livekit-plugins-google 0.9.0__tar.gz → 0.10.0__tar.gz

{livekit_plugins_google-0.9.0 → livekit_plugins_google-0.10.0}/PKG-INFO

@@ -1,6 +1,6 @@
-Metadata-Version: 2.1
+Metadata-Version: 2.2
 Name: livekit-plugins-google
-Version: 0.9.0
+Version: 0.10.0
 Summary: Agent Framework plugin for services from Google Cloud
 Home-page: https://github.com/livekit/agents
 License: Apache-2.0
@@ -22,8 +22,18 @@ Description-Content-Type: text/markdown
 Requires-Dist: google-auth<3,>=2
 Requires-Dist: google-cloud-speech<3,>=2
 Requires-Dist: google-cloud-texttospeech<3,>=2
-Requires-Dist: google-genai>=0.3.0
+Requires-Dist: google-genai==0.5.0
 Requires-Dist: livekit-agents>=0.12.3
+Dynamic: classifier
+Dynamic: description
+Dynamic: description-content-type
+Dynamic: home-page
+Dynamic: keywords
+Dynamic: license
+Dynamic: project-url
+Dynamic: requires-dist
+Dynamic: requires-python
+Dynamic: summary
 
 # LiveKit Plugins Google
 

{livekit_plugins_google-0.9.0 → livekit_plugins_google-0.10.0}/livekit/plugins/google/__init__.py

@@ -13,11 +13,12 @@
 # limitations under the License.
 
 from . import beta
+from .llm import LLM
 from .stt import STT, SpeechStream
 from .tts import TTS
 from .version import __version__
 
-__all__ = ["STT", "TTS", "SpeechStream", "__version__", "beta"]
+__all__ = ["STT", "TTS", "SpeechStream", "__version__", "beta", "LLM"]
 from livekit.agents import Plugin
 
