Embedded voice recognition control module and equipment interaction system thereof

Abstract

The invention provides an embedded voice recognition control module and an equipment interaction system thereof. The embedded voice recognition control module comprises a voice acquisition sub-module, a voice preprocessing sub-module, a voice recognition and command analysis sub-module, a control command generation sub-module, an actuator and feedback sub-module and a self-adaptive learning and personalization sub-module; a deep learning speech recognition technology and a natural language processing technology are integrated in the small embedded speech recognition control module, so that speech control instruction recognition and feedback can be completed with low power consumption, dependence of speech control on a network is reduced, the response speed is increased, and the speech control efficiency is improved. And moreover, the privacy of the user can be better protected, and the deep learning speech recognition model can be adaptively trained based on the character use habit of the user and the accent of the user, so that the speech recognition accuracy is continuously improved.

Description

Technical Field

[0001] The present invention relates to the field of speech recognition and device control, and in particular to an embedded speech recognition control module and a device interaction system thereof. Background Art

[0002] With the development of the Internet of Things, people have higher and higher requirements for the intelligence and miniaturization of devices. In many scenarios, people's hands are occupied, such as when driving, cooking, and holding things in their hands. Voice control is extremely important. People hope to complete various tasks through voice commands, such as querying information, controlling devices, etc., without having to put down the items in their hands to operate manually.

[0003] Compared with manually operated devices, voice control can complete tasks more quickly. Traditional voice recognition systems usually rely on cloud servers for calculations, which have problems such as network delay, privacy security, high power consumption, and complex interaction with devices. Therefore, the present invention proposes an embedded voice recognition control module and its device interaction system. Summary of the invention

[0004] The present invention provides an embedded voice recognition control module and a device interaction system thereof, which integrates deep learning voice recognition technology and natural language processing technology in a compact embedded voice recognition control module, so that voice control instruction recognition and feedback can be completed with only low power consumption, while also reducing the dependence of voice control on the network, improving the response speed, and being able to better protect the privacy of users.

[0005] The present invention provides an embedded speech recognition control module, characterized in that it comprises:

[0006] A voice collection submodule consisting of one or more microphone arrays, used to collect the user's voice signal;

[0007] The speech preprocessing submodule is used to preprocess the speech signal to obtain a high-quality speech signal;

[0008] The speech recognition and command parsing submodule is used to convert high-quality speech signals into text data based on a deep learning speech recognition model, and to perform semantic parsing of the text data through natural language processing technology to determine user control commands;

[0009] A control instruction generation submodule is used to generate corresponding device control instructions based on user control commands and send them to lower-level devices;

[0010] The actuator and feedback submodule is used to generate corresponding feedback voice based on the control instruction execution result of the actuator of the lower device;

[0011] The adaptive learning and personalization submodule is used to adaptively train the deep learning speech recognition model based on the user's text usage habits and accent.

[0012] Preferably, in an embedded speech recognition control module, the speech collection submodule includes:

[0013] A direction sensing unit is used to determine the source direction of each sound source signal based on the time difference and phase difference of the sound signals received by each microphone in the microphone array, and to determine the effective coverage range corresponding to each channel signal according to the size of the different channel signals received by each microphone;

[0014] A sound data analysis unit, used to determine an effective coverage area corresponding to a human voice signal as a target effective coverage range, and confirm a target microphone array area corresponding to the target effective coverage range;

[0015] Acquire a first target human voice signal collected by each microphone in each target microphone array area, perform target sound fusion based on the first target human voice signal, obtain a second target human voice signal, and send the second target human voice signal as the collected voice signal to the voice preprocessing submodule;

[0016] A sensitivity control unit, used to compare the multiple first target human voice signals contained in different second target human voice signals with preset thresholds respectively, to obtain an acquisition error rate, and based on the acquisition error rate and in combination with the position information of the person generating the sound, to control the sensitivity of the microphone corresponding to each target human voice signal respectively;

[0017] When any microphone is within the target microphone array area corresponding to two or more channel signals, the final sensitivity of the microphone is based on the sensitivity corresponding to the channel signal with greater sensitivity.

[0018] Preferably, in an embedded speech recognition control module, the speech preprocessing submodule includes:

[0019] A first preprocessing unit, used for performing preliminary processing on the speech signal based on Gaussian filtering and normalized least mean square algorithm to obtain a denoised speech signal;

[0020] The second preprocessing unit is used to obtain the volume characteristics corresponding to each denoised speech signal, and perform volume normalization processing on the denoised speech signal based on the volume characteristics to obtain a high-quality speech signal.

