feat: Add support for store audio and transcription into events/sessions and artifacts

PiperOrigin-RevId: 798011660

contributing/samples/live_bidi_streaming_single_agent/agent.py

@@ -65,8 +65,8 @@ async def check_prime(nums: list[int]) -> str:
 
 
 root_agent = Agent(
-    model='gemini-2.0-flash-live-preview-04-09',  # for Vertex project
-    # model='gemini-2.0-flash-live-001',  # for AI studio key
+    # model='gemini-2.0-flash-live-preview-04-09',  # for Vertex project
+    model='gemini-2.0-flash-live-001',  # for AI studio key
     name='hello_world_agent',
     description=(
         'hello world agent that can roll a dice of 8 sides and check prime'

src/google/adk/agents/invocation_context.py

@@ -40,6 +40,25 @@ class LlmCallsLimitExceededError(Exception):
   """Error thrown when the number of LLM calls exceed the limit."""
 
 
+class RealtimeCacheEntry(BaseModel):
+  """Store audio data chunks for caching before flushing."""
+
+  model_config = ConfigDict(
+      arbitrary_types_allowed=True,
+      extra="forbid",
+  )
+  """The pydantic model config."""
+
+  role: str
+  """The role that created this audio data, typically "user" or "model"."""
+
+  data: types.Blob
+  """The audio data chunk."""
+
+  timestamp: float
+  """Timestamp when the audio chunk was received."""
+
+
 class _InvocationCostManager(BaseModel):
   """A container to keep track of the cost of invocation.
 
@@ -156,6 +175,12 @@ class InvocationContext(BaseModel):
   live_session_resumption_handle: Optional[str] = None
   """The handle for live session resumption."""
 
+  input_realtime_cache: Optional[list[RealtimeCacheEntry]] = None
+  """Caches input audio chunks before flushing to session and artifact services."""
+
+  output_realtime_cache: Optional[list[RealtimeCacheEntry]] = None
+  """Caches output audio chunks before flushing to session and artifact services."""
+
   run_config: Optional[RunConfig] = None
   """Configurations for live agents under this invocation."""
 

src/google/adk/flows/llm_flows/audio_cache_manager.py

@@ -0,0 +1,264 @@
+# Copyright 2025 Google LLC
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+from __future__ import annotations
+
+import logging
+import time
+from typing import TYPE_CHECKING
+
+from google.genai import types
+
+from ...agents.invocation_context import RealtimeCacheEntry
+from ...events.event import Event
+
+if TYPE_CHECKING:
+  from ...agents.invocation_context import InvocationContext
+
+logger = logging.getLogger('google_adk.' + __name__)
+
+
+class AudioCacheManager:
+  """Manages audio caching and flushing for live streaming flows."""
+
+  def __init__(self, config: AudioCacheConfig | None = None):
+    """Initialize the audio cache manager.
+
+    Args:
+      config: Configuration for audio caching behavior.
+    """
+    self.config = config or AudioCacheConfig()
+
+  def cache_audio(
+      self,
+      invocation_context: InvocationContext,
+      audio_blob: types.Blob,
+      cache_type: str,
+  ) -> None:
+    """Cache incoming user or outgoing model audio data.
+
+    Args:
+      invocation_context: The current invocation context.
+      audio_blob: The audio data to cache.
+      cache_type: Type of audio to cache, either 'input' or 'output'.
+
+    Raises:
+      ValueError: If cache_type is not 'input' or 'output'.
+    """
+    if cache_type == 'input':
+      if not invocation_context.input_realtime_cache:
+        invocation_context.input_realtime_cache = []
+      cache = invocation_context.input_realtime_cache
+      role = 'user'
+    elif cache_type == 'output':
+      if not invocation_context.output_realtime_cache:
+        invocation_context.output_realtime_cache = []
+      cache = invocation_context.output_realtime_cache
+      role = 'model'
+    else:
+      raise ValueError("cache_type must be either 'input' or 'output'")
+
+    audio_entry = RealtimeCacheEntry(
+        role=role, data=audio_blob, timestamp=time.time()
+    )
+    cache.append(audio_entry)
+
+    logger.debug(
+        'Cached %s audio chunk: %d bytes, cache size: %d',
+        cache_type,
+        len(audio_blob.data),
+        len(cache),
+    )
+
+  async def flush_caches(
+      self,
+      invocation_context: InvocationContext,
+      flush_user_audio: bool = True,
+      flush_model_audio: bool = True,
+  ) -> None:
+    """Flush audio caches to session and artifact services.
+
+    The multimodality data is saved in artifact service in the format of
+    audio file. The file data reference is added to the session as an event.
+    The audio file follows the naming convention: artifact_ref =
+    f"artifact://{invocation_context.app_name}/{invocation_context.user_id}/
+    {invocation_context.session.id}/_adk_live/{filename}#{revision_id}"
+
+    Note: video data is not supported yet.
+
+    Args:
+      invocation_context: The invocation context containing audio caches.
+      flush_user_audio: Whether to flush the input (user) audio cache.
+      flush_model_audio: Whether to flush the output (model) audio cache.
+    """
+    if flush_user_audio and invocation_context.input_realtime_cache:
+      success = await self._flush_cache_to_services(
+          invocation_context,
+          invocation_context.input_realtime_cache,
+          'input_audio',
+      )
+      if success:
+        invocation_context.input_realtime_cache = []
+        logger.debug('Flushed input audio cache')
+
+    if flush_model_audio and invocation_context.output_realtime_cache:
+      success = await self._flush_cache_to_services(
+          invocation_context,
+          invocation_context.output_realtime_cache,
+          'output_audio',
+      )
+      if success:
+        invocation_context.output_realtime_cache = []
+        logger.debug('Flushed output audio cache')
+
+  async def _flush_cache_to_services(
+      self,
+      invocation_context: InvocationContext,
+      audio_cache: list[RealtimeCacheEntry],
+      cache_type: str,
+  ) -> bool:
+    """Flush a list of audio cache entries to session and artifact services.
+
+    The artifact service stores the actual blob. The session stores the
+    reference to the stored blob.
+
+    Args:
+      invocation_context: The invocation context.
+      audio_cache: The audio cache to flush.
+      cache_type: Type identifier for the cache ('input_audio' or 'output_audio').
+
+    Returns:
+      True if the cache was successfully flushed, False otherwise.
+    """
+    print('flush cache')
+    if not invocation_context.artifact_service or not audio_cache:
+      logger.debug('Skipping cache flush: no artifact service or empty cache')
+      return False
+
+    try:
+      # Combine audio chunks into a single file
+      combined_audio_data = b''
+      mime_type = audio_cache[0].data.mime_type if audio_cache else 'audio/pcm'
+
+      for entry in audio_cache:
+        combined_audio_data += entry.data.data
+
+      # Generate filename with timestamp from first audio chunk (when recording started)
+      timestamp = int(audio_cache[0].timestamp * 1000)  # milliseconds
+      filename = f"adk_live_audio_storage_{cache_type}_{timestamp}.{mime_type.split('/')[-1]}"
+
+      # Save to artifact service
+      combined_audio_part = types.Part(
+          inline_data=types.Blob(data=combined_audio_data, mime_type=mime_type)
+      )
+
+      revision_id = await invocation_context.artifact_service.save_artifact(
+          app_name=invocation_context.app_name,
+          user_id=invocation_context.user_id,
+          session_id=invocation_context.session.id,
+          filename=filename,
+          artifact=combined_audio_part,
+      )
+
+      # Create artifact reference for session service
+      artifact_ref = f'artifact://{invocation_context.app_name}/{invocation_context.user_id}/{invocation_context.session.id}/_adk_live/{filename}#{revision_id}'
+
+      # Create event with file data reference to add to session
+      audio_event = Event(
+          id=Event.new_id(),
+          invocation_id=invocation_context.invocation_id,
+          author=audio_cache[0].role,
+          content=types.Content(
+              role=audio_cache[0].role,
+              parts=[
+                  types.Part(
+                      file_data=types.FileData(
+                          file_uri=artifact_ref, mime_type=mime_type
+                      )
+                  )
+              ],
+          ),
+          timestamp=audio_cache[0].timestamp,
+      )
+
+      # Add to session
+      await invocation_context.session_service.append_event(
+          invocation_context.session, audio_event
+      )
+
+      logger.debug(
+          'Successfully flushed %s cache: %d chunks, %d bytes, saved as %s',
+          cache_type,
+          len(audio_cache),
+          len(combined_audio_data),
+          filename,
+      )
+      return True
+
+    except Exception as e:
+      logger.error('Failed to flush %s cache: %s', cache_type, e)
+      return False
+
+  def get_cache_stats(
+      self, invocation_context: InvocationContext
+  ) -> dict[str, int]:
+    """Get statistics about current cache state.
+
+    Args:
+      invocation_context: The invocation context.
+
+    Returns:
+      Dictionary containing cache statistics.
+    """
+    input_count = len(invocation_context.input_realtime_cache or [])
+    output_count = len(invocation_context.output_realtime_cache or [])
+
+    input_bytes = sum(
+        len(entry.data.data)
+        for entry in (invocation_context.input_realtime_cache or [])
+    )
+    output_bytes = sum(
+        len(entry.data.data)
+        for entry in (invocation_context.output_realtime_cache or [])
+    )
+
+    return {
+        'input_chunks': input_count,
+        'output_chunks': output_count,
+        'input_bytes': input_bytes,
+        'output_bytes': output_bytes,
+        'total_chunks': input_count + output_count,
+        'total_bytes': input_bytes + output_bytes,
+    }
+
+
+class AudioCacheConfig:
+  """Configuration for audio caching behavior."""
+
+  def __init__(
+      self,
+      max_cache_size_bytes: int = 10 * 1024 * 1024,  # 10MB
+      max_cache_duration_seconds: float = 300.0,  # 5 minutes
+      auto_flush_threshold: int = 100,  # Number of chunks
+  ):
+    """Initialize audio cache configuration.
+
+    Args:
+      max_cache_size_bytes: Maximum cache size in bytes before auto-flush.
+      max_cache_duration_seconds: Maximum duration to keep data in cache.
+      auto_flush_threshold: Number of chunks that triggers auto-flush.
+    """
+    self.max_cache_size_bytes = max_cache_size_bytes
+    self.max_cache_duration_seconds = max_cache_duration_seconds
+    self.auto_flush_threshold = auto_flush_threshold

