WO2021159902A1 - 年龄识别方法、装置、设备及计算机可读存储介质 - Google Patents

年龄识别方法、装置、设备及计算机可读存储介质 Download PDF

Info

Publication number
WO2021159902A1
WO2021159902A1 PCT/CN2021/071262 CN2021071262W WO2021159902A1 WO 2021159902 A1 WO2021159902 A1 WO 2021159902A1 CN 2021071262 W CN2021071262 W CN 2021071262W WO 2021159902 A1 WO2021159902 A1 WO 2021159902A1
Authority
WO
WIPO (PCT)
Prior art keywords
voice
age
target
feature
network model
Prior art date
Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
Ceased
Application number
PCT/CN2021/071262
Other languages
English (en)
French (fr)
Inventor
马坤
赵之砚
施奕明
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
OneConnect Smart Technology Co Ltd
Original Assignee
OneConnect Smart Technology Co Ltd
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by OneConnect Smart Technology Co Ltd filed Critical OneConnect Smart Technology Co Ltd
Publication of WO2021159902A1 publication Critical patent/WO2021159902A1/zh
Anticipated expiration legal-status Critical
Ceased legal-status Critical Current

Links

Images

Classifications

    • G—PHYSICS
    • G10—MUSICAL INSTRUMENTS; ACOUSTICS
    • G10L—SPEECH ANALYSIS TECHNIQUES OR SPEECH SYNTHESIS; SPEECH RECOGNITION; SPEECH OR VOICE PROCESSING TECHNIQUES; SPEECH OR AUDIO CODING OR DECODING
    • G10L25/00—Speech or voice analysis techniques not restricted to a single one of groups G10L15/00 - G10L21/00
    • G10L25/48—Speech or voice analysis techniques not restricted to a single one of groups G10L15/00 - G10L21/00 specially adapted for particular use
    • G10L25/51—Speech or voice analysis techniques not restricted to a single one of groups G10L15/00 - G10L21/00 specially adapted for particular use for comparison or discrimination
    • G—PHYSICS
    • G06—COMPUTING OR CALCULATING; COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/04—Architecture, e.g. interconnection topology
    • G06N3/045—Combinations of networks
    • G—PHYSICS
    • G06—COMPUTING OR CALCULATING; COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/04—Architecture, e.g. interconnection topology
    • G06N3/048—Activation functions
    • G—PHYSICS
    • G06—COMPUTING OR CALCULATING; COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/08—Learning methods
    • G—PHYSICS
    • G10—MUSICAL INSTRUMENTS; ACOUSTICS
    • G10L—SPEECH ANALYSIS TECHNIQUES OR SPEECH SYNTHESIS; SPEECH RECOGNITION; SPEECH OR VOICE PROCESSING TECHNIQUES; SPEECH OR AUDIO CODING OR DECODING
    • G10L25/00—Speech or voice analysis techniques not restricted to a single one of groups G10L15/00 - G10L21/00
    • G10L25/03—Speech or voice analysis techniques not restricted to a single one of groups G10L15/00 - G10L21/00 characterised by the type of extracted parameters
    • G10L25/18—Speech or voice analysis techniques not restricted to a single one of groups G10L15/00 - G10L21/00 characterised by the type of extracted parameters the extracted parameters being spectral information of each sub-band
    • G—PHYSICS
    • G10—MUSICAL INSTRUMENTS; ACOUSTICS
    • G10L—SPEECH ANALYSIS TECHNIQUES OR SPEECH SYNTHESIS; SPEECH RECOGNITION; SPEECH OR VOICE PROCESSING TECHNIQUES; SPEECH OR AUDIO CODING OR DECODING
    • G10L25/00—Speech or voice analysis techniques not restricted to a single one of groups G10L15/00 - G10L21/00
    • G10L25/27—Speech or voice analysis techniques not restricted to a single one of groups G10L15/00 - G10L21/00 characterised by the analysis technique
    • G10L25/30—Speech or voice analysis techniques not restricted to a single one of groups G10L15/00 - G10L21/00 characterised by the analysis technique using neural networks