 from .log import logger

livekit_plugins_google-0.10.0/livekit/plugins/google/_utils.py

@@ -0,0 +1,202 @@
+from __future__ import annotations
+
+import base64
+import inspect
+import json
+from typing import Any, Dict, List, Optional, get_args, get_origin
+
+from livekit import rtc
+from livekit.agents import llm, utils
+from livekit.agents.llm.function_context import _is_optional_type
+
+from google.genai import types
+
+JSON_SCHEMA_TYPE_MAP: dict[type, types.Type] = {
+    str: "STRING",
+    int: "INTEGER",
+    float: "NUMBER",
+    bool: "BOOLEAN",
+    dict: "OBJECT",
+    list: "ARRAY",
+}
+
+__all__ = ["_build_gemini_ctx", "_build_tools"]
+
+
+def _build_parameters(arguments: Dict[str, Any]) -> types.Schema | None:
+    properties: Dict[str, types.Schema] = {}
+    required: List[str] = []
+
+    for arg_name, arg_info in arguments.items():
+        prop = types.Schema()
+        if arg_info.description:
+            prop.description = arg_info.description
+
+        _, py_type = _is_optional_type(arg_info.type)
+        origin = get_origin(py_type)
+        if origin is list:
+            item_type = get_args(py_type)[0]
+            if item_type not in JSON_SCHEMA_TYPE_MAP:
+                raise ValueError(f"Unsupported type: {item_type}")
+            prop.type = "ARRAY"
+            prop.items = types.Schema(type=JSON_SCHEMA_TYPE_MAP[item_type])
+
+            if arg_info.choices:
+                prop.items.enum = arg_info.choices
+        else:
+            if py_type not in JSON_SCHEMA_TYPE_MAP:
+                raise ValueError(f"Unsupported type: {py_type}")
+
+            prop.type = JSON_SCHEMA_TYPE_MAP[py_type]
+
+            if arg_info.choices:
+                prop.enum = arg_info.choices
+                if py_type is int:
+                    raise ValueError(
+                        f"Parameter '{arg_info.name}' uses integer choices, not supported by this model."
+                    )
+
+        properties[arg_name] = prop
+
+        if arg_info.default is inspect.Parameter.empty:
+            required.append(arg_name)
+
+    if properties:
+        parameters = types.Schema(type="OBJECT", properties=properties)
+        if required:
+            parameters.required = required
+
+        return parameters
+
+    return None
+
+
+def _build_tools(fnc_ctx: Any) -> List[types.FunctionDeclaration]:
+    function_declarations: List[types.FunctionDeclaration] = []
+    for fnc_info in fnc_ctx.ai_functions.values():
+        parameters = _build_parameters(fnc_info.arguments)
+
+        func_decl = types.FunctionDeclaration(
+            name=fnc_info.name,
+            description=fnc_info.description,
+            parameters=parameters,
+        )
+
+        function_declarations.append(func_decl)
+    return function_declarations
+
+
+def _build_gemini_ctx(
+    chat_ctx: llm.ChatContext, cache_key: Any
+) -> tuple[list[types.Content], Optional[types.Content]]:
+    turns: list[types.Content] = []
+    system_instruction: Optional[types.Content] = None
+    current_role: Optional[str] = None
+    parts: list[types.Part] = []
+
+    for msg in chat_ctx.messages:
+        if msg.role == "system":
+            if isinstance(msg.content, str):
+                system_instruction = types.Content(parts=[types.Part(text=msg.content)])
+            continue
+
+        if msg.role == "assistant":
+            role = "model"
+        elif msg.role == "tool":
+            role = "user"
+        else:
+            role = "user"
+
+        # If role changed, finalize previous parts into a turn
+        if role != current_role:
+            if current_role is not None and parts:
+                turns.append(types.Content(role=current_role, parts=parts))
+            current_role = role
+            parts = []
+
+        if msg.tool_calls:
+            for fnc in msg.tool_calls:
+                parts.append(
+                    types.Part(
+                        function_call=types.FunctionCall(
+                            id=fnc.tool_call_id,
+                            name=fnc.function_info.name,
+                            args=fnc.arguments,
+                        )
+                    )
+                )
+
+        if msg.role == "tool":
+            if msg.content:
+                if isinstance(msg.content, dict):
+                    parts.append(
+                        types.Part(
+                            function_response=types.FunctionResponse(
+                                id=msg.tool_call_id,
+                                name=msg.name,
+                                response=msg.content,
+                            )
+                        )
+                    )
+                elif isinstance(msg.content, str):
+                    parts.append(
+                        types.Part(
+                            function_response=types.FunctionResponse(
+                                id=msg.tool_call_id,
+                                name=msg.name,
+                                response={"result": msg.content},
+                            )
+                        )
+                    )
+        else:
+            if msg.content:
+                if isinstance(msg.content, str):
+                    parts.append(types.Part(text=msg.content))
+                elif isinstance(msg.content, dict):
+                    parts.append(types.Part(text=json.dumps(msg.content)))
+                elif isinstance(msg.content, list):
+                    for item in msg.content:
+                        if isinstance(item, str):
+                            parts.append(types.Part(text=item))
+                        elif isinstance(item, llm.ChatImage):
+                            parts.append(_build_gemini_image_part(item, cache_key))
+
+    # Finalize last role's parts if any remain
+    if current_role is not None and parts:
+        turns.append(types.Content(role=current_role, parts=parts))
+
+    return turns, system_instruction
+
+
+def _build_gemini_image_part(image: llm.ChatImage, cache_key: Any) -> types.Part:
+    if isinstance(image.image, str):
+        # Check if the string is a Data URL
+        if image.image.startswith("data:image/jpeg;base64,"):
+            # Extract the base64 part after the comma
+            base64_data = image.image.split(",", 1)[1]
+            try:
+                image_bytes = base64.b64decode(base64_data)
+            except Exception as e:
+                raise ValueError("Invalid base64 data in image URL") from e
+
+            return types.Part.from_bytes(data=image_bytes, mime_type="image/jpeg")
+        else:
+            # Assume it's a regular URL
+            return types.Part.from_uri(file_uri=image.image, mime_type="image/jpeg")
+
+    elif isinstance(image.image, rtc.VideoFrame):
+        if cache_key not in image._cache:
+            opts = utils.images.EncodeOptions()
+            if image.inference_width and image.inference_height:
+                opts.resize_options = utils.images.ResizeOptions(
+                    width=image.inference_width,
+                    height=image.inference_height,
+                    strategy="scale_aspect_fit",
+                )
+            encoded_data = utils.images.encode(image.image, opts)
+            image._cache[cache_key] = base64.b64encode(encoded_data).decode("utf-8")
+
+        return types.Part.from_bytes(
+            data=image._cache[cache_key], mime_type="image/jpeg"
+        )
+    raise ValueError(f"Unsupported image type: {type(image.image)}")

{livekit_plugins_google-0.9.0 → livekit_plugins_google-0.10.0}/livekit/plugins/google/beta/realtime/__init__.py