src/google/adk/flows/llm_flows/base_llm_flow.py

@@ -47,6 +47,8 @@ from ...telemetry import tracer
 from ...tools.base_toolset import BaseToolset
 from ...tools.tool_context import ToolContext
 from ...utils.context_utils import Aclosing
+from .audio_cache_manager import AudioCacheManager
+from .transcription_manager import TranscriptionManager
 
 if TYPE_CHECKING:
   from ...agents.llm_agent import LlmAgent
@@ -58,6 +60,14 @@ logger = logging.getLogger('google_adk.' + __name__)
 
 _ADK_AGENT_NAME_LABEL_KEY = 'adk_agent_name'
 
+# Timing configuration
+DEFAULT_REQUEST_QUEUE_TIMEOUT = 0.25
+DEFAULT_TRANSFER_AGENT_DELAY = 1.0
+DEFAULT_TASK_COMPLETION_DELAY = 1.0
+
+# Statistics configuration
+DEFAULT_ENABLE_CACHE_STATISTICS = False
+
 
 class BaseLlmFlow(ABC):
   """A basic flow that calls the LLM in a loop until a final response is generated.
@@ -69,6 +79,10 @@ class BaseLlmFlow(ABC):
     self.request_processors: list[BaseLlmRequestProcessor] = []
     self.response_processors: list[BaseLlmResponseProcessor] = []
 
+    # Initialize configuration and managers
+    self.audio_cache_manager = AudioCacheManager()
+    self.transcription_manager = TranscriptionManager()
+
   async def run_live(
       self,
       invocation_context: InvocationContext,
@@ -169,7 +183,7 @@ class BaseLlmFlow(ABC):
                     and event.content.parts[0].function_response.name
                     == 'transfer_to_agent'
                 ):
-                  await asyncio.sleep(1)
+                  await asyncio.sleep(DEFAULT_TRANSFER_AGENT_DELAY)
                   # cancel the tasks that belongs to the closed connection.
                   send_task.cancel()
                   await llm_connection.close()
@@ -181,7 +195,7 @@ class BaseLlmFlow(ABC):
                     == 'task_completed'
                 ):
                   # this is used for sequential agent to signal the end of the agent.
-                  await asyncio.sleep(1)
+                  await asyncio.sleep(DEFAULT_TASK_COMPLETION_DELAY)
                   # cancel the tasks that belongs to the closed connection.
                   send_task.cancel()
                   return
@@ -218,7 +232,7 @@ class BaseLlmFlow(ABC):
         # event loop to process events.
         # TODO: revert back(remove timeout) once we move off streamlit.
         live_request = await asyncio.wait_for(
-            live_request_queue.get(), timeout=0.25
+            live_request_queue.get(), timeout=DEFAULT_REQUEST_QUEUE_TIMEOUT
         )
         # duplicate the live_request to all the active streams
         logger.debug(
@@ -253,6 +267,12 @@ class BaseLlmFlow(ABC):
           invocation_context.transcription_cache.append(
               TranscriptionEntry(role='user', data=live_request.blob)
           )
+
+        # Cache input audio chunks before flushing
+        self.audio_cache_manager.cache_audio(
+            invocation_context, live_request.blob, cache_type='input'
+        )
+
         await llm_connection.send_realtime(live_request.blob)
 
