From d219392e5b58e2f03f1d54e3be05adb3af5e5c8f Mon Sep 17 00:00:00 2001
From: LauraGPT <18321252+LauraGPT@users.noreply.github.com>
Date: Sat, 18 Jul 2026 21:33:07 +0000
Subject: [PATCH] fix(stt): anchor Qwen3-ASR envelope stripping

---
 tests/tools/test_transcription_tools.py | 24 ++++++++++++++++++++++++
 tools/transcription_tools.py            |  2 +-
 2 files changed, 25 insertions(+), 1 deletion(-)

diff --git a/tests/tools/test_transcription_tools.py b/tests/tools/test_transcription_tools.py
index c0186513a0e..9cfe47a50c8 100644
--- a/tests/tools/test_transcription_tools.py
+++ b/tests/tools/test_transcription_tools.py
@@ -1850,6 +1850,30 @@ class TestTranscribeAudioElevenLabsDispatch:
         assert mock_elevenlabs.call_args[0][1] == "scribe_v2"
 
 
+# ============================================================================
+# _extract_transcript_text
+# ============================================================================
+
+class TestExtractTranscriptText:
+    def test_strips_qwen3_asr_language_envelope(self):
+        from tools.transcription_tools import _extract_transcript_text
+
+        result = _extract_transcript_text(
+            "language zh\n<audio_language>zh</audio_language>\n<asr_text>你好，世界",
+        )
+
+        assert result == "你好，世界"
+
+    def test_keeps_non_envelope_marker_literal(self):
+        from tools.transcription_tools import _extract_transcript_text
+
+        result = _extract_transcript_text(
+            "The user literally said <asr_text> while reading markup.",
+        )
+
+        assert result == "The user literally said <asr_text> while reading markup."
+
+
 # Shell safety — shlex.split on auto-detected templates
 # ============================================================================
 class TestShellSafety:
diff --git a/tools/transcription_tools.py b/tools/transcription_tools.py
index 0b14e34bbcf..725020ed369 100644
--- a/tools/transcription_tools.py
+++ b/tools/transcription_tools.py
@@ -2095,7 +2095,7 @@ def _extract_transcript_text(transcription: Any) -> str:
         text = str(transcription).strip()
 
     marker = "<asr_text>"
-    if marker in text:
+    if text.lstrip().lower().startswith("language") and marker in text:
         text = text.split(marker, 1)[1].strip()
 
     return text