Speechmatics built a model for voice agents. The benchmark says "accurate" has more than one meaning.

Speechmatics built a model for voice agents. The benchmark says "accurate" has more than one meaning.

On August 27, Speechmatics and LiveKit announced Linden, a speech-to-text model built specifically for voice agents. The press release points at an open benchmark. That benchmark is more interesting than the announcement, because it shows the lowest word error rate and the slowest response time sitting on the same row.

%1,07 POOLED WER, LOWEST ON THE BOARD

%83,2 PERFECT TRANSCRIPTS (TOP: 84.7%)

495 ms MEDIAN LATENCY (FASTEST: 221MS)

55+ LANGUAGES LINDEN COVERS

What was announced

Speechmatics and LiveKit put out a joint release on August 27, 2026. Linden is the first Speechmatics model built specifically for voice agents, and it is now available through LiveKit Inference, the layer where LiveKit gathers voice models in one place. Teams already building there can switch over from the interface or in code in under a minute.

The pitch is aimed at a specific set of failures: strong accents, speakers using a second language, and alphanumeric strings like account or card numbers read out over a poor line. Linden covers 55 or more languages. The release also brings speaker diarization to LiveKit agents for the first time, so an agent can tell apart who said what rather than treating a call as one undivided stream of words.

LiveKit Inference keeps the parts around the model: the pipeline, routing, billing, and the turn detection that decides when the agent starts speaking.

The benchmark, read properly

The release points to the Pipecat STT benchmark, an open evaluation of speech-to-text services for real-time agents. It runs 1,000 samples and reports four families of metric: semantic word error rate, time to final segment, transcription success rate, and the share of transcripts that come back perfect. Semantic WER only counts errors that change meaning for the model downstream, so punctuation, filler words and number formatting do not inflate the score.

Here is what the board actually shows for Speechmatics: 1.07% pooled WER, 1.40% mean WER, 83.2% perfect transcripts, and latency of 495ms median, 676ms at P95, 736ms at P99.

That 1.07% is the lowest pooled error rate on the board, and it is a real result. Read one column over and the ranking changes. AssemblyAI's universal-3-5-pro returns 84.7% perfect transcripts against Speechmatics' 83.2%. Cartesia ink-2 sits at 84.2%, Soniox stt-rt-v4 at 84.1%, AssemblyAI u3-rt-pro at 83.9% and Soniox stt-rt-v5 at 83.3%, which puts Speechmatics sixth on that column. The model with the fewest total errors gets fewer calls exactly right than five others.

Latency splits it further. NVIDIA's Nemotron 3.0 ASR, an English-only model, posts a 221ms median, Deepgram nova-3-general 247ms, Soniox 249ms. Speechmatics sits at 495ms, roughly twice the fastest. On a phone call that gap is audible. It lands in the pause before the agent answers, which is the part callers judge without knowing they are judging it.

A vendor number and a benchmark number, again

Speechmatics states on its own site, in the article introducing the LiveKit integration, that it is 25% better than the next provider in real-world tests. The open benchmark it points to does not produce that figure.

On pooled WER the next best score belongs to Azure at 1.18%, a relative gap of about 9%. Against AssemblyAI at 1.22% the gap is about 12%. Both figures are my own arithmetic on published numbers.

Both can be true if "real-world tests" means something other than this benchmark. The company does not say which test produced the 25%, so we are not going to guess. The point for anyone choosing a model is narrow and practical: when a company cites an open benchmark and also quotes a number, check whether the number came from that benchmark.

This keeps happening. We ran into the same shape in July, when a vendor-published score and an independent lab's measurement of the same benchmark family sat 15 points apart.

What this looks like on a real phone line

At TAO AI LAB we build voice assistants that answer real phone lines for businesses, in Turkish and English. We run on Vapi, with its Squads feature, a separate assistant per department. We are not on LiveKit, so this release is not a switch we flip. The reason it matters to us is the layer it touches, and that layer exists on every platform.

Transcription is where a call quietly goes wrong. The caller has an accent the model has heard less often. They are speaking their second language. They read out a reservation number while a television plays behind them. None of that shows up as an error message. It shows up as an assistant that answers a question nobody asked, and the caller decides the system does not work.

So the benchmark's split is the useful part, more than the headline. A voice assistant needs the transcript to be right and the reply to start quickly, and this board shows those two goals pulling in different directions. My read: the choice is a trade, and it should be made per use case rather than by picking whoever is top of a list. A line where people dictate numbers can pay for accuracy with a few hundred milliseconds. A line built on quick back and forth cannot.

The other thing worth taking from this is what a benchmark cannot do. It runs 1,000 samples chosen by someone else. It has never heard your caller's accent, your product names, or the noise on your line. Numbers like these are useful for narrowing a shortlist and weak for making the final call.

What this does not settle

Three limits. First, we could not open the press release page directly, so the announcement details here come from the coverage of it and from LiveKit and Speechmatics documentation. Second, the benchmark's own repository does not state the date of the run behind these figures, and results move as providers ship, so treat the table as a snapshot. Third, none of these numbers were produced on Turkish audio, and our assistants are built for Turkish as well as English, so for us the board is a starting point rather than an answer.

Frequently asked questions

What is Linden?
A speech-to-text model Speechmatics built specifically for voice agents, announced with LiveKit on August 27, 2026 and available through LiveKit Inference. It covers 55 or more languages.

Does the lowest word error rate mean the best model?
Not by itself. On the Pipecat benchmark Speechmatics has the lowest pooled WER at 1.07%, while five models return more perfect transcripts, AssemblyAI leading at 84.7%, and several services answer in roughly half the time.

What is semantic WER?
A word error rate that counts only mistakes changing meaning for the model that reads the transcript. Punctuation, filler words and number formatting are ignored.

Why does transcription accuracy matter more on a phone line?
Because accents, second-language speakers, dictated numbers and background noise all arrive at once, and a transcription mistake reaches the caller as a wrong answer rather than an error.

Voice assistants are judged on the parts nobody lists in a spec sheet, and the layer that holds a conversation together is what we build.

One question for you: if you had to give up 250 milliseconds of response time for a clearly better transcript, would you take the trade? Tell me in the comments.

Sources

• Speechmatics and LiveKit, "Speechmatics and LiveKit target the accuracy gap breaking voice agents in production" (August 27, 2026)
• Pipecat STT benchmark, metrics and results table (semantic WER, TTFS, perfect transcription rate)
• LiveKit documentation, Speechmatics STT
• Speechmatics, speech technology for AI voice agents