Definitions

  • This application relates to the field of artificial intelligence technology, and in particular to an age identification method, device, equipment, and computer-readable storage medium.
  • the main purpose of this application is to provide an age identification method, device, equipment, and computer-readable storage medium, aiming to solve the traditional technical problem of low accuracy in age identification.
  • an embodiment of the present application provides an age identification method, and the age identification method includes:
  • the depth feature of the input spectrogram is extracted through the age recognition network model, and the target age group to which the target user belongs is determined according to the depth feature.
  • an embodiment of the present application further provides an age identification device, the age identification device including:
  • the sample expansion module is used to obtain real voice samples from a preset database, and perform sample expansion on the real voice samples based on the generative confrontation network GAN to obtain expanded voice samples;
  • a model training module is used to obtain an age recognition network model through the extended speech sample training
  • the voice conversion module is used to obtain the target voice of the target user and convert the target voice into a corresponding input spectrogram
  • the age determination module is configured to extract the depth characteristics of the input spectrogram through the age recognition network model, and determine the target age group to which the target user belongs according to the depth characteristics.
  • an embodiment of the present application further provides an age identification device, the age identification device includes a processor, a memory, and a computer program stored on the memory and executable by the processor, wherein When the computer program is executed by the processor, the age identification method as described above is realized, and the age identification method includes the following steps:
  • the depth feature of the input spectrogram is extracted through the age recognition network model, and the target age group to which the target user belongs is determined according to the depth feature.
  • the embodiments of the present application also provide a computer-readable storage medium having a computer program stored on the computer-readable storage medium, and when the computer program is executed by a processor, the above-mentioned age An identification method, the age identification method includes the following steps:
  • the depth feature of the input spectrogram is extracted through the age recognition network model, and the target age group to which the target user belongs is determined according to the depth feature.
  • the embodiments of the present application can improve the generalization ability of age recognition and improve the accuracy of age recognition.
  • FIG. 1 is a schematic diagram of the hardware structure of the age recognition device involved in the solution of the embodiment of the application;
  • FIG. 2 is a schematic flowchart of the first embodiment of the age identification method of this application.
  • FIG. 3 is a schematic flowchart of a second embodiment of the age identification method of this application.
  • FIG. 4 is a schematic flowchart of a third embodiment of the age identification method of this application.
  • the technical solution of this application can be applied to the fields of artificial intelligence, smart city, digital medical care, blockchain and/or big data technology.
  • the data involved in this application such as voice samples, depth features, and/or determined age group information, can be stored in a database, or can be stored in a blockchain, such as distributed storage through a blockchain.
  • the application is not limited.
  • the age identification method involved in the embodiments of the present application is mainly applied to an age identification device, and the age identification device may be a device with a data processing function such as a server, a personal computer (PC), or a notebook computer.
  • a data processing function such as a server, a personal computer (PC), or a notebook computer.
  • FIG. 1 is a schematic diagram of the hardware structure of the age recognition device involved in the solution of the embodiment of this application.
  • the age recognition device may include a processor and a memory.
  • the age identification device may also include a communication bus, a user interface, and/or a network interface.
  • the age identification device includes a processor 1001 (for example, a central processing unit, a CPU), a communication bus 1002, a user interface 1003, a network interface 1004, and a memory 1005.
  • the communication bus 1002 is used to realize the connection and communication between these components;
  • the user interface 1003 may include a display (Display), an input unit such as a keyboard (Keyboard);
  • the network interface 1004 may optionally include a standard wired interface, a wireless interface (Such as wireless fidelity WIreless-FIdelity, WI-FI interface);
  • the memory 1005 can be a high-speed random access memory (random access memory, RAM), or a stable memory (non-volatile memory), such as a disk memory, a memory
  • 1005 may also be a storage device independent of the foregoing processor 1001.
  • the hardware structure shown in FIG. 1 does not constitute a limitation to the present application, and may include more or less components than those shown in the figure, or a combination of certain components, or different component arrangements.
  • the memory 1005 as a computer-readable storage medium in FIG. 1 may include an operating system, a network communication module, and a computer program.
  • the network communication module can be used to connect to a preset database and perform data communication with the database; and the processor 1001 can call a computer program stored in the memory 1005 and execute the age identification method provided in the embodiment of the present application.
  • the embodiment of the present application provides an age identification method.
  • FIG. 2 is a schematic flowchart of the first embodiment of the age identification method of this application.
  • the age identification method includes the following steps:
  • Step S10 Obtain real speech samples from a preset database, and perform sample expansion on the real speech samples based on the generative confrontation network GAN to obtain expanded speech samples;
  • loan companies often identify the user’s age based on the user’s voice during the conversation, and then adopt different collection methods for collection according to the user’s age.
  • Traditional voice age recognition methods are mostly based on the statistical analysis of the voice signal characteristics of the voice to determine the age of the speaker; however, due to the limitation of the voice signal characteristics, this method has insufficient generalization ability.
  • the recognition accuracy rate in the application is low, and the application effect is not good.
  • this embodiment proposes an age identification method. First, large-scale data samples are obtained by data expansion of the generative confrontation network GAN.
  • the data samples can be more in line with the distribution of real data (ie Ensure the quality of the sample), and then use a sufficient number of real enough data samples to train and build an end-to-end network model, so that the model training process can more accurately understand the hidden laws of the data samples, improve the performance of the resulting network model, and then Subsequent use of the network model for the accuracy of age recognition; then the target speech to be recognized is converted into a spectrogram, and feature extraction is performed on the spectrogram through the obtained network model to obtain the in-depth characteristics of the target speech.
  • these deep features include more features, and can pay more attention to the difficult-to-recognizable age attribute representations in the target speech.
  • the age identification method in this embodiment is implemented by an age identification device.
  • the age identification device may be a server, a personal computer, a notebook computer, or other devices.
  • a server is used as an example for description.
  • the server in this embodiment may be a server in a collection system.
  • the server is connected to a preset database.
  • the database stores a number of real voice samples collected in advance. These real voice samples can be in the form of original voice or frequency spectrum. In the form of a graph; these real voice samples include corresponding sample annotations, and the annotation content includes the age group of the user to which the real voice sample belongs (of course, the annotation content can also include other information).
  • an age recognition network model for recognizing age.
  • the age recognition network model is constructed based on a deep neural network of machine learning. Considering that the real voice samples that can be obtained in practice may have the problem of data imbalance, and the number and quality of the samples have a greater impact on the training results (model capabilities) of the model. For this, in this embodiment, the real voice needs to be The sample is expanded by sample to obtain expanded voice samples, thereby obtaining large-scale data samples.
