Open-source speech recognition toolkit for training, inference, streaming ASR, VAD, punctuation, speaker diarization pipelines, and OpenAI-compatible/MCP serving.
Python asraudiochineseemotion-recognitionfunasrmcp-servermultilingual-asropenai-compatible-apiparaformerpunctuationpytorchreal-time-asrspeaker-diarizationspeech-recognitionspeech-to-textstreaming-asrtranscriptionvllmvoice-activity-detectionwhisper-alternative
docs(vllm): clarify native realtime handshake boundary
LauraGPTa047362837e3
diff --git a/docs/vllm_guide_zh_v2.md b/docs/vllm_guide_zh_v2.mdindex ad0c76208..1e5bbf75e 100644--- a/docs/vllm_guide_zh_v2.md+++ b/docs/vllm_guide_zh_v2.md@@ -100,7 +100,10 @@ native checkpoint 传给 `AutoModelVLLM` 或官方 `model.pt`、`config.yaml` 与 `Qwen3-0.6B/` 目录,因而会拒绝 native layout。原生 `vllm serve` 的转写 API 不会自动获得 FunASR WebSocket 服务的 VAD、partial-预览、会话状态或说话人处理。需要实时行为时,请按 vLLM 的+预览、会话状态或说话人处理,因此 native Fun-ASR 模型没有可用的实时转写会话。+即使对 `/v1/realtime` 发起 WebSocket 握手,也不能据此判断模型支持实时转写;当前+vLLM 服务会拒绝这类握手,Uvicorn 通常显示 `403 Forbidden`。这是当前实现的拒绝表现,+不是稳定的 API 契约,也不应据此编写支持性探测。需要实时行为时,请按 vLLM 的[realtime speech-to-text 示例](https://docs.vllm.ai/en/latest/examples/speech_to_text/realtime/)在应用侧实现音频分段、请求调度与可替换预览;需要 FunASR 已实现的 WebSocket协议时,继续使用上文的官方 checkpoint + `serve_realtime_ws.py` 路径。diff --git a/tests/test_vllm_model_source_docs.py b/tests/test_vllm_model_source_docs.pyindex 14c43c3ff..578e2202c 100644--- a/tests/test_vllm_model_source_docs.py+++ b/tests/test_vllm_model_source_docs.py@@ -39,6 +39,8 @@ def test_v2_guide_keeps_native_vllm_and_funasr_realtime_paths_separate():"vllm serve FunAudioLLM/Fun-ASR-Nano-2512-vllm --port 8000","serve_realtime_ws.py","realtime speech-to-text",+ "没有可用的实时转写会话",+ "不是稳定的 API 契约",]for marker in required_markers: