New speech recognition model promises to end streaming text stutter

Most streaming ASR models spit out text quickly, but they tend to self-correct after the fact — words already on screen suddenly change when more audio context arrives. Confucius4-R2T2, a 1.7-billion-parameter model released open-source, was built specifically to prevent that behavior. It operates in an append-only mode: once text is emitted it stays locked, which is critical when the output feeds directly into voice agents, live captions, or downstream NLP pipelines. The model is based on Qwen3-ASR and adds a training paradigm called Longest Stable Prefix, which lets it decide dynamically when a text prefix is stable enough to emit and when it needs more audio.

Under the hood the team used three dedicated data-construction techniques: samples with stable prefixes, forced time-alignment, and token-level audio segmentation. Combined with the LSP paradigm — a detailed technical report is forthcoming and has not yet been peer-reviewed — the model learns to expose only prefixes it is confident about, thereby supplying quality context for the next prediction without risking what has already been written. The approach enables flexible decoding chunks ranging from 80 milliseconds to 2 seconds, so the latency-accuracy trade-off can be tuned to the use case.

According to the developers' figures, R2T2 achieves accuracy approaching that of full offline recognition with average latency of 200 to 600 milliseconds. On a Pareto plot comparing English WER and Chinese CER against latency (chunk-wise mean fuzzy latency, measured retrospectively), the model places itself in the preferred bottom-left quadrant among open models and claims competitiveness with leading closed systems. A side-by-side visual comparison against GPT-Live-Transcribe on the same audio shows the competitor correcting backward while R2T2 keeps its text stable. The results are self-reported; the developers have invited other model maintainers to challenge or verify them via the repository's issue tracker.

The package ships with a vLLM backend for high-throughput inference and a Hugging Face Transformers backend for the standard path. Both support offline and true streaming modes. The model natively supports context prompts and hotwords, targets primarily Chinese and English but covers a range of additional languages, and the recommended installation route is Docker with a pre-configured CUDA environment, with isolated Conda or uv environments offered for local development.

What is missing from the picture is the technical report that will explain the LSP paradigm in depth — it has not yet been published, so the community will have to wait to evaluate the method thoroughly. In the meantime the code and weights are available on GitHub and Hugging Face, and the open invitation for external verification is a healthy step when SOTA claims are involved. If you are building a product that requires stable real-time text without flicker or corrections, it is worth running an independent evaluation on your own data before committing.