EMBEDDED AI-BASED CONTACTLESS VOICE CONTROL MODULE FOR SURGICAL LIGHTING SYSTEMS

Abstract

The invention relates to a contactless voice control module that enables the control of light intensity, focus area, spot width, color temperature, and endoscopy mode functions of surgical lighting systems used in operating rooms via voice commands without requiring physical contact, performing voice processing and artificial intelligence-based command recognition locally on the embedded hardware associated with the surgical lighting system.

Description

EMBEDDED AI-BASED FOR SURGICAL LIGHTING SYSTEMS CONTACTLESS VOICE CONTROL MODULE TECHNICAL FIELD The invention transforms the lighting systems used in surgical lighting in operating rooms. physical functions of intensity, focus area, spot width, color temperature and endoscopy mode voice control that allows control via voice commands without requiring physical contact. processing and artificial intelligence-based command recognition operations with surgical lighting system a contactless voice control performed locally on the associated embedded hardware It is related to the module. PREVIOUS TECHNIQUE Surgical lighting systems are designed according to the type of operation and the different surgical procedures. depending on the stages, light intensity, focus area, spot width, color temperature and There is a need to change lighting parameters such as endoscopy mode. It is heard. In current practices, these settings are mostly sterile handles, Control units located on the surgical lamp body or touch control This is accomplished using control panels. In this type of control method, the user a control surface of the surgical lighting system during the surgical procedure and physical Interaction is required. The user needs to change the lighting parameters. for the purpose of interrupting his surgical practice and moving to a control unit, This can cause disruption to the workflow. Physical contact The control structures required are also important in terms of maintaining sterile working conditions. Additional attention is required. Gesture recognition in some systems aims to reduce the need for physical contact. These methods utilize either general-purpose speech recognition solutions. However, the operating room environment has different acoustics than general use environments. The conditions are such that the use of a surgical mask prevents the acoustics of speech. You can change its features, from extractor fans and monitor alarms to air conditioning. sounds originating from systems and multiple people talking in the same room This can make it difficult to understand voice commands. General-purpose voice recognition systems are relevant in the operating room environment. noise conditions, masked speech, Turkish surgical command structure, and embedded equipment. 1 Not developed taking into account both processing capacity and memory limitations. This is possible. Therefore, general purpose solutions should not be used for direct surgical illumination. implementation of voice commands into systems, accurate perception of voice commands and specified the command is safely transferred to the appropriate function of the surgical lighting system. It may not be sufficient in terms of quality. In addition, closed source third parties The use of software, the artificial intelligence model used in surgical lighting because it cannot be retrained with application-specific data, and the source code is inaccessible, licensing costs, the privacy of audio data, and the long-term nature of the software infrastructure used. This can lead to limitations regarding its long-term sustainability. Therefore, surgery Enables lighting systems to be controlled without requiring physical contact. Adapted to speaking conditions with and without masks, in an operating room environment. Taking noise conditions into account, and capable of processing Turkish surgical lighting commands, embedded without the need for an external voice recognition service or internet connection. Surgical lighting can operate locally on the hardware and execute detected commands. a control system is needed that can transfer the functions of the system in a controlled manner. It is located. THE PURPOSE OF THE INVENTION The purpose of the invention is to provide lighting used in surgical lighting systems. Its functions can be performed audibly by surgeons or other users without requiring physical contact. an embedded AI-based controller that allows it to be controlled via commands It is the implementation of the module. Another purpose of the invention is to control the voice function with a predetermined wake-up call. By activating it depending on the perception of the expression, the ambient conversations are directly The goal is to provide a control structure that prevents the device from converting the command into a signal. Another aim of the invention is to reduce noise and clip silence from the received audio data, after being subjected to sampling standardization and feature extraction processes The goal is to enable evaluation by an artificial intelligence-based command recognition model. Another objective of the invention is to obtain results under masked and unmasked speech conditions. domain-specific Turkish voice data and noise representing the operating room environment. A command recognition model trained using these conditions for surgical illumination The aim is to ensure that it is used under the control of the system. Another purpose of the invention is to determine which device function corresponds to the voice command. After its determination, the command in question is intended as a device control command. 2 conversion and specified device control intent only corresponding surgical The goal is to ensure it is matched to the lighting function. Another aim of the invention is to prevent commands that carry a risk of false triggers during surgery. It undergoes an additional verification process before being transferred to the lighting system. by ensuring that ambient conversations, similar expressions or noise are avoided The aim is to minimize potential unwanted device operations. Another aim of the invention is to improve speech processing, command recognition, intention determination, and commands. decision-making processes are performed locally on the embedded AI processing unit. Thanks to this, the system can operate without needing an internet connection and externally. The goal is to enable it to function without relying on a voice recognition service. Another aim of the invention is to increase the light intensity of surgical lighting systems, to improve light quality. Intensity reduction, focus adjustment, spot