[Test] Add long text output correctness test for Qwen3-Omni (#3539)

Signed-off-by: ZeldaHuang <hzm414167@alibaba-inc.com>
Signed-off-by: tzhouam <tzhouam@connect.ust.hk>
Co-authored-by: tzhouam <tzhouam@connect.ust.hk>

tests/helpers/assertions.py

@@ -477,6 +477,7 @@ def assert_omni_response(response: Any, request_config: dict[str, Any], run_leve
         # Verify similarity (Whisper transcript vs streamed/detokenized text)
         if "audio" in modalities:
             audio_ref_text = request_config.get("audio_ref_text")
+            similarity_threshold = request_config.get("similarity_threshold", 0.8)
             if "text" in modalities:
                 transcript = (response.audio_content or "").strip()
                 text_output = (response.text_content or "").strip()
@@ -484,14 +485,14 @@ def assert_omni_response(response: Any, request_config: dict[str, Any], run_leve
                     transcript.lower(),
                     text_output.lower(),
                 )
-                assert similarity > 0.9, "The audio content is not same as the text"
                 print(f"similarity is: {similarity}")
+                assert similarity > similarity_threshold, "The audio content is not same as the text"
             if audio_ref_text:
                 audio_similarity = cosine_similarity_text(
                     response.audio_content.lower(),
                     str(audio_ref_text).lower(),
                 )
-                assert audio_similarity > 0.9, (
+                assert audio_similarity > similarity_threshold, (
                     f"The audio content does not match reference text: similarity={audio_similarity:.3f}"
                 )