Signal processing method and terminal equipment

Abstract

The present invention relates to the field of speech recognition technology, and provides a signal processing method and a terminal device, wherein the signal processing method comprises: extracting feature information from an input signal; and determining whether the input signal is an abnormal speech signal according to the extracted feature information. In the present invention, by determining whether the input signal is an abnormal speech signal according to the feature information extracted from the input signal, effective recognition of abnormal speech signals is achieved, the accuracy of recognition is improved, and security is provided for the user's speech interaction, thereby improving the user's experience of use.

Description

Technical Field

[0001] The present invention relates to the technical field of speech recognition, and in particular to a signal processing method and terminal equipment. Background Art

[0002] Ultrasonic attack (also known as "dolphin attack") was proposed at the 2017 ACM CCS (The ACM Conference on Computer and Communications Security) conference. It is said that ultrasonic attack can break into the voice assistant application of the terminal, which makes people question the security of voice interaction.

[0003] Ultrasonic attacks utilize the basic principles of ultrasound. The implementation idea is: in the attack signal modulation stage, the attack voice signal (here called the baseband signal) in the first frequency range (normal frequency range) is modulated to an ultrasonic range that the user cannot hear. Then in the attack signal transmission stage, an ultrasonic transmitter is used to transmit the modulated attack voice signal to the device under test (terminal devices such as mobile phones and smart home devices). In the voice signal acquisition stage, the attack voice signal will be collected by the voice control system of the device under test. The voice signal acquisition stage further includes a voice signal demodulation stage, which automatically demodulates the hardware vulnerabilities of the voice acquisition system on the device under test to restore the baseband signal. After that, the demodulated baseband signal passes through the amplifier, low-pass filter and analog-to-digital converter (ADC) in the voice signal acquisition stage, and is recognized and responded to by the voice control system of the terminal device, thereby achieving control of the device. The specific process is as follows: Figure 1 shown.

[0004] Regarding the security issues of speaker recognition systems, current research only focuses on how to distinguish between replayed sounds and real sounds. The solution is generally to extract sound features such as silence features, Mel-Frequency Cepstrum Coefficients (MFCC), and then train the extracted features using training models such as the GMM-UBM (Gaussian Mixture Model-Universal Background Model) model, feature support vector machine (SVM) or K-nearest neighbors classification algorithm (KNN), and then use the above trained models for classification to distinguish between replayed sounds and real sounds. The above GMM-UBM-based speaker recognition system is Figure 2As shown in FIG. 1 , the system based on SVM for distinguishing between reproduced sound and real sound is as follows: Figure 3 described.

[0005] The existing solutions to the above-mentioned voice security, such as the speaker recognition system (GMM-UBM), or the support vector machine (SVM) based on the Mel-frequency cepstral coefficient (MFCC) feature, or the K-nearest neighbor classification algorithm (KNN), can provide certain security guarantees, but there are still the following problems:

[0006] 1) The existing speaker recognition system based on Gaussian mixture model (GMM-UBM) cannot accurately distinguish between real speech and recorded speech, and thus cannot resist ultrasonic attacks.

[0007] 2) The existing Gaussian mixture model-based speaker recognition system (GMM-UBM) can only work in a quiet environment and cannot work properly in other complex environments.

[0008] 3) The existing support vector machine (SVM) or K-nearest neighbor classification algorithm (KNN) based on Mel-frequency cepstral coefficient (MFCC) features can only process low-quality attack voice information recorded by mobile phones or laptops, and cannot process high-fidelity recording attacks. If the voice information used for the attack is a high-fidelity recording, it cannot be effectively recognized and processed.

[0009] In summary, existing voice security solutions cannot effectively solve the above-mentioned ultrasonic attack problem, resulting in great security risks in user voice interaction. Summary of the invention

[0010] The present invention provides a signal processing method and a terminal device to achieve effective recognition of abnormal voice signals, improve recognition accuracy, and provide security for user voice interaction, thereby enhancing the user's experience.

[0011] The signal processing method comprises:

[0012] Extract feature information from the input signal;

[0013] According to the extracted feature information, it is determined whether the input signal is an abnormal speech signal.

[0014] Preferably, the characteristic information includes energy characteristics of the signal and / or periodic characteristics of the signal.

[0015] Preferably, the energy characteristic includes a short-time energy characteristic; and / or the period characteristic includes a short-time zero-crossing rate characteristic.

[0016] Preferably, determining whether the input signal is an abnormal speech signal according to the extracted feature information includes:

[0017] Based on the extracted feature information, it is determined whether the input signal is an abnormal speech signal through machine learning.

[0018] Preferably, extracting characteristic information from the input signal includes:

[0019] Extracts characteristic information from a set frequency range of an input signal.

[0020] Preferably, determining whether the input signal is an abnormal speech signal according to the extracted feature information includes:

[0021] According to the changes of the extracted feature information, it is determined whether the input signal is an abnormal speech signal.

[0022] Preferably, the characteristic information includes energy characteristics of the signal and / or directional characteristics of the signal.

[0023] Preferably, determining whether the input signal is an abnormal speech signal according to the extracted feature information includes:

[0024] Determine whether the characteristic information matches the wake-up characteristic information corresponding to the current dynamic wake-up instruction; if not, determine that the voice signal is an abnormal voice signal.

[0025] Preferably, the feature information includes Mel-frequency cepstral coefficient (MFCC) features and / or a dynamic verification code.

[0026] Preferably, the characteristic information includes a voice command,

[0027] Determining whether the input signal is an abnormal speech signal based on the extracted feature information includes:

[0028] Determine whether the voice command matches the wake-up word content corresponding to the current dynamic wake-up image; if not, determine that the voice signal is an abnormal voice signal.

[0029] Preferably, the characteristic information includes background noise information;

[0030] Determining whether the input signal is an abnormal speech signal based on the extracted feature information includes:

[0031] According to the background noise information of the signal, it is determined whether the input signal is an abnormal speech signal.

[0032] Preferably, determining whether the input signal is an abnormal speech signal according to the background noise information of the signal includes:

[0033] The current environment information is determined according to the background noise of the signal, and whether it is an abnormal speech signal is determined according to the current environment information.

[0034] Preferably, the characteristic information includes sound source location information;

[0035] Determining whether the input signal is an abnormal speech signal based on the extracted feature information includes:

[0036] It is determined whether the sound source position information extracted from the input voice signal matches the detected user mouth position information; if not, the voice signal is determined to be an abnormal voice signal.

[0037] Preferably, the method further comprises: sending an abnormal voice interference signal within a preset frequency range.

[0038] Preferably, the abnormal voice interference signal includes an ultrasonic random interference signal.

[0039] Preferably, before extracting the characteristic information from the input signal, the method further includes:

[0040] According to the sampling device whose sampling frequency is higher than the set frequency threshold, the input signal is sampled;

[0041] Perform low-pass filtering on the sampled signal.

[0042] Preferably, the method further comprises: when determining that the input signal is an abnormal voice signal, sending corresponding prompt information.

