🎧 DuplexJev — stop transcribing to decide

The decision layer for voice agents. Give it a clip and the questions your agent needs answered — has the user finished? should I stop talking? are they angry? Every answer comes from one forward pass: no ASR, no decoding. 10 decisions in 92 ms, 21× faster than ASR → LLM (paper).

⭐ GitHub · 🤗 Models · 🔌 Pipecat: pip install pipecat-duplexjev · LiveKit: pip install livekit-plugins-duplexjev · 👁️ New: voice + picture together → DuplexJev-Vision demo · 🏢 Commercial licence & hosted API: jiejin@adventists.ai

⏱️ The first run after the Space has been idle takes a few seconds while Hugging Face attaches a GPU; after that each run takes about 0.3 s on the GPU. Free GPU time per visitor is limited by Hugging Face — log in for more runs. 首次运行需要几秒分配 GPU，之后每次约 0.3 秒；登录 Hugging Face 可获得更多免费次数。
Speech language (picks the question wording) / 语音语言
🇬🇧 English speech examples
🇨🇳 Chinese speech examples · 中文语音示例 (answers in English · 中文)

Results appear here

Model: DuplexJev-32B-Para (weights CC BY-NC 4.0), one full ZeroGPU card. Audio is processed for this request only and not stored. Edit the questions freely: one per line, name | question | option 1; option 2; …. GPU time shown excludes the queue and network. 音频只用于本次判断，不保存。