Clinical accuracy.
Real-world speed.
The latest speech models, tested on the same medical audio. Compare batch and realtime transcription, including how quickly the text arrives.
1,513 English recordings · 7.2 hours · Speed for every API timed from the US (AWS Ohio), October 2026
Batch transcription
The latest speech APIs alongside Omi’s open model, tested on the same medical audio.
Swipe the table to compare accuracy and speed.
| # | Model | ||||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|
| 1 | Omi Medical v1.1 | 0.87%25 / 2,872 | 5.77% | 0.23%1 / 442 | 96.81%91 / 94 | 3.19%3 / 94 | 98.91% | 97.85% | 131 ms | 222 ms | 200 |
| 2 | Omi Medical-1 Flash | 1.04%30 / 2,872 | 5.98% | 0.68%3 / 442 | 96.81%91 / 94 | 3.19%3 / 94 | 97.85% | 97.33% | —Async only · see note | — | 0 |
| 3 | Voxtral Small Transcribe 3 | 1.04%30 / 2,872 | 5.58% | 0.68%3 / 442 | 84.04%79 / 94 | 15.96%15 / 94 | 91.86% | 87.78% | 742 ms | 1,526 ms | 200 |
| 4 | ElevenLabs Scribe v2 Medical | 1.08%31 / 2,872 | 5.87% | 0.45%2 / 442 | 86.17%81 / 94 | 13.83%13 / 94 | 86.17% | 86.17% | 279 ms | 455 ms | 200 |
| 5 | Google Chirp 3 | 1.11%32 / 2,872 | 6.31% | 1.13%5 / 442 | 81.91%77 / 94 | 18.09%17 / 94 | 90.59% | 86.03% | 790 ms | 1,124 ms | 200 |
| 6 | ElevenLabs Scribe v2 | 1.15%33 / 2,872 | 6.49% | 0.68%3 / 442 | 81.91%77 / 94 | 18.09%17 / 94 | 86.52% | 84.15% | 417 ms | 683 ms | 200 |
| 7 | Microsoft MAI-Transcribe-2 | 1.18%34 / 2,872 | 5.64% | 2.04%9 / 442 | 85.11%80 / 94 | 14.89%14 / 94 | 89.89% | 87.43% | 200 ms | 474 ms | 200 |
| 8 | Speechmatics Enhanced Medical | 1.18%34 / 2,872 | 6.37% | 0.90%4 / 442 | 80.85%76 / 94 | 19.15%18 / 94 | 84.44% | 82.61% | 1,428 ms | 2,124 ms | 200 |
| 9 | Voxtral Mini Transcribe 3 | 1.22%35 / 2,872 | 5.69% | 1.36%6 / 442 | 79.79%75 / 94 | 20.21%19 / 94 | 90.36% | 84.75% | 603 ms | 1,120 ms | 200 |
| 10 | Deepgram Nova-3 Pharma | 1.22%35 / 2,872 | 6.26% | 1.13%5 / 442 | 80.85%76 / 94 | 19.15%18 / 94 | 93.83% | 86.86% | 373 ms | 876 ms | 200 |
| 11 | OpenAI GPT Transcribe | 1.32%38 / 2,872 | 7.23% | 0.68%3 / 442 | 79.79%75 / 94 | 20.21%19 / 94 | 88.24% | 83.80% | 584 ms | 992 ms | 200 |
| 12 | AssemblyAI Universal 3.5 Pro Medical | 1.50%43 / 2,872 | 6.45% | 1.13%5 / 442 | 85.11%80 / 94 | 14.89%14 / 94 | 84.21% | 84.66% | 2,692 ms | 3,876 ms | 200 |
| 13 | Soniox async v5 | 1.57%45 / 2,872 | 6.63% | 2.26%10 / 442 | 82.98%78 / 94 | 17.02%16 / 94 | 84.78% | 83.87% | 2,214 ms | 2,775 ms | 200 |
| 14 | Meta Muse Voice Transcribe 1.0 | 1.64%47 / 2,872 | 6.43% | 1.13%5 / 442 | 85.11%80 / 94 | 14.89%14 / 94 | 93.02% | 88.89% | 1,405 ms | 2,498 ms | 200 |
| 15 | Grok Voice Transcribe 2.0 | 1.74%50 / 2,872 | 6.53% | 1.36%6 / 442 | 76.60%72 / 94 | 23.40%22 / 94 | 92.31% | 83.72% | 323 ms | 663 ms | 200 |
| 16 | omi-medical-edge-1Open weightsEarlier edition · v7Download model ↗ | 2.16%62 / 2,872 | 6.64% | 4.52%20 / 442 | 87.23%82 / 94 | 12.77%12 / 94 | 93.18% | 90.11% | — | — | — |
| 17 | Deepgram Nova-3 Medical | 2.26%65 / 2,872 | 7.09% | 1.36%6 / 442 | 73.40%69 / 94 | 26.60%25 / 94 | 88.46% | 80.23% | 347 ms | 674 ms | 200 |
| 18 | Gemini 3.5 Transcribe | 2.44%70 / 2,872 | 7.89% | 1.36%6 / 442 | 84.04%79 / 94 | 15.96%15 / 94 | 92.94% | 88.27% | 1,260 ms | 1,705 ms | 200 |
| 19 | Corti STT | 2.65%76 / 2,872 | 8.06% | 2.94%13 / 442 | 76.60%72 / 94 | 23.40%22 / 94 | 84.71% | 80.45% | 2,739 ms | 3,382 ms | 200 |
| 20 | Speechmatics Oak-1 | 2.92%84 / 2,872 | 7.61% | 5.88%26 / 442 | 75.53%71 / 94 | 24.47%23 / 94 | 89.87% | 82.08% | 1,371 ms | 1,675 ms | 200 |
