Module livekit.agents.beta

class EndCallTool(Toolset):
    def __init__(
        self,
        *,
        extra_description: str = "",
        delete_room: bool = True,
        end_instructions: str | None = "say goodbye to the user",
        ignore_on_enter: bool = False,
        on_tool_called: Callable[[Toolset.ToolCalledEvent], Awaitable[None]] | None = None,
        on_tool_completed: Callable[[Toolset.ToolCompletedEvent], Awaitable[None]] | None = None,
    ):
        """
        This tool allows the agent to end the call and disconnect from the room.

        Args:
            extra_description: Additional description to add to the end call tool.
            delete_room: Whether to delete the room when the user ends the call. deleting the room disconnects all remote users, including SIP callers.
            end_instructions: Tool output to the LLM for generating the tool response. With
                None the session ends once the turn that called the tool is done, which suits
                a model that speaks while calling it, such as GPT-Live.
            ignore_on_enter: Hide the tool during ``on_enter`` so the model can't end the call while greeting.
            on_tool_called: Callback to call when the tool is called.
            on_tool_completed: Callback to call when the tool is completed.
        """
        end_call_tool = function_tool(
            self._end_call,
            name="end_call",
            description=f"{END_CALL_DESCRIPTION}\n{extra_description}",
            flags=ToolFlag.IGNORE_ON_ENTER if ignore_on_enter else ToolFlag.NONE,
        )
        super().__init__(id="end_call", tools=[end_call_tool])

        self._delete_room = delete_room
        self._extra_description = extra_description

        self._end_instructions = end_instructions
        self._on_tool_called = on_tool_called
        self._on_tool_completed = on_tool_completed

        self._shutdown_session_task: asyncio.Task[None] | None = None

    async def _end_call(self, ctx: RunContext) -> Any | None:
        logger.debug("end_call tool called")
        activity = ctx.session.current_agent._get_activity_or_raise()
        llm_v = activity.llm
        output: Any = self._end_instructions

        def _on_speech_done(_: SpeechHandle) -> None:
            # read auto_tool_reply_generation from the active realtime session so a fallback
            # adapter reports the model actually in use
            rt_session = activity.realtime_llm_session
            auto_tool_reply = (
                rt_session is not None and rt_session.capabilities.auto_tool_reply_generation
            )
            # a result of None asks for no reply, so there is nothing to wait for. GPT-Live
            # answers it anyway (see its warning), and the answer can be silent
            if not isinstance(llm_v, RealtimeModel) or not auto_tool_reply or output is None:
                # tool reply will reuse the same speech handle, so we can shutdown the session
                # directly after this speech handle is done
                ctx.session.shutdown()
            else:
                self._shutdown_session_task = asyncio.create_task(
                    self._delayed_session_shutdown(ctx)
                )

        ctx.speech_handle.add_done_callback(_on_speech_done)
        ctx.session.once("close", self._on_session_close)

        if self._on_tool_called:
            await self._on_tool_called(Toolset.ToolCalledEvent(ctx=ctx, arguments={}))

        completed_ev = Toolset.ToolCompletedEvent(ctx=ctx, output=self._end_instructions)
        if self._on_tool_completed:
            await self._on_tool_completed(completed_ev)

        output = completed_ev.output
        return output

    async def _delayed_session_shutdown(self, ctx: RunContext) -> None:
        """Shutdown the session after the tool reply is played out"""
        speech_created_fut = asyncio.Future[SpeechHandle]()

        @ctx.session.once("speech_created")
        def _on_speech_created(ev: SpeechCreatedEvent) -> None:
            if not speech_created_fut.done():
                speech_created_fut.set_result(ev.speech_handle)

        try:
            speech_handle = await asyncio.wait_for(speech_created_fut, timeout=5.0)
            await speech_handle
        except asyncio.TimeoutError:
            logger.warning("tool reply timed out, shutting down session")
        finally:
            ctx.session.off("speech_created", _on_speech_created)
            ctx.session.shutdown()

    def _on_session_close(self, ev: CloseEvent) -> None:
        """Close the job process when AgentSession is closed"""
        if self._shutdown_session_task:
            # cleanup
            self._shutdown_session_task.cancel()
            self._shutdown_session_task = None

        job_ctx = get_job_context()

        if self._delete_room:

            async def _on_shutdown() -> None:
                logger.info("deleting the room because the user ended the call")
                await job_ctx.delete_room()

            job_ctx.add_shutdown_callback(_on_shutdown)

        # shutdown the job process
        job_ctx.shutdown(reason=ev.reason.value)

class Instructions:
    """Instructions with optional modality-specific additions.

