Module livekit.plugins.azure.realtime
Azure Voice Live Realtime API support for LiveKit Agents.
Sub-modules
livekit.plugins.azure.realtime.realtime_modellivekit.plugins.azure.realtime.utils
Classes
class RealtimeModel (*,
endpoint: str | None = None,
model: str | None = None,
voice: str = 'en-US-AvaMultilingualNeural',
input_audio_transcription: NotGivenOr[AudioInputTranscriptionOptions | None] = NOT_GIVEN,
modalities: "NotGivenOr[list[Literal['text', 'audio']]]" = NOT_GIVEN,
turn_detection: NotGivenOr[TurnDetection | None] = NOT_GIVEN,
tool_choice: NotGivenOr[llm.ToolChoice | None] = NOT_GIVEN,
temperature: NotGivenOr[float] = NOT_GIVEN,
max_output_tokens: NotGivenOr[int] = NOT_GIVEN,
api_key: str | None = None,
use_default_credential: bool = False,
conn_options: APIConnectOptions = APIConnectOptions(max_retry=3, retry_interval=2.0, timeout=10.0))-
Expand source code
class RealtimeModel(llm.RealtimeModel): def __init__( self, *, endpoint: str | None = None, model: str | None = None, voice: str = DEFAULT_VOICE, input_audio_transcription: NotGivenOr[AudioInputTranscriptionOptions | None] = NOT_GIVEN, modalities: NotGivenOr[list[Literal["text", "audio"]]] = NOT_GIVEN, turn_detection: NotGivenOr[TurnDetection | None] = NOT_GIVEN, tool_choice: NotGivenOr[llm.ToolChoice | None] = NOT_GIVEN, temperature: NotGivenOr[float] = NOT_GIVEN, max_output_tokens: NotGivenOr[int] = NOT_GIVEN, api_key: str | None = None, use_default_credential: bool = False, conn_options: APIConnectOptions = DEFAULT_API_CONNECT_OPTIONS, ) -> None: """ Initialize Azure Voice Live Realtime model. Requires the optional dependencies: ``pip install 'livekit-plugins-azure[realtime]'``. Args: endpoint: Azure Voice Live endpoint URL. If None, reads from AZURE_VOICE_LIVE_ENDPOINT. model: Model name. If None, reads from AZURE_VOICE_LIVE_MODEL (default: "gpt-realtime"). voice: Voice for audio responses (default: "en-US-AvaMultilingualNeural"). input_audio_transcription: Configuration for input audio transcription. If NOT_GIVEN, uses azure-speech for the non-multimodal (text) models such as gpt-4.1 and for phi4-mm-realtime, which don't support whisper-1, and whisper-1 for the other models such as gpt-realtime. Set to None to disable transcription. Use AudioInputTranscriptionOptions to configure model and language, OpenAI transcription models such as whisper-1 take a single language, azure-speech up to 10 languages ("en-US,zh-CN"). modalities: List of modalities to enable (default: ["text", "audio"]). turn_detection: Turn detection configuration. Accepts ServerVad, AzureSemanticVad, AzureSemanticVadEn, or AzureSemanticVadMultilingual (default: ServerVad with threshold=0.5). tool_choice: Tool selection policy (default: "auto"). temperature: Sampling temperature (default: 0.8). max_output_tokens: Maximum output tokens (default: 4096). api_key: Azure API key. If None, reads from AZURE_VOICE_LIVE_API_KEY. use_default_credential: Use DefaultAzureCredential for auth instead of API key. conn_options: Connection retry and timeout options. Example: ```python from livekit.plugins.azure.realtime import RealtimeModel from azure.ai.voicelive.models import AudioInputTranscriptionOptions, ServerVad # English-only session (recommended for reliable language detection) model = RealtimeModel( endpoint=os.getenv("AZURE_VOICE_LIVE_ENDPOINT"), api_key=os.getenv("AZURE_VOICE_LIVE_API_KEY"), model="gpt-realtime", voice="en-US-AvaNeural", input_audio_transcription=AudioInputTranscriptionOptions( model="whisper-1", language="en-US", # Constrains transcription to English ), turn_detection=ServerVad(threshold=0.5, silence_duration_ms=500), ) # Multi-language session, azure-speech takes up to 10 languages, the first is primary model = RealtimeModel( endpoint=os.getenv("AZURE_VOICE_LIVE_ENDPOINT"), api_key=os.getenv("AZURE_VOICE_LIVE_API_KEY"), model="gpt-4.1", voice="en-US-AvaMultilingualNeural", input_audio_transcription=AudioInputTranscriptionOptions( model="azure-speech", language="en-US,zh-CN,ja-JP", # Allow English, Chinese, and Japanese ), ) ``` """ modalities_list: Sequence[Modality | str] = ( [Modality.TEXT if m == "text" else Modality.AUDIO for m in modalities] if is_given(modalities) else DEFAULT_MODALITIES ) turn_detection_val = to_turn_detection(turn_detection) # Get model from environment if not provided, the transcription default depends on it model_val = model or os.environ.get("AZURE_VOICE_LIVE_MODEL") or "gpt-realtime" input_audio_transcription_val = to_audio_transcription( input_audio_transcription, model=model_val ) super().__init__( capabilities=llm.RealtimeCapabilities( message_truncation=False, turn_detection=turn_detection_val is not None, user_transcription=input_audio_transcription_val is not None, auto_tool_reply_generation=False, # Tool responses handled via generate_reply audio_output=Modality.AUDIO in modalities_list, manual_function_calls=True, mutable_chat_context=False, mutable_instructions=True, mutable_tools=True, per_response_tool_choice=True, ) ) # Get endpoint from environment if not provided endpoint_val = endpoint or os.environ.get("AZURE_VOICE_LIVE_ENDPOINT") if not endpoint_val: raise ValueError( "Azure Voice Live endpoint must be provided via 'endpoint' parameter " "or AZURE_VOICE_LIVE_ENDPOINT environment variable" ) # Get API key if not using default credential api_key_val = api_key if not use_default_credential: api_key_val = api_key or os.environ.get("AZURE_VOICE_LIVE_API_KEY") if not api_key_val: raise ValueError( "Azure Voice Live API key must be provided via 'api_key' parameter " "or AZURE_VOICE_LIVE_API_KEY environment variable, " "or set use_default_credential=True" ) tool_choice_val: llm.ToolChoice | None = ( cast(llm.ToolChoice, tool_choice) if is_given(tool_choice) else None ) self._opts = _RealtimeOptions( endpoint=endpoint_val, model=model_val, voice=voice, input_audio_transcription=input_audio_transcription_val, tool_choice=tool_choice_val, turn_detection=turn_detection_val, input_audio_format=DEFAULT_INPUT_AUDIO_FORMAT, output_audio_format=DEFAULT_OUTPUT_AUDIO_FORMAT, modalities=modalities_list, temperature=temperature if is_given(temperature) else DEFAULT_TEMPERATURE, max_output_tokens=max_output_tokens if is_given(max_output_tokens) else DEFAULT_MAX_OUTPUT_TOKENS, api_key=api_key_val, use_default_credential=use_default_credential, conn_options=conn_options, ) self._sessions = weakref.WeakSet[RealtimeSession]() @property def model(self) -> str: return self._opts.model @property def provider(self) -> str: return "azure-voicelive" def session(self, *, turn_detection_disabled: bool = False) -> RealtimeSession: # manual turn-taking is unsupported (can_disable_turn_detection=False) sess = RealtimeSession(self) self._sessions.add(sess) return sess async def aclose(self) -> None: passInitialize Azure Voice Live Realtime model.
