Device wake-up related method, device and story machine

By determining the wake-up word probability of speech frames and adjusting the filter coefficients, suppressing human voice noise, and updating the beamforming filter coefficients with wake-up feedback, the problem of traditional VAD failure under vocal interference is solved, and the wake-up performance of the device is improved.

CN113744732BActive Publication Date: 2024-11-05ALIBABA GROUP HOLDING LTD
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202010481877.2
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2020-05-28
Publication Date
2024-11-05
Estimated Expiration
2040-05-28

AI Technical Summary

Technical Problem

In the prior art, under the interference of highly noisy vocals, the voice wake-up effect has dropped sharply, and the failure of traditional VAD results in the reduction of the wake-up performance of the device.

Method used

By determining the wake-up word probability of the speech frame, adjusting the filter coefficient, performing speech enhancement processing to suppress human voice noise, and updating the beamforming filter coefficients in combination with wake-up feedback to distinguish between target voice and human voice noise.

Benefits of technology

It effectively improves the wake-up performance of the device under highly noisy vocal interference, ensuring that the device can accurately recognize the wake-up words and adjust them to the wake-up state.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN113744732B_ABST
    Figure CN113744732B_ABST
Patent Text Reader

Abstract

This application discloses a device wake-up system, method, apparatus, and device. The device wake-up method includes: determining the probability of a wake-up word in the previous speech frame; adjusting filter coefficients based on the wake-up word probability and the previous speech frame; performing speech enhancement processing on the current speech frame based on the adjusted filter coefficients to suppress human voice noise other than the target sound source in the current speech frame; and identifying a wake-up word based on the current speech frame after suppressing human voice noise to adjust the device to a wake-up state. This processing method, by combining wake-up feedback from the previous speech frame with the updated beamforming filter coefficients, can distinguish between target speech and human voice noise, achieving a reliable noise reduction effect; therefore, it can effectively improve the device's wake-up performance under highly noisy human voice interference.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present application relates to the field of automation control technology, and specifically to a device wake-up system, method and apparatus, a voice conference minutes system, method and apparatus, a service startup system, method and apparatus, a story machine, a smart speaker, and an electronic device. Background Art

[0002] In recent years, with the advancement of speech recognition technology, intelligent storytelling machines with voice wake-up function have been widely used. However, the existing acoustic model technology cannot effectively overcome the influence of foam noise and human voice interference. Under highly noisy human voice interference, the voice wake-up effect will drop sharply.

[0003] Microphone array signal processing technology can significantly improve the signal-to-noise ratio and the performance of the voice system. Beamforming is a commonly used array signal processing algorithm with advantages such as low computational complexity and easy deployment, and is suitable for story machines with limited hardware performance. At present, beamforming technology mainly requires voice activity detection (VAD) to distinguish between noise and target sound sources to provide information input.

[0004] However, in the process of implementing the present invention, the inventors found that the technical solution has at least the following problems: the traditional VAD will fail under human voice noise, which seriously reduces the wake-up performance of the storytelling machine under human voice interference noise. In summary, how to improve the beamforming scheme to distinguish human voice noise from the target sound source, so as to improve the wake-up performance of the storytelling machine under highly noisy human voice interference, has become a technical problem that technicians in this field urgently need to solve. Summary of the invention

[0005] The present application provides a device wake-up method to solve the problem of low wake-up performance under highly noisy human voice interference in the prior art. The present application also provides a device wake-up system and device, a voice conference minutes system, method and device, a service startup system, method and device, a story machine, a smart speaker, and an electronic device.

[0006] The present application provides a device wake-up method, comprising:

[0007] Determine the wake-up word probability of the previous speech frame;

[0008] Adjust the filter coefficient according to the wake-up word probability and the previous speech frame;

[0009] Performing speech enhancement processing on the current speech frame according to the adjusted filter coefficient to suppress human voice noise other than the target sound source in the current speech frame;

[0010] The wake-up word is recognized according to the current voice frame after suppressing human voice noise, so as to adjust the device to the wake-up state.

[0011] Optionally, determining the probability of a wake-up word of a previous speech frame includes:

[0012] Performing speech enhancement processing on the previous speech frame according to the filter coefficient before adjustment;

[0013] Determine the acoustic probability of the wake-up word related acoustic unit in the previous speech frame after speech enhancement;

[0014] The wake-up word probability is determined according to the acoustic probability of the acoustic unit related to the wake-up word.

[0015] Optionally, determining the wake-up word probability according to the acoustic probability of the acoustic unit related to the wake-up word includes:

[0016] The maximum acoustic probability is taken as the wake-up word probability.

[0017] Optionally, also include:

[0018] According to the acquisition time of the voice frames, each voice frame is sequentially stored in a cache queue;

[0019] According to the processing time length of determining the acoustic probability, a previous speech frame is read from the cache queue.

[0020] Optionally, adjusting the filter coefficient according to the wake-up word probability and the previous voice frame includes:

[0021] The wake-up word probability is used as the weight of the previous speech frame to determine the target covariance matrix and the noise covariance matrix;

[0022] The adjusted filter coefficients are determined by a beamforming algorithm according to a target covariance matrix and a noise covariance matrix.

[0023] Optionally, the previous speech frame includes: a speech frame adjacent to the current speech frame, or a speech frame not adjacent to the current speech frame.

[0024] Optionally, the speech frames include speech frames collected by multiple microphones.

[0025] The present application also provides a device waking up a device, comprising:

[0026] A wake-up word probability determination unit, used to determine the probability of the wake-up word of the previous speech frame;

[0027] A filter coefficient adjustment unit, used to adjust the filter coefficient according to the wake-up word probability and the previous speech frame;

[0028] A human voice noise suppression unit, used to perform voice enhancement processing on the current voice frame according to the adjusted filter coefficient to suppress human voice noise other than the target sound source in the current voice frame;

[0029] The wake-up unit is used to recognize the wake-up word according to the speech frame after suppressing the human voice noise, so as to adjust the device to the wake-up state.

[0030] Optionally, the wake-up word probability determination unit includes:

[0031] A speech enhancement subunit, used for performing speech enhancement processing on a previous speech frame according to the filter coefficient before adjustment;

[0032] An acoustic probability determination subunit, used to determine the acoustic probability of the wake-up word related acoustic unit in the previous speech frame after speech enhancement;

[0033] The wake-up word probability determination subunit is used to determine the wake-up word probability according to the acoustic probability of the acoustic unit related to the wake-up word.

[0034] Optionally, the filter coefficient adjustment unit includes:

[0035] A weighting subunit, used to use the wake-up word probability as the weight of the previous speech frame, and determine the target covariance matrix and the noise covariance matrix according to the weighted previous speech frame;

[0036] The filter coefficient determination subunit is used to determine the adjusted filter coefficient according to the target covariance matrix and the noise covariance matrix through a beamforming algorithm.

[0037] The present application also provides a story machine, comprising:

[0038] Processor; and

[0039] A memory is used to store a program for implementing a device wake-up method. After the device is powered on and runs the program of the method through the processor, the following steps are performed: determining the probability of a wake-up word in the previous voice frame; adjusting the filter coefficient according to the probability of the wake-up word and the previous voice frame; performing voice enhancement processing on the current voice frame according to the adjusted filter coefficient to suppress human voice noise other than the target sound source in the current voice frame; identifying the wake-up word based on the voice frame after suppressing the human voice noise to adjust the story machine to the wake-up state.

[0040] The present application also provides an electronic device, including:

[0041] Processor; and

[0042] A memory for storing a program for implementing a device wake-up method. After the device is powered on and runs the program of the method through the processor, the following steps are performed: determining the probability of a wake-up word in the previous voice frame; adjusting the filter coefficient according to the probability of the wake-up word and the previous voice frame; performing voice enhancement processing on the current voice frame according to the adjusted filter coefficient to suppress human voice noise other than the target sound source in the current voice frame; identifying the wake-up word according to the voice frame after suppressing the human voice noise, so as to adjust the device to the wake-up state.

[0043] Optionally, the device includes: a smart speaker and a smart TV.

[0044] The present application also provides a method for recording minutes of a voice conference, including:

[0045] Determine the probability of the recording service start word of the previous speech frame;

[0046] Adjust the filter coefficient according to the probability of the starting word and the previous speech frame;

[0047] Performing speech enhancement processing on the current speech frame according to the adjusted filter coefficient to suppress human voice noise other than the target sound source in the current speech frame;

[0048] The start word is identified based on the speech frame after suppressing human voice noise to start the recording service and record the voice meeting minutes.

[0049] Optionally, also include:

[0050] Filter out sound sources whose speaking duration is greater than the duration threshold.

[0051] The present application also provides a voice conference minutes device, including:

[0052] A recording service start word probability determination unit, used to determine the probability of the recording service start word of the previous speech frame;

[0053] A filter coefficient adjustment unit, used to adjust the filter coefficient according to the probability of the start word and the previous speech frame;

[0054] A human voice noise suppression unit, used to perform voice enhancement processing on the current voice frame according to the adjusted filter coefficient to suppress human voice noise other than the target sound source in the current voice frame;

[0055] The recording unit is used to recognize the start word according to the voice frame after suppressing the human voice noise, so as to start the recording service and record the voice conference minutes.

[0056] The present application also provides a smart speaker, comprising:

[0057] Processor; and

[0058] The memory is used to store a program for implementing a device wake-up method. After the device is powered on and runs the program of the method through the processor, the following steps are performed: determining the probability of a recording service start word in the previous voice frame; adjusting the filter coefficient according to the start word probability and the previous voice frame; performing voice enhancement processing on the current voice frame according to the adjusted filter coefficient to suppress human voice noise other than the target sound source in the current voice frame; identifying the start word according to the voice frame after suppressing the human voice noise to start the recording service and record the voice meeting minutes.