[0043] The present invention also provides a terminal device, comprising:

[0044] Processor; and

[0045] The memory is configured to store machine-readable instructions, which, when executed by the processor, enable the processor to perform the above method.

[0046] The present invention also provides a signal processing method, comprising:

[0047] Determine the control operation corresponding to the input signal;

[0048] According to the control operation corresponding to the input signal, it is determined whether the input signal is an abnormal voice signal.

[0049] Preferably, determining whether the input signal is an abnormal voice signal according to the control operation corresponding to the input signal includes:

[0050] Determine currently feasible control actions;

[0051] If the control operation corresponding to the input signal does not match the currently feasible control operation, it is determined that the input signal is an abnormal speech signal.

[0052] The present invention also provides a terminal device, comprising:

[0053] Processor; and

[0054] The memory is configured to store machine-readable instructions, which, when executed by the processor, enable the processor to perform the above method.

[0055] The present invention also provides a signal processing method, comprising:

[0056] Receive input signal;

[0057] Send abnormal voice interference signals within the preset frequency range.

[0058] Preferably, the abnormal voice interference signal includes an ultrasonic random interference signal.

[0059] The present invention also provides a terminal device, comprising:

[0060] Processor; and

[0061] The memory is configured to store machine-readable instructions, which, when executed by the processor, enable the processor to perform the above method.

[0062] The present invention also provides a signal processing method, comprising:

[0063] Receive input signal;

[0064] According to the sampling device whose sampling frequency is higher than the set frequency threshold, the input signal is sampled;

[0065] Perform low-pass filtering on the sampled signal.

[0066] The present invention also provides a terminal device, comprising:

[0067] Processor; and

[0068] The memory is configured to store machine-readable instructions, which, when executed by the processor, enable the processor to perform the above method.

[0069] In the present invention, by determining whether the input signal is an abnormal voice signal based on feature information extracted from the input signal, effective recognition of abnormal voice signals is achieved, the accuracy of recognition is improved, and security is provided for the user's voice interaction, thereby enhancing the user's usage experience. BRIEF DESCRIPTION OF THE DRAWINGS

[0070] Figure 1 It is a schematic diagram of the process of ultrasonic attack principle in the prior art;

[0071] Figure 2 It is a flowchart of the processing of the GMM-UBM speaker recognition system in the prior art;

[0072] Figure 3 It is a flow chart of a system process for distinguishing between retransmitted sound and real sound based on SVM in the prior art;

[0073] Figure 4 It is a flow chart of identifying abnormal speech signals based on machine learning provided by the present invention;

[0074] Figure 5 It is a schematic diagram comparing the spectrogram information of a normal speech signal and an abnormal speech signal within a fixed frequency range provided by the present invention;

[0075] Figure 6 It is a schematic diagram comparing characteristic information of a normal speech signal and an abnormal speech signal provided by the present invention;

[0076] Figure 7 A schematic diagram of voice signal recognition for transforming a wake-up instruction provided by the present invention;

[0077] Figure 8 A schematic diagram of sound collection by the terminal device provided by the present invention;

[0078] Fig. 9 A schematic diagram of identifying abnormal speech signals according to feature changes provided by the present invention;

[0079] Fig.10 A schematic diagram of the spectrum at different sampling frequencies provided by the present invention;

[0080] Fig.11 A schematic diagram of the high sampling frequency analog-to-digital conversion process provided by the present invention;

[0081] Fig.12 A schematic flow chart of a first signal processing method provided by the present invention;

[0082] Fig.13 A schematic diagram of the framework of machine learning for identifying abnormal speech signals provided by the present invention;

[0083] Fig.14 A schematic diagram of displaying a dynamic wake-up instruction provided in the second embodiment of the present invention;

[0084] Fig.15 A schematic diagram of a wake-up process provided in Embodiment 2 of the present invention;

[0085] Fig.16 A schematic diagram of a flow chart for comprehensively judging abnormal voice signals according to user behavior provided in Embodiment 7 of the present invention;

[0086] Fig.17 A schematic diagram of a flow chart of determining an abnormal voice signal according to surrounding environment information provided in Embodiment 8 of the present invention;

[0087] Fig.18 A schematic diagram of the structure of the terminal device provided by the present invention;

[0088] Fig.19 A schematic flow chart of a second signal processing method provided by the present invention;

[0089] Fig. 20 A schematic flow chart of a third signal processing method provided by the present invention;

[0090] Fig.21 A schematic flow chart of a fourth signal processing method provided by the present invention;

[0091] Fig. 22 A schematic block diagram of a computing system provided by the present invention;

[0092] Fig.23 A schematic diagram of voice signal recognition combined with dynamic verification code verification wake-up instruction provided by the present invention;

[0093] Fig.24 A schematic diagram of voice signal recognition combined with dynamic verification code verification voice command provided by the present invention;

[0094] Fig.25 A schematic diagram of embedding voice commands into the dynamic wallpaper provided by the present invention;

[0095] Fig.26 A schematic diagram of determining the mouth position by the position of a handheld terminal device provided by the present invention;

[0096] Fig. 27 The present invention provides a flow chart of a method for verifying a voice command by combining a dynamic verification code. DETAILED DESCRIPTION

[0097] The present invention provides a signal processing method and a terminal device. The specific implementation manner of the present invention is described in detail below with reference to the accompanying drawings.

[0098] Embodiments of the present invention are described in detail below, examples of which are shown in the accompanying drawings, wherein the same or similar reference numerals throughout represent the same or similar elements or elements having the same or similar functions. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention, and cannot be interpreted as limiting the present invention.

[0099] It will be understood by those skilled in the art that, unless expressly stated, the singular forms "one", "said", and "the" used herein may also include plural forms. It should be further understood that the term "comprising" used in the specification of the present invention refers to the presence of the features, integers, steps, operations, elements and / or components, but does not exclude the presence or addition of one or more other features, integers, steps, operations, elements, components and / or groups thereof. It should be understood that when we refer to an element as being "connected" or "coupled" to another element, it may be directly connected or coupled to the other element, or there may be intermediate elements. In addition, the "connection" or "coupling" used herein may include wireless connection or wireless coupling. The term "and / or" used herein includes all or any unit and all combinations of one or more associated listed items.

[0100] It will be understood by those skilled in the art that, unless otherwise defined, all terms (including technical and scientific terms) used herein have the same meaning as generally understood by those skilled in the art in the art to which the present invention belongs. It should also be understood that terms such as those defined in general dictionaries should be understood to have meanings consistent with the meanings in the context of the prior art, and will not be interpreted with idealized or overly formal meanings unless specifically defined as herein.

[0101] For the signal processing method to be implemented by the present invention, the present invention provides the following processing methods, and explains each processing method accordingly:

[0102] 1) How to distinguish whether it is a normal speech signal based on the characteristics of the signal

[0103] In the voice interaction state, the signal received by the terminal device may be a voice signal or an abnormal voice signal such as an ultrasonic signal. Figure 1 The speech signal acquisition stage is processed in the process, and then the set feature information can be extracted from the processed speech signal. The set feature information may include energy features for characterizing signal energy and / or period features for characterizing signal period. According to the extracted feature information, it can be distinguished whether it is a normal speech signal, that is, whether it is real speech.