  • the real speech samples used should be in the form of spectrograms (including three-dimensional information of time, frequency, and amplitude). For real speech samples in the form of original speech, they must first pass the short-time Fourier The leaf transform (or other methods) converts it into the corresponding spectrogram.
  • the generative adversarial network GAN includes two sub-networks, which can be called generator G (Generator) and discriminator D (Discriminator);
  • G is a network that generates extended samples, which can pass a random noise Generate a simulated sample that follows the distribution of real speech samples as much as possible, denoted as G(z)
  • D is a discriminant network that determines whether the input sample is "real", if the output is 1, it means real, and the output is 0, it means no It may be real; in the training process, G's goal is to generate real simulation samples as much as possible to deceive D, and D's goal is to separate the simulation samples generated by G from the real voice samples as much as possible.
  • G can generate a simulated sample G(z) that is enough to "make it fake".
  • Step S20 Obtain an age recognition network model through the extended speech sample training
  • the server when the expanded voice sample is obtained, the server will train through the expanded voice sample to obtain the age recognition network model; for the convenience of subsequent processing, the age recognition network model can be set in an end-to-end form, that is, the age recognition
  • the input of the network model is speech
  • the output is the age group the speech belongs to.
  • the end-to-end approach does not require every process Separate labeling is helpful to reduce the workload of labeling, and at the same time, it is helpful to improve the accuracy of age recognition.
  • a deep network model can be used. For example, it can be implemented based on the classic deep residual network ResNet50. Of course, the ResNet50 architecture is used. At the same time, part of the structure can also be adjusted according to the actual situation.
  • Step S30 Obtain the target voice of the target user, and convert the target voice into a corresponding input spectrogram
  • the server can perform voice age recognition processing in the collection process through the age recognition network model. Specifically, when a collection item needs to be collected, first obtain the collection item information corresponding to the item, such as the loan user (target user) of a certain loan, contact information, etc., and then call the collection phone number according to the collection item information ; When the phone is connected, a general greeting voice can be played first to confirm its identity; when the target user makes a voice reply via the phone, the server can obtain the target voice of the target user, and then use the short-time Fourier transform The target voice is converted into the corresponding input spectrogram for subsequent analysis and processing.
  • the collection item information corresponding to the item such as the loan user (target user) of a certain loan, contact information, etc.
  • Step S40 Extract the depth feature of the input spectrogram through the age recognition network model, and determine the target age group to which the target user belongs according to the depth feature.
  • the input spectrogram when the input spectrogram is obtained, the input spectrogram can be extracted through the age recognition network model to obtain the corresponding depth feature, and then the target age group to which the target user belongs is determined according to the depth feature.
  • an attention mechanism can be introduced into the age recognition network model, or it can be considered as constructing an attention module to be embedded in the age recognition network model model, such as inserting into the age recognition network model.
  • a certain feature layer in the middle is then used for feature optimization (refine feature), and then the obtained optimized feature is subjected to subsequent processing (such as inputting the next layer, or using the feature as the final feature).
  • the age recognition network model in this embodiment includes an intermediate feature layer and a feature optimization layer, and the step S30 includes:
  • the age recognition network model in this embodiment includes an intermediate feature layer and a feature optimization layer.
  • the intermediate feature layer has the feature extraction functions of the general network intermediate layer (including convolution, pooling, etc.), and the feature optimization layer is based on attention Force mechanism construction.
  • the server may first extract the original features of the input spectrogram through the intermediate feature layer of the age recognition network model to obtain the corresponding original features.
  • this original feature it can be considered that it includes all the features of the input spectrogram, but these features are not necessarily related to age. If all of them are used for age recognition, the accuracy of recognition may be affected; at the same time, the amount of calculation is too large. This will affect the recognition speed; therefore, this embodiment will also refine these original features.
  • the original feature is optimized based on the attention mechanism to obtain the corresponding optimized feature, and the optimized feature is determined as the depth feature of the input spectrogram.
  • the server optimizes the original features through the feature optimization layer of the age recognition network model and based on the attention mechanism to obtain the corresponding optimized features.
  • the original feature it can be represented by the original feature map feature map and denoted as F; while the optimized feature can be optimized for the feature map representation, denoted as F".
  • F the feature map representation
  • R is the feature image (spectrogram) space
  • C is the image (spectrogram) channel number
  • H is the image high height
  • W is the image width.
  • the corresponding one-dimensional channel attention map channel attention map can be calculated, denoted as M C (F), and
  • each channel channel of F can be regarded as a feature vector feature detector, channel attention channel attention mainly focuses on what is meaningful in the input; and in order to calculate channel attention efficiently, this embodiment uses maximum pooling and Average pooling compresses F in the spatial dimension, and then obtains two different spatial background descriptions and Then use a shared network composed of a multi-layer perceptron MLP to calculate the two different spatial background descriptions to obtain the M C (F) of F, which is
  • ⁇ is the sigmoid function
  • W 0 ⁇ R C/r*C W 1 ⁇ R C*C/r
  • r is the compression ratio
  • W 0 uses Relu as the activation function.
  • H and W are as described above.
  • the main focus is on location information.
  • maximum pooling and average pooling to obtain two different feature descriptions for F'in the dimension of the channel. and Then use concatenation to merge the two feature descriptions, and use the convolution operation to generate M S (F') of F', that is
  • ⁇ is the sigmoid function
  • f 7*7 represents the 7*7 convolutional layer
  • F can be considered as an optimized feature.
  • the server can determine it as the depth feature of the input spectrogram, and perform age recognition processing based on the depth feature.
  • age The intermediate feature layer of the recognition network model can be two or more layers (here "above” includes the number, the same below), and the optimized feature layer can be after any intermediate feature layer.
  • the intermediate feature layer includes two layers, starting from the input Called the first layer and the second layer in turn; the feature optimization layer can be after the first layer and before the second layer.
  • the input of the feature optimization layer is the output of the first layer
  • the output of the feature optimization layer is the optimized feature of the output
  • the optimized feature layer can also be after the second layer, at this time the input of the optimized feature is the output of the second layer, and
  • the optimized features output by the feature optimization layer will be directly used as the final depth features for age recognition.
  • the target age group to which the target user belongs can be determined according to the depth feature.
  • the age recognition network model in this embodiment is an end-to-end form, so the age recognition process can be implemented in the output layer of the age recognition network model; in the output layer, the server will calculate the depth features and the data of each expanded voice sample.
  • the span range of the age group can be set according to the actual situation.
  • the server when the server determines the target age group to which the target user belongs, it can query the preset speech library according to the target age group to obtain the corresponding target speech template.
  • the target speech template may be pre-defined by relevant managers. Set and stored, there will be different speech templates for different age groups; when the server obtains the target speech template, it can broadcast the voice according to the target speech template, so as to perform voice collection on the target user.
