diff --git a/opennmt/tests/tokenizer_test.py b/opennmt/tests/tokenizer_test.py
index 869985cbd..bf3d627a5 100644
--- a/opennmt/tests/tokenizer_test.py
+++ b/opennmt/tests/tokenizer_test.py
@@ -103,6 +103,15 @@ def testOpenNMTTokenizer(self):
         [["Hello", "world", "￭!"], ["Test"], ["My", "name"]],
         ["Hello world!", "Test", "My name"])
 
+  def testOpenNMTTokenizerEmptyTensor(self):
+    tokenizer = tokenizers.OpenNMTTokenizer()
+    tokens = tokenizer.tokenize(tf.constant(""))
+    self.assertIs(tokens.dtype, tf.string)
+    self.assertListEqual(tokens.shape.as_list(), [0])
+    text = tokenizer.detokenize(tokens)
+    self.assertIs(text.dtype, tf.string)
+    self.assertListEqual(text.shape.as_list(), [])
+
   def testOpenNMTTokenizerArguments(self):
     tokenizer = tokenizers.OpenNMTTokenizer(
         mode="aggressive", spacer_annotate=True, spacer_new=True)
diff --git a/opennmt/tokenizers/tokenizer.py b/opennmt/tokenizers/tokenizer.py
index 23b9cb888..f7b817c6c 100644
--- a/opennmt/tokenizers/tokenizer.py
+++ b/opennmt/tokenizers/tokenizer.py
@@ -159,7 +159,7 @@ def _tokenize_tensor(self, text):
     def _python_wrapper(string_t):
       string = tf.compat.as_text(string_t.numpy())
       tokens = self._tokenize_string(string)
-      return tf.constant(tokens)
+      return tf.constant(tokens, dtype=tf.string)
     tokens = tf.py_function(_python_wrapper, [text], tf.string)
     tokens.set_shape([None])
     return tokens