    Construction::

        # Simple — same instructions for all modalities
        Instructions("You are a helpful assistant.")

        # With modality-specific additions
        Instructions(
            "You are a helpful assistant.",
            audio="Keep responses short for voice.",
            text="Use markdown formatting.",
        )

        # With per-call context kept out of the cached prefix
        Instructions(
            "You are a helpful assistant.",
            dynamic=f"Current time: {now}. Caller: {caller_id}.",
        )

    Rendering::

        instr.render()                              # → common text
        instr.render(modality="audio")               # → common + audio addition
        instr.render(modality="text", name="Alex")   # → common + text, with {name} filled

    ``dynamic`` holds text that changes from call to call, such as the time or the
    caller's number. :meth:`render` appends it after a single newline. The voice pipeline
    stores it as its own system message right after the instructions message, so the
    instructions message stays identical across calls and can end a cached prompt prefix
    (see :func:`livekit.agents.voice.generation.update_instructions`).
    """

    def __init__(
        self,
        common: str = "",
        *,
        audio: str | None = None,
        text: str | None = None,
        dynamic: str | None = None,
    ) -> None:
        self.common = common
        self.audio = audio
        self.text = text
        self.dynamic = dynamic

    def render(
        self,
        *,
        modality: Literal["audio", "text"] | None = None,
        data: dict[str, object] | None = None,
    ) -> str:
        """Render instructions to a plain string.

        Args:
            modality: If given, appends the modality-specific addition to the common text.
            data: Template variables to fill. Missing placeholders log a warning
                and are replaced with empty strings.
        """
        # one newline, not the section separator: the pipeline stores dynamic as the next
        # message, and provider formatters join messages' text with one newline
        parts = [self.render_static(modality=modality), self.dynamic or ""]
        result = "\n".join(p for p in parts if p)

        if data:
            result = utils.misc.safe_render(result, data)

        return result

    def render_static(self, *, modality: Literal["audio", "text"] | None = None) -> str:
        """Render the common text and the modality addition, without ``dynamic``."""
        parts = [self.common]
        if modality is not None:
            addition = self.audio if modality == "audio" else self.text
            if addition:
                parts.append(addition)
        return "\n\n".join(p for p in parts if p)

    @staticmethod
    def resolve_template(template: str, **kwargs: object) -> Instructions:
        """Fill a template string, producing an ``Instructions`` with modality variants.

        If any kwarg value is an ``Instructions`` object, its ``common``/``audio``/``text``
        parts are substituted into the matching variant of the result. This is used by
        workflow tasks to build modality-aware instructions from a single template.
        ``dynamic`` is not templated: the result never carries one.
        """
        any_instructions = any(isinstance(v, Instructions) for v in kwargs.values())
        if any_instructions:
            audio_kw: dict[str, object] = {
                # an explicit "" removes the section; only None falls back to common
                k: (v.audio if v.audio is not None else str(v))
                if isinstance(v, Instructions)
                else v
                for k, v in kwargs.items()
            }
            text_kw: dict[str, object] = {
                k: (v.text if v.text is not None else str(v)) if isinstance(v, Instructions) else v
                for k, v in kwargs.items()
            }
            # audio/text hold fully rendered variants of the whole template, so they go in
            # place of common (which render() would otherwise prepend, doubling the template).
            audio = utils.misc.safe_render(template, audio_kw)
            text = utils.misc.safe_render(template, text_kw)
            if audio == text:
                # no modality-specific differences; a single common variant renders correctly
                # with or without a modality
                return Instructions(common=audio)
            return Instructions(common="", audio=audio, text=text)
        else:
            rendered = utils.misc.safe_render(template, kwargs)
            return Instructions(common=rendered)

    def __str__(self) -> str:
        return self.common

    def __repr__(self) -> str:
        if self.dynamic is None:
            return f"Instructions({self.common!r})"
        return f"Instructions({self.common!r}, dynamic={self.dynamic!r})"

    def __hash__(self) -> int:
        return hash((self.common, self.audio, self.text, self.dynamic))

    def __eq__(self, other: object) -> bool:
        if isinstance(other, Instructions):
            return (
                self.common == other.common
                and self.audio == other.audio
                and self.text == other.text
                and self.dynamic == other.dynamic
            )
        if isinstance(other, str):
            # a plain string has no sections, so only section-free instructions equal it;
            # the realtime handoff path reuses a session on this equality
            return self.common == other and not self.audio and not self.text and not self.dynamic
        return NotImplemented

class TranslationModel(ABC):
    """A model that hears speech in one language and speaks it in another.

