Corti launches clinical-grade speech model beating OpenAI in accuracy

🔑 增強重點摘要

•Corti's Symphony platform extends beyond speech-to-text to include "Symphony for Medical Coding," an agentic AI system that treats coding as a reasoning task and has demonstrated over 25% higher clinical accuracy than generalist models from OpenAI, Anthropic, Amazon, Oracle, and Google in clinical accuracy benchmarks.
•The underlying research for Corti's medical coding approach, named "Code Like Humans," is a multi-agent framework accepted at EMNLP 2025, one of machine learning's top conferences, and was developed from the largest medical coding study of its kind, involving 5.8 million patient encounters.
•Corti's Symphony for Speech-to-Text is designed as an API-level infrastructure for speech-enabled clinical workflows, offering endpoints for stateless real-time dictation, stateful conversational transcription, and asynchronous batch processing, with features like command-and-control and contextual correction.
•Corti's Symphony significantly outperforms OpenAI's own tailored clinical product, ChatGPT for Clinicians, and other major LLM providers on the "HealthBench Professional" benchmark, particularly in clinical reasoning and safety under adversarial 'red teaming' conditions.
•The models are trained on a vast corpus of medical data, combining publicly available and proprietary speech/text-to-speech data, and are augmented with synthetically generated examples to enhance robustness and coverage of rare medical terminology, abbreviations, medications, and dosages.

📊 競品分析▸ Show

Feature/Benchmark	Corti Symphony	OpenAI (Generalist)	Whisper	Anthropic	Amazon	Google	ElevenLabs	Parakeet	ChatGPT for Clinicians
Speech-to-Text (WER on Medical Terminology)	1.4%	17.7%	17.4%	N/A	N/A	N/A	18.1%	18.9%	N/A
Medical Coding (F1 Accuracy)	0.74 (74%)	0.48 (48%)	N/A	0.59 (59%)	0.58 (58%)	0.43 (43%)	N/A	N/A	N/A
Agent Reasoning (HealthBench Professional Score)	60.5%	48.1% (GPT-5.4)	N/A	N/A	N/A	N/A	N/A	N/A	59.0%
Red Teaming (Safety Score on HealthBench)	87.7%	30.3% (GPT-5.4)	N/A	N/A	N/A	N/A	N/A	N/A	N/A

🛠️ 技術深入

Symphony for Speech-to-Text decomposes the transcription process into specialized components for recognition, formatting, and contextual correction to optimize medical term recall and produce clinically structured text.
It supports three API endpoints for different use cases: /transcribe for stateless real-time dictation, /streams for stateful real-time conversational transcription, and /transcripts for asynchronous batch audio processing.
The system's common processing pipeline performs audio ingress and routing, diarization, transcription, formatting, contextual correction, and generates structured transcripts.
Symphony for Medical Coding is built on a multi-agent framework called "Code Like Humans," which employs four sequential agents: an evidence extractor, an index navigator, a tabular validator, and a code reconciler, mirroring the decision process of expert human coders.
This medical coding architecture reasons from codified logic rather than relying solely on patterns learned from training data, enabling it to support various coding systems (e.g., ICD-10-CM, ICD-10-PCS, CPT, ICD-10-UK, ICD-10-International) without requiring local retraining.
Corti's broader Agentic Framework is a modular AI system designed for developers to build advanced AI agents for clinical and operational tasks, addressing limitations of general LLMs by enabling reliable access to clinical data through external tools and providing a controlled execution layer for safe interaction with real systems.

🔮 前景展望AI analysis grounded in cited sources

Specialized AI models will increasingly dominate niche, highly regulated industries like healthcare.

Corti's significant outperformance over generalist models in medical speech-to-text, coding, and clinical reasoning demonstrates the critical value of domain-specific training and architectural design for accuracy and safety in healthcare.

The 'agentic era' of healthcare AI will be driven by multi-agent systems that mimic human reasoning processes.

Corti's success with its multi-agent 'Code Like Humans' framework for medical coding, which treats coding as a reasoning task, indicates a shift towards more sophisticated AI architectures that can interpret complex clinical contexts and guidelines.

Improved clinical-grade AI will lead to more accurate public health data and better resource allocation.

Corti's ability to identify previously missed diagnoses and conditions, such as suicide attempts, through more accurate coding highlights how enhanced AI can provide a more reliable data layer for epidemiological research, public health surveillance, and resource allocation decisions.

⏳ 時間線

2014

Corti founded in Copenhagen, Denmark.

2015-12

Corti's first funding round.

2023-09

Secured $60 million in Series B funding, co-led by Prosus Ventures and Atomico.

2025

The 'Code Like Humans' multi-agent framework, foundational to Symphony for Medical Coding, accepted at EMNLP 2025.

2026-04-01

Launched Symphony for Medical Coding, an agentic AI model for medical coding automation.

2026-05-20

Launched Symphony for Speech-to-Text, a clinical-grade speech model for healthcare environments.

Corti 推出醫療級語音模型，準確度超越 OpenAI

⚡ 30-Second TL;DR

關鍵要點

🧠 深度解析

🔑 增強重點摘要

🛠️ 技術深入

🔮 前景展望AI analysis grounded in cited sources

⏳ 時間線

📎 來源 (17)

👉相關動態

GM 透過圍繞 AI 代理重新設計工作流程，使合併請求數量成長三倍

Runway 將 AI 影片模型的錯誤轉化為新功能