  • real voice samples are obtained from a preset database, and the real voice samples are sample-expanded based on the generative confrontation network GAN to obtain expanded voice samples; the age recognition network model is obtained through training of the expanded voice samples; The target voice of the target user, and convert the target voice into a corresponding input spectrogram; extract the depth features of the input spectrogram through the age recognition network model, and determine the to which the target user belongs according to the depth characteristics Target age group.
  • this embodiment first obtains large-scale data samples by means of data expansion of the generative confrontation network GAN. While increasing the number of data samples, the data samples can be more in line with the distribution of real data (that is, the quality of the samples is guaranteed).
  • the model training process can more accurately understand the hidden laws of data samples, improve the performance of the resulting network model, and then use the network model for subsequent
  • the accuracy of age recognition then the target speech to be recognized is converted into a spectrogram, and the spectrogram is feature extracted through the obtained network model to obtain the depth characteristics of the target speech.
  • These in-depth features include more features, and can pay more attention to the age attribute representations that are not easily recognized in the target speech.
  • FIG. 3 is a schematic flowchart of a second embodiment of the age identification method of this application.
  • step S30 includes:
  • Step S31 Obtain the target voice of the target user, and determine whether the voice duration of the target voice is greater than a preset duration threshold;
  • the server obtains the target voice of the target user, it can first determine whether the voice duration of the target voice is greater than a preset duration threshold, and the preset duration threshold may be set according to actual needs.
  • Step S32 If the voice duration is greater than the preset duration threshold, perform voice cutting on the target voice to obtain two or more voice segments, and convert each voice segment into a corresponding segment spectrogram;
  • the server will perform voice cutting on the target voice to obtain more than two voice segments, and then convert each voice segment into a corresponding segment spectrogram.
  • the duration of each voice segment can be determined by defining different rules according to the actual situation. For example, when the voice duration of the target voice is greater than the preset duration threshold, different voice durations can correspond to the number of different voice fragments. For example, the preset duration threshold is 3 seconds, and the voice duration is greater than 3 seconds and not greater than 4 seconds.
  • the speech duration longer than 4 seconds corresponds to 3 speech fragments, and then the target speech can be averagely cut according to the determined speech fragments so that the fragment duration of each speech fragment is the same; for another example, when the speech of the target speech
  • the duration is greater than the preset duration threshold, it can be cut every other preset film length. If the preset duration threshold is 3 seconds, it can be cut every 3 seconds. For a voice duration of 5 seconds, cut It is two speech fragments, corresponding to the duration of 3 seconds and 2 seconds respectively. Of course, other cutting methods can also be used in practice. It is worth noting that if the voice duration of the target voice is less than or equal to the preset duration threshold, the entire target voice can be directly converted into the corresponding input spectrogram and the age recognition process in step S40 is performed.
  • the step S40 includes:
  • Step S41 Extract the depth features of each segment of the spectrogram through the age recognition network model, and respectively determine the segment age corresponding to each segment of the spectrogram according to the depth feature of each segment of the spectrogram;
  • the depth feature of each segment of the spectrogram can be extracted through the age recognition network model, and the segment age corresponding to each segment of the spectrogram can be determined according to the depth feature of each segment of the spectrogram.
  • the feature extraction process of each segment of the spectrogram and the segment age determination process please refer to the above step S40, which will not be repeated here.
  • Step S42 Determine the target age group to which the target user belongs according to the segment age group corresponding to each segment spectrogram.
  • the target age group to which the target user belongs can be determined.
  • the same segment age range is determined as the target age group to which the target user belongs; and if the segment age range corresponding to each segment spectrogram is not the same, it can be determined according to the actual.
  • the situation defines a voting decision rule to determine the target age group to which the target user belongs according to the rule and a plurality of determined segment age groups. For example, for the defined voting decision rule, the median average can be used.
  • the target voice corresponds to three fragment spectrograms, one of which corresponds to the age range of 22 to 24 years old, one corresponds to the age range of 26 to 28 years old, and one corresponds to If the age group is 28 to 30 years old, the median value of the three age groups is 23, 27, 29, and then the average of the three median values is 26.3 as the target age of the target user to determine the target age to which the target user belongs Segment; for the definition of voting decision rules, it can also be a majority determination method, such as the target voice corresponds to three segment spectrograms, two of which correspond to the age range of 22 to 24 years old, and one corresponds to the age range of 26 to 28 years old, then The age group 22 to 24 years old with the largest number of spectrograms corresponding to the segment can be determined as the target age group to which the target user belongs.
  • voting decision rules can also be used, such as calculating the credibility of each age group based on the speech duration of each segment of speech, and taking the age group with the highest credibility as the target age to which the target user belongs. Duan etc.
  • multiple segments of speech are obtained by cutting the speech with a long duration, and the age recognition is performed on each segment of the speech respectively, thereby reducing the amount of calculation and improving the efficiency of age recognition. Determining the target age group to which the target user belongs according to the recognition results is also beneficial to reduce the recognition error caused by accidental factors in the recognition process and improve the accuracy of age recognition.
  • FIG. 4 is a schematic flowchart of a third embodiment of an age identification method according to this application.
  • the method further includes:
  • Step S50 when receiving a collection instruction, dial a collection call according to the collection instruction, and obtain a corresponding connection voice after the call is connected;
  • the server when the server receives the collection instruction, it can obtain the corresponding collection item information according to the collection instruction, such as the loan user (target user) of a certain loan, contact information, etc., and then call the collection based on the collection item information Call, and get the connected voice of the other party after the call is connected.
  • the collection instruction can be triggered by a manager through a certain terminal, or a related collection plan stored in the server. When the time reaches the collection time set by the collection plan, the collection instruction is automatically triggered.
  • Step S60 judging whether there are more than two user voices in the connected voice
  • the server When the connected voice is obtained, the server will determine whether there are more than two user voices in the connected voice.
  • Step S70 If there are more than two user voices in the connected voice, the target voice of the target user is determined in each user voice according to the voice duration and/or voice volume of each user voice.
  • the target user when the target user answers the call, it may be in a noisy environment or talking with a person. At this time, there will be more than two user voices in the connected voice obtained by the server. There are more than two user voices in the voice, and the server needs to determine the target voice of the target user from them to accurately identify the age of the target user. Specifically, the server can distinguish each user's voice according to frequency, and then obtain the voice attributes of each user's voice, the voice attributes include voice frequency, voice duration, voice volume, etc., and then determine the target voice of the target user based on the voice attributes.
  • the target user is always holding the phone to make a call, so the user's voice with the longest voice duration can be determined as the target voice of the target user; for another example, the target user is the user closest to the phone.
  • the target user’s voice is relatively loudest, so the user’s voice with the largest voice volume can be determined as the target voice of the target user; of course, the above two factors can also be combined.
