aifollow.news
返回 IT之家 科技新闻
IT之家 科技新闻· · 原发布时间 AI 评分73

微软将 MAI-Transcribe-2 用于 Dragon Copilot 医患对话转写

AI 辅助摘要

微软称,Dragon Copilot Transcribe 已全面采用 MAI-Transcribe-2,每月处理约 1000 万次患者就诊场景,多语言工作流的错误率下降约 50%。该模型支持语言混说识别、说话人区分及逐字或精简输出;每小时音频 0.10 美元为限时价格。

正文 · AI 翻译

IT Home reported on November 18 that at the Ignite 2026 event held yesterday (November 17), Microsoft announced that its healthcare AI product Dragon Copilot Transcribe has officially adopted its in-house speech recognition model, MAI-Transcribe-2.

Microsoft says the model is currently the fastest and most accurate multilingual speech transcription model and now handles 100% of patient-clinician conversation transcription in Dragon Copilot.

In terms of deployment at scale, Dragon Copilot Transcribe processes about 10 million patient visits each month and has reduced error rates by about 50% in multilingual workflows. About one-third of physicians across the United States are already using it. Microsoft says the model can recognize multiple languages and common code-switching scenarios, such as Chinese-English and Hindi-English, while maintaining high recognition quality in noisy environments.

In practice, this means consultation recordings can be converted into structured clinical records faster and more accurately, reducing the time spent manually organizing medical records, checking terminology and translating information. For multilingual patients, more reliable transcription can also help reduce the risk of medical information being distorted by language recognition errors.

MAI-Transcribe-2 supports speaker diarization, word-level timestamps, keyword biasing and configurable transcription styles. It can also detect languages automatically and offers two output modes, “verbatim” and “concise,” for different needs such as compliance archiving, analysis and medical record writing.

Microsoft says MAI-Transcribe-2 achieved leading performance across 60 languages in the FLEURS benchmark, with an average word error rate of 5.2%; it also ranks among the leaders in Artificial Analysis’s combined assessment of accuracy and latency. The company also says the model can process long audio up to 10 times as fast as some competing products and has introduced a limited-time price of US$0.10 per hour of audio (IT Home note: approximately 0.67 yuan at current exchange rates).

发现内容有误?提交纠错