Speechmatics

Speechmatics is best understood as three products wearing one API. First, a general transcription engine with a model ladder you choose from: Melia 1 at $0.12/hr for multilingual and mixed-language audio, Oak 1 Medical at $0.15/hr for clinical vocabulary, Standard at $0.23/hr for high-volume cost-sensitive work, and Enhanced at $0.38/hr when single-language accuracy matters more than money. Second, a real-time layer quoted at 350ms with WebSocket and REST APIs, speaker diarization, language labeling, language hints and a custom dictionary — the part broadcasters, contact centres and court reporters buy. Third, Linden 1, a real-time-only voice-agent model with turn detection built in, aimed at teams wiring STT into conversational agents rather than transcribing files. The honest strengths: the multilingual model doesn't require you to declare a language up front, and it handles speakers switching mid-sentence, which removes an entire preprocessing step from multilingual pipelines. Diarization and custom vocabulary are included rather than gated. The deployment matrix is unusually broad for this category — SaaS, containers, virtual appliance, on-device and hybrid — and there is a shipped precedent for on-device work in Adobe Premiere and Stenograph CATalyst VP, plus EU/US/AUS processing location choice and no model-training logging by default. SOC 2 Type II, ISO/IEC 27001:2022, GDPR and HIPAA are in place, which is what makes the healthcare and legal use cases sellable at all. The honest weaknesses: everything is metered by the audio hour, and the enrichment add-ons are where budgets break. Translation at $0.49/hr is four times the Melia 1 base rate; add chapters, topics and summaries and you have left the cheap tier behind. Several add-ons are marked "on request" — PII redaction, entity detection, audio alignment — so you can't self-serve those and shouldn't plan a pipeline around them without asking. Add-on availability also varies by model and by pre-recorded versus real-time, so a feature that works on Enhanced may not exist on Melia 1. Self-serve tops out at 50 concurrent real-time sessions and 10 files per second; above that it's an Enterprise contract, and on-prem, on-device and hybrid deployments are enterprise conversations rather than a signup flow. SSO is a $280/mo add-on rather than something included. Where it fits: contact centres and voice-agent builders who need speaker-aware streaming across many languages, media and broadcast teams captioning live events, healthcare teams using Oak 1 Medical for dictation and ambient scribes, legal transcription where accent coverage matters, and any enterprise whose security review kills pure-cloud vendors before the demo. Where it doesn't: anyone transcribing a handful of files a month, teams without engineering capacity to consume an API, and buyers whose pipeline enriches every hour with translation and summaries — at that point you are comparing enriched-hour prices, not $0.12/hr.