[0104] For example, Figure 4As shown, during the voice signal collection stage of the terminal device, a signal is collected, and a bandpass filter is used to extract a voice signal within a set frequency range (such as 2000 Hz to 3000 Hz) from the voice signal obtained after analog-to-digital conversion (ADC), and feature information of the voice signal is extracted. A machine learning method is used to identify whether the voice signal is a real normal voice signal or an abnormal voice signal demodulated for attack based on the extracted voice signal feature information.

[0105] Among them, the energy feature of the above-mentioned speech signal can be short-time energy, and its periodic feature can be short-time zero-crossing rate; of course, for the energy feature, it is not limited to short-time energy, and any feature information that can represent the energy feature of the signal is within the protection scope of the present invention; for the periodic feature, it is not limited to short-time zero-crossing rate, and any feature information that can represent the periodic feature of the signal is within the protection scope of the present invention.

[0106] In this treatment method, if Figure 5 The figure shows the comparison of the spectrogram information of the normal speech signal and the abnormal speech signal used for attack at 2000 Hz and 3000 Hz. In the figure, the A position and the B position of the horizontal axis are two corresponding positions selected in the spectrogram of the normal speech signal and the abnormal speech signal used for attack, respectively. The vertical axis is the brightness value, where the higher the brightness value of the color, the stronger the energy. From the comparison of the color brightness values ​​at the same position in the figure, it can be seen that the energy of normal speech is significantly higher than that of attacking speech. Therefore, for the input speech information after analog-to-digital conversion, it is necessary to use Figure 4 The bandpass filter in is used to extract the speech signal whose frequency range is 2000 Hz to 3000 Hz.

[0107] Among them, for the use Figure 4 The bandpass filter in the above example extracts the normal speech signal (frequency range is 2000 Hz to 3000 Hz) and the abnormal speech signal used for attack, whose characteristic information includes short-time zero-crossing rate and short-time energy, such as Figure 6 As shown in the figure, it can be seen that the difference between short-time energy and short-time zero-crossing rate is quite obvious. The short-time energy (0~5x10 -6 ) is significantly lower than the short-term energy of normal speech signals (0 to 1x10 -4 ), while the short-term zero-crossing rate of the abnormal speech signal used for attack is about 0.1, which is significantly higher than the short-term zero-crossing rate of the normal speech signal of about 0.06.

[0108] Among them, the short-time zero-crossing rate indicates the number of times the speech signal waveform in a frame of speech crosses the horizontal axis (zero level), reflecting the periodic characteristics of speech. The short-time energy indicates the energy of a frame of speech signal.

[0109] Based on the feature information of the extracted signal, machine learning is used to identify whether the voice signal is a real normal voice signal or an abnormal voice signal demodulated for attack. Machine learning includes but is not limited to support vector machine (SVM), deep neural network (DNN) and K nearest neighbor classification algorithm (KNN). The overall block diagram of using machine learning to identify abnormal voice signals used for attack is shown below. Figure 4 As shown. The role of the machine learning classifier is to identify whether the voice signal is an abnormal voice signal used for attack based on the extracted signal features (periodic features and / or energy features). If the voice signal is judged to be a real normal voice signal, the corresponding voice command is executed. If the voice signal is judged to be an abnormal voice signal used for attack, the corresponding voice command is ignored or a corresponding prompt is given to the user.

[0110] 2) Processing method based on dynamic wake-up instructions

[0111] At present, the voice wake-up commands of terminal devices are fixed. For example, when a user inputs the wake-up word 'HeyXXX' or 'Hi XXX' to the terminal device, the terminal device will be awakened and enter the voice interaction state after detecting that the wake-up word entered by the user is the preset wake-up word. These wake-up words are originally defined by the product and cannot be changed.

[0112] The attacker can easily obtain the voice files of these wake-up words, for example, through recording or speech synthesis. Therefore, even if the terminal device is next to the user, the attacker can wake up the user's terminal device through inaudible ultrasound.

[0113] In view of the above problems, the present invention proposes to provide users with dynamic wake-up instructions, such as Figure 7 Example shown.

[0114] Dynamic wake-up commands mean that the voice wake-up commands are changing, and only the user knows about the changes. In this way, attackers cannot obtain the wake-up commands of the user's device, and cannot wake up the user's terminal device through ultrasound.

[0115] 3) Distinguish attack signals by changes in signal characteristics

[0116] Usually when a user uses a terminal device, the distance between the user and the terminal device is very short. When the user holds the terminal device, a slight change in the position between the user and the terminal device will have a significant impact on the characteristics of the voice signal (directional characteristics and / or energy characteristics). The sound used for attack is slightly far away from the terminal device, and the placement of the ultrasonic signal playback device is generally fixed. Even if it changes, the characteristics of the signal can be considered unchanged because it is slightly far away from the terminal device. Figures 8-9 shown. Figure 8 In the example, it is assumed that the terminal device has two sound collection devices, such as sound collection point 1 and sound collection point 2 in the figure. The user's sound source is close to the terminal device, and the position change between the sound source and the terminal device has a greater impact on the characteristics of the signal it propagates. The attack sound source is far away from the terminal device, and the position change between the sound source and the terminal device has little impact on the characteristics of the signal it propagates. Fig. 9 In the present invention, the characteristics of the user's normal voice signal can change, while the characteristics of the voice signal used for attack are fixed. Therefore, the terminal device can distinguish the attack signal from the user's normal voice signal by detecting the changes in the characteristics.

[0117] In order to change the directional characteristics of the signal when the user inputs a voice command, the user can change the directional characteristics by shaking the terminal device or changing the location of the sound source when inputting a voice command to the terminal device. In addition, the energy characteristics can also be changed by shaking the terminal device or making gestures (such as waving hands). However, attackers often do not know the gestures and therefore cannot change the characteristics of the signal.

[0118] The device can extract energy features and / or directional features of the voice signal after analog-to-digital conversion, and then use machine learning to distinguish between abnormal voice signals used for attack and normal voice signals of users based on changes in the energy features and / or directional features of the signals.

[0119] 4) Use an ADC (analog-to-digital converter) with a higher sampling frequency

[0120] From the experimental phenomenon, it is found that when the sampling frequency (Fs) of the recording pen is 44.1KHz, the ultrasonic attack signal can be received and analyzed normally. When the sampling frequency is 96KHz, the ultrasonic attack signal cannot be analyzed.

[0121] Since most terminal devices support a sampling frequency of 44.1KHz, simulation results show that a 44.1KHz sampling frequency will cause spectrum leakage and superposition, thus mixing the ultrasonic attack signal into the voice command, such as Fig.10 The spectrum diagram of different sampling frequencies is shown in .