@@ -1,7 +1,6 @@
 from .api_proto import (
     ClientEvents,
     LiveAPIModels,
-    ResponseModality,
     Voice,
 )
 from .realtime_api import RealtimeModel
@@ -10,6 +9,5 @@ __all__ = [
     "RealtimeModel",
     "ClientEvents",
     "LiveAPIModels",
-    "ResponseModality",
     "Voice",
 ]

livekit_plugins_google-0.10.0/livekit/plugins/google/beta/realtime/api_proto.py

@@ -0,0 +1,24 @@
+from __future__ import annotations
+
+from typing import Literal, Sequence, Union
+
+from google.genai import types
+
+from ..._utils import _build_gemini_ctx, _build_tools
+
+LiveAPIModels = Literal["gemini-2.0-flash-exp"]
+
+Voice = Literal["Puck", "Charon", "Kore", "Fenrir", "Aoede"]
+
+__all__ = ["_build_tools", "ClientEvents", "_build_gemini_ctx"]
+
+ClientEvents = Union[
+    types.ContentListUnion,
+    types.ContentListUnionDict,
+    types.LiveClientContentOrDict,
+    types.LiveClientRealtimeInput,
+    types.LiveClientRealtimeInputOrDict,
+    types.LiveClientToolResponseOrDict,
+    types.FunctionResponseOrDict,
+    Sequence[types.FunctionResponseOrDict],
+]

{livekit_plugins_google-0.9.0 → livekit_plugins_google-0.10.0}/livekit/plugins/google/beta/realtime/realtime_api.py

@@ -1,7 +1,6 @@
 from __future__ import annotations
 
 import asyncio
-import base64
 import json
 import os
 from dataclasses import dataclass
@@ -11,14 +10,22 @@ from livekit import rtc
 from livekit.agents import llm, utils
 from livekit.agents.llm.function_context import _create_ai_function_info
 
-from google import genai  # type: ignore
-from google.genai.types import (  # type: ignore
+from google import genai
+from google.genai._api_client import HttpOptions
+from google.genai.types import (
+    Blob,
+    Content,
     FunctionResponse,
-    GenerationConfigDict,
+    GenerationConfig,
+    LiveClientContent,
+    LiveClientRealtimeInput,
     LiveClientToolResponse,
-    LiveConnectConfigDict,
+    LiveConnectConfig,
+    Modality,
+    Part,
     PrebuiltVoiceConfig,
     SpeechConfig,
+    Tool,
     VoiceConfig,
 )
 
@@ -26,10 +33,11 @@ from ...log import logger
 from .api_proto import (
     ClientEvents,
     LiveAPIModels,
-    ResponseModality,
     Voice,
+    _build_gemini_ctx,
     _build_tools,
 )
+from .transcriber import TranscriberSession, TranscriptionContent
 
 EventTypes = Literal[
     "start_session",
@@ -39,6 +47,9 @@ EventTypes = Literal[
     "function_calls_collected",
     "function_calls_finished",
     "function_calls_cancelled",
+    "input_speech_transcription_completed",
+    "agent_speech_transcription_completed",
+    "agent_speech_stopped",
 ]
 
 
@@ -55,6 +66,12 @@ class GeminiContent:
     content_type: Literal["text", "audio"]
 
 
+@dataclass
+class InputTranscription:
+    item_id: str
+    transcript: str
+
+
 @dataclass
 class Capabilities:
     supports_truncate: bool
@@ -65,7 +82,7 @@ class ModelOptions:
     model: LiveAPIModels | str
     api_key: str | None
     voice: Voice | str
-    response_modalities: ResponseModality
+    response_modalities: list[Modality] | None
     vertexai: bool
     project: str | None
     location: str | None
@@ -76,18 +93,22 @@ class ModelOptions:
     top_k: int | None
     presence_penalty: float | None
     frequency_penalty: float | None
-    instructions: str
+    instructions: Content | None
+    enable_user_audio_transcription: bool
+    enable_agent_audio_transcription: bool
 
 
 class RealtimeModel:
     def __init__(
         self,
         *,
-        instructions: str = "",
+        instructions: str | None = None,
         model: LiveAPIModels | str = "gemini-2.0-flash-exp",
         api_key: str | None = None,
         voice: Voice | str = "Puck",
-        modalities: ResponseModality = "AUDIO",
+        modalities: list[Modality] = ["AUDIO"],
+        enable_user_audio_transcription: bool = True,
+        enable_agent_audio_transcription: bool = True,
         vertexai: bool = False,
         project: str | None = None,
         location: str | None = None,
@@ -103,15 +124,24 @@ class RealtimeModel:
         """
         Initializes a RealtimeModel instance for interacting with Google's Realtime API.
 
