Create

Sign in to ReadmeX

or

How we handle your data: Privacy Policy

ReadmeX

Your AI reading companion

What would you like to catch up on?

Pick a briefing, or ask your own question.

Your community and people briefings stay personal to you.

Story

Confucius4-R2T2 Report Details Stable Streaming ASR

AI summary

NetEase Youdao published the technical report for Confucius4-R2T2, a real-time ASR model built on Qwen3-ASR-1.7B and post-trained on about 1,500 hours of Chinese and English audio. The report says its longest stable prefix supervision, explicit WAIT behavior, and streaming-submission strategy support incremental transcription without revising committed text. In the report’s evaluation, the quality-first configuration had a 5.76% macro-average error rate across 13 main test sets, though the benchmark results were not independently verified in the supplied reports. The source also reported that the model and its quantized versions had reached 200,000 global downloads four days after release.

In-depth insight

The approach gives developers a documented method for balancing recognition quality, output timing, and text stability in real-time captions, simultaneous interpretation, and voice applications.

开源仅4 天(9月21日),网易有道子曰 Confucius4-R2T2 就登顶 Hugging Face 自动语音识别(ASR)热榜第一。目前,模型及各量化版本全球累计下载量已达到20 万次。

围绕R2T2,社区陆续推出量化版本与不同运行环境的适配,让更多开发者能够体验高质量的实时语音转写能力。近日,Confucius4-R2T2 的技术报告也上线论文预印本平台arXiv。

这份关注背后,是实时语音应用对“快、准、稳”的共同需求。

对实时字幕、同声传译和语音智能体而言,识别结果不仅要准确,还要及时、稳定。一段已经交给下游处理的文字,如果随后被改写,就可能引发字幕跳动、译文重写或重复推理。

#Confucius4-R2T2#NetEase Youdao#Qwen3-ASR-1.7B#Hugging Face

Original · 9 min readRead the original
Group chat

No comments yet. Start the conversation.