width adjustment, color temperature change, and endoscopy. The goal is to enable the control of mode functions via voice commands. DETAILED DESCRIPTION OF THE INVENTION The invention enables the control of surgical lighting systems without requiring physical contact. connecting a surgical lighting system to its control infrastructure, enabling its operation. It is an embedded AI-based voice control module. This module... Basically, it's a sound pickup unit, a sound preprocessing unit, a wake-up detection unit, and an artificial intelligence unit. intelligence-based command recognition model, intent determination unit, secure command mapping unit, critical command verification unit, embedded artificial intelligence processing unit, and surgical lighting. It includes the control interface. The sound receiving unit used within the scope of the invention captures the user's voice and sound. By detecting the voice commands contained within it, the system processes the resulting voice data. It transfers the audio to the audio preprocessing unit for processing. The audio obtained from the audio receiving unit... The voice data is not used directly in controlling device function and is primarily used in voice pre-processing. It is processed by the processing unit. In one application of the invention, the sound pickup unit is a digital MEMS microphone. The detected audio signal is transmitted over the I2S bus to an embedded AI processing unit. It transmits and buffers audio samples at a sampling frequency of 16 kHz and a resolution of 16 bits. is written to memory. The audio preprocessing unit adjusts the amplitude of the buffered samples. It normalizes and divides the audio signal into overlapping short time frames. Speech initiation is determined by applying the sound activity detection process on the frames. and endpoints are identified, non-dialogue sections are trimmed, and 3 Spectral analysis using the noise spectrum extracted from frames without speech. Filtering is performed. The resulting instruction segment is selected according to the model's input time. It is being cut or completed with zero samples. Segment windowing and fast The Fourier transform is applied, and then the log-transform is processed through the Mel filter bank. A Mel spectrogram is being generated. Alternatively, discrete values ​​of log-Mel coefficients can be used. MFCC coefficients can also be obtained by applying the cosine transform. These processes... The result is an audio data feature matrix with fixed time and frequency dimensions. transformed into a wake-up detection unit and a command recognition model. is being transferred. The feature matrix obtained from the audio preprocessing unit is used by the wake-up detection unit. the environment containing positive sound samples related to the expression of awakening Its speech was trained using negative voice samples containing similar phrases and noise. It is processed in the form of sliding time windows by a binary classification model. In one application, this classification model is called log-Mel or MFCC. It is in a DS-CNN or compact CRNN structure that processes its features. The model is always A normalized confidence score for the wake-up phrase class for the window. It produces this confidence score, which is determined using validation data and embedded. exceeding the wake-up decision threshold held in memory and the threshold overshoot being predetermined If verified within a number of consecutive time windows, the predetermined The wake-up call phrase is assumed to be present within the audio data. The input to the command recognition model is T number of time points obtained from the speech preprocessing unit. a two-dimensional feature matrix consisting of a frame and F frequency coefficients It forms this matrix. In one application, this matrix performs separate filtering on each channel. 1×1 that combines inter-channel information with in-depth convolutional layers that perform this function. It is processed in DS-CNN blocks consisting of point convolution layers. Another The emergence of convolutional layers in the application form, GRU combining temporal relationships Alternatively, it is fed to the LSTM layer to create a compact CRNN. ​​The output of the model... layer; increasing light intensity, decreasing light intensity, changing focus adjustment, spotlight Change width, change color temperature, and control endoscopy mode. Normalized confidence scores for classes and for non-command or unknown class. It produces. The class with the highest score is determined using validation data. and is recognized as a command when it exceeds the class-specific decision threshold held in memory. This is done. If the decision threshold is not exceeded, the output is either out of command or unknown. 4 It is marked as accepted and the device verification process is not initiated. Identity and trust score are transferred to the intent assessment unit. Domain-specific Turkish language skills in training an AI-based command recognition model. The audio dataset is used. This audio dataset includes both masked and unmasked speech. It includes voice data obtained under speech conditions. Training data. Noise conditions representing an operating room environment are included, thus enabling the model to... The use of surgical masks and different sound sources in the operating room environment being able to distinguish surgical lighting commands under the given conditions that is intended. The lightweight structure of the command recognition model, designed to run on embedded hardware; The use of deeply separable convolutions instead of standard multichannel convolutions, The selection of the number of layers and channels based on computational load and memory limits, model from 32-bit floating-point representation of weights and intermediate activations to 8-bit integer quantization of the representation and weights whose contribution is below the determined threshold value or is created by pruning the channels. The trained model is in ONNX format. from there to a suitable inference form for the microcontroller, for example TensorFlow Lite It is converted to a binary model compatible with Micro or STM32Cube.AI. Model weights The intermediate activation used during extraction is stored in external QSPI Flash memory. Buffers are statically allocated on SRAM and between successive layers. It is being reused. In one application, the model uses a 216 MHz Arm Cortex-M7. It will run offline on the STM32F767VIT6 microcontroller, which has its core. The total model storage footprint is 20 MB, and the time from command completion to device control... The time until the decision is made shall not exceed 