[0059] The present application also provides an electronic device, including:

[0060] Processor; and

[0061] The memory is used to store a program for implementing a device wake-up method. After the device is powered on and runs the program of the method through the processor, the following steps are performed: determining the probability of a recording service start word in the previous voice frame; adjusting the filter coefficient according to the start word probability and the previous voice frame; performing voice enhancement processing on the current voice frame according to the adjusted filter coefficient to suppress human voice noise other than the target sound source in the current voice frame; identifying the start word according to the voice frame after suppressing the human voice noise to start the recording service and record the voice meeting minutes.

[0062] The present application also provides a service startup method, including:

[0063] Determine the target service start word probability of the previous speech frame;

[0064] Adjust the filter coefficient according to the probability of the starting word and the previous speech frame;

[0065] Performing speech enhancement processing on the current speech frame according to the adjusted filter coefficient to suppress human voice noise other than the target sound source in the current speech frame;

[0066] The start word is identified according to the speech frame after suppressing human voice noise to start the target service.

[0067] The present application also provides a service startup device, comprising:

[0068] A service start word probability determination unit, used to determine the target service start word probability of the previous speech frame;

[0069] A filter coefficient adjustment unit, used to adjust the filter coefficient according to the probability of the start word and the previous speech frame;

[0070] A human voice noise suppression unit, used to perform voice enhancement processing on the current voice frame according to the adjusted filter coefficient to suppress human voice noise other than the target sound source in the current voice frame;

[0071] The service start unit is used to recognize the start word according to the speech frame after suppressing human voice noise to start the target service.

[0072] The present application also provides a smart speaker, comprising:

[0073] Processor; and

[0074] The memory is used to store a program for implementing a service startup method. After the device is powered on and runs the program of the method through the processor, the following steps are performed: determining the probability of a target service startup word in the previous voice frame; adjusting the filter coefficient according to the startup word probability and the previous voice frame; performing voice enhancement processing on the current voice frame according to the adjusted filter coefficient to suppress human voice noise other than the target sound source in the current voice frame; and identifying the startup word according to the voice frame after suppressing the human voice noise to start the target service.

[0075] The present application also provides an electronic device, including:

[0076] Processor; and

[0077] The memory is used to store a program for implementing a service startup method. After the device is powered on and runs the program of the method through the processor, the following steps are performed: determining the probability of a target service startup word in the previous voice frame; adjusting the filter coefficient according to the startup word probability and the previous voice frame; performing voice enhancement processing on the current voice frame according to the adjusted filter coefficient to suppress human voice noise other than the target sound source in the current voice frame; and identifying the startup word according to the voice frame after suppressing the human voice noise to start the target service.

[0078] The present application also provides a device wake-up system, including:

[0079] The terminal device is used to receive the wake-up word probability of the last voice frame collected by the terminal device and sent by the server, and adjust the filter coefficient according to the last voice frame and the wake-up word probability; according to the adjusted filter coefficient, perform voice enhancement processing on the current voice frame collected by the terminal device to suppress human voice noise other than the target sound source in the current voice frame; if the server recognizes the wake-up word according to the current voice frame after suppressing the human voice noise, the terminal device is adjusted to the wake-up state;

[0080] The server is used to determine the wake-up word probability of the previous speech frame; and to identify the wake-up word based on the current speech frame after suppressing human voice noise.

[0081] The present application also provides a device wake-up method, including:

[0082] Receive the wake-up word probability of the last voice frame collected by the terminal device sent by the server;

[0083] Adjust the filter coefficient according to the previous speech frame and the probability of the wake-up word;

[0084] According to the adjusted filter coefficient, voice enhancement processing is performed on the current voice frame collected by the terminal device to suppress human voice noise other than the target sound source in the current voice frame;

[0085] If the server recognizes the wake-up word based on the current voice frame after suppressing human voice noise, the terminal device is adjusted to the wake-up state.

[0086] The present application also provides a device wake-up method, including:

[0087] Determine the wake-up word probability of the last voice frame collected by the terminal device, and send the wake-up word probability back to the terminal device, so that the terminal device can adjust the filter coefficient according to the last voice frame and the wake-up word probability, and perform voice enhancement processing on the current voice frame collected by the terminal device according to the adjusted filter coefficient to suppress human voice noise other than the target sound source in the current voice frame;

[0088] Identify the wake-up word based on the current voice frame sent by the terminal device after suppressing human voice noise;

[0089] Notify the terminal device to recognize the wake-up word so that the terminal device can adjust to the wake-up state.

[0090] The present application also provides a voice conference minutes system, including:

[0091] The terminal device is used to receive the recording service start word probability of the last voice frame collected by the terminal device and sent by the server, and adjust the filter coefficient according to the last voice frame and the start word probability; perform voice enhancement processing on the current voice frame collected by the terminal device according to the adjusted filter coefficient to suppress human voice noise other than the target sound source in the current voice frame; if the server recognizes the start word according to the current voice frame after suppressing the human voice noise, the recording service is started to record the voice meeting minutes;

[0092] The server is used to receive the last voice frame sent by the terminal device and determine the probability of the start word of the last voice frame; and receive the current voice frame after suppressing human voice noise sent by the terminal device, and identify the start word based on the current voice frame after suppressing human voice noise.

[0093] The present application also provides a method for recording minutes of a voice conference, including:

[0094] The probability of the recording service start word of the last voice frame collected by the terminal device sent by the receiving server;

[0095] Adjust the filter coefficient according to the previous speech frame and the probability of the starting word;

[0096] According to the adjusted filter coefficient, voice enhancement processing is performed on the current voice frame collected by the terminal device to suppress human voice noise other than the target sound source in the current voice frame;

[0097] If the server recognizes the start word based on the current voice frame after suppressing human voice noise, the recording service is started to record the voice conference minutes.

[0098] The present application also provides a method for recording minutes of a voice conference, including:

[0099] Determine the recording service start word probability of the previous voice frame according to the previous voice frame collected by the terminal device, and send the start word probability back to the terminal device, so that the terminal device can adjust the filter coefficient according to the previous voice frame and the start word probability, and perform voice enhancement processing on the current voice frame collected by the terminal device according to the adjusted filter coefficient to suppress human voice noise other than the target sound source in the current voice frame;

[0100] Identify the start word according to the current voice frame after suppressing human voice noise sent by the terminal device;

[0101] The terminal device is notified to recognize the start word so as to start the recording service of the terminal device to record the voice conference minutes.

[0102] The present application also provides a service startup system, comprising:

[0103] The terminal device is used to receive the probability of the target service start word of the previous voice frame collected by the terminal device and sent by the server, and adjust the filter coefficient according to the previous voice frame and the start word probability; perform voice enhancement processing on the current voice frame collected by the terminal device according to the adjusted filter coefficient to suppress human voice noise other than the target sound source in the current voice frame; if the server recognizes the start word according to the current voice frame after suppressing the human voice noise, the target service is started;

[0104] The server is used to receive the last voice frame sent by the terminal device and determine the probability of the start word of the last voice frame; and receive the current voice frame after suppressing human voice noise sent by the terminal device, and identify the start word based on the current voice frame after suppressing human voice noise.

[0105] The present application also provides a service startup method, including:

[0106] Receive the target service start word probability of the last voice frame collected by the terminal device and sent by the server;

[0107] Adjust the filter coefficient according to the previous speech frame and the probability of the starting word;

[0108] According to the adjusted filter coefficient, voice enhancement processing is performed on the current voice frame collected by the terminal device to suppress human voice noise other than the target sound source in the current voice frame;

[0109] If the server recognizes the start word based on the current speech frame after suppressing human voice noise, the target service is started.

[0110] The present application also provides a service startup method, including:

[0111] Determine the target service start word probability of the previous voice frame according to the previous voice frame collected by the terminal device, and send the start word probability back to the terminal device so that the terminal device can adjust the filter coefficient according to the previous voice frame and the start word probability, and perform voice enhancement processing on the current voice frame collected by the terminal device according to the adjusted filter coefficient to suppress human voice noise other than the target sound source in the current voice frame;

[0112] Identify the start word according to the current voice frame after suppressing human voice noise sent by the terminal device;

[0113] The terminal device is notified to recognize the activation word so as to activate the target service of the terminal device.

[0114] The present application also provides a computer-readable storage medium, wherein the computer-readable storage medium stores instructions, and when the computer-readable storage medium is run on a computer, the computer executes the above-mentioned various methods.

[0115] The present application also provides a computer program product comprising instructions, which, when executed on a computer, enables the computer to execute the above-mentioned various methods.

[0116] Compared with the prior art, this application has the following advantages:

[0117] The device wake-up method provided in the embodiment of the present application determines the wake-up word probability of the previous voice frame; adjusts the filter coefficient according to the wake-up word probability and the previous voice frame; performs voice enhancement processing on the current voice frame according to the adjusted filter coefficient to suppress the human voice noise other than the target sound source in the current voice frame; recognizes the wake-up word according to the current voice frame after suppressing the human voice noise to adjust the device to the wake-up state; this processing method makes it possible to update the beamforming filter coefficient in combination with the wake-up feedback of the previous voice frame, so that the target voice and human voice noise can be distinguished to obtain a reliable noise reduction effect; therefore, the wake-up performance of the device under highly noisy human voice interference can be effectively improved.

[0118] The voice conference minutes method provided in the embodiment of the present application determines the probability of the recording service start word of the previous voice frame; adjusts the filter coefficient according to the start word probability and the previous voice frame; performs voice enhancement processing on the current voice frame according to the adjusted filter coefficient to suppress the human voice noise other than the target sound source in the current voice frame; recognizes the start word according to the voice frame after suppressing the human voice noise to start the recording service and record the voice conference minutes; this processing method makes it possible to update the beamforming filter coefficient in combination with the recording service start feedback of the previous voice frame, so that the target voice and human voice noise can be distinguished to obtain a reliable recording service start effect; therefore, the recording service start performance of the device under highly noisy human voice interference can be effectively improved.

