 cases, without French.
Every measure
| Measure | Average | Run range | Rank |
|---|---|---|---|
| Time to a signable noteWait from pressing stop to the graded note, plus review time: how long until the note can be signed if it is reviewed straight away. | 2m 6s | 1m 59s–2m 12s | 3 of 7 |
| Notes ready to sign as writtenShare of notes that captured every checklist fact in full with no hallucination flags, so nothing needed editing. | 2% | 0%–10% | 6 of 7 |
| Editing effort0% means no editing required; 100% means writing the full note by hand. Fixes are weighted by the work they take: partial fact 1, missing fact 2, hallucination flag 3, and +2 when a fact is captured wrongly. | 22% | 20%–23% | 6 of 7 |
| Review time per noteEstimated clinician time to proofread and fix one note (190 wpm reading, 5 s to recall and 40 wpm to type each missing or partial fact, 12 s to verify and delete each hallucination flag, including a wrong fact before it is retyped). | 1m 46s | 1m 40s–1m 49s | 3 of 7 |
| Worst note, any runThe single slowest note to review and fix across every case and run. | 3m 7s | — | 4 of 7 |
| Time saved per noteEstimated time to write the note yourself (5 s to recall each checklist fact, typing it at 40 wpm) minus review time. | 2m 39s | 2m 35s–2m 45s | 4 of 7 |
| Significant note errors per runMissed facts, partial facts and hallucinations that a blind AI rater judged could change diagnosis, treatment, follow-up, safety or the legal record, across 11 cases. Significance is a judgement call; the flat count is in Hallucination flags. | 9.8 | 8.0–12.0 | 6 of 7 |
| Hallucination flags per runEvery claim in the note with no support in the consultation, counted flat: each one counts, however minor. No severity judgement. Across 11 cases. | 10.2 | 9.0–11.0 | 7 of 7 |
| Edit actions per noteMissing items, partial items and hallucination flags a clinician would need to fix, per note. | 2.54 | 2.30–2.70 | 6 of 7 |
| Clinical coverageShare of each case checklist captured by the note (partial counts as half). | 89% | 88%–90% | 4 of 7 |
| Significant transcript errors per runMis-heard words that change clinical meaning, rated blind, across 11 cases. | 28.4 | 27.0–30.0 | 4 of 7 |
| Word error ratePooled word error rate per run against the human reference. Every word counts equally, including fillers. | 16.7% | 16.3%–18.0% | 4 of 7 |
| Stop to prompted noteMedian time from pressing stop to the note written to our instruction, per run. | 20s | 19s–22s | 2 of 6 |
| Data per consultAverage network data per consultation, per run. | 19.2 MB | 18.2 MB–19.5 MB | 4 of 7 |
Definitions and weights are in the methodology. Compare Lyrebird head to head ↗
Consultations
CONSULTATION