Confucius4-R2T2 introduces stable-prefix streaming ASR with high accuracy and low latency, enabling real-time transcription without output revisions.
Researchers have unveiled Confucius4-R2T2, a stable-prefix streaming automatic speech recognition (ASR) model designed for real-time transcription. Published on November 10, 2026, on arXiv and highlighted on HuggingFace Daily Papers, the model introduces a significant advancement in streaming ASR by ensuring that once text is emitted, it remains immutable, all while maintaining high accuracy and low latency.
Confucius4-R2T2, short for Real Real-Time Transcription, adapts the Qwen3-ASR-1.7B model through post-training on 1,500 hours of Chinese and English audio. The training process employs the Longest Stable Prefix (LSP) learning method, a novel concept introduced by the researchers. LSP ensures that the model's output is monotonic, consistent with full-audio transcription, and maximally stable. This approach allows the model to convert from a full-context ASR to a stable-prefix streaming ASR without architectural modifications.
The training process freezes the pre-trained audio encoder and projector, updating only the language decoder through two-stage fine-tuning using Supervised Fine-Tuning (SFT) and Gradient Projection Optimization (GRPO). Additionally, the model incorporates an explicit 'Wait' action to determine when more audio is needed before committing to a transcript.
Confucius4-R2T2 delivers substantial improvements in streaming recognition accuracy. At a latency of 160 milliseconds with one provisional token, the model achieves a recognition error rate of 5.76%, significantly outperforming the base model's 21.01% error rate under the same conditions. This positions R2T2 as a leading open-source solution for low-latency, high-accuracy transcription.
The model's stable-prefix output is particularly beneficial for applications requiring immediate and reliable text processing, such as live captioning, simultaneous speech translation, and downstream natural language processing pipelines.
Source link
