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
fix: respect explicit VAD silence threshold
游雁ef4605505664
diff --git a/docs/vllm_guide.md b/docs/vllm_guide.mdindex f180ea7c3..d6729640e 100644--- a/docs/vllm_guide.md+++ b/docs/vllm_guide.md@@ -45,6 +45,14 @@ cd /path/to/FunASR && pip install -e .**Hardware**: GPU ≥ 8 GB VRAM, CUDA ≥ 11.8. 16 GB+ recommended.+Install a PyTorch/torchaudio/vLLM combination that matches your NVIDIA driver and+CUDA runtime. Do not blindly keep the newest wheel if it was built for a newer+CUDA runtime than your driver supports; PyTorch can fail during CUDA+initialization with `The NVIDIA driver on your system is too old` before FunASR+starts. If that happens, reinstall compatible PyTorch, torchaudio, and vLLM+wheels for the CUDA version reported by `nvidia-smi`, or update the NVIDIA+driver first.+---## 2. vLLM Engine Architecturediff --git a/docs/vllm_guide_zh.md b/docs/vllm_guide_zh.mdindex bce42db5e..81e44c314 100644--- a/docs/vllm_guide_zh.md+++ b/docs/vllm_guide_zh.md@@ -44,6 +44,13 @@ cd /path/to/FunASR && pip install -e .**硬件**:GPU ≥ 8GB VRAM,CUDA ≥ 11.8。推荐 16GB+。+请根据 NVIDIA 驱动和 `nvidia-smi` 显示的 CUDA 版本选择匹配的+PyTorch/torchaudio/vLLM 组合,不要无条件保留 pip 拉到的最新 wheel。若+vLLM 或 PyTorch wheel 依赖的 CUDA runtime 高于当前驱动支持范围,可能在+FunASR 启动前就报 `The NVIDIA driver on your system is too old`。遇到该错误时,+优先重装与当前驱动/CUDA 匹配的 PyTorch、torchaudio、vLLM wheel,或先升级+NVIDIA 驱动。+---## 2. vLLM 推理引擎架构diff --git a/docs/vllm_guide_zh_v2.md b/docs/vllm_guide_zh_v2.mdindex 80a00c8d3..0e4a6ea2b 100644--- a/docs/vllm_guide_zh_v2.md+++ b/docs/vllm_guide_zh_v2.md@@ -45,6 +45,13 @@ cd /path/to/FunASR && pip install -e .**硬件**:GPU ≥ 8GB VRAM,CUDA ≥ 11.8。推荐 16GB+。+请根据 NVIDIA 驱动和 `nvidia-smi` 显示的 CUDA 版本选择匹配的+PyTorch/torchaudio/vLLM 组合,不要无条件保留 pip 拉到的最新 wheel。若+vLLM 或 PyTorch wheel 依赖的 CUDA runtime 高于当前驱动支持范围,可能在+FunASR 启动前就报 `The NVIDIA driver on your system is too old`。遇到该错误时,+优先重装与当前驱动/CUDA 匹配的 PyTorch、torchaudio、vLLM wheel,或先升级+NVIDIA 驱动。+---## 2. vLLM 推理引擎架构diff --git a/funasr/models/fsmn_vad_streaming/model.py b/funasr/models/fsmn_vad_streaming/model.pyindex 9874affd5..ae5f43aaa 100644--- a/funasr/models/fsmn_vad_streaming/model.py+++ b/funasr/models/fsmn_vad_streaming/model.py@@ -938,8 +938,10 @@ class FsmnVADStreaming(nn.Module):n = int(len(audio_sample) // chunk_stride_samples + int(_is_final))m = int(len(audio_sample) % chunk_stride_samples * (1 - int(_is_final)))segments = []- # Dynamic silence threshold- dynamic_silence = kwargs.get("dynamic_silence", True)+ # Keep explicit fixed-threshold requests from being overwritten by the dynamic schedule.+ dynamic_silence = kwargs.get(+ "dynamic_silence", kwargs.get("max_end_silence_time") is None+ )silence_schedule = kwargs.get("silence_schedule", DEFAULT_SILENCE_SCHEDULE)speech_to_sil_ms = self.vad_opts.speech_to_sil_time_thresaccumulated_ms = cache.get("_dynamic_accumulated_ms", 0)diff --git a/tests/test_fsmn_vad_dynamic_silence.py b/tests/test_fsmn_vad_dynamic_silence.pynew file mode 100644index 000000000..f396614f9--- /dev/null+++ b/tests/test_fsmn_vad_dynamic_silence.py@@ -0,0 +1,68 @@+import unittest+from types import SimpleNamespace+from unittest.mock import patch++import torch++from funasr.models.fsmn_vad_streaming import model as vad_model+++class TestFsmnVadDynamicSilence(unittest.TestCase):+ def _run_inference(self, **kwargs):+ vad = vad_model.FsmnVADStreaming.__new__(vad_model.FsmnVADStreaming)+ vad.vad_opts = SimpleNamespace(speech_to_sil_time_thres=100)+ vad.forward = lambda **batch: []++ cache = {+ "frontend": {},+ "prev_samples": torch.empty(0),+ "encoder": {},+ "stats": SimpleNamespace(+ vad_state_machine=vad_model.VadStateMachine.kVadInStateInSpeechSegment,+ max_end_sil_frame_cnt_thresh=200,+ speech_noise_thres=0.6,+ ),+ }+ frontend = SimpleNamespace(fs=16000, frame_shift=10, lfr_n=1)++ def fake_extract_fbank(*args, **kwargs):+ cache["frontend"]["waveforms"] = torch.zeros(1, 16000)+ return torch.zeros(1, 1, 80), torch.tensor([100])++ with (+ patch.object(+ vad_model,+ "load_audio_text_image_video",+ return_value=[torch.zeros(16000)],+ ),+ patch.object(vad_model, "extract_fbank", side_effect=fake_extract_fbank),+ ):+ vad_model.FsmnVADStreaming.inference(+ vad,+ torch.zeros(16000),+ frontend=frontend,+ cache=cache,+ key=["utt"],+ chunk_size=1000,+ is_final=False,+ device="cpu",+ **kwargs,+ )++ return cache++ def test_explicit_max_end_silence_time_keeps_fixed_threshold_by_default(self):+ cache = self._run_inference(max_end_silence_time=300)++ self.assertEqual(cache["stats"].max_end_sil_frame_cnt_thresh, 200)+ self.assertNotIn("_dynamic_accumulated_ms", cache)++ def test_explicit_dynamic_silence_still_enables_schedule(self):+ cache = self._run_inference(max_end_silence_time=300, dynamic_silence=True)++ self.assertEqual(cache["stats"].max_end_sil_frame_cnt_thresh, 1900)+ self.assertEqual(cache["_dynamic_accumulated_ms"], 1000)+++if __name__ == "__main__":+ unittest.main()