  • For a user’s voice according to the duration of the voice and the voice The volume can obtain the corresponding time length points and volume points respectively, and then add the time length points and the volume points to obtain a comprehensive score, and determine the user voice with the highest comprehensive score as the target voice of the target user. It is worth noting that if there is only one user voice in the connected voice, the user voice can be directly used as the target voice.
  • the relevant age recognition processing of steps S30 and S40 can be executed.
  • the server of this embodiment when the server of this embodiment is performing call collection, if there are more than two user voices in the connected voice, the target voice can be determined first, and then the subsequent age recognition processing can be performed to avoid the voice of multiple users.
  • the method further includes:
  • the server when it obtains the target age group to which the current target user belongs, it may also store it.
  • these historical target age ranges can be summarized and counted to obtain the corresponding collection age distribution. For example, among 100 historical target users, 30 belong to the age range of 26 to 28, and 70 belong to the age range of 30 to 32; for example, in the collection cycle last month, there were a total of 100 historical target users. Among them, 30 are in the age group of 26 to 28, and 70 are in the age group of 30 to 32.
  • the collection age distribution it is determined whether there is an age group with an abnormal number of users, and if there is, the corresponding abnormal prompt information is sent to the corresponding management terminal.
  • the server when it obtains the collection age distribution, it can determine whether there is an age group with an abnormal number of users based on the collection age distribution. For the abnormality judgment, the corresponding abnormality rule can be set in advance, and then the abnormality judgment is made based on the rule, such as the threshold or proportion of the abnormal number corresponding to the age group, when the number of users corresponding to a certain age group exceeds the threshold or the proportion of the abnormal number , It can be considered that the number of users in this age group is abnormal. In this embodiment, if there is an age group with an abnormal number of users, it may be that the current loan business has a certain risk in this age group, or it may be that the recognition ability of the age recognition network model is reduced, causing too many users to be identified as such. Age group, at this time, the server can send corresponding abnormal prompt information to the corresponding management terminal to prompt the relevant management personnel to inspect and deal with it in time.
  • the corresponding abnormal prompt information to the corresponding management terminal to prompt the relevant management personnel to inspect and deal with it in time
  • the server of this embodiment can analyze the collection age distribution of historical target users, and analyze whether there is an abnormal situation according to the collection age distribution, which is conducive to timely detection of abnormal situations, reducing business risks and the stability of the age identification network model sex.
  • an embodiment of the present application also provides an age identification device.
  • the age identification device includes:
  • the sample expansion module is used to obtain real voice samples from a preset database, and perform sample expansion on the real voice samples based on the generative confrontation network GAN to obtain expanded voice samples;
  • a model training module is used to obtain an age recognition network model through the extended speech sample training
  • the voice conversion module is used to obtain the target voice of the target user and convert the target voice into a corresponding input spectrogram
  • the age determination module is configured to extract the depth characteristics of the input spectrogram through the age recognition network model, and determine the target age group to which the target user belongs according to the depth characteristics.
  • each virtual function module of the above-mentioned age recognition device is stored in the memory 1005 of the age recognition device shown in FIG.
  • the age recognition network model includes an intermediate feature layer and a feature optimization layer
  • the age determination module includes:
  • a feature extraction unit configured to perform original feature extraction on the input spectrogram through the intermediate feature layer of the age recognition network model to obtain corresponding original features
  • the feature optimization unit is used to perform feature optimization on the original feature based on the attention mechanism through the feature optimization layer of the age recognition network model to obtain the corresponding optimized feature, and determine the optimized feature as the input spectrogram The depth characteristics.
  • the original feature includes an original feature map F
  • the optimized feature includes an optimized feature map F".
  • the feature optimization unit is further configured to obtain the feature optimization layer through the age recognition network model.
  • the original feature map F of the original feature calculate the one-dimensional channel attention map corresponding to the F channel attention map; perform element-wise multiplication on the F and the channel attention map to obtain the corresponding intermediate feature map F'
  • Calculate the two-dimensional channel attention map spatial attention map corresponding to the F' perform element-wise multiplication on the F'and the spatial attention map to obtain the corresponding optimized feature map F.
  • the voice conversion module includes:
  • the duration judging unit is configured to obtain the target voice of the target user, and determine whether the voice duration of the target voice is greater than a preset duration threshold;
  • the voice segmentation unit is configured to perform voice cutting on the target voice if the voice duration is greater than the preset duration threshold to obtain two or more voice segments, and respectively convert each voice segment into a corresponding segment spectrogram ;
  • the age determination module is further configured to extract the depth characteristics of each segment of the spectrogram through the age recognition network model, and determine the segment age corresponding to each segment of the spectrogram according to the depth characteristics of each segment of the spectrogram;
  • the segment age group corresponding to the spectrogram determines the target age group to which the target user belongs.
  • the age identification device further includes:
  • the voice acquisition module is configured to, when receiving the collection instruction, dial a collection call according to the collection instruction, and obtain the corresponding connected voice after the call is connected;
  • the voice judgment module is used to judge whether there are more than two user voices in the connected voice
  • the voice determining module is configured to determine the target voice of the target user in each user voice according to the voice duration and/or voice volume of each user voice if there are more than two user voices in the connected voice.
  • the age identification device further includes:
  • the voice collection module is used to obtain the corresponding target speech template according to the target age group, and perform voice collection on the target user according to the target speech template.
  • the age identification device further includes:
  • the distribution acquisition module is used to acquire the historical target age range of the historical target users of a preset number or a preset period, and obtain the collection age distribution according to the historical target age range;
  • the abnormality judgment module is used for judging whether there is an age group with an abnormal number of users based on the collection age distribution, and if it exists, sending corresponding abnormality prompt information to the corresponding management terminal.
  • each module in the above-mentioned age recognition device corresponds to each step in the above-mentioned embodiment of the age recognition method, and the functions and realization processes thereof will not be repeated here.
  • the embodiment of the present application also provides a computer-readable storage medium.
  • the computer-readable storage medium of the present application stores a computer program, and when the computer program is executed by a processor, the steps of the above-mentioned age identification method are realized.
  • the storage medium involved in this application such as a computer-readable storage medium, may be non-volatile or volatile.
  • the technical solution of this application essentially or the part that contributes to the existing technology can be embodied in the form of a software product, and the computer software product is stored in a storage medium (such as ROM/RAM) as described above. , Magnetic disks, optical disks), including several instructions to make a terminal device (which can be a mobile phone, a computer, a server, an air conditioner, or a network device, etc.) execute the method described in each embodiment of the present application.
  • a terminal device which can be a mobile phone, a computer, a server, an air conditioner, or a network device, etc.

