TWI817897B - Low-noise voiceprint identification device for financial transaction system and method thereof - Google Patents

Abstract

A low-noise voiceprint identification device for a financial transaction system and method thereof are provided, which improve the quality of voiceprint recognition and reduce noise by adding a recognition buffer mechanism and processing sound signals through filtering. It can be used as an auxiliary authentication function for financial transaction behavior on electronic devices. The voiceprint identification device includes an audio receiving unit, a voiceprint recognition unit, a storage unit, and an arithmetic unit. The arithmetic unit compares at least one voiceprint feature extracted from audio data with the first feature data and the second feature data of the storage unit, utilizes deep learning and neural network methods to train the first feature data and the second feature data, in order to security verification.

Description

Low-noise voiceprint recognition device and method used in financial trading system

The present invention relates to a voiceprint recognition device and a method thereof, and in particular to a low-noise voiceprint recognition device and a method used in a financial transaction system.

Voiceprints are unique, unique and not easily tampered with. In addition, they do not involve user privacy issues. Therefore, voiceprint recognition can be used to extract individual differences from audio signals and extract information that can reflect who the user is. The basic principle of user identification is that each user creates a module that can describe the user's personality characteristics as a description of the user's personality characteristics.

Voiceprint recognition is currently used in identity recognition as a tool to determine the user's identity. However, the disadvantage of voiceprint recognition is that the voice is easily affected by factors such as age, emotion, or physical condition, resulting in reduced recognition performance. Therefore, the accuracy of current voiceprint recognition technology still needs to be improved. In addition, factors such as the equipment’s sound quality and environmental noise will also affect the recognition performance of voiceprint recognition.

In view of the above shortcomings of the prior art, one aspect of the present invention provides a low-noise voiceprint recognition device. The device includes: an audio receiving unit for retrieving audio data; a voiceprint recognition unit Connected to the audio receiving unit, the voiceprint recognition unit receives the audio from the audio receiving unit audio data, and the voiceprint recognition unit identifies at least one voiceprint feature of the audio data; a storage unit stores a first characteristic data; a computing unit is connected to the voiceprint recognition unit and the storage unit, the The computing unit receives the at least voiceprint characteristics from the voiceprint recognition unit and the first characteristic data from the storage unit, and the computing unit includes: a verification calculator, the count value of the verification calculator in an initial state is 0 units, wherein the computing unit determines whether the at least one voiceprint feature matches the first feature data based on the first feature data; wherein, if the computing unit determines that the at least one voiceprint feature matches the first feature data, Then the computing unit stores the at least one voice fingerprint feature in a temporary storage space of the storage unit, and trains the first feature data based on the at least one voice fingerprint feature in the temporary storage space to generate an adjusted first feature data, and perform security verification based on the adjusted first feature data; if the computing unit determines that the at least one voiceprint feature does not match the first feature data, the computing unit sends a verification request, and completes the verification request After confirmation, the computing unit stores the at least one voice fingerprint characteristic in the temporary storage space, but the storage unit does not change the first characteristic data based on the at least one voice fingerprint characteristic to perform security operations based on the first characteristic data. Verification, wherein the verification request includes identification of user identity characteristics; if the computing unit determines that the at least voiceprint characteristics do not match the first characteristic data, and the count value reaches the predetermined value, then the computing unit determines according to the temporary storage space The at least one voiceprint feature and the first feature data in the generated second feature data; and after the second feature data is generated, the verification calculator adjusts the count value to 0 units, and the computing unit converts the first The characteristic data and the second characteristic data are stored in the storage unit, so as to determine whether the at least one voiceprint characteristic matches the first characteristic data and the second characteristic data based on the first characteristic data and the second characteristic data, so as to perform Security verification includes: if the computing unit determines that the at least one voiceprint feature matches the second feature data, the computing unit transmits the at least one voiceprint feature and stores it in the temporary storage space, and according to the temporary storage space The at least voiceprint characteristics train the second characteristic data to generate an adjusted second characteristic data, and perform security operations based on the adjusted second characteristic data. sexual verification; if the computing unit determines that the at least one voiceprint characteristic does not match the second characteristic data, but matches the first characteristic data, the computing unit removes the second characteristic data from the storage unit to obtain the result according to the first characteristic data. The characteristic data performs security verification; and after the second characteristic data is generated, if the computing unit determines that the at least one voiceprint characteristic does not match the second characteristic data and does not match the first characteristic data, the verification calculator The value increases by 1 unit, and when the count value is less than a predetermined value, the computing unit sends a verification request, and after the verification request is confirmed, the computing unit stores the at least one voice fingerprint feature in the temporary storage space , but the storage unit does not change the first characteristic data and the second characteristic data based on the at least audio fingerprint characteristics, so as to perform security verification based on the first characteristic data and the second characteristic data, wherein the content of the audio data is identified. The at least one voiceprint feature further includes at least one of the following operations: filtering, reducing noise, suppressing background noise, amplifying specific voiceprint features, calculating vector parameters of the audio data, detecting specific keywords, and detecting specific sound waves. band, detect specific sound waveforms, and detect specific sound wave frequencies.