Requires the optional dependencies:
pip install 'livekit-plugins-azure[realtime]'.Args
endpoint- Azure Voice Live endpoint URL. If None, reads from AZURE_VOICE_LIVE_ENDPOINT.
model- Model name. If None, reads from AZURE_VOICE_LIVE_MODEL (default: "gpt-realtime").
voice- Voice for audio responses (default: "en-US-AvaMultilingualNeural").
input_audio_transcription- Configuration for input audio transcription. If NOT_GIVEN, uses azure-speech for the non-multimodal (text) models such as gpt-4.1 and for phi4-mm-realtime, which don't support whisper-1, and whisper-1 for the other models such as gpt-realtime. Set to None to disable transcription. Use AudioInputTranscriptionOptions to configure model and language, OpenAI transcription models such as whisper-1 take a single language, azure-speech up to 10 languages ("en-US,zh-CN").
modalities- List of modalities to enable (default: ["text", "audio"]).
turn_detection- Turn detection configuration. Accepts ServerVad, AzureSemanticVad, AzureSemanticVadEn, or AzureSemanticVadMultilingual (default: ServerVad with threshold=0.5).
tool_choice- Tool selection policy (default: "auto").
temperature- Sampling temperature (default: 0.8).
max_output_tokens- Maximum output tokens (default: 4096).
api_key- Azure API key. If None, reads from AZURE_VOICE_LIVE_API_KEY.
use_default_credential- Use DefaultAzureCredential for auth instead of API key.
conn_options- Connection retry and timeout options.
Example
from livekit.plugins.azure.realtime import RealtimeModel from azure.ai.voicelive.models import AudioInputTranscriptionOptions, ServerVad # English-only session (recommended for reliable language detection) model = RealtimeModel( endpoint=os.getenv("AZURE_VOICE_LIVE_ENDPOINT"), api_key=os.getenv("AZURE_VOICE_LIVE_API_KEY"), model="gpt-realtime", voice="en-US-AvaNeural", input_audio_transcription=AudioInputTranscriptionOptions( model="whisper-1", language="en-US", # Constrains transcription to English ), turn_detection=ServerVad(threshold=0.5, silence_duration_ms=500), ) # Multi-language session, azure-speech takes up to 10 languages, the first is primary model = RealtimeModel( endpoint=os.getenv("AZURE_VOICE_LIVE_ENDPOINT"), api_key=os.getenv("AZURE_VOICE_LIVE_API_KEY"), model="gpt-4.1", voice="en-US-AvaMultilingualNeural", input_audio_transcription=AudioInputTranscriptionOptions( model="azure-speech", language="en-US,zh-CN,ja-JP", # Allow English, Chinese, and Japanese ), )Ancestors
- livekit.agents.llm.realtime.RealtimeModel
Instance variables
prop model : str-
Expand source code
@property def model(self) -> str: return self._opts.model prop provider : str-
Expand source code
@property def provider(self) -> str: return "azure-voicelive"
Methods
async def aclose(self) ‑> None-
Expand source code
async def aclose(self) -> None: pass def session(self, *, turn_detection_disabled: bool = False) ‑> RealtimeSession-
Expand source code
def session(self, *, turn_detection_disabled: bool = False) -> RealtimeSession: # manual turn-taking is unsupported (can_disable_turn_detection=False) sess = RealtimeSession(self) self._sessions.add(sess) return sessCreate a new session, optionally with server-side turn detection disabled.
turn_detection_disabledis honored only by plugins reportingcan_disable_turn_detection; the model itself is left unchanged and reusable.
class RealtimeSession (realtime_model: RealtimeModel)-
Expand source code
class RealtimeSession( llm.RealtimeSession[Literal["azure_server_event_received", "azure_client_event_sent"]] ): """ Azure Voice Live Realtime API session. Manages WebSocket connection to Azure Voice Live and handles: - Audio streaming (input/output) - Text generation - Function calling - Turn detection (VAD) - Session management Emits additional events: - azure_server_event_received: Raw server events from Azure - azure_client_event_sent: Raw client events sent to Azure """ def __init__(self, realtime_model: RealtimeModel) -> None: super().__init__(realtime_model) self._realtime_model: RealtimeModel = realtime_model # per-session copy, so update_options only affects this session self._opts = replace(realtime_model._opts) self._tools = llm.ToolContext.empty() self._instructions: str | None = None # every client event goes through this channel, so they reach Azure in call order self._msg_ch = utils.aio.Chan[ClientEvent]() self._credential: DefaultAzureCredential | None = None # set once Azure confirmed the configuration and the replayed conversation of the # current connection, a connection that fails later was healthy until then self._session_confirmed = False self._connection_established = False # the requests of the current connection, and those a lost one left unconfirmed self._requests: _ConnectionRequests | None = None self._resend: list[ClientEvent] = [] self._current_generation: _ResponseGeneration | _DiscardedGeneration | None = None self._remote_chat_ctx = llm.remote_chat_context.RemoteChatContext() # items of discarded responses, deleted from the Azure conversation self._dropped_item_ids: set[str] = set() # requests whose errors are expected and only logged self._ignored_error_event_ids: set[str] = set() # generate_reply requests by client event id, resolved by the matching response.created self._response_created_futures: dict[str, asyncio.Future[llm.GenerationCreatedEvent]] = {} # sent generate_reply requests that timed out or were cancelled before their response self._discarded_event_ids: set[str] = set() # conversation.item.create requests of update_chat_ctx, by item id and by event id self._item_create_futures: dict[str, asyncio.Future[None]] = {} self._item_create_events: dict[str, str] = {} self._pending_items: dict[str, llm.ChatItem] = {} self._update_chat_ctx_lock = asyncio.Lock() self._update_fnc_ctx_lock = asyncio.Lock() # Audio buffering for input, sent in 100ms chunks self._bstream = utils.audio.AudioByteStream( SAMPLE_RATE, NUM_CHANNELS, samples_per_channel=SAMPLE_RATE // 10 ) self._input_resampler: rtc.AudioResampler | None = None self._video_warned = False self._main_atask = asyncio.create_task( self._main_task(), name="AzureRealtimeSession._main_task" ) @property def chat_ctx(self) -> llm.ChatContext: return self._remote_chat_ctx.to_chat_ctx() @property def tools(self) -> llm.ToolContext: return self._tools @property def tools_ctx(self) -> llm.ToolContext: return self._tools def _send(self, event: ClientEvent) -> None: with contextlib.suppress(utils.aio.channel.ChanClosed): self._msg_ch.send_nowait(event) @utils.log_exceptions(logger=logger) async def _main_task(self) -> None: """Main task that manages the Azure Voice Live WebSocket connection.""" num_retries: int = 0 max_retries = self._opts.conn_options.max_retry reconnecting = False try: while not self._msg_ch.closed: self._session_confirmed = False self._connection_established = False try: await self._run_connection(reconnecting=reconnecting) except APIError as e: if self._connection_established: # the connection was healthy before it dropped (e.g. an idle timeout) num_retries = 0 if max_retries == 0 or not e.retryable: self._emit_error(e, recoverable=False) logger.error("Azure Voice Live connection failed", exc_info=e) return if num_retries >= max_retries: self._emit_error( APIConnectionError( f"Azure Voice Live connection failed after {num_retries} attempts" ), recoverable=False, ) logger.error( f"Azure Voice Live connection failed after {num_retries} attempts", exc_info=e, ) return self._emit_error(e, recoverable=True) retry_interval = self._opts.conn_options._interval_for_retry(num_retries) logger.warning( f"Azure Voice Live connection failed, retrying in {retry_interval}s", exc_info=e, extra={"attempt": num_retries, "max_retries": max_retries}, ) await asyncio.sleep(retry_interval) num_retries += 1 except Exception as e: self._emit_error(e, recoverable=False) logger.exception("Azure Voice Live session failed") return reconnecting = True finally: # nothing can be sent anymore: settle everything still waiting on the connection self._msg_ch.close() self._close_current_generation() self._fail_pending_requests("Azure Voice Live session closed") if self._credential is not None: with contextlib.suppress(Exception): await