Confucius4-R2T2
A low-latency, streaming automatic speech recognition model that commits transcribed text permanently as it arrives, without revising earlier words, for uses such as live captioning and voice-driven agents.
Confucius4-R2T2, short for Real Real-Time Transcription, is a streaming automatic speech recognition model built on Qwen3-ASR. It runs in an append-only mode, committing transcript text permanently rather than revising it later, which avoids the flickering and disruptive corrections that some streaming transcription systems show, and is useful for applications such as live captioning, feeding downstream language model agents, and simultaneous speech translation.
It is trained with data-construction techniques including stable-prefix data, forced time-alignment data, and token-level audio segmentation, combined with a learning approach that decides when a stable prefix of the transcript can safely be emitted versus when more audio context is still needed. Decoding chunk size is configurable from 80 milliseconds to 2 seconds, trading latency against accuracy, and the model remains competitive with offline recognition despite its streaming design.
It supports a high-throughput vLLM backend as well as a Hugging Face Transformers backend, is optimized for Chinese and English with broader multilingual support, and natively accepts context and hotword prompts.