What is Speechmatics Oak 1?
Oak 1 is Speechmatics' speech-to-text model for healthcare audio, launched on 14 October 2026 and available for both recorded and real-time audio. It is built for clinical conversations in which patients, doctors, nurses, interpreters and relatives switch languages mid-sentence: you do not choose a language first, and every word in the transcript is labelled with the language it was spoken in. Speaker labels can be carried into clinical summaries.
Speechmatics says it covers medical speech in more than 55 languages, with its extra medical-vocabulary accuracy applied to a listed subset and plain transcription elsewhere. A later batch update, it says, cut medical keyterm errors by a further 25% and errors on numbers, IDs and dosages by about half.
Oak 1 runs on Speechmatics' cloud in every region, in a private cloud or on premises through its containers. It is a separate model from the single-language Enhanced Medical option, which keeps features such as custom dictionaries and confidence scores.
How much does Speechmatics Oak 1 cost?
| Plan | Price | What you get |
|---|---|---|
| Pay as you go | $0.30 an hour batch, $0.46 real-time | No discounts; $100 sign-up credit, 2 concurrent real-time sessions on free credit |
| Annual commitment | Up to 25% off | Largest annual tier; stacks with the 33% training opt-in discount |
Oak 1 lists at $0.30 an hour for recorded audio and $0.46 for real-time, billed per second. 33% off if you allow training on your data; annual commitments save up to 25%. $100 free credit for 12 months.
Is Speechmatics Oak 1 worth it?
Oak 1 answers a problem that English-first benchmarks hide. In Speechmatics' tests on 19 medical datasets in 11 languages, it caught 91% of medical terms on average, against 74% for eight rivals including AWS, Google, Microsoft, OpenAI, Deepgram and ElevenLabs. The spread matters more than the average: in Norwegian it scored 93.3% where the next best managed 74.5%, and Deepgram fell from 84.0% in English to 26.1% in Danish.
In English the lead is thin. Oak scored 88.5%, against 87.7% for ElevenLabs Scribe v2 Medical, which is a rounding error for a clinic that works only in English. Where Oak pulls clear is mixed-language speech: 15.1% error on Arabic-English conversations, against 33.2% for the next model tested.
The catch: every comparison is Speechmatics' own, run with default settings and no custom vocabulary, and Artificial Analysis has not yet scored Oak. The model also gives up conveniences the Enhanced models keep, including custom dictionaries, confidence scores, translation and automatic end-of-utterance detection, so real-time clients must detect silence themselves. On-premises deployments do not get real-time speaker labels.
Our take: for a health system serving patients in many languages, Oak 1 is the first model worth piloting before anything general-purpose. A clinic that only sees English speakers gains little over the medical tiers it may already pay for.
Where it earns its place
An ambient-scribe start-up serving clinics in Norway and Sweden
Draft visit notes from consultations that switch between Norwegian, Swedish and English
Send recordings to oak-1 with language set to multi; Speechmatics says an hour of audio is processed in about 18 seconds. 10,000 consultation hours cost about $3,000 at list price.
A London hospital trust's interpreting service
Keep a record of who said what when an interpreter joins a consultation
Run Oak 1 in real time with speaker labels through the SaaS API, or in on-premises containers if audio cannot leave the trust, accepting that diarization is SaaS-only.
Before you commit
- Setup
- Set "model": "oak-1" and "language": "multi" in a batch job or real-time StartRecognition message; auto is rejected. Real-time clients send ForceEndOfUtterance after their own voice-activity detection.
- Integrations
- Speechmatics batch and real-time APIs in every region, on-premises and private-cloud containers, LiveKit plugin.
- Data and privacy
- If you do not opt in to training, real-time audio is never stored and recorded audio is deleted after 7 days or sooner by API. Opting in earns 33% off. HIPAA, SOC 2 Type II, ISO/IEC 27001:2022 and GDPR.
- Cost at scale
- 100,000 hours of recorded audio cost about $30,000 at list price, or roughly $15,000 with the training opt-in and the top annual commitment combined.
- Limits
- Medical-term uplift only for listed languages. No speaker identification, custom dictionary, translation, audio events or confidence scores. Real-time diarization is SaaS-only.
What’s good
- 91% medical-term recall across 11 languages, per Speechmatics
- No language selection; word-level language labels
- Cloud, private cloud or on-premises containers
- HIPAA, SOC 2 Type II and ISO 27001; real-time audio not stored by default
What’s not
- All accuracy comparisons are vendor-run
- English lead over Scribe v2 Medical is 0.8 points
- No custom dictionary, confidence scores or translation
- Real-time clients must send their own end-of-utterance signal
Best for
Hospitals, medical-scribe vendors and telehealth services handling consultations in several languages, especially where audio must stay on premises.
Speechmatics Oak 1 alternatives
Speechmatics Oak 1: common questions
- Is Speechmatics Oak 1 worth it?
- Built for the consulting room where three languages share one appointment, and on Speechmatics' own figures it misses fewer drug names than anyone. We rate it 4.1 out of 5.
- How much does Speechmatics Oak 1 cost?
- Oak 1 lists at $0.30 an hour for recorded audio and $0.46 for real-time, billed per second. 33% off if you allow training on your data; annual commitments save up to 25%. $100 free credit for 12 months.
- What are the downsides of Speechmatics Oak 1?
- All accuracy comparisons are vendor-run; english lead over Scribe v2 Medical is 0.8 points; no custom dictionary, confidence scores or translation; real-time clients must send their own end-of-utterance signal.
- Who is Speechmatics Oak 1 best for?
- Hospitals, medical-scribe vendors and telehealth services handling consultations in several languages, especially where audio must stay on premises.
- How much does Speechmatics Oak 1 cost for a team or in production?
- 100,000 hours of recorded audio cost about $30,000 at list price, or roughly $15,000 with the training opt-in and the top annual commitment combined.
- Is Speechmatics Oak 1 safe for company data?
- If you do not opt in to training, real-time audio is never stored and recorded audio is deleted after 7 days or sooner by API. Opting in earns 33% off. HIPAA, SOC 2 Type II, ISO/IEC 27001:2022 and GDPR.
- What are the best Speechmatics Oak 1 alternatives?
- The closest alternatives we have reviewed are ElevenLabs Scribe v2, AssemblyAI STT-3.7 (batch, Realtime and Realtime-Agent), Voxtral 3 (Transcribe and TTS) and NVIDIA Nemotron 3.5 Transcribe.
Answers checked Nov 21, 2026.