+        Environment Requirements:
+        - For VertexAI: Set the `GOOGLE_APPLICATION_CREDENTIALS` environment variable to the path of the service account key file.
+        The Google Cloud project and location can be set via `project` and `location` arguments or the environment variables
+        `GOOGLE_CLOUD_PROJECT` and `GOOGLE_CLOUD_LOCATION`. By default, the project is inferred from the service account key file,
+        and the location defaults to "us-central1".
+        - For Google Gemini API: Set the `api_key` argument or the `GOOGLE_API_KEY` environment variable.
+
         Args:
             instructions (str, optional): Initial system instructions for the model. Defaults to "".
-            api_key (str or None, optional): OpenAI API key. If None, will attempt to read from the environment variable OPENAI_API_KEY
-            modalities (ResponseModality): Modalities to use, such as ["TEXT", "AUDIO"]. Defaults to ["AUDIO"].
+            api_key (str or None, optional): Google Gemini API key. If None, will attempt to read from the environment variable GOOGLE_API_KEY.
+            modalities (list[Modality], optional): Modalities to use, such as ["TEXT", "AUDIO"]. Defaults to ["AUDIO"].
             model (str or None, optional): The name of the model to use. Defaults to "gemini-2.0-flash-exp".
             voice (api_proto.Voice, optional): Voice setting for audio outputs. Defaults to "Puck".
+            enable_user_audio_transcription (bool, optional): Whether to enable user audio transcription. Defaults to True
+            enable_agent_audio_transcription (bool, optional): Whether to enable agent audio transcription. Defaults to True
             temperature (float, optional): Sampling temperature for response generation. Defaults to 0.8.
             vertexai (bool, optional): Whether to use VertexAI for the API. Defaults to False.
-                project (str or None, optional): The project to use for the API. Defaults to None. (for vertexai)
+                project (str or None, optional): The project id to use for the API. Defaults to None. (for vertexai)
                 location (str or None, optional): The location to use for the API. Defaults to None. (for vertexai)
             candidate_count (int, optional): The number of candidate responses to generate. Defaults to 1.
             top_p (float, optional): The top-p value for response generation
@@ -130,21 +160,38 @@ class RealtimeModel:
         self._model = model
         self._loop = loop or asyncio.get_event_loop()
         self._api_key = api_key or os.environ.get("GOOGLE_API_KEY")
-        self._vertexai = vertexai
-        self._project_id = project or os.environ.get("GOOGLE_PROJECT")
-        self._location = location or os.environ.get("GOOGLE_LOCATION")
-        if self._api_key is None and not self._vertexai:
-            raise ValueError("GOOGLE_API_KEY is not set")
+        self._project = project or os.environ.get("GOOGLE_CLOUD_PROJECT")
+        self._location = location or os.environ.get("GOOGLE_CLOUD_LOCATION")
+        if vertexai:
+            if not self._project or not self._location:
+                raise ValueError(
+                    "Project and location are required for VertexAI either via project and location or GOOGLE_CLOUD_PROJECT and GOOGLE_CLOUD_LOCATION environment variables"
+                )
+            self._api_key = None  # VertexAI does not require an API key
+
+        else:
+            self._project = None
+            self._location = None
+            if not self._api_key:
+                raise ValueError(
+                    "API key is required for Google API either via api_key or GOOGLE_API_KEY environment variable"
+                )
+
+        instructions_content = (
+            Content(parts=[Part(text=instructions)]) if instructions else None
+        )
 
         self._rt_sessions: list[GeminiRealtimeSession] = []
         self._opts = ModelOptions(
             model=model,
-            api_key=api_key,
+            api_key=self._api_key,
             voice=voice,
+            enable_user_audio_transcription=enable_user_audio_transcription,
+            enable_agent_audio_transcription=enable_agent_audio_transcription,
             response_modalities=modalities,
             vertexai=vertexai,
-            project=project,
-            location=location,
+            project=self._project,
+            location=self._location,
             candidate_count=candidate_count,
             temperature=temperature,
             max_output_tokens=max_output_tokens,
@@ -152,7 +199,7 @@ class RealtimeModel:
             top_k=top_k,
             presence_penalty=presence_penalty,
             frequency_penalty=frequency_penalty,
-            instructions=instructions,
+            instructions=instructions_content,
         )
 