500 milliseconds. It is being structured. The output from the command recognition model should include at least one class ID and a confidence score. and as a command log including the detection time to the intent determination unit is being transmitted. The intention setting unit primarily determines whether the wake-up session is active or not. that it is not, the time window defined after the wake-up message of the command log whether it occurs within it and whether the confidence score exceeds the decision threshold for the relevant class. It checks whether it has passed. Failure to meet any of these conditions In this case, a "no transaction" result is generated. If the conditions are met. The class ID is searched for in the authorized instruction table stored in embedded memory, and the corresponding value is found. The incoming function code consists of the operation type and, if applicable, the target value or increment / decrement step. A device control intent is being created. In this data structure, the function code includes light intensity. Focus, spot width, temperature, or endoscopy mode; the type of procedure is increased, It indicates the process of reducing, adjusting, enabling, or disabling. The generated device control intent is secure without being used as a physical control signal. The command is transferred to the command mapping unit. Commands that do not have a corresponding entry in the authorized command table are also transferred. Model outputs are not transmitted to the surgical lighting control interface. Device control intent, generated by the intent determination unit, secure command. The secure command is transferred to the mapping unit. The secure command mapping unit controls the specified device. its intention with the corresponding function from the previously defined surgical illumination functions It relates to this. In this context, the intention regarding increasing light intensity is light intensity. the function of increasing light intensity, the intention of decreasing light intensity, the function of decreasing light intensity, The intention regarding focus change is related to the focus adjustment function, and the intention regarding changing the spot width is related to... intention spot width adjustment function, intention color regarding color temperature change intention regarding temperature function and endoscopy mode endoscopy mode function It is being matched. Within the scope of the invention, each device in the authorized command panel has the intention of being controlled, A verification flag is held to indicate whether the intention in question is critical. An intent-safe command mapping that provides a non-critical and relevant trust threshold. The situation arises when an intention is created with a critical flag activated, as the process moves to the next stage. The machine enters the "awaiting approval" state. During this stage, the candidate device is checked. The intention is stored in temporary memory, with a timer for a predetermined approval period. is being initiated and any control data is being sent to the surgical lighting control interface. It is not being sent. The predefined approval statement within the approval period. detection or recognition of the same command a second time and above the relevant confidence threshold In this case, the verification result is compared with the candidate intent in temporary memory. If a match is found, the state machine switches to the "application" state and The verified device control intent is transmitted to the surgical lighting control interface. Timeout occurred, a different command was detected, or the confidence score fell below the threshold. If the candidate remains or the wake-up session ends, the intention is cancelled. And the system returns to "listening" mode. Thus, a single classification result... It prevents a critical device function from being directly triggered. Secure instruction mapping and, when necessary, critical instruction verification. The control decision made after completion is surgical illumination control. It is transferred to the interface. The surgical lighting control interface is AI-based. Surgical lighting device control decision created in the voice control module 6 It transmits the information to the system's control infrastructure. Thus, as a result of the voice command... The specified function is implemented on the surgical lighting system. Voice preprocessing, wake-up call detection, command recognition, device procedures for determining the intention to control and forming the control decision. This is implemented locally on the embedded AI processing unit. This structure thanks to which voice data can be sent to an external server or an external audio system for command recognition purposes. There is no need to send it to the recognition service and voice control processes. This is done independently of an internet connection. In the invention's operating principle, the user first enters the wake-up phrase. This statement is detected by the speech acquisition unit and processed through speech preprocessing. After being processed in the unit, it is evaluated by the wake-up detection unit. Surgery performed by the user after the awakening expression is recognized. The lighting command is received and pre-processed. Pre-processed audio data. The relevant surgical lighting function is evaluated by the command recognition model. is determined. The determined command is assigned to the device control intention in the intention determination unit. The surgical lighting system is being converted and integrated into the secure command mapping unit. It is associated with its corresponding function. The command carries a risk of critical or false triggering. If there is a command carrying a critical command, it is added by the command verification unit. A check is performed. After it is determined that the command is executable... The relevant control decision regarding surgical lighting is made via the surgical lighting control interface. It is transferred to the system. Thanks to this working structure, the surgical lighting system as a general-purpose and free speech-evaluating voice assistant not working, activated by a predetermined wake-up phrase and surgical Limited task area for evaluating device control commands specific to lighting systems It operates as an embedded control system. Thus, wake-up detection, Domain-specific command recognition, intent determination, secure matching, and critical commands when needed. By applying verification processes sequentially, a sound is detected directly by the device. This prevents it from turning into a matter of control. 7

