feat: Granular Per Agent Speech Configuration

Merge #3170

Addresses Feature Request: #3116

This PR adds a `speech_config` to the **LLM Agent configuration** for the **live use case**. When an **asynchronous LLM** call is made to the **Gemini Live API**, it prioritizes the most specific agent configuration's speech_config. If that is null, it then uses the run configuration's speech_config. Unit tests have been added to verify this behavior.

COPYBARA_INTEGRATE_REVIEW=#3170 from qyuo:bidi_agent_speech_config af1bd27
PiperOrigin-RevId: 822305427

Author: Luis Pabon

contributing/samples/live_bidi_streaming_multi_agent/agent.py

@@ -16,6 +16,7 @@
 
 from google.adk.agents.llm_agent import Agent
 from google.adk.examples.example import Example
+from google.adk.models.google_llm import Gemini
 from google.adk.tools.example_tool import ExampleTool
 from google.genai import types
 
@@ -28,6 +29,17 @@ def roll_die(sides: int) -> int:
 
 roll_agent = Agent(
     name="roll_agent",
+    model=Gemini(
+        # model="gemini-2.0-flash-live-preview-04-09",  # for Vertex project
+        model="gemini-live-2.5-flash-preview",  # for AI studio key
+        speech_config=types.SpeechConfig(
+            voice_config=types.VoiceConfig(
+                prebuilt_voice_config=types.PrebuiltVoiceConfig(
+                    voice_name="Kore",
+                )
+            )
+        ),
+    ),
     description="Handles rolling dice of different sizes.",
     instruction="""
       You are responsible for rolling dice based on the user's request.
@@ -69,6 +81,17 @@ def check_prime(nums: list[int]) -> str:
 
 prime_agent = Agent(
     name="prime_agent",
+    model=Gemini(
+        # model="gemini-2.0-flash-live-preview-04-09",  # for Vertex project
+        model="gemini-live-2.5-flash-preview",  # for AI studio key
+        speech_config=types.SpeechConfig(
+            voice_config=types.VoiceConfig(
+                prebuilt_voice_config=types.PrebuiltVoiceConfig(
+                    voice_name="Puck",
+                )
+            )
+        ),
+    ),
     description="Handles checking if numbers are prime.",
     instruction="""
       You are responsible for checking whether numbers are prime.
@@ -100,8 +123,17 @@ def get_current_weather(location: str):
 
 root_agent = Agent(
     # find supported models here: https://google.github.io/adk-docs/get-started/streaming/quickstart-streaming/
-    model="gemini-2.0-flash-live-preview-04-09",  # for Vertex project
-    # model="gemini-live-2.5-flash-preview",  # for AI studio key
+    model=Gemini(
+        # model="gemini-2.0-flash-live-preview-04-09",  # for Vertex project
+        model="gemini-live-2.5-flash-preview",  # for AI studio key
+        speech_config=types.SpeechConfig(
+            voice_config=types.VoiceConfig(
+                prebuilt_voice_config=types.PrebuiltVoiceConfig(
+                    voice_name="Zephyr",
+                )
+            )
+        ),
+    ),
     name="root_agent",
     instruction="""
       You are a helpful assistant that can check time, roll dice and check if numbers are prime.

src/google/adk/agents/run_config.py

@@ -35,7 +35,10 @@ class StreamingMode(Enum):
 
 
 class RunConfig(BaseModel):
-  """Configs for runtime behavior of agents."""
+  """Configs for runtime behavior of agents.
+
+  The configs here will be overriden by agent-specific configurations.
+  """
 
   model_config = ConfigDict(
       extra='forbid',

src/google/adk/models/google_llm.py

@@ -60,6 +60,8 @@ class Gemini(BaseLlm):
 
   model: str = 'gemini-2.5-flash'
 
+  speech_config: Optional[types.SpeechConfig] = None
+
   retry_options: Optional[types.HttpRetryOptions] = None
   """Allow Gemini to retry failed responses.
 
@@ -269,6 +271,9 @@ async def connect(self, llm_request: LlmRequest) -> BaseLlmConnection:
           self._live_api_version
       )
 
+    if self.speech_config is not None:
+      llm_request.live_connect_config.speech_config = self.speech_config
+
     llm_request.live_connect_config.system_instruction = types.Content(
         role='system',
         parts=[

