From 7d2b8a3cade08ddf2cb20a689b66ee26407d8c5e Mon Sep 17 00:00:00 2001
From: LauraGPT <18321252+LauraGPT@users.noreply.github.com>
Date: Sun, 19 Jul 2026 05:48:41 +0000
Subject: [PATCH] fix(stt): anchor qwen asr envelope stripping

---
 tests/tools/test_transcription_tools.py |  9 +++++++++
 tools/transcription_tools.py            | 11 ++++++++---
 2 files changed, 17 insertions(+), 3 deletions(-)

diff --git a/tests/tools/test_transcription_tools.py b/tests/tools/test_transcription_tools.py
index 9cfe47a50c83..63b9ad696405 100644
--- a/tests/tools/test_transcription_tools.py
+++ b/tests/tools/test_transcription_tools.py
@@ -1873,6 +1873,15 @@ class TestExtractTranscriptText:
 
         assert result == "The user literally said <asr_text> while reading markup."
 
+    def test_keeps_language_sentence_with_marker_literal(self):
+        from tools.transcription_tools import _extract_transcript_text
+
+        result = _extract_transcript_text(
+            "Language teachers may say <asr_text> when discussing markup.",
+        )
+
+        assert result == "Language teachers may say <asr_text> when discussing markup."
+
 
 # Shell safety — shlex.split on auto-detected templates
 # ============================================================================
diff --git a/tools/transcription_tools.py b/tools/transcription_tools.py
index 725020ed3699..f359d454a66d 100644
--- a/tools/transcription_tools.py
+++ b/tools/transcription_tools.py
@@ -30,6 +30,7 @@ Usage::
 import logging
 import os
 import platform
+import re
 import shlex
 import shutil
 import subprocess
@@ -2094,8 +2095,12 @@ def _extract_transcript_text(transcription: Any) -> str:
     if text is None:
         text = str(transcription).strip()
 
-    marker = "<asr_text>"
-    if text.lstrip().lower().startswith("language") and marker in text:
-        text = text.split(marker, 1)[1].strip()
+    match = re.match(
+        r"\s*language\s+[\w.-]+(?:\s*<audio_language>[^<]*</audio_language>)?\s*<asr_text>\s*(?P<text>.*)",
+        text,
+        flags=re.IGNORECASE | re.DOTALL,
+    )
+    if match:
+        text = match.group("text").strip()
 
     return text