       if live_request.content:
@@ -303,6 +323,23 @@ class BaseLlmFlow(ABC):
                 invocation_id=invocation_context.invocation_id,
                 author=get_author_for_event(llm_response),
             )
+
+            # Handle transcription events ONCE per llm_response, outside the event loop
+            if llm_response.input_transcription:
+              await self.transcription_manager.handle_input_transcription(
+                  invocation_context, llm_response.input_transcription
+              )
+
+            if llm_response.output_transcription:
+              await self.transcription_manager.handle_output_transcription(
+                  invocation_context, llm_response.output_transcription
+              )
+
+            # Flush audio caches based on control events using configurable settings
+            await self._handle_control_event_flush(
+                invocation_context, llm_response
+            )
+
             async with Aclosing(
                 self._postprocess_live(
                     invocation_context,
@@ -330,6 +367,24 @@ class BaseLlmFlow(ABC):
                           role=event.content.role, data=event.content
                       )
                   )
+                # Cache output audio chunks from model responses
+                # TODO: support video data
+                if (
+                    event.content
+                    and event.content.parts
+                    and event.content.parts[0].inline_data
+                    and event.content.parts[0].inline_data.mime_type.startswith(
+                        'audio/'
+                    )
+                ):
+                  audio_blob = types.Blob(
+                      data=event.content.parts[0].inline_data.data,
+                      mime_type=event.content.parts[0].inline_data.mime_type,
+                  )
+                  self.audio_cache_manager.cache_audio(
+                      invocation_context, audio_blob, cache_type='output'
+                  )
+
                 yield event
         # Give opportunity for other tasks to run.
         await asyncio.sleep(0)
@@ -512,12 +567,14 @@ class BaseLlmFlow(ABC):
 
     # Skip the model response event if there is no content and no error code.
     # This is needed for the code executor to trigger another loop.
-    # But don't skip control events like turn_complete.
+    # But don't skip control events like turn_complete or transcription events.
     if (
         not llm_response.content
         and not llm_response.error_code
         and not llm_response.interrupted
         and not llm_response.turn_complete
+        and not llm_response.input_transcription
+        and not llm_response.output_transcription
     ):
       return
 
@@ -805,6 +862,42 @@ class BaseLlmFlow(ABC):
 
     return model_response_event
 
+  async def _handle_control_event_flush(
+      self, invocation_context: InvocationContext, llm_response: LlmResponse
+  ) -> None:
+    """Handle audio cache flushing based on control events.
+
+    Args:
+      invocation_context: The invocation context containing audio caches.
+      llm_response: The LLM response containing control event information.
+    """
+    if llm_response.interrupted:
+      # user interrupts so the model will stop. we can flush model audio here
+      await self.audio_cache_manager.flush_caches(
+          invocation_context,
+          flush_user_audio=False,
+          flush_model_audio=True,
+      )
+    elif llm_response.turn_complete:
+      # turn completes so we can flush both user and model
+      await self.audio_cache_manager.flush_caches(
+          invocation_context,
+          flush_user_audio=True,
+          flush_model_audio=True,
+      )
+    elif getattr(llm_response, 'generation_complete', False):
+      # model generation complete so we can flush model audio
+      await self.audio_cache_manager.flush_caches(
+          invocation_context,
+          flush_user_audio=False,
+          flush_model_audio=True,
+      )
+
+    # Log cache statistics if enabled
+    if DEFAULT_ENABLE_CACHE_STATISTICS:
+      stats = self.audio_cache_manager.get_cache_stats(invocation_context)
+      logger.debug('Audio cache stats: %s', stats)
+
   async def _run_and_handle_error(
       self,
       response_generator: AsyncGenerator[LlmResponse, None],

src/google/adk/flows/llm_flows/transcription_manager.py

@@ -0,0 +1,141 @@
+# Copyright 2025 Google LLC
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+from __future__ import annotations
+
+import logging
+import time
+from typing import TYPE_CHECKING
+
+from google.genai import types
+
+from ...events.event import Event
+
+if TYPE_CHECKING:
+  from ...agents.invocation_context import InvocationContext
+
+logger = logging.getLogger('google_adk.' + __name__)
+
+
+class TranscriptionManager:
+  """Manages transcription events for live streaming flows."""
+
+  async def handle_input_transcription(
+      self,
+      invocation_context: InvocationContext,
+      transcription: types.Transcription,
+  ) -> None:
+    """Handle user input transcription events.
+
+    Args:
+      invocation_context: The current invocation context.
+      transcription: The transcription data from user input.
+    """
+    await self._create_and_save_transcription_event(
+        invocation_context=invocation_context,
+        transcription=transcription,
+        author='user',
+        is_input=True,
+    )
+
+  async def handle_output_transcription(
+      self,
+      invocation_context: InvocationContext,
+      transcription: types.Transcription,
+  ) -> None:
+    """Handle model output transcription events.
+
+    Args:
+      invocation_context: The current invocation context.
+      transcription: The transcription data from model output.
+    """
+    await self._create_and_save_transcription_event(
+        invocation_context=invocation_context,
+        transcription=transcription,
+        author=invocation_context.agent.name,
+        is_input=False,
+    )
+
+  async def _create_and_save_transcription_event(
+      self,
+      invocation_context: InvocationContext,
+      transcription: types.Transcription,
+      author: str,
+      is_input: bool,
+  ) -> None:
+    """Create and save a transcription event to session service.
+
+    Args:
+      invocation_context: The current invocation context.
+      transcription: The transcription data.
+      author: The author of the transcription event.
+      is_input: Whether this is an input (user) or output (model) transcription.
+    """
+    try:
+      transcription_event = Event(
+          id=Event.new_id(),
+          invocation_id=invocation_context.invocation_id,
+          author=author,
+          input_transcription=transcription if is_input else None,
+          output_transcription=transcription if not is_input else None,
+          timestamp=time.time(),
+      )
+
+      # Save transcription event to session
+      await invocation_context.session_service.append_event(
+          invocation_context.session, transcription_event
+      )
+
+      logger.debug(
+          'Saved %s transcription event for %s: %s',
+          'input' if is_input else 'output',
+          author,
+          transcription.text
+          if hasattr(transcription, 'text')
+          else 'audio transcription',
+      )
+
+    except Exception as e:
+      logger.error(
+          'Failed to save %s transcription event: %s',
+          'input' if is_input else 'output',
+          e,
+      )
+      raise
+
+  def get_transcription_stats(
+      self, invocation_context: InvocationContext
+  ) -> dict[str, int]:
+    """Get statistics about transcription events in the session.
+
+    Args:
+      invocation_context: The current invocation context.
+
+    Returns:
+      Dictionary containing transcription statistics.
+    """
+    input_count = 0
+    output_count = 0
+
+    for event in invocation_context.session.events:
+      if hasattr(event, 'input_transcription') and event.input_transcription:
+        input_count += 1
+      if hasattr(event, 'output_transcription') and event.output_transcription:
+        output_count += 1
+
+    return {
+        'input_transcriptions': input_count,
+        'output_transcriptions': output_count,
+        'total_transcriptions': input_count + output_count,
+    }