Landscapes

  • Engineering & Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • Computational Linguistics (AREA)
  • Health & Medical Sciences (AREA)
  • Theoretical Computer Science (AREA)
  • Artificial Intelligence (AREA)
  • Evolutionary Computation (AREA)
  • Computing Systems (AREA)
  • Software Systems (AREA)
  • Biophysics (AREA)
  • General Health & Medical Sciences (AREA)
  • Molecular Biology (AREA)
  • Biomedical Technology (AREA)
  • General Engineering & Computer Science (AREA)
  • General Physics & Mathematics (AREA)
  • Mathematical Physics (AREA)
  • Data Mining & Analysis (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Multimedia (AREA)
  • Signal Processing (AREA)
  • Audiology, Speech & Language Pathology (AREA)
  • Human Computer Interaction (AREA)
  • Acoustics & Sound (AREA)
  • Spectroscopy & Molecular Physics (AREA)
  • Information Retrieval, Db Structures And Fs Structures Therefor (AREA)

Abstract

一种年龄识别方法、装置、设备及计算机可读存储介质，涉及人工智能技术领域，方法包括：从预设数据库中获取真实语音样本，并基于生成式对抗网络GAN对所述真实语音样本进行样本扩充，得到扩充语音样本（S10）；通过所述扩充语音样本训练得到年龄识别网络模型（S20）；获取目标用户的目标语音，并将所述目标语音转换为对应的输入频谱图（S30）；通过所述年龄识别网络模型提取所述输入频谱图的深度特征，并根据所述深度特征确定所述目标用户所属的目标年龄段（S40）。利用上述方法可提高年龄识别的准确性。