Claims

EMBEDDED AI-BASED FOR I CONTACTLESS VOICE CONTROL MODULE The invention transforms the lighting systems used in surgical lighting in operating rooms. physical functions of intensity, focus area, spot width, color temperature and endoscopy mode voice control that allows control via voice commands without requiring physical contact. processing and artificial intelligence-based command recognition operations with surgical lighting system a contactless voice control performed locally on the associated embedded hardware It is related to the module. 17 REQUESTS 1. Surgical lighting systems with voice commands that do not require physical contact. a surgical device structured to allow control through The embedded artificial light system is connected to the control infrastructure of the lighting system. It is an intelligence-based voice control module, the feature of which is; - at least one audio receiver that detects user-generated audio data. unit, - normalizes the audio data received from the audio pickup unit, time dividing into frames, identifying sections containing speech, noise which applies a reduction process and provides a fixed time and data from the audio data in question. at least one audio prefix that forms a feature matrix with frequency dimensions processing unit, - the feature matrix obtained from the audio preprocessor, into the wake-up phrase. ambient speech with positive sound samples, similar expressions and noise. a binary classification system trained using negative sound samples each of which operates in the form of sliding time windows through the model a normalized wake-up expression class for the time window confidence score is created and verification data of that confidence score a wake-up decision determined using and stored in embedded memory exceeding the threshold, and the number of such threshold exceedances determined in advance. audio data if verified within a consecutive time window that it contains a predetermined wake-up phrase by identifying at least one that enables the command evaluation process. wake-up detection unit, - by processing the feature matrix received from the audio preprocessing unit, at least one class ID relating to surgical lighting function and a class ID belonging to this class ID at least one AI-based command recognition that generates a trust score model, - the class ID obtained from the command recognition model, of the wake-up session the status, predetermined following the wake-up phrase of the command its occurrence within the time window and the confidence score belonging to the relevant class 8 a device control by evaluating it according to the conditions of exceeding the decision threshold. at least one intention-setting unit that transforms it into an intention, - the device in question controls the intended surgical lighting as defined in the pre-defined guidelines. a command that associates one of the functions with the corresponding function and is permitted surgical device control intentions not found in the table at least one safe command that prevents it from being transferred to the lighting system. the mapping unit, - a device control intent previously defined as critical directly by preventing its implementation, the intention to control the device in question at least one that performs an additional verification process before implementation critical command verification unit, - voice preprocessing, wake-up detection, command recognition, intent determination, Secure command mapping and critical command verification processes are performed locally. at least one embedded artificial intelligence processing unit that performs this function, and - surgical control decision corresponding to verified device control intention at least one surgical device that transmits data to the lighting system's control infrastructure. It is characterized by including a lighting control interface.

2. It is an embedded AI-based voice control module according to claim 1, and its feature is; containing a digital MEMS microphone and by that MEMS microphone The detected audio signal is processed by embedded AI via the I2S bus. by including an audio receiving unit configured to transmit to the unit It is characteristic.