self._credential.close() self._credential = None def _get_credential(self) -> AzureKeyCredential | DefaultAzureCredential: if self._opts.use_default_credential: # one credential per session, so its token cache survives reconnections if self._credential is None: self._credential = DefaultAzureCredential() return self._credential assert self._opts.api_key is not None, "API key must be set when not using credentials" return AzureKeyCredential(self._opts.api_key) async def _run_connection(self, *, reconnecting: bool) -> None: """Connect, configure the session, and exchange events until the connection ends.""" async with contextlib.AsyncExitStack() as stack: try: conn = await asyncio.wait_for( stack.enter_async_context( connect( endpoint=self._opts.endpoint, credential=self._get_credential(), model=self._opts.model, ) ), self._opts.conn_options.timeout, ) except asyncio.TimeoutError as e: raise APIConnectionError("Azure Voice Live connection timed out") from e except Exception as e: raise APIConnectionError("failed to connect to Azure Voice Live") from e requests = _ConnectionRequests() self._requests = requests # receive from the start, so the replay's confirmations never back up the socket tasks = [ asyncio.create_task(self._recv_task(conn), name="AzureRealtimeSession._recv_task") ] try: try: await self._send_direct( conn, ClientEventSessionUpdate(session=self._create_session_config()) ) if reconnecting: await self._replay_conversation(conn, requests) except Exception as e: raise APIConnectionError( "failed to configure the Azure Voice Live session" ) from e if reconnecting: self.emit("session_reconnected", llm.RealtimeSessionReconnectedEvent()) requests.replayed = True self._check_established() tasks.append( asyncio.create_task( self._send_task(conn, requests), name="AzureRealtimeSession._send_task" ) ) done, _ = await asyncio.wait(tasks, return_when=asyncio.FIRST_COMPLETED) for task in done: if task.cancelled(): raise APIConnectionError("Azure Voice Live connection was cancelled") task.result() finally: await utils.aio.cancel_and_wait(*tasks) self._on_connection_closed(requests) async def _send_task(self, conn: VoiceLiveConnection, requests: _ConnectionRequests) -> None: async for event in self._msg_ch: await self._send_event(conn, requests, event) async def _send_event( self, conn: VoiceLiveConnection, requests: _ConnectionRequests, event: ClientEvent ) -> None: if isinstance(event, ClientEventResponseCreate): event_id = event.event_id or "" reply_fut = self._response_created_futures.get(event_id) if reply_fut is None or reply_fut.done(): # generate_reply timed out or was cancelled before it could be sent self._response_created_futures.pop(event_id, None) self._discarded_event_ids.discard(event_id) return requests.unconfirmed[event_id] = event elif isinstance(event, ClientEventConversationItemCreate) and event.item and event.item.id: item_id = event.item.id if item_id in requests.item_events: # created, or being created, in this conversation (e.g. replayed after a # reconnection): the waiter is settled when that creation is confirmed or rejected item_fut = self._item_create_futures.get(item_id) if self._remote_chat_ctx.get(item_id) is not None and item_fut is not None: self._item_create_futures.pop(item_id, None) if not item_fut.done(): item_fut.set_result(None) return event_id = event.event_id or "" requests.item_events[item_id] = event_id requests.unconfirmed[event_id] = event elif isinstance(event, ClientEventInputAudioBufferAppend): requests.input_audio_samples += _decoded_size(event.audio) // BYTES_PER_SAMPLE requests.input_audio.append(event) elif isinstance(event, ClientEventInputAudioBufferCommit): if not requests.input_audio_samples: # nothing to commit, e.g. the turn was committed already return if requests.input_audio_samples < _MIN_COMMIT_SAMPLES: # Azure rejects committing less than 100ms: the turn is dropped, rather than # left in the buffer to merge with the next one event = ClientEventInputAudioBufferClear() else: event_id = event.event_id or utils.shortuuid("commit_") event.event_id = event_id requests.unconfirmed[event_id] = event requests.committed_audio[event_id] = list(requests.input_audio) requests.input_audio_samples = 0 requests.input_audio.clear() elif isinstance(event, ClientEventInputAudioBufferClear): requests.input_audio_samples = 0 requests.input_audio.clear() try: await self._send_direct(conn, event) except Exception as e: raise APIConnectionError("failed to send an event to Azure Voice Live") from e async def _send_direct(self, conn: VoiceLiveConnection, event: ClientEvent) -> None: await conn.send(event) self.emit("azure_client_event_sent", event) async def _recv_task(self, conn: VoiceLiveConnection) -> None: try: async for event in conn: self._handle_server_event(event) except Exception as e: raise APIConnectionError("failed to receive events from Azure Voice Live") from e if not self._msg_ch.closed: raise APIConnectionError("Azure Voice Live connection closed unexpectedly") def _on_connection_closed(self, requests: _ConnectionRequests) -> None: """Settle what the closed connection can no longer complete.""" self._requests = None self._close_current_generation(None if self._msg_ch.closed else "connection lost") # the deletions of dropped items went down with the conversation self._dropped_item_ids.clear() self._ignored_error_event_ids.clear() # a new connection starts a new, empty conversation: what Azure didn't confirm is sent # again once the conversation is replayed, the waiting callers are none the wiser unconfirmed: list[ClientEvent] = [] for event_id, event in requests.unconfirmed.items(): if isinstance(event, ClientEventResponseCreate): # a response of this conversation can't arrive anymore self._discarded_event_ids.discard(event_id) fut = self._response_created_futures.get(event_id) if fut is not None and not fut.done(): unconfirmed.append(event) elif isinstance(event, ClientEventInputAudioBufferCommit): # the turn went down with the input audio buffer, it's committed again unconfirmed.extend(requests.committed_audio.get(event_id, ())) unconfirmed.append(event) elif event.item and event.item.id and self._remote_chat_ctx.get(event.item.id) is None: unconfirmed.append(event) # so is the audio of the turn in progress, it doesn't belong to a conversation yet unconfirmed.extend(requests.input_audio) self._resend = unconfirmed + requests.unsent + self._resend def _fail_pending_requests(self, reason: str) -> None: self._resend = [] error = llm.RealtimeError(reason) for response_fut in self._response_created_futures.values(): if not response_fut.done(): response_fut.set_exception(error) self._response_created_futures.clear() self._discarded_event_ids.clear() for item_fut in self._item_create_futures.values(): if not item_fut.done(): item_fut.set_exception(error) self._item_create_futures.clear() self._item_create_events.clear() self._pending_items.clear() async def _replay_conversation( self, conn: VoiceLiveConnection, requests: _ConnectionRequests ) -> None: """Re-create the conversation on a new connection, which starts empty. Replayed items are mirrored right away, the conversation.item.created events Azure answers them with are then no-ops. Requests the lost connection left unconfirmed follow them, with the audio its input audio buffer held. """ chat_ctx = self.chat_ctx.copy( exclude_empty_message=True, exclude_handoff=True, exclude_config_update=True, ) replayed = llm.remote_chat_context.RemoteChatContext() azure_items: list[AzureConversationItem] = [] for item in chat_ctx.items: # a function call cut off before its arguments were complete if item.type == "function_call" and not item.arguments: continue try: azure_items.append(livekit_item_to_azure_item(item)) except ValueError: continue replayed.insert(replayed.tail_id, item) # the mirror is the conversation every connection starts from, a failed replay is redone self._remote_chat_ctx = replayed for azure_item in azure_items: event_id = utils.shortuuid("replay_") if azure_item.id: requests.item_events[azure_item.id] = event_id requests.replay_events[event_id] = azure_item.id await self._send_direct( conn, ClientEventConversationItemCreate(event_id=event_id, item=azure_item) ) resend, self._resend = self._resend, [] for i, event in enumerate(resend): try: await self._send_event(conn, requests, event) except BaseException: # a request that fails to send is still unconfirmed, the rest is sent next time requests.unsent = resend[i + 1 :] raise def _voice_config(self) -> str | AzureStandardVoice: voice = self._opts.voice # Azure voice names contain a hyphen, e.g. "en-US-AvaNeural" if isinstance(voice, str) and "-" in voice and voice not in _OPENAI_VOICES: transcription = self._opts.input_audio_transcription language = transcription.language if transcription else None # a voice can only be pinned to a single locale, a language list stays auto-detected locale = language if language and "," not in language else None return AzureStandardVoice(name=voice, locale=locale) return voice def _create_session_config(self) -> RequestSession: """Configure the Azure Voice Live session with the current settings.""" session = RequestSession( modalities=list(self._opts.modalities), instructions=( self._instructions if self._instructions is not None else "You are a helpful assistant." ), voice=self._voice_config(), input_audio_format=self._opts.input_audio_format, output_audio_format=self._opts.output_audio_format, turn_detection=self._opts.turn_detection, input_audio_transcription=self._opts.input_audio_transcription, tool_choice=to_azure_tool_choice(self._opts.tool_choice), temperature=self._opts.temperature, max_response_output_tokens=self._opts.max_output_tokens, ) if tools := livekit_tools_to_azure_tools(self._tools.flatten()): session.tools = tools return session def _output_modalities(self) -> list[Literal["text", "audio"]]: return ["audio", "text"] if self._realtime_model.capabilities.audio_output else ["text"] def _handle_server_event(self, event: ServerEvent) -> None: """Handle events from Azure Voice Live.""" self.emit("azure_server_event_received", event) try: if isinstance(event, ServerEventSessionUpdated): self._handle_session_updated(event) elif isinstance(event, ServerEventInputAudioBufferSpeechStarted): self._handle_input_speech_started(event) elif isinstance(event, ServerEventInputAudioBufferSpeechStopped): self._handle_input_speech_stopped(event) elif isinstance(event, ServerEventInputAudioBufferCommitted): self._handle_input_audio_buffer_committed(event) elif isinstance(event, ServerEventConversationItemCreated): self._handle_conversation_item_created(event) elif isinstance(event, ServerEventConversationItemDeleted): self._handle_conversation_item_deleted(event) elif isinstance(event, ServerEventConversationItemInputAudioTranscriptionCompleted): self._handle_input_audio_transcription_completed(event) elif isinstance(event, ServerEventResponseCreated): self._handle_response_created(event) elif isinstance(event, ServerEventResponseOutputItemAdded): self._handle_output_item_added(event) elif isinstance(event, ServerEventResponseContentPartAdded): self._handle_content_part_added(event) elif isinstance(event, ServerEventResponseAudioDelta): self._handle_audio_delta(event) elif isinstance(event, ServerEventResponseAudioTranscriptDelta): self._handle_text_delta(event, is_transcript=True) elif isinstance(event, ServerEventResponseTextDelta): # text-only responses (modalities=["text"]) self._handle_text_delta(event, is_transcript=False) elif isinstance(event, ServerEventResponseFunctionCallArgumentsDelta): self._handle_function_call_arguments_delta(event) elif isinstance(event, ServerEventResponseFunctionCallArgumentsDone): self._handle_function_call_arguments_done(event) elif isinstance(event, ServerEventResponseDone): self._handle_response_done(event) elif isinstance(event, ServerEventError): self._handle_error(event) except Exception: logger.exception( "failed to handle Azure Voice Live event", extra={"event_type": event.type} ) def _check_established(self) -> None: requests = self._requests if ( requests is not None and requests.replayed and not requests.replay_events and self._session_confirmed ): self._connection_established = True def _handle_session_updated(self, event: ServerEventSessionUpdated) -> None: # a session.updated only confirms a configuration change, it never means a reconnection self._session_confirmed = True self._check_established() logger.debug( "Azure Voice Live session updated", extra={"session_id": getattr(event.session, "id", None)}, ) def _handle_input_speech_started(self, _: ServerEventInputAudioBufferSpeechStarted) -> None: # interrupting the reply is left to the agent, which knows if it can be interrupted self.emit("input_speech_started", llm.InputSpeechStartedEvent()) def _handle_input_speech_stopped(self, _: ServerEventInputAudioBufferSpeechStopped) -> None: self.emit( "input_speech_stopped", llm.InputSpeechStoppedEvent( user_transcription_enabled=self._opts.input_audio_transcription is not None ), ) def _handle_input_audio_buffer_committed(self, _: ServerEventInputAudioBufferCommitted) -> None: if (requests := self._requests) is None: return # confirmed: the turn doesn't need to be committed again after a reconnection. Commits # are confirmed in order, without their event id commit_id = next(iter(requests.committed_audio), None) if commit_id is not None: requests.committed_audio.pop(commit_id) requests.unconfirmed.pop(commit_id, None) else: # committed by the turn detection of Azure requests.input_audio.clear() def _handle_conversation_item_created(self, event: ServerEventConversationItemCreated) -> None: item = event.item if item is None or not item.id: return item_id = item.id created_by_client = False if (requests := self._requests) is not None: created_by_client = item_id in requests.item_events if requests.replay_events.pop(requests.item_events.get(item_id, ""), None): self._check_established() # confirmed: the item doesn't need to be sent again after a reconnection requests.unconfirmed.pop(requests.item_events.get(item_id, ""), None) # items of update_chat_ctx are mirrored as the caller built them lk_item: llm.ChatItem | None = self._pending_items.pop(item_id, None) if lk_item is None: try: lk_item = azure_item_to_livekit_item(item) except ValueError: logger.debug( "ignoring an Azure Voice Live conversation item", extra={"item_type": item.type}, ) if ( not created_by_client and isinstance(self._current_generation, _DiscardedGeneration) and lk_item is not None and ( lk_item.type == "function_call" or (lk_item.type == "message" and lk_item.role == "assistant") ) ): # an output of the discarded response, created ahead of its output_item.added self._drop_item(item_id) return if ( lk_item is not None and item_id not in self._dropped_item_ids and self._remote_chat_ctx.get(item_id) is None ): previous_item_id = event.previous_item_id if previous_item_id is None or self._remote_chat_ctx.get(previous_item_id) is None: # Azure appends an item it isn't told where to insert previous_item_id = self._remote_chat_ctx.tail_id self._remote_chat_ctx.insert(previous_item_id, lk_item) self.emit( "remote_item_added", llm.RemoteItemAddedEvent(previous_item_id=previous_item_id, item=lk_item), ) if (fut := self._item_create_futures.pop(item_id, None)) and not fut.done(): fut.set_result(None) def _handle_conversation_item_deleted(self, event: ServerEventConversationItemDeleted) -> None: self._dropped_item_ids.discard(event.item_id) if event.item_id and self._remote_chat_ctx.get(event.item_id) is not None: self._remote_chat_ctx.delete(event.item_id) def _handle_input_audio_transcription_completed( self, event: ServerEventConversationItemInputAudioTranscriptionCompleted ) -> None: item_id = event.item_id or "" transcript = event.transcript or "" remote_item = self._remote_chat_ctx.get(item_id) if ( transcript and remote_item is not None and isinstance(remote_item.item, llm.ChatMessage) and transcript not in remote_item.item.content ): remote_item.item.content.append(transcript) self.emit( "input_audio_transcription_completed", llm.InputTranscriptionCompleted(item_id=item_id, transcript=transcript, is_final=True), ) def _close_generation(self, generation: _ResponseGeneration) -> None: """Close all channels and futures of a generation, so its consumers never hang.""" for message in generation.messages.values(): if