← Sticker Wall AGENT · DETAIL

Corti Emergency AI: Real-time telephone voice analysis with 95% accuracy in cardiac arrest detection, adopted by emergency dispatch centers across multiple European countries.

Workflow: The input is a real-time voice stream from an emergency call; processing is handled by speech recognition combined with

AGENT

Key Fields

FIELD STAMPS

IndustryContent / Creator Economy

RegionEurope

ScaleSME

ChannelOnline

🔧 Workflow

The input is a real-time voice stream from an emergency call; processing is handled by speech recognition combined with Corti's proprietary language model for context understanding and cardiac arrest pattern matching. Rule-wise, it only prompts the dispatcher and does not issue direct diagnoses; the output consists of call-risk warnings and structured records. The frequency is triggered in real-time per call, and human judgment is finalized by trained dispatchers. According to Corti's official statements, its system handles over 1 million interactions per week, and emergency dispatch scenarios can increase out-of-hospital cardiac arrest detection rates by 13%.

🛠 Setup Requirements

Building a similar system requires: speech recognition, natural language processing, medical knowledge graphs, model training experience, and a scalable cloud audio processing architecture. Tools include Deepgram or Whisper for speech transcription, Hugging Face or custom-built models for inference, and cloud service providers such as AWS or Azure for deployment. Development to launch takes 3-6 months, and the team must include medical advisors, AI engineers, and emergency operations experts.

🧰 Toolchain

🔧 Corti Symphony API
🔧 Deepgram Speech Recognition
🔧 AWS Lambda Real-time Audio Processing
🔧 Hugging Face Model Hub

💰 Revenue

① Emergency center enterprise deployment (primary company revenue): Emergency centers pay on a project or licensing basis, with annual deployment fees running into hundreds of thousands of euros. Combined annual revenue from multiple emergency centers such as Copenhagen reaches the tens of millions of dollars, with a company valuation of approximately $100 million (Provider metric $100M). The exact proportion of total revenue is unstated (media estimate, independently unverified, figures current up to 2026). ② Pay-per-call API usage: Developers pay-as-you-go, sold in batches starting at 1,000 credits; unit prices and usage volumes are not publicly disclosed (based on company disclosures), and the revenue share is unspecified. ③ Replicator integration service fees: Individual developers charge clinics and training institutions per project; neither pricing nor customer counts are available, and the percentage of total revenue is unmentioned. ④ Opportunity item - Emergency voice data annotation and evaluation: Charged per usage, exact revenue share yet to be determined.

💸 Cost

Main costs: GPU training and cloud inference fees (ranging from thousands to tens of thousands of dollars per month), medical data compliance and expert advisor consulting fees, and sales and implementation team salaries.

⏱ Time Investment

Core team full-time commitment, 8+ hours per day; based on project cycles, emergency center implementation from pilot to full launch takes 3-12 months.

🚀 Getting Started

Beginner guide: First, understand emergency workflows and medical terminology, and participate in Corti or similar open-source emergency datasets (such as SUDDEN cardiac arrest audio) for learning. Step two, build a prototype using Whisper plus a rule engine, test cardiac arrest keywords and breathing sounds using small samples, and then gradually incorporate model optimization.

🔑 Keys to Success

✅ Real-time and low latency: Deliver judgments within the first 90 seconds of the call connecting.
✅ Medical-grade accuracy: Requires clinical validation and regulatory compliance.
✅ Deep integration with emergency workflows: Does not replace humans; acts solely as an auxiliary prompt.
✅ Clinical interpretability and trust: The AI must clearly explain the basis of its judgment to emergency personnel (such as identifying specific breathing sounds or keywords) to gain front-line trust and adoption.

⚠️ 风险

⚠️ Medical liability risk: AI misjudgment may lead to legal disputes, necessitating clear final decision-making authority by humans.
⚠️ Data privacy risk: Emergency calls contain sensitive personal information and must comply with regulations such as GDPR.
⚠️ Model bias risk: Different accents, dialects, and background noise may affect recognition accuracy.

📌 Real Cases

📌 The Copenhagen Emergency Medical Services has used Corti since December 2016. Studies show that Corti's cardiac arrest recognition accuracy reaches 95%, compared to 73% for human dispatchers. The service has now expanded to emergency centers in multiple European countries, including the Netherlands and the UK.
📌 According to AgentMarketCap data, Corti's Symphony model stack is trained on over 1.5 million hours of clinical audio, scoring higher than OpenAI on the HealthBench Professional benchmark and providing real-time decision support for medical institutions worldwide.
📌 According to SmartCitiesHub, Corti originated in Copenhagen and, as of 2026, has expanded to emergency dispatch centers across multiple European countries, providing real-time emergency medical decision support and being integrated into urban smart healthcare infrastructure.

Sources