[0122] According to the Nyquist sampling criterion, the sampling frequency Fs ≥ 2*Fh (the highest input frequency), if the ADC sampling frequency of 96KHz is used, the ultrasonic attack signal of up to 48KHz can be sampled. Due to the cost and equipment size of sending ultrasonic signals above 25KHz, it is difficult to perform an actual attack operation. Therefore, in the voice signal acquisition stage, the use of a high-frequency 96KHz ADC sampling frequency will suppress spectrum aliasing. This provides conditions for further filtering out ultrasonic attack signals. The specific implementation plan is as follows: Fig.11 As shown, the analog-to-digital sampling frequency is increased to 96KHz, and then passes through a digital low-pass filter. Of course, the software also needs to support a sampling frequency of 96KHz.

[0123] 5) Use interference signal

[0124] When the terminal device detects a signal input, it can make the speaker on the device side hop or send an ultrasonic random noise signal at full frequency within a set frequency range (such as 20KHz to 48KHz), and then detect the input voice command. If it is a normal voice command input, the terminal device can correctly parse it. Because normal voice commands and the ultrasonic random noise signals of the speaker are at different frequencies, they are not interfered by the ultrasonic random noise signals. On the contrary, if it is an ultrasonic attack signal, the ultrasonic voice command used for the attack cannot be parsed normally. The ultrasonic attack signal will interfere with the ultrasonic random noise signal sent by the speaker of the terminal device. Finally, the signal received by the terminal device is the superposition of the ultrasonic random noise signal and the ultrasonic attack signal. The terminal device is ultimately unable to correctly parse the ultrasonic attack signal, thereby achieving the purpose of defending against ultrasonic attacks. In order to save power consumption, this solution can be used in the working state after wake-up.

[0125] In addition, since the ultrasonic random noise signal may affect other terminal devices, the transmitted ultrasonic random noise signal should reduce the output power appropriately to prevent ultrasonic pollution to other terminal devices. In order to save power as much as possible, the terminal device speaker should try to use frequency hopping or full-frequency output within the device frequency range (such as 20KHz to 48KHz).

[0126] 6) Comprehensive judgment based on surrounding environment information and learned user behavior

[0127] The voice control system comprehensively judges whether the current input voice command content is normal user usage behavior based on the user's daily usage behavior and surrounding environmental information learned by the voice control system. If the current input signal content is judged to be abnormal user usage behavior, new verification information is required to be entered, such as requiring special instructions to be spoken, or operating the terminal device as required for verification. Only after passing this additional verification process will the corresponding instruction be executed.

[0128] 7) Combine dynamic verification code for wake-up and recognition

[0129] Current terminal devices use fixed voice commands. For example, when a user inputs "Hello XXX" or "Hello, XXX" into the terminal device, the terminal device will be awakened after detecting that the wake-up word entered by the user is a preset wake-up word. After the terminal device is awakened, when the user inputs "Open XXX application" or "Call XXX" into the terminal device, the terminal device recognizes the correct voice content and executes the corresponding voice interaction command.

[0130] An attacker can use inaudible ultrasound to wake up the user's terminal device or send execution instructions to the user's terminal device.

[0131] In view of the above problems, the present invention proposes a method for waking up by combining a dynamic verification code, such as Fig.23 As shown, when the user inputs the voice command "Hi XX", the terminal device detects that the current voice command is consistent with the preset wake-up command, and then the terminal device screen will immediately display a dynamic verification code (such as a random number string, i.e., a random verification code), such as "015". The user then inputs the voice or text command "015", and the terminal device detects whether the verification code input by the user is consistent with the generated dynamic verification code. If they are consistent, the voice input by the user can be considered to be a normal voice signal, and the device can be successfully awakened. If they are inconsistent, the voice input by the user can be considered to be an abnormal voice signal. In addition, the terminal device can also display the dynamic verification code to the user before waking up. When waking up the terminal device, the user can enter the wake-up command and the dynamic verification code at the same time or successively. When the terminal device detects that the input wake-up command is consistent with the preset wake-up command, and the input dynamic verification code is consistent with the generated dynamic verification code, the voice input by the user can be considered to be a normal voice signal, and the terminal device can be successfully awakened.

[0132] The present invention also proposes a method for executing a voice command in combination with a dynamic verification code, such as Fig.24 After the terminal device wakes up, a random string, such as "756", will be displayed on the voice command input interface. The user can enter the execution command and the dynamic verification code simultaneously or successively. The terminal device can only execute the user's command when it detects that the entered dynamic verification code is consistent with the generated dynamic verification code. For example, the user can successfully open the XX application by entering "open XX application 756".

[0133] In the above scheme, the dynamic verification code can be regarded as the characteristic information of the voice signal. The terminal device can extract the dynamic verification code input by the user from the received voice signal, and then determine whether the extracted dynamic verification code matches the generated dynamic verification code (that is, the wake-up characteristic information corresponding to the current dynamic wake-up instruction); if there is no match, it is determined that the input voice signal is an abnormal voice signal.

[0134] 8) Dynamic wallpaper embedded with voice commands

[0135] In order to solve the problem that the terminal device has a single wake-up command and is easily awakened by attackers, the present invention proposes a method for waking up by embedding a dynamic wallpaper with a voice command. The user can select several images as dynamic wallpapers for the mobile phone and register a voice command for each wallpaper. Fig.25 As shown, the wallpaper of the terminal device will automatically change when it is in use. When the user wants to wake up the terminal device, he can enter the voice command corresponding to the current wallpaper. When the terminal device detects that the voice command entered by the user is consistent with the voice command (i.e., the content of the wake-up word) corresponding to the current wallpaper (also called a dynamic wake-up image), the terminal device is determined to be a normal voice signal, and the terminal device can be awakened. If they are inconsistent, the input voice signal is determined to be an abnormal voice signal. For example, when the current mobile phone wallpaper displays picture 2, and the user's voice command received is "Hi, little dog", the terminal device will be awakened.

[0136] 9) Determine mouth position and sound source location

[0137] The source of the voice command sound should be the user's mouth. The user's mouth position can be detected by using a camera or determined by the position of the user's handheld terminal device, such as Fig.26 As shown in the figure, by the position of the handheld terminal device, it can be determined that the voice command should come from the back, front, and left side of the terminal device respectively. The sound source localization method can determine the sound point of the sound command, and determine whether the voice command comes from the real user or the attacker by judging whether the mouth position is consistent with the sound source position.

[0138] Based on the several processing methods provided above, the present invention provides a speech signal processing method, such as Fig.12 As shown, including:

[0139] Step 1201, extracting feature information from an input signal.

[0140] The feature information is extracted from the input signal, including:

[0141] Extracts characteristic information from a set frequency range of an input signal.

[0142] The above-mentioned input signal may represent a speech signal obtained after being processed in the signal acquisition stage.

[0143] Before this step, it also includes:

[0144] According to the sampling device whose sampling frequency is higher than the set frequency threshold, the input signal is sampled;

[0145] Perform low-pass filtering on the sampled signal.