not message.modalities.done(): message.modalities.set_result(self._output_modalities()) message.text_ch.close() message.audio_ch.close() generation.message_ch.close() generation.function_ch.close() def _close_current_generation(self, reason: str | None = None) -> None: generation = self._current_generation self._current_generation = None if isinstance(generation, _ResponseGeneration): self._close_generation(generation) if reason: logger.warning(f"in-progress Azure Voice Live generation closed due to {reason}") def _generation_for(self, response_id: str | None) -> _ResponseGeneration | None: """The current generation, if the event of `response_id` belongs to it.""" generation = self._current_generation if not isinstance(generation, _ResponseGeneration): return None if response_id and generation.response_id and response_id != generation.response_id: return None return generation def _drop_item(self, item_id: str) -> None: """Remove an item of a discarded response, nobody heard it, from the conversation.""" if item_id in self._dropped_item_ids: # being deleted already, e.g. its conversation.item.created came first return self._dropped_item_ids.add(item_id) if self._remote_chat_ctx.get(item_id) is not None: self._remote_chat_ctx.delete(item_id) event_id = utils.shortuuid("drop_item_") # e.g. the cancelled response never kept it self._ignored_error_event_ids.add(event_id) self._send(ClientEventConversationItemDelete(item_id=item_id, event_id=event_id)) def _handle_response_created(self, event: ServerEventResponseCreated) -> None: response = event.response response_id = response.id if response else None metadata = response.metadata if response else None client_event_id = metadata.get("client_event_id") if isinstance(metadata, dict) else None # never leave a previous generation open, e.g. when its response.done is still in flight self._close_current_generation() if client_event_id and self._requests is not None: # confirmed: the request doesn't need to be sent again after a reconnection self._requests.unconfirmed.pop(client_event_id, None) # generate_reply tags its request with metadata that Azure echoes back, so a response # created by server-side turn detection can't be mistaken for the requested one fut = self._response_created_futures.pop(client_event_id, None) if client_event_id else None if client_event_id and ( client_event_id in self._discarded_event_ids or (fut is not None and fut.done()) ): # its generate_reply timed out or was cancelled, nobody is waiting for it anymore self._discarded_event_ids.discard(client_event_id) self._send(ClientEventResponseCancel(response_id=response_id)) self._current_generation = _DiscardedGeneration(response_id=response_id) logger.warning( "discarding an Azure Voice Live response created after its generate_reply " "timed out or was cancelled" ) return generation = _ResponseGeneration( response_id=response_id, message_ch=utils.aio.Chan[llm.MessageGeneration](), function_ch=utils.aio.Chan[llm.FunctionCall](), created_timestamp=time.time(), ) self._current_generation = generation generation_ev = llm.GenerationCreatedEvent( message_stream=generation.message_ch, function_stream=generation.function_ch, user_initiated=fut is not None, response_id=response_id, ) if fut is not None: fut.set_result(generation_ev) self.emit("generation_created", generation_ev) def _handle_output_item_added(self, event: ServerEventResponseOutputItemAdded) -> None: item = event.item if item is None or not item.id: return discarded = self._current_generation if isinstance(discarded, _DiscardedGeneration) and ( not event.response_id or not discarded.response_id or event.response_id == discarded.response_id ): self._drop_item(item.id) return generation = self._generation_for(event.response_id) if generation is None: return if item.type == "message": message = _MessageGeneration( message_id=item.id, text_ch=utils.aio.Chan[str](), # unbounded: the agent can start reading a reply only once it may be played audio_ch=utils.aio.Chan[rtc.AudioFrame](), modalities=asyncio.Future[list[Literal["text", "audio"]]](), ) if not self._realtime_model.capabilities.audio_output: message.audio_ch.close() message.modalities.set_result(["text"]) generation.messages[item.id] = message generation.message_ch.send_nowait( llm.MessageGeneration( message_id=item.id, text_stream=message.text_ch, audio_stream=message.audio_ch, modalities=message.modalities, ) ) elif item.type == "function_call": # emitted once its arguments are complete, see _handle_function_call_arguments_done generation.function_calls[item.id] = _FunctionCallGeneration( item_id=item.id, call_id=getattr(item, "call_id", None) or "", name=getattr(item, "name", None) or "", ) def _handle_content_part_added(self, event: ServerEventResponseContentPartAdded) -> None: generation = self._generation_for(event.response_id) if generation is None or not (message := generation.messages.get(event.item_id)): return part_type = getattr(event.part, "type", None) modalities: list[Literal["text", "audio"]] if part_type == "audio": modalities = ["audio", "text"] elif part_type == "text": modalities = ["text"] else: return with contextlib.suppress(asyncio.InvalidStateError): message.modalities.set_result(modalities) def _handle_audio_delta(self, event: ServerEventResponseAudioDelta) -> None: generation = self._generation_for(event.response_id) if generation is None or not (message := generation.messages.get(event.item_id)): return # the SDK decodes the base64 payload, accept a raw string in case it doesn't data = event.delta if isinstance(data, str): data = base64.b64decode(data) if not data or message.audio_ch.closed: return if generation.first_token_timestamp is None: generation.first_token_timestamp = time.time() with contextlib.suppress(asyncio.InvalidStateError): message.modalities.set_result(["audio", "text"]) message.audio_ch.send_nowait( rtc.AudioFrame( data=data, sample_rate=SAMPLE_RATE, num_channels=NUM_CHANNELS, samples_per_channel=len(data) // BYTES_PER_SAMPLE, ) ) def _handle_text_delta( self, event: ServerEventResponseAudioTranscriptDelta | ServerEventResponseTextDelta, *, is_transcript: bool, ) -> None: generation = self._generation_for(event.response_id) if generation is None or not (message := generation.messages.get(event.item_id)): return delta = event.delta if not delta or message.text_ch.closed: return # transcripts trail their audio, while text deltas are model output on their own, # including the text fallback of an audio session if not is_transcript and generation.first_token_timestamp is None: generation.first_token_timestamp = time.time() message.text_ch.send_nowait(delta) message.audio_transcript += delta def _handle_function_call_arguments_delta( self, event: ServerEventResponseFunctionCallArgumentsDelta ) -> None: generation = self._generation_for(event.response_id) if generation is None or not ( function_call := generation.function_calls.get(event.item_id) ): return if generation.first_token_timestamp is None: generation.first_token_timestamp = time.time() function_call.arguments += event.delta or "" def _handle_function_call_arguments_done( self, event: ServerEventResponseFunctionCallArgumentsDone ) -> None: generation = self._generation_for(event.response_id) if generation is None or not event.item_id: return pending = generation.function_calls.pop(event.item_id, None) call_id = event.call_id or (pending.call_id if pending else "") name = event.name or (pending.name if pending else "") arguments = ( event.arguments if event.arguments is not None else (pending.arguments if pending else "") ) if not call_id or not name: logger.warning( "ignoring an Azure Voice Live function call without a name or call id", extra={"item_id": event.item_id}, ) return remote_item = self._remote_chat_ctx.get(event.item_id) if remote_item is not None and isinstance(remote_item.item, llm.FunctionCall): remote_item.item.arguments = arguments generation.function_ch.send_nowait( llm.FunctionCall(id=event.item_id, call_id=call_id, name=name, arguments=arguments) ) logger.debug( "Azure Voice Live function call completed", extra={"function": name, "call_id": call_id}, ) def _handle_response_done(self, event: ServerEventResponseDone) -> None: response = event.response response_id = response.id if response else None generation = self._current_generation if isinstance(generation, _DiscardedGeneration): if ( not response_id or not generation.response_id or response_id == generation.response_id ): self._current_generation = None return if generation is None or ( response_id and generation.response_id and response_id != generation.response_id ): return # mirror what the assistant said, so reconnections replay it for item_id, message in generation.messages.items(): transcript = message.audio_transcript remote_item = self._remote_chat_ctx.get(item_id) if ( transcript and remote_item is not None and isinstance(remote_item.item, llm.ChatMessage) and transcript not in remote_item.item.content ): remote_item.item.content.append(transcript) self._close_generation(generation) self._current_generation = None status = getattr(response, "status", None) self._emit_response_metrics(generation, event) if status == "failed": error = getattr(getattr(response, "status_details", None), "error", None) self._emit_error( APIError("Azure Voice Live response failed", body=error, retryable=True), recoverable=True, ) elif status in ("cancelled", "incomplete"): logger.debug( f"Azure Voice Live response {status}", extra={ "response_id": response_id, "reason": getattr(getattr(response, "status_details", None), "reason", None), }, ) def _emit_response_metrics( self, generation: _ResponseGeneration, event: ServerEventResponseDone ) -> None: response = event.response usage = getattr(response, "usage", None) input_details = getattr(usage, "input_token_details", None) output_details = getattr(usage, "output_token_details", None) created_timestamp = generation.created_timestamp ttft = ( generation.first_token_timestamp - created_timestamp if generation.first_token_timestamp else -1 ) duration = time.time() - created_timestamp output_tokens = _token_count(usage, "output_tokens") self.emit( "metrics_collected", RealtimeModelMetrics( timestamp=created_timestamp, request_id=getattr(response, "id", None) or "", ttft=ttft, duration=duration, cancelled=getattr(response, "status", None) == "cancelled", label=self._realtime_model.label, input_tokens=_token_count(usage, "input_tokens"), output_tokens=output_tokens, total_tokens=_token_count(usage, "total_tokens"), tokens_per_second=output_tokens / duration if duration > 0 else 0, input_token_details=RealtimeModelMetrics.InputTokenDetails( audio_tokens=_token_count(input_details, "audio_tokens"), text_tokens=_token_count(input_details, "text_tokens"), image_tokens=_token_count(input_details, "image_tokens"), cached_tokens=_token_count(input_details, "cached_tokens"), cached_tokens_details=None, ), output_token_details=RealtimeModelMetrics.OutputTokenDetails( text_tokens=_token_count(output_details, "text_tokens"), audio_tokens=_token_count(output_details, "audio_tokens"), image_tokens=0, ), metadata=Metadata( model_name=self._realtime_model.model, model_provider=self._realtime_model.provider, ), ), ) def _handle_error(self, event: ServerEventError) -> None: error = event.error message = getattr(error, "message", None) or "unknown error" code = getattr(error, "code", None) event_id = getattr(error, "event_id", None) if event_id: if (requests := self._requests) is not None: if (replayed_id := requests.replay_events.pop(event_id, None)) is not None: requests.item_events.pop(replayed_id, None) if self._remote_chat_ctx.get(replayed_id) is not None: self._remote_chat_ctx.delete(replayed_id) logger.warning( f"Azure Voice Live rejected a replayed conversation item: {message}", extra={"item_id": replayed_id, "code": code}, ) self._check_established() return # a rejected request is settled, it's never sent again after a reconnection rejected = requests.unconfirmed.pop(event_id, None) requests.committed_audio.pop(event_id, None) if ( isinstance(rejected, ClientEventConversationItemCreate) and rejected.item and rejected.item.id ): rejected_id = rejected.item.id # the item doesn't exist, so it can be created again requests.item_events.pop(rejected_id, None) self._item_create_events.pop(event_id, None) self._pending_items.pop(rejected_id, None) # fails whoever waits for the item now, e.g. a retry joining this creation item_fut = self._item_create_futures.pop(rejected_id, None) if item_fut is not None and not item_fut.done(): item_fut.set_exception(llm.RealtimeError(message, code=code)) return if event_id in self._ignored_error_event_ids: self._ignored_error_event_ids.discard(event_id) logger.debug(f"Azure Voice Live (ignored): {message}") return # a rejected conversation.item.create fails its update_chat_ctx if (item_id := self._item_create_events.pop(event_id, None)) is not None: self._pending_items.pop(item_id, None) if (fut := self._item_create_futures.pop(item_id, None)) and not fut.done(): fut.set_exception(llm.RealtimeError(message, code=code)) return # a rejected response.create never gets a response.created if (reply_fut := self._response_created_futures.pop(event_id, None)) and not ( reply_fut.done() ): reply_fut.set_exception(llm.RealtimeError(message, code=code)) # cancelling a response that already ended is a harmless race lowered = message.lower() if code == "response_cancel_not_active" or "no active response" in lowered: logger.debug(f"Azure Voice Live (suppressed): {message}") return logger.error(f"Azure Voice Live error: {message}", extra={"code": code}) self._emit_error( APIError(f"Azure Voice Live error: {message}", body=error, retryable=True), recoverable=True, ) def push_audio(self, frame: rtc.AudioFrame) -> None: """Push audio frame to Azure Voice Live.""" for resampled_frame in self._resample_audio(frame): for audio_frame in self._bstream.push(resampled_frame.data.tobytes()): self._send_audio(audio_frame) def _send_audio(self, frame: rtc.AudioFrame) -> None: self._send( ClientEventInputAudioBufferAppend( audio=base64.b64encode(frame.data).decode("utf-8"), ) ) def _resample_audio(self, frame: rtc.AudioFrame) -> Iterator[rtc.AudioFrame]: """Resample audio to target sample rate and channel count if needed.""" if self._input_resampler: if frame.sample_rate != self._input_resampler._input_rate: # input audio changed to a different sample rate self._input_resampler = None if self._input_resampler is None and ( frame.sample_rate != SAMPLE_RATE or frame.num_channels != NUM_CHANNELS ): self._input_resampler = rtc.AudioResampler( input_rate=frame.sample_rate, output_rate=SAMPLE_RATE, num_channels=NUM_CHANNELS, ) if self._input_resampler: # TODO(long): flush the resampler when the input source is changed yield from self._input_resampler.push(frame) else: yield frame def push_video(self, frame: rtc.VideoFrame) -> None: """Push video frame (not supported by Azure Voice Live).""" if not self._video_warned: self._video_warned = True logger.warning("push_video() is not supported by Azure Voice Live") def update_options(self, *, tool_choice: NotGivenOr[llm.ToolChoice | None] = NOT_GIVEN) -> None: """Update session options and send session.update to Azure server.""" if not is_given(tool_choice) or tool_choice == self._opts.tool_choice: return self._opts.tool_choice = tool_choice self._send( ClientEventSessionUpdate( session=RequestSession(tool_choice=to_azure_tool_choice(tool_choice)) ) ) def generate_reply( self, *, instructions: NotGivenOr[str] = NOT_GIVEN, tool_choice: NotGivenOr[llm.ToolChoice] = NOT_GIVEN, tools: NotGivenOr[list[llm.Tool]] = NOT_GIVEN, ) -> asyncio.Future[llm.GenerationCreatedEvent]: """Generate a reply from the model. Returns a Future that resolves to GenerationCreatedEvent when the response.created event carrying this request's client_event_id is received from Azure. """ fut = asyncio.Future[llm.GenerationCreatedEvent]() if self._msg_ch.closed: fut.set_exception(llm.RealtimeError("Azure Voice Live session is closed")) return fut event_id = utils.shortuuid("response_create_") params = ResponseCreateParams(metadata={"client_event_id": event_id}) if is_given(tool_choice): params.tool_choice = to_azure_response_tool_choice(tool_choice) if is_given(tools): params.tools = livekit_tools_to_azure_tools(tools) self._response_created_futures[event_id] = fut