[0021] Preferably, in an embedded speech recognition control module, the sensitivity control unit includes:

[0022] The first intelligent control subunit is used to compare the total coverage area of ​​all target effective coverage areas with the full coverage area of ​​the microphone array, determine the invalid vocal area, and adjust the sensitivity value of the microphone corresponding to the invalid vocal area to a minimum threshold;

[0023] The speaker localization subunit is used to obtain the distribution characteristics of the microphones in the target microphone array area corresponding to each human voice signal, and determine the relative distance between the speaker and the embedded speech recognition control module based on the distribution characteristics and the sound propagation characteristics;

[0024] The second intelligent control subunit is used to compare the relative distance with a preset standard value to obtain a position difference, and to compare a plurality of first target human voice signals contained in different second target human voice signals with preset thresholds to obtain an acquisition error rate;

[0025] According to the positive and negative values ​​corresponding to the position difference, the sensitivity adjustment direction of the microphone and the corresponding adjustment coefficient are determined, and based on the acquisition error and the standard sensitivity of the microphone, the sensitivity adjustment amplitude of the microphone is obtained;

[0026] According to the sensitivity adjustment amplitude and the adjustment coefficient corresponding to the sensitivity adjustment direction, an actual adjustment value is obtained, and based on the actual adjustment value and according to the sensitivity adjustment direction corresponding to the microphone, the sensitivity of the target microphone is adjusted.

[0027] Preferably, in an embedded speech recognition control module, the speech recognition and command parsing submodule includes:

[0028] Model recognition unit, used to recognize and convert high-quality voice signals into text data based on deep learning voice recognition model, and to perform semantic analysis on text data based on natural language technology to determine user semantics and user intentions;

[0029] A command screening unit, which is used to screen high-quality voice signals based on user semantics and user intent in combination with a preset control command data set to obtain effective control voice;

[0030] The command determination unit is used to determine the user control command according to the correspondence between the effective control voice and the control vocabulary in the preset control instruction data set and the control instruction corresponding to the control vocabulary.

[0031] Preferably, in an embedded speech recognition control module, the control instruction generation submodule includes:

[0032] An instruction generation unit, used to generate corresponding control instructions based on user commands and in combination with the communication protocol between the embedded speech recognition control module and the lower device;

[0033] The instruction sending unit is used to send control instructions to the lower-level device through a preset communication channel.

[0034] Preferably, in an embedded speech recognition control module, the adaptive learning and personalization submodule includes:

[0035] A user data processing unit is used to obtain all valid historical wake-up records of the voice recognition control module, and to define a wake-up cycle from the initial wake-up to the successful recognition of the control command by the voice recognition control module, and to divide the voice recognition control module into multiple successful control wake-up cycles to obtain multiple successful control wake-up cycle data sets;

[0036] Respectively obtain invalid voice contents of the user in a plurality of successful control wake-up cycles, and when the number of the invalid voice contents is zero, remove the successful control wake-up cycle;

[0037] A data identification unit is used to generate a corresponding instruction tag based on the successful control instruction corresponding to the successful voice wake-up cycle and add it to the successful control wake-up cycle data set, and monitor invalid voices, or true invalid voices and pseudo invalid voices based on the keywords of the successful control instruction corresponding to the successful control wake-up cycle data set;

[0038] A usage habit analysis unit is used to respectively obtain the unique voice features corresponding to each pseudo invalid voice, compare the unique voice features with the voice features of the standard control instruction corresponding to the control command, and obtain the accent difference between the user's actual voice and the pre-selected voice and the user's text usage habits;

[0039] Based on the voiceprint features of the voice signals in different successful control wake-up cycles, the successful control wake-up cycles are clustered to obtain the voice control data clusters corresponding to the same user;

[0040] Compare the accent differences corresponding to the pseudo-invalid voices in the same voice control data cluster and the user's text usage habits to obtain the current user's personalized accent and usage habits, and compare the personalized accents and usage habits between different users to obtain the user voice similarity;

[0041] A data merging unit, used to merge the voice control data clusters whose user voice similarity is greater than a preset value to obtain a merged data set, and generate different batches of training sets according to the merged data set and the unmerged voice control data clusters;

[0042] The adaptive training unit is used to input the corresponding training set into the deep learning speech recognition model for adaptive training according to the training set batches, until all batches of training are completed and the adaptive training is ended.