3. An embedded AI-based voice control module according to claim 1 or 2, Its feature is that it processes audio samples at a sampling frequency of 16 kHz and a resolution of 16 bits. It includes an audio preprocessing unit configured to write to a buffer memory. It is characteristic.

4. Embedded AI-based voice control according to any of the previous requests. It is a module whose function is to normalize the amplitude of buffered audio samples. to divide the sound samples into overlapping short time frames, sound Identifying the start and end points of the conversation through the process of determining the activity. to cut out sections that do not contain dialogue and sections without dialogue Spectral filtering using the noise spectrum extracted from the frames by including an audio preprocessing unit configured to perform this task. It is characteristic. 9 5. Embedded AI-based voice control according to any of the previous requests. It is a module whose feature is to interrupt the command segment according to the model's input time or completing with zero instances, windowing and fast Applying the Fourier transform and log-Mel through the Mel filter bank. by including an audio preprocessing unit configured to generate the spectrogram It is characteristic.

6. It is an embedded AI-based voice control module according to claim 5, and its feature is; MFCC coefficients by applying discrete cosine transform to log-Mel coefficients It is characterized by containing an audio preprocessing unit structured to create sound.

7. Embedded AI-based voice control according to any of the previous requests. It is a module whose characteristics include T time frames and F frequency coefficients. The resulting two-dimensional feature matrix is ​​filtered separately on each channel. inter-channel information through in-depth convolutional layers via DS-CNN blocks containing connecting 1×1 point convolution layers an AI-based command recognition model structured for processing It is characterized by its inclusion.

8. Embedded AI-based voice control according to any of the previous requests. It is a module whose feature is the temporal relationships of the output of the convolution layers. to process through at least one GRU or LSTM layer that integrates AI-based command recognition model with structured CRNN architecture It is characterized by its inclusion.

9. Embedded AI-based voice control according to any of the previous requests. It is a module whose features include increasing light intensity, decreasing light intensity, and adjusting focus. changing, changing spot width, changing color temperature, and endoscopy normalize the mode for non-command or unknown class with control classes AI-based commands structured to generate established confidence scores It is characterized by its inclusion of a recognition model.

10. According to claim 9, it is an embedded AI-based voice control module, the feature of which is; The class with the highest confidence score, determined using validation data, and The recognized command is triggered when the class-specific decision threshold held in memory is exceeded. to accept as such and the output if the decision threshold in question is not exceeded AI configured to mark as unauthorized or unknown. It is characterized by its inclusion of a command recognition model based on the command recognition model.

11. Embedded AI-based voice control according to any of the previous requests. It is a module whose feature is that it can be used in both masked and unmasked speech conditions. audio data and noise conditions representing the operating room environment AI-based system trained using a domain-specific Turkish voice dataset. It is characterized by its inclusion of a command recognition model.

12. Embedded AI-based voice control according to any of the previous requests. It is a module whose feature is that it uses in-depth processing instead of standard multi-channel convolution. separable convolutions and layer and channel numbers affect processing load and memory. by including an AI-based command recognition model defined according to its limitations It is characteristic.

13. Embedded AI-based voice control according to any of the previous requests. It is a module whose feature is that model weights and intermediate activations are 32-bit floating point. quantized AI-based command from representation to 8-bit integer representation It is characterized by its inclusion of a recognition model.

14. Embedded AI-based voice control according to any of the previous requests. It is a module whose characteristic is that its contribution remains below a predetermined threshold value. AI-based command recognition model with trimmed weights or channels It is characterized by its inclusion.

15. Embedded AI-based voice control according to any of the previous requests. It is a module whose feature is that it converts ONNX format to TensorFlow Lite Micro or Artificial intelligence converted into STM32Cube.AI compatible microcontroller inference format It is characterized by its inclusion of an intelligence-based command recognition model.

16. Embedded AI-based voice according to any of the previous requests. It is a control module whose feature is that it stores model weights in external QSPI Flash memory. SRAM intermediate activation buffers used during retention and inference to allocate statically and reuse between successive layers It is characterized by containing an embedded artificial intelligence processing unit structured accordingly.