    Its own primitive, beside :class:`~livekit.agents.llm.LLM`,
    :class:`~livekit.agents.llm.RealtimeModel` and :class:`~livekit.agents.llm.DuplexModel`: a
    translator holds no conversation, so it takes no instructions, carries no chat context,
    calls no tools and never decides to reply. What it has instead is a language it translates
    into. Run one with
    :class:`~livekit.agents.beta.translation.TranslationSession`.
    """

    def __init__(
        self,
        *,
        capabilities: TranslationCapabilities,
        target_language: str,
        source_language: str | None = None,
    ) -> None:
        """
        Args:
            capabilities: What this provider can do.
            target_language: The language the model speaks its translation in.
            source_language: The language the model expects to hear, or ``None`` when it
                detects the language itself.
        """
        self._capabilities = capabilities
        self._target_language = target_language
        self._source_language = source_language
        self._label = f"{type(self).__module__}.{type(self).__name__}"

    @property
    def model(self) -> str:
        return "unknown"

    @property
    def provider(self) -> str:
        return "unknown"

    @property
    def capabilities(self) -> TranslationCapabilities:
        return self._capabilities

    @property
    def label(self) -> str:
        return self._label

    @property
    def target_language(self) -> str:
        """The language the model speaks its translation in."""
        return self._target_language

    @property
    def source_language(self) -> str | None:
        """The language the model expects to hear, or ``None`` when it detects it itself."""
        return self._source_language

    def update_options(self, *, target_language: str) -> None:
        """Point the model at another language.

        Takes effect on the sessions opened after it. To retarget a running session, call
        :meth:`TranslationModelSession.update_options` on the session itself, which a
        provider only serves when ``capabilities.retargetable``.
        """
        self._target_language = target_language

    def audio_gate(self) -> AudioGate | None:
        """The gate that tells this model's speech from its silence, or None for the default.

        A provider can stream audio without pause, its silence included; the session ends a
        translation once the gate hears it go quiet, rather than once frames stop arriving.
        Return a gate measured on the provider's own silence, as a duplex model does.
        """
        return None

    @abstractmethod
    def session(self) -> TranslationModelSession:
        """Open a session; the caller pushes audio into it and reads its translation out."""

    @abstractmethod
    async def aclose(self) -> None: ...

    async def __aenter__(self) -> TranslationModel:
        return self

    async def __aexit__(
        self,
        exc_type: type[BaseException] | None,
        exc: BaseException | None,
        exc_tb: TracebackType | None,
    ) -> None:
        await self.aclose()

class TranslationSession(rtc.EventEmitter[EventTypes], Generic[Userdata_T]):
    """Runs a :class:`TranslationModel` against a room: speech in, translation out.

    Its own primitive rather than an ``AgentSession``: a translator holds no conversation. It
    takes no instructions, keeps no chat context, calls no tools, and never decides to reply —
    so there is no ``Agent`` to supply and no turn to take. What is left it does carry: room
    I/O, both transcripts, user and agent state, and a close that reports why.

    A model reports the speaker's turns itself, so no VAD runs here, and nothing waits for the
    speaker to stop before speaking over them.

    Example:
        ```python
        session = TranslationSession(model=VoiceModel(target_language="es"))
        await session.start(room=ctx.room)
        ```
    """

    def __init__(
        self,
        *,
        model: TranslationModel,
        userdata: NotGivenOr[Userdata_T] = NOT_GIVEN,
        output_segment_timeout: float = DEFAULT_OUTPUT_SEGMENT_TIMEOUT,
        loop: asyncio.AbstractEventLoop | None = None,
    ) -> None:
        """
        Args:
            model: The translation model this session runs.
            userdata: Arbitrary state carried on the session.
            output_segment_timeout: Seconds of silence from the model that end a segment of
                its speech. Used only when the model's ``capabilities.output_segmentation``
                is unset; a model that marks its own boundaries ends a segment on the mark,
                so a pause inside a sentence does not split it.
            loop: Event loop to run on.
        """
        super().__init__()
        self._loop = loop or asyncio.get_event_loop()
        self._model = model
        self._userdata: Userdata_T | None = userdata if is_given(userdata) else None
        self._output_segment_timeout = output_segment_timeout