[0146] The sampling device may be, but is not limited to, an ADC.

[0147] Step 1202: Determine whether the input signal is an abnormal speech signal based on the extracted feature information.

[0148] The process of determining whether the signal is abnormal based on the extracted feature information includes the following processing methods:

[0149] (1) The process of determining an abnormal voice signal includes:

[0150] Based on the extracted feature information, it is determined whether the input signal is an abnormal speech signal through machine learning.

[0151] The above-mentioned characteristic information includes energy characteristics of the signal and / or periodic characteristics of the signal.

[0152] The energy characteristic specifically includes a short-time energy characteristic; the period characteristic specifically includes a short-time zero-crossing rate characteristic.

[0153] Furthermore, when determining whether the input signal is an abnormal speech signal by machine learning, it can be determined based on energy features only, or based on period features only, or of course it can be determined based on a combination of energy features and period features.

[0154] The energy feature and / or the periodic feature of the signal may be extracted within a set frequency range (eg, a frequency range of 2000 Hz to 3000 Hz).

[0155] (2) The process of determining the abnormal voice signal includes:

[0156] According to the changes of the extracted feature information, it is determined whether the input signal is an abnormal speech signal.

[0157] The characteristic information includes energy characteristics of the signal and / or directional characteristics of the signal.

[0158] (3) The process of determining the abnormal voice signal includes:

[0159] Determine whether the characteristic information matches the wake-up characteristic information corresponding to the current dynamic wake-up instruction; if not, determine that the voice signal is an abnormal voice signal.

[0160] Preferably, the wake-up feature information includes Mel-frequency cepstral coefficient (MFCC) features and / or a dynamic verification code.

[0161] (4) The extracted feature information includes background noise information;

[0162] The process of determining an abnormal voice signal includes:

[0163] According to the background noise information of the signal, it is determined whether the input signal is an abnormal speech signal.

[0164] Wherein, determining whether the input signal is an abnormal speech signal according to the background noise information of the signal includes:

[0165] The current environment information is determined according to the background noise of the signal, and whether it is an abnormal speech signal is determined according to the current environment information.

[0166] If the current environment information does not match the environment information corresponding to the control operation corresponding to the input signal, the input signal may be considered to be an abnormal voice signal.

[0167] (5) The extracted feature information includes voice commands,

[0168] The process of determining an abnormal voice signal includes:

[0169] It is possible but not limited to determining whether the voice command matches the wake-up word content corresponding to the current dynamic wake-up image by using keyword detection technology; if not, determining that the voice signal is an abnormal voice signal.

[0170] (6) The extracted feature information includes the location information of the sound source;

[0171] The process of determining an abnormal voice signal includes:

[0172] It is determined whether the sound source position information extracted from the input voice signal matches the detected user mouth position information; if not, the voice signal is determined to be an abnormal voice signal.

[0173] Of course, when determining an abnormal voice signal, the following processing can also be used to interfere with the abnormal voice signal, thereby achieving the effect of directly filtering out the abnormal voice signal, that is, sending an abnormal voice interference signal within a preset frequency range, so as to interfere with the abnormal voice signal through the abnormal voice interference signal, thereby destroying the voice of the abnormal voice signal and making it unable to be recognized by the terminal device.

[0174] The abnormal voice interference signal includes an ultrasonic random interference signal. The abnormal voice interference signal is not limited to the ultrasonic random interference signal. Any signal that can interfere with the voice of the abnormal voice signal and thus destroy the voice is within the protection scope of the present invention.

[0175] Furthermore, when it is determined that the input signal is an abnormal voice signal, a corresponding prompt message is sent.

[0176] With respect to the signal processing method and corresponding processing modes provided by the present invention, each of the above-mentioned processing modes is specifically explained in the following in the form of specific embodiments.

[0177] Embodiment 1

[0178] The embodiment of the present invention discloses a method for identifying abnormal voice signals used for attack by extracting voice signal features and distinguishing the voice signal features using machine learning. The detailed system block diagram of this embodiment is shown in FIG. Fig.13 shown.

[0179] step:

[0180] 1. Prepare a large amount of normal speech signals and abnormal speech signal data used for attack to train the network model and classifier of machine learning. The larger the amount of data, the more stable the network model and the more accurate the classifier can be.

[0181] 2. Through the bandpass filter, the normal voice signal and the demodulated abnormal voice signal used for attack are extracted, and the frequency range of the voice information is 2000 Hz to 3000 Hz.

[0182] 3. Extract the short-time zero-crossing rate and short-time energy characteristics of normal speech signals and abnormal speech signals after passing through the bandpass filter.

[0183] 4. The extracted speech feature information (short-time zero-crossing rate and short-time energy features) is input into the machine learning network for network model training and classifier training until the machine learning classifier can accurately identify abnormal speech information and normal speech information used for attack that are similar to the trained data.

[0184] 5. Use the trained machine learning classifier to collect the voice information collected by the terminal device microphone as the Fig.13The test voice in the image is limited to a frequency range of 2000 Hz to 3000 Hz by a bandpass filter, and feature information of the test voice information including short-time zero-crossing rate and short-time energy is further extracted. The feature information is then directly input into a trained machine learning classifier to determine the category of the voice information collected by the microphone, that is, to determine whether the test voice information is normal voice or abnormal voice used for attack.

[0185] Embodiment 2

[0186] Dynamic wake-up instructions are generated and notified to the user.

[0187] Dynamic instructions can be customized by users. For example, users can set the words or sentences for wake-up instructions themselves. The system can also set a set of user wake-up words or sentences by default. The system can also detect the user's preferences and concerns during this period, and then extract keywords as wake-up instructions.

[0188] Currently, many terminal devices can display floating information on the screen when in standby mode. Fig.14 As shown, the wake-up word can be displayed on the screen in standby mode or on the screen of a wearable device connected thereto. The user sees the current wake-up command and then wakes up the terminal device by reading the wake-up command.

[0189] For example, the terminal device currently randomly selects one of the previously defined wake-up words and displays it on the screen. When the user sees the wake-up word displayed on the screen, he / she speaks the current wake-up word.

[0190] The terminal device extracts MFCC features from the voice signal after the voice signal acquisition stage, and matches the extracted features with the wake-up features (MFCC features) corresponding to the current dynamic wake-up command (such as the wake-up word displayed on the screen). If there is no match, the voice signal is determined to be an abnormal voice signal, and the device remains in standby mode or gives a corresponding prompt message. If there is a match, the voice signal is a normal voice signal and the screen is woken up. Fig.15 As shown, the terminal device first performs wake-up word detection in the above manner. If a normal voice signal is detected, speaker recognition is performed. If the recognition passes, the screen is woken up.

[0191] Embodiment 3

[0192] Attack signals are distinguished by changes in signal characteristics.

[0193] This method allows the user to change the characteristics of the user's voice signal by slightly shaking the terminal device left and right, up and down, or adding gestures between the terminal device and the user when initiating a voice command to the terminal device. The terminal device distinguishes whether it is a normal voice signal based on the change in the signal characteristics, and only responds to the voice command if it is a normal voice signal.

