NetEase Youdao open-sources two real-time interpretation models
NetEase's education arm put its simultaneous-interpretation stack into the open — a 200-millisecond speech recognizer and a 14B translation model — while researchers published a five-level ladder for the thing everyone keeps arguing about: AI that improves itself.
NetEase Youdao released and open-sourced two models built for simultaneous interpretation: Ziyue R2T2, a true streaming speech recognition model, and Ziyue4-T3PO, a 14-billion-parameter text-to-text simultaneous machine translation model. R2T2's average latency is 200 to 600 milliseconds, and Youdao says both its latency and its recognition quality reach state of the art among open models — competitive with closed ones. The training recipe is the interesting part: stable-prefix data, forced time-alignment data and token-level audio segmentation, plus a longest-stable-prefix learning scheme that keeps offline recognition accuracy intact while the model still outputs as a stream.
T3PO is fully streaming and exposes a latency dial — low-latency through high-quality modes depending on the setting — and pairs with an external streaming ASR to run speech-to-text live. In a live demo the pair translated unusually fast speech into Chinese in near-real time, which is the case most interpretation systems decline to demonstrate.
Why it matters: latency is the metric that decides whether simultaneous interpretation is usable, and it has been the part of the stack labs kept closed while advertising translation quality. Youdao is giving away the recognizer and the translation model and monetizing the product layer above them — the reverse of the API-first playbook. The 200-to-600-millisecond window is a company claim; nobody outside NetEase has reproduced it yet.
A large multi-institution team posted a survey arguing the field needs a shared vocabulary for recursive self-improvement — and supplied one, as a five-level ladder. "The Last AI Built by Humans: Toward Genuine Recursive Self-Improvement" defines RSI as a system turning experience and feedback into persistent changes to itself across interaction rounds, then grades autonomy: L1 executes a human-designed improvement procedure; L2 chooses how to improve a specified target; L3 lets the learner's evolving state shape what it learns next; L4 keeps reusable memory or skills that change later behavior; L5 rewrites the improvement process itself, or recursive meta-improvement.
The team sorted 491 existing papers onto that ladder and found the field still shallow: roughly 75 percent land at L1 or L2, under 6 percent reach L5. It also introduces a "Headroom-Closed Index" to show where current models still depend on costly, externally coordinated interventions, and argues coding is the clearest path to genuine RSI because fixes can be tested instantly, while robotics, science and medicine face slow, expensive feedback. A taxonomy paper is not a capability claim — but it is exactly what the past two weeks of RSI argument needed, and we covered how unsettled that fight still is in Google won't confirm the RSI rumor.
What to watch: independent latency benchmarks for R2T2 on fast, accented speech, and whether any lab publishes a self-assessment against the L1–L5 ladder. Self-scoring would be the tell.
If self-improvement gets graded on autonomy levels, who should get to write the rubric? Tell us in the comments.
