Netease Youdao released the streaming ASR model Confucius4-R2T2

Netease Youdao company introduced Confucius4-R2T2 — a specialized model for automatic speech recognition (ASR) with low latency. The solution is developed based on the Qwen3-ASR architecture and is focused on high-precision processing of audio streams in real-time.

The model supports flexible adjustment of chunk size (chunking) in the range from 80 ms to 2 s, which allows adapting the system to the requirements of specific usage scenarios. The architecture supports append-only output mode, ensuring stable operation during data streaming. In performance tests, the model showed a result of 16 points with a processing time of 1 hour.

The technical implementation of Confucius4-R2T2 is optimized for working with Chinese and English languages. The developers provided native support for integration with vLLM, which simplifies the deployment of the model in high-load server environments. The source code and model weights are available for download on the Hugging Face platform.

The implementation of such tools allows reducing delays in voice control and transcription systems. The use of the optimized Qwen3-ASR engine makes it possible to integrate the model into existing audio processing pipelines without the need for deep infrastructure overhaul.