NetEase Youdao Open-Sources Two Speech Models Topping Hugging Face Trending

Around September 23, NetEase Youdao open-sourced two voice models, both hitting #1 on Hugging Face's Trending charts in their respective categories: the streaming ASR model Confucius4-R2T2 and the Chinese-English simultaneous interpretation model Confucius4-T3PO. Multiple posters noted that the two combined form a fully open-source, locally deployable real-time speech translation/interpretation stack.

Confirmed

• R2T2: 2B parameters, fine-tuned from Qwen3-ASR-1.7B; true streaming ASR with configurable decoding chunks of 80ms–2s, 160ms latency, and near-offline accuracy; uses an append-only output mode (text is never revised once emitted), making it well-suited for live captioning and voice agents; licensed under a NetEase custom license.
• T3PO: 14B parameters, a Chinese-English simultaneous interpretation model capable of autonomous decisions (e.g., segmentation timing); released under Apache 2.0.
• Both model weights are available on Hugging Face; combined, they form a fully open-source, locally deployable real-time interpretation stack.

Why it matters

• Real-time speech translation has largely relied on closed-source services; R2T2 + T3PO offer the first fully open-source, locally deployable end-to-end solution, with direct practical value for developers of captioning, interpretation, and voice agents.
• R2T2's no-revision output design eliminates the "text jitter" common in streaming transcription—a key trade-off aimed at production use.

2026-09-23 ~ 2026-09-23 · 5 related posts

• [source] NetEase Youdao open-sources 2B streaming ASR and 14B Chinese-English simultaneous translation models — AdinaYakup · 2026-09-23
• [source] NetEase Youdao's open-source ASR and translation models top Hugging Face trending — thetripathi58 · 2026-09-23
• NetEase Youdao open-sources Confucius4-R2T2 and T3PO, a full real-time voice AI stack — thetripathi58 · 2026-09-23

2 near-duplicate retellings: AdinaYakup · thetripathi58