Description

年龄识别方法、装置、设备及计算机可读存储介质
本申请要求于2020年2月12日提交中国专利局、申请号为202010094834.9，发明名称为“年龄识别方法、装置、设备及计算机可读存储介质”的中国专利申请的优先权，其全部内容通过引用结合在本申请中。
技术领域
本申请涉及人工智能技术领域，尤其涉及一种年龄识别方法、装置、设备及计算机可读存储介质。
背景技术
发明人发现，目前，贷款公司在进行催收时，为了增强用户体验和催收效果，往往会根据对话过程中的用户语音对用户年龄进行识别，然后根据用户年龄采用不同的催收方式进行催收。发明人意识到，传统的语音年龄识别的方法，大多是基于声音的语音信号学特征进行统计学分析，进而确定说话者的年龄；但这种方法由于语音信号学特征的限制，其泛化能力不足，在实际应用中识别准确率低，应用效果不佳。
发明内容
本申请的主要目的在于提供一种年龄识别方法、装置、设备及计算机可读存储介质，旨在解决传统的年龄识别准确性低的技术问题。
为实现上述目的，本申请实施例提供一种年龄识别方法，所述年龄识别方法包括：
从预设数据库中获取真实语音样本，并基于生成式对抗网络GAN对所述真实语音样本进行样本扩充，得到扩充语音样本；
通过所述扩充语音样本训练得到年龄识别网络模型；
获取目标用户的目标语音，并将所述目标语音转换为对应的输入频谱图；
通过所述年龄识别网络模型提取所述输入频谱图的深度特征，并根据所述深度特征确定所述目标用户所属的目标年龄段。
此外，为实现上述目的，本申请实施例还提供一种年龄识别装置，所述年龄识别装置包括：
样本扩充模块，用于从预设数据库中获取真实语音样本，并基于生成式对抗网络GAN对所述真实语音样本进行样本扩充，得到扩充语音样本；
模型训练模块，用于通过所述扩充语音样本训练得到年龄识别网络模型；
语音转换模块，用于获取目标用户的目标语音，并将所述目标语音转换为对应的输入频谱图；
年龄确定模块，用于通过所述年龄识别网络模型提取所述输入频谱图的深度特征，并根据所述深度特征确定所述目标用户所属的目标年龄段。
此外，为实现上述目的，本申请实施例还提供一种年龄识别设备，所述年龄识别设备包括处理器、存储器、以及存储在所述存储器上并可被所述处理器执行的计算机程序，其中所述计算机程序被所述处理器执行时，实现如上述的年龄识别方法，该年龄识别方法包括以下步骤：
从预设数据库中获取真实语音样本，并基于生成式对抗网络GAN对所述真实语音样本进行样本扩充，得到扩充语音样本；
通过所述扩充语音样本训练得到年龄识别网络模型；
获取目标用户的目标语音，并将所述目标语音转换为对应的输入频谱图；
通过所述年龄识别网络模型提取所述输入频谱图的深度特征，并根据所述深度特征确定所述目标用户所属的目标年龄段。
此外，为实现上述目的，本申请实施例还提供一种计算机可读存储介质，所述计算机可读存储介质上存储有计算机程序，其中所述计算机程序被处理器执行时，实现如上述的 年龄识别方法，该年龄识别方法包括以下步骤：
从预设数据库中获取真实语音样本，并基于生成式对抗网络GAN对所述真实语音样本进行样本扩充，得到扩充语音样本；
通过所述扩充语音样本训练得到年龄识别网络模型；
获取目标用户的目标语音，并将所述目标语音转换为对应的输入频谱图；
通过所述年龄识别网络模型提取所述输入频谱图的深度特征，并根据所述深度特征确定所述目标用户所属的目标年龄段。
本申请实施例能够提高年龄识别的泛化能力，提高年龄识别的准确性。
附图说明
图1为本申请实施例方案中涉及的年龄识别设备的硬件结构示意图；
图2为本申请年龄识别方法第一实施例的流程示意图；
图3为本申请年龄识别方法第二实施例的流程示意图；
图4为本申请年龄识别方法第三实施例的流程示意图。
本申请目的的实现、功能特点及优点将结合实施例，参照附图做进一步说明。
具体实施方式
应当理解，此处所描述的具体实施例仅仅用以解释本申请，并不用于限定本申请。
本申请的技术方案可应用于人工智能、智慧城市、数字医疗、区块链和/或大数据技术领域。可选的，本申请涉及的数据如语音样本、深度特征和/或确定出的年龄段信息等可存储于数据库中，或者可以存储于区块链中，比如通过区块链分布式存储，本申请不做限定。
本申请实施例涉及的年龄识别方法主要应用于年龄识别设备，该年龄识别设备可以是服务器、个人计算机(personal computer，PC)、笔记本电脑等具有数据处理功能的设备。
参照图1，图1为本申请实施例方案中涉及的年龄识别设备的硬件结构示意图。本申请实施例中，该年龄识别设备可以包括处理器和存储器。可选的，该年龄识别设备还可包括通信总线、用户接口和/或网络接口。例如，该年龄识别设备包括处理器1001(例如中央处理器Central Processing Unit，CPU)，通信总线1002，用户接口1003，网络接口1004，存储器1005。其中，通信总线1002用于实现这些组件之间的连接通信；用户接口1003可以包括显示屏(Display)、输入单元比如键盘(Keyboard)；网络接口1004可选的可以包括标准的有线接口、无线接口(如无线保真WIreless-FIdelity，WI-FI接口)；存储器1005可以是高速随机存取存储器(random access memory，RAM)，也可以是稳定的存储器(non-volatile memory)，例如磁盘存储器，存储器1005可选的还可以是独立于前述处理器1001的存储装置。本领域技术人员可以理解，图1中示出的硬件结构并不构成对本申请的限定，可以包括比图示更多或更少的部件，或者组合某些部件，或者不同的部件布置。
继续参照图1，图1中作为一种计算机可读存储介质的存储器1005可以包括操作系统、网络通信模块以及计算机程序。在图1中，网络通信模块可用于连接预设数据库，与数据库进行数据通信；而处理器1001可以调用存储器1005中存储的计算机程序，并执行本申请实施例提供的年龄识别方法。
基于上述的硬件架构，提出本申请年龄识别方法的各实施例。
本申请实施例提供了一种年龄识别方法。
参照图2，图2为本申请年龄识别方法第一实施例的流程示意图。
本实施例中，所述年龄识别方法包括以下步骤：
步骤S10，从预设数据库中获取真实语音样本，并基于生成式对抗网络GAN对所述真实语音样本进行样本扩充，得到扩充语音样本；
目前，贷款公司在进行催收时，为了增强用户体验和催收效果，往往会根据对话过程中的用户语音对用户年龄进行识别，然后根据用户年龄采用不同的催收方式进行催收。传 统的语音的年龄识别的方法，大多是基于声音的语音信号学特征进行统计学分析，进而确定说话者的年龄；但这种方法由于语音信号学特征的限制，其泛化能力不足，在实际应用中识别准确率低，应用效果不佳。对此，本实施例提出一种年龄识别方法，首先通过生成式对抗网络GAN数据扩充的方式获得大规模数据样本，在提升数据样本数量的同时，使得数据样本能够更加符合真实数据的分布(即保证样本的质量)，再利用足够数量、足够真实的数据样本训练构建端对端的网络模型，从而使得模型训练过程能够更准确地理解数据样本的隐藏规律，提高所得到的网络模型的性能，进而后续使用该网络模型进行年龄识别的准确性；然后将待识别的目标语音转化为频谱图，通过得到的网络模型对该频谱图进行特征提取，获得目标语音的深度特征，与传统的基于信号学特征进行年龄识别相比，这些深度特征所包括的特征数更多，且更能关注到目标语音中不易识别的年龄属性表征，通过使用该深度特征识别目标用户所属的目标年龄段，有利于准确把握年龄与语音之间的关联关系，进而提高年龄识别的泛化能力，提高年龄识别的准确性。
本实施例中的年龄识别方法是由年龄识别设备实现的，该年龄识别设备可以是服务器、个人计算机、笔记本电脑等设备，本实施例中以服务器为例进行说明。本实施例中的服务器可以是催收系统中的服务器，服务器与一预设数据库连接，该数据库中存储有预先收集的若干真实语音样本，这些真实语音样本可以是原始语音的形式，也可以是频谱图的形式；这些真实语音样本包括有对应的样本批注，批注内容包括真实语音样本所属用户的年龄段(当然批注内容还可以包括其它信息)。
本实施例中，在进行年龄识别前，首先需要训练构建识别年龄所用的年龄识别网络模型，该年龄识别网络模型是基于机器学习的深度神经网络的方式构建。考虑到在实际中能获得的真实语音样本可能存在数据不平衡的问题，而样本数量和质量对模型的训练结果(模型能力)具有较大的影响，对此，本实施例中需要对真实语音样本进行样本扩充，得到扩充语音样本，从而得到大规模数据样本。
本实施例在进行样本扩充时，为了提高样本扩充的效率、并保证样本扩充后的样本指令，可基于生成式对抗网络GAN的方式进行。值得说明的是，进行样本扩充时，所使用的真实语音样本应该是频谱图的形式(包括时间、频率、振幅三维信息)，对于原始语音的形式的真实语音样本，首先要通过短时傅里叶变换(或其它方式)将其转换为对应的频谱图。其中，生成式对抗网络GAN(Generative Adversarial Networks)包括两个子网络，可分别称为生成器G(Generator)和判别器D(Discriminator)；G是一个生成扩充样本的网络，可通过一个随机的噪声生成一个尽量服从真实语音样本分布的模拟样本，记做G(z)，D是一个判别网络，判别输入的样本是否“真实”，如输出为1，则代表真实，输出为0，则代表不可能真实；在训练过程中，G的目标就是尽量生成真实的模拟样本去欺骗D，而D的目标就是尽量把G生成的模拟样本和真实语音样本分别开来。在最理想的状态下，G可以生成足以“以假乱真”的模拟样本G(z)，对于D来说，它难以判定G生成的模拟样本G(z)究竟是不是真实的，也即D(G(z))＝0.5；当满足该条件时，即可认为得到一个训练完成的G(也即GAN训练完成)，并用以对真实语音样本进行样本扩展，得到扩充语音样本。
步骤S20，通过所述扩充语音样本训练得到年龄识别网络模型；
