Model reference — AI Medical · Audio
On clinical speech, accuracy is decided by vocabulary and recording conditions — not by the headline error rate on read speech.
Most deployments combine three parts. A transcription model produces the words. A diarization model marks who spoke when, which is what makes a consultation transcript readable. Where text has to line up with the audio for review, a forced-alignment step is added. Sizing then depends on whether you are keeping pace with a live stream or working through a recorded archive.
Every model on this page runs as part of a managed AI pipeline in our GPU clusters, with a dedicated private cluster in our cloud or an on-premise installation where medical governance requires it. Output is decision support for a qualified professional to review, not a diagnosis.
Each specification table gives three hardware tiers — Minimum, the smallest setup on which the model runs correctly; Medium, the usual production configuration; and High, a configuration sized for peak volume. The rate is what one server of that tier processes per hour. Use these figures for initial sizing only. Before production we benchmark your own data to confirm accuracy, latency, throughput and cost.
Vendor: OpenAI
What it does: the accuracy reference for general transcription across many languages, and the baseline everything else is measured against on clinical audio.
| Requirement | Minimum | Medium | High |
|---|---|---|---|
| GPU type | RTX 3060 | RTX 4090 | A100 80 GB |
| VRAM | 8 GB | 24 GB | 80 GB |
| vCPUs | 6 | 12 | 24 |
| RAM | 16 GB | 32 GB | 64 GB |
| Server | 1× RTX 3060 12 GB | 1× RTX 4090 24 GB | 1× A100 SXM 80 GB |
| Rate (audio hours/hour) | ≈ 30 | ≈ 110 | ≈ 300 |
Vendor: SYSTRAN
What it does: the same model reimplemented for throughput, several times faster per GPU at matching accuracy. The usual choice for a recorded archive.
| Requirement | Minimum | Medium | High |
|---|---|---|---|
| GPU type | RTX 3060 | RTX 4090 | A100 80 GB |
| VRAM | 8 GB | 24 GB | 80 GB |
| vCPUs | 6 | 12 | 24 |
| RAM | 16 GB | 32 GB | 64 GB |
| Server | 1× RTX 3060 12 GB | 1× RTX 4090 24 GB | 1× A100 SXM 80 GB |
| Rate (audio hours/hour) | ≈ 90 | ≈ 320 | ≈ 900 |
Vendor: University of Oxford
What it does: adds word-level timing and speaker assignment, so a transcript can be played back against the audio at the exact phrase under review.
| Requirement | Minimum | Medium | High |
|---|---|---|---|
| GPU type | RTX 3060 | RTX 4090 | A100 80 GB |
| VRAM | 8 GB | 24 GB | 80 GB |
| vCPUs | 6 | 12 | 24 |
| RAM | 16 GB | 32 GB | 64 GB |
| Server | 1× RTX 3060 12 GB | 1× RTX 4090 24 GB | 1× A100 SXM 80 GB |
| Rate (audio hours/hour) | ≈ 60 | ≈ 200 | ≈ 560 |
Vendor: NVIDIA
What it does: very fast English transcription with strong punctuation, suited to high-volume dictation on modest hardware.
| Requirement | Minimum | Medium | High |
|---|---|---|---|
| GPU type | RTX 3060 | RTX 4090 | A100 80 GB |
| VRAM | 8 GB | 24 GB | 80 GB |
| vCPUs | 6 | 12 | 24 |
| RAM | 16 GB | 32 GB | 64 GB |
| Server | 1× RTX 3060 12 GB | 1× RTX 4090 24 GB | 1× A100 SXM 80 GB |
| Rate (audio hours/hour) | ≈ 200 | ≈ 700 | ≈ 2,000 |
Vendor: NVIDIA
What it does: transcribes and translates between several European languages in one pass, for clinics where the consultation is not in the record language.
| Requirement | Minimum | Medium | High |
|---|---|---|---|
| GPU type | RTX 3090 | L40S 48 GB | A100 80 GB |
| VRAM | 24 GB | 48 GB | 80 GB |
| vCPUs | 8 | 16 | 32 |
| RAM | 32 GB | 64 GB | 128 GB |
| Server | 1× RTX 3090 24 GB | 1× L40S 48 GB | 1× A100 SXM 80 GB |
| Rate (audio hours/hour) | ≈ 70 | ≈ 240 | ≈ 680 |
Vendor: pyannote
What it does: marks who spoke when, separating clinician from patient from relative. Run alongside a transcription model rather than instead of one.
| Requirement | Minimum | Medium | High |
|---|---|---|---|
| GPU type | RTX 3060 | RTX 4090 | A100 80 GB |
| VRAM | 8 GB | 24 GB | 80 GB |
| vCPUs | 6 | 12 | 24 |
| RAM | 16 GB | 32 GB | 64 GB |
| Server | 1× RTX 3060 12 GB | 1× RTX 4090 24 GB | 1× A100 SXM 80 GB |
| Rate (audio hours/hour) | ≈ 120 | ≈ 420 | ≈ 1,200 |
These models are trained or tuned specifically on medical dictation and clinical speech, where general transcription loses accuracy on drug names, dosages and anatomy. Tiers and rates read as above; use them for initial sizing only.