17. Embedded AI-based voice control according to any of the previous requests. It is a module, and its feature is; a total model storage footprint of 20 MB and the command The time elapsed from completion to the formation of the device control decision is 500. AI-based commands structured to not exceed milliseconds It is characterized by its inclusion of a recognition model.

18. Embedded AI-based voice control according to any of the previous requests. It is a module whose feature is to indicate whether the wake-up session is active, and the command... 11 within the time window defined after the wake-up message of the recording whether it has occurred and whether the confidence score has passed the decision threshold for the relevant class. to check that it has not been met and that none of the conditions in question have been met It is configured to generate a "no transaction" result if it is not provided. It is characterized by including a unit for expressing intent.

19. It is an embedded AI-based voice control module according to claim 18, and its feature is; Searching for the class ID in the permissioned instruction table stored in embedded memory and corresponding incoming function code, operation type, and if applicable, target value or increment / decrement. The intention is structured to create the device control intention, consisting of several steps. It is characterized by containing a unit of determination.

20. It is an embedded AI-based voice control module according to Claim 19, and its feature is; function code for light intensity, focus, spot width, color temperature or endoscopy with one of the modes and the type of operation to increase, decrease, adjust, enable or configured to be associated with one of the disable operations It is characterized by including a unit for expressing intent.

21. An embedded AI-based voice control module according to claim 19 or 20, Its feature allows surgical model outputs that do not have a corresponding entry in the permissioned command table. configured to prevent transmission to the lighting control interface. It is characterized by including a unit for expressing intent.

22. Embedded AI-based voice control according to any of the previous requests. It is a module whose feature is that the device controls the intended surgical procedure in a predefined manner. to associate the lighting functions with the corresponding function It is characterized by containing a structured secure instruction mapping unit.

23. Embedded AI-based voice control according to any of the previous requests. It is a module whose feature is that each device in the authorized command table has a control intention. a verification that shows whether the intention to control the device in question is critical Embedded memory containing a permissions instruction table configured to hold the flag. It is characterized by its inclusion.

24. According to claim 23, it is an embedded artificial intelligence-based voice control module, the feature of which is; A device with the critical flag enabled is awaiting approval when a control intent is created. to switch to this state, temporarily save the candidate device control intent to memory, and configured to start a timer for a predetermined approval period. It is characterized by containing a critical command verification unit. 12 25. According to claim 24, it is an embedded artificial intelligence-based voice control module, the feature of which is; detection of the predefined approval statement within the approval period or if the same command is recognized a second time above the relevant confidence threshold The verification result is stored in temporary memory along with the candidate device control intent. Compare and verify the device if a match is found. configured to transmit its intent to the surgical lighting control interface. It is characterized by containing a critical command verification unit.

26. An embedded AI-based voice control module according to claim 24 or 25, Features include: timeout, detection of a different command, and trust score. if it falls below the threshold or the wake-up session ends To delete the candidate device control intent stored in temporary memory and to listen to the system. Critical command verification unit configured to return to its current state. It is characterized by its inclusion.

27. Embedded AI-based voice control according to any of the previous requests. It is a module whose features include: voice preprocessing, wake-up call detection, and command recognition. recognition, determination of device control intent, and formation of control decision related operations to transmitting voice data to an external server or external voice recognition system. to perform locally and offline without sending it to the service It is characterized by containing a structured embedded artificial intelligence processing unit.

28. Surgical lighting system with voice commands that do not require physical contact. It is a method for controlling it through, and its characteristic feature is; - user-generated audio data via the audio receiving unit perception, - The detected audio data is processed by the audio preprocessing unit at a fixed time and Converting it into a feature matrix with frequency dimensions, - the feature matrix obtained from the audio preprocessing unit, to the wake-up phrase ambient speech with positive sound samples, similar expressions and noise. a binary classification system trained using negative sound samples processing by the model in the form of sliding time windows, each a normalized wake-up expression class for the time window the creation of a trust score, and the verification of that trust score. a wake-up signal determined using data and stored in embedded memory Determining whether the decision threshold has been exceeded and anticipating the threshold breach beforehand. if verified within a specified number of consecutive time windows 13 the predefined wake-up phrase within the voice data acceptance of its existence, - If the wake-up call is detected, follow the wake-up call. the voice command is recognized by an AI-based command recognition model by classifying and creating at least one class identity and trust score, - the class identity in question is determined by the intent unit through device control. transforming it into its intention, - the device control intent is pre-determined by the secure command mapping unit. with the corresponding function from the defined surgical illumination functions matching, - Critical command verification of the critically identified device control intent. subject to further verification by the unit and applicable the control decision corresponding to the intention to control the device that was determined surgical lighting via surgical lighting control interface It is characterized by including the steps involved in transferring it to the system.

