Module livekit.plugins.ai_coustics.plugin

def to_native_buffer(data: memoryview) -> tuple[np.ndarray, NativeAudioBufferMut]:
    """
    Convert frame.data (int16 memoryview) to NativeAudioBufferMut (f32 pointer).
    Returns both the numpy array (to keep it alive) and the NativeAudioBufferMut.
    """
    # Convert int16 to float32 in range [-1.0, 1.0]
    # astype() creates a copy, which is writable by default
    samples = (
        np.frombuffer(data, dtype=np.int16).astype(np.float32, copy=True) / 32768.0
    )

    # Get the memory address directly from the numpy array
    ptr_value = samples.ctypes.data

    # Create NativeAudioBufferMut pointing to the numpy memory
    native_buffer = NativeAudioBufferMut(
        ptr=ptr_value,
        len=len(samples),  # Number of f32 samples
    )

    return samples, native_buffer

class AICousticsAudioEnhancer(rtc.FrameProcessor[rtc.AudioFrame]):

    def __init__(
        self,
        *,
        model: EnhancerModel,
        vad_settings: VadSettings,
        model_parameters: Optional[ModelParameters] = None,
        auth: Optional[AuthBase] = None,
    ) -> None:
        self._model = model
        self._vad_settings = vad_settings
        self._model_parameters = model_parameters
        self._auth = auth or Auth.livekit_cloud()
        self._last_error_msg: Optional[str] = None

        self._enhancer: Enhancer | None = None
        self._info: StreamInfo | None = None
        self._credentials: Credentials | None = None
        self._settings: EnhancerSettings | None = None
        self._enabled = True

    @property
    def enabled(self) -> bool:
        return self._enabled

    @enabled.setter
    def enabled(self, value: bool) -> None:
        self._enabled = value

    def update_model_parameters(self, model_parameters: ModelParameters):
        """
        Updates the model parameters on the running model.

        The native core must already exist (i.e. at least one audio frame must
        have been processed) for the update to take effect; otherwise the call
        is a no-op and a warning is logged. The new parameters are also stored
        so they are reapplied if the native core is later recreated (e.g. on a
        sample-rate or channel change).
        """
        if not self._enhancer:
            logger.warning("update_model_parameters: Native core not yet initialized, skipping. Process at least one audio frame first.")
            return
        new_uniffi = model_parameters._to_uniffi()
        current_uniffi = (
            self._model_parameters._to_uniffi()
            if self._model_parameters is not None
            else ModelParametersUniffi(bypass=None, enhancement_level=None)
        )
        if model_parameters_equal(new_uniffi, current_uniffi):
            return
        self._model_parameters = model_parameters
        self._enhancer.update_model_parameters(new_uniffi)

    def _on_stream_info_updated(
        self, *, room_name: str, participant_identity: str, publication_sid: str
    ):
        self._info = StreamInfo(
            room_id="",
            room_name=room_name,
            participant_identity=participant_identity,
            participant_id="",
            track_id=publication_sid,
        )
        if self._enhancer is not None:
            self._enhancer.update_stream_info(self._info)

    def _on_credentials_updated(self, *, token: str, url: str):
        self._credentials = Credentials(token=token, url=url)
        if self._enhancer is not None:
            self._enhancer.update_credentials(self._credentials)

    def _process(self, frame: rtc.AudioFrame) -> rtc.AudioFrame:
        """
        Processes a single audio frame.

        If the frame processor is disabled or processing fails, the original frame is
        returned unchanged.
        """
        if not self.enabled:
            return frame

        auth_mode = self._auth._to_auth_mode(self._credentials)
        if auth_mode is None:
            self._log_process_frame_error("Missing auth mode")
            return frame

        if self._auth_mode_requires_credentials() and not self._credentials:
            self._log_process_frame_error("Missing credentials")
            return frame

        if self._auth_mode_requires_stream_info() and self._info is None:
            self._log_process_frame_error("Missing stream info")
            return frame

        ## lazily create enhancer
        if self._enhancer is None or (
            ## implicitly recreate audio enhancer on sample rate or channel changes
            self._settings is not None
            and (
                self._settings.sample_rate != frame.sample_rate
                or self._settings.num_channels != frame.num_channels
                or self._settings.samples_per_channel != frame.samples_per_channel
            )
        ):
            self._settings = EnhancerSettings(
                sample_rate=frame.sample_rate,
                num_channels=frame.num_channels,
                samples_per_channel=frame.samples_per_channel,
                model=self._model,
                model_parameters=self._model_parameters._to_uniffi() if self._model_parameters else ModelParametersUniffi(bypass=None, enhancement_level=None),
                vad=self._vad_settings
            )
            try:
                self._enhancer = Enhancer(auth_mode, self._settings)
            except EnhancerError as e:
                self._log_process_frame_error(f"Failed to initialize plugin core: {e} - Disabling noise cancellation for all following audio frames.")
                self._enhancer = None
                self._enabled = False
                return frame
            if self._info is not None:
                self._enhancer.update_stream_info(self._info)

        # Convert frame.data to NativeAudioBufferMut (f32)
        # Keep samples alive during the process call
        samples, native_buffer = to_native_buffer(frame.data)

        # Process in-place (modifies samples array)
        try:
            vad_data = self._enhancer.process_with_vad(native_buffer)
        except EnhancerError as e:
            self._log_process_frame_error(f"Processing failed: {e}")
            return frame

        # Convert back to int16 and create new frame
        processed_int16 = (np.clip(samples, -1.0, 1.0) * 32767.0).astype(np.int16)

        output_frame = rtc.AudioFrame(
            data=processed_int16.tobytes(),
            sample_rate=frame.sample_rate,
            num_channels=frame.num_channels,
            samples_per_channel=frame.samples_per_channel,
            userdata=frame.userdata,
        )
        output_frame.userdata[FRAME_USERDATA_AIC_VAD_ATTRIBUTE] = vad_data
        return output_frame

    def _auth_mode_requires_stream_info(self) -> bool:
        """Does the given auth mode require update_stream_info be called?"""
        return isinstance(self._auth, LiveKitCloud)

    def _auth_mode_requires_credentials(self) -> bool:
        """
        Does the given auth mode require update_credentials be called?

        Note that this is just here to provide helpful warnings to users,
        the actual auth layer is in the rust core.
        """
        return isinstance(self._auth, LiveKitCloud)

    def _log_process_frame_error(self, msg: str):
        """
        Logs a new error to the screen when processing a frame.
        Only shows logs which were newly introduced as compared with the
        last processed frame.
        """
        if self._last_error_msg == msg:
            return
        self._last_error_msg = msg
        logger.error(msg)

    def _close(self):
        if self._enhancer is not None:
            self._enhancer = None

@dataclass
class ModelParameters:
    enhancement_level: Optional[float] = None
    bypass: Optional[float] = None

    def _to_uniffi(self):
        return ModelParametersUniffi(
            enhancement_level=self.enhancement_level,
            bypass=self.bypass,
        )