From 471070c2c86ea23214d8cc2e2ef3c25a7ab0b13b Mon Sep 17 00:00:00 2001
From: Lynn <lynn_inf@hotmail.com>
Date: Fri, 14 Aug 2026 16:27:34 +0800
Subject: [PATCH] Fix: return error when no valid speech text in
 intranscription response (#18277)

---
 api/apps/restful_apis/chat_api.py                |  2 ++
 rag/llm/sequence2txt_model.py                    | 16 ++++++++++++++--
 .../unit_test/rag/llm/test_sequence2txt_model.py |  2 +-
 3 files changed, 17 insertions(+), 3 deletions(-)

diff --git a/api/apps/restful_apis/chat_api.py b/api/apps/restful_apis/chat_api.py
index 1d9658595a..92b32ddc88 100644
--- a/api/apps/restful_apis/chat_api.py
+++ b/api/apps/restful_apis/chat_api.py
@@ -1151,6 +1151,8 @@ async def transcription():
             os.remove(temp_audio_path)
         except Exception as e:
             logging.error(f"Failed to remove temp audio file: {str(e)}")
+        if "**ERROR**" in text:
+            return get_data_error_result(message=text)
         return get_json_result(data={"text": text})
 
     async def event_stream():
diff --git a/rag/llm/sequence2txt_model.py b/rag/llm/sequence2txt_model.py
index 2c96dce3d0..cb57f9f407 100644
--- a/rag/llm/sequence2txt_model.py
+++ b/rag/llm/sequence2txt_model.py
@@ -230,7 +230,13 @@ class QWenSeq2txt(Base):
         resp = dashscope.MultiModalConversation.call(model=self.model_name, messages=messages, result_format="message", asr_options={"enable_lid": True, "enable_itn": False})
 
         try:
-            text = resp["output"]["choices"][0]["message"].content[0]["text"]
+            choices = resp["output"].get("choices") or []
+            if not choices:
+                raise ValueError("no valid speech content in response")
+            content = (choices[0].get("message") or {}).get("content") or []
+            if not content or "text" not in content[0]:
+                raise ValueError("no speech text in response")
+            text = content[0]["text"]
         except Exception as e:
             text = "**ERROR**: " + str(e)
         return text, num_tokens_from_string(text)
@@ -369,7 +375,13 @@ class QWenSeq2txt(Base):
         full = ""
         for chunk in stream:
             try:
-                piece = chunk["output"]["choices"][0]["message"].content[0]["text"]
+                choices = chunk["output"].get("choices") or []
+                if not choices:
+                    raise ValueError("no valid speech content in response")
+                content = (choices[0].get("message") or {}).get("content") or []
+                if not content or "text" not in content[0]:
+                    raise ValueError("no speech text in response")
+                piece = content[0]["text"]
                 full = piece
                 yield {"event": "delta", "text": piece}
             except Exception as e:
diff --git a/test/unit_test/rag/llm/test_sequence2txt_model.py b/test/unit_test/rag/llm/test_sequence2txt_model.py
index 42f3c16bfe..0eed80cb11 100644
--- a/test/unit_test/rag/llm/test_sequence2txt_model.py
+++ b/test/unit_test/rag/llm/test_sequence2txt_model.py
@@ -46,7 +46,7 @@ def test_fun_asr_flash_uses_native_request_format(tmp_path):
 
 
 def test_qwen_audio_asr_keeps_existing_dashscope_path():
-    response = {"output": {"choices": [{"message": MagicMock(content=[{"text": "legacy text"}])}]}}
+    response = {"output": {"choices": [{"message": {"content": [{"text": "legacy text"}]}}]}}
 
     with patch("dashscope.MultiModalConversation.call", return_value=response) as call:
         model = QWenSeq2txt("test-key", "qwen-audio-asr")