Another aspect of the present invention provides a low-noise voiceprint recognition method applied to a voiceprint recognition device. The voiceprint recognition device includes an arithmetic unit and a verification calculator, and a count value of the verification calculator is An initial state is 0 units, and the voiceprint recognition method includes controlling the computing unit to perform the following steps: acquiring an audio data; identifying at least one voiceprint feature in the audio data; providing a first feature data and a temporary storage space and a verification request; based on the first characteristic data, compare whether the at least one voiceprint feature matches the first feature data; if the at least one voiceprint feature matches the first feature data, store the at least one voiceprint feature Characterize in the temporary storage space, and train the first characteristic data based on the at least voiceprint characteristics in the temporary storage space to generate an adjusted first characteristic data to perform security operations based on the adjusted first characteristic data. Sexual verification; if the at least one voiceprint feature does not match the first feature data, the verification request is sent. After the verification request is confirmed, the at least one voiceprint feature is stored in the temporary storage space, but the at least one voiceprint feature is not stored in the temporary storage space according to the at least one voiceprint feature. The voiceprint characteristic changes the first characteristic data to perform installation according to the first characteristic data. Completeness verification, wherein the verification request includes identification of user identity characteristics; if the count value reaches the predetermined value, a second characteristic is generated based on the at least one fingerprint characteristic and the first characteristic data in the temporary storage space data; and after the second characteristic data is generated, adjust the count value of the verification calculator to 0 units, and compare the at least one voiceprint characteristic with the third characteristic data based on the first characteristic data and the second characteristic data. Whether a characteristic data and the second characteristic data are consistent for security verification includes: if the computing unit determines that the at least one voice fingerprint characteristic matches the second characteristic data, the computing unit sends the at least one voice fingerprint characteristic to store in the temporary storage space, and train the second characteristic data based on the at least one voiceprint characteristic in the temporary storage space to generate an adjusted second characteristic data, and perform security operations based on the adjusted second characteristic data Verify; if the computing unit determines that the at least one voiceprint characteristic does not match the second characteristic data but matches the first characteristic data, the computing unit removes the second characteristic data from the storage unit to determine the first characteristic data based on the first characteristic data. The data is subjected to security verification; and after the second characteristic data is generated, if the computing unit determines that the at least one voiceprint characteristic does not match the second characteristic data and does not match the first characteristic data, the verification calculator calculates the count value Increase by 1 unit, and when the count value is less than a predetermined value, the computing unit sends a verification request, and after the verification request is confirmed, the computing unit stores the at least one fingerprint feature in the temporary storage space, However, the storage unit does not change the first characteristic data and the second characteristic data based on the at least audio fingerprint characteristics to perform security verification based on the first characteristic data and the second characteristic data, wherein the audio data is identified. The at least one voiceprint feature further includes at least one of the following operations: filtering, reducing noise, suppressing background noise, amplifying specific voiceprint features, calculating vector parameters of the audio data, detecting specific keywords, and detecting specific sound wave bands. , detect specific sound wave waveforms and detect specific sound wave frequencies.

To sum up, when users log into the system environment and want to perform financial transactions and other actions, they read the voiceprint characteristics recorded when the device is used, combined with deep learning, neural network and other training The voiceprint feature model can improve the quality of authentication, allowing the model to more accurately identify the user, and further apply this authentication to various financial transaction systems.

100,200: Voiceprint recognition device

300,400: Voiceprint recognition method

10: Audio receiving unit

20: Voiceprint recognition unit

30:Arithmetic unit

32: Validation Calculator

40:Storage unit

42: Temporary storage space

44: First characteristic data

46: Second characteristic data

S1~S14: steps

S1’, S2’: steps

FIG. 1 is a block diagram of a voiceprint recognition device according to an embodiment of the present invention.

FIG. 2 is a block diagram of a voiceprint recognition device according to another embodiment of the present invention.

FIG. 3 is a flow chart illustrating a voiceprint recognition method according to an embodiment of the present invention.

FIG. 4 is a flow chart illustrating a voiceprint recognition method according to another embodiment of the present invention.

The following description is a preferred implementation manner for completing the invention, and its purpose is to describe the basic spirit of the invention, but is not intended to limit the invention. For the actual invention, reference must be made to the following claims.

It must be understood that the words "including" and "including" used in this specification are used to indicate the existence of specific technical features, values, method steps, work processes, components and/or components, but do not exclude the possibility of Plus further technical features, values, method steps, processes, components, assemblies, or any combination of the above.