[0119] The service startup method provided in the embodiment of the present application determines the probability of the target service startup word of the previous voice frame; adjusts the filter coefficient according to the startup word probability and the previous voice frame; performs voice enhancement processing on the current voice frame according to the adjusted filter coefficient to suppress the human voice noise other than the target sound source in the current voice frame; identifies the startup word according to the voice frame after suppressing the human voice noise to start the target service; this processing method combines the service startup feedback of the previous voice frame to update the beamforming filter coefficient, so that the target voice and human voice noise can be distinguished to obtain a reliable noise reduction effect; therefore, the service startup performance of the device under highly noisy human voice interference can be effectively improved. BRIEF DESCRIPTION OF THE DRAWINGS

[0120] Figure 1 A flowchart of an embodiment of a device wake-up method provided by the present application;

[0121] Figure 2 A signal model diagram of an embodiment of a device wake-up method provided by the present application;

[0122] Figure 3 A specific flow chart of an embodiment of a device wake-up method provided by the present application;

[0123] Figure 4 A schematic diagram of determining the probability of a wake-up word in an embodiment of a device wake-up method provided in the present application;

[0124] Figure 5 A schematic diagram of beamforming in an embodiment of a device wake-up method provided by the present application;

[0125] Figure 6 The present application provides a specific flow chart of an embodiment of a device wake-up method. DETAILED DESCRIPTION

[0126] Many specific details are described in the following description to facilitate a full understanding of the present application. However, the present application can be implemented in many other ways than those described herein, and those skilled in the art can make similar generalizations without violating the connotation of the present application, so the present application is not limited by the specific implementation disclosed below.

[0127] In this application, a device wake-up system, method and apparatus, a voice conference minutes system, method and apparatus, a service startup system, method and apparatus, a story machine, a smart speaker, and an electronic device are provided. Each solution is described in detail in the following embodiments.

[0128] First embodiment

[0129] Please refer to Figure 1 , which is a flow chart of an embodiment of the device wake-up method of the present application. The execution subject of the method includes but is not limited to a smart storytelling machine, and may also be other devices with a wake-up function, such as a smart speaker, a smart repeater, etc. This embodiment takes a storytelling machine as an example to illustrate the method. In this embodiment, the method may include the following steps:

[0130] Step S101: Determine the wake-up word probability of the previous speech frame.

[0131] The storytelling machine has a voice wake-up function, and the storytelling machine can be woken up directly by voice without the user pressing a button. The storytelling machine can collect voice data through multiple microphones (microphone arrays), a segment of voice data can include multiple voice frames, and a voice frame can be a segment of 10ms of voice data. The storytelling machine can process each voice frame in turn.

[0132] For ease of description, the present application refers to a frame of speech that is undergoing noise reduction processing as the current speech frame, a frame of speech at the previous moment of the current speech frame as the previous speech frame, and a frame of speech at the next moment of the current speech frame as the next speech frame. The previous speech frame may be a speech frame adjacent to the current speech frame, or a speech frame that is not adjacent to the current speech frame but is separated by n frames (such as 1 frame or 2 frames). In the case where the previous speech frame is not adjacent to the current speech frame, there is no need to adjust the filter coefficient of each speech frame, which can effectively save computing resources and improve wake-up efficiency.

[0133] The wake-up word probability may be the probability that a frame of speech includes an acoustic unit related to the wake-up word. A wake-up word (such as Tmall Genie Story Machine) usually involves multiple speech frames, each of which may include acoustic units related to the wake-up word, acoustic units unrelated to the wake-up word, or both acoustic units related to the wake-up word and acoustic units unrelated to the wake-up word.

[0134] In the method provided in the embodiment of the present application, when performing noise reduction processing on the current speech frame, the filter coefficient is related to the probability of the wake-up word of the previous speech frame. That is to say, the filter coefficient of the current speech frame is determined in combination with the probability of the wake-up word of the previous speech frame.

[0135] Please refer to Figure 2 , which is a signal model diagram of an embodiment of the device wake-up method of the present application. In this embodiment, the noise reduction processing of the story machine involves the following modules: a beamforming (PMWF) module, a filter module, a wake-up module, a wake-up VAD module, and a cache queue. Figure 2 It can be seen that after the original multi-microphone signal (mic in) collected by the story machine enters the system, it can be sent to the filter module and the cache queue at the same time. Among them, the filter converts the mic in signal into an enhanced signal (enhanced speech) through filtering operation, and then sends it to the wake-up device. The wake-up device outputs the corresponding wake-up signal to the external module (such as the application-level module), and at the same time outputs the corresponding acoustic signal (posteriors) to the wake-up VAD module. Then the PMWF module combines the wake-up VAD and the corresponding original audio signal spit out by the cache queue to update the filter, as shown by the dotted line in the figure, forming a feedback loop here.

[0136] Please refer to Figure 3 , which is a specific flow chart of an embodiment of the device wake-up method of the present application. In one example, step S101 may include the following sub-steps:

[0137] Step S1011: Perform speech enhancement processing on the previous speech frame according to the filter coefficient before adjustment.

[0138] In the method provided in the embodiment of the present application, when the current voice frame is subjected to noise reduction processing, the filter coefficient is related to the probability of the wake-up word of the previous voice frame, so the filter coefficients of different voice frames may not be the same. The filter coefficient of the previous voice frame is referred to as the filter coefficient before adjustment, and the filter coefficient of the current voice frame is referred to as the filter coefficient after adjustment. The filter coefficient before adjustment can be determined based on the probability of the wake-up word of the previous voice frame.

[0139] Speech enhancement refers to the technology of extracting useful speech signals from the noise background and suppressing and reducing noise interference when the speech signal is interfered with or even drowned by various noises.

[0140] In specific implementation, the last speech frame can be subjected to speech enhancement processing by a filter. The filter converts the last speech frame signal into an enhanced signal by filtering operation. Since the filter belongs to a relatively mature prior art, it will not be described here.

[0141] Step S1013: Determine the acoustic probability of the acoustic unit related to the wake-up word in the previous speech frame after speech enhancement.

[0142] After the filter converts the previous speech frame signal into an enhanced signal through a filtering operation, the enhanced signal can be sent to the wake-up device, and the wake-up device determines the acoustic probability of the acoustic unit related to the wake-up word in the previous speech frame after speech enhancement.

[0143] By adopting the method provided in the embodiment of the present application, after the wake-up device recognizes the wake-up signal, it can not only output the corresponding wake-up signal to the external module, but also output the corresponding acoustic signal to the wake-up VAD module at the same time. A frame of speech may include acoustic units related to the wake-up word, and may also include acoustic units unrelated to the wake-up word. For example, if the wake-up word is "Hello TV", the modeling units may include ni, hao, dian, shi, wo, qu, na, etc., then ni, hao, dian, shi are related units, and wo, qu, na are unrelated units. In this embodiment, the wake-up device can output the acoustic probability of each acoustic unit to the wake-up VAD module.

[0144] In this embodiment, the Figure 4 The wake-up device based on the feedforward neural network (FSMN) shown in the figure. The audio signal is sent to the wake-up model based on the FSMN through feature extraction, and the probability of each acoustic unit is given. The acoustic units here can be divided into wake-up word-independent units and wake-up word-related units.

[0145] FSMN is a time series modeling neural network architecture that can effectively use historical and future input information to determine the acoustic probability of each acoustic unit in a frame of speech; this processing method has a high accuracy in wake-up word recognition. In specific implementation, wake-ups with other structures can also be used. Since wake-ups are relatively mature existing technologies, they will not be described here.

[0146] Step S1015: Determine the wake-up word probability according to the acoustic probability of the acoustic unit related to the wake-up word.

[0147] After the wake-up device outputs the acoustic probability of each acoustic unit to the wake-up VAD module, the wake-up VAD module can determine the wake-up word probability according to the acoustic probability of the acoustic unit related to the wake-up word.

[0148] In this embodiment, the maximum value of the acoustic probabilities in all wake-up word related acoustic units is used as the wake-up word probability. The wake-up VAD module converts the input acoustic model score into the wake-up word probability through the following formula: t (W i ) represents the probability output of the i-th wake-up word related unit at time t:

[0149] W i ∈Wake-up word related unit

[0150] In specific implementation, other methods may also be used to determine the wake-up word probability, such as calculating the average value of the acoustic probabilities of all acoustic units related to the wake-up word, and the like.

[0151] After determining the wake-up word probability of the previous speech frame, we can proceed to the next step and adjust the filter coefficient based on the wake-up feedback.

[0152] Step S103: adjusting the filter coefficient according to the wake-up word probability and the previous speech frame.

[0153] like Figure 5 As shown, the method provided in the embodiment of the present application determines the filter coefficient of the current data frame through a beamforming module according to the wake-up word probability and multi-microphone signal of the previous voice frame (multi-microphone signal). In other words, the filter coefficient of the current data frame is related to the wake-up word information of the previous voice frame. However, the traditional beamforming module performs voice activity detection on the current data frame through a traditional VAD module, and determines the filter coefficient of the current data frame according to the voice activity detection result and the multi-microphone signal of the current data frame. This coefficient has nothing to do with the previous data frame, let alone the wake-up word information of the previous data frame, and there is no concept of the probability of the wake-up word.

[0154] Please refer to Figure 6 , which is a specific flow chart of an embodiment of the device wake-up method of the present application. In this embodiment, step S103 may include the following sub-steps:

[0155] Step S1031: Using the wake-up word probability as the weight of the previous speech frame, and determining the target covariance matrix and the noise covariance matrix.

[0156] Step S1033: Determine the adjusted filter coefficients according to the target covariance matrix and the noise covariance matrix through a beamforming algorithm.

[0157] use Figure 6 The processing method shown makes it possible to enhance the output energy of the target sound source by weighting the previous speech frame with its corresponding wake-up word probability and obtaining the beam coefficient through the PMWF formula when calculating the target covariance matrix and the noise covariance matrix.

