This new 2B open-source speech model transcribes in real time with 200ms latency 🤯 It’s called Confucius R2T2 by NetEase Youdao and commits each word the moment it's recognized and never rewrites it. 100% open source.

Almost every streaming ASR system is built to emit something at every step. When the audio is ambiguous, it falls back on language priors and produces its best guess, because staying silent isn't really an option in the architecture. R2T2 introduces an explicit wait state instead. If there isn't enough acoustic evidence yet, it holds. When evidence arrives, it commits.

What I find interesting is how this reframes the latency conversation. The field keeps chasing "predict faster." This design argues the harder skill is knowing when a prediction is premature and would cost you more downstream than a short pause. Link: Open source, vLLM backend. Built for voice agents.

Here are the links to the repo: and Hugging Face:

Awesome find! Benchmarking it to see if it's a good fit for Fluidvoice and my users ;)

What is the inference cost ? we have already NLLB model there right ? so what is the special about this model ?

Nice real time transcription, @Build_Vertical.

Wow, thanks for the information 🤯

the 'never rewrites' claim is the real tradeoff here. most streaming models correct earlier words as context rolls in. genuinely asking how WER compares to Whisper on phonetically ambiguous audio

200ms immutable token emission cuts voice agent turn-taking latency in half.

Not completely open source. Companies with 1B revenues need a specific license.

damn that's awesome, trying it right away

the wait state is the clever bit guessing less can beat speaking sooner

How can we get it?

Here:

big if true, esp for ai voice agents. how is it compared to deepgram in terms of accuracy?

Committing each word without rewriting is a bold design. How does it handle names or phrases that only become clear from later context?

That's great!

finally an open source model that actually stays bullish