[0194] In this method, the characteristics of the input voice signal will change regularly. If the terminal device detects that the change pattern of the characteristics of the input voice signal is consistent with the set change pattern, the voice signal is considered to be a normal voice signal, otherwise, it is considered to be an abnormal voice signal.

[0195] It is also possible to pre-train a machine learning network model for the above-mentioned change pattern, and distinguish whether it is a normal voice signal based on the characteristics of the input voice signal and the trained machine learning network model.

[0196] Embodiment 4

[0197] The embodiment of the invention discloses a method for preventing ultrasonic wave attacks, which includes hardware and software prevention measures.

[0198] The purpose is to ensure the user's information security and prevent potential malicious attacks.

[0199] step:

[0200] 1. During the voice signal acquisition phase, the ultrasonic signal is sampled by a higher frequency ADC during the voice sampling process of the electromechanical microphone, thereby effectively suppressing spectrum aliasing. If an attacker uses a 25KHz ultrasonic signal to attack the device, and the terminal device uses a 96KHz ADC for sampling, spectrum aliasing will not occur.

[0201] 2. Since there is no spectrum aliasing, the ultrasonic attack signal is completely retained in the corresponding frequency band. The mixed ultrasonic attack signal can be filtered out by the digital LPF (Low-Pass Filter), thereby ensuring the safety of the equipment. For the specific implementation process, see Fig.11 The high sampling frequency analog-to-digital conversion flow chart is shown. Through the method of interfering with ultrasonic attack disclosed in the embodiment of the present invention, the user's information security is guaranteed and potential malicious attacks are prevented.

[0202] Embodiment 5

[0203] In this embodiment, the purpose of attack defense is achieved by adopting a frequency hopping method to send ultrasonic random noise signals.

[0204] step:

[0205] 1. Wait for the voice input prompt or microphone (Microphone, Mic) working prompt.

[0206] 2. If there is a voice input prompt, the speaker is triggered to hop frequency to emit random noise. The frequency hopping range is (20KHz ~ 48KHz) to send ultrasonic random noise signals, interfering with the ultrasonic signal in the frequency band.

[0207] 3. Since the human voice signal and the ultrasonic random noise signal emitted by the speaker are not in the same frequency band, the human voice signal will not be interfered by the ultrasonic random noise signal emitted by the speaker, and the normal instructions of the human voice can be analyzed normally by the device.

[0208] 4. If the input is an ultrasonic attack signal, it will be interfered by the ultrasonic random noise signal with a frequency hopping range of (20KHz ~ 48KHz), which will interfere with the demodulation of the attack signal and destroy the voice of the attack signal, making it impossible for the device to recognize it, thereby achieving the purpose of attack defense.

[0209] In the above embodiment, the ultrasonic random noise signal sending method is exemplified as follows: within the frequency band range of 20KHz to 48KHz, firstly, a frequency of 20KHz to 24KHz is selected to send an ultrasonic random noise signal, and then a frequency of 44KHz to 48KHz is selected to send an ultrasonic random noise signal, and then a frequency of 24KHz to 28KHz is selected to send an ultrasonic random noise signal, and this cycle is repeated.

[0210] Embodiment 6

[0211] Similar to the fifth embodiment above, in this embodiment, in order to reduce the complexity of software and hardware, an ultrasonic random noise signal is sent within the entire ultrasonic frequency band (20 KHz to 48 KHz).

[0212] step:

[0213] 1. Wait for the voice input prompt or the microphone working prompt.

[0214] 2. If there is a voice input prompt, the speaker is triggered to emit noise. The frequency range of the noise is (20KHz ~ 48KHz) to send ultrasonic random noise signals to interfere with the ultrasonic signals in the frequency band.

[0215] 3. Since the human voice signal and the ultrasonic random noise signal emitted by the speaker are not in the same frequency band, the human voice signal will not be interfered by the ultrasonic random noise signal emitted by the speaker, and the normal instructions of the human voice can be analyzed normally by the device.

[0216] 4. If the input is an ultrasonic attack signal, it will be interfered by the ultrasonic random noise signal within the frequency band (20KHz~48KHz), which will interfere with the demodulation of the attack signal and destroy the voice of the attack signal, making it impossible for the device to recognize it, thereby achieving the purpose of attack defense.

[0217] In the above embodiment, the ultrasonic random noise signal sending method is exemplified as follows: within the frequency band range of 20KHz to 48KHz, a frequency band of 20KHz to 48KHz is selected to send an ultrasonic random noise signal, and after 5ms, the frequency band of 20KHz to 48KHz is selected again to send an ultrasonic random noise signal, and this cycle is repeated.

[0218] Embodiment 7

[0219] This embodiment discloses a method for comprehensively judging whether the current voice command is an abnormal voice used for attack by learning user behavior through a voice assistant.

[0220] Determine the control operation corresponding to the input signal, and determine whether the input signal is an abnormal voice signal according to the control operation corresponding to the input signal. Wherein, the currently feasible control operation can be determined, and if the control operation corresponding to the input signal does not match the currently feasible control operation, then determine that the input signal is an abnormal voice signal.

[0221] The currently feasible control operation may be determined based on, but not limited to, user behavior (such as historical user operation behavior).

[0222] like Fig.16 The specific steps include:

[0223] While the user is using the voice control system normally, the voice control system will conduct comprehensive learning of the user's usage behavior.

[0224] When the voice assistant receives a new command, it comprehensively identifies the current command based on the user's usage behavior learned by the voice assistant to determine whether it is normal usage behavior.

[0225] If the current user behavior is identified as normal usage behavior, the current voice signal is considered to be a real normal voice signal, and the instruction operation is executed.

[0226] If the current user behavior is identified as abnormal use behavior, the user may be required to perform additional verification information verification, including but not limited to speaking special instructions, or performing terminal device operations as required. If the additional verification information verification is passed, the current voice signal is considered to be a real voice signal and the corresponding instruction operation is executed. Otherwise, it is considered to be an abnormal voice signal used for attack, and the voice instruction is ignored.

[0227] For example, when a user transfers money to others through a voice assistant, they usually make a phone call to confirm, and then use the voice assistant to issue the command "transfer money to someone". When others use abnormal voice signals used for attack, they may directly use the voice assistant to issue the command "transfer money to someone". Because there is no process of making a phone call, this behavior is regarded as abnormal user behavior, and additional verification is required, such as saying the phrase displayed on the screen to the terminal device at this time or shaking the terminal device as required, etc. If the verification information cannot be completed, the command is regarded as an abnormal voice signal used for attack, and the command "transfer money to someone" is directly ignored. If the verification information can be successfully completed, the command is regarded as a real normal voice signal, and the command "transfer money to someone" is executed.

[0228] Embodiment 8

[0229] This embodiment discloses a method for comprehensively judging whether a current voice command is an attack voice based on surrounding environment information through a voice assistant learning, such as Fig.17 shown.