        self._input = io.AgentInput(self._on_video_input_changed, self._on_audio_input_changed)
        self._output = io.AgentOutput(
            self._on_video_output_changed,
            self._on_audio_output_changed,
            self._on_text_output_changed,
        )

        self._room_io: TranslationRoomIO | None = None
        self._model_session: TranslationModelSession | None = None

        self._user_state: UserState = "listening"

        self._started = False
        self._closing_task: asyncio.Task[None] | None = None
        self._job_context_cb_registered = False
        self._lock = asyncio.Lock()

        self._forward_input_atask: asyncio.Task[None] | None = None
        self._forward_output_atask: asyncio.Task[None] | None = None
        self._output_transcript_ch: aio.Chan[OutputTranscript | object] | None = None
        # the translation transcribed so far, for the event; the model sends it in pieces
        self._output_transcript_text = ""
        self._output_transcript_atask: asyncio.Task[None] | None = None
        self._segment_timer: asyncio.TimerHandle | None = None
        # tells the model's speech from its silence, which a provider can stream without pause
        self._gate: AudioGate | None = None
        self._gate_open = False
        # set while a quiet segment waits for its transcript to go final before it ends
        self._awaiting_text = False
        self._text_wait_timeout = TEXT_WAIT_TIMEOUT
        self._segment_tasks: set[asyncio.Task[None]] = set()
        # serialises capturing a frame against ending a segment, so a tail the resampler
        # holds cannot be flushed out from under a frame being captured
        self._output_lock = asyncio.Lock()
        self._resampler: rtc.AudioResampler | None = None
        self._resampler_key: tuple[int, int, int | None] | None = None
        self._resampler_sink: io.AudioOutput | None = None
        # the language the room labels the translation with. It follows the language stamped
        # on the speech itself, which trails a retarget until the provider applies it, and
        # follows the request only for a provider that stamps nothing
        self._output_language: str | None = None
        self._output_language_stamped = False
        # whether the open segment has any audio in it; an empty one is not worth ending
        self._segment_has_audio = False
        # whether translation text has arrived since the last segment ended, which the
        # synchronizer holds for audio still to come
        self._segment_has_text = False
        # whether that text is still going: some has arrived, and its final has not
        self._text_open = False
        # a label to apply once the open segment ends: requested while it held text waiting for
        # audio that has not come, which is still in the old language
        self._relabel_at_segment_end: str | None = None

    @property
    def model(self) -> TranslationModel:
        """The translation model this session runs."""
        return self._model

    @property
    def input(self) -> io.AgentInput:
        return self._input

    @property
    def output(self) -> io.AgentOutput:
        return self._output

    @property
    def user_state(self) -> UserState:
        return self._user_state

    @property
    def room_io(self) -> TranslationRoomIO | None:
        return self._room_io

    @property
    def userdata(self) -> Userdata_T:
        if self._userdata is None:
            raise ValueError("TranslationSession userdata is not set")

        return self._userdata

    @userdata.setter
    def userdata(self, value: Userdata_T) -> None:
        self._userdata = value

    def set_target_language(self, language: str) -> None:
        """Point the model at another language.

        Before :meth:`start` this only changes what the next session is opened with. On a
        running session it reaches the model, which a provider serves only when its
        ``capabilities.retargetable``.

        Raises:
            RuntimeError: If the session is running and its model cannot be retargeted.
        """
        if self._model_session is not None and not self._model.capabilities.retargetable:
            raise RuntimeError(
                f"{self._model.label} translates into a language fixed for the session's "
                "life; start a new session to translate into another"
            )

        # the running session first: it can refuse a language the capability allows in
        # general, and the model setting must not claim a language being spoken
        if self._model_session is not None:
            self._model_session.update_options(target_language=language)

        self._model.update_options(target_language=language)
        if self._model_session is None:
            self._relabel_output(language)
        elif not self._output_language_stamped:
            # nothing stamps the speech with its language, so the label follows the request.
            # The open segment ends first: speech that never pauses would otherwise keep it
            # open, and the old label with it
            task = asyncio.create_task(self._close_segment_and_relabel(language))
            self._segment_tasks.add(task)
            task.add_done_callback(self._segment_tasks.discard)

    async def start(
        self,
        *,
        room: NotGivenOr[rtc.Room] = NOT_GIVEN,
        room_options: NotGivenOr[TranslationRoomOptions] = NOT_GIVEN,
    ) -> None:
        """Start translating.

