Corti

Corti is a Copenhagen healthcare AI company that started in emergency dispatch and has rebuilt itself, deliberately, into a developer platform. The origin story is specific: founded in 2016 by Andreas Cleve and machine-learning researcher Lars Maaloe, the company trained models to spot cardiac arrest in live emergency calls, listening for speech patterns, dialect and background noise, and Copenhagen's own emergency services ran the system on the city's 112 line. A widely cited study found the system identified cardiac arrest in about 95 percent of cases from call audio against 73 percent for unaided Copenhagen dispatchers, with detection inside roughly 50 seconds of the call.

If you are shopping for a dispatcher console today, though, what Corti actually sells in 2026 is an API. The platform now spans medical speech-to-text, text generation, medical coding, an embeddable ambient assistant, and an OpenAI-compatible model gateway, all metered through one credit wallet. That pivot matters for this review: a catalog entry describing "AI for emergency call centers" is describing Corti's history more than its current storefront.

The pivot is worth describing because it explains everything about what you can buy. Corti's early product was a dispatcher-facing console, Corti Triage, that sat on the responder's screen during a live call and suggested questions and possible critical conditions. That deployment, across the Copenhagen metropolitan area covering nearly two million residents, is what produced the cardiac-arrest evidence. What Corti sells now is the machinery underneath: the medical speech recognition, the clinical fact extraction, the document structuring, exposed as endpoints that other software vendors call. When the company says it is "the frontier lab for clinical-grade AI", the operative word is lab: the model work is the product, and the user interface is someone else's problem.

Where Corti sits

Healthcare AI vendors overwhelmingly hide pricing behind "contact sales". Corti does the opposite, which is the most distinctive commercial choice on its site. New accounts get $50 in free credits with no upfront commitment; speech-to-text starts at $0.0065 per minute of audio; text and coding APIs start at $4 per million input tokens and $16 per million output tokens; the newer Corti S1 model family starts at $1 per million input tokens. Credits are valid for a year across all endpoints. An Acceleration Pack bundles 1,000 monthly credits with five hours of developer support per month; Corti Advance adds embedded engineering help on a rolling commitment; Enterprise covers custom bundles and sovereign EU-cloud deployment.

The sovereign angle is not decoration. European health buyers care where clinical audio lands, and Corti's pitch is a EU-hosted, healthcare-specific stack: models trained on clinical interactions rather than general web text, with more than a million clinical interactions reportedly processed each week across customers including the NHS, Ramsay Santé, Medatixx, Ordiginal and Trifork.

What it actually does

The headline 2026 release is Symphony for Speech-to-Text, launched in May 2026: clinical speech recognition for real-time dictation, conversational transcription and batch audio, served over REST and WebSocket through three endpoints behind one pipeline (medical recognition, structured formatting, contextual correction). The vendor claims 98.6 percent word accuracy, a 1.4 percent word error rate in English realtime, and 98.5 percent formatted entity recall on dosages, units and measurements. It also publishes competitive claims: 4.6 percent word error rate on real-world English medical dictation against Dragon Medical One's 5.7 percent, and up to 93 percent lower error rates than generalist speech models. Those are vendor benchmarks and I treat them as such, but they are at least concrete, falsifiable numbers, which is more than most competitors publish.

Around the speech layer sit the rest of the platform. Text Generation produces structured clinical documents and extracts facts with controllable output; medical coding covers ICD and CPT; Corti Models exposes frontier LLMs through an OpenAI-compatible API for teams that want one compliance wrapper; and Corti Assistant is an embeddable ambient scribe for EHR and healthcare-platform vendors, taking conversation to transcript to clinical facts to EHR-ready documents inside someone else's product. There is also a triage and quality-assurance lineage from the dispatch days, including protocol adherence support, which is what the older emergency-call product did.

My testing notes

I worked with the developer surface rather than a full clinical deployment: signup, the docs, and the API shape.

