Speechmatics

Speech-to-text benchmarks are usually run on audio nobody actually has: one speaker, a good microphone, a neutral accent, no interruptions. Every engine looks excellent on that. The differences appear the moment you feed in a real meeting, a doctor's consultation, or two people talking over each other in two languages.

Speechmatics is built for that second case, and it is the rare vendor in this space with an independent result to point at rather than a self-reported number. In Pipecat's August 2026 benchmark of twelve services, it recorded a 1.07% pooled word error rate — the lowest in the test.

This review covers what that benchmark does and does not tell you, why code-switching is the capability worth caring about, what on-device deployment changes for regulated buyers, and where a transcription API stops being the hard part of your problem.

Why Engineers Pick This One

• Independently benchmarked, not self-reported. Pipecat's August 2026 test put it at a 1.07% pooled word error rate — lowest of twelve services, ahead of Deepgram, AWS and Azure.
• Code-switching is the standout. It handles a speaker changing language mid-sentence, which most engines break on and which is normal speech in much of the world.
• 55+ languages with speaker diarization and custom vocabulary built in rather than bolted on.
• Runs on-device and on-premises, not just in the cloud — the deciding factor for healthcare, legal and anything that cannot send audio out.
• ISO 27001:2022, GDPR, HIPAA and SOC 2 Type II, with data non-logging as standard.

What A 1.07% Word Error Rate Actually Means

This is the strongest claim on the page, so it deserves to be read carefully rather than quoted.

A 1.07% pooled word error rate means roughly one word in a hundred is wrong — inserted, deleted or substituted. Against eleven competitors including Deepgram, AWS and Azure, that is a real result, and the fact it comes from Pipecat rather than from Speechmatics is what makes it worth citing at all. Nearly every accuracy claim in this category is the vendor grading its own homework.

What it does not tell you is how it performs on your audio. 'Pooled' means averaged across a test set, and word error rate is exquisitely sensitive to domain: medical terminology, product names, industry jargon and heavy background noise all move it substantially. A service that wins a pooled benchmark can still be second-best on your particular recordings.

The other thing WER hides is which words were wrong. One error in a hundred is excellent if the misses are 'the' and 'a', and much worse if they are drug names or dollar amounts. Custom vocabulary exists precisely because the words that matter most are the ones a general model gets wrong most.

An independent benchmark tells you the engine is good. Only your own audio tells you it is good at the words you cannot afford to lose.

Code-Switching, And Why It Is The Real Differentiator

Most engines handle multiple languages by asking you to pick one per file. That model breaks on how enormous numbers of people actually speak.

A speaker moving between Hindi and English inside one sentence, or Spanish and English, or Arabic and French, is not an edge case — it is ordinary conversation across much of the world, and a single-language engine transcribes half of it as nonsense.

Handling that mid-sentence is genuinely hard, because the model cannot settle on a language and stay there. It is also the capability most likely to matter if your users are outside a monolingual market, and the one least visible in an English-only demo.

Speaker diarization sits alongside it: labelling who said what. For meetings, interviews and contact-centre calls, an undiarized transcript of four people is close to useless — you have the words and none of the structure.

On-Device And On-Premises Changes Who Can Buy It

Cloud, on-premises and on-device deployment, with data non-logging as standard, plus ISO 27001:2022, GDPR, HIPAA and SOC 2 Type II.

In consumer software that list is procurement boilerplate. In healthcare and legal it is the whole conversation. A hospital cannot send patient consultations to a third-party cloud, and a law firm cannot send privileged material, regardless of how accurate the transcription is.

On-device is the strongest version of that answer: audio never leaves the machine, so the compliance question largely evaporates rather than being managed. It also removes latency and network dependence, which matters for live captioning where a two-second lag is the difference between usable and not.

The named customers — Adobe, LiveKit, Ubisoft, Stenograph — are a useful signal here. Stenograph in particular is a court-reporting company, which is about as unforgiving an accuracy environment as exists.

Choosing Between Transcription Vendors Sensibly

The evaluation everyone should run, and almost nobody does, takes an afternoon.

1. Collect twenty minutes of your genuinely worst audio — background noise, accents, people talking over each other. Not your best sample.
2. Run it through two or three vendors on their free tiers rather than reading comparison tables.
3. Score the errors by importance, not by count. Ten wrong articles matter less than one wrong dosage.
4. Load your domain terms into custom vocabulary before judging, since that is how the product is meant to be used.
5. Test diarization on your hardest case — a four-person call with interruptions, not a two-person interview.
6. If you have multilingual speakers, test a code-switched sample specifically. This is where engines diverge most.
7. Only then compare cost, and compare it at your real monthly volume rather than per minute.

Where Transcription Stops Being The Problem

It is worth being clear about scope. Speechmatics is infrastructure — an API that turns audio into accurate text. It is not a meeting-notes product, a podcast editor, or an analytics platform.

If you want summaries, action items, searchable archives or speaker analytics, you are building that on top, or you want a product like Otter or Rev that ships the whole workflow with transcription inside it. Choosing an API when you wanted an app is the most common mismatch in this category.

The flip side is that infrastructure is the right choice when transcription quality is the constraint and you control the product around it. A voice AI agent, a live captioning system, a clinical documentation tool — these live or die on the accuracy of the text, and a two-percent difference in error rate compounds through everything downstream.

Finally, accuracy on paper is not accuracy in your pipeline. Audio capture quality, sample rate and how you chunk streaming input can cost you more than the difference between the top three vendors.

The Four Terms This Category Runs On

Comparing transcription vendors is impossible without these.

Word error rate (WER)

The share of words wrong — inserted, deleted or substituted. Lower is better. It counts errors without weighting them, so it treats a missed 'the' and a missed drug name identically.

Diarization

Labelling who spoke each segment. Without it, a four-person meeting transcript has all the words and none of the structure.

Code-switching

Changing language mid-sentence. Ordinary speech in much of the world, and the thing single-language engines fail on most visibly.

Custom vocabulary

Domain terms supplied in advance — product names, drug names, jargon. It exists because the words that matter most to you are the ones a general model gets wrong most often.

Coverage, Deployment And Compliance