diff --git a/dashscope/legacy_1x/audio/tts/speech_synthesizer.py b/dashscope/legacy_1x/audio/tts/speech_synthesizer.py
index c3328681..5597c224 100644
--- a/dashscope/legacy_1x/audio/tts/speech_synthesizer.py
+++ b/dashscope/legacy_1x/audio/tts/speech_synthesizer.py
@@ -180,6 +180,14 @@ def call(  # pylint: disable=R1702,too-many-branches  # type: ignore[override] #
                             part,
                         )
                     else:
+                        if (
+                            not isinstance(part.output, dict)
+                            or "sentence" not in part.output
+                        ):
+                            # Non-timestamp message (e.g. extra messages returned
+                            # in the SSML/bgm scenario). Skip it safely instead of
+                            # raising KeyError on part.output["sentence"].
+                            continue
                         if _callback is not None:
                             sentence = SpeechSynthesisResult(
                                 None,  # type: ignore[arg-type]