AI Briefing
KOSign in

NetEase Youdao releases Confucius4-R2T2, a stable-prefix streaming ASR model

·2026.11.09 23:13

Key point

Adapts Qwen3-ASR-1.7B with Longest Stable Prefix learning on 1,500 hours of audio to enable monotonic, real-time transcription.

Details

  • NetEase Youdao releases Confucius4-R2T2, a low-latency, high-accuracy stable-prefix streaming automatic speech recognition (ASR) model.
  • Supports configurable decoding chunks from 80 ms to 2 s and operates in a monotonic output mode that commits transcript text permanently without revising previous words.
  • Adapts Qwen3-ASR-1.7B using Longest Stable Prefix (LSP) learning on 1,500 hours of Chinese and English audio.
  • Two-stage training with SFT and GRPO updates only the language decoder, keeping the pre-trained audio encoder and projector frozen with no architectural changes.
  • LSP targets and an explicit Wait action teach the model what it can commit and when it needs more audio.
  • Code is released under Apache License 2.0, while model weights are released under the NetEase Model Use License Agreement.
open-sourcestreamingqwen3-asrasrspeech-recognitionnetease-youdao
Read originalShare on Threads

This summary was generated automatically by AI. Check the original for the author's claims and context. Copyright belongs to the original author.

Our guide explains how the AI works. Report summary errors, attribution issues, or removal requests via Contact.

Related

RedditSalesforce releases SOV-Transcribe-1, an open-weight speech recognition modelRedditadventists-ai/Sophon-Realtime-v1 · Hugging FacePyTorchKRjev-chat: Uses Jev model to assess counterpart intent and inspect LLM reply candidatesPyTorchKRllama.cpp v1.0.0, First Major Release in Three and a Half YearsGeekNewsGitHub Next releases Ace, a local collaborative coding environment for humans and AI agentsCloudflare BlogCloudflare open-sources Kitesurf, its stateless browser for AI agents