Vendor: Google Health
What it does: transcribes physician dictation with the lowest published error rate on this page — 4.6 percent with its language model — because it was built for medical speech rather than adapted to it.
RAD-DICT WER 6.6 percent with greedy decoding and 4.6 percent with the published 6-gram language model; other medical dictation sets about 5.8–9.3 percent WER depending on decoding and domain.
| Requirement | Minimum | Medium | High |
|---|---|---|---|
| GPU type | RTX 3060 | RTX 4090 | A100 80 GB |
| VRAM | 12 GB | 24 GB | 80 GB |
| vCPUs | 6 | 12 | 24 |
| RAM | 16 GB | 32 GB | 64 GB |
| Server | 1× RTX 3060 12 GB | 1× RTX 4090 24 GB | 1× A100 SXM 80 GB |
| Rate (audio hours/hour) | ≈ 90 | ≈ 315 | ≈ 810 |
Vendor: NVIDIA
What it does: transcribes clean dictation at 5–7 percent error and translates across EU languages in the same pass, which suits multilingual clinics without a second model.
WER about 5–7 percent on clean English dictation, with multilingual EU support (developer-reported, Open ASR Leaderboard).
| Requirement | Minimum | Medium | High |
|---|---|---|---|
| GPU type | RTX 3090 | L40S 48 GB | A100 80 GB |
| VRAM | 24 GB | 48 GB | 80 GB |
| vCPUs | 8 | 16 | 32 |
| RAM | 32 GB | 64 GB | 128 GB |
| Server | 1× RTX 3090 24 GB | 1× L40S 48 GB | 1× A100 SXM 80 GB |
| Rate (audio hours/hour) | ≈ 30 | ≈ 105 | ≈ 270 |
Vendor: NVIDIA
What it does: transcribes long-form English fast and with strong punctuation. Accented clinical speech and drug names are where it loses ground, which is exactly what we measure on your recordings.
WER about 6 percent general English; degrades on accented clinical speech and drug names (independent).
| Requirement | Minimum | Medium | High |
|---|---|---|---|
| GPU type | RTX 3060 | RTX 4090 | A100 80 GB |
| VRAM | 12 GB | 24 GB | 80 GB |
| vCPUs | 6 | 12 | 24 |
| RAM | 16 GB | 32 GB | 64 GB |
| Server | 1× RTX 3060 12 GB | 1× RTX 4090 24 GB | 1× A100 SXM 80 GB |
| Rate (audio hours/hour) | ≈ 90 | ≈ 315 | ≈ 810 |
Vendor: OpenAI (weights released openly)
What it does: is the multilingual baseline everything else is compared against, at 99 languages. Known to hallucinate on silence, so a voice-activity step goes in front of it.
WER about 7–10 percent on medical dictation; 15–25 percent on noisy multi-speaker consultations; known hallucination risk on silence (independent).
| Requirement | Minimum | Medium | High |
|---|---|---|---|
| GPU type | RTX 3060 | RTX 4090 | A100 80 GB |
| VRAM | 12 GB | 24 GB | 80 GB |
| vCPUs | 6 | 12 | 24 |
| RAM | 16 GB | 32 GB | 64 GB |
| Server | 1× RTX 3060 12 GB | 1× RTX 4090 24 GB | 1× A100 SXM 80 GB |
| Rate (audio hours/hour) | ≈ 90 | ≈ 315 | ≈ 810 |
Vendor: Community (multiple fine-tunes)
What it does: cuts errors on drug and anatomy terms by 20–40 percent against base Whisper. Quality varies sharply between published fine-tunes, so selection and licence checking is part of the work.
20–40 percent relative WER reduction on drug and anatomy terms against base Whisper; quality varies sharply by fine-tune — licence and provenance checked per model.
| Requirement | Minimum | Medium | High |
|---|---|---|---|
| GPU type | RTX 3060 | RTX 4090 | A100 80 GB |
| VRAM | 12 GB | 24 GB | 80 GB |
| vCPUs | 6 | 12 | 24 |
| RAM | 16 GB | 32 GB | 64 GB |
| Server | 1× RTX 3060 12 GB | 1× RTX 4090 24 GB | 1× A100 SXM 80 GB |
| Rate (audio hours/hour) | ≈ 90 | ≈ 315 | ≈ 810 |
Volume, latency and language mix decide the combination. Live dictation needs a model that holds real-time pace on one card; an archive of recorded consultations is better served by the fastest batch option. We measure error rate on your own recordings, against your own terminology, before sizing anything.
At the start of a project we run a short proof of concept on a sample of your own data, measuring the accuracy and the throughput the model actually achieves on your material. The estimates on this page are replaced with real figures, so the cost and the schedule for the full engagement are known before anything is committed.
Send a representative sample, your expected volume and your latency target for medical speech to text. We benchmark the shortlisted models, recommend the lowest-cost GPU configuration that meets the target, and scale it from pilot capacity to a dedicated production cluster — with the cost per unit of work known before you commit.