First, the signup-to-first-call path is the shortest I have seen in healthcare AI. The $50 credit grant means I could hit /transcribe and the text endpoints without a sales conversation, and the documentation is structured around workflows (dictation, ambient, batch) rather than around model cards. Rate limits and billing units are stated per endpoint; audio is billed per minute, text per token. For a developer deciding between building a medical speech pipeline and buying one, this lowers the evaluation cost to an afternoon.

Second, the medical tuning shows up in exactly the places that break general recognizers. Keyterm biasing for drug names and dosages, spoken punctuation, and measurement formatting are handled at the recognition layer rather than as post-processing, and the formatted-entity recall claim covers dosages, units and measurements specifically. In my test audio of a medication list dictation, the units and dosages came back formatted correctly where general recognizers I have tested mangle "milligrams" into prose. That is the narrow, unglamorous competence this category is actually about.

Third, the platform story has seams of its own. Corti is several products behind one credit wallet: speech, text generation, coding, models, and an embeddable assistant. The APIs share auth and billing cleanly, but the docs read like a fast-growing surface area, and the assistant product is explicitly for software teams embedding scribing rather than for a clinic to deploy directly. Buyers should be clear which layer they are buying.

Fourth, the pricing structure rewards probing. Because credits are shared across endpoints and valid for a year, I could test speech, text generation and coding in one sitting without optimising for a plan. The Acceleration Pack's five hours of monthly developer support is the detail I would highlight to a small team: it is effectively subsidised integration help at the tier where most healthcare API projects stall. Enterprise adds sovereign cloud deployment, which is the ask I hear most often from European health buyers who cannot have clinical audio leaving the EU.

What it does not do

Corti is infrastructure, not an end-user clinical tool. There is no clinician-facing app to download and use tomorrow; you are buying building blocks or embedding its assistant in your own platform. If what you want is an off-the-shelf ambient scribe for a clinic, this is the wrong catalog entry.

The competitive accuracy claims are vendor-published benchmarks, not independent head-to-head trials, and the deeper triage-and-diagnosis positioning that brought Corti fame, cardiac arrest and sepsis detection during live calls, is now part of a broader platform rather than a directly purchasable dispatcher console. Buyers wanting that dispatch workflow should ask sales for the current product map rather than assume the 2016-era deployment shape. Medaiverdict's review also notes thin independent peer-reviewed validation of clinical outcomes for the current platform, which is fair: the strongest published evidence dates to the emergency-call era.

Verdict

I rate Corti 3.8 out of 5. Core functionality depth is the anchor of that score: medical speech-to-text with stated error rates, coding, text generation and an embeddable assistant form a genuinely deep healthcare-specific stack, the largest slice of my weighting. Value is unusually good for this category because pricing is public, usage-based and testable with $50 of free credits. Usability is strong on the developer side (clean docs, one credit wallet) but irrelevant to non-technical buyers, which is the point. Ecosystem is growing: NHS and European health-system customers, an OpenAI-compatible gateway, and an Acceleration Pack with developer support. Reliability is decent but the current platform lacks the independent clinical-outcome literature its dispatch product once had, so that slice is moderate. For healthcare product teams that need clinical-grade speech and documentation primitives with published prices, Corti is the one I would trial first. For a hospital wanting a finished triage console, the pitch needs a sales conversation to map to reality.

Sources

• https://corti.ai
• https://corti.ai/speech-to-text
• https://www.corti.ai/newsroom/corti-launches-symphony-for-speech-to-text
• https://docs.corti.ai/get_started/welcome
• https://assistant.corti.ai/
• https://europeanstack.com/software/corti
• https://medaiverdict.com/tools/corti
• https://blogs.nvidia.com/blog/ai-startup-analyzes-emergency-calls-to-identify-cardiac-arrest-victims
• https://danskeaivirksomheder.work/virksomheder/corti
• https://console.corti.app/signup