[0043] Preferably, in an embedded speech recognition control module, the user data processing unit includes:

[0044] The voiceprint file creation subunit is used to create a voiceprint file based on the voiceprint features corresponding to the voice signal. When a new voiceprint feature is detected, a new voiceprint file is automatically generated;

[0045] When confirming that the current user's voice is a voiceprint object, determine whether the current user's voice control command is a single feedback command. If so, add an invalid label to the historical wake-up record of the current user;

[0046] Otherwise, add a valid tag to the historical wakeup record of the current user.

[0047] Preferably, in an embedded speech recognition control module, the user data processing unit further includes:

[0048] The intelligent comparison subunit is used to compare the current voice segment collected by the profiled user in the current successful control wake-up cycle with the historical voice segment in the invalid voiceprint file corresponding to the latest similar control command of the profiled user when a pseudo invalid voice appears in the voice control command recognition process of the profiled user, and determine the usage habit difference and accent difference of the profiled user according to the comparison result;

[0049] The audio corresponding to the current speech segment is highlighted according to the usage habit difference or the pronunciation difference, and a training control group is generated based on the highlighted speech segment and the historical speech segment. The training control group is used to perform adaptive training on the deep learning speech recognition model.

[0050] The present invention provides a device interaction system of an embedded voice recognition control module, characterized in that a communication connection is established between any one of the embedded voice recognition control modules and its corresponding lower-level device.

[0051] Other features and advantages of the present invention will be described in the following description, and partly become apparent from the description, or understood by practicing the present invention. The purpose and other advantages of the present invention can be realized and obtained by the structures specifically pointed out in this application document.

[0052] The technical solution of the present invention is further described in detail below through the accompanying drawings and embodiments. BRIEF DESCRIPTION OF THE DRAWINGS

[0053] The accompanying drawings are used to provide a further understanding of the present invention and constitute a part of the specification. Together with the embodiments of the present invention, they are used to explain the present invention and do not constitute a limitation of the present invention. In the accompanying drawings:

[0054] Figure 1This is a structural diagram of an embedded speech recognition control module in an embodiment of the present invention;

[0055] Figure 2 This is a structural diagram of a speech collection submodule of an embedded speech recognition control module in an embodiment of the present invention;

[0056] Figure 3 This is a structural diagram of a speech preprocessing submodule of an embedded speech recognition control module in an embodiment of the present invention;

[0057] Figure 4 A structural diagram of a speech recognition and command parsing submodule of an embedded speech recognition control module in an embodiment of the present invention;

[0058] Figure 5 A structural diagram of a control instruction generation submodule of an embedded speech recognition control module in an embodiment of the present invention;

[0059] Figure 6 It is a structural diagram of an adaptive learning and personalization submodule of an embedded speech recognition control module in an embodiment of the present invention. DETAILED DESCRIPTION

[0060] The preferred embodiments of the present invention are described below in conjunction with the accompanying drawings. It should be understood that the preferred embodiments described herein are only used to illustrate and explain the present invention, and are not used to limit the present invention.

[0061] Embodiment 1:

[0062] The present invention provides an embedded speech recognition control module, such as Figure 1 As shown, including:

[0063] A voice collection submodule consisting of one or more microphone arrays, used to collect the user's voice signal;

[0064] The speech preprocessing submodule is used to preprocess the speech signal to obtain a high-quality speech signal;

[0065] The speech recognition and command parsing submodule is used to convert high-quality speech signals into text data based on a deep learning speech recognition model, and to perform semantic parsing of the text data through natural language processing technology to determine user control commands;

[0066] A control instruction generation submodule is used to generate corresponding device control instructions based on user control commands and send them to lower-level devices;

[0067] The actuator and feedback submodule is used to generate corresponding feedback voice based on the control instruction execution result of the actuator of the lower device;

[0068] The adaptive learning and personalization submodule is used to adaptively train the deep learning speech recognition model based on the user's text usage habits and accent.

[0069] In this embodiment, the feedback voice can be a simple operation confirmation (such as "the light is turned on"), or a prompt of operation failure (such as "the curtain opening failed, please check the device status"). In this way, the interaction experience between the user and the device is enhanced, and the user can intuitively know whether the device has been operated according to his or her wishes.