src/google/adk/models/gemini_llm_connection.py

@@ -164,14 +164,8 @@ class GeminiLlmConnection(BaseLlmConnection):
               message.server_content.input_transcription
               and message.server_content.input_transcription.text
           ):
-            user_text = message.server_content.input_transcription.text
-            parts = [
-                types.Part.from_text(
-                    text=user_text,
-                )
-            ]
             llm_response = LlmResponse(
-                content=types.Content(role='user', parts=parts)
+                input_transcription=message.server_content.input_transcription,
             )
             yield llm_response
           if (
@@ -186,13 +180,8 @@ class GeminiLlmConnection(BaseLlmConnection):
             # We rely on other control signals to determine when to yield the
             # full text response(turn_complete, interrupted, or tool_call).
             text += message.server_content.output_transcription.text
-            parts = [
-                types.Part.from_text(
-                    text=message.server_content.output_transcription.text
-                )
-            ]
             llm_response = LlmResponse(
-                content=types.Content(role='model', parts=parts), partial=True
+                output_transcription=message.server_content.output_transcription
             )
             yield llm_response
 

src/google/adk/models/llm_response.py

@@ -40,6 +40,8 @@ class LlmResponse(BaseModel):
     interrupted: Flag indicating that LLM was interrupted when generating the
       content. Usually it's due to user interruption during a bidi streaming.
     custom_metadata: The custom metadata of the LlmResponse.
+    input_transcription: Audio transcription of user input.
+    output_transcription: Audio transcription of model output.
   """
 
   model_config = ConfigDict(
@@ -97,6 +99,12 @@ class LlmResponse(BaseModel):
   ] = None
   """The session resumption update of the LlmResponse"""
 
+  input_transcription: Optional[types.Transcription] = None
+  """Audio transcription of user input."""
+
+  output_transcription: Optional[types.Transcription] = None
+  """Audio transcription of model output."""
+
   @staticmethod
   def create(
       generate_content_response: types.GenerateContentResponse,