| 21 | Amazon Transcribe Medical | 4.28%123 / 2,872 | 6.19% | 2.26%10 / 442 | 79.79%75 / 94 | 20.21%19 / 94 | 90.36% | 84.75% | 1,831 ms | 1,928 ms | 200 |
Speed includes upload, processing and response. 200 short clips, one pass. Speed for every API timed from the US (AWS Ohio), October 2026.
21 modelsOpen-model rows use earlier v7 accuracy results. Dosage scores use the current 94-event panel; comparable API speed is unavailable.
Omi Medical-1 Flash (omi-medical-1-flash) runs English async jobs in the US only, so it is not in the 200-clip speed test. On a single RTX PRO 6000, a 1-hour file is transcribed in a median 3.89 s (one request per file, median of 6 runs). About Flash
Lower is better for error rates and speed. Higher is better for dosage / unit matches. Medical-WER ties are ordered by dose / unit matches, highest first. Counts appear below each clinical metric. p95 is the delay within which 95% of measured responses arrive.
Additional result: AssemblyAI Sync Dictation
Short recordings only (1,465 clips), so this endpoint is separate from the full-dataset ranking. WER 5.31%; medical WER 1.19% (16 / 1,347); drug errors 0.77% (2 / 259); dosage / unit matches 81.82% (63 / 77). Median response 0.344 s; p95 0.567 s.
Realtime transcription
13 models on the same medical audio. Confirmed stream segments.
Swipe the table to compare accuracy and speed.
| # | Model | ||||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|
| 1 | Omi Medical v1.1 | 1.04%30 / 2,872 | 6.17% | 0.68%3 / 442 | 95.74%90 / 94 | 4.26%4 / 94 | 100.00% | 97.83% | 86 ms | 169 ms | 200 |
| 2 | Speechmatics Linden-1 | 1.29%37 / 2,872 | 6.67% | 0.68%3 / 442 | 72.34%68 / 94 | 27.66%26 / 94 | 79.07% | 75.56% | 238 ms | 387 ms | 200 |
| 3 | OpenAI GPT Live Transcribe | 1.32%38 / 2,872 | 6.83% | 0.90%4 / 442 | 82.98%78 / 94 | 17.02%16 / 94 | 87.64% | 85.25% | 580 ms | 704 ms | 200 |
| 4 | Microsoft MAI-Transcribe-2 Streaming | 1.32%38 / 2,872 | 5.73% | 1.81%8 / 442 | 85.11%80 / 94 | 14.89%14 / 94 | 87.91% | 86.49% | 137 ms | 206 ms | 199 |
| 5 | Voxtral Mini Transcribe Realtime 3 | 1.36%39 / 2,872 | 7.44% | 1.36%6 / 442 | 75.53%71 / 94 | 24.47%23 / 94 | 92.21% | 83.04% | 396 ms | 717 ms | 199 |
| 6 | ElevenLabs Scribe v2 Realtime | 1.39%40 / 2,872 | 6.62% | 1.58%7 / 442 | 80.85%76 / 94 | 19.15%18 / 94 | 90.48% | 85.39% | 83 ms | 152 ms | 200 |
| 7 | AssemblyAI Universal 3.6 Pro | 1.39%40 / 2,872 | 6.83% | 0.68%3 / 442 | 82.98%78 / 94 | 17.02%16 / 94 | 84.78% | 83.87% | 87 ms | 152 ms | 200 |
| 8 | Soniox v5 | 1.43%41 / 2,872 | 7.28% | 2.71%12 / 442 | 79.79%75 / 94 | 20.21%19 / 94 | 93.75% | 86.21% | 30 ms | 71 ms | 200 |
| 9 | Meta Muse Voice Transcribe 1.0 | 1.60%46 / 2,872 | 6.10% | 1.58%7 / 442 | 81.91%77 / 94 | 18.09%17 / 94 | 91.67% | 86.52% | 115 ms | 149 ms | 200 |
| 10 | Grok Voice Transcribe 2.0 | 2.19%63 / 2,872 | 7.05% | 2.26%10 / 442 | 79.79%75 / 94 | 20.21%19 / 94 | 90.36% | 84.75% | 145 ms | 222 ms | 200 |
| 11 | Deepgram Nova-3 Pharma | 2.89%83 / 2,872 | 8.31% | 2.94%13 / 442 | 72.34%68 / 94 | 27.66%26 / 94 | 89.47% | 80.00% | 86 ms | 138 ms | 200 |
| 12 | Deepgram Nova-3 Medical | 3.31%95 / 2,872 | 8.42% | 4.07%18 / 442 | 74.47%70 / 94 | 25.53%24 / 94 | 93.33% | 82.84% | 81 ms | 126 ms | 200 |
| 13 | Gemini 3.5 Transcribe Live | 3.34%96 / 2,872 | 8.20% | 3.62%16 / 442 | 79.79%75 / 94 | 20.21%19 / 94 | 93.75% | 86.21% | 286 ms | 424 ms | 200 |
Speed is the time from the end of speech to the final transcript. Each clip is cut at a shared detected end of speech, where every provider gets its finish signal. 200 short clips, one pass. Every API timed from the US (AWS Ohio), October 2026.
