From e8ea59a02c9362ae78e6730ca574862f5aedd6c7 Mon Sep 17 00:00:00 2001
From: LauraGPT <lauragpt@users.noreply.github.com>
Date: Fri, 14 Aug 2026 07:35:21 +0000
Subject: [PATCH] fix(vad): preserve Silero ONNX CPU inputs

---
 funasr/models/silero_vad/model.py | 16 ++++++++--------
 tests/test_silero_vad_adapter.py  | 21 ++++++++++++++++++++-
 2 files changed, 28 insertions(+), 9 deletions(-)

diff --git a/funasr/models/silero_vad/model.py b/funasr/models/silero_vad/model.py
index c5f120d46..7f90ceeda 100644
--- a/funasr/models/silero_vad/model.py
+++ b/funasr/models/silero_vad/model.py
@@ -26,7 +26,8 @@ class SileroVad(torch.nn.Module):
                 '`python -m pip install "funasr[silero]"` or '
                 "`python -m pip install silero-vad`."
             ) from error
-        self.model = load_silero_vad(onnx=kwargs.get("silero_onnx", False))
+        self.onnx = bool(kwargs.get("silero_onnx", False))
+        self.model = load_silero_vad(onnx=self.onnx)
         self.get_speech_timestamps = get_speech_timestamps
 
     @staticmethod
@@ -34,8 +35,10 @@ class SileroVad(torch.nn.Module):
         if not max_single_segment_time:
             return segments
         limit_ms = int(max_single_segment_time)
-        if limit_ms < 0:
-            raise ValueError("max_single_segment_time must be non-negative")
+        if limit_ms <= 0:
+            raise ValueError(
+                "max_single_segment_time must resolve to a positive millisecond value"
+            )
         split = []
         for start, end in segments:
             while end - start > limit_ms:
@@ -60,11 +63,8 @@ class SileroVad(torch.nn.Module):
         started = time.perf_counter()
         results = []
         for index, audio in enumerate(audio_list):
-            waveform = (
-                torch.as_tensor(audio, dtype=torch.float32)
-                .flatten()
-                .to(self.anchor.device)
-            )
+            device = torch.device("cpu") if self.onnx else self.anchor.device
+            waveform = torch.as_tensor(audio, dtype=torch.float32).flatten().to(device)
             timestamps = self.get_speech_timestamps(
                 waveform,
                 self.model,
diff --git a/tests/test_silero_vad_adapter.py b/tests/test_silero_vad_adapter.py
index 2effe6f49..3399e1381 100644
--- a/tests/test_silero_vad_adapter.py
+++ b/tests/test_silero_vad_adapter.py
@@ -72,10 +72,29 @@ class TestSileroVadAdapter(unittest.TestCase):
         results, _ = model.inference(data_in=[torch.zeros(16000)], key=["sample"])
         self.assertEqual(results, [{"key": "sample", "value": []}])
 
+    @patch("silero_vad.get_speech_timestamps")
+    @patch("silero_vad.load_silero_vad")
+    def test_onnx_waveform_stays_on_cpu(self, load_model, timestamps):
+        load_model.return_value = object()
+
+        def timestamps_stub(waveform, model, sampling_rate, **options):
+            self.assertEqual(waveform.device.type, "cpu")
+            return []
+
+        timestamps.side_effect = timestamps_stub
+        model = SileroVad(silero_onnx=True).to("meta")
+        results, _ = model.inference(data_in=[torch.zeros(16000)], key=["sample"])
+        self.assertEqual(results, [{"key": "sample", "value": []}])
+        load_model.assert_called_once_with(onnx=True)
+
     def test_rejects_negative_max_segment_length(self):
-        with self.assertRaisesRegex(ValueError, "non-negative"):
+        with self.assertRaisesRegex(ValueError, "positive"):
             SileroVad._split_long_segments([[100, 1100]], -500)
 
+    def test_rejects_sub_millisecond_max_segment_length(self):
+        with self.assertRaisesRegex(ValueError, "positive"):
+            SileroVad._split_long_segments([[100, 1100]], 0.5)
+
 
 if __name__ == "__main__":
     unittest.main()