[0070] In this embodiment, the lower-level devices include but are not limited to home appliances, automobiles, industrial automation equipment, and wearable devices.

[0071] The beneficial effects of the above technical solution are as follows: the present invention effectively improves the collection quality of voice signals through a voice collection submodule composed of one or more microphone arrays. The microphone array can utilize the spatial position relationship between multiple microphones to enhance the reception of sounds in a specific direction, while suppressing interference noises in other directions, effectively improving the clarity of the collected voice signals; the voice preprocessing submodule preprocesses the voice signal to obtain a high-quality voice signal, which is conducive to improving the accuracy of voice recognition; the voice recognition and command parsing submodule utilizes a deep learning voice recognition model to convert the high-quality voice signal into text data. The deep learning model has a powerful feature extraction capability and can automatically learn complex features in the voice signal. Then, the semantics of the text data can be accurately determined by natural language processing technology to understand the user's control command and the user's intention. The control instruction generation submodule can generate corresponding device control instructions based on the user control command and send them to the lower device. The precise instruction generation and transmission mechanism ensures that the user's voice instruction can be accurately converted into the actual operation of the device, improving the convenience and intelligence of the device control; the actuator and feedback submodule generates corresponding feedback voice based on the control instruction execution result of the lower device actuator, so that the user can understand the operation of the device in time. Finally, the deep learning speech recognition model is adaptively trained based on the user's text usage habits and user accents through the adaptive learning and personalized submodules, so that the speech recognition system can better adapt to the characteristics of each user. As the number of times the user uses it increases, the system will become more and more familiar with the user's accent, vocabulary usage habits, etc., thereby continuously improving the accuracy of speech recognition. The present invention integrates deep learning speech recognition technology and natural language processing technology in a compact embedded speech recognition control module, so that voice control command recognition and feedback can be completed with only low power consumption, while also reducing the dependence of voice control on the network, improving the response speed, and better protecting the privacy of users.

[0072] Embodiment 2:

[0073] Based on Example 1, the voice collection submodule, such as Figure 2 As shown, including:

[0074] A direction sensing unit is used to determine the source direction of each sound source signal based on the time difference and phase difference of the sound signals received by each microphone in the microphone array, and to determine the effective coverage range corresponding to each channel signal according to the size of the different channel signals received by each microphone;

[0075] A sound data analysis unit, used to determine an effective coverage area corresponding to a human voice signal as a target effective coverage range, and confirm a target microphone array area corresponding to the target effective coverage range;

[0076] Acquire a first target human voice signal collected by each microphone in each target microphone array area, perform target sound fusion based on the first target human voice signal, obtain a second target human voice signal, and send the second target human voice signal as the collected voice signal to the voice preprocessing submodule;

[0077] A sensitivity control unit, used to compare the multiple first target human voice signals contained in different second target human voice signals with preset thresholds respectively, to obtain an acquisition error rate, and based on the acquisition error rate and in combination with the position information of the person generating the sound, to control the sensitivity of the microphone corresponding to each target human voice signal respectively;

[0078] When any microphone is within the target microphone array area corresponding to two or more channel signals, the final sensitivity of the microphone is based on the sensitivity corresponding to the channel signal with greater sensitivity.

[0079] The beneficial effects of the above technical solution: the present invention determines the effective coverage range of each sound signal through the direction perception unit, and then determines the effective coverage range of the human voice signal through the sound data analysis unit and confirms the target microphone array area corresponding to the target effective coverage range, and at the same time, generates a second target human voice signal with higher clarity by fusing the human voice signals collected by different microphones. At the same time, in the process of human voice signal collection, the sensitivity of the microphone is adjusted according to the signal strength of the first target human voice signal collected by each microphone, thereby enhancing the microphone array's reception of sound in a specific direction and suppressing interference noise from other directions, thereby effectively improving the quality of the human voice signal and providing a basic guarantee for the accuracy of speech recognition.

[0080] Embodiment 3:

[0081] Based on Example 1, the speech preprocessing submodule, such as Figure 3 As shown, including:

[0082] A first preprocessing unit, used for performing preliminary processing on the speech signal based on Gaussian filtering and normalized least mean square algorithm to obtain a denoised speech signal;

[0083] The second preprocessing unit is used to obtain the volume characteristics corresponding to each denoised speech signal, and perform volume normalization processing on the denoised speech signal based on the volume characteristics to obtain a high-quality speech signal.