[0158] In specific implementation, the following pmwf formula can be used to estimate the filter coefficient:

[0159]

[0160] Among them, Φ s =E{p 2 xx H},Φn =E{(1-p) 2 xx H}; p represents the probability of the wake-up word; x is the speech input; E{} is the expectation, which can be achieved by the average value; tr{} is the trace of the matrix, that is, the sum of the matrix diagonals; β is the parameter for controlling speech distortion.

[0161] In one example, the method may further include the following steps: storing each speech frame in sequence according to the acquisition time of the speech frame. Figure 2 In the buffer queue shown in FIG. 1 , the previous speech frame is read from the buffer queue according to the processing time of determining the acoustic probability. In this processing mode, the multi-microphone signal is processed by Figure 2 The FIFO queue shown in the figure is aligned in time with the wake-up word probability output by the wake-up VAD module and then input into the beamforming module together. Since the wake-up word probability is calculated immediately after the acoustic unit is output, the wake-up word probability output delay is the same as the wake-up word model delay, which can be calculated by Figure 4 In the FSMN model, roder control is used to achieve a relatively low latency.

[0162] Step S105: performing speech enhancement processing on the current speech frame according to the adjusted filter coefficient to suppress human voice noise other than the target sound source in the current speech frame.

[0163] Since the adjusted filter coefficient can enhance the output energy of the target sound source, performing speech enhancement processing on the current speech frame according to the filter coefficient can effectively suppress human voice noise other than the target sound source in the current speech frame.

[0164] Step S107: Identify the wake-up word based on the voice frame after suppressing the human voice noise to adjust the story machine to the wake-up state.

[0165] After suppressing the human voice noise other than the target sound source in the current speech frame, the wake-up device can identify the wake-up word based on the speech frame after suppressing the human voice noise. After identifying the wake-up word, the story machine can be adjusted to the wake-up state.

[0166] It can be seen from the above embodiments that the device wake-up method provided in the embodiments of the present application determines the wake-up word probability of the previous voice frame; adjusts the filter coefficient according to the wake-up word probability and the previous voice frame; performs voice enhancement processing on the current voice frame according to the adjusted filter coefficient to suppress the human voice noise other than the target sound source in the current voice frame; recognizes the wake-up word according to the voice frame after suppressing the human voice noise to adjust the device to the wake-up state; this processing method combines the wake-up feedback of the previous voice frame to update the beamforming filter coefficient, so that the target voice and human voice noise can be distinguished to obtain a reliable noise reduction effect; therefore, the wake-up performance of the device under highly noisy human voice interference can be effectively improved.

[0167] Second embodiment

[0168] In the above embodiment, a device awakening method is provided, and correspondingly, the present application also provides a device awakening apparatus, which corresponds to the embodiment of the above method.

[0169] The parts of this embodiment that are the same as those of the first embodiment are not described in detail. Please refer to the corresponding parts in the first embodiment. A device wake-up device provided in the present application includes:

[0170] A wake-up word probability determination unit, used to determine the probability of the wake-up word of the previous speech frame;

[0171] A filter coefficient adjustment unit, used to adjust the filter coefficient according to the wake-up word probability and the previous speech frame;

[0172] A human voice noise suppression unit, used to perform voice enhancement processing on the current voice frame according to the adjusted filter coefficient to suppress human voice noise other than the target sound source in the current voice frame;

[0173] The wake-up unit is used to recognize the wake-up word according to the speech frame after suppressing the human voice noise, so as to adjust the device to the wake-up state.

[0174] In one example, the wake-up word probability determination unit includes:

[0175] A speech enhancement subunit, used for performing speech enhancement processing on a previous speech frame according to the filter coefficient before adjustment;

[0176] An acoustic probability determination subunit, used to determine the acoustic probability of the wake-up word related acoustic unit in the previous speech frame after speech enhancement;

[0177] The wake-up word probability determination subunit is used to determine the wake-up word probability according to the acoustic probability of the acoustic unit related to the wake-up word.

[0178] In one example, the filter coefficient adjustment unit includes:

[0179] A weighting subunit, used to use the wake-up word probability as the weight of the previous speech frame, and determine the target covariance matrix and the noise covariance matrix according to the weighted previous speech frame;

[0180] The filter coefficient determination subunit is used to determine the adjusted filter coefficient according to the target covariance matrix and the noise covariance matrix through a beamforming algorithm.

[0181] Third embodiment

[0182] The present application also provides a storytelling machine. Since the device embodiment is basically similar to the method embodiment, the description is relatively simple, and the relevant parts can be referred to the partial description of the method embodiment. The device embodiment described below is only illustrative.

[0183] A storytelling machine according to the present embodiment includes: a processor and a memory; the memory is used to store a program for implementing a device wake-up method. After the device is powered on and the program of the method is run through the processor, the following steps are performed: determining the probability of a wake-up word in the previous voice frame; adjusting the filter coefficient according to the probability of the wake-up word and the previous voice frame; performing voice enhancement processing on the current voice frame according to the adjusted filter coefficient to suppress human voice noise other than the target sound source in the current voice frame; identifying the wake-up word according to the voice frame after suppressing the human voice noise, so as to adjust the storytelling machine to the wake-up state.

[0184] Fourth embodiment

[0185] The present application also provides an electronic device embodiment. Since the device embodiment is basically similar to the method embodiment, the description is relatively simple, and the relevant parts can be referred to the partial description of the method embodiment. The device embodiment described below is only illustrative.

[0186] An electronic device of this embodiment includes: a processor and a memory; the memory is used to store a program for implementing a device wake-up method. After the device is powered on and the program of the method is run through the processor, the following steps are performed: determining the probability of a wake-up word in a previous voice frame; adjusting a filter coefficient according to the probability of the wake-up word and the previous voice frame; performing voice enhancement processing on a current voice frame according to the adjusted filter coefficient to suppress human voice noise other than a target sound source in the current voice frame; and identifying a wake-up word according to the voice frame after suppressing the human voice noise to adjust the device to a wake-up state.

[0187] Fifth embodiment

[0188] In the above embodiment, a device wake-up method is provided, and correspondingly, the present application also provides a device wake-up system. The system corresponds to the embodiment of the above method. The parts of this embodiment that are the same as the first embodiment are not repeated here, please refer to the corresponding parts in the first embodiment.

[0189] A device wake-up system provided in the present application includes: a server and a terminal device.

[0190] The terminal device is a device with a wake-up function, such as a smart story machine, a smart speaker, a smart repeater, etc.

[0191] The terminal device is used to receive the wake-up word probability of the previous voice frame collected by the terminal device and sent by the server, and adjust the filter coefficient according to the previous voice frame and the wake-up word probability; according to the adjusted filter coefficient, perform voice enhancement processing on the current voice frame collected by the terminal device to suppress the human voice noise other than the target sound source in the current voice frame; if the server recognizes the wake-up word based on the current voice frame after suppressing the human voice noise, the terminal device is adjusted to the wake-up state; accordingly, the server is used to receive the previous voice frame sent by the terminal device, determine the wake-up word probability of the previous voice frame; and, receive the current voice frame after suppressing the human voice noise sent by the terminal device, and identify the wake-up word based on the current voice frame after suppressing the human voice noise.

[0192] In this embodiment, the terminal device can collect voice data through multiple microphones (microphone array), and the terminal device can perform noise reduction processing on each voice frame through a filter, and send the processed voice frame to the server; the server recognizes the wake-up word through the wake-up device, outputs the corresponding wake-up signal to the external module, and outputs the corresponding acoustic signal to the wake-up VAD module, and determines the wake-up word probability of the voice frame through the wake-up VAD module, and sends the probability back to the terminal device; the terminal device determines the filter coefficient of the next voice frame according to the wake-up word probability and the voice frame, and performs noise reduction processing on the next collected voice frame according to the adjusted filter coefficient.

[0193] It can be seen that the differences between the system provided in this embodiment and the method provided in Example 1 include: the system provided in this embodiment is a system in which the server performs wake-up word recognition and wake-up word probability determination, which can further reduce the performance requirements for the terminal device, and a wake-up device updated in real time by the server can be used; therefore, the hardware cost of the terminal device can be effectively reduced, and the accuracy of wake-up word recognition can be effectively improved, thereby effectively improving the device wake-up performance.

[0194] It can be seen from the above embodiments that the device wake-up system provided by the embodiments of the present application sends a voice frame to a server through a terminal device, and the server determines the wake-up word probability of the previous voice frame; the terminal device adjusts the filter coefficient according to the wake-up word probability and the previous voice frame; according to the adjusted filter coefficient, a voice enhancement process is performed on the current voice frame to suppress the human voice noise other than the target sound source in the current voice frame, and the current voice frame after the human voice noise is suppressed is sent to the server; the server recognizes the wake-up word based on the current voice frame after the human voice noise is suppressed, and if the server recognizes the wake-up word, the terminal device is adjusted to the wake-up state; this processing method enables the beamforming filter coefficient to be updated in combination with the wake-up feedback of the previous voice frame, so that the target voice and human voice noise can be distinguished, and a reliable noise reduction effect can be obtained; therefore, the wake-up performance of the device under highly noisy human voice interference can be effectively improved.

[0195] Sixth embodiment

[0196] In the above-mentioned embodiment, a device wake-up system is provided. Correspondingly, the present application also provides a device wake-up method, and the execution subject of the method includes but is not limited to a terminal device. The method corresponds to the embodiment of the above-mentioned system. The parts of this embodiment that are the same as those in the fifth embodiment are not repeated here. Please refer to the corresponding parts in the fifth embodiment.

[0197] A device wake-up method provided in this application includes the following steps:

[0198] Step 1: Receive the wake-up word probability of the last voice frame collected by the terminal device sent by the server;

[0199] Step 2: Adjust the filter coefficient according to the previous speech frame and the wake-up word probability;

[0200] Step 3: Perform speech enhancement processing on the current speech frame collected by the terminal device according to the adjusted filter coefficient to suppress human voice noise other than the target sound source in the current speech frame;

[0201] Step 4: If the server recognizes the wake-up word based on the current voice frame after suppressing human voice noise, the terminal device is adjusted to the wake-up state.