tests/unittests/flows/llm_flows/test_audio_cache_manager.py

@@ -0,0 +1,389 @@
+# Copyright 2025 Google LLC
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import time
+from unittest.mock import AsyncMock
+from unittest.mock import Mock
+
+from google.adk.flows.llm_flows.audio_cache_manager import AudioCacheConfig
+from google.adk.flows.llm_flows.audio_cache_manager import AudioCacheManager
+from google.genai import types
+import pytest
+
+from ... import testing_utils
+
+
+class TestAudioCacheConfig:
+  """Test the AudioCacheConfig class."""
+
+  def test_default_values(self):
+    """Test that default configuration values are set correctly."""
+    config = AudioCacheConfig()
+    assert config.max_cache_size_bytes == 10 * 1024 * 1024  # 10MB
+    assert config.max_cache_duration_seconds == 300.0  # 5 minutes
+    assert config.auto_flush_threshold == 100
+
+  def test_custom_values(self):
+    """Test that custom configuration values are set correctly."""
+    config = AudioCacheConfig(
+        max_cache_size_bytes=5 * 1024 * 1024,
+        max_cache_duration_seconds=120.0,
+        auto_flush_threshold=50,
+    )
+    assert config.max_cache_size_bytes == 5 * 1024 * 1024
+    assert config.max_cache_duration_seconds == 120.0
+    assert config.auto_flush_threshold == 50
+
+
+class TestAudioCacheManager:
+  """Test the AudioCacheManager class."""
+
+  def setup_method(self):
+    """Set up test fixtures."""
+    self.config = AudioCacheConfig()
+    self.manager = AudioCacheManager(self.config)
+
+  @pytest.mark.asyncio
+  async def test_cache_input_audio(self):
+    """Test caching input audio data."""
+    invocation_context = await testing_utils.create_invocation_context(
+        testing_utils.create_test_agent()
+    )
+
+    audio_blob = types.Blob(data=b'test_audio_data', mime_type='audio/pcm')
+
+    # Initially no cache
+    assert invocation_context.input_realtime_cache is None
+
+    # Cache audio
+    self.manager.cache_audio(invocation_context, audio_blob, 'input')
+
+    # Verify cache is created and populated
+    assert invocation_context.input_realtime_cache is not None
+    assert len(invocation_context.input_realtime_cache) == 1
+
+    entry = invocation_context.input_realtime_cache[0]
+    assert entry.role == 'user'
+    assert entry.data == audio_blob
+    assert isinstance(entry.timestamp, float)
+
+  @pytest.mark.asyncio
+  async def test_cache_output_audio(self):
+    """Test caching output audio data."""
+    invocation_context = await testing_utils.create_invocation_context(
+        testing_utils.create_test_agent()
+    )
+
+    audio_blob = types.Blob(data=b'test_model_audio', mime_type='audio/wav')
+
+    # Initially no cache
+    assert invocation_context.output_realtime_cache is None
+
+    # Cache audio
+    self.manager.cache_audio(invocation_context, audio_blob, 'output')
+
+    # Verify cache is created and populated
+    assert invocation_context.output_realtime_cache is not None
+    assert len(invocation_context.output_realtime_cache) == 1
+
+    entry = invocation_context.output_realtime_cache[0]
+    assert entry.role == 'model'
+    assert entry.data == audio_blob
+    assert isinstance(entry.timestamp, float)
+
+  @pytest.mark.asyncio
+  async def test_multiple_audio_caching(self):
+    """Test caching multiple audio chunks."""
+    invocation_context = await testing_utils.create_invocation_context(
+        testing_utils.create_test_agent()
+    )
+
+    # Cache multiple input audio chunks
+    for i in range(3):
+      audio_blob = types.Blob(data=f'input_{i}'.encode(), mime_type='audio/pcm')
+      self.manager.cache_audio(invocation_context, audio_blob, 'input')
+
+    # Cache multiple output audio chunks
+    for i in range(2):
+      audio_blob = types.Blob(
+          data=f'output_{i}'.encode(), mime_type='audio/wav'
+      )
+      self.manager.cache_audio(invocation_context, audio_blob, 'output')
+
+    # Verify all chunks are cached
+    assert len(invocation_context.input_realtime_cache) == 3
+    assert len(invocation_context.output_realtime_cache) == 2
+
+  @pytest.mark.asyncio
+  async def test_flush_caches_both(self):
+    """Test flushing both input and output caches."""
+    invocation_context = await testing_utils.create_invocation_context(
+        testing_utils.create_test_agent()
+    )
+
+    # Set up mock artifact service
+    mock_artifact_service = AsyncMock()
+    mock_artifact_service.save_artifact.return_value = 123
+    invocation_context.artifact_service = mock_artifact_service
+
+    # Cache some audio
+    input_blob = types.Blob(data=b'input_data', mime_type='audio/pcm')
+    output_blob = types.Blob(data=b'output_data', mime_type='audio/wav')
+    self.manager.cache_audio(invocation_context, input_blob, 'input')
+    self.manager.cache_audio(invocation_context, output_blob, 'output')
+
+    # Flush caches
+    await self.manager.flush_caches(invocation_context)
+
+    # Verify caches are cleared
+    assert invocation_context.input_realtime_cache == []
+    assert invocation_context.output_realtime_cache == []
+
+    # Verify artifact service was called twice (once for each cache)
+    assert mock_artifact_service.save_artifact.call_count == 2
+
+  @pytest.mark.asyncio
+  async def test_flush_caches_selective(self):
+    """Test selectively flushing only one cache."""
+    invocation_context = await testing_utils.create_invocation_context(
+        testing_utils.create_test_agent()
+    )
+
+    # Set up mock artifact service
+    mock_artifact_service = AsyncMock()
+    mock_artifact_service.save_artifact.return_value = 123
+    invocation_context.artifact_service = mock_artifact_service
+
+    # Cache some audio
+    input_blob = types.Blob(data=b'input_data', mime_type='audio/pcm')
+    output_blob = types.Blob(data=b'output_data', mime_type='audio/wav')
+    self.manager.cache_audio(invocation_context, input_blob, 'input')
+    self.manager.cache_audio(invocation_context, output_blob, 'output')
+
+    # Flush only input cache
+    await self.manager.flush_caches(
+        invocation_context, flush_user_audio=True, flush_model_audio=False
+    )
+
+    # Verify only input cache is cleared
+    assert invocation_context.input_realtime_cache == []
+    assert len(invocation_context.output_realtime_cache) == 1
+
+    # Verify artifact service was called once
+    assert mock_artifact_service.save_artifact.call_count == 1
+
+  @pytest.mark.asyncio
+  async def test_flush_empty_caches(self):
+    """Test flushing when caches are empty."""
+    invocation_context = await testing_utils.create_invocation_context(
+        testing_utils.create_test_agent()
+    )
+
+    # Set up mock artifact service
+    mock_artifact_service = AsyncMock()
+    invocation_context.artifact_service = mock_artifact_service
+
+    # Flush empty caches (should not error)
+    await self.manager.flush_caches(invocation_context)
+
+    # Verify artifact service was not called
+    mock_artifact_service.save_artifact.assert_not_called()
+