[0230] step:

[0231] When the user is using the voice assistant normally, the voice assistant will judge the surrounding environmental information (which can be determined based on background noise), combine the user's usage behavior with the surrounding environmental information, and conduct comprehensive learning.

[0232] When the voice assistant receives a new command, it will comprehensively identify the current command based on the surrounding environment information to determine whether it is a normal usage behavior.

[0233] If the current user behavior is identified as normal usage behavior, the current voice signal is considered to be a real normal voice signal, and the instruction operation is executed.

[0234] If the current user behavior is identified as abnormal use behavior, the user is required to perform additional verification information verification, including but not limited to speaking special instructions, or performing terminal device operations as required. If the additional verification information verification is passed, the current voice signal is considered to be a real normal voice signal and the command operation is executed. Otherwise, it is considered to be an abnormal voice signal used for attack, and the voice command is ignored.

[0235] For example, when a user transfers money to others through a voice assistant, it is usually done in a relatively quiet environment. When others use abnormal voice signals for attack, the occasion may be a relatively noisy environment such as a subway or station, and the background noise is also relatively noisy. At this time, the voice instruction for transfer is regarded as abnormal user behavior, and additional verification is required, such as saying the phrase displayed on the screen to the terminal device at this time or shaking the terminal device as required, etc. If the verification information cannot be completed, the voice instruction is regarded as an abnormal voice signal for attack, and the voice instruction for transfer is directly ignored. If the verification information can be successfully completed, the voice instruction is regarded as a real normal voice signal, and the subsequent transfer operation is performed.

[0236] Embodiment 9

[0237] This embodiment discloses a method for preventing attack voice by using a dynamic verification code, such as Fig. 27 shown.

[0238] step:

[0239] After the terminal device receives the voice command, the voice wake-up module or the voice assistant performs keyword detection (also called command word detection) to determine whether the voice command is a preset wake-up word or a voice command executable by the terminal device;

[0240] If the detected voice command is not a preset wake-up word, nor a voice command that can be executed by the voice assistant, it is judged to be an abnormal voice and may be an attack signal;

[0241] If the detected voice command is a preset wake-up word, or a voice command executable by the voice assistant, the terminal device screen will display a dynamic verification code (which can be a random verification code);

[0242] The user reads the dynamic verification code displayed on the terminal device;

[0243] The terminal device detects the dynamic verification code read by the user. If the dynamic verification code read by the user is consistent with the dynamic verification code displayed on the screen, it is judged as a real normal voice signal and executes the command operation (such as wake-up). Otherwise, it is judged as an abnormal voice and may be an attack signal.

[0244] Embodiment 10

[0245] This embodiment discloses a method for preventing attacking voice by embedding voice commands into dynamic wallpaper.

[0246] step:

[0247] Users can select several images and register their own voice commands for each image;

[0248] The terminal device sets the selected image as a dynamic wallpaper of the terminal device;

[0249] When the user wants to wake up the terminal device, he can issue the voice command registered by the current dynamic wallpaper;

[0250] Perform keyword detection on the user's voice command. If the detected content is consistent with the registered command, it is judged as real voice and wakes up the terminal device. Otherwise, it is judged as attack voice.

[0251] Embodiment 11

[0252] This embodiment discloses a method for determining whether the speech is an attacking speech by tracking the mouth position and sound source localization.

[0253] The source of the voice command sound should be the user's mouth. The sound source point of the voice can be determined by sound source positioning. If the mouth position is consistent with the sound source position, it is judged to be normal voice, otherwise it is abnormal voice.

[0254] step:

[0255] After the input speech signal is detected, the sound source localization algorithm can be used to obtain the sound source position of the received speech signal, that is, to extract the sound source position information from the input speech signal;

[0256] The terminal device can also determine the user's mouth position. The user's mouth position can be directly detected by a camera, or by the user holding the terminal device, or by gravity sensing to determine the face position, thereby obtaining the user's voice direction;

[0257] Determine whether the mouth position is consistent with the sound source position. If the positions are consistent, it is determined that the voice command comes from a real user and the input voice signal is a normal voice signal. Otherwise, it is an attack voice and the input voice signal is an abnormal voice signal.

[0258] Based on the voice signal processing method provided by the present invention, the present invention also provides a terminal device, such as Fig.18 As shown, including:

[0259] Processor 1801; and

[0260] The memory 1802 is configured to store machine-readable instructions, which, when executed by the processor, enable the processor to perform the above method.

[0261] The present invention also provides a signal processing method, such as Fig.19 As shown, including:

[0262] 1901, determine a control operation corresponding to an input signal.

[0263] 1902. Determine whether the input signal is an abnormal speech signal according to a control operation corresponding to the input signal.

[0264] Preferably, determining whether the input signal is an abnormal voice signal according to the control operation corresponding to the input signal includes:

[0265] Determine currently feasible control actions;

[0266] If the control operation corresponding to the input signal does not match the currently feasible control operation, it is determined that the input signal is an abnormal speech signal.

[0267] The currently feasible control operation may be determined based on, but not limited to, user behavior (such as historical user operation behavior).

[0268] Based on the above signal processing method, the present invention also provides a terminal device, such as Fig.18 As shown, including:

[0269] Processor 1801; and

[0270] The memory 1802 is configured to store machine-readable instructions, which, when executed by the processor, enable the processor to perform the above method.

[0271] The present invention also provides a signal processing method, such as Fig. 20 As shown, including:

[0272] 2001, receiving an input signal.

[0273] 2002, sending an abnormal voice interference signal within a preset frequency range.

[0274] Preferably, the abnormal voice interference signal includes an ultrasonic random interference signal.

[0275] Based on the above signal processing method, the present invention also provides a terminal device, such as Fig.18 As shown, including:

[0276] Processor 1801; and

[0277] The memory 1802 is configured to store machine-readable instructions, which, when executed by the processor, enable the processor to perform the above method.

[0278] The present invention also provides a signal processing method, such as Fig.21 As shown, including:

[0279] 2101, receiving an input signal.

[0280] 2102, sampling an input signal according to a sampling device having a sampling frequency higher than a set frequency threshold.

[0281] 2103, low-pass filtering is performed on the sampled signal.

[0282] Based on the above signal processing method, the present invention also provides a terminal device, such as Fig.18 As shown, including:

[0283] Processor 1801; and

[0284] The memory 1802 is configured to store machine-readable instructions, which, when executed by the processor, enable the processor to perform the above method.

[0285] Fig. 22 A block diagram of a computing system that can be used to implement a base station or user equipment according to an embodiment of the present disclosure is schematically shown.

[0286] like Fig. 22 As shown, the computing system 600 includes a processor 610, a computer-readable storage medium 620, an output interface 630, and an input interface 640. The computing system 600 can execute the above reference Fig.12 ,as well as Figure 19-21 The method described is used to determine whether the input signal is an abnormal speech signal.