self._send( ClientEventResponseCreate( event_id=event_id, response=params, additional_instructions=instructions if is_given(instructions) else None, ) ) def _on_timeout() -> None: if not fut.done(): # a sent request is cancelled once Azure creates it, see _handle_response_created if self._response_created_futures.pop(event_id, None) is not None: self._discarded_event_ids.add(event_id) fut.set_exception(llm.RealtimeError("generate_reply timed out.")) # one deadline covers waiting for the connection, sending, and response.created handle = asyncio.get_running_loop().call_later(_GENERATE_REPLY_TIMEOUT, _on_timeout) def _on_done(_: asyncio.Future[llm.GenerationCreatedEvent]) -> None: handle.cancel() # still registered: the caller cancelled it before its response.created arrived if self._response_created_futures.pop(event_id, None) is not None: self._discarded_event_ids.add(event_id) fut.add_done_callback(_on_done) return fut def interrupt(self) -> None: """Interrupt the current response.""" if isinstance(self._current_generation, _ResponseGeneration) or ( self._response_created_futures ): self._send(ClientEventResponseCancel()) def truncate( self, *, message_id: str, modalities: list[Literal["text", "audio"]], audio_end_ms: int, audio_transcript: NotGivenOr[str] = NOT_GIVEN, ) -> None: """Truncate conversation history (not supported by Azure Voice Live).""" logger.warning("truncate() is not supported by Azure Voice Live") async def update_instructions(self, instructions: str) -> None: """Update system instructions.""" self._instructions = instructions self._send(ClientEventSessionUpdate(session=RequestSession(instructions=instructions))) async def update_chat_ctx(self, chat_ctx: llm.ChatContext) -> None: """Create the new items of `chat_ctx` in the Azure conversation. Raises: llm.RealtimeError: if Azure rejects an item or doesn't confirm it in time. """ async with self._update_chat_ctx_lock: # Filter out internal framework items that Azure doesn't understand chat_ctx = chat_ctx.copy( exclude_handoff=True, exclude_config_update=True, ) # Remove instruction messages (already sent via the session configuration) remove_instructions(chat_ctx) remote_ctx = self._remote_chat_ctx.to_chat_ctx() known_ids = {item.id for item in remote_ctx.items} # empty messages are only placeholders, unless they already exist remotely chat_ctx = llm.ChatContext( [ item for item in chat_ctx.items if item.type != "message" or item.content or item.id in known_ids ] ) diff_ops = llm.utils.compute_chat_ctx_diff(remote_ctx, chat_ctx) # Azure manages the conversation history internally, only additions are synced if diff_ops.to_remove or diff_ops.to_update: logger.debug( "Azure Voice Live ignores removed and updated chat items", extra={ "to_remove": len(diff_ops.to_remove), "to_update": len(diff_ops.to_update), }, ) if not diff_ops.to_create: return if self._msg_ch.closed: raise llm.RealtimeError("Azure Voice Live session is closed") futs: list[asyncio.Future[None]] = [] event_ids: list[str] = [] for previous_item_id, item_id in diff_ops.to_create: chat_item = chat_ctx.get_by_id(item_id) assert chat_item is not None try: azure_item = livekit_item_to_azure_item(chat_item) except ValueError: logger.warning( "skipping a chat item Azure Voice Live doesn't support", extra={"item_type": chat_item.type}, ) continue event_id = utils.shortuuid("chat_ctx_create_") fut = asyncio.Future[None]() self._item_create_futures[item_id] = fut self._item_create_events[event_id] = item_id self._pending_items[item_id] = chat_item futs.append(fut) event_ids.append(event_id) self._send( ClientEventConversationItemCreate( event_id=event_id, # without an anchor Azure appends the item previous_item_id=previous_item_id if previous_item_id in known_ids else None, item=azure_item, ) ) known_ids.add(item_id) if not futs: return try: results = await asyncio.wait_for( asyncio.gather(*futs, return_exceptions=True), timeout=_UPDATE_CHAT_CTX_TIMEOUT, ) except asyncio.TimeoutError: raise llm.RealtimeError("update_chat_ctx timed out.") from None finally: for event_id in event_ids: if (pending_id := self._item_create_events.pop(event_id, None)) is not None: self._item_create_futures.pop(pending_id, None) self._pending_items.pop(pending_id, None) if errors := [r for r in results if isinstance(r, BaseException)]: raise llm.RealtimeError( f"Azure Voice Live rejected {len(errors)} of {len(results)} chat items: " f"{errors[0]}" ) async def update_tools(self, tools: list[llm.Tool]) -> None: """Update available tools.""" async with self._update_fnc_ctx_lock: self._tools = llm.ToolContext(tools) # an empty list clears the tools of the session self._send( ClientEventSessionUpdate( session=RequestSession( tools=livekit_tools_to_azure_tools(self._tools.flatten()) ) ) ) def commit_audio(self) -> None: """Commit the audio buffer. Azure can't commit less than 100ms of audio, such a short turn is cleared instead. """ # the buffered tail belongs to this turn, send it before the commit for audio_frame in self._bstream.flush(): self._send_audio(audio_frame) # checked against the buffer of the connection when it's sent self._send(ClientEventInputAudioBufferCommit(event_id=utils.shortuuid("commit_"))) def clear_audio(self) -> None: """Clear the audio buffer.""" self._bstream.clear() self._send(ClientEventInputAudioBufferClear()) def commit_user_turn(self) -> None: logger.warning("commit_user_turn is not supported by Azure Realtime API.") def _emit_error(self, error: Exception, recoverable: bool) -> None: """Emit an error event.""" self.emit( "error", llm.RealtimeModelError( timestamp=time.time(), label=self.realtime_model.label, error=error, recoverable=recoverable, ), ) async def aclose(self) -> None: """Close the session.""" self._msg_ch.close() await utils.aio.cancel_and_wait(self._main_atask) # the main task may have been cancelled before it started, and couldn't clean up self._close_current_generation() self._fail_pending_requests("Azure Voice Live session closed") if self._credential is not None: with contextlib.suppress(Exception): await self._credential.close() self._credential = NoneAzure Voice Live Realtime API session.
Manages WebSocket connection to Azure Voice Live and handles: - Audio streaming (input/output) - Text generation - Function calling - Turn detection (VAD) - Session management
Emits additional events: - azure_server_event_received: Raw server events from Azure - azure_client_event_sent: Raw client events sent to Azure
Ancestors
- livekit.agents.llm.realtime.RealtimeSession
- abc.ABC
- EventEmitter
- typing.Generic
Instance variables
prop chat_ctx : llm.ChatContext-
Expand source code
@property def chat_ctx(self) -> llm.ChatContext: return self._remote_chat_ctx.to_chat_ctx() prop tools : llm.ToolContext-
Expand source code
@property def tools(self) -> llm.ToolContext: return self._tools prop tools_ctx : llm.ToolContext-
Expand source code
@property def tools_ctx(self) -> llm.ToolContext: return self._tools
Methods
async def aclose(self) ‑> None-
Expand source code
async def aclose(self) -> None: """Close the session.""" self._msg_ch.close() await utils.aio.cancel_and_wait(self._main_atask) # the main task may have been cancelled before it started, and couldn't clean up self._close_current_generation() self._fail_pending_requests("Azure Voice Live session closed") if self._credential is not None: with contextlib.suppress(Exception): await self._credential.close() self._credential = NoneClose the session.
def clear_audio(self) ‑> None-
Expand source code
def clear_audio(self) -> None: """Clear the audio buffer.""" self._bstream.clear() self._send(ClientEventInputAudioBufferClear())Clear the audio buffer.
def commit_audio(self) ‑> None-
Expand source code
def commit_audio(self) -> None: """Commit the audio buffer. Azure can't commit less than 100ms of audio, such a short turn is cleared instead. """ # the buffered tail belongs to this turn, send it before the commit for audio_frame in self._bstream.flush(): self._send_audio(audio_frame) # checked against the buffer of the connection when it's sent self._send(ClientEventInputAudioBufferCommit(event_id=utils.shortuuid("commit_")))Commit the audio buffer.
