diff --git a/src/google/adk/runners.py b/src/google/adk/runners.py
index 626b6ee5d1..39a8529b95 100644
--- a/src/google/adk/runners.py
+++ b/src/google/adk/runners.py
@@ -2379,18 +2379,24 @@ def _new_invocation_context_for_live(
     run_config = run_config or RunConfig()
 
     # For live multi-agents system, we need model's text transcription as
-    # context for the transferred agent.
+    # context for the transferred agent. Only fill fields the caller left
+    # unset: an explicit None is an opt-out and stays None. Both fields
+    # default to an AudioTranscriptionConfig() via default_factory, so
+    # `model_fields_set` is what separates "unset" from "explicitly off".
+    # Writes go on a copy, never the caller's own RunConfig (see the same
+    # treatment for response_modalities in run_live).
     if hasattr(self.agent, 'sub_agents') and self.agent.sub_agents:
+      updates: dict[str, Any] = {}
       if (
           run_config.response_modalities
           and types.Modality.AUDIO in run_config.response_modalities
+          and 'output_audio_transcription' not in run_config.model_fields_set
       ):
-        if not run_config.output_audio_transcription:
-          run_config.output_audio_transcription = (
-              types.AudioTranscriptionConfig()
-          )
-      if not run_config.input_audio_transcription:
-        run_config.input_audio_transcription = types.AudioTranscriptionConfig()
+        updates['output_audio_transcription'] = types.AudioTranscriptionConfig()
+      if 'input_audio_transcription' not in run_config.model_fields_set:
+        updates['input_audio_transcription'] = types.AudioTranscriptionConfig()
+      if updates:
+        run_config = run_config.model_copy(update=updates)
     return self._new_invocation_context(
         session,
         live_request_queue=live_request_queue,
diff --git a/tests/unittests/test_runners.py b/tests/unittests/test_runners.py
index 05460f50f2..ef6417f7f8 100644
--- a/tests/unittests/test_runners.py
+++ b/tests/unittests/test_runners.py
@@ -2983,5 +2983,115 @@ async def test_base_agent_run_live_does_not_leak_context():
     otel_context.detach(token)
 
 
+def _live_runner(agent) -> Runner:
+  return Runner(
+      app_name=TEST_APP_ID,
+      agent=agent,
+      session_service=InMemorySessionService(),
+      artifact_service=InMemoryArtifactService(),
+  )
+
+
+def _multi_agent() -> LlmAgent:
+  return LlmAgent(
+      name="live_root",
+      model="gemini-2.0-flash-live-001",
+      sub_agents=[
+          LlmAgent(name="live_child", model="gemini-2.0-flash-live-001")
+      ],
+  )
+
+
+def _live_session() -> Session:
+  return Session(id=TEST_SESSION_ID, app_name=TEST_APP_ID, user_id=TEST_USER_ID)
+
+
+def test_live_context_respects_explicit_transcription_opt_out():
+  """An explicit None for transcription stays None for the run."""
+  from google.adk.agents.live_request_queue import LiveRequestQueue
+
+  runner = _live_runner(_multi_agent())
+  config = RunConfig(
+      response_modalities=[types.Modality.AUDIO],
+      input_audio_transcription=None,
+      output_audio_transcription=None,
+  )
+
+  ic = runner._new_invocation_context_for_live(
+      _live_session(),
+      live_request_queue=LiveRequestQueue(),
+      run_config=config,
+  )
+
+  assert ic.run_config.input_audio_transcription is None
+  assert ic.run_config.output_audio_transcription is None
+  # The caller's own RunConfig is never written to.
+  assert config.input_audio_transcription is None
+  assert config.output_audio_transcription is None
+
+
+def test_live_context_keeps_default_transcription_enabled():
+  """A default RunConfig keeps its default-on transcription configs."""
+  from google.adk.agents.live_request_queue import LiveRequestQueue
+
+  runner = _live_runner(_multi_agent())
+  config = RunConfig(response_modalities=[types.Modality.AUDIO])
+
+  ic = runner._new_invocation_context_for_live(
+      _live_session(),
+      live_request_queue=LiveRequestQueue(),
+      run_config=config,
+  )
+
+  assert ic.run_config.input_audio_transcription is not None
+  assert ic.run_config.output_audio_transcription is not None
+
+
+def test_live_context_without_sub_agents_passes_config_through():
+  """Without sub_agents the config is passed through untouched."""
+  from google.adk.agents.live_request_queue import LiveRequestQueue
+
+  runner = _live_runner(
+      LlmAgent(name="live_solo", model="gemini-2.0-flash-live-001")
+  )
+  config = RunConfig(
+      response_modalities=[types.Modality.AUDIO],
+      input_audio_transcription=None,
+      output_audio_transcription=None,
+  )
+
+  ic = runner._new_invocation_context_for_live(
+      _live_session(),
+      live_request_queue=LiveRequestQueue(),
+      run_config=config,
+  )
+
+  assert ic.run_config is config
+  assert config.input_audio_transcription is None
+  assert config.output_audio_transcription is None
+
+
+def test_live_context_opt_out_survives_unset_modalities():
+  """Explicit None is respected even when response_modalities is unset."""
+  from google.adk.agents.live_request_queue import LiveRequestQueue
+
+  runner = _live_runner(_multi_agent())
+  config = RunConfig(
+      input_audio_transcription=None,
+      output_audio_transcription=None,
+  )
+
+  ic = runner._new_invocation_context_for_live(
+      _live_session(),
+      live_request_queue=LiveRequestQueue(),
+      run_config=config,
+  )
+
+  assert ic.run_config.input_audio_transcription is None
+  assert ic.run_config.output_audio_transcription is None
+  assert config.input_audio_transcription is None
+  assert config.output_audio_transcription is None
+
+
 if __name__ == "__main__":
   pytest.main([__file__])