The use of words such as "first" and "second" in the claims is used to modify the elements in the claims, and is not used to indicate a priority order, precedence relationship, or that one element precedes another element. , or the time sequence when executing method steps, is only used to distinguish components with the same name.

Please refer to FIGS. 1 and 2 together. FIG. 1 is a block diagram of a voiceprint recognition device according to one embodiment of the present invention. FIG. 2 is a block diagram of a voiceprint recognition device according to another embodiment of the present invention. .

In one embodiment, as shown in FIG. 1 , the voiceprint recognition device 100 includes an audio receiving unit 10 , a voiceprint recognition unit 20 , a computing unit 30 and a storage unit 40 .

The audio receiving unit 10 captures audio data. The voiceprint recognition unit 20 is connected to the audio receiving unit 10 to receive the audio data, and can identify at least one voiceprint feature in the audio data. In some embodiments, the above-mentioned at least one sound fingerprint feature may be a specific keyword, a sound wave band, a sound wave waveform, a sound wave frequency, etc. In one embodiment, the voiceprint recognition unit 20 is used to identify the at least one voiceprint feature in the audio data, and the voiceprint recognition unit 20 may include at least one of the following operations, such as filtering, noise reduction, and background suppression. Noise, amplify specific voiceprint features, calculate a vector parameter of audio data, detect specific keywords, detect specific sound wave bands, detect specific sound wave waveforms, detect specific sound wave frequencies, etc.

The storage unit 40 can store the first characteristic data 44 . In some embodiments, the first characteristic data 44 may be any relevant voiceprint information that matches the user, such as specific keywords, sound wave bands, sound wave waveforms, sound wave frequencies, etc. In some embodiments, the first characteristic data 44 may be an analog signal, a digital signal, an analog/digital hybrid signal, or other data storage modes.

The computing unit 30 is connected to the voiceprint recognition unit 20 and the storage unit 40, and can receive the at least one voiceprint feature from the voiceprint recognition unit 20 and the first feature data 44 received from the storage unit 40, and determine in the computing unit 30 Whether the above-mentioned at least one voiceprint feature matches the first feature data 44.

If the computing unit 30 determines that the at least one voiceprint feature matches the first feature data 44, the computing unit 30 stores the at least one voiceprint feature in a temporary storage space 42 of the storage unit 40, and based on the data in the temporary storage space 42 The at least one voiceprint feature is used to train the first feature data 44 to generate an adjusted The first characteristic data 44 is adjusted, and security verification is performed based on the adjusted first characteristic data 44 . For example, the above security verification can be executed immediately after the adjusted first characteristic data 44 is created to further achieve the effect of a fast security verification.

In another example, the above security verification can wait until a new voiceprint feature is subsequently received, and then compare the new voiceprint feature with the adjusted first feature data 44, so that the security verification can also be achieved. Purpose, compared with the previous example, can reduce the probability of misjudgment as "fail", improve the smoothness of the overall process, and thereby improve the user experience.

In some embodiments, at least the voiceprint characteristics determined above may correspond to a normal voiceprint of the user, or an abnormal voiceprint affected by factors such as age, emotion, or physical condition, and the first characteristic data 44 can be viewed It is the user's original preset voiceprint feature. If at least the voiceprint feature can match the first feature data 44, then the adjusted first feature data 44 is used as the voiceprint recognition standard for security verification, and the first feature The data 44 can be continuously trained to continuously update the acceptable voiceprint range of the first characteristic data 44 .

In one embodiment, as shown in Figure 2, the computing unit 30 of the voiceprint recognition device 200 may further include a verification calculator 32. The count value of the verification calculator 32 is 0 units in an initial state. If the computing unit 30 determines If the at least one voiceprint characteristic does not match the first characteristic data 44, the count value of the verification calculator 32 is increased by 1 unit.

If the count value is less than a predetermined value, and if the computing unit 30 determines that the at least one voiceprint feature does not match the first feature data 44, the computing unit 30 will send a verification request to the user and wait for the verification request to be confirmed by the user. Afterwards, the computing unit 30 stores the at least one voice fingerprint feature in the temporary storage space 42 , but the storage unit 40 does not change the first feature data 44 based on the at least one voice fingerprint feature, but still uses the first feature data 44 for processing. Security verification. In some embodiments, the at least one voiceprint feature may correspond to the user's abnormal voiceprint that is affected by factors such as age, emotion, or physical condition, and the first feature data 44 may be regarded as the user's original preset voiceprint. Characteristics. If the abnormal voiceprint is at least the above If the voiceprint characteristics do not match the first characteristic data 44, the present invention still retains the first characteristic data 44, and only additionally stores the abnormal voiceprint in the temporary storage space 42, which can be used as a recognition buffering mechanism. The above recognition buffering The mechanism mainly avoids the possibility that at least the above-mentioned voiceprint characteristics of the user are only temporary abnormalities in the voiceprint, or that the voiceprint recognition device 100 makes an error in judgment.

