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
feat: finalize vLLM service + benchmark matrix
游雁6a18244602df
diff --git a/docs/vllm_guide.md b/docs/vllm_guide.mdindex 7ad6daf57..6ca5f726c 100644--- a/docs/vllm_guide.md+++ b/docs/vllm_guide.md@@ -371,20 +371,24 @@ Client Server### Benchmark 结果(184 files, 11541s audio, 1947 VAD segments)-| 模型 | 方法 | RTFx | CER | 加速 |-|------|------|------|-----|------|-| Fun-ASR-Nano | PyTorch (fixed VAD) | 19.6 | 8.94% | 基准 |-| Fun-ASR-Nano | **vLLM batch (fixed VAD)** | **393.9** | **8.91%** | **20.7x** |-| Fun-ASR-Nano | **vLLM server (dynamic VAD)** | **68.3** | **8.14%** | **3.5x** |-| Fun-ASR-Nano | yuekaizhang vLLM | 273.0 | 17.07% | 14.3x |-| GLM-ASR-Nano | PyTorch | 34.4 | 12.94% | 基准 |-| GLM-ASR-Nano | **vLLM (ours)** | **263.2** | **12.92%** | **7.6x** |+| 模型 | 引擎 | VAD | RTFx | CER | 备注 |+|------|------|-----|------|-----|------|+| Fun-ASR-Nano | PyTorch | fixed | 17.1 | 8.90% | 基准 |+| Fun-ASR-Nano | PyTorch | dynamic | 21.2 | 8.06% | |+| Fun-ASR-Nano | **vLLM batch** | fixed | **371** | 8.88% | 21.7x 加速 |+| Fun-ASR-Nano | **vLLM batch** | dynamic | **340** | **8.20%** | 19.9x 加速 |+| Fun-ASR-Nano | **Service (no SPK)** | dynamic | **102** | 8.14% | 含 VAD + 时间戳 |+| Fun-ASR-Nano | **Service (with SPK)** | dynamic | **46** | 8.19% | 含 VAD + 时间戳 + 说话人 |+| Fun-ASR-Nano | yuekaizhang vLLM | fixed | 273 | 17.07% | 第三方实现 |+| GLM-ASR-Nano | **vLLM batch** | fixed | **265** | 12.93% | 7.6x 加速 |+| GLM-ASR-Nano | vLLM batch | dynamic | - | - | 不支持长音频推理 |**关键结论:**-- vLLM batch 与 PyTorch CER 完全一致(差 < 0.05%)-- 新动态 VAD(保留长段 ≤60s)CER 降至 **8.14%**,优于 PyTorch 的 8.94%-- Fun-ASR-Nano batch 加速 20.7x,服务模式加速 3.5x-- 比第三方 yuekaizhang/Fun-ASR-vllm 快 44%,且 CER 优 9%+- vLLM 与 PyTorch CER 完全一致(Fun-ASR-Nano 差 < 0.02%)+- 新动态 VAD 对 Fun-ASR-Nano 有效(CER 8.90% → 8.20%),不适用于 GLM-ASR(不支持长音频)+- Fun-ASR-Nano vLLM batch 加速 21.7x,Service 模式加速 6x+- SPK 默认关闭(RTFx 102),开启后 RTFx 46(增加说话人聚类开销)+- 比第三方 yuekaizhang/Fun-ASR-vllm 快 36%,且 CER 优 8%### 统一 vLLM 服务(serve_vllm.py)