Open voice models that hear how you speak.
We build the listening and speaking layers for full-duplex voice agents: knowing when a person has finished, when they want to cut in, and what their tone, stress and laughter mean. Our voice stack already runs in 300+ AI hardware products.
| What it is | Status | |
|---|---|---|
| DuplexJev | Speech decision model for voice agents: turn-taking, barge-in, emotion, stress, laughter and more, read from audio in one forward pass (10 decisions in ~0.09 s). #1 on public paralinguistic benchmarks (LISTEN, StressTest, MSPB). 32B and 4B, served with vLLM. | Open weights · paper · demo |
| Sophon-Realtime | Open full-duplex voice model that hears your tone and speaks with one: laughs, sighs and stresses words when the conversation calls for it. #1 among open full-duplex models on VoiceBench (82.8) and Big Bench Audio (81.8%). Tag and Flash builds run on one GPU with vLLM-Omni; OpenAI GPT-Live compatible Live mode. | Open weights · HF · ModelScope · pip install sophon-realtime |
| heteronym-bench | TTS heteronym reading benchmark (English / Japanese): test sets, automatic judges and baselines. | Open |
- Models: huggingface.co/adventists-ai
- Website: adventists.cn
- Questions: DuplexJev Discussions
- Contact: jiejin@adventists.ai
DuplexJev is joint work with Zhejiang University and the University of Virginia.