+  @pytest.mark.asyncio
+  async def test_flush_without_artifact_service(self):
+    """Test flushing when no artifact service is available."""
+    invocation_context = await testing_utils.create_invocation_context(
+        testing_utils.create_test_agent()
+    )
+
+    # No artifact service
+    invocation_context.artifact_service = None
+
+    # Cache some audio
+    input_blob = types.Blob(data=b'input_data', mime_type='audio/pcm')
+    self.manager.cache_audio(invocation_context, input_blob, 'input')
+
+    # Flush should not error but should not clear cache either
+    await self.manager.flush_caches(invocation_context)
+
+    # Cache should remain (no actual flushing happened)
+    assert len(invocation_context.input_realtime_cache) == 1
+
+  @pytest.mark.asyncio
+  async def test_flush_artifact_creation(self):
+    """Test that artifacts are created correctly during flush."""
+    invocation_context = await testing_utils.create_invocation_context(
+        testing_utils.create_test_agent()
+    )
+
+    # Set up mock services
+    mock_artifact_service = AsyncMock()
+    mock_artifact_service.save_artifact.return_value = 456
+    mock_session_service = AsyncMock()
+
+    invocation_context.artifact_service = mock_artifact_service
+    invocation_context.session_service = mock_session_service
+
+    # Cache audio with specific data
+    test_data = b'specific_test_audio_data'
+    audio_blob = types.Blob(data=test_data, mime_type='audio/pcm')
+    self.manager.cache_audio(invocation_context, audio_blob, 'input')
+
+    # Flush cache
+    await self.manager.flush_caches(invocation_context)
+
+    # Verify artifact was saved with correct data
+    mock_artifact_service.save_artifact.assert_called_once()
+    call_args = mock_artifact_service.save_artifact.call_args
+    saved_artifact = call_args.kwargs['artifact']
+    assert saved_artifact.inline_data.data == test_data
+    assert saved_artifact.inline_data.mime_type == 'audio/pcm'
+
+    # Verify session event was created
+    mock_session_service.append_event.assert_called_once()
+
+  def test_get_cache_stats_empty(self):
+    """Test getting statistics for empty caches."""
+    invocation_context = Mock()
+    invocation_context.input_realtime_cache = None
+    invocation_context.output_realtime_cache = None
+
+    stats = self.manager.get_cache_stats(invocation_context)
+
+    expected = {
+        'input_chunks': 0,
+        'output_chunks': 0,
+        'input_bytes': 0,
+        'output_bytes': 0,
+        'total_chunks': 0,
+        'total_bytes': 0,
+    }
+    assert stats == expected
+
+  @pytest.mark.asyncio
+  async def test_get_cache_stats_with_data(self):
+    """Test getting statistics for caches with data."""
+    invocation_context = await testing_utils.create_invocation_context(
+        testing_utils.create_test_agent()
+    )
+
+    # Cache some audio data of different sizes
+    input_blob1 = types.Blob(data=b'12345', mime_type='audio/pcm')  # 5 bytes
+    input_blob2 = types.Blob(
+        data=b'1234567890', mime_type='audio/pcm'
+    )  # 10 bytes
+    output_blob = types.Blob(data=b'abc', mime_type='audio/wav')  # 3 bytes
+
+    self.manager.cache_audio(invocation_context, input_blob1, 'input')
+    self.manager.cache_audio(invocation_context, input_blob2, 'input')
+    self.manager.cache_audio(invocation_context, output_blob, 'output')
+
+    stats = self.manager.get_cache_stats(invocation_context)
+
+    expected = {
+        'input_chunks': 2,
+        'output_chunks': 1,
+        'input_bytes': 15,  # 5 + 10
+        'output_bytes': 3,
+        'total_chunks': 3,
+        'total_bytes': 18,  # 15 + 3
+    }
+    assert stats == expected
+
+  @pytest.mark.asyncio
+  async def test_error_handling_in_flush(self):
+    """Test error handling during cache flush operations."""
+    invocation_context = await testing_utils.create_invocation_context(
+        testing_utils.create_test_agent()
+    )
+
+    # Set up mock artifact service that raises an error
+    mock_artifact_service = AsyncMock()
+    mock_artifact_service.save_artifact.side_effect = Exception(
+        'Artifact service error'
+    )
+    invocation_context.artifact_service = mock_artifact_service
+
+    # Cache some audio
+    audio_blob = types.Blob(data=b'test_data', mime_type='audio/pcm')
+    self.manager.cache_audio(invocation_context, audio_blob, 'input')
+
+    # Flush should not raise exception but should log error and retain cache
+    await self.manager.flush_caches(invocation_context)
+
+    # Cache should remain since flush failed
+    assert len(invocation_context.input_realtime_cache) == 1
+
+  @pytest.mark.asyncio
+  async def test_filename_uses_first_chunk_timestamp(self):
+    """Test that the filename timestamp comes from the first audio chunk, not flush time."""
+    invocation_context = await testing_utils.create_invocation_context(
+        testing_utils.create_test_agent()
+    )
+
+    # Set up mock services
+    mock_artifact_service = AsyncMock()
+    mock_artifact_service.save_artifact.return_value = 789
+    mock_session_service = AsyncMock()
+
+    invocation_context.artifact_service = mock_artifact_service
+    invocation_context.session_service = mock_session_service
+
+    # Cache multiple audio chunks with specific timestamps
+    first_timestamp = 1234567890.123  # First chunk timestamp
+    second_timestamp = 1234567891.456  # Second chunk timestamp (later)
+
+    # Manually create audio cache entries with specific timestamps
+    invocation_context.input_realtime_cache = []
+
+    from google.adk.agents.invocation_context import RealtimeCacheEntry
+
+    first_entry = RealtimeCacheEntry(
+        role='user',
+        data=types.Blob(data=b'first_chunk', mime_type='audio/pcm'),
+        timestamp=first_timestamp,
+    )
+
+    second_entry = RealtimeCacheEntry(
+        role='user',
+        data=types.Blob(data=b'second_chunk', mime_type='audio/pcm'),
+        timestamp=second_timestamp,
+    )
+
+    invocation_context.input_realtime_cache.extend([first_entry, second_entry])
+
+    # Sleep briefly to ensure current time is different from first timestamp
+    time.sleep(0.01)
+
+    # Flush cache
+    await self.manager.flush_caches(invocation_context)
+
+    # Verify artifact was saved
+    mock_artifact_service.save_artifact.assert_called_once()
+    call_args = mock_artifact_service.save_artifact.call_args
+    filename = call_args.kwargs['filename']
+
+    # Extract timestamp from filename (format: input_audio_{timestamp}.pcm)
+    expected_timestamp_ms = int(first_timestamp * 1000)
+    assert (
+        f'adk_live_audio_storage_input_audio_{expected_timestamp_ms}.pcm'
+        == filename
+    )
+
+    # Verify the timestamp in filename matches first chunk, not current time
+    current_timestamp_ms = int(time.time() * 1000)
+    assert expected_timestamp_ms != current_timestamp_ms  # Should be different
+    assert filename.startswith(
+        f'adk_live_audio_storage_input_audio_{expected_timestamp_ms}'
+    )