Azure can't commit less than 100ms of audio, such a short turn is cleared instead.
def commit_user_turn(self) ‑> None-
Expand source code
def commit_user_turn(self) -> None: logger.warning("commit_user_turn is not supported by Azure Realtime API.") def generate_reply(self,
*,
instructions: NotGivenOr[str] = NOT_GIVEN,
tool_choice: NotGivenOr[llm.ToolChoice] = NOT_GIVEN,
tools: NotGivenOr[list[llm.Tool]] = NOT_GIVEN) ‑> _asyncio.Future[livekit.agents.llm.realtime.GenerationCreatedEvent]-
Expand source code
def generate_reply( self, *, instructions: NotGivenOr[str] = NOT_GIVEN, tool_choice: NotGivenOr[llm.ToolChoice] = NOT_GIVEN, tools: NotGivenOr[list[llm.Tool]] = NOT_GIVEN, ) -> asyncio.Future[llm.GenerationCreatedEvent]: """Generate a reply from the model. Returns a Future that resolves to GenerationCreatedEvent when the response.created event carrying this request's client_event_id is received from Azure. """ fut = asyncio.Future[llm.GenerationCreatedEvent]() if self._msg_ch.closed: fut.set_exception(llm.RealtimeError("Azure Voice Live session is closed")) return fut event_id = utils.shortuuid("response_create_") params = ResponseCreateParams(metadata={"client_event_id": event_id}) if is_given(tool_choice): params.tool_choice = to_azure_response_tool_choice(tool_choice) if is_given(tools): params.tools = livekit_tools_to_azure_tools(tools) self._response_created_futures[event_id] = fut self._send( ClientEventResponseCreate( event_id=event_id, response=params, additional_instructions=instructions if is_given(instructions) else None, ) ) def _on_timeout() -> None: if not fut.done(): # a sent request is cancelled once Azure creates it, see _handle_response_created if self._response_created_futures.pop(event_id, None) is not None: self._discarded_event_ids.add(event_id) fut.set_exception(llm.RealtimeError("generate_reply timed out.")) # one deadline covers waiting for the connection, sending, and response.created handle = asyncio.get_running_loop().call_later(_GENERATE_REPLY_TIMEOUT, _on_timeout) def _on_done(_: asyncio.Future[llm.GenerationCreatedEvent]) -> None: handle.cancel() # still registered: the caller cancelled it before its response.created arrived if self._response_created_futures.pop(event_id, None) is not None: self._discarded_event_ids.add(event_id) fut.add_done_callback(_on_done) return futGenerate a reply from the model.
Returns a Future that resolves to GenerationCreatedEvent when the response.created event carrying this request's client_event_id is received from Azure.
def interrupt(self) ‑> None-
Expand source code
def interrupt(self) -> None: """Interrupt the current response.""" if isinstance(self._current_generation, _ResponseGeneration) or ( self._response_created_futures ): self._send(ClientEventResponseCancel())Interrupt the current response.
def push_audio(self, frame: rtc.AudioFrame) ‑> None-
Expand source code
def push_audio(self, frame: rtc.AudioFrame) -> None: """Push audio frame to Azure Voice Live.""" for resampled_frame in self._resample_audio(frame): for audio_frame in self._bstream.push(resampled_frame.data.tobytes()): self._send_audio(audio_frame)Push audio frame to Azure Voice Live.
def push_video(self, frame: rtc.VideoFrame) ‑> None-
Expand source code
def push_video(self, frame: rtc.VideoFrame) -> None: """Push video frame (not supported by Azure Voice Live).""" if not self._video_warned: self._video_warned = True logger.warning("push_video() is not supported by Azure Voice Live")Push video frame (not supported by Azure Voice Live).
def truncate(self,
*,
message_id: str,
modalities: "list[Literal['text', 'audio']]",
audio_end_ms: int,
audio_transcript: NotGivenOr[str] = NOT_GIVEN) ‑> None-
Expand source code
def truncate( self, *, message_id: str, modalities: list[Literal["text", "audio"]], audio_end_ms: int, audio_transcript: NotGivenOr[str] = NOT_GIVEN, ) -> None: """Truncate conversation history (not supported by Azure Voice Live).""" logger.warning("truncate() is not supported by Azure Voice Live")Truncate conversation history (not supported by Azure Voice Live).
async def update_chat_ctx(self, chat_ctx: llm.ChatContext) ‑> None-
Expand source code
async def update_chat_ctx(self, chat_ctx: llm.ChatContext) -> None: """Create the new items of `chat_ctx` in the Azure conversation. Raises: llm.RealtimeError: if Azure rejects an item or doesn't confirm it in time. """ async with self._update_chat_ctx_lock: # Filter out internal framework items that Azure doesn't understand chat_ctx = chat_ctx.copy( exclude_handoff=True, exclude_config_update=True, ) # Remove instruction messages (already sent via the session configuration) remove_instructions(chat_ctx) remote_ctx = self._remote_chat_ctx.to_chat_ctx() known_ids = {item.id for item in remote_ctx.items} # empty messages are only placeholders, unless they already exist remotely chat_ctx = llm.ChatContext( [ item for item in chat_ctx.items if item.type != "message" or item.content or item.id in known_ids ] ) diff_ops = llm.utils.compute_chat_ctx_diff(remote_ctx, chat_ctx) # Azure manages the conversation history internally, only additions are synced if diff_ops.to_remove or diff_ops.to_update: logger.debug( "Azure Voice Live ignores removed and updated chat items", extra={ "to_remove": len(diff_ops.to_remove), "to_update": len(diff_ops.to_update), }, ) if not diff_ops.to_create: return if self._msg_ch.closed: raise llm.RealtimeError("Azure Voice Live session is closed") futs: list[asyncio.Future[None]] = [] event_ids: list[str] = [] for previous_item_id, item_id in diff_ops.to_create: chat_item = chat_ctx.get_by_id(item_id) assert chat_item is not None try: azure_item = livekit_item_to_azure_item(chat_item) except ValueError: logger.warning( "skipping a chat item Azure Voice Live doesn't support", extra={"item_type": chat_item.type}, ) continue event_id = utils.shortuuid("chat_ctx_create_") fut = asyncio.Future[None]() self._item_create_futures[item_id] = fut self._item_create_events[event_id] = item_id self._pending_items[item_id] = chat_item futs.append(fut) event_ids.append(event_id) self._send( ClientEventConversationItemCreate( event_id=event_id, # without an anchor Azure appends the item previous_item_id=previous_item_id if previous_item_id in known_ids else None, item=azure_item, ) ) known_ids.add(item_id) if not futs: return try: results = await asyncio.wait_for( asyncio.gather(*futs, return_exceptions=True), timeout=_UPDATE_CHAT_CTX_TIMEOUT, ) except asyncio.TimeoutError: raise llm.RealtimeError("update_chat_ctx timed out.") from None finally: for event_id in event_ids: if (pending_id := self._item_create_events.pop(event_id, None)) is not None: self._item_create_futures.pop(pending_id, None) self._pending_items.pop(pending_id, None) if errors := [r for r in results if isinstance(r, BaseException)]: raise llm.RealtimeError( f"Azure Voice Live rejected {len(errors)} of {len(results)} chat items: " f"{errors[0]}" )Create the new items of
chat_ctxin the Azure conversation.Raises
llm.RealtimeError- if Azure rejects an item or doesn't confirm it in time.
async def update_instructions(self, instructions: str) ‑> None-
Expand source code
async def update_instructions(self, instructions: str) -> None: """Update system instructions.""" self._instructions = instructions self._send(ClientEventSessionUpdate(session=RequestSession(instructions=instructions)))Update system instructions.
def update_options(self, *, tool_choice: NotGivenOr[llm.ToolChoice | None] = NOT_GIVEN) ‑> None-
Expand source code
def update_options(self, *, tool_choice: NotGivenOr[llm.ToolChoice | None] = NOT_GIVEN) -> None: """Update session options and send session.update to Azure server.""" if not is_given(tool_choice) or tool_choice == self._opts.tool_choice: return self._opts.tool_choice = tool_choice self._send( ClientEventSessionUpdate( session=RequestSession(tool_choice=to_azure_tool_choice(tool_choice)) ) )Update session options and send session.update to Azure server.
async def update_tools(self, tools: list[llm.Tool]) ‑> None-
Expand source code
async def update_tools(self, tools: list[llm.Tool]) -> None: """Update available tools.""" async with self._update_fnc_ctx_lock: self._tools = llm.ToolContext(tools) # an empty list clears the tools of the session self._send( ClientEventSessionUpdate( session=RequestSession( tools=livekit_tools_to_azure_tools(self._tools.flatten()) ) ) )Update available tools.
Inherited members