Raise in vad on non-finite input instead of reporting silence (closes #4216)

src/torchaudio/functional/filtering.py

@@ -1694,6 +1694,19 @@ def vad(
             flushedLen_ns = (measures_len - num_measures_to_flush) * measure_period_ns
             break
     # end for window
+    if not has_triggered and not bool(torch.isfinite(waveform).all()):
+        # Nothing triggered. Either the signal really is silent, or the running trigger
+        # measure was poisoned: one non-finite sample makes it non-finite for the rest of
+        # the waveform, so every later comparison against trigger_level is false and the
+        # whole input is discarded as silence. Only the second case is an error, and the
+        # scan sits in this branch so a waveform that triggers never pays for it.
+        raise ValueError(
+            "waveform must be finite everywhere, but it contains NaN or infinite values. "
+            "Vad tracks a running measure of the signal, and a single non-finite sample "
+            "makes every later comparison against trigger_level false, so the whole input "
+            "would be discarded as silence."
+        )
+
     if not has_triggered and shape[-1] >= fixed_pre_trigger_len_ns:
         return waveform[..., :fixed_pre_trigger_len_ns].view(shape[:-1] + torch.Size([fixed_pre_trigger_len_ns]))
 

test/torchaudio_unittest/transforms/transforms_test_impl.py

@@ -498,3 +498,18 @@ def test_vad_on_zero_audio(self, input_shape, output_shape, sample_rate: int, pr
         expected_output = torch.zeros(output_shape, dtype=self.dtype, device=self.device)
         result = T.Vad(sample_rate, pre_trigger_time=pre_trigger_time)(inpt)
         self.assertEqual(result, expected_output)
+
+    @parameterized.expand([(float("nan"),), (float("inf"),), (float("-inf"),)])
+    def test_vad_rejects_non_finite_audio(self, bad_value: float):
+        """VAD should raise on non-finite input rather than report the signal as silence.
+
+        The trigger measure is a running mean, so one non-finite sample poisons it and
+        every later comparison against trigger_level is false. Nothing triggers and the
+        whole waveform is discarded, returning an empty Tensor. See
+        https://github.com/pytorch/audio/issues/4216.
+        """
+        sample_rate = 16000
+        waveform = torch.zeros(sample_rate, dtype=self.dtype, device=self.device)
+        waveform[100] = bad_value
+        with self.assertRaisesRegex(ValueError, "finite"):
+            T.Vad(sample_rate)(waveform)