[0084] Beneficial effects of the above technical solution: The present invention performs denoising and echo elimination on the voice signal through the first preprocessing unit, effectively reducing the interference of noise on the human voice signal while also reducing the interference between human voice signals, which can effectively improve the voice recognition efficiency and accuracy of the module, thereby improving the response efficiency of user voice control. Then, the volume characteristics corresponding to each denoised voice signal are obtained through the second preprocessing unit, and the denoised voice signal is volume normalized based on the volume characteristics to eliminate the differences between the various audios and obtain a high-quality voice signal.

[0085] Embodiment 4:

[0086] Based on Example 2, the sensitivity control unit includes:

[0087] The first intelligent control subunit is used to compare the total coverage area of ​​all target effective coverage areas with the full coverage area of ​​the microphone array, determine the invalid vocal area, and adjust the sensitivity value of the microphone corresponding to the invalid vocal area to a minimum threshold;

[0088] The speaker localization subunit is used to obtain the distribution characteristics of the microphones in the target microphone array area corresponding to each human voice signal, and determine the relative distance between the speaker and the embedded speech recognition control module based on the distribution characteristics and the sound propagation characteristics;

[0089] The second intelligent control subunit is used to compare the relative distance with a preset standard value to obtain a position difference, and to compare a plurality of first target human voice signals contained in different second target human voice signals with preset thresholds to obtain an acquisition error rate;

[0090] According to the positive and negative values ​​corresponding to the position difference, the sensitivity adjustment direction of the microphone and the corresponding adjustment coefficient are determined, and based on the acquisition error and the standard sensitivity of the microphone, the sensitivity adjustment amplitude of the microphone is obtained;

[0091] According to the sensitivity adjustment amplitude and the adjustment coefficient corresponding to the sensitivity adjustment direction, an actual adjustment value is obtained, and based on the actual adjustment value and according to the sensitivity adjustment direction corresponding to the microphone, the sensitivity of the target microphone is adjusted.

[0092] Beneficial effects of the above technical solution: The present invention utilizes the distance between the speaker and the embedded book module and the actual sound receiving effect of each microphone to enhance the reception of sound in a specific direction, while suppressing interference noise from other directions, effectively improving the clarity of the collected voice signal.

[0093] Embodiment 5:

[0094] Based on Example 1, the speech recognition and command parsing submodule, such as Figure 4 As shown, including:

[0095] Model recognition unit, used to recognize and convert high-quality voice signals into text data based on deep learning voice recognition model, and to perform semantic analysis on text data based on natural language technology to determine user semantics and user intentions;

[0096] A command screening unit, which is used to screen high-quality voice signals based on user semantics and user intent in combination with a preset control command data set to obtain effective control voice;

[0097] The command determination unit is used to determine the user control command according to the correspondence between the effective control voice and the control vocabulary in the preset control instruction data set and the control instruction corresponding to the control vocabulary.

[0098] Beneficial effects of the above technical solution: The present invention uses a deep learning speech recognition model to convert high-quality speech signals into text data. The deep learning model has a powerful feature extraction capability and can automatically learn complex features in speech signals, such as speech patterns under different accents, speaking speeds, and intonations. Then, the semantics of the text data is parsed by natural language processing technology to accurately determine the user control command. This enables the system to understand the user's intentions, whether it is a simple command (such as "turn on the lights") or a complex command (such as "adjust the air conditioning temperature in the living room to 26 degrees and turn on the air purification mode") can be effectively parsed.

[0099] Embodiment 6:

[0100] Based on Example 1, the control instruction generation submodule is as follows: Figure 5 As shown, including:

[0101] An instruction generation unit, used to generate corresponding control instructions based on user commands and in combination with the communication protocol between the embedded speech recognition control module and the lower device;

[0102] The instruction sending unit is used to send control instructions to the lower-level device through a preset communication channel.

[0103] Beneficial effects of the above technical solution: The present invention can generate corresponding device control instructions based on user control commands and send them to lower-level devices. The precise instruction generation and transmission mechanism ensures that the user's voice instructions can be accurately converted into actual operations of the device. For example, in a smart home system, the user's voice instructions can directly control various smart home appliances, such as lamps, curtains, electrical appliances, etc., which improves the convenience and intelligence of device control.