        Args:
            room: The room to listen to and speak into. Omit it to drive the session's
                ``input``/``output`` yourself.
            room_options: How the session reaches that room.
        """
        async with self._lock:
            if self._started:
                return

            job_ctx = get_job_context(required=False)

            if is_given(room) and self._room_io is None:
                options = room_options if is_given(room_options) else TranslationRoomOptions()

                if self.input.audio is not None and options.audio_input:
                    logger.warning(
                        "room audio input is enabled but input.audio is already set, ignoring.."
                    )
                    options.audio_input = False

                if self.output.audio is not None and options.audio_output:
                    logger.warning(
                        "room audio output is enabled but output.audio is already set, ignoring.."
                    )
                    options.audio_output = False

                self._room_io = TranslationRoomIO(self, room, options=options)
                await self._room_io.start()

                if job_ctx:
                    await job_ctx.connect()

            if job_ctx and not self._job_context_cb_registered:
                job_ctx.add_shutdown_callback(
                    lambda: self._aclose_impl(reason=CloseReason.JOB_SHUTDOWN)
                )
                self._job_context_cb_registered = True

            self._model_session = self._model.session()
            self._output_language = self._model.target_language
            self._output_language_stamped = False
            self._reset_segment_state()
            # digital silence unless the model knows its own; its quiet period is the segment's
            self._gate = self._model.audio_gate() or FixedGate(
                0.0, min_silence_duration=self._output_segment_timeout
            )
            self._gate_open = False
            self._model_session.on("input_speech_started", self._on_input_speech_started)
            self._model_session.on("input_speech_stopped", self._on_input_speech_stopped)
            self._model_session.on("input_transcript", self._on_input_transcript)
            self._model_session.on("output_transcript", self._on_output_transcript)
            self._model_session.on("error", self._on_model_error)

            self._output_transcript_ch = aio.Chan[OutputTranscript | object]()
            self._output_transcript_text = ""
            self._output_transcript_atask = asyncio.create_task(
                self._forward_output_transcript_task(self._output_transcript_ch),
                name="_forward_output_transcript_task",
            )

            self._closing_task = None
            self._started = True

            if self.input.audio is not None:
                self._forward_input_atask = asyncio.create_task(
                    self._forward_input_audio_task(), name="_forward_input_audio_task"
                )

            self._forward_output_atask = asyncio.create_task(
                self._forward_output_audio_task(self._model_session),
                name="_forward_output_audio_task",
            )

    async def aclose(self) -> None:
        """Stop translating and release the model and the room."""
        await self._aclose_impl(reason=CloseReason.USER_INITIATED)

    def shutdown(self) -> None:
        """Close the session without waiting for it."""
        self._close_soon(reason=CloseReason.USER_INITIATED)

    def _close_soon(self, *, reason: CloseReason) -> None:
        if self._closing_task:
            return

        self._closing_task = asyncio.create_task(self._aclose_impl(reason=reason))

    @log_exceptions(logger=logger)
    async def _aclose_impl(self, *, reason: CloseReason) -> None:
        async with self._lock:
            if not self._started:
                return

            self._started = False
            self._cancel_segment_timer()
            if self._segment_tasks:
                await aio.cancel_and_wait(*self._segment_tasks)
                self._segment_tasks.clear()
            # a label deferred for this run's segments must not outlive them
            self._reset_segment_state()

            await aio.cancel_and_wait(
                *[
                    task
                    for task in (self._forward_input_atask, self._forward_output_atask)
                    if task is not None
                ]
            )
            self._forward_input_atask = self._forward_output_atask = None

            if self._output_transcript_ch:
                self._output_transcript_ch.close()
                self._output_transcript_ch = None
            if self._output_transcript_atask:
                await aio.cancel_and_wait(self._output_transcript_atask)
                self._output_transcript_atask = None

            if self._model_session:
                await self._model_session.aclose()
                self._model_session = None

