Metadata-Version: 2.4
Name: sophon-realtime
Version: 1.0.0
Summary: Sophon-Realtime v1: open full-duplex voice model that hears your tone and speaks with one (Tag build, one GPU, vLLM-Omni)
Author: Adventists AI
License-Expression: Apache-2.0
Project-URL: Homepage, https://github.com/adventists-ai/sophon-realtime
Project-URL: ModelScope, https://www.modelscope.cn/models/AdventistsAI/Sophon-Realtime-v1
Project-URL: Hugging Face, https://huggingface.co/adventists-ai/Sophon-Realtime-32B-Tag
Requires-Python: >=3.10
Description-Content-Type: text/markdown
License-File: LICENSE
License-File: NOTICE
Requires-Dist: vllm==0.30.0
Requires-Dist: vllm-omni==0.30.0
Requires-Dist: duplexjev-vllm==0.3.0
Requires-Dist: modelscope
Provides-Extra: trt
Requires-Dist: tensorrt-cu13==10.13.3.9; extra == "trt"
Dynamic: license-file

# sophon-realtime

**Sophon-Realtime v1**: an open full-duplex voice model that hears your tone and speaks with one. This package installs the recommended **Tag** build as a vLLM-Omni plugin, plus a small command line to download and serve it on one 80 GB GPU.

```bash
pip install sophon-realtime          # Python 3.12, Linux, CUDA; installs vLLM 0.30 + vLLM-Omni 0.30
sophon-realtime serve                # downloads Sophon-Realtime-32B-Tag (~42 GB) on first run, then serves on port 8000
```

- Full duplex: `ws://HOST:8000/v1/realtime` (turn-taking, barge-in, back-channels; OpenAI-Realtime-style)
- `POST /v1/chat/completions` (audio in, text + audio out)
- Options: `--port`, `--gpu`, `--model DIR`, `--source modelscope|hf`; `sophon-realtime download` only downloads.
- Optional TensorRT for faster speech: `pip install "sophon-realtime[trt]"`.

VoiceBench 82.8 and Big Bench Audio 81.8% (#1 among open full-duplex models), #1 in paralinguistic understanding, NVV-SuperBench 0.86 (zh) / 0.80 (en). Client examples, the Flash build, Live mode (OpenAI GPT-Live compatible) and fine-tuning: https://github.com/adventists-ai/sophon-realtime

Apache-2.0. Third-party notices: NOTICE.