If the count value reaches the predetermined value, the computing unit 30 generates a second feature data 46 based on the at least one voice fingerprint feature and the first feature data 44 in the temporary storage space 42 . In some embodiments, the second characteristic data 46 may be an analog signal, a digital signal, an analog/digital mixed signal, or other data storage modes.

After the second characteristic data 46 is generated, the verification calculator 32 adjusts the count value to 0 units, and the computing unit 30 stores the first characteristic data 44 and the second characteristic data 46 in the storage unit 40 and modifies them according to the first characteristic data 44 , the second characteristic data 46, so that the computing unit 30 determines whether the subsequently received at least one voiceprint characteristic matches the first characteristic data 44 and the second characteristic data 46 for security verification. In some embodiments, the at least one voiceprint feature may correspond to the user's abnormal voiceprint that is affected by factors such as age, emotion, or physical condition, and the first feature data 44 may be regarded as the user's original preset voiceprint. Characteristics. If the above-mentioned at least one voiceprint feature of the abnormal voiceprint has reached a certain predetermined number of times (predetermined time) and cannot match the first feature data 44, the present invention still retains the first feature data 44, and additionally based on the first feature data 44 stored in the temporary storage space 42 The above-mentioned at least voiceprint characteristics and the first characteristic data 44 generate a second characteristic data 46, which can be used as another recognition buffering mechanism. The above-mentioned recognition buffering mechanism mainly increases the user's age, mood, or physical condition and other factors for a certain period of time. Influenced voiceprint feature identification data.

In one embodiment, when the computing unit 30 determines whether the at least one voiceprint feature matches the first feature data 44 and the second feature data 46, if the at least one voiceprint feature matches the second feature data 46, the calculation unit 30 The at least one voiceprint feature is transmitted and stored in the temporary storage space 42, and the second feature data 46 is trained according to the at least one voiceprint feature in the temporary storage space 42 to generate an adjusted second feature. Data 46, and security verification is performed based on the adjusted second characteristic data 46. For example, the above security verification can be executed immediately after the adjusted second characteristic data 46 is created to further achieve the effect of a fast security verification.

In another example, the above security verification can wait until a new voiceprint feature is subsequently received, and then compare the new voiceprint feature with the adjusted second feature data 46, so that the security verification can also be achieved. Purpose, compared with the previous example, can reduce the probability of misjudgment as "fail", improve the smoothness of the overall process, and thereby improve the user experience.

In some embodiments, the above-mentioned at least voiceprint characteristics may correspond to the user's abnormal voiceprint after being affected by factors such as age, emotion, or physical condition, and the second characteristic data 46 may be regarded as the user's abnormal voiceprint after being affected by abnormality for a certain period of time. Voiceprint feature identification data. If the above-mentioned at least one voiceprint feature of the abnormal voiceprint has reached a certain predetermined number of times (predetermined time) but matches the second feature data 46, the present invention still retains the first feature data 44, and only determines that the user has reached a certain predetermined number of times (predetermined time). The stable condition after the abnormal change of voiceprint, and the adjusted second characteristic data 46 is used as the voiceprint recognition standard of this stable condition for security verification, and the second characteristic data 46 can be continuously trained to continuously update the second characteristic. Data 46 Acceptable range of voiceprints.

If the computing unit 30 determines that the at least one voiceprint characteristic does not match the second characteristic data 46 but matches the first characteristic data 44 , the computing unit 30 removes the second characteristic data 46 from the storage unit 40 . After the second characteristic data 46 is removed, the security verification is performed again based on the first characteristic data 44 . In some embodiments, the above-mentioned at least voiceprint characteristics may correspond to abnormal voiceprint characteristics of the user affected by factors such as age, emotion, or physical condition, and the first characteristic data 44 may be regarded as the user's original preset voiceprint. Characteristics, the second characteristic data 46 can be regarded as the user's voiceprint characteristic identification data after being affected by abnormality for a certain period of time. If the user recovers from the situation that caused the abnormal voiceprint characteristics, it can be determined that the second characteristic data 46 is no longer applicable. As the standard for voiceprint recognition, the second characteristic data 46 is removed and the first characteristic data 44 is restored as the standard for voiceprint recognition to perform security verification.

If the computing unit 30 determines that the at least one voiceprint feature does not match the second feature data 46 and the first feature data 44 , the verification calculator 32 increases the count value by 1 unit. If the count value is less than a predetermined value, the computing unit 30 sends a verification request to the user, and after waiting for the verification request to be confirmed by the user, the computing unit 30 stores the at least one voice fingerprint feature in the temporary storage space 42 , but the storage unit 40 does not change the first characteristic data 44 and the second characteristic data 46 according to the at least one voice fingerprint characteristic, so as to perform security verification based on the first characteristic data and the second characteristic.