            # detach what room io made, since it closes those on the way out and a restart
            # would otherwise reuse the closed devices. A caller's own stay attached: they
            # were never closed here, and losing the reference would strand them
            if (room_io := self._room_io) is not None:
                if self.input.audio is room_io.audio_input:
                    self.input.audio = None
                if self.output.audio is room_io.audio_output:
                    self.output.audio = None
                if self.output.transcription is room_io.transcription_output:
                    self.output.transcription = None

            self.emit("close", CloseEvent(reason=reason))

            if self._room_io:
                await self._room_io.aclose()
                self._room_io = None

            self._discard_resampler()
            self._user_state = "listening"

    # -- model session --

    @log_exceptions(logger=logger)
    async def _forward_input_audio_task(self) -> None:
        audio_input = self.input.audio
        if audio_input is None:
            return

        async for frame in audio_input:
            if (model_session := self._model_session) is not None:
                model_session.push_audio(frame)

    @log_exceptions(logger=logger)
    async def _forward_output_audio_task(self, model_session: TranslationModelSession) -> None:
        marks_segments = self._model.capabilities.output_segmentation
        try:
            async for ev in model_session.audio_stream:
                if isinstance(ev, OutputSpeechStopped):
                    # read in order behind the audio it closes, so nothing it ended is
                    # still waiting to be captured. Honoured even while the output is
                    # muted: whatever was captured before the mute is still a segment, and
                    # dropping the mark would merge the next translation into it
                    await self._close_output_segment()
                    continue

                if ev.language is not None:
                    self._output_language_stamped = True
                    # the speech says what language it is in, so that drives the label
                    self._relabel_at_segment_end = None
                    if ev.language != self._output_language:
                        # the old language's segment ends before speech in the new one, so
                        # neither is labelled with the other's language
                        await self._close_output_segment()
                        self._relabel_output(ev.language)

                if self.output.audio is None or not self.output.audio_enabled:
                    continue

                was_open = self._gate_open
                self._gate_open = self._gate.update(ev.frame) if self._gate else True
                if not self._gate_open and not self._segment_has_audio:
                    # silence between translations, from a provider that streams it: there
                    # is nothing to play, and played it would hold the next segment open
                    continue

                async with self._output_lock:
                    if (audio_output := self.output.audio) is None:
                        continue

                    if not self._segment_has_audio:
                        # a segment starts once the one before has played, as the sink
                        # requires; the model's speech waits in its stream meanwhile
                        await audio_output.wait_for_playout()

                    await self._capture_output_frame(ev.frame, audio_output)

                if not marks_segments:
                    if self._gate_open:
                        # the pauses inside a translation still play, and the timer still
                        # ends it for a provider that stops sending rather than go quiet
                        self._arm_segment_timer()
                    elif was_open:
                        # the gate has heard the speech stay quiet long enough
                        self._cancel_segment_timer()
                        self._on_output_silence()
        finally:
            self._cancel_segment_timer()
            async with self._output_lock:
                await self._drain_resampler()

    @log_exceptions(logger=logger)
    async def _forward_output_transcript_task(
        self, event_ch: aio.Chan[OutputTranscript | object]
    ) -> None:
        async for ev in event_ch:
            transcription_output = self.output.transcription
            if transcription_output is None or not self.output.transcription_enabled:
                continue

            if ev is _END_OF_SEGMENT:
                transcription_output.flush()
                continue

            assert isinstance(ev, OutputTranscript)
            if ev.text:
                await transcription_output.capture_text(ev.text)
            if ev.is_final and not self._text_follows_audio():
                transcription_output.flush()

    def _text_follows_audio(self) -> bool:
        """Whether the translation's text segments end with its audio segments.

        A synchronizer pairs each audio segment with one text segment, and releases the text
        only once both have ended. Ended on its own schedule, the text would close after its
        audio had played, and the words not yet released would be lost. So while audio is
        going out, the text ends where the audio does, marked in the transcript stream so it
        stays behind the text that came before it.
        """
        return self.output.audio is not None and self.output.audio_enabled

    def _relabel_output(self, language: str) -> None:
        """Label the translation in the room with the language it is now spoken in."""
        self._output_language = language
        if self._room_io is not None:
            self._room_io._on_output_language_changed(language)

    def _on_input_speech_started(self, _: InputSpeechStarted) -> None:
        self._update_user_state("speaking")

    def _on_input_speech_stopped(self, _: InputSpeechStopped) -> None:
        self._update_user_state("listening")

    def _on_input_transcript(self, ev: InputTranscript) -> None:
        self.emit(
            "user_input_transcribed",
            UserInputTranscribedEvent(
                transcript=ev.text,
                is_final=ev.is_final,
                language=ev.language,
                speaker_id=ev.speaker_id,
            ),
        )

    def _on_output_transcript(self, ev: OutputTranscript) -> None:
        if ev.text:
            self._segment_has_text = True
            self._text_open = True
        if ev.is_final:
            self._text_open = False
        if self._output_transcript_ch is not None:
            self._output_transcript_ch.send_nowait(ev)

        if ev.is_final and self._awaiting_text:
            # the quiet segment was only waiting for this; it ends behind it
            self._cancel_segment_timer()
            self._close_output_segment_soon()

