Module livekit.agents.beta.translation.translation_session

Classes

class TranslationSession (*,
model: TranslationModel,
userdata: NotGivenOr[Userdata_T] = NOT_GIVEN,
output_segment_timeout: float = 0.5,
loop: asyncio.AbstractEventLoop | None = None)

Expand source code

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] | None = None
        self._output_transcript_atask: asyncio.Task[None] | None = None
        self._segment_timer: asyncio.TimerHandle | None = None
        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

    @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)

    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._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]()
            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()

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

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

                    await self._capture_output_frame(ev.frame, audio_output)

                if not marks_segments:
                    self._arm_segment_timer()
        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]) -> None:
        async for ev in event_ch:
            transcription_output = self.output.transcription
            if transcription_output is None or not self.output.transcription_enabled:
                continue

            await transcription_output.capture_text(ev.text)
            if ev.is_final:
                transcription_output.flush()

    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 self._output_transcript_ch is not None:
            self._output_transcript_ch.send_nowait(ev)

        self.emit(
            "translation_transcribed",
            TranslationTranscribedEvent(
                transcript=ev.text,
                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)
            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)

    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)

    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()

    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._close_output_segment_soon
        )

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

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

session = TranslationSession(model=VoiceModel(target_language="es"))
await session.start(room=ctx.room)

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.

Ancestors

• EventEmitter
• typing.Generic

Instance variables

prop input : io.AgentInput

Expand source code

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

prop model : TranslationModel

Expand source code

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

The translation model this session runs.

prop output : io.AgentOutput

Expand source code

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

prop room_io : TranslationRoomIO | None

Expand source code

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

prop user_state : UserState

Expand source code

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

prop userdata : Userdata_T

Expand source code

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

    return self._userdata

Methods

async def aclose(self) ‑> None

Expand source code

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

Stop translating and release the model and the room.

def set_target_language(self, language: str) ‑> None

Expand source code

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)

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.

def shutdown(self) ‑> None

Expand source code

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

Close the session without waiting for it.

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

Expand source code

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._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]()
        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",
        )

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.

Inherited members

• EventEmitter:
  • emit
  • off
  • on
  • once

class TranslationTranscribedEvent (**data: Any)

Expand source code

class TranslationTranscribedEvent(BaseModel):
    """The translation the model spoke, in the language it translates into."""

    type: Literal["translation_transcribed"] = "translation_transcribed"
    transcript: str
    is_final: bool
    language: str | None = None
    created_at: float = Field(default_factory=time.time)

The translation the model spoke, in the language it translates into.

Create a new model by parsing and validating input data from keyword arguments.

Raises [ValidationError][pydantic_core.ValidationError] if the input data cannot be validated to form a valid model.

self is explicitly positional-only to allow self as a field name.

Ancestors

• pydantic.main.BaseModel

Class variables

var created_at : float

var is_final : bool

var language : str | None

var model_config

var transcript : str

var type : Literal['translation_transcribed']