[0202] Seventh embodiment

[0203] In the above-mentioned embodiment, a device wake-up system is provided. Correspondingly, the present application also provides a device wake-up method. The execution subject of the method includes but is not limited to the server, and may also be other devices capable of implementing the method. The method corresponds to the embodiment of the above-mentioned system. The parts of this embodiment that are the same as those in the fifth embodiment will not be repeated here. Please refer to the corresponding parts in the fifth embodiment.

[0204] A device wake-up method provided in this application includes the following steps:

[0205] Step 1: Determine the wake-up word probability of the previous voice frame collected by the terminal device, and send the wake-up word probability back to the terminal device so that the terminal device can adjust the filter coefficient according to the previous voice frame and the wake-up word probability, and perform voice enhancement processing on the current voice frame collected by the terminal device according to the adjusted filter coefficient to suppress human voice noise other than the target sound source in the current voice frame;

[0206] Step 2: Identify the wake-up word based on the current voice frame sent by the terminal device after suppressing human voice noise;

[0207] Step 3: Notify the terminal device to recognize the wake-up word so that the terminal device can adjust to the wake-up state.

[0208] Eighth embodiment

[0209] In the above-mentioned embodiment, a device wake-up method is provided, and correspondingly, the present application also provides a voice conference minutes method. The following first describes the background technology related to the method.

[0210] A typical recording scenario is that a multi-person meeting is being held in a conference room. When the speaker is speaking, participants or conference assistants can use voice to start the recording function of the terminal device at any time to record the conference voice content. Terminal devices are usually loaded with multiple services, such as recording services, song ordering services, and IOT device control services in smart speakers. Each service has a corresponding start word, such as the start word of the recording service is "Tmall Genie, start recording", and the start word of the song ordering service is "I want to order a song". The way the terminal device responds to the user's voice command and starts the recording service is to recognize the start word of the recording service from the user's voice. When the start word of the service is recognized, the recording service is automatically started.

[0211] In order to ensure accurate activation of the recording service under the highly noisy human voice interference in the venue, the terminal device usually uses microphone array signal processing technology to enhance the voice (target sound source) that issues the voice command to activate the recording service, suppress environmental noise and human voice noise in the venue, so as to accurately identify the recording service activation words. In the prior art, a typical microphone array signal processing technology is beamforming technology, which mainly uses voice activity detection (VAD) to distinguish between noise and target sound sources to provide information input.

[0212] However, in the process of implementing the present invention, the inventors found that the technical solution has at least the following problems: the traditional VAD will fail under human voice noise, which seriously reduces the recording service startup performance of the terminal device under human voice interference noise. In summary, how to improve the beamforming solution to improve the recording service startup performance of the terminal device under highly noisy human voice interference has become a technical problem that those skilled in the art urgently need to solve.

[0213] In order to solve this problem, the present application also provides a method for recording minutes of a voice conference. The execution subject of the method includes but is not limited to terminal devices, such as smart speakers, smart TVs, etc. The method corresponds to an embodiment of the above method. The parts of this embodiment that are the same as the first embodiment are not repeated here, please refer to the corresponding parts in the first embodiment.

[0214] A method for recording minutes of a voice conference provided by the present application comprises the following steps:

[0215] Step 1: Determine the probability of the recording service start word of the previous speech frame.

[0216] The recording service start word probability may be the probability that an acoustic unit related to the recording service start word is included in a frame of speech. A recording service start word (such as "Tmall Genie, start recording") usually involves multiple speech frames, each of which may include acoustic units related to the recording service start word, may include acoustic units unrelated to the recording service start word, or may include both acoustic units related to the recording service start word and acoustic units unrelated to the recording service start word.

[0217] In the method provided in the embodiment of the present application, when performing noise reduction processing on the current speech frame, the filter coefficient is related to the probability of the recording service startup word of the previous speech frame, that is, the filter coefficient of the current speech frame is determined in combination with the probability of the recording service startup word of the previous speech frame.

[0218] In one example, step 1 may include the following sub-steps:

[0219] Step 1-1: Perform speech enhancement processing on the previous speech frame according to the filter coefficients before adjustment.

[0220] In the method provided in the embodiment of the present application, when the current voice frame is subjected to noise reduction processing, the filter coefficient is related to the probability of the recording service startup word of the previous voice frame, so the filter coefficients of different voice frames may not be the same. The filter coefficient of the previous voice frame is referred to as the filter coefficient before adjustment, and the filter coefficient of the current voice frame is referred to as the filter coefficient after adjustment. The filter coefficient before adjustment can be determined based on the probability of the recording service startup word of the previous voice frame of the previous voice frame.

[0221] In a specific implementation, the last speech frame can be subjected to speech enhancement processing by a filter, and the filter converts the last speech frame signal into an enhanced signal by filtering operation.

[0222] Step 1-3: Determine the acoustic probability of the acoustic unit related to the recording service start word in the previous speech frame after speech enhancement.

[0223] After the filter converts the previous speech frame signal into an enhanced signal through a filtering operation, the enhanced signal can be sent to the service initiator, and the service initiator determines the acoustic probability of the acoustic unit related to the recorded service initiation word in the previous speech frame after speech enhancement.

[0224] By adopting the method provided in the embodiment of the present application, after the service initiator recognizes the recording service initiation signal, it can not only output the corresponding recording service initiation signal to the external module, but also output the corresponding acoustic signal to the service initiation VAD module at the same time. A frame of speech may include acoustic units related to the recording service initiation word, and may also include acoustic units unrelated to the recording service initiation word. In this embodiment, the service initiator can output the acoustic probability of each acoustic unit to the service initiation VAD module.

[0225] In this embodiment, a service initiator based on a feedforward neural network (FSMN) is used. The audio signal is sent to the service initiation model based on the FSMN through feature extraction to give the probability of each acoustic unit. The acoustic units here can be divided into units unrelated to the recorded service initiation words and units related to the recorded service initiation words.

[0226] Step 1-5: Determine the probability of the recording service activation word according to the acoustic probability of the acoustic unit related to the recording service activation word.

[0227] After the service initiator outputs the acoustic probability of each acoustic unit to the service initiation VAD module, the service initiation VAD module may determine the probability of the recording service initiation word according to the acoustic probability of the acoustic unit related to the recording service initiation word.

[0228] In this embodiment, the maximum value of the acoustic probabilities in all acoustic units related to the recording service activation word is used as the recording service activation word probability.

[0229] After determining the probability of the recording service start word of the previous speech frame, you can proceed to the next step and adjust the filter coefficient based on the recording service start feedback.

[0230] Step 2: Adjust the filter coefficients based on the startup word probability and the previous speech frame.

[0231] The method provided in the embodiment of the present application determines the filter coefficient of the current data frame through a beamforming module according to the probability of the recording service startup word of the previous voice frame (multi-microphone signal) and the multi-microphone signal. In other words, the filter coefficient of the current data frame is related to the recording service startup word information of the previous voice frame. However, the traditional beamforming module performs voice activity detection on the current data frame through a traditional VAD module, and determines the filter coefficient of the current data frame according to the voice activity detection result and the multi-microphone signal of the current data frame. The coefficient has nothing to do with the previous data frame, let alone the recording service startup word information of the previous data frame, and there is no concept of the probability of the recording service startup word.

[0232] In this embodiment, step 2 may include the following sub-steps:

[0233] Step 2-1: Use the probability of the recording service start word as the weight of the previous speech frame to determine the target covariance matrix and the noise covariance matrix.

[0234] Step 2-3: Determine the adjusted filter coefficients according to the target covariance matrix and the noise covariance matrix through a beamforming algorithm.

[0235] In one example, when calculating the target covariance matrix and the noise covariance matrix, the output energy of the target sound source is enhanced by weighting the previous speech frame with the probability of its corresponding recording service start word and obtaining the beam coefficient through the PMWF formula.

[0236] In one example, the method may further include the following steps: storing each speech frame in sequence into a cache queue according to the acquisition time of the speech frame; and reading the previous speech frame from the cache queue according to the processing time for determining the acoustic probability. With this processing method, multi-microphone signals are aligned in time with the recording service start word probability output by the service start VAD module through a first-in-first-out queue, and then input into the beamforming module together. Since the recording service start word probability is calculated immediately after the acoustic unit is output, the recording service start word probability output delay is the same as the recording service start word model delay, which can be controlled by the FSMN model roder, and the delay can be relatively low.

[0237] Step 3: Perform speech enhancement processing on the current speech frame according to the adjusted filter coefficient to suppress human voice noise other than the target sound source in the current speech frame.

[0238] Since the adjusted filter coefficient can enhance the output energy of the target sound source, performing speech enhancement processing on the current speech frame according to the filter coefficient can effectively suppress human voice noise other than the target sound source in the current speech frame.

[0239] Step 4: Identify the start word based on the voice frame after suppressing human voice noise to start the recording service and record the voice meeting minutes.

[0240] After suppressing the human voice noise other than the target sound source in the current voice frame, the service launcher can identify the recording service start word based on the voice frame after suppressing the human voice noise. After identifying the recording service start word, the recording service function of the terminal device can be turned on.

[0241] In one example, the speaker of a conference usually concentrates on speaking, and it is less likely that the speaker will issue a voice command to start the recording service. Usually, the conference assistant or other participants will issue a voice command to start the recording service. Therefore, for starting the recording service, the speaker's voice can be filtered out as human voice noise; accordingly, the method can also include the following steps: filtering out sound sources whose speaking duration is greater than a duration threshold. This processing method can effectively improve the speed and accuracy of starting the recording service.

[0242] The duration threshold can be determined according to actual needs. In specific implementation, the speaking duration of each sound source can be determined first, such as recording the speaking duration of various voiceprints; then, the sound source with a speaking duration greater than the duration threshold is regarded as the sound source of human voice noise and filtered out.

