Speechmatics Review

Most engineering teams start with OpenAI Whisper because it's free, then discover the accuracy gap the first time a call-center recording gets garbled. That's the opening Speechmatics has quietly held for years. Founded in 2006 out of Cambridge, they raised a $62M Series B in 2022 and still ship their own Ursa engine rather than reselling someone else's.

Vendor risk is low here. This isn't a two-year-old wrapper — it's a two-decade speech-recognition specialist with real enterprise logos in healthcare and finance. The on-premises and on-device deployment options matter for regulated buyers who can't send audio to a shared cloud.

The catch is category positioning: Deepgram raised far more and markets harder, so Speechmatics wins on accuracy and quiet trust rather than mindshare. Pilot it against your worst audio for 30 days, then decide.

Every voice-AI roadmap eventually hits a data-residency wall, and that's where cloud-only ASR vendors fall out of contention. Speechmatics answers it directly: the Ursa engine runs on-premises and on-device, not just in a hosted API. For a Head of AI standardizing a speech layer across products, owning where audio is processed is the whole ballgame.

The strategic depth shows in Flow, which stitches ASR, an LLM, and text-to-speech into one conversational agent runtime with LiveKit and Pipecat integrations. That positions them as an infrastructure layer, not a single-feature transcription vendor.

The tradeoff is model control: unlike self-hosting OpenAI Whisper, you're committing to their proprietary engine and its 56+ language roadmap. But their code-switching and sub-500ms streaming are genuinely ahead of what an open-weights stack gives you out of the box.

Eight free hours a month is a real evaluation budget, not a demo. Pro starts at $0.24 per hour, billed to the second. No credit card to start.

Run the math. 500 hours a month of audio, at $0.24, is $120 monthly before volume discounts kick in. Amazon Transcribe lists around $1.44 per hour at standard rates — six times the sticker. Speech Intelligence summarization and sentiment ride the same meter, no separate SKU.

The catch is the ceiling. Pro caps at 6,000 hours a month; past that you're in Enterprise custom pricing with no published rate. Text-to-speech runs $0.011 per 1,000 characters. Transparent tiers mean procurement won't fight this, but the enterprise invoice is a sales call, not a page.

Sub-500ms partials mean the transcript often lands before the speaker finishes the sentence. For a live captioning or agent-assist workflow, that latency floor is the difference between usable and laggy. Real-time and batch share one API, so you're not maintaining two integrations.

Diarization is built in, not a paid add-on, with controls for speaker sensitivity and max speaker count. That's the daily-friction win — AssemblyAI and most rivals treat speaker labeling as a separate call. Custom Dictionary biases recognition toward product names and jargon, which is where generic ASR quietly mangles your domain terms.

The friction is observability. The docs describe the endpoints well, but there's no public word-error-rate benchmark by language, so you're testing accuracy on your own audio. Code-switching handling, however, is genuinely strong for mixed-language calls.

Most transcription tools quietly assume everyone in the meeting sticks to one language. Real conversations don't work that way, and Code-Switching handling here actually follows a speaker who slides between languages mid-sentence. That's the kind of thing you don't notice until a tool gets it wrong.

Getting started is refreshingly low-friction. The free tier is 8 hours a month with no credit card, so you can wire up a real prototype before anyone signs anything. Full feature access on free, too — not a crippled trial.

Reliability-wise, a 56+ language engine that also runs on-device is reassuring for spotty connections. The catch: this is a developer API, so there's no polished mobile app the way Otter.ai or Google's consumer apps feel. But for building, not just recording, that's the right call.

The scariest word in speech-to-text right now is 'free.' OpenAI Whisper commoditized decent transcription overnight. So the real question is why anyone pays.

Speechmatics has an answer, and it's not marketing. Founded 2006, still independent while Nuance got swallowed by Microsoft. The Ursa engine's accuracy edge is real on hard audio — noisy, accented, multilingual — where Whisper drifts. On-prem and on-device deployment gives a clean exit story: your data never had to leave.

'The most accurate' is the kind of superlative that ages poorly, and there's no public per-language benchmark to check it. And the last funding round was 2022's $62M Series B — no fresh capital signal since. But 20 years and real enterprise use is a track record most rivals can't fake.