     @property
@@ -208,16 +255,16 @@ class GeminiRealtimeSession(utils.EventEmitter[EventTypes]):
         self._chat_ctx = chat_ctx
         self._fnc_ctx = fnc_ctx
         self._fnc_tasks = utils.aio.TaskSet()
+        self._is_interrupted = False
 
         tools = []
         if self._fnc_ctx is not None:
             functions = _build_tools(self._fnc_ctx)
-            tools.append({"function_declarations": functions})
+            tools.append(Tool(function_declarations=functions))
 
-        self._config = LiveConnectConfigDict(
-            model=self._opts.model,
+        self._config = LiveConnectConfig(
             response_modalities=self._opts.response_modalities,
-            generation_config=GenerationConfigDict(
+            generation_config=GenerationConfig(
                 candidate_count=self._opts.candidate_count,
                 temperature=self._opts.temperature,
                 max_output_tokens=self._opts.max_output_tokens,
@@ -237,7 +284,7 @@ class GeminiRealtimeSession(utils.EventEmitter[EventTypes]):
             tools=tools,
         )
         self._client = genai.Client(
-            http_options={"api_version": "v1alpha"},
+            http_options=HttpOptions(api_version="v1alpha"),
             api_key=self._opts.api_key,
             vertexai=self._opts.vertexai,
             project=self._opts.project,
@@ -246,12 +293,22 @@ class GeminiRealtimeSession(utils.EventEmitter[EventTypes]):
         self._main_atask = asyncio.create_task(
             self._main_task(), name="gemini-realtime-session"
         )
-        # dummy task to wait for the session to be initialized # TODO: sync chat ctx
-        self._init_sync_task = asyncio.create_task(
-            asyncio.sleep(0), name="gemini-realtime-session-init"
-        )
+        if self._opts.enable_user_audio_transcription:
+            self._transcriber = TranscriberSession(
+                client=self._client, model=self._opts.model
+            )
+            self._transcriber.on("input_speech_done", self._on_input_speech_done)
+        if self._opts.enable_agent_audio_transcription:
+            self._agent_transcriber = TranscriberSession(
+                client=self._client, model=self._opts.model
+            )
+            self._agent_transcriber.on("input_speech_done", self._on_agent_speech_done)
+        # init dummy task
+        self._init_sync_task = asyncio.create_task(asyncio.sleep(0))
         self._send_ch = utils.aio.Chan[ClientEvents]()
         self._active_response_id = None
+        if chat_ctx:
+            self.generate_reply(chat_ctx)
 
     async def aclose(self) -> None:
         if self._send_ch.closed:
@@ -269,32 +326,97 @@ class GeminiRealtimeSession(utils.EventEmitter[EventTypes]):
         self._fnc_ctx = value
 
     def _push_audio(self, frame: rtc.AudioFrame) -> None:
-        data = base64.b64encode(frame.data).decode("utf-8")
-        self._queue_msg({"mime_type": "audio/pcm", "data": data})
+        if self._opts.enable_user_audio_transcription:
+            self._transcriber._push_audio(frame)
+        realtime_input = LiveClientRealtimeInput(
+            media_chunks=[Blob(data=frame.data.tobytes(), mime_type="audio/pcm")],
+        )
+        self._queue_msg(realtime_input)
 
-    def _queue_msg(self, msg: dict) -> None:
+    def _queue_msg(self, msg: ClientEvents) -> None:
         self._send_ch.send_nowait(msg)
 
+    def generate_reply(
+        self,
+        ctx: llm.ChatContext | llm.ChatMessage,
+        turn_complete: bool = True,
+    ) -> None:
+        if isinstance(ctx, llm.ChatMessage) and isinstance(ctx.content, str):
+            new_chat_ctx = llm.ChatContext()
+            new_chat_ctx.append(text=ctx.content, role=ctx.role)
+        elif isinstance(ctx, llm.ChatContext):
+            new_chat_ctx = ctx
+        else:
+            raise ValueError("Invalid chat context")
+        turns, _ = _build_gemini_ctx(new_chat_ctx, id(self))
+        client_content = LiveClientContent(
+            turn_complete=turn_complete,
+            turns=turns,
+        )
+        self._queue_msg(client_content)
+
     def chat_ctx_copy(self) -> llm.ChatContext:
         return self._chat_ctx.copy()
 
     async def set_chat_ctx(self, ctx: llm.ChatContext) -> None:
         self._chat_ctx = ctx.copy()
 
