diff --git a/skl2onnx/_parse.py b/skl2onnx/_parse.py
index b64430ebe..2d07e9588 100644
--- a/skl2onnx/_parse.py
+++ b/skl2onnx/_parse.py
@@ -287,7 +287,7 @@ def _parse_sklearn_feature_union(scope, model, inputs, custom_parsers=None):
     # Encode each transform as our IR object
     for name, transform in model.transformer_list:
         transformed_result_names.append(
-            _parse_sklearn_simple_model(
+            _parse_sklearn(
                 scope, transform, inputs,
                 custom_parsers=custom_parsers)[0])
         if (model.transformer_weights is not None and name in
diff --git a/tests/test_sklearn_feature_union.py b/tests/test_sklearn_feature_union.py
index f183eb7b3..b40774ad1 100644
--- a/tests/test_sklearn_feature_union.py
+++ b/tests/test_sklearn_feature_union.py
@@ -38,6 +38,30 @@ def test_feature_union_default(self):
         dump_data_and_model(X_test, model, model_onnx,
                             basename="SklearnFeatureUnionDefault")
 
+    @unittest.skipIf(
+        pv.Version(ort_version) <= pv.Version('0.4.0'),
+        reason="onnxruntime too old")
+    def test_feature_union_nested(self):
+        data = load_iris()
+        X, y = data.data, data.target
+        X = X.astype(np.float32)
+        X_train, X_test, y_train, y_test = train_test_split(
+                X, y, test_size=0.5, random_state=42)
+        model = FeatureUnion([
+            ('features', FeatureUnion([
+                ('standard', StandardScaler()),
+                ])
+             ),
+            ]).fit(X_train)
+
+        model_onnx = convert_sklearn(
+            model, 'feature union',
+            [('input', FloatTensorType([None, X_test.shape[1]]))],
+            target_opset=TARGET_OPSET)
+        self.assertTrue(model_onnx is not None)
+        dump_data_and_model(X_test, model, model_onnx,
+                            basename="SklearnFeatureUnionNested")
+
     @unittest.skipIf(
         pv.Version(ort_version) <= pv.Version('0.4.0'),
         reason="onnxruntime too old")