Introduction

On October 14, Speechmatics launched Oak (also styled Oak 1), a closed medical speech-to-text model aimed at clinical conversations that hop languages and speakers mid-visit. It is live today on the Speechmatics API — cloud, private cloud, or on-prem — with Batch support for Oak 1. There are no open weights.

Oak’s pitch is different from generic multilingual STT: follow language switches without picking a language up front, track who spoke, and protect drug names, diagnoses, and procedures that ordinary word-error-rate scores treat like any other token. Speechmatics reports 91% medical-term recall on average across 11 languages, ranking first against eight other providers in their October 2026 eval. Coverage is 55+ languages for medical speech.

Primary source: the Speechmatics Oak launch article (October 14). Vendor X posted the same day.

What Shipped

  • Model: Oak / Oak 1 — dedicated medical STT, not a general-purpose rename.
  • Access: closed API; deploy in cloud, private cloud, or on-premises; Batch for Oak 1 workflows today.
  • Languages: medical speech across 55+ languages; the published medical-term recall bench covers English plus Arabic, Danish, Dutch, Finnish, French, German, Norwegian, Polish, Spanish, and Swedish.
  • Conversation behavior: automatic mid-conversation language switching, speaker tracking, and language metadata on the transcript — no language picker and no hand-routing between models.
  • Clinical focus: preserve drug names, conditions, procedures, and side effects rather than optimizing only overall WER.
  • Batch speed claim: about 18 seconds to batch-transcribe one hour of multilingual audio (vendor figure).

This is a distinct lane from Deepgram Nova-3 Medical Multilingual and other medical STT SKUs: Oak’s day-0 story is multilingual consultation flow (code-switch + speakers + medical-term recall), not a single-language medical WER race.

Medical Term Recall, Not Just WER

Speechmatics argues WER flattens “the” and “amlodipine” into the same error budget. Their primary metric here is medical term recall — how often drug names, conditions, procedures, and side effects actually land in the transcript.

In their internal evaluation (same audio, default settings, no custom vocabulary or keyterm prompting, each provider’s best available model including dedicated medical / pharma variants where they existed):

  • Oak averaged 91% medical-term recall across 11 languages, first in every language tested; the other eight providers averaged about 74%.
  • English: Oak 88.5% vs ElevenLabs Scribe v2 Medical 87.7% (AssemblyAI’s medical model 85.6% in the same write-up).
  • Larger gaps show up outside English — for example Spanish 96.3% Oak vs ~93% next-best / ~86.8% average of others; Norwegian 93.3% vs 74.5% next-best.

Code-switching was scored separately (no language hints): Speechmatics cites mixed error rates of 15.1% for Arabic↔English and 14.1% for Mandarin↔English in published pairs — examples, not the full language matrix. Treat all of this as vendor benchmarking until you run your own clinical audio.

Who It’s For

Oak targets multilingual medical scribes, multi-person visits (patient + clinician + interpreter or relative), and teams that deploy across markets without wanting a weaker transcript the moment English leaves the room. Speechmatics already points to healthcare production context (clinical scribes, multilingual intake, NHS 111 call transcription via partners) and lists HIPAA / SOC 2 Type II / ISO 27001 / GDPR posture plus zero data retention options — relevant when patient audio cannot wander.

Getting started is through the Speechmatics API and docs / Academy examples; healthcare teams can also send sample audio for a head-to-head against their current provider.

Conclusion

If your product has to follow a consult that flips languages and speakers while keeping clinical terms intact, Oak is a same-day closed medical STT ship with a clear API path and a medical-term recall story that goes beyond English-only medical models. If you need open weights or fully on-device medical ASR, this is not that drop — watch open EU / domain ASR stacks separately (including today’s TranscribePsy for general on-device EU speech). For clinical builders already on Speechmatics, the next step is simple: run Oak Batch on a real multilingual consult tape and score the terms that matter to your note pipeline.

—Anabel ♡