[0104] Embodiment 7:

[0105] Based on Example 1, the adaptive learning and personalization submodule, such as Figure 6 As shown, including:

[0106] A user data processing unit is used to obtain all valid historical wake-up records of the voice recognition control module, and to define a wake-up cycle from the initial wake-up to the successful recognition of the control command by the voice recognition control module, and to divide the voice recognition control module into multiple successful control wake-up cycles to obtain multiple successful control wake-up cycle data sets;

[0107] Respectively obtain invalid voice contents of the user in a plurality of successful control wake-up cycles, and when the number of the invalid voice contents is zero, remove the successful control wake-up cycle;

[0108] A data identification unit is used to generate a corresponding instruction tag based on the successful control instruction corresponding to the successful voice wake-up cycle and add it to the successful control wake-up cycle data set, and monitor invalid voices, or true invalid voices and pseudo invalid voices based on the keywords of the successful control instruction corresponding to the successful control wake-up cycle data set;

[0109] A usage habit analysis unit is used to respectively obtain the unique voice features corresponding to each pseudo invalid voice, compare the unique voice features with the voice features of the standard control instruction corresponding to the control command, and obtain the accent difference between the user's actual voice and the pre-selected voice and the user's text usage habits;

[0110] Based on the voiceprint features of the voice signals in different successful control wake-up cycles, the successful control wake-up cycles are clustered to obtain the voice control data clusters corresponding to the same user;

[0111] Compare the accent differences corresponding to the pseudo-invalid voices in the same voice control data cluster and the user's text usage habits to obtain the current user's personalized accent and usage habits, and compare the personalized accents and usage habits between different users to obtain the user voice similarity;

[0112] A data merging unit, used to merge the voice control data clusters whose user voice similarity is greater than a preset value to obtain a merged data set, and generate different batches of training sets according to the merged data set and the unmerged voice control data clusters;

[0113] The adaptive training unit is used to input the corresponding training set into the deep learning speech recognition model for adaptive training according to the training set batches, until all batches of training are completed and the adaptive training is ended.

[0114] Wherein, the user data processing unit includes:

[0115] The voiceprint file creation subunit is used to create a voiceprint file based on the voiceprint features corresponding to the voice signal. When a new voiceprint feature is detected, a new voiceprint file is automatically generated;

[0116] When confirming that the current user's voice is a voiceprint object, determine whether the current user's voice control command is a single feedback command. If so, add an invalid label to the historical wake-up record of the current user;

[0117] Otherwise, add a valid tag to the historical wake-up record of the current user;

[0118] The intelligent comparison subunit is used to compare the current voice segment collected by the profiled user in the current successful control wake-up cycle with the historical voice segment in the invalid voiceprint file corresponding to the latest similar control command of the profiled user when a pseudo invalid voice appears in the voice control command recognition process of the profiled user, and determine the usage habit difference and accent difference of the profiled user according to the comparison result;

[0119] The audio corresponding to the current speech segment is highlighted according to the usage habit difference or the pronunciation difference, and a training control group is generated based on the highlighted speech segment and the historical speech segment. The training control group is used to perform adaptive training on the deep learning speech recognition model.

[0120] In this embodiment, the unique voice features include the user's accent, speaking speed, tone, word order and other characteristics.

[0121] Beneficial effects of the above technical solution: The present invention adaptively trains the deep learning speech recognition model based on the user's text usage habits and user accent, so that the speech recognition system can better adapt to the characteristics of each user. As the number of times the user uses it increases, the system will become more and more familiar with the user's accent, vocabulary usage habits, etc., thereby continuously improving the accuracy of speech recognition.

[0122] Embodiment 8:

[0123] The present invention provides a device interaction system of an embedded voice recognition control module, which establishes a communication connection between the embedded voice recognition control module described in any one of 1-7 and its corresponding lower-level device.

[0124] In this embodiment, the communication connection includes a wired communication connection and a wireless communication connection:

[0125] In terms of wired communication, the communication protocols of interfaces such as SPI, I2C and UART have been optimized. The SPI interface uses adaptive clock technology, which can automatically adjust the clock frequency according to the data transmission rate to improve data transmission efficiency; the I2C interface adds conflict detection and automatic retransmission mechanisms to enhance communication stability; the UART interface supports multiple baud rate adaptations to facilitate connection with different devices.

[0126] In terms of wireless communications, the Wi-Fi 6 module supports MU-MIMO technology, which can perform high-speed data transmission with multiple devices at the same time; the Bluetooth 5.0 module uses low-power broadcast technology to extend the battery life of the device and enhance the signal coverage range; the ZigBee 3.0 module has self-organizing and self-healing functions to ensure the reliability of the network in complex environments.