tests/unittests/models/test_google_llm.py

@@ -1858,3 +1858,189 @@ def mock_model_dump(*args, **kwargs):
   # Should still succeed using repr()
   assert "Config:" in log_output
   assert "GenerateContentConfig" in log_output
+
+
+@pytest.mark.asyncio
+async def test_connect_uses_gemini_speech_config_when_request_is_none(
+    gemini_llm, llm_request
+):
+  """Tests that Gemini's speech_config is used when live_connect_config's is None."""
+  # Arrange: Set a speech_config on the Gemini instance with the voice "Kore"
+  gemini_llm.speech_config = types.SpeechConfig(
+      voice_config=types.VoiceConfig(
+          prebuilt_voice_config=types.PrebuiltVoiceConfig(
+              voice_name="Kore",
+          )
+      )
+  )
+  llm_request.live_connect_config = (
+      types.LiveConnectConfig()
+  )  # speech_config is None
+
+  mock_live_session = mock.AsyncMock()
+
+  with mock.patch.object(gemini_llm, "_live_api_client") as mock_live_client:
+
+    class MockLiveConnect:
+
+      async def __aenter__(self):
+        return mock_live_session
+
+      async def __aexit__(self, *args):
+        pass
+
+    mock_live_client.aio.live.connect.return_value = MockLiveConnect()
+
+    # Act
+    async with gemini_llm.connect(llm_request) as connection:
+      # Assert
+      mock_live_client.aio.live.connect.assert_called_once()
+      call_args = mock_live_client.aio.live.connect.call_args
+      config_arg = call_args.kwargs["config"]
+
+      # Verify the speech_config from the Gemini instance was used
+      assert config_arg.speech_config is not None
+      assert (
+          config_arg.speech_config.voice_config.prebuilt_voice_config.voice_name
+          == "Kore"
+      )
+      assert isinstance(connection, GeminiLlmConnection)
+
+
+@pytest.mark.asyncio
+async def test_connect_uses_request_speech_config_when_gemini_is_none(
+    gemini_llm, llm_request
+):
+  """Tests that request's speech_config is used when Gemini's is None."""
+  # Arrange: Set a speech_config on the request instance with the voice "Kore"
+  gemini_llm.speech_config = None
+  request_speech_config = types.SpeechConfig(
+      voice_config=types.VoiceConfig(
+          prebuilt_voice_config=types.PrebuiltVoiceConfig(
+              voice_name="Kore",
+          )
+      )
+  )
+  llm_request.live_connect_config = types.LiveConnectConfig(
+      speech_config=request_speech_config
+  )
+
+  mock_live_session = mock.AsyncMock()
+
+  with mock.patch.object(gemini_llm, "_live_api_client") as mock_live_client:
+
+    class MockLiveConnect:
+
+      async def __aenter__(self):
+        return mock_live_session
+
+      async def __aexit__(self, *args):
+        pass
+
+    mock_live_client.aio.live.connect.return_value = MockLiveConnect()
+
+    # Act
+    async with gemini_llm.connect(llm_request) as connection:
+      # Assert
+      mock_live_client.aio.live.connect.assert_called_once()
+      call_args = mock_live_client.aio.live.connect.call_args
+      config_arg = call_args.kwargs["config"]
+
+      # Verify the speech_config from the request instance was used
+      assert config_arg.speech_config is not None
+      assert (
+          config_arg.speech_config.voice_config.prebuilt_voice_config.voice_name
+          == "Kore"
+      )
+      assert isinstance(connection, GeminiLlmConnection)
+
+
+@pytest.mark.asyncio
+async def test_connect_request_gemini_config_overrides_speech_config(
+    gemini_llm, llm_request
+):
+  """Tests that live_connect_config's speech_config is preserved even if Gemini has one."""
+  # Arrange: Set different speech_configs on both the Gemini instance ("Puck") and the request ("Zephyr")
+  gemini_llm.speech_config = types.SpeechConfig(
+      voice_config=types.VoiceConfig(
+          prebuilt_voice_config=types.PrebuiltVoiceConfig(
+              voice_name="Puck",
+          )
+      )
+  )
+  request_speech_config = types.SpeechConfig(
+      voice_config=types.VoiceConfig(
+          prebuilt_voice_config=types.PrebuiltVoiceConfig(
+              voice_name="Zephyr",
+          )
+      )
+  )
+  llm_request.live_connect_config = types.LiveConnectConfig(
+      speech_config=request_speech_config
+  )
+
+  mock_live_session = mock.AsyncMock()
+
+  with mock.patch.object(gemini_llm, "_live_api_client") as mock_live_client:
+
+    class MockLiveConnect:
+
+      async def __aenter__(self):
+        return mock_live_session
+
+      async def __aexit__(self, *args):
+        pass
+
+    mock_live_client.aio.live.connect.return_value = MockLiveConnect()
+
+    # Act
+    async with gemini_llm.connect(llm_request) as connection:
+      # Assert
+      mock_live_client.aio.live.connect.assert_called_once()
+      call_args = mock_live_client.aio.live.connect.call_args
+      config_arg = call_args.kwargs["config"]
+
+      # Verify the speech_config from the request ("Zephyr") was overwritten by Gemini's speech_config ("Puck")
+      assert config_arg.speech_config is not None
+      assert (
+          config_arg.speech_config.voice_config.prebuilt_voice_config.voice_name
+          == "Puck"
+      )
+      assert isinstance(connection, GeminiLlmConnection)
+
+
+@pytest.mark.asyncio
+async def test_connect_speech_config_remains_none_when_both_are_none(
+    gemini_llm, llm_request
+):
+  """Tests that speech_config is None when neither Gemini nor the request has it."""
+  # Arrange: Ensure both Gemini instance and request have no speech_config
+  gemini_llm.speech_config = None
+  llm_request.live_connect_config = (
+      types.LiveConnectConfig()
+  )  # speech_config is None
+
+  mock_live_session = mock.AsyncMock()
+
+  with mock.patch.object(gemini_llm, "_live_api_client") as mock_live_client:
+
+    class MockLiveConnect:
+
+      async def __aenter__(self):
+        return mock_live_session
+
+      async def __aexit__(self, *args):
+        pass
+
+    mock_live_client.aio.live.connect.return_value = MockLiveConnect()
+
+    # Act
+    async with gemini_llm.connect(llm_request) as connection:
+      # Assert
+      mock_live_client.aio.live.connect.assert_called_once()
+      call_args = mock_live_client.aio.live.connect.call_args
+      config_arg = call_args.kwargs["config"]
+
+      # Verify the final speech_config is still None
+      assert config_arg.speech_config is None
+      assert isinstance(connection, GeminiLlmConnection)