How Accurate Is Speechmatics in 2026? Melia-1, Enhanced & Standard WER Benchmarks

It depends on the tier and the audio. Speechmatics sells three tiers — Melia-1 ($0.24/hr, mid-2026), Standard ($0.30/hr) and Enhanced ($0.75/hr) — and its own published claims are the best-documented figures available: an 11.97% average WER across 21 open-source test sets for the Ursa/Enhanced generation. Treat any single WER number with caution: results move substantially with audio condition, language and domain. For cross-vendor comparisons measured under one published methodology, see the Hugging Face Open ASR Leaderboard.

They are close, and the answer depends on what you are transcribing. Speechmatics positions Melia-1 on multilingual accuracy and price ($0.24/hr); AssemblyAI positions Universal-3.5 Pro on English and on its entity and formatting features. Neither vendor's published figures are measured on the same audio, so a like-for-like claim is not available from vendor material alone. The Hugging Face Open ASR Leaderboard is the closest thing to a neutral comparison.

Melia-1 for most workloads. It is the cheapest tier at $0.24/hr and is the model Speechmatics positions for high-volume batch work with strong multilingual coverage. Enhanced ($0.75/hr) is aimed at English broadcast content where the vendor's accuracy claims are highest and the price premium is acceptable. Standard ($0.30/hr) sits between them and is rarely the right pick — at that price the gap to Melia-1 is small. Run your own audio through both before committing.

Speechmatics is one of the few vendors that ships diarization as part of the transcription response rather than as a separate model, and it publishes diarization behaviour in its documentation. Diarization Error Rate varies enormously by audio type — broadcast audio with clean turn-taking scores far better than multi-speaker meetings with crosstalk. For published DER benchmarks across models, the pyannote model cards and the DIHARD challenge reports are the standard references.

Yes. Speechmatics offers a real-time streaming API alongside its batch API, with partial and final transcript events. Streaming and batch are priced and tuned differently, so batch accuracy figures should not be extrapolated to the streaming endpoint — that is true of every vendor, not just Speechmatics.

Different strengths. Whisper Large-v3 is open source (MIT), free to self-host, and has the widest community tooling; Speechmatics is a commercial API with diarization built in, an SLA, and EU data residency options. Whisper's published WER on standard datasets is well documented on the Hugging Face Open ASR Leaderboard, which also lists commercial vendors. For self-hosted, private, or zero-marginal-cost work Whisper wins on economics; for managed infrastructure with support, Speechmatics is the stronger fit.

They target different things. Deepgram markets Nova-3 on streaming latency and on its keyterm prompting; Speechmatics markets Melia-1 on multilingual coverage and price. Deepgram's per-hour pricing is competitive with Melia-1 at volume. Neither vendor publishes WER measured on the other's test set, so pick based on which capability matches your workload — streaming latency, language coverage, or diarization — rather than a single accuracy figure.

Ursa is the model generation behind the Enhanced tier. Speechmatics published its Ursa accuracy claims (11.97% average WER across 21 open-source test sets, a 38% relative lead over Google and 22% over Microsoft) when the generation launched. Melia-1 is the newer, cheaper model introduced in mid-2026 and positioned for multilingual batch work.

Because 'accuracy' in vendor marketing usually means 1 minus WER on the vendor's own chosen test audio, which is typically clean read speech. A 92-93% accuracy claim corresponds to roughly 7-8% WER on that audio. Real-world audio — meetings with crosstalk, phone-quality recordings, accented speech, technical vocabulary — routinely lands well outside that range for every vendor. Treat marketing accuracy figures as best-case ceilings, not expectations.

Because 92% and 93% are different metrics than WER. G2's 92% Speechmatics rating is aggregate user satisfaction — how happy customers are, not how many words are correct. Speechmatics' 93% medical accuracy claim is roughly 100% minus WER on the vendor's medical test set (a separate product from the batch tiers evaluated on this page). Ursa's marketed accuracy is (roughly) 100% minus 11.97% = 88% on the vendor's 21-set composite. These numbers are not directly comparable to each other, or to WER measured on different audio. When comparing engines, always compare like-for-like: WER on the same audio, scored the same way. For figures measured across vendors under one published methodology, see the Hugging Face Open ASR Leaderboard.