[0127] The beneficial effects of the above technical solution: the present invention establishes multiple communication connections between the embedded voice recognition control module and the lower-level devices through the device interaction system, so that the embedded voice recognition control module can be seamlessly connected and stably interact with various types of devices. It has strong compatibility and extensibility, effectively expanding the scope of use of the embedded voice recognition control module while breaking the network's restrictions on voice interaction. Users can freely use voice commands to control the device whether the device is connected to the network or not.

[0128] Obviously, those skilled in the art can make various changes and modifications to the present invention without departing from the spirit and scope of the present invention. Thus, if these modifications and variations of the present invention fall within the scope of the claims of the present invention and their equivalents, the present invention is also intended to include these modifications and variations.

Claims

1. An embedded speech recognition control module, characterized in that: include: A voice collection submodule consisting of one or more microphone arrays, used to collect the user's voice signal; The speech preprocessing submodule is used to preprocess the speech signal to obtain a high-quality speech signal; The speech recognition and command parsing submodule is used to convert high-quality speech signals into text data based on a deep learning speech recognition model, and to perform semantic parsing of the text data through natural language processing technology to determine user control commands; A control instruction generation submodule is used to generate corresponding device control instructions based on user control commands and send them to lower-level devices; The actuator and feedback submodule is used to generate corresponding feedback voice based on the control instruction execution result of the actuator of the lower device; The adaptive learning and personalization submodule is used to adaptively train the deep learning speech recognition model based on the user's text usage habits and accent.

2. An embedded speech recognition control module according to claim 1, characterized in that: The voice collection submodule includes: A direction sensing unit is used to determine the source direction of each sound source signal based on the time difference and phase difference of the sound signals received by each microphone in the microphone array, and to determine the effective coverage range corresponding to each channel signal according to the size of the different channel signals received by each microphone; A sound data analysis unit, used to determine an effective coverage area corresponding to a human voice signal as a target effective coverage range, and confirm a target microphone array area corresponding to the target effective coverage range; Acquire a first target human voice signal collected by each microphone in each target microphone array area, perform target sound fusion based on the first target human voice signal, obtain a second target human voice signal, and send the second target human voice signal as the collected voice signal to the voice preprocessing submodule; A sensitivity control unit, used to compare the multiple first target human voice signals contained in different second target human voice signals with preset thresholds respectively, to obtain an acquisition error rate, and based on the acquisition error rate and in combination with the position information of the person generating the sound, to control the sensitivity of the microphone corresponding to each target human voice signal respectively; When any microphone is within the target microphone array area corresponding to two or more channel signals, the final sensitivity of the microphone is based on the sensitivity corresponding to the channel signal with greater sensitivity.

3. The embedded speech recognition control module according to claim 1, characterized in that: The speech preprocessing submodule includes: A first preprocessing unit, used for performing preliminary processing on the speech signal based on Gaussian filtering and normalized least mean square algorithm to obtain a denoised speech signal; The second preprocessing unit is used to obtain the volume characteristics corresponding to each denoised speech signal, and perform volume normalization processing on the denoised speech signal based on the volume characteristics to obtain a high-quality speech signal.

4. The embedded speech recognition control module according to claim 2, characterized in that: Sensitivity control unit, comprising: The first intelligent control subunit is used to compare the total coverage area of ​​all target effective coverage areas with the full coverage area of ​​the microphone array, determine the invalid vocal area, and adjust the sensitivity value of the microphone corresponding to the invalid vocal area to a minimum threshold; The speaker localization subunit is used to obtain the distribution characteristics of the microphones in the target microphone array area corresponding to each human voice signal, and determine the relative distance between the speaker and the embedded speech recognition control module based on the distribution characteristics and the sound propagation characteristics; The second intelligent control subunit is used to compare the relative distance with a preset standard value to obtain a position difference, and to compare a plurality of first target human voice signals contained in different second target human voice signals with preset thresholds to obtain an acquisition error rate; According to the positive and negative values ​​corresponding to the position difference, the sensitivity adjustment direction of the microphone and the corresponding adjustment coefficient are determined, and based on the acquisition error and the standard sensitivity of the microphone, the sensitivity adjustment amplitude of the microphone is obtained; According to the sensitivity adjustment amplitude and the adjustment coefficient corresponding to the sensitivity adjustment direction, an actual adjustment value is obtained, and based on the actual adjustment value and according to the sensitivity adjustment direction corresponding to the microphone, the sensitivity of the target microphone is adjusted.