[0287] Specifically, the processor 610 may include, for example, a general-purpose microprocessor, an instruction set processor and / or a related chipset and / or a dedicated microprocessor (e.g., an application-specific integrated circuit (ASIC)), etc. The processor 610 may also include an onboard memory for cache purposes. The processor 610 may be a processor for executing reference Fig.12 ,as well as Figure 19-21 A single processing unit or multiple processing units may be used for different actions of the described method flow.

[0288] Computer readable storage medium 620, for example, can be any medium capable of containing, storing, conveying, propagating or transmitting instructions. For example, readable storage media can include, but are not limited to, electrical, magnetic, optical, electromagnetic, infrared or semiconductor systems, devices, devices or propagation media. Specific examples of readable storage media include: magnetic storage devices, such as magnetic tapes or hard disks (HDD); optical storage devices, such as compact disks (CD-ROMs); memory, such as random access memory (RAM) or flash memory; and / or wired / wireless communication links.

[0289] The computer readable storage medium 620 may include a computer program 621, which may include code / computer executable instructions, which when executed by the processor 610 causes the processor 610 to perform, for example, the above-mentioned combination of Fig.12 ,as well as Figure 19-21 The described method sequences and any variations thereof.

[0290] The computer program 621 may be configured to have, for example, computer program codes including computer program modules. For example, in an exemplary embodiment, the codes in the computer program 621 may include one or more program modules, for example, including 621A, module 621B, ... It should be noted that the division method and number of modules are not fixed, and those skilled in the art may use appropriate program modules or program module combinations according to actual conditions. When these program module combinations are executed by the processor 610, the processor 610 may execute, for example, the above combination of Fig.12 ,as well as Figure 19-21 The described method sequences and any variations thereof.

[0291] According to an embodiment of the present disclosure, the processor 610 may use the output interface 630 and the input interface 640 to perform the above combined Fig.12 ,as well as Figure 19-21 The described method sequences and any variations thereof.

[0292] In the present invention, by determining whether the input signal is an abnormal voice signal based on feature information extracted from the input signal, effective recognition of abnormal voice signals is achieved, the accuracy of recognition is improved, and security is provided for the user's voice interaction, thereby enhancing the user's usage experience.

[0293] Those skilled in the art will appreciate that each block in these structure diagrams and / or block diagrams and / or flow charts and combinations of blocks in these structure diagrams and / or block diagrams and / or flow charts can be implemented using computer program instructions. Those skilled in the art will appreciate that these computer program instructions can be provided to a general-purpose computer, a professional computer, or a processor of other programmable data processing methods to implement, thereby executing the scheme specified in the block or multiple blocks of the structure diagram and / or block diagram and / or flow chart disclosed in the present invention through the processor of the computer or other programmable data processing method.

[0294] The modules of the device of the present invention can be integrated into one or deployed separately. The modules can be combined into one module or further divided into multiple sub-modules.

[0295] Those skilled in the art will appreciate that the accompanying drawings are merely schematic diagrams of a preferred embodiment, and the modules or processes in the accompanying drawings are not necessarily required to implement the present invention.

[0296] Those skilled in the art can understand that the modules in the device in the embodiment can be distributed in the device in the embodiment according to the description of the embodiment, or can be changed accordingly and located in one or more devices different from the present embodiment. The modules in the above embodiment can be combined into one module, or can be further divided into multiple sub-modules.

[0297] The above serial numbers of the present invention are only for description and do not represent the advantages and disadvantages of the embodiments.

[0298] The above disclosures are only several specific embodiments of the present invention; however, the present invention is not limited thereto, and any changes that can be conceived by those skilled in the art should fall within the protection scope of the present invention.

Claims

1. A signal processing method, characterized in that: include: According to the sampling device whose sampling frequency is higher than the set frequency threshold, the input signal is sampled; Perform low-pass filtering on the sampled signal; Extracting feature information from the processed input signal, wherein the feature information includes background noise information; Determining whether the input signal is an abnormal voice signal used to attack a normal voice signal according to the extracted feature information includes: Determining whether the input signal is the abnormal speech signal according to the background noise information of the signal; When it is determined that the input signal is the abnormal voice signal, an abnormal voice interference signal is sent within a preset frequency range, wherein the abnormal voice interference signal includes an ultrasonic random interference signal.

2. The method according to claim 1, characterized in that The characteristic information also includes energy characteristics of the signal and / or periodic characteristics of the signal.

3. The method according to claim 2, characterized in that The energy characteristics include short-term energy characteristics; And / or, the periodic characteristic includes a short-time zero-crossing rate characteristic.

4. The method according to claim 2 or 3, characterized in that Determining whether the input signal is an abnormal voice signal used to attack a normal voice signal according to the extracted feature information further includes: Based on the extracted feature information, it is determined whether the input signal is the abnormal speech signal by means of machine learning.

5. The method according to any one of claims 2 to 4, characterized in that: Extract feature information from the input signal, including: Extracts characteristic information from a set frequency range of an input signal.

6. The method according to claim 1, characterized in that Determining whether the input signal is an abnormal voice signal used to attack a normal voice signal according to the extracted feature information further includes: According to the change of the extracted characteristic information, it is determined whether the input signal is the abnormal speech signal.

7. The method according to claim 6, characterized in that The characteristic information also includes energy characteristics of the signal and / or directional characteristics of the signal.

8. The method according to claim 1, characterized in that Determining whether the input signal is an abnormal voice signal used to attack a normal voice signal according to the extracted feature information further includes: Determine whether the characteristic information matches the wake-up characteristic information corresponding to the current dynamic wake-up instruction; if not, determine that the voice signal is the abnormal voice signal.

9. The method according to claim 8, characterized in that The feature information also includes Mel-frequency cepstral coefficient (MFCC) features and / or dynamic verification codes.

10. The method according to claim 1, characterized in that The feature information also includes voice commands, Determining whether the input signal is an abnormal voice signal used to attack a normal voice signal according to the extracted feature information includes: Determine whether the voice command matches the wake-up word content corresponding to the current dynamic wake-up image; if not, determine that the voice signal is the abnormal voice signal.

11. The method according to claim 1, characterized in that Determining whether the input signal is the abnormal speech signal according to the background noise information of the signal includes: The current environment information is determined according to the background noise of the signal, and whether it is the abnormal speech signal is determined according to the current environment information.

12. The method according to claim 1, characterized in that The characteristic information also includes sound source location information; Determining whether the input signal is an abnormal voice signal used to attack a normal voice signal according to the extracted feature information includes: determining whether the sound source position information extracted from the input speech signal matches the detected user mouth position information; If there is no match, it is determined that the voice signal is the abnormal voice signal.

13. The method according to any one of claims 1 to 12, characterized in that Also includes: When it is determined that the input signal is an abnormal voice signal, a corresponding prompt message is sent.

14. A terminal device, comprising: processor; as well as A memory configured to store machine-readable instructions, wherein when the instructions are executed by the processor, the processor executes the method according to any one of claims 1 to 13.

Images