        # the model sends pieces; the event carries all of it, so a final is never empty
        self._output_transcript_text += ev.text
        # the model's pieces carry their leading spaces, which open the transcript too
        transcript = self._output_transcript_text.lstrip()
        if ev.is_final:
            self._output_transcript_text = ""

        self.emit(
            "translation_transcribed",
            TranslationTranscribedEvent(
                transcript=transcript,
                is_final=ev.is_final,
                language=ev.language if ev.language is not None else self._model.target_language,
            ),
        )

    def _on_model_error(self, ev: TranslationModelError) -> None:
        self.emit("error", ErrorEvent(error=ev, source=self._model))

        if not ev.recoverable:
            logger.error(
                "translation model error, closing the session",
                extra={"label": ev.label, "error": str(ev.error)},
            )
            self._close_soon(reason=CloseReason.ERROR)

    # -- output segments --

    async def _capture_output_frame(
        self, frame: rtc.AudioFrame, audio_output: io.AudioOutput
    ) -> None:
        """Hand one frame to the sink, at the rate that sink plays at.

        The resampler is rebuilt whenever the model's format or the sink changes, since one
        built for the old pair would tag frames with a rate the new one does not play.
        """
        target_rate = audio_output.sample_rate
        key = (frame.sample_rate, frame.num_channels, target_rate)

        if self._resampler is not None and (
            self._resampler_key != key or self._resampler_sink is not audio_output
        ):
            if self._resampler_sink is audio_output:
                # same sink, new format: its tail still belongs there, in this segment
                await self._drain_resampler()
            else:
                # the sink was replaced: its segment ends on the sink that received it
                logger.debug("audio output replaced, closing the segment it was given")
                await self._close_segment_locked()

        if target_rate is None or frame.sample_rate == target_rate:
            await audio_output.capture_frame(frame)
            self._segment_has_audio = True
            return

        if self._resampler is None:
            self._resampler = rtc.AudioResampler(
                input_rate=frame.sample_rate,
                output_rate=target_rate,
                num_channels=frame.num_channels,
            )
            self._resampler_key = key
            self._resampler_sink = audio_output

        for resampled in self._resampler.push(frame):
            await audio_output.capture_frame(resampled)
            self._segment_has_audio = True

    async def _drain_resampler(self) -> None:
        """Push what the resampler still holds to the sink it was built for, then drop it.

        Not to whatever sink is current: those frames were converted for the old one, and
        another sink would replay them at its own rate.
        """
        resampler, sink = self._resampler, self._resampler_sink
        self._discard_resampler()
        if resampler is None or sink is None:
            return

        for frame in resampler.flush():
            await sink.capture_frame(frame)
            self._segment_has_audio = True

    def _discard_resampler(self) -> None:
        """Forget the resampler and the pair it was built for, dropping anything it holds.

        Used where that tail cannot be placed, or once :meth:`_drain_resampler` has emptied it.
        """
        self._resampler = None
        self._resampler_key = None
        self._resampler_sink = None

    async def _close_output_segment(self) -> None:
        """End the segment the model has been speaking into."""
        self._cancel_segment_timer()
        async with self._output_lock:
            await self._close_segment_locked()

    async def _close_segment_locked(self) -> None:
        """End the segment on the sink that actually received its audio.