5. The embedded speech recognition control module according to claim 1, characterized in that: Speech recognition and command parsing submodules include: Model recognition unit, used to recognize and convert high-quality voice signals into text data based on deep learning voice recognition model, and to perform semantic analysis on text data based on natural language technology to determine user semantics and user intentions; A command screening unit, which is used to screen high-quality voice signals based on user semantics and user intent in combination with a preset control command data set to obtain effective control voice; The command determination unit is used to determine the user control command according to the correspondence between the effective control voice and the control vocabulary in the preset control instruction data set and the control instruction corresponding to the control vocabulary.

6. The embedded speech recognition control module according to claim 1, characterized in that: The control instruction generation submodule includes: An instruction generation unit, used to generate corresponding control instructions based on user commands and in combination with the communication protocol between the embedded speech recognition control module and the lower device; The instruction sending unit is used to send control instructions to the lower-level device through a preset communication channel.

7. The embedded speech recognition control module according to claim 1, characterized in that: Adaptive learning and personalization submodules, including: A user data processing unit, used to obtain all valid historical wake-up records of the voice recognition control module, define the initial wake-up to the successful recognition of the control command by the voice recognition control module as one wake-up cycle, divide the voice recognition control module into multiple successful control wake-up cycles, and obtain multiple successful control wake-up cycle data sets; Respectively obtain invalid voice contents of the user in a plurality of successful control wake-up cycles, and when the number of the invalid voice contents is zero, remove the successful control wake-up cycle; A data identification unit is used to generate a corresponding instruction tag based on the successful control instruction corresponding to the successful voice wake-up cycle and add it to the successful control wake-up cycle data set, and monitor invalid voices, or true invalid voices and pseudo invalid voices based on the keywords of the successful control instruction corresponding to the successful control wake-up cycle data set; A usage habit analysis unit is used to respectively obtain the unique voice features corresponding to each pseudo invalid voice, compare the unique voice features with the voice features of the standard control instruction corresponding to the control command, and obtain the accent difference between the user's actual voice and the pre-selected voice and the user's text usage habits; Based on the voiceprint features of the voice signals in different successful control wake-up cycles, the successful control wake-up cycles are clustered to obtain the voice control data clusters corresponding to the same user; Compare the accent differences corresponding to the pseudo-invalid voices in the same voice control data cluster and the user's text usage habits to obtain the current user's personalized accent and usage habits, and compare the personalized accents and usage habits between different users to obtain the user voice similarity; A data merging unit, used to merge the voice control data clusters whose user voice similarity is greater than a preset value to obtain a merged data set, and generate different batches of training sets according to the merged data set and the unmerged voice control data clusters; The adaptive training unit is used to input the corresponding training set into the deep learning speech recognition model for adaptive training according to the training set batches, until all batches of training are completed and the adaptive training is ended.

8. An embedded speech recognition control module according to claim 7, characterized in that: The user data processing unit comprises: The voiceprint file creation subunit is used to create a voiceprint file based on the voiceprint features corresponding to the voice signal. When a new voiceprint feature is detected, a new voiceprint file is automatically generated; When confirming that the current user's voice is a voiceprint object, determine whether the current user's voice control command is a single feedback command. If so, add an invalid label to the historical wake-up record of the current user; Otherwise, add a valid tag to the historical wakeup record of the current user.

9. An embedded speech recognition control module according to claim 8, characterized in that: The user data processing unit further includes: The intelligent comparison subunit is used to compare the current voice segment collected by the profiled user in the current successful control wake-up cycle with the historical voice segment in the invalid voiceprint file corresponding to the latest similar control command of the profiled user when a pseudo invalid voice appears in the voice control command recognition process of the profiled user, and determine the usage habit difference and accent difference of the profiled user according to the comparison result; The audio corresponding to the current speech segment is highlighted according to the usage habit difference or the pronunciation difference, and a training control group is generated based on the highlighted speech segment and the historical speech segment. The training control group is used to perform adaptive training on the deep learning speech recognition model.

10. A device interaction system with an embedded speech recognition control module, characterized in that: A communication connection is established between the embedded speech recognition control module described in any one of claims 1-9 and its corresponding lower-level device.

Images