In one embodiment, if the count value reaches the predetermined value, the computing unit 30 trains the second feature data 46 according to the at least one voice fingerprint feature in the temporary storage space 42 to generate an adjusted second feature data 46 . After the adjusted second characteristic data 46 is generated, the verification calculator 32 adjusts the count value to 0 units, and the computing unit 30 determines whether at least the voice fingerprint characteristic is consistent with the first characteristic data 44 and the adjusted second characteristic data 46, so as to Perform security verification. For example, the above security verification can be executed immediately after the adjusted second characteristic data 46 is created to further achieve the effect of a fast security verification.

In another example, the above security verification can wait until a new voiceprint feature is subsequently received, and then compare the new voiceprint feature with the first feature data 44 and the adjusted second feature data 46, and so on. It can achieve the purpose of security verification, and compared with the previous example, it can reduce the probability of misjudgment as "failed", improve the smoothness of the overall process, and thereby improve the user experience.

In some embodiments, the above-mentioned at least voiceprint characteristics may correspond to abnormal voiceprint characteristics of the user affected by factors such as age, emotion, or physical condition, and the conditions causing the abnormal voiceprint characteristics may be difficult to recover. If the above-mentioned at least one voiceprint feature of the abnormal voiceprint has reached a certain predetermined number of times (predetermined time) and cannot match the first feature data 44 and the second feature data 46, it can be determined to be the second feature data 46 The voiceprint recognition standard for abnormal changes in voiceprint to reach a stable state is no longer applicable. Therefore, the voiceprint features in the temporary storage space 42 are used to retrain the second feature data 46 to generate an adjusted second feature data 46 to retrain it. Used as a standard for voiceprint recognition for security verification.

In one embodiment, the voiceprint recognition devices 100 and 200 can be installed on personal computers, laptops, cellular phones, camera phones, smart phones, personal digital assistants, media players, and navigation systems that perform financial transactions over the Internet. device, email device, game console, tablet, wearable, or a combination of any of these devices.

In one embodiment, the computing unit 30 may be a central processing unit (CPU), or may be configured as other components with sufficient computing capabilities.

In one embodiment, the storage unit 40 may be a server for storing voiceprint recognition databases of one or more companies, banks or institutions. In another embodiment, the storage unit 40 may be any type of fixed or removable random access memory (RAM), read-only memory (ROM), Flash memory (flash memory), hard disk drive (HDD), solid state drive (SSD) or similar components or a combination of the above components.

In one embodiment, the identification method of the user identity included in the above verification request may be SMS, fax, email, application, digital signature, or based on other biological characteristics (such as facial image, fingerprint, Iris recognition), etc.

In one embodiment, the storage unit 40 may include a deep learning model for training the first feature data 44 and the second feature data 46. The above-mentioned deep learning model may include at least one of various types of neural network models, The various types of neural network models include convolutional neural network (CNN), region with convolutional neural network (region with convolution) neural network (R-CNN), region proposal network (RPN), recurrent neural network (RNN), stacking-based deep neural network (S-DNN) ), state-space dynamic neural network (S-SDNN), deconvolution network (deconvolution network), deep belief network (DBN), restricted Boltzmann machine (restricted Boltzmann machine, RBM), fully convolutional network, long short-term memory (LSTM) network and classification network.

Please refer to FIG. 3 and FIG. 4 together. FIG. 3 is a flow chart illustrating a voiceprint identification method according to one embodiment of the present invention. FIG. 4 is a flow chart illustrating a voiceprint identification method according to another embodiment of the present invention. flowchart.

In one embodiment, as shown in FIG. 3 , in step S1 , audio data is captured by the audio receiving unit 10 .

In step S2, the voiceprint recognition unit 20 receives the audio data transmitted by the audio receiving unit 10, and identifies at least one voiceprint feature in the audio data.

In step S3, the computing unit 30 receives the at least one voiceprint feature sent by the voiceprint recognition unit 20 and the first feature data 44 sent by the storage unit 40 to compare the voiceprint feature with the first feature data. 44 is consistent.

In step S4, if the at least one voiceprint feature matches the first feature data 44, the computing unit 30 stores the at least one voiceprint feature in the temporary storage space 42, and the computing unit 30 stores the at least one voiceprint feature in the temporary storage space 42 according to the above feature data in the temporary storage space 42. At least the voiceprint feature trains the first characteristic data 44 to generate an adjusted first characteristic data 44 for security verification based on the adjusted first characteristic data 44 . For example, the above security The security verification can be executed immediately after the adjusted first characteristic data 44 is created to further achieve the effect of a fast security verification.

In another example, the above security verification can wait until a new voiceprint feature is subsequently received, and then compare the new voiceprint feature with the adjusted first feature data 44, so that the security verification can also be achieved. Purpose, compared with the previous example, can reduce the probability of misjudgment as "fail", improve the smoothness of the overall process, and thereby improve the user experience.