+    def cancel_response(self) -> None:
+        raise NotImplementedError("cancel_response is not supported yet")
+
+    def create_response(
+        self,
+        on_duplicate: Literal[
+            "cancel_existing", "cancel_new", "keep_both"
+        ] = "keep_both",
+    ) -> None:
+        raise NotImplementedError("create_response is not supported yet")
+
+    def commit_audio_buffer(self) -> None:
+        raise NotImplementedError("commit_audio_buffer is not supported yet")
+
+    def server_vad_enabled(self) -> bool:
+        return True
+
+    def _on_input_speech_done(self, content: TranscriptionContent) -> None:
+        if content.response_id and content.text:
+            self.emit(
+                "input_speech_transcription_completed",
+                InputTranscription(
+                    item_id=content.response_id,
+                    transcript=content.text,
+                ),
+            )
+
+        # self._chat_ctx.append(text=content.text, role="user")
+        # TODO: implement sync mechanism to make sure the transcribed user speech is inside the chat_ctx and always before the generated agent speech
+
+    def _on_agent_speech_done(self, content: TranscriptionContent) -> None:
+        if not self._is_interrupted and content.response_id and content.text:
+            self.emit(
+                "agent_speech_transcription_completed",
+                InputTranscription(
+                    item_id=content.response_id,
+                    transcript=content.text,
+                ),
+            )
+            # self._chat_ctx.append(text=content.text, role="assistant")
+
     @utils.log_exceptions(logger=logger)
     async def _main_task(self):
         @utils.log_exceptions(logger=logger)
         async def _send_task():
             async for msg in self._send_ch:
-                await self._session.send(msg)
+                await self._session.send(input=msg)
 
-            await self._session.send(".", end_of_turn=True)
+            await self._session.send(input=".", end_of_turn=True)
 
         @utils.log_exceptions(logger=logger)
         async def _recv_task():
             while True:
                 async for response in self._session.receive():
                     if self._active_response_id is None:
+                        self._is_interrupted = False
                         self._active_response_id = utils.shortuuid()
                         text_stream = utils.aio.Chan[str]()
                         audio_stream = utils.aio.Chan[rtc.AudioFrame]()
@@ -307,7 +429,7 @@ class GeminiRealtimeSession(utils.EventEmitter[EventTypes]):
                             audio=[],
                             text_stream=text_stream,
                             audio_stream=audio_stream,
-                            content_type=self._opts.response_modalities,
+                            content_type="audio",
                         )
                         self.emit("response_content_added", content)
 
@@ -326,6 +448,8 @@ class GeminiRealtimeSession(utils.EventEmitter[EventTypes]):
                                         samples_per_channel=len(part.inline_data.data)
                                         // 2,
                                     )
+                                    if self._opts.enable_agent_audio_transcription:
+                                        self._agent_transcriber._push_audio(frame)
                                     content.audio_stream.send_nowait(frame)
 
                         if server_content.interrupted or server_content.turn_complete:
@@ -333,10 +457,8 @@ class GeminiRealtimeSession(utils.EventEmitter[EventTypes]):
                                 if isinstance(stream, utils.aio.Chan):
                                     stream.close()
 
-                            if server_content.interrupted:
-                                self.emit("input_speech_started")
-                            elif server_content.turn_complete:
-                                self.emit("response_content_done", content)
+                            self.emit("agent_speech_stopped")
+                            self._is_interrupted = True
 
                             self._active_response_id = None
 
@@ -387,6 +509,10 @@ class GeminiRealtimeSession(utils.EventEmitter[EventTypes]):
             finally:
                 await utils.aio.gracefully_cancel(*tasks)
                 await self._session.close()
+                if self._opts.enable_user_audio_transcription:
+                    await self._transcriber.aclose()
+                if self._opts.enable_agent_audio_transcription:
+                    await self._agent_transcriber.aclose()
 
     @utils.log_exceptions(logger=logger)
     async def _run_fnc_task(self, fnc_call_info: llm.FunctionCallInfo, item_id: str):
@@ -419,6 +545,6 @@ class GeminiRealtimeSession(utils.EventEmitter[EventTypes]):
                     )
                 ]
             )
-            await self._session.send(tool_response)
+            await self._session.send(input=tool_response)
 
             self.emit("function_calls_finished", [called_fnc])