29. The method according to claim 28, its characteristic is; audio preprocessing of perceived audio data. a feature matrix with fixed time and frequency dimensions by the unit During the conversion process, the amplitude of the audio data is normalized, word The subject is the decomposition of audio data into overlapping short time frames, audio activity. Determining the start and end points of the conversation through a process of identification. trimming of non-dialogue sections and sections without dialogue Spectral filtering using the noise spectrum extracted from the frames It is characterized by including the steps involved in its implementation.

30. The method is according to claim 28 or 29, and its characteristic is that the instruction segment is the model's input. interruption according to duration or completion with zero instances, the command in question Segment windowing and fast Fourier transform application and Mel filter. generating a log-Mel spectrogram via a database or log-Mel The MFCC coefficients are obtained by applying a discrete cosine transform to the coefficients. It is characterized by including the steps involved in its implementation.

31. A method according to either of claims 28-30, characterized by its inclusion of audio data. wake-up call to check if a pre-determined wake-up call phrase is present During the stage of determination by the sensing unit, the wake-up call If not detected, active information regarding the functions of the surgical lighting system. failure to initiate the command evaluation phase and the wake-up call 14 If detected, the artificial voice command that follows the wake-up phrase It is characterized by its transition to an intelligence-based command recognition model.

32. A method according to either of claims 28-31, characterized by its function being the expression of awakening. the subsequent voice command is recognized by an AI-based command recognition model by classifying and creating at least one class identity and trust score At this stage, the class with the highest confidence score is specific to that class. If it exceeds the decision threshold, it is accepted as a recognized command and the decision If the threshold is not exceeded, the output will be either unintentional or unknown. It is characterized by the device control process not being initiated by marking it.

33. The method according to either of claims 28-32, its characteristic being; the intention of class identity. During the stage where the device control intent is converted by the determination unit, Checking whether the wake-up session is active, the wake-up command log. it occurs within the time window defined after the expression whether it has not occurred and whether the confidence score exceeds the decision threshold for the relevant class. monitoring, failure to meet any of the aforementioned conditions in this case, the result is "no action" and the conditions are met. In this case, the class ID is searched in the authorized command panel and the corresponding device is found. It is characterized by the formation of an intention to control.

34. This method, according to claim 33, has the characteristic of being: the class ID in the allowed instruction table. During the search phase, a model was found that did not have a counterpart in the authorized command table. no device control intent is created for the outputs and the model in question It is characterized by the failure to transmit the outputs to the surgical lighting control interface.

35. The method is based on either of claims 28-34 and its characteristic is that it is critical. The specified device control intent is added by the critical command verification unit. During the verification process, the device in question is checked. Saving your intention to temporary memory puts the system into a pending approval state. and within the predetermined approval period, the predefined the detection of the confirmation message or the second time the same command is given, the relevant confidence threshold It is characterized by the expectation of being recognized.

36. The method is based on Claim 35 and its characteristic feature is the control of devices identified as critical. During the stage when the intent is subject to further verification, the approval period obtained from the perceived confirmation message or the same command recognized a second time the obtained verification result is stored in temporary memory along with the candidate device control intent. The device is checked for verification if a match is found after comparison. It is characterized by the transmission of the intention to the surgical lighting control interface.

37. The method is in accordance with claim 35 or 36, and its characteristic is that the device is designated as critical. During the stage where the control intent is subjected to additional verification, time exceeding the threshold, different command being detected, confidence score below the threshold at least one of the following situations: remaining or the wake-up session ending If this happens, the candidate device control intent stored in temporary memory It is characterized by being deleted and the system returning to listening mode.

38. A method according to either of claims 28-37, characterized by: perceived sound. preprocessing of the data, determining the wake-up phrase, voice command classification, conversion of class identity to device control intent, device Matching the control intent with the surgical illumination function, critical device control. embedded in the stages of verifying intent and forming a control decision On the artificial intelligence processing unit, voice data is sent to an external server or an external device. Locally and offline without being sent to a voice recognition service. It is characterized by its implementation. 16