tests/unittests/flows/llm_flows/test_transcription_manager.py

@@ -0,0 +1,312 @@
+# Copyright 2025 Google LLC
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+from unittest.mock import AsyncMock
+from unittest.mock import Mock
+
+from google.adk.flows.llm_flows.transcription_manager import TranscriptionManager
+from google.genai import types
+import pytest
+
+from ... import testing_utils
+
+
+class TestTranscriptionManager:
+  """Test the TranscriptionManager class."""
+
+  def setup_method(self):
+    """Set up test fixtures."""
+    self.manager = TranscriptionManager()
+
+  @pytest.mark.asyncio
+  async def test_handle_input_transcription(self):
+    """Test handling user input transcription events."""
+    invocation_context = await testing_utils.create_invocation_context(
+        testing_utils.create_test_agent()
+    )
+
+    # Set up mock session service
+    mock_session_service = AsyncMock()
+    invocation_context.session_service = mock_session_service
+
+    # Create test transcription
+    transcription = types.Transcription(text='Hello from user')
+
+    # Handle transcription
+    await self.manager.handle_input_transcription(
+        invocation_context, transcription
+    )
+
+    # Verify session service was called
+    mock_session_service.append_event.assert_called_once()
+
+    # Check the event that was created
+    call_args = mock_session_service.append_event.call_args
+    event = call_args[0][1]  # Second argument is the event
+
+    assert event.author == 'user'
+    assert event.input_transcription == transcription
+    assert event.output_transcription is None
+    assert event.invocation_id == invocation_context.invocation_id
+    assert isinstance(event.timestamp, float)
+
+  @pytest.mark.asyncio
+  async def test_handle_output_transcription(self):
+    """Test handling model output transcription events."""
+    agent = testing_utils.create_test_agent()
+    invocation_context = await testing_utils.create_invocation_context(agent)
+
+    # Set up mock session service
+    mock_session_service = AsyncMock()
+    invocation_context.session_service = mock_session_service
+
+    # Create test transcription
+    transcription = types.Transcription(text='Hello from model')
+
+    # Handle transcription
+    await self.manager.handle_output_transcription(
+        invocation_context, transcription
+    )
+
+    # Verify session service was called
+    mock_session_service.append_event.assert_called_once()
+
+    # Check the event that was created
+    call_args = mock_session_service.append_event.call_args
+    event = call_args[0][1]  # Second argument is the event
+
+    assert event.author == agent.name
+    assert event.input_transcription is None
+    assert event.output_transcription == transcription
+    assert event.invocation_id == invocation_context.invocation_id
+    assert isinstance(event.timestamp, float)
+
+  @pytest.mark.asyncio
+  async def test_handle_multiple_transcriptions(self):
+    """Test handling multiple transcription events."""
+    invocation_context = await testing_utils.create_invocation_context(
+        testing_utils.create_test_agent()
+    )
+
+    # Set up mock session service
+    mock_session_service = AsyncMock()
+    invocation_context.session_service = mock_session_service
+
+    # Handle multiple input transcriptions
+    for i in range(3):
+      transcription = types.Transcription(text=f'User message {i}')
+      await self.manager.handle_input_transcription(
+          invocation_context, transcription
+      )
+
+    # Handle multiple output transcriptions
+    for i in range(2):
+      transcription = types.Transcription(text=f'Model response {i}')
+      await self.manager.handle_output_transcription(
+          invocation_context, transcription
+      )
+
+    # Verify session service was called for each transcription
+    assert mock_session_service.append_event.call_count == 5
+
+  @pytest.mark.asyncio
+  async def test_error_handling_input_transcription(self):
+    """Test error handling during input transcription processing."""
+    invocation_context = await testing_utils.create_invocation_context(
+        testing_utils.create_test_agent()
+    )
+
+    # Set up mock session service that raises an error
+    mock_session_service = AsyncMock()
+    mock_session_service.append_event.side_effect = Exception(
+        'Session service error'
+    )
+    invocation_context.session_service = mock_session_service
+
+    # Create test transcription
+    transcription = types.Transcription(text='Test transcription')
+
+    # Handle transcription should raise the exception
+    with pytest.raises(Exception, match='Session service error'):
+      await self.manager.handle_input_transcription(
+          invocation_context, transcription
+      )
+
+  @pytest.mark.asyncio
+  async def test_error_handling_output_transcription(self):
+    """Test error handling during output transcription processing."""
+    invocation_context = await testing_utils.create_invocation_context(
+        testing_utils.create_test_agent()
+    )
+
+    # Set up mock session service that raises an error
+    mock_session_service = AsyncMock()
+    mock_session_service.append_event.side_effect = Exception(
+        'Session service error'
+    )
+    invocation_context.session_service = mock_session_service
+
+    # Create test transcription
+    transcription = types.Transcription(text='Test transcription')
+
+    # Handle transcription should raise the exception
+    with pytest.raises(Exception, match='Session service error'):
+      await self.manager.handle_output_transcription(
+          invocation_context, transcription
+      )
+
+  def test_get_transcription_stats_empty_session(self):
+    """Test getting transcription statistics for empty session."""
+    invocation_context = Mock()
+    invocation_context.session.events = []
+
+    stats = self.manager.get_transcription_stats(invocation_context)
+
+    expected = {
+        'input_transcriptions': 0,
+        'output_transcriptions': 0,
+        'total_transcriptions': 0,
+    }
+    assert stats == expected
+
+  def test_get_transcription_stats_with_events(self):
+    """Test getting transcription statistics for session with events."""
+    invocation_context = Mock()
+
+    # Create mock events
+    input_event1 = Mock()
+    input_event1.input_transcription = types.Transcription(text='User 1')
+    input_event1.output_transcription = None
+
+    input_event2 = Mock()
+    input_event2.input_transcription = types.Transcription(text='User 2')
+    input_event2.output_transcription = None
+
+    output_event = Mock()
+    output_event.input_transcription = None
+    output_event.output_transcription = types.Transcription(
+        text='Model response'
+    )
+
+    regular_event = Mock()
+    regular_event.input_transcription = None
+    regular_event.output_transcription = None
+
+    invocation_context.session.events = [
+        input_event1,
+        output_event,
+        input_event2,
+        regular_event,
+    ]
+
+    stats = self.manager.get_transcription_stats(invocation_context)
+
+    expected = {
+        'input_transcriptions': 2,
+        'output_transcriptions': 1,
+        'total_transcriptions': 3,
+    }
+    assert stats == expected
+
+  def test_get_transcription_stats_missing_attributes(self):
+    """Test getting transcription statistics when events don't have transcription attributes."""
+    invocation_context = Mock()
+
+    # Create mock events and explicitly set transcription attributes to None
+    event1 = Mock()
+    event1.input_transcription = None
+    event1.output_transcription = None
+
+    event2 = Mock()
+    event2.input_transcription = None
+    event2.output_transcription = None
+
+    invocation_context.session.events = [event1, event2]
+
+    stats = self.manager.get_transcription_stats(invocation_context)
+
+    expected = {
+        'input_transcriptions': 0,
+        'output_transcriptions': 0,
+        'total_transcriptions': 0,
+    }
+    assert stats == expected
+
+  @pytest.mark.asyncio
+  async def test_transcription_event_fields(self):
+    """Test that transcription events have correct field values."""
+    invocation_context = await testing_utils.create_invocation_context(
+        testing_utils.create_test_agent()
+    )
+
+    # Set up mock session service
+    mock_session_service = AsyncMock()
+    invocation_context.session_service = mock_session_service
+
+    # Create test transcription with specific content
+    transcription = types.Transcription(
+        text='Test transcription content', finished=True
+    )
+
+    # Handle input transcription
+    await self.manager.handle_input_transcription(
+        invocation_context, transcription
+    )
+
+    # Verify the event structure
+    call_args = mock_session_service.append_event.call_args
+    event = call_args[0][1]
+
+    # Check all required fields are present
+    assert hasattr(event, 'id')
+    assert hasattr(event, 'invocation_id')
+    assert hasattr(event, 'author')
+    assert hasattr(event, 'input_transcription')
+    assert hasattr(event, 'output_transcription')
+    assert hasattr(event, 'timestamp')
+
+    # Check values
+    assert event.id is not None
+    assert event.invocation_id == invocation_context.invocation_id
+    assert event.author == 'user'
+    assert event.input_transcription == transcription
+    assert event.output_transcription is None
+
+  @pytest.mark.asyncio
+  async def test_transcription_with_different_data_types(self):
+    """Test handling transcriptions with different data types."""
+    invocation_context = await testing_utils.create_invocation_context(
+        testing_utils.create_test_agent()
+    )
+
+    # Set up mock session service
+    mock_session_service = AsyncMock()
+    invocation_context.session_service = mock_session_service
+
+    # Test with transcription that has basic fields only
+    transcription = types.Transcription(
+        text='Advanced transcription', finished=True
+    )
+
+    # Handle transcription
+    await self.manager.handle_input_transcription(
+        invocation_context, transcription
+    )
+
+    # Verify the transcription object is preserved as-is
+    call_args = mock_session_service.append_event.call_args
+    event = call_args[0][1]
+
+    assert event.input_transcription == transcription
+    assert event.input_transcription.text == 'Advanced transcription'