In step S5, the computing unit 30 may include a verification calculator 32. The count value of the verification calculator 32 is 0 units in an initial state. If the above-mentioned voiceprint feature does not match the first feature data 44, the verification calculator 32 will be increased. The count value is 1 unit. If the count value is less than a predetermined value, the computing unit 30 will send a verification request to the user. When the verification request is confirmed by the user, the computing unit 30 will store the at least one fingerprint feature in the temporary storage space 42 within, but does not change the first characteristic data 44 based on the at least one voiceprint characteristic, so as to perform security verification based on the first characteristic data 44 .

In step S6 , if the at least one voice fingerprint feature does not match the first feature data 44 and the count value reaches a predetermined value, the computing unit 30 compares the at least one voice fingerprint feature in the temporary storage space 42 with the first feature data 44 A second characteristic data 46 is generated.

In step S7, after the second characteristic data 46 is generated, the verification calculator 32 adjusts the count value to 0 units, and the computing unit 30 is modified to compare at least the voiceprint characteristics according to the first characteristic data 44 and the second characteristic data 46. Whether it is consistent with the first characteristic data 44 and the second characteristic data 46 for security verification.

In one embodiment, as shown in FIG. 4 , in step S8 , the computing unit 30 is modified to compare at least one voice fingerprint feature with the first feature data 44 and the second feature data 46 based on the first feature data 44 and the second feature data 46 . Whether the two feature data 46 are consistent, and the computing unit 30 first compares at least the voice fingerprint feature and the second feature data 46 based on the second feature data 46 to see whether they are consistent.

In step S9, if the voiceprint feature matches the second feature data 46, the computing unit 30 stores the at least one voiceprint feature in the temporary storage space 42, and the storage unit 40 stores the voiceprint feature according to the at least one feature in the temporary storage space 42. The voiceprint feature trains the second feature data 46 to generate an adjusted second feature data 46, and security verification is performed based on the adjusted second feature data 46. For example, the above security verification can be executed immediately after the adjusted second characteristic data 46 is created to further achieve the effect of a fast security verification.

In another example, the above security verification can wait until a new voiceprint feature is subsequently received, and then compare the new voiceprint feature with the adjusted second feature data 46, so that the security verification can also be achieved. Purpose, compared with the previous example, can reduce the probability of misjudgment as "fail", improve the smoothness of the overall process, and thereby improve the user experience.

In step S10 , if the voiceprint feature does not match the second feature data 46 , the computing unit 30 then compares the voiceprint feature with the first feature data 44 based on the first feature data 44 .

In step S11 , if the voiceprint feature does not match the second feature data 46 but matches the first feature data 44 , the computing unit 30 removes the second feature data 46 to perform security verification based on the first feature data 44 .

In step S12, the count value of the verification calculator 32 is 0 units in an initial state. If the voiceprint feature does not match the second feature data 46 and does not match the first feature data 44, then the count value of the verification calculator 32 is increased by 1 unit. . If the count value is less than a predetermined value, a verification request is sent to the user. After the verification request is confirmed by the user, the computing unit 30 sends the voiceprint feature to the temporary storage space. 42, but the storage unit 40 does not change the first characteristic data 44 and the second characteristic data 46 according to the at least one voice fingerprint characteristic, so as to perform security verification based on the first characteristic data 44 and the second characteristic data 46.

In step S13 , if the at least one voiceprint feature does not match the second feature data 46 and the first feature data 44 , and the count value reaches a predetermined value, the computing unit 30 calculates the at least one voiceprint feature based on the at least one voiceprint feature in the temporary storage space 42 . Feature training second feature data 46 to generate adjusted second feature data 46 .

In step S14, after the adjusted second characteristic data 46 is generated, the verification calculator adjusts the count value to 0 units, and the computing unit 30 is modified to determine at least one based on the first characteristic data 44 and the adjusted second characteristic data 46. Whether the voiceprint characteristics match the first characteristic data 44 and the adjusted second characteristic data 46 is used for security verification. For example, the above security verification can be executed immediately after the adjusted second characteristic data 46 is created to further achieve the effect of a fast security verification.

In another example, the above security verification can wait until a new voiceprint feature is subsequently received, and then compare the new voiceprint feature with the first feature data 44 and the adjusted second feature data 46, and so on. It can achieve the purpose of security verification, and compared with the previous example, it can reduce the probability of misjudgment as "failed", improve the smoothness of the overall process, and thereby improve the user experience.

In one embodiment, the voiceprint recognition methods 300 and 400 can be configured on an electronic financial transaction system, such as an automated teller machine (ATM), online banking, or other financial transaction systems.

In one embodiment, the identification method of the user identity included in the above verification request may be SMS, fax, email, application, digital signature, or based on other biological characteristics (such as facial image, fingerprint, Iris recognition), etc.

