GitHub radar
Youdao’s model transcribes your speech in real time
A streaming ASR model from NetEase Youdao with less than 1 sec latency. Free installation to transcribe speech in near realtime.
It’s a low-latency and high-accuracy true streaming Automatic Speech Recognition (ASR) model called Confucius4-R2T2. It’s created by NetEase Youdao, an online Chinese dictionary and translator. Confucius4-R2T2 transcribes the audio stream in real-time at frame level. Each frame has a length between 80ms and 2s. It doesn’t need to wait for the whole sentence to end and doesn’t post-edit the results. With an average latency of 200–600ms, Confucius4-R2T2 achieves similar performance compared to the offline ASR model. Confucius4-R2T2 is built upon Alibaba’s Qwen3-ASR (a 1.7 billion parameter model) and finetuned by Youdao team. Besides Chinese and English, it also supports about a dozen other languages including Russian, etc. Confucius4-R2T2 has about 2 billion parameters and can run on your PC.
Why a vibe-coder should care
Confucius4-R2T2 helps us process our contents, especially for subtitling and live captioning purposes. Because it provides subtitles with low latency (<1 sec), we can use it in live streaming and quickly transcribe videos. Low latency is one of the strengths of Confucius4-R2T2. Most streaming ASR models have high latency. However, Confucius4-R2T2 is mainly optimized for Chinese and English. So if you want to use it for other languages (e.g., Russian), there’s no assurance about its accuracy. Feel free to install it on your own. No one’s API key is needed. But you need to set up a Docker environment and make some modifications.
How to install
Copy this and send it to your agent — Claude Code, Codex, any of them:
Set up the speech recognition model Confucius4-R2T2 by NetEase Youdao for me: https://huggingface.co/netease-youdao/Confucius4-R2T2 — deploy it locally following the card's instructions (via Docker) and show me how to send it audio and get text back almost in real time.
A regular laptop with 8 GB of RAM or more — the model itself takes about 4 GB.
Open on Hugging Face▌ More finds