Speechmatics

Our full take

Speechmatics is a speech-to-text engine first, and that is the whole point to get straight. It listens to audio and writes down the words. It now also offers its own text-to-speech, but it does not generate a reply (there is no language model) and it does not dial a phone. So if you are shopping for a finished voice agent that answers your calls, this is not that. It is one of the parts you would build that agent from, and it is a good one.

Where it earns its place is languages. Speechmatics transcribes 56+ languages off a single model, which means you get the regional accents and dialects (Brazilian Portuguese, Canadian French, and so on) without bolting on a separate pack for each. Most of the cheaper speech-to-text engines top out around seven or ten languages. Deepgram, the closest building-block vendor we cover, lists seven. If your callers speak Tagalog, Welsh, Swahili or Urdu, that gap is the entire reason to look here.

The second reason is where it runs. Most speech-to-text APIs only run in the vendor’s cloud, you send them audio and they send back text. Speechmatics will also run inside your own data centre, as a container on your own hardware (CPU or GPU), on Kubernetes, or as a pre-built virtual machine they call a Virtual Appliance. For a hospital or a bank that cannot let call audio leave the building, that on-premises option (meaning it runs on your own servers, not someone else’s cloud) is often a hard requirement, not a nice-to-have. It is the kind of thing you cannot retrofit, so it matters that it is there from the start.

Now the pricing, and here is the bit that trips people up twice over. Speechmatics bills per hour of audio, not per minute like the agent platforms. And the rate table you see is not the list price: it renders with a “Model Training” toggle switched on by default, a setting that shares your audio with Speechmatics to train its models in exchange for a discount of about a third. Every figure on the default page assumes a data deal you have not agreed to. We switched the toggle off on 23 July 2026 and read the real list rates, and those are the figures this page now stores. The list floor for the established models is $0.36 an hour on Standard, about $0.006 a minute, the figure shown at the top of this page. Treat that as the floor, not the average. The real rate climbs with the model you pick and the mode you run.

Here is how the list rates split, read with the discount toggle off. Real-time transcription (live, as the audio streams in) is $0.36 an hour on the Standard model and $0.64 on the higher-accuracy Enhanced model. Batch transcription (a recorded file after the fact) is $0.36 Standard and $0.60 Enhanced. In per-minute terms that is about $0.006 to $0.0107 depending on model and mode. Leave the model-training toggle on and the displayed rates drop to $0.24/$0.43 real-time and $0.24/$0.40 batch, the discounted figures we ourselves stored before this correction, which is rather the point about how easy the default is to mistake for the price. The rate table loads through JavaScript, so we read these figures from a dated capture rather than a plain page fetch. The free tier remains 3,000 minutes a month (50 hours), split 1,200 real-time and 1,800 batch, no card needed, plus around a million free text-to-speech characters, and a separate 20% volume discount once you cross 500 hours a month.

The new model is Melia 1, launched 2026-06-17. It transcribes multilingual audio in one pass, including speakers who switch language mid-sentence, across 55+ languages by Speechmatics’ own description, with no need to pick a language up front. Two caveats before you build on it. First, it is a batch-only production preview for now (real-time is promised, not shipped), so it cannot power a live phone agent yet. Second, the advertised as low as $0.129 an hour, which is also the headline price on the Pro plan card, is the model-training-discounted figure; the list rate with the toggle off is $0.192 an hour. Speechmatics’ launch post also claims Melia beats Deepgram, Microsoft and AssemblyAI on the FLEURS multilingual test set. That is the vendor marking its own homework, so treat it as a claim, not a result.

One honest caveat on cost. That per-minute number looks tiny next to a $0.06-a-minute agent platform, and it is, but it is not comparing like for like. Speechmatics is charging you for one job, the transcription. The platforms are charging for transcription plus the language model plus the voice plus the phone line bundled together. To build a full phone agent on Speechmatics you still have to pay for an LLM and telephony separately, though it now has its own text-to-speech ($0.011 per 1,000 characters) so the voice no longer has to come from elsewhere. Add those up and the real per-minute cost lands a lot closer to the bundled platforms than the $0.006 headline suggests.

On compliance, Speechmatics is unusually well-documented for a building block. Its own security page states SOC 2 Type II, ISO/IEC 27001:2022, GDPR and full HIPAA compliance, with AES 256 encryption at rest and TLS 1.2 or higher in transit, plus a public trust centre where you can pull the actual reports. We have ticked HIPAA, SOC 2 Type II and GDPR here because the vendor states them directly. We left SOC 2 Type I unticked: the page names Type II, not Type I, and we do not assume one from the other. For a regulated buyer, that combination of on-prem deployment plus written certifications is the strong card.

My read: Speechmatics is the one you reach for when language coverage or on-premises deployment is non-negotiable, and you have the engineering to assemble the rest of the agent around it. The voice-quality and ease-of-use scores sit lower here than for a finished platform, and that is fair, this is infrastructure, not a product you switch on. If you just want calls answered without standing up your own stack, a bundled platform will get you there faster. If you need to transcribe twenty languages, or keep the audio on your own servers, very little else competes.

The 1 to 10 scores on this page are an editorial preview, our provisional read to get the framework in place, not a measured result. We have not run Speechmatics through our own test calls yet, so there is no Voxrater latency figure here. The pricing, language, deployment and compliance detail is sourced from Speechmatics’ own pricing, security, languages and deployments pages, first captured 2026-05-31 and most recently re-captured 2026-07-23 (when we also read the rate table with the model-training discount toggled off to confirm the list rates).