tests/unittests/streaming/test_streaming_audio_storage.py

@@ -0,0 +1,241 @@
+# Copyright 2025 Google LLC
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import asyncio
+import time
+
+from google.adk.agents import Agent
+from google.adk.agents import LiveRequestQueue
+from google.adk.agents.invocation_context import RealtimeCacheEntry
+from google.adk.agents.run_config import RunConfig
+from google.adk.events.event import Event
+from google.adk.models import LlmResponse
+from google.genai import types
+import pytest
+
+from .. import testing_utils
+
+
+def test_audio_caching_direct():
+  """Test audio caching logic directly without full live streaming."""
+  # This test directly verifies that our audio caching logic works
+  audio_data = b'\x00\xFF\x01\x02\x03\x04\x05\x06'
+  audio_mime_type = 'audio/pcm'
+
+  # Create mock responses for successful completion
+  responses = [
+      LlmResponse(
+          content=types.Content(
+              role='model',
+              parts=[types.Part.from_text(text='Processing audio...')],
+          ),
+          turn_complete=False,
+      ),
+      LlmResponse(turn_complete=True),  # This should trigger flush
+  ]
+
+  mock_model = testing_utils.MockModel.create(responses)
+  mock_model.model = 'gemini-2.0-flash-exp'  # For CFC support
+
+  root_agent = Agent(
+      name='test_agent',
+      model=mock_model,
+      tools=[],
+  )
+
+  # Test our implementation by directly calling it
+  async def test_caching():
+    # Create context similar to what would be created in real scenario
+    invocation_context = await testing_utils.create_invocation_context(
+        root_agent, run_config=RunConfig(support_cfc=True)
+    )
+
+    # Import our caching classes
+    from google.adk.agents.invocation_context import RealtimeCacheEntry
+    from google.adk.flows.llm_flows.base_llm_flow import BaseLlmFlow
+
+    # Create a mock flow to test our methods
+    flow = BaseLlmFlow()
+
+    # Test adding audio to cache
+    invocation_context.input_realtime_cache = []
+    audio_entry = RealtimeCacheEntry(
+        role='user',
+        data=types.Blob(data=audio_data, mime_type=audio_mime_type),
+        timestamp=1234567890.0,
+    )
+    invocation_context.input_realtime_cache.append(audio_entry)
+
+    # Verify cache has data
+    assert len(invocation_context.input_realtime_cache) == 1
+    assert invocation_context.input_realtime_cache[0].data.data == audio_data
+
+    # Test flushing cache
+    await flow._handle_control_event_flush(invocation_context, responses[-1])
+
+    # Verify cache was cleared
+    assert len(invocation_context.input_realtime_cache) == 0
+
+    # Check if artifacts were created
+    artifact_keys = (
+        await invocation_context.artifact_service.list_artifact_keys(
+            app_name=invocation_context.app_name,
+            user_id=invocation_context.user_id,
+            session_id=invocation_context.session.id,
+        )
+    )
+
+    # Should have at least one audio artifact
+    audio_artifacts = [key for key in artifact_keys if 'audio' in key.lower()]
+    assert (
+        len(audio_artifacts) > 0
+    ), f'Expected audio artifacts, found: {artifact_keys}'
+
+    # Verify artifact content
+    if audio_artifacts:
+      artifact = await invocation_context.artifact_service.load_artifact(
+          app_name=invocation_context.app_name,
+          user_id=invocation_context.user_id,
+          session_id=invocation_context.session.id,
+          filename=audio_artifacts[0],
+      )
+      assert artifact.inline_data.data == audio_data
+
+    return True
+
+  # Run the async test
+  result = asyncio.run(test_caching())
+  assert result is True
+
+
+def test_transcription_handling():
+  """Test that transcriptions are properly handled and saved to session service."""
+
+  # Create mock responses with transcriptions
+  input_transcription = types.Transcription(
+      text='Hello, this is transcribed input', finished=True
+  )
+  output_transcription = types.Transcription(
+      text='This is transcribed output', finished=True
+  )
+
+  responses = [
+      LlmResponse(
+          content=types.Content(
+              role='model', parts=[types.Part.from_text(text='Processing...')]
+          ),
+          turn_complete=False,
+      ),
+      LlmResponse(input_transcription=input_transcription, turn_complete=False),
+      LlmResponse(
+          output_transcription=output_transcription, turn_complete=False
+      ),
+      LlmResponse(turn_complete=True),
+  ]
+
+  mock_model = testing_utils.MockModel.create(responses)
+  mock_model.model = 'gemini-2.0-flash-exp'
+
+  root_agent = Agent(
+      name='test_agent',
+      model=mock_model,
+      tools=[],
+  )
+
+  async def test_transcription():
+    # Create context
+    invocation_context = await testing_utils.create_invocation_context(
+        root_agent, run_config=RunConfig(support_cfc=True)
+    )
+
+    from google.adk.events.event import Event
+    from google.adk.flows.llm_flows.base_llm_flow import BaseLlmFlow
+
+    flow = BaseLlmFlow()
+
+    # Test processing transcription events
+    session_events_before = len(invocation_context.session.events)
+
+    # Simulate input transcription event
+    input_event = Event(
+        id=Event.new_id(),
+        invocation_id=invocation_context.invocation_id,
+        author='user',
+        input_transcription=input_transcription,
+    )
+
+    # Simulate output transcription event
+    output_event = Event(
+        id=Event.new_id(),
+        invocation_id=invocation_context.invocation_id,
+        author=invocation_context.agent.name,
+        output_transcription=output_transcription,
+    )
+
+    # Save transcription events to session
+    await invocation_context.session_service.append_event(
+        invocation_context.session, input_event
+    )
+    await invocation_context.session_service.append_event(
+        invocation_context.session, output_event
+    )
+
+    # Verify transcriptions were saved to session
+    session_events_after = len(invocation_context.session.events)
+    assert session_events_after == session_events_before + 2
+
+    # Check that transcription events were saved
+    transcription_events = [
+        event
+        for event in invocation_context.session.events
+        if hasattr(event, 'input_transcription')
+        and event.input_transcription
+        or hasattr(event, 'output_transcription')
+        and event.output_transcription
+    ]
+    assert len(transcription_events) >= 2
+
+    # Verify input transcription
+    input_transcription_events = [
+        event
+        for event in invocation_context.session.events
+        if hasattr(event, 'input_transcription') and event.input_transcription
+    ]
+    assert len(input_transcription_events) >= 1
+    assert (
+        input_transcription_events[0].input_transcription.text
+        == 'Hello, this is transcribed input'
+    )
+    assert input_transcription_events[0].author == 'user'
+
+    # Verify output transcription
+    output_transcription_events = [
+        event
+        for event in invocation_context.session.events
+        if hasattr(event, 'output_transcription') and event.output_transcription
+    ]
+    assert len(output_transcription_events) >= 1
+    assert (
+        output_transcription_events[0].output_transcription.text
+        == 'This is transcribed output'
+    )
+    assert (
+        output_transcription_events[0].author == invocation_context.agent.name
+    )
+
+    return True
+
+  # Run the async test
+  result = asyncio.run(test_transcription())
+  assert result is True

tests/unittests/testing_utils.py

@@ -41,6 +41,18 @@ from google.genai.types import Part
 from typing_extensions import override
 
 
+def create_test_agent(name: str = 'test_agent') -> LlmAgent:
+  """Create a simple test agent for use in unit tests.
+
+  Args:
+    name: The name of the test agent.
+
+  Returns:
+    A configured LlmAgent instance suitable for testing.
+  """
+  return LlmAgent(name=name)
+
+
 class UserContent(types.Content):
 
   def __init__(self, text_or_part: str):