In one embodiment, the voiceprint recognition device in the voiceprint recognition methods 300 and 400 may include a deep learning model for training the first feature data 44 and the second feature data 46. The above-mentioned deep learning model may include various types of At least one of the neural network models, the various types of neural network models Types include convolutional neural network (CNN), region with convolutional neural network (R-CNN), region proposal network (RPN), recurrent neural network ( recurrent neural network (RNN), stacking-based deep neural network (S-DNN), state-space dynamic neural network (S-SDNN), deconvolution network deconvolution network, deep belief network (DBN), restricted Boltzmann machine (RBM), fully convolutional network, long short-term memory (long short-term memory) short-term memory, LSTM) network and classification network.

The present invention can be used as an auxiliary authentication function for financial transaction systems on electronic devices. When users log into the system environment and want to perform financial transactions and other actions, they read the voiceprint pattern recorded when the device is used, and train the voiceprint pattern model through deep learning and adding recursive neural networks to improve the quality of authentication. , so that the model can more accurately identify the user's own instructions, and then apply this authentication to various financial transaction systems, such as mobile banking (online banking), etc.

However, the above are only examples of the present invention, and should not be used to limit the scope of the present invention. All simple equivalent changes and modifications made based on the patent scope of the present invention and the content of the patent specification are still within the scope of the present invention. Within the scope covered by the patent of this invention.

400: Voiceprint identification method

S1~S14: steps

S1’, S2’: steps

Claims (6)