13 modelsLower is better for error rates and speed. Higher is better for dosage / unit matches. Medical-WER ties are ordered by dose / unit matches, highest first. Counts appear below each clinical metric. p95 is the delay within which 95% of measured responses arrive.
Same audio. Clear measurements.
Omi-run evaluation. Same recordings and scorer for every model; a fixed subset for speed. Accuracy and speed are measured separately.
What changed in Omi Medical v1.1
v1.1 is the October 2026 release of our medical transcription service. The API model ID stays omi-medical-1. The gains are in live transcription; batch medical-term and drug-name results are unchanged.
| Measure | v1 (September) | v1.1 |
|---|---|---|
| Medical-term errors | 38 / 2,872 | 30 / 2,872 |
| Drug-name errors | 5 / 442 | 3 / 442 |
| Overall WER | 6.55% | 6.17% |
| Dose / unit matches | 82 / 94 | 90 / 94 |
Realtime speed is now measured from a shared end of speech, so we don't compare it with the earlier method.
Safeguards in v1.1
New guards help prevent stray text from digital silence and tiny audio fragments at the end of a turn. Improved speech detection helps recover speech masked by electrical hum; only speech detection is filtered, and the recognizer hears the original audio.
These reduce specific failure modes. Mumbled or heavily masked speech can still be misheard, so review medical transcripts before use.
Dataset and accuracy
1,513 English recordings, 7.2 hours. Every model uses the same audio and scorer. Medical WER covers 2,872 medical words; drug error covers 442 occurrences. Dosage / unit matching uses 94 reference events.
Omi accuracy comes from a US run on 17–18 October 2026 (api.us.omi.health; realtime recordings four streams at a time). Omi Medical-1 Flash accuracy comes from a US production run on 5 November 2026 (async jobs, two at a time). Other providers’ accuracy comes from September 2026; Speechmatics Oak-1 and MAI-Transcribe-2 Streaming were added in October 2026.
Realtime accuracy uses confirmed segments, including the final pieces at Stop. Drafts and Omi’s separate closing-record revisions are excluded. Scores describe the tested configurations, not clinical safety; small differences do not establish statistical significance.
How speed is measured
Where: batch APIs were timed on 18 October and realtime APIs on 19 October 2026 from one AWS client in Ohio (us-east-2), using each provider’s US endpoint where one is offered. Corti was reached through its EU endpoint, the region of our account. MAI-Transcribe-2 Streaming ran on an Azure Sweden Central deployment.
Sample: 200 short clips (<30 seconds), one pass, one request or stream at a time per model; models ran side by side. One warm-up request per model is excluded. Failed attempts are excluded, never replaced by retries.
Batch: request start to usable transcript, including upload and processing. Async endpoints are polled every 500 ms, so smaller differences between them are not meaningful.
Realtime: each clip is cut at a shared detected end of speech and streamed at normal speed. At that point the client sends the provider’s documented finish signal, and the clock runs until the last final text arrives. Accuracy is measured separately on all 1,513 recordings.
Scorer and citation
Scorer on GitHub · Diarization benchmark
Omi Health, Medical Speech-to-Text Benchmark. October 2026, edition 2026-10-18.1 (extends 2026-09-26.1).
Try it on your medical audio.
Start with the API. Compare the results that matter to your application.