[0243] It can be seen from the above embodiments that the voice conference minutes method provided in the embodiments of the present application determines the probability of the recording service start word of the previous voice frame; adjusts the filter coefficient according to the start word probability and the previous voice frame; performs voice enhancement processing on the current voice frame according to the adjusted filter coefficient to suppress the human voice noise other than the target sound source in the current voice frame; identifies the start word according to the voice frame after suppressing the human voice noise to start the recording service and record the voice conference minutes; this processing method combines the recording service start feedback of the previous voice frame to update the beamforming filter coefficient, so that the target voice and human voice noise can be distinguished to obtain a reliable recording service start effect; therefore, the recording service start performance of the device under highly noisy human voice interference can be effectively improved.

[0244] Ninth embodiment

[0245] In the above embodiment, a voice conference minutes method is provided, and correspondingly, the present application also provides a voice conference minutes system. The system corresponds to the embodiment of the above method. The parts of this embodiment that are the same as the first embodiment are not repeated here, please refer to the corresponding parts in the first embodiment.

[0246] The present application provides a voice conference minutes system including: a server and a terminal device.

[0247] The terminal device is a device with a recording service function, such as a smart speaker.

[0248] The terminal device is used to receive the recording service start word probability of the previous voice frame collected by the terminal device and sent by the server, and adjust the filter coefficient according to the previous voice frame and the start word probability; according to the adjusted filter coefficient, perform voice enhancement processing on the current voice frame collected by the terminal device to suppress the human voice noise other than the target sound source in the current voice frame; if the server recognizes the start word based on the current voice frame after suppressing the human voice noise, the recording service is started to record the voice meeting minutes; accordingly, the server is used to receive the previous voice frame sent by the terminal device, determine the start word probability of the previous voice frame; and, receive the current voice frame after suppressing the human voice noise sent by the terminal device, and recognize the start word based on the current voice frame after suppressing the human voice noise.

[0249] In this embodiment, the terminal device can collect voice data through multiple microphones (microphone array), and the terminal device can perform noise reduction processing on each voice frame through a filter, and send the processed voice frame to the server; the server recognizes the recording service start-up word through the service starter, outputs the corresponding recording service start-up signal to the external module, and simultaneously outputs the corresponding acoustic signal to the service start-up VAD module, and determines the recording service start-up word probability of the voice frame through the service start-up VAD module, and sends the probability back to the terminal device; the terminal device determines the filter coefficient of the next voice frame according to the recording service start-up word probability and the voice frame, and performs noise reduction processing on the next collected voice frame according to the adjusted filter coefficient.

[0250] It can be seen that the differences between the system provided in this embodiment and the method provided in Example 8 include: the system provided in this embodiment is a process in which the server side performs recording service startup word recognition and recording service startup word probability determination, which can further reduce the performance requirements for the terminal device, and can use a service launcher that is updated in real time by the server side; therefore, the hardware cost of the terminal device can be effectively reduced, and the accuracy of recording service startup word recognition can be effectively improved, thereby effectively improving the device recording service startup performance.

[0251] It can be seen from the above embodiments that the voice conference minutes system provided by the embodiments of the present application sends voice frames to the server through the terminal device, and the server determines the probability of the recording service start-up word of the previous voice frame; the terminal device adjusts the filter coefficient according to the probability of the recording service start-up word and the previous voice frame; according to the adjusted filter coefficient, the current voice frame is subjected to voice enhancement processing to suppress the human voice noise other than the target sound source in the current voice frame, and the current voice frame after the human voice noise is suppressed is sent to the server; the server recognizes the recording service start-up word according to the current voice frame after the human voice noise is suppressed, and if the server recognizes the recording service start-up word, the terminal device starts the recording service to record the voice conference minutes; this processing method enables the beamforming filter coefficient to be updated in combination with the recording service start-up feedback of the previous voice frame, so that the target voice and human voice noise can be distinguished, and a reliable noise reduction effect can be obtained; therefore, the recording service startup performance of the device under highly noisy human voice interference can be effectively improved.

[0252] Tenth embodiment

[0253] In the above embodiment, a voice conference minutes system is provided. Correspondingly, the present application also provides a voice conference minutes method, and the execution subject of the method includes but is not limited to a terminal device. The method corresponds to the embodiment of the above system. The parts of this embodiment that are the same as the ninth embodiment are not repeated here. Please refer to the corresponding parts in the ninth embodiment.

[0254] A method for recording minutes of a voice conference provided by the present application comprises the following steps:

[0255] Step 1: Receive the recording service start word probability of the last voice frame collected by the terminal device and sent by the server;

[0256] Step 2: Adjust the filter coefficient according to the previous speech frame and the probability of the start word;

[0257] Step 3: Perform speech enhancement processing on the current speech frame collected by the terminal device according to the adjusted filter coefficient to suppress human voice noise other than the target sound source in the current speech frame;

[0258] Step 4: If the server recognizes the start word based on the current voice frame after suppressing human voice noise, the recording service is started to record the voice conference minutes.

[0259] Eleventh Embodiment

[0260] In the above embodiment, a voice conference minutes system is provided. Correspondingly, the present application also provides a voice conference minutes method, and the execution subject of the method includes but is not limited to the server. The method corresponds to the embodiment of the above system. The parts of this embodiment that are the same as the ninth embodiment are not repeated here. Please refer to the corresponding parts in the ninth embodiment.

[0261] A method for recording minutes of a voice conference provided by the present application comprises the following steps:

[0262] Step 1: According to the last voice frame collected by the terminal device, determine the probability of the recording service start word of the last voice frame, and send the start word probability back to the terminal device, so that the terminal device can adjust the filter coefficient according to the last voice frame and the start word probability, and perform voice enhancement processing on the current voice frame collected by the terminal device according to the adjusted filter coefficient to suppress human voice noise other than the target sound source in the current voice frame;

[0263] Step 2: Identify the start word according to the current voice frame after suppressing human voice noise sent by the terminal device;

[0264] Step 3: Notify the terminal device to recognize the start word so as to start the recording service of the terminal device to record the voice conference minutes.

[0265] Twelfth Embodiment

[0266] In the above embodiment, a voice conference method is provided. Correspondingly, the present application also provides a service startup method, the execution subject of which includes but is not limited to a terminal device, such as a smart speaker, etc. The method corresponds to the embodiment of the above method. The parts of this embodiment that are the same as the eighth embodiment are not repeated here, please refer to the corresponding parts in the eighth embodiment.

[0267] A service startup method provided by this application includes the following steps:

[0268] Step 1: Determine the probability of the target service start word in the previous speech frame;

[0269] Step 2: Adjust the filter coefficient according to the probability of the starting word and the previous speech frame;

[0270] Step 3: Perform speech enhancement processing on the current speech frame according to the adjusted filter coefficient to suppress human voice noise other than the target sound source in the current speech frame;

[0271] Step 4: Identify the start word based on the speech frame after suppressing human voice noise to start the target service.

[0272] It can be seen from the above embodiments that the service startup method provided in the embodiments of the present application determines the probability of the target service startup word of the previous voice frame; adjusts the filter coefficient according to the startup word probability and the previous voice frame; performs voice enhancement processing on the current voice frame according to the adjusted filter coefficient to suppress the human voice noise other than the target sound source in the current voice frame; identifies the startup word according to the voice frame after suppressing the human voice noise to start the target service; this processing method combines the service startup feedback of the previous voice frame to update the beamforming filter coefficient, so that the target voice and human voice noise can be distinguished to obtain a reliable noise reduction effect; therefore, the service startup performance of the device under highly noisy human voice interference can be effectively improved.

[0273] Thirteenth Embodiment

[0274] In the above embodiment, a service startup method is provided, and correspondingly, the present application also provides a service startup system. The system corresponds to the embodiment of the above method. The parts of this embodiment that are the same as the first embodiment are not repeated here, please refer to the corresponding parts in the first embodiment.

[0275] A service startup system provided in the present application includes: a server and a terminal device.

[0276] The terminal device is a device with a wake-up function, such as a smart speaker.

[0277] The terminal device is used to receive the probability of the target service start word of the previous voice frame collected by the terminal device and sent by the server, and adjust the filter coefficient according to the previous voice frame and the start word probability; according to the adjusted filter coefficient, perform voice enhancement processing on the current voice frame collected by the terminal device to suppress the human voice noise other than the target sound source in the current voice frame; if the server recognizes the start word based on the current voice frame after suppressing the human voice noise, the target service is started; accordingly, the server is used to receive the previous voice frame sent by the terminal device, determine the start word probability of the previous voice frame; and, receive the current voice frame after suppressing the human voice noise sent by the terminal device, and identify the start word based on the current voice frame after suppressing the human voice noise.

[0278] In this embodiment, the terminal device can collect voice data through multiple microphones (microphone array), and the terminal device can perform noise reduction processing on each voice frame through a filter, and send the processed voice frame to the server; the server identifies the target service startup word through the service launcher, outputs the corresponding wake-up signal to the external module, and simultaneously outputs the corresponding acoustic signal to the service launch VAD module, and determines the target service startup word probability of the voice frame through the service launch VAD module, and sends the probability back to the terminal device; the terminal device determines the filter coefficient of the next voice frame based on the target service startup word probability and the voice frame, and performs noise reduction processing on the next collected voice frame based on the adjusted filter coefficient.

[0279] It can be seen that the differences between the system provided in this embodiment and the method provided in Example 12 include: the system provided in this embodiment is a process in which the service start-up word recognition and service start-up word probability determination are performed by the server, which can further reduce the performance requirements for the terminal device, and a service launcher updated in real time by the server can be used; therefore, the hardware cost of the terminal device can be effectively reduced, and the accuracy of service start-up word recognition can be effectively improved, thereby effectively improving the device service startup performance.