A low-noise voiceprint recognition device for financial trading systems, including: an audio receiving unit, used to capture an audio data; a voiceprint recognition unit, connected to the audio receiving unit, the voiceprint recognition unit The audio data is received by the audio receiving unit, and the voiceprint recognition unit identifies at least one voiceprint feature of the audio data; a storage unit stores a first feature data; a computing unit is connected to the voiceprint recognition unit and the storage unit, the computing unit receives the at least voiceprint characteristics from the voiceprint recognition unit and the first characteristic data from the storage unit, and the computing unit includes: a verification calculator, the verification calculator The count value is 0 units in an initial state, wherein the computing unit determines whether the at least one voiceprint feature matches the first feature data based on the first feature data; wherein, if the computing unit determines that the at least one voiceprint feature is consistent with the first feature data, If the first characteristic data is met, the computing unit stores the at least one voice fingerprint feature in a temporary storage space of the storage unit, and trains the first feature data based on the at least one voice fingerprint feature in the temporary storage space, To generate an adjusted first characteristic data, and perform security verification based on the adjusted first characteristic data; if the computing unit determines that the at least one voice fingerprint characteristic does not match the first characteristic data, the verification calculator calculates the calculated The value increases by 1 unit, and when the count value is less than a predetermined value, the computing unit sends a verification request, and after the verification request is confirmed, the computing unit stores the at least one voice fingerprint feature in the temporary storage space , but the storage unit does not change the first characteristic data based on the at least one voiceprint characteristic, so as to perform security verification based on the first characteristic data, wherein the verification request includes identification of the user's identity characteristics; If the computing unit determines that the at least one voiceprint feature does not match the first feature data, and the count value reaches the predetermined value, then the calculation unit determines that the at least one voiceprint feature in the temporary storage space and the first feature data Generate a second characteristic data; and after the second characteristic data is generated, the verification calculator adjusts the count value to 0 units, and the computing unit stores the first characteristic data and the second characteristic data in the storage unit, Determining whether the at least one voiceprint feature matches the first feature data and the second feature data based on the first feature data and the second feature data for security verification, including: if the computing unit determines that the at least one voiceprint feature is consistent with the first feature data and the second feature data. If the voiceprint feature matches the second feature data, the computing unit transmits the at least voiceprint feature and stores it in the temporary storage space, and trains the second feature data based on the at least voiceprint feature in the temporary storage space. , to generate an adjusted second characteristic data, and perform security verification based on the adjusted second characteristic data; if the computing unit determines that the at least one voice fingerprint characteristic does not match the second characteristic data, but matches the first characteristic data , then the computing unit removes the second characteristic data from the storage unit to perform security verification based on the first characteristic data; and after the second characteristic data is generated, if the computing unit determines that the at least one voiceprint characteristic does not match the second characteristic data and does not match the first characteristic data, the verification calculator increases the count value by 1 unit, and when the count value is less than a predetermined value, the computing unit sends a verification request, and in the verification After the request is completed and confirmed, the computing unit stores the at least one voiceprint feature in the temporary storage space, but the storage unit does not change the first feature data and the second feature data based on the at least one voiceprint feature. The first characteristic data and the second characteristic data are subject to security verification, wherein identifying the at least one voiceprint feature in the audio data further includes at least one of the following operations: filtering, reducing noise, suppressing background noise, amplifying specific voiceprint features, calculating vector parameters of the audio data, and detecting specific keys words, detect specific sound wave bands, detect specific sound wave waveforms, and detect specific sound wave frequencies. The low-noise voiceprint recognition device of claim 1, wherein the computing unit determines whether the at least voiceprint feature is consistent with the first feature data and the second feature data further includes: after the second feature data is generated, If the count value reaches the predetermined value, the computing unit trains the second feature data based on the at least one voice fingerprint feature in the temporary storage space to generate an adjusted second feature data; and after generating the adjusted second feature data After obtaining the two characteristic data, the verification calculator adjusts the count value to 0 units, and the computing unit determines the characteristics of the at least voiceprint and the first characteristic data and the adjusted second characteristic data based on the first characteristic data and the adjusted second characteristic data. Check whether the adjusted second characteristic data is consistent for security verification. The low-noise voiceprint recognition device of claim 1, wherein the voiceprint recognition device further includes: a deep learning algorithm for training the first feature data and the second feature data, and the deep learning algorithm includes a A Deep Neural Network (DNN) model, a Convolutional Neural Network (CNN) model, a Recurrent Neural Network (RNN), and one of their combinations. A low-noise voiceprint recognition method for financial trading systems, applied to a voiceprint recognition device. The voiceprint recognition device includes an arithmetic unit and a verification calculator. The count value of the verification calculator is in an initial state. is 0 units. The voiceprint recognition method includes controlling the computing unit to perform the following steps: Retrieve an audio data; identify at least one voiceprint feature in the audio data; provide a first feature data, a temporary storage space and a verification request; based on the first feature data, compare the at least one voiceprint feature with Whether the first characteristic data matches; if the at least one voice fingerprint characteristic matches the first characteristic data, the at least one voice fingerprint characteristic is stored in the temporary storage space, and the at least one voice fingerprint characteristic in the temporary storage space is Feature training the first feature data to generate an adjusted first feature data for security verification based on the adjusted first feature data; if the at least one voiceprint feature does not match the first feature data, the verification The count value of the calculator is increased by 1 unit, and when the count value is less than a predetermined value, the verification request is sent. After the verification request is confirmed, the at least one fingerprint feature is stored in the temporary storage space, but no The first feature data is modified based on the at least voiceprint feature to perform security verification based on the first feature data, where the verification request includes identification of user identity features; if the count value reaches the predetermined value, based on the The at least one voiceprint feature and the first feature data in the temporary storage space generate a second feature data; and after the second feature data is generated, the count value of the verification calculator is adjusted to 0 units, and according to the The first characteristic data and the second characteristic data compare whether the at least one voiceprint characteristic is consistent with the first characteristic data and the second characteristic data for security verification, including: If the computing unit determines that the at least one voiceprint feature matches the second feature data, the computing unit transmits the at least one voiceprint feature and stores it in the temporary storage space, and based on the at least one voiceprint feature in the temporary storage space Feature training the second feature data to generate an adjusted second feature data, and perform security verification based on the adjusted second feature data; if the computing unit determines that the at least one voiceprint feature does not match the second feature data, However, if the first characteristic data is met, the computing unit removes the second characteristic data from a storage unit to perform security verification based on the first characteristic data; and after the second characteristic data is generated, if the computing unit If it is determined that the at least one voiceprint feature does not match the second feature data and does not match the first feature data, the verification calculator increases the count value by 1 unit, and when the count value is less than a predetermined value, the computing unit sends a Verification request, and after the verification request is completed and confirmed, the computing unit stores the at least one voice fingerprint feature in the temporary storage space, but the storage unit does not change the first feature data and the first feature data based on the at least one voice fingerprint feature. The second characteristic data is used to perform security verification based on the first characteristic data and the second characteristic data, wherein identifying the at least one fingerprint characteristic in the audio data further includes at least one of the following operations: filtering, noise reduction , suppress background noise, amplify specific voiceprint features, calculate vector parameters of the audio data, detect specific keywords, detect specific sound wave bands, detect specific sound wave waveforms, and detect specific sound wave frequencies. The low-noise voiceprint identification method of claim 4, wherein comparing whether the at least one voiceprint feature matches the first feature data and the second feature data further includes the following steps: After the second characteristic data is generated and the count value reaches the predetermined value, the second characteristic data is trained according to the at least one voiceprint characteristic in the temporary storage space to generate an adjusted second characteristic data; and After the adjusted second characteristic data is generated, the count value of the verification calculator is adjusted to 0 units, and based on whether the at least one voiceprint characteristic is consistent with the first characteristic data and the adjusted second characteristic data, the process is performed Security verification. The low-noise voiceprint recognition method of claim 4, wherein the voiceprint recognition device further includes: a deep learning algorithm for training the first feature data and the second feature data, and the deep learning algorithm includes a A Deep Neural Network (DNN) model, a Convolutional Neural Network (CNN) model, a Recurrent Neural Network (RNN), and one of their combinations.

Images