        Drains the resampler first: a partial frame it is still holding belongs to the
        segment being closed, not to whatever the model says next. The sink it was built
        for owns the segment, so replacing ``output.audio`` mid-translation closes the old
        sink's segment rather than flushing the replacement's.
        """
        sink = self._resampler_sink or self.output.audio
        await self._drain_resampler()
        if sink is not None:
            sink.flush()

        had_audio, self._segment_has_audio = self._segment_has_audio, False
        self._segment_has_text = False
        self._text_open = False
        if self._gate is not None:
            # the next speech opens a segment of its own, however the last one ended
            self._gate.deactivate()
        self._gate_open = False
        if had_audio and self._output_transcript_ch is not None:
            # the text segment ends with its audio, behind the text already sent
            self._output_transcript_ch.send_nowait(_END_OF_SEGMENT)
        if had_audio and (language := self._relabel_at_segment_end) is not None:
            # the text that was waiting on this segment's audio has gone out with it
            self._relabel_at_segment_end = None
            self._relabel_output(language)

    def _arm_segment_timer(self) -> None:
        """Restart the silence that ends a segment, for a model that marks none itself.

        Armed after every frame, so the segment ends only once the model has gone quiet for
        ``output_segment_timeout``.
        """
        self._cancel_segment_timer()
        self._segment_timer = self._loop.call_later(
            self._output_segment_timeout, self._on_output_silence
        )

    def _cancel_segment_timer(self) -> None:
        """Disarm the silence timer, or the wait for text, leaving the current segment open."""
        self._awaiting_text = False
        if self._segment_timer is not None:
            self._segment_timer.cancel()
            self._segment_timer = None

    def _on_output_silence(self) -> None:
        """End the segment the model went quiet in, once its transcript has ended too.

        Text still arriving, or not arrived yet, belongs to this segment's audio; ended now,
        the segment would leave it to the next one, or to none. So it waits for the final,
        up to ``TEXT_WAIT_TIMEOUT``, and speech resuming meanwhile carries the segment on.
        """
        self._segment_timer = None
        if self._waits_for_text() and (self._text_open or not self._segment_has_text):
            self._awaiting_text = True
            self._segment_timer = self._loop.call_later(
                self._text_wait_timeout, self._close_output_segment_soon
            )
            return

        self._close_output_segment_soon()

    def _waits_for_text(self) -> bool:
        """Whether a segment's end waits on its transcript: the model sends one, and it is
        published paired with the audio."""
        return (
            self._model.capabilities.output_transcription
            and self._text_follows_audio()
            and self.output.transcription is not None
            and self.output.transcription_enabled
        )

    async def _close_segment_and_relabel(self, language: str) -> None:
        """End the open segment, then label the translation with ``language``.

        For a provider that does not stamp its speech: the segment the old label covers has
        to end, or its playout never finishes and the new label never lands. A segment with
        text but no audio yet is left open, since ending it would discard that text, and the
        label changes once it ends. With neither, the label changes now.
        """
        async with self._output_lock:
            # speech the resampler still holds is the old language's too
            await self._drain_resampler()
            if self._segment_has_audio:
                self._cancel_segment_timer()
                await self._close_segment_locked()
            elif self._segment_has_text:
                # the old-language text waits for its audio; the label waits for both
                self._relabel_at_segment_end = language
                return

        self._relabel_at_segment_end = None
        self._relabel_output(language)

    def _reset_segment_state(self) -> None:
        """Forget what the open segment held, and any label deferred until it ends."""
        self._segment_has_audio = False
        self._segment_has_text = False
        self._text_open = False
        self._relabel_at_segment_end = None

    def _close_output_segment_soon(self) -> None:
        """Close the segment from the timer, which cannot await the drain itself.

        The work runs as a task, tracked so closing the session can wait on it.
        """
        self._segment_timer = None
        self._awaiting_text = False
        task = asyncio.create_task(self._close_output_segment())
        self._segment_tasks.add(task)
        task.add_done_callback(self._segment_tasks.discard)

    # -- state --

    def _update_user_state(self, state: UserState) -> None:
        if self._user_state == state:
            return

        old_state = self._user_state
        self._user_state = state
        self.emit("user_state_changed", UserStateChangedEvent(old_state=old_state, new_state=state))

    # -- io --

    def _on_audio_input_changed(self) -> None:
        if not self._started:
            return

        if self._forward_input_atask is not None:
            self._forward_input_atask.cancel()
            self._forward_input_atask = None

        if self.input.audio is not None:
            self._forward_input_atask = asyncio.create_task(
                self._forward_input_audio_task(), name="_forward_input_audio_task"
            )

    def _on_video_input_changed(self) -> None:
        if self.input.video is not None:
            logger.warning("a translation model takes no video input, ignoring..")

    def _on_audio_output_changed(self) -> None: ...

    def _on_video_output_changed(self) -> None: ...

    def _on_text_output_changed(self) -> None: ...