[0280] It can be seen from the above embodiments that the service startup system provided by the embodiments of the present application sends a voice frame to a server through a terminal device, and the server determines the probability of a target service startup word in a previous voice frame; the terminal device adjusts the filter coefficient according to the target service startup word probability and the previous voice frame; according to the adjusted filter coefficient, a voice enhancement process is performed on the current voice frame to suppress human voice noise other than the target sound source in the current voice frame, and the current voice frame after suppressing the human voice noise is sent to the server; the server recognizes the target service startup word based on the current voice frame after suppressing the human voice noise, and if the server recognizes the target service startup word, the target service is started; this processing method enables the beamforming filter coefficient to be updated in combination with the wake-up feedback of the previous voice frame, so that the target voice and human voice noise can be distinguished, and a reliable noise reduction effect can be obtained; therefore, the service startup performance of the device under highly noisy human voice interference can be effectively improved.

[0281] Fourteenth Embodiment

[0282] In the above-mentioned embodiment, a service startup system is provided. Correspondingly, the present application also provides a service startup method, and the execution subject of the method includes but is not limited to a terminal device. The method corresponds to the embodiment of the above-mentioned system. The parts of this embodiment that are the same as the thirteenth embodiment are not repeated here. Please refer to the corresponding parts in the thirteenth embodiment.

[0283] A service startup method provided by this application includes the following steps:

[0284] Step 1: Receive the target service start word probability of the last voice frame collected by the terminal device and sent by the server;

[0285] Step 2: Adjust the filter coefficient according to the previous speech frame and the probability of the start word;

[0286] Step 3: Perform speech enhancement processing on the current speech frame collected by the terminal device according to the adjusted filter coefficient to suppress human voice noise other than the target sound source in the current speech frame;

[0287] Step 4: If the server recognizes the start word based on the current speech frame after suppressing human voice noise, the target service is started.

[0288] Fifteenth Embodiment

[0289] In the above-mentioned embodiment, a service startup system is provided. Correspondingly, the present application also provides a service startup method, and the execution subject of the method includes but is not limited to a terminal device. The method corresponds to the embodiment of the above-mentioned system. The parts of this embodiment that are the same as the thirteenth embodiment are not repeated here. Please refer to the corresponding parts in the thirteenth embodiment.

[0290] A service startup method provided by this application includes the following steps:

[0291] Step 1: According to the last voice frame collected by the terminal device, determine the probability of the target service start word of the last voice frame, and send the start word probability back to the terminal device, so that the terminal device can adjust the filter coefficient according to the last voice frame and the start word probability, and perform voice enhancement processing on the current voice frame collected by the terminal device according to the adjusted filter coefficient to suppress human voice noise other than the target sound source in the current voice frame;

[0292] Step 2: Identify the start word according to the current voice frame after suppressing human voice noise sent by the terminal device;

[0293] Step 3: Notify the terminal device to recognize the activation word so as to activate the target service of the terminal device.

[0294] Although the present application is disclosed as above in the form of a preferred embodiment, it is not intended to limit the present application. Any technical personnel in this field may make possible changes and modifications without departing from the spirit and scope of the present application. Therefore, the scope of protection of the present application shall be based on the scope defined by the claims of the present application.

[0295] In a typical configuration, a computing device includes one or more processors (CPU), input / output interfaces, network interfaces, and memory.

[0296] The memory may include non-permanent storage in a computer-readable medium, random access memory (RAM) and / or non-volatile memory in the form of read-only memory (ROM) or flash RAM. The memory is an example of a computer-readable medium.

[0297] 1. Computer-readable media includes permanent and non-permanent, removable and non-removable media that can be used to store information by any method or technology. Information can be computer-readable instructions, data structures, program modules or other data. Examples of computer storage media include, but are not limited to, phase change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory or other memory technology, compact disk read-only memory (CD-ROM), digital versatile disk (DVD) or other optical storage, magnetic cassettes, magnetic tape magnetic disk storage or other magnetic storage devices or any other non-transmission media that can be used to store information that can be accessed by a computing device. As defined herein, computer-readable media does not include non-transitory media such as modulated data signals and carrier waves.

[0298] 2. Those skilled in the art should understand that the embodiments of the present application can be provided as methods, systems or computer program products. Therefore, the present application can take the form of a complete hardware embodiment, a complete software embodiment or an embodiment combining software and hardware. Moreover, the present application can take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

Claims

1. A device wake-up method, It is characterized in that include: Determine the wake-up word probability of the previous speech frame; The wake-up word probability is used as the weight of the previous speech frame to determine the target covariance matrix and the noise covariance matrix; Through the beamforming algorithm, the filter coefficients are adjusted according to the target covariance matrix and the noise covariance matrix; According to the adjusted filter coefficient, a speech enhancement process is performed on the current speech frame to suppress human voice noise other than the target sound source in the current speech frame; The wake-up word is recognized according to the current voice frame after suppressing human voice noise, so as to adjust the device to the wake-up state.

2. The method according to claim 1, It is characterized in that The determining of the wake-up word probability of the previous speech frame includes: Performing speech enhancement processing on the previous speech frame according to the filter coefficient before adjustment; Determine the acoustic probability of the wake-up word related acoustic unit in the previous speech frame after speech enhancement; The wake-up word probability is determined according to the acoustic probability of the acoustic unit related to the wake-up word.

3. The method according to claim 2, It is characterized in that The determining the wake-up word probability according to the acoustic probability of the acoustic unit related to the wake-up word includes: The maximum acoustic probability is taken as the wake-up word probability.

4. The method according to claim 2, It is characterized in that Also includes: According to the acquisition time of the voice frames, each voice frame is sequentially stored in a cache queue; According to the processing time length of determining the acoustic probability, a previous speech frame is read from the cache queue.

5. The method according to claim 1, It is characterized in that The previous speech frame includes: a speech frame adjacent to the current speech frame, or a speech frame not adjacent to the current speech frame.

6. The method according to claim 1, It is characterized in that The speech frames include speech frames collected by a plurality of microphones.

7. A device wake-up device, It is characterized in that include: A wake-up word probability determination unit, used to determine the probability of the wake-up word of the previous speech frame; A filter coefficient adjustment unit is used to use the wake-up word probability as the weight of the previous speech frame to determine the target covariance matrix and the noise covariance matrix; and adjust the filter coefficient according to the target covariance matrix and the noise covariance matrix through a beamforming algorithm; A human voice noise suppression unit, used to perform voice enhancement processing on the current voice frame according to the adjusted filter coefficient to suppress human voice noise other than the target sound source in the current voice frame; The wake-up unit is used to recognize the wake-up word according to the speech frame after suppressing the human voice noise, so as to adjust the device to the wake-up state.

8. The device according to claim 7, It is characterized in that The wake-up word probability determination unit includes: A speech enhancement subunit, used for performing speech enhancement processing on a previous speech frame according to the filter coefficient before adjustment; An acoustic probability determination subunit, used to determine the acoustic probability of the wake-up word related acoustic unit in the previous speech frame after speech enhancement; The wake-up word probability determination subunit is used to determine the wake-up word probability according to the acoustic probability of the acoustic unit related to the wake-up word.

9. A story machine, It is characterized in that include: processor; as well as A memory for storing a program for implementing a device wake-up method. After the device is powered on and the program of the method is run through the processor, the following steps are performed: determining the probability of a wake-up word of a previous voice frame; using the probability of the wake-up word as a weight of the previous voice frame to determine a target covariance matrix and a noise covariance matrix; adjusting the filter coefficient according to the target covariance matrix and the noise covariance matrix through a beamforming algorithm; According to the adjusted filter coefficient, speech enhancement processing is performed on the current speech frame to suppress human voice noise other than the target sound source in the current speech frame; the wake-up word is recognized according to the speech frame after suppressing the human voice noise to adjust the story machine to the wake-up state.

10. An electronic device, It is characterized in that include: processor; as well as A memory for storing a program for implementing a device wake-up method. After the device is powered on and the program of the method is run through the processor, the following steps are performed: determining the probability of a wake-up word of a previous voice frame; using the probability of the wake-up word as a weight of the previous voice frame to determine a target covariance matrix and a noise covariance matrix; adjusting the filter coefficient according to the target covariance matrix and the noise covariance matrix through a beamforming algorithm; According to the adjusted filter coefficient, speech enhancement processing is performed on the current speech frame to suppress human voice noise other than the target sound source in the current speech frame; the wake-up word is recognized according to the speech frame after suppressing the human voice noise to adjust the device to the wake-up state.

11. A method for recording minutes of a voice conference. It is characterized in that include: Determine the probability of the recording service start word of the previous speech frame; The probability of the start word is used as the weight of the previous speech frame to determine the target covariance matrix and the noise covariance matrix; Through the beamforming algorithm, the filter coefficients are adjusted according to the target covariance matrix and the noise covariance matrix; According to the adjusted filter coefficient, a speech enhancement process is performed on the current speech frame to suppress human voice noise other than the target sound source in the current speech frame; The start word is identified based on the speech frame after suppressing human voice noise to start the recording service and record the voice meeting minutes.

12. The method according to claim 11, It is characterized in that Also includes: Filter out sound sources whose speaking duration is greater than the duration threshold.

13. A voice conference minutes device, It is characterized in that include: A recording service start word probability determination unit, used to determine the probability of the recording service start word of the previous speech frame; A filter coefficient adjustment unit is used to use the probability of the start word as the weight of the previous speech frame to determine the target covariance matrix and the noise covariance matrix; and adjust the filter coefficient according to the target covariance matrix and the noise covariance matrix through a beamforming algorithm; A human voice noise suppression unit, used to perform voice enhancement processing on the current voice frame according to the adjusted filter coefficient to suppress human voice noise other than the target sound source in the current voice frame; The recording unit is used to recognize the start word according to the voice frame after suppressing the human voice noise, so as to start the recording service and record the voice conference minutes.

14. A smart speaker, It is characterized in that include: processor; as well as The memory is used to store a program for implementing a device wake-up method. After the device is powered on and the program of the method is run through the processor, the following steps are performed: determining the probability of a recording service start word of the previous voice frame; using the start word probability as the weight of the previous voice frame to determine a target covariance matrix and a noise covariance matrix; adjusting the filter coefficient according to the target covariance matrix and the noise covariance matrix through a beamforming algorithm; According to the adjusted filter coefficient, a speech enhancement process is performed on the current speech frame to suppress human voice noise other than the target sound source in the current speech frame; The start word is identified based on the speech frame after suppressing human voice noise to start the recording service and record the voice meeting minutes.

