Speechmatics is a speech-to-text API for pre-recorded and real-time audio. It sells itself to regulated industries like healthcare, and to teams building voice agents. Instead of one general model, it now offers several: Melia 1 for multilingual work, Oak 1 for medical vocabulary and Linden 1 for real-time voice agents. New accounts get $100 in credit with no card required, and the self-serve rates are public.
This review covers the documentation, the pricing page and the vendor's own claims. scored.tools did not run audio through the API. Any accuracy or latency figure below is something Speechmatics says about itself, and it is labelled that way.
Scores at a glance
| Criterion | Score (out of 10) | Why |
|---|---|---|
| Overall rating | 7 | Strong pricing and developer setup, quality unproven outside vendor claims |
| Output quality | 5 | Only vendor evidence, including a claimed 350ms latency, case studies and Enhanced model accuracy claims |
| Ease of use | 8 | $100 credit with no card, API key from the portal, pip-installable SDKs, docs with examples |
| Pricing value | 8 | Published hourly rates, $25 credit packs, volume discounts up to 25% |
| API and integration | 8 | Batch, real-time and search APIs, plus LiveKit, Pipecat, Vapi, Twilio and Zapier |
| Problem fit | 6 | Compliance and on-prem options match the pitch, but these are vendor-stated and agent STT is cloud-only |
The output quality score of 5 reflects missing evidence. It says nothing about whether the transcripts are bad. Speechmatics has not been shown to be weak; it just hasn't been shown to be strong by anyone outside the company.
Key features
Batch and real-time transcription
The API handles pre-recorded files (batch) and live streams (real-time). The docs also describe a search API. Real-time costs more than batch on every model that offers both, so it makes sense to route archived audio through batch even if your app also streams.
Melia 1: multilingual
Melia 1 is the general multilingual model. Speechmatics says it transcribes 55+ languages and can switch language mid-sentence. Code-switching is a real problem in call centres and media in places like India, Singapore and the Gulf. If your audio mixes languages, this is the claim to test first.
Oak 1: medical
Oak 1 targets clinical vocabulary, such as drug names, dosages and procedure terms. General models tend to mangle these, and a wrong dosage in a clinical note costs more than a typo. Speechmatics lists HIPAA, GDPR, SOC 2 and ISO 27001 compliance on its site, along with cloud and on-prem deployment. All of that is vendor-stated, so ask for the paperwork during procurement.
Linden 1: voice agents
Linden 1 is a real-time model with built-in turn detection, which is the part of a voice agent that decides when the user has stopped talking. Getting turn detection wrong makes an agent either interrupt people or sit in awkward silence. Speechmatics cites 350ms latency on its site. Per the tool data, agent STT is cloud-only, so on-prem buyers can't use this model.
SDKs and integrations
The docs list pip-installable SDKs and integrations with LiveKit, Pipecat, Vapi, Twilio and Zapier. For voice agent builders, LiveKit and Pipecat support matters most, because it means you can plug Speechmatics into a pipeline you already run without writing a custom transport layer. There's also a GitHub academy repo with tutorials (168 stars). It's a sample repo and tells you nothing about transcription quality.
Pricing breakdown
Pricing is the strongest part of the offer, but it takes some reading. Speechmatics advertises Melia 1 "from $0.12/hr" pre-recorded and $0.20/hr real-time. Those are the lowest rates you can reach, not the list price. The pricing page shows these list rates per hour:
| Model | Pre-recorded | Real-time |
|---|---|---|
| Melia 1 | $0.24 | $0.40 |
| Oak 1 (medical) | $0.30 | $0.46 |
| Linden 1 (voice agent) | Not offered | $0.30 (promo, normally $0.40) |
| Standard | $0.45 | $0.45 |
| Enhanced | $0.75 | $0.80 |
Two discounts close the gap to $0.12. The pricing page offers 33% off if you opt into a model-training program (off by default), and up to 25% off on the top annual subscription. Stack both and Melia 1 batch lands at roughly $0.12/hr, with Enhanced at about $0.38 to $0.40/hr. The training opt-in means Speechmatics can use your audio to train models. For a hospital or a law firm that's often a non-starter, so budget at or near list price if you're in a regulated field.
Plans
- Free: $0 with $100 credit, no card. At the Melia 1 list rate that's over 400 hours of batch audio, enough for a proper evaluation.
- Pro (pay as you go): one-off credit packs from $25, valid for 12 months, with no recurring commitment. Annual subscriptions start at $120/yr for 126 credits (5% off) and go up to $15,000/yr for 20,000 credits (25% off).
- Enterprise: custom per-hour rates and higher concurrency limits. SSO is a paid add-on at $280/mo.
The pricing page also lists text to speech with the first 1 million characters free, which suggests Speechmatics is moving toward a full voice stack. This review covers speech-to-text only.
The Enhanced model is the odd one out. At $0.75/hr list for batch, it costs about three times Melia 1. Speechmatics' accuracy claims for Enhanced are its own, so check whether it beats Melia 1 on your audio before paying the premium.
Pros and cons
Pros
- Self-serve rates are published per hour for each model, so you can estimate a monthly bill from your audio volume.
- $100 starting credit with no card makes a real trial cheap.
- Separate medical and voice agent models instead of one general model asked to do everything.
- SDKs and documented integrations with LiveKit, Pipecat, Vapi and Twilio cut setup time for voice agent stacks.
- On-prem deployment is offered for teams that can't send audio to a third-party cloud.
Cons
- Rates differ by model and by batch versus real-time, and the headline $0.12/hr needs stacked discounts to reach.
- The cheapest rates depend partly on opting into model training, which many regulated buyers won't accept.
- Enhanced is much more expensive than Melia 1, with only vendor evidence to justify the gap.
- Accuracy, latency and customer results on the site are the vendor's claims and haven't been independently checked.
- Voice agent STT is cloud-only, which weakens the on-prem story for agent builders.
Who is it for
Speechmatics fits developers building in healthcare, where Oak 1 and the compliance list are the draw. It also fits teams handling multilingual or code-switched audio, if Melia 1 performs on their recordings the way Speechmatics says it does. Voice agent builders on LiveKit or Pipecat get a model with turn detection and a ready integration, which saves real plumbing work.
It's a weaker fit if you want a finished transcription app with an editor and human review. That's closer to what Rev sells. If you have GPUs and engineering time, self-hosting Whisper costs nothing per hour, though you take on the ops work and won't get a medical model or turn detection out of the box. For general English transcription at scale, Deepgram and AssemblyAI are the obvious alternatives to price against.
Verdict
Speechmatics scores 7 out of 10. It's a developer-focused speech-to-text API with public per-hour pricing, a generous free credit and specialised models for medical and voice agent work. The setup and pricing look solid on paper. The big unknown is quality, because every accuracy and latency figure comes from Speechmatics itself.
If you work in a regulated industry or with multilingual audio, put Speechmatics on your shortlist and spend the $100 credit on your own recordings. Run the same files through Deepgram and AssemblyAI and compare word error rates on the terms that matter to you, such as drug names, accents or mid-sentence language switches. Price it at list rates unless you're comfortable opting into model training. If it wins on your audio, the pricing is easy to plan around. If it doesn't, you've lost nothing.