fix: Improve handling of partial and complete transcriptions in live calls

In `gemini_llm_connection.py`, accumulate partial transcription texts and emit `LlmResponse` with `partial=True` for each chunk. When the transcription is marked as `finished`, emit a final `LlmResponse` with the full accumulated text and `partial=False`.

In `runners.py`, modify `_should_append_to_history` to only add transcription events to the history when they are fully finished, preventing partial transcriptions from being added.

Co-authored-by: Hangfei Lin <hangfei@google.com>
PiperOrigin-RevId: 829029715

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

@@ -587,21 +587,19 @@ class BaseLlmFlow(ABC):
 
     # Handle transcription events ONCE per llm_response, outside the event loop
     if llm_response.input_transcription:
-      input_transcription_event = (
-          await self.transcription_manager.handle_input_transcription(
-              invocation_context, llm_response.input_transcription
-          )
+      model_response_event.input_transcription = (
+          llm_response.input_transcription
       )
-      yield input_transcription_event
+      model_response_event.partial = llm_response.partial
+      yield model_response_event
       return
 
     if llm_response.output_transcription:
-      output_transcription_event = (
-          await self.transcription_manager.handle_output_transcription(
-              invocation_context, llm_response.output_transcription
-          )
+      model_response_event.output_transcription = (
+          llm_response.output_transcription
       )
-      yield output_transcription_event
+      model_response_event.partial = llm_response.partial
+      yield model_response_event
       return
 
     # Flush audio caches based on control events using configurable settings

src/google/adk/models/gemini_llm_connection.py

@@ -35,6 +35,8 @@ class GeminiLlmConnection(BaseLlmConnection):
 
   def __init__(self, gemini_session: live.AsyncSession):
     self._gemini_session = gemini_session
+    self._input_transcription_text: str = ''
+    self._output_transcription_text: str = ''
 
   async def send_history(self, history: list[types.Content]):
     """Sends the conversation history to the gemini model.
@@ -166,15 +168,49 @@ class GeminiLlmConnection(BaseLlmConnection):
               text = ''
             yield llm_response
           if message.server_content.input_transcription:
-            llm_response = LlmResponse(
-                input_transcription=message.server_content.input_transcription,
-            )
-            yield llm_response
+            if message.server_content.input_transcription.text:
+              self._input_transcription_text += (
+                  message.server_content.input_transcription.text
+              )
+              yield LlmResponse(
+                  input_transcription=types.Transcription(
+                      text=message.server_content.input_transcription.text,
+                      finished=False,
+                  ),
+                  partial=True,
+              )
+            # finished=True and partial transcription may happen in the same
+            # message.
+            if message.server_content.input_transcription.finished:
+              yield LlmResponse(
+                  input_transcription=types.Transcription(
+                      text=self._input_transcription_text,
+                      finished=True,
+                  ),
+                  partial=False,
+              )
+              self._input_transcription_text = ''
           if message.server_content.output_transcription:
-            llm_response = LlmResponse(
-                output_transcription=message.server_content.output_transcription
-            )
-            yield llm_response
+            if message.server_content.output_transcription.text:
+              self._output_transcription_text += (
+                  message.server_content.output_transcription.text
+              )
+              yield LlmResponse(
+                  output_transcription=types.Transcription(
+                      text=message.server_content.output_transcription.text,
+                      finished=False,
+                  ),
+                  partial=True,
+              )
+            if message.server_content.output_transcription.finished:
+              yield LlmResponse(
+                  output_transcription=types.Transcription(
+                      text=self._output_transcription_text,
+                      finished=True,
+                  ),
+                  partial=False,
+              )
+              self._output_transcription_text = ''
           if message.server_content.turn_complete:
             if text:
               yield self.__build_full_text_response(text)
@@ -188,10 +224,12 @@ class GeminiLlmConnection(BaseLlmConnection):
           # in case it's an interrupted message, we merge the previous partial
           # text. Other we don't merge. because content can be none when model
           # safety threshold is triggered
-          if message.server_content.interrupted and text:
-            yield self.__build_full_text_response(text)
-            text = ''
-          yield LlmResponse(interrupted=message.server_content.interrupted)
+          if message.server_content.interrupted:
+            if text:
+              yield self.__build_full_text_response(text)
+              text = ''
+            else:
+              yield LlmResponse(interrupted=message.server_content.interrupted)
         if message.tool_call:
           if text:
             yield self.__build_full_text_response(text)

src/google/adk/runners.py

@@ -588,6 +588,10 @@ class Runner:
     # Don't append audio response from model in live mode to session.
     # The data is appended to artifacts with a reference in file_data in the
     # event.
+    # We should append non-partial events only.For example, non-finished(partial)
+    # transcription events should not be appended.
+    # Function call and function response events should be appended.
+    # Other control events should be appended.
     if is_live_call and contents._is_live_model_audio_event(event):
       return False
     return True