15. An electronic device, It is characterized in that include: processor; as well as The memory is used to store a program for implementing a device wake-up method. After the device is powered on and the program of the method is run through the processor, the following steps are performed: determining the probability of a recording service start word of the previous voice frame; using the start word probability as the weight of the previous voice frame to determine a target covariance matrix and a noise covariance matrix; adjusting the filter coefficient according to the target covariance matrix and the noise covariance matrix through a beamforming algorithm; According to the adjusted filter coefficient, a speech enhancement process is performed on the current speech frame to suppress human voice noise other than the target sound source in the current speech frame; The start word is identified based on the speech frame after suppressing human voice noise to start the recording service and record the voice meeting minutes.

16. A service startup method, It is characterized in that include: Determine the target service start word probability of the previous speech frame; The probability of the start word is used as the weight of the previous speech frame to determine the target covariance matrix and the noise covariance matrix; Through the beamforming algorithm, the filter coefficients are adjusted according to the target covariance matrix and the noise covariance matrix; According to the adjusted filter coefficient, a speech enhancement process is performed on the current speech frame to suppress human voice noise other than the target sound source in the current speech frame; The start word is identified according to the speech frame after suppressing human voice noise to start the target service.

17. A service startup device, It is characterized in that include: A service start word probability determination unit, used to determine the target service start word probability of the previous speech frame; A filter coefficient adjustment unit is used to use the probability of the start word as the weight of the previous speech frame to determine the target covariance matrix and the noise covariance matrix; and adjust the filter coefficient according to the target covariance matrix and the noise covariance matrix through a beamforming algorithm; A human voice noise suppression unit, used to perform voice enhancement processing on the current voice frame according to the adjusted filter coefficient to suppress human voice noise other than the target sound source in the current voice frame; The service start unit is used to recognize the start word according to the speech frame after suppressing human voice noise to start the target service.

18. A smart speaker, It is characterized in that include: processor; as well as The memory is used to store a program for implementing the service startup method. After the smart speaker is powered on and runs the program of the method through the processor, the following steps are performed: determining the probability of the target service startup word of the previous voice frame; using the startup word probability as the weight of the previous voice frame to determine the target covariance matrix and the noise covariance matrix; adjusting the filter coefficient according to the target covariance matrix and the noise covariance matrix through a beamforming algorithm; According to the adjusted filter coefficient, a speech enhancement process is performed on the current speech frame to suppress human voice noise other than the target sound source in the current speech frame; The start word is identified according to the speech frame after suppressing human voice noise to start the target service.

19. An electronic device, It is characterized in that include: processor; as well as The memory is used to store a program for implementing the service startup method. After the device is powered on and the program of the method is run through the processor, the following steps are performed: determining the probability of the target service startup word of the previous voice frame; using the startup word probability as the weight of the previous voice frame to determine the target covariance matrix and the noise covariance matrix; adjusting the filter coefficient according to the target covariance matrix and the noise covariance matrix through a beamforming algorithm; According to the adjusted filter coefficient, a speech enhancement process is performed on the current speech frame to suppress human voice noise other than the target sound source in the current speech frame; The start word is identified according to the speech frame after suppressing human voice noise to start the target service.

20. A device wake-up system, It is characterized in that include: The terminal device is used to receive the wake-up word probability of the previous voice frame collected by the terminal device and sent by the server, use the wake-up word probability as the weight of the previous voice frame, determine the target covariance matrix and the noise covariance matrix; and adjust the filter coefficient according to the target covariance matrix and the noise covariance matrix through the beamforming algorithm; According to the adjusted filter coefficient, voice enhancement processing is performed on the current voice frame collected by the terminal device to suppress human voice noise other than the target sound source in the current voice frame; If the server recognizes the wake-up word based on the current voice frame after suppressing human voice noise, the terminal device is adjusted to the wake-up state; The server is used to determine the wake-up word probability of the previous speech frame; and to identify the wake-up word based on the current speech frame after suppressing human voice noise.

21. A device wake-up method, It is characterized in that include: Receive the wake-up word probability of the last voice frame collected by the terminal device sent by the server; The wake-up word probability is used as the weight of the previous speech frame to determine the target covariance matrix and the noise covariance matrix; Through the beamforming algorithm, the filter coefficients are adjusted according to the target covariance matrix and the noise covariance matrix; According to the adjusted filter coefficient, voice enhancement processing is performed on the current voice frame collected by the terminal device to suppress human voice noise other than the target sound source in the current voice frame; If the server recognizes the wake-up word based on the current voice frame after suppressing human voice noise, the terminal device is adjusted to the wake-up state.

22. A device wake-up method, It is characterized in that include: Determine the wake-up word probability of the last voice frame collected by the terminal device, and send the wake-up word probability back to the terminal device so that the terminal device uses the wake-up word probability as the weight of the last voice frame to determine the target covariance matrix and the noise covariance matrix; adjust the filter coefficient according to the target covariance matrix and the noise covariance matrix through the beamforming algorithm, and perform voice enhancement processing on the current voice frame collected by the terminal device according to the adjusted filter coefficient to suppress human voice noise other than the target sound source in the current voice frame; Identify the wake-up word based on the current voice frame sent by the terminal device after suppressing human voice noise; Notify the terminal device to recognize the wake-up word so that the terminal device can adjust to the wake-up state.

23. A voice conference minutes system, It is characterized in that include: The terminal device is used to receive the recording service start word probability of the last voice frame collected by the terminal device and sent by the server, use the start word probability as the weight of the last voice frame, determine the target covariance matrix and the noise covariance matrix; adjust the filter coefficient according to the target covariance matrix and the noise covariance matrix through the beamforming algorithm; According to the adjusted filter coefficient, voice enhancement processing is performed on the current voice frame collected by the terminal device to suppress human voice noise other than the target sound source in the current voice frame; If the server recognizes the start word according to the current voice frame after suppressing human voice noise, the recording service is started to record the voice conference minutes; The server is used to receive the last speech frame sent by the terminal device and determine the probability of the start word of the last speech frame; And, a current speech frame after suppressing human voice noise sent by a receiving terminal device, and identifying the start word according to the current speech frame after suppressing human voice noise.

24. A method for recording minutes of a voice conference. It is characterized in that include: The probability of the recording service start word of the last voice frame collected by the terminal device sent by the receiving server; The probability of the start word is used as the weight of the previous speech frame to determine the target covariance matrix and the noise covariance matrix; Through the beamforming algorithm, the filter coefficients are adjusted according to the target covariance matrix and the noise covariance matrix; According to the adjusted filter coefficient, voice enhancement processing is performed on the current voice frame collected by the terminal device to suppress human voice noise other than the target sound source in the current voice frame; If the server recognizes the start word based on the current voice frame after suppressing human voice noise, the recording service is started to record the voice conference minutes.

25. A method for recording minutes of a voice conference. It is characterized in that include: According to the last voice frame collected by the terminal device, determine the probability of the recording service start word of the last voice frame, and send the start word probability back to the terminal device, so that the terminal device uses the start word probability as the weight of the last voice frame to determine the target covariance matrix and the noise covariance matrix; adjust the filter coefficient according to the target covariance matrix and the noise covariance matrix through the beamforming algorithm, and perform voice enhancement processing on the current voice frame collected by the terminal device according to the adjusted filter coefficient, so as to suppress the human voice noise other than the target sound source in the current voice frame; Identify the start word according to the current voice frame after suppressing human voice noise sent by the terminal device; The terminal device is notified to recognize the start word so as to start the recording service of the terminal device to record the voice conference minutes.

26. A service startup system, It is characterized in that include: The terminal device is used to receive the probability of the target service start word of the previous voice frame collected by the terminal device and sent by the server, use the start word probability as the weight of the previous voice frame, determine the target covariance matrix and the noise covariance matrix; adjust the filter coefficient according to the target covariance matrix and the noise covariance matrix through the beamforming algorithm; According to the adjusted filter coefficient, voice enhancement processing is performed on the current voice frame collected by the terminal device to suppress human voice noise other than the target sound source in the current voice frame; If the server recognizes the start word based on the current voice frame after suppressing human voice noise, the target service is started; The server is used to receive the last voice frame sent by the terminal device and determine the probability of the start word of the last voice frame; and receive the current voice frame after suppressing human voice noise sent by the terminal device, and identify the start word based on the current voice frame after suppressing human voice noise.

27. A service startup method, It is characterized in that include: Receive the target service start word probability of the last voice frame collected by the terminal device and sent by the server; The probability of the start word is used as the weight of the previous speech frame to determine the target covariance matrix and the noise covariance matrix; Through the beamforming algorithm, the filter coefficients are adjusted according to the target covariance matrix and the noise covariance matrix; According to the adjusted filter coefficient, voice enhancement processing is performed on the current voice frame collected by the terminal device to suppress human voice noise other than the target sound source in the current voice frame; If the server recognizes the start word based on the current speech frame after suppressing human voice noise, the target service is started.

28. A service startup method, It is characterized in that include: According to the last voice frame collected by the terminal device, the probability of the target service start word of the last voice frame is determined, and the start word probability is sent back to the terminal device, so that the terminal device uses the start word probability as the weight of the last voice frame to determine the target covariance matrix and the noise covariance matrix; through the beamforming algorithm, the filter coefficient is adjusted according to the target covariance matrix and the noise covariance matrix, and according to the adjusted filter coefficient, the current voice frame collected by the terminal device is performed The speech enhancement processing is performed to suppress the human voice noise other than the target sound source in the current voice frame; Identify the start word according to the current voice frame after suppressing human voice noise sent by the terminal device; The terminal device is notified to recognize the activation word so as to activate the target service of the terminal device.