本实施例中，在得到扩充语音样本时，服务器将通过扩充语音样本进行训练，得到年龄识别网络模型；为了后续处理方便，该年龄识别网络模型可设置为端对端的形式，也即该年龄识别网络模型的输入为语音，输出为该语音所属的年龄段，与传统的“通过一模型进行特征提取、再通过另一模型进行分类”的方式相比，端对端的方式不需要对每一个过程分别进行标注，有利于减少标注的工作量，同时有利于提高年龄识别的准确性。而对于本实施例的年龄识别网络模型，为了提高模型泛化性和识别准确性，可采用深度网络模型的方式，例如可基于经典深度残差网络ResNet50的方式实现，当然，在采用ResNet50的 架构时，也可以根据实际情况对部分结构进行调整。
步骤S30，获取目标用户的目标语音，并将所述目标语音转换为对应的输入频谱图；
本实施例中，在得到年龄识别网络模型时，服务器即可通过该年龄识别网络模型在催收过程进行语音年龄识别处理。具体的，当需要对某一催收项目进行催收时，首先可获取该项目对应的催收项目信息，例如某一笔贷款的贷款用户(目标用户)、联系方式等，然后根据催收项目信息拨打催收电话；当电话接通时，可先播放一段通用的问候语音，以确认其身份；当目标用户通过电话进行语音回复时，服务器即可获取到目标用户的目标语音，然后通过短时傅里叶变换的方式将该目标语音转换为对应的输入频谱图，用以进行后续分析处理。
步骤S40，通过所述年龄识别网络模型提取所述输入频谱图的深度特征，并根据所述深度特征确定所述目标用户所属的目标年龄段。
本实施例中，在得到输入频谱图时，即可通过年龄识别网络模型对输入频谱图进行特征提取，得到对应的深度特征，然后根据深度特征确定目标用户所属的目标年龄段。
进一步的，对于语音信号而言，同时具有时域和频域的属性，这两类属性在频谱图中会有相应的体现，并对应了各种特征，这些特征可能与年龄有关，可能与年龄无关(如环境噪音特征)。为了提高年龄识别的准确性，本实施例中在进行特征提取时，可在年龄识别网络模型中引入注意力机制，或认为是构造一个注意力模块嵌入到年龄识别网络模型模型中，如插入到中间的某个特征层之后用于特征优化(refine feture)，然后再将得到的优化特征进行后续处理(如输入下一层，或将该特征作为最终特征)。本实施例中的年龄识别网络模型包括中间特征层和特征优化层，所述步骤S30包括：
通过所述年龄识别网络模型的中间特征层对所述输入频谱图进行原始特征提取，得到对应的原始特征；
本实施例中的年龄识别网络模型包括中间特征层和特征优化层，其中，中间特征层具有一般网络中间层的特征提取功能(包括卷积、池化等)，而特征优化层则是基于注意力机制构建。在得到输入频谱图时，服务器可先通过年龄识别网络模型的中间特征层对输入频谱图进行原始特征提取，得到对应的原始特征。对于该原始特征，可认为是包括了输入频谱图的所有特征，但这些特征不一定与年龄都有关，若全部使用进行年龄识别，可能会影响识别的准确性；同时由于计算量过大，也将影响识别速度；因此本实施例还将对这些原始特征进行特征优化(refine feture)。
通过所述年龄识别网络模型的特征优化层、基于注意力机制对所述原始特征进行特征优化，得到对应的优化特征，并将所述优化特征确定为所述输入频谱图的深度特征。
在得到原始特征时，服务器将通过年龄识别网络模型的特征优化层、基于注意力机制对所述原始特征进行特征优化，得到对应的优化特征。具体的，在对于原始特征，可以是以原始特征图feature map的方式进行表示，并记为F；而优化特征则可以优化特征图表示，记为F”。在进行特征优化时，可通过年龄识别网络模型的特征优化层获取原始特征的原始特征图F，且有
F∈R C*H*W
其中，R为特征图像(频谱图)空间，C为图像(频谱图)channel通道数，H为图像high高度，W为图像width宽度。
对于该F，可计算得到其对应的一维通道注意力图channel attention map，记为M C(F)，且有
M C(F)∈R C*1*1
其中，F的每个通道channel都可视为一个特征向量feature detector，通道注意力channel attention主要关注于输入中什么是有意义的；而为了高效计算channel attention，本实施例使用分别最大池化和平均池化对F在空间维度上进行压缩，进而得到两个不同的空间背景描述
Figure PCTCN2021071262-appb-000001
和
Figure PCTCN2021071262-appb-000002
然后使用由多层感知机MLP组成的共享网络对这两个不同的空间背景描述进行计算得到F的M C(F)，也即
Figure PCTCN2021071262-appb-000003
其中，σ为sigmoid函数，W 0∈R C/r*C，W 1∈R C*C/r，r为压缩率，W 0使用Relu作为激活函数。
在得到channel attention map(M C(F))时，可对所述F和所述channel attention map进行元素相乘element-wise multiplication，得到对应的中间特征图F'，也即
Figure PCTCN2021071262-appb-000004
其中，
Figure PCTCN2021071262-appb-000005
为element-wise multiplication。
在得到F'时，可计算得到其对应的二维通道注意力图spatial attention map，记为M S(F')，且有
M S(F')∈R 1*H*W
其中，H、W含义见上述。对于该spatial attention主要关注位置信息。在计算spatial attention时，首先在频道channel的维度对F'分别使用最大池化和平均池化得到两个不同的特征描述
Figure PCTCN2021071262-appb-000006
和
Figure PCTCN2021071262-appb-000007
然后使用使用concatenation将两个特征描述合并，并使用卷积操作生成F'的M S(F')，也即
Figure PCTCN2021071262-appb-000008
其中，σ为sigmoid函数，f 7*7表示7*7的卷积层，
Figure PCTCN2021071262-appb-000009
为F'在channel维度的最大池化的特征描述，
Figure PCTCN2021071262-appb-000010
为F'在channel维度的平均池化的特征描述。
在得到spatial attention map(M S(F'))时，可对F'和spatial attention map(M S(F'))进行element-wise multiplication，得到对应的优化特征图F”，也即
Figure PCTCN2021071262-appb-000011
上式中F”即可认为是优化特征，对于该优化特征，服务器可将其确定为输入频谱图 的深度特征，并根据该深度特征进行年龄识别处理。值得说明的是，在实际中，年龄识别网络模型的中间特征层可以是两层以上(此处“以上”包括本数，下同)，而优化特征层可以是在任一中间特征层之后。例如，中间特征层包括两层，从输入开始依次称为第一层和第二层；特征优化层可以是在第一层之后、第二层之前，此时特征优化层的输入为第一层的输出，而特征优化层的输出的优化特征将作为第二层的输入，经第二层处理后得到最终的深度特征用以进行年龄识别；优化特征层也可以是第二层之后，此时优化特征的输入为第二层的输出，而特征优化层的输出的优化特征将直接作为最终的深度特征用以进行年龄识别。
在得到深度特征时，即可根据该深度特征确定目标用户所属的目标年龄段。本实施例中的年龄识别网络模型是端对端的形式，因而年龄识别的过程可以是在年龄识别网络模型的输出层实现的；在该输出层中，服务器会计算深度特征与各扩充语音样本的样本特征的空间距离，并将空间距离最小的扩充语音样本作为与输入频谱图(目标语音)匹配的目标样本；然后可查询目标样本的样本批注，从而确定目标样本对应的样本年龄，该样本年龄也可认为是目标语音对应的语音年龄，根据该语音年龄即可确定目标用户所属的目标年龄段。当然，对于年龄段的跨度范围，可以根据实际情况进行设置。
进一步的，服务器在确定目标用户所属的目标年龄段时，即可根据该目标年龄段查询预设话术库，以获取对应的目标话术模板，该目标话术模板可以是由相关管理人员预先设置并存储的，对于不同的年龄段会对于有不同的话术模板；服务器在得到目标话术模板时，即可根据目标话术模板进行语音播报，从而对目标用户进行语音催收。
本实施例通过从预设数据库中获取真实语音样本，并基于生成式对抗网络GAN对所述真实语音样本进行样本扩充，得到扩充语音样本；通过所述扩充语音样本训练得到年龄识别网络模型；获取目标用户的目标语音，并将所述目标语音转换为对应的输入频谱图；通过所述年龄识别网络模型提取所述输入频谱图的深度特征，并根据所述深度特征确定所述目标用户所属的目标年龄段。通过以上方式，本实施例首先通过生成式对抗网络GAN数据扩充的方式获得大规模数据样本，在提升数据样本数量的同时，使得数据样本能够更加符合真实数据的分布(即保证样本的质量)，再利用足够数量、足够真实的数据样本训练构建端对端的网络模型，从而使得模型训练过程能够更准确地理解数据样本的隐藏规律，提高所得到的网络模型的性能，进而后续使用该网络模型进行年龄识别的准确性；然后将待识别的目标语音转化为频谱图，通过得到的网络模型对该频谱图进行特征提取，获得目标语音的深度特征，与传统的基于信号学特征进行年龄识别相比，这些深度特征所包括的特征数更多，且更能关注到目标语音中不易识别的年龄属性表征，通过使用该深度特征识别目标用户所属的目标年龄段，有利于准确把握年龄与语音之间的关联关系，进而提高年龄识别的泛化能力，提高年龄识别的准确性。
基于上述图2所示实施例，提出本申请年龄识别方法第二实施例。
参照图3，图3为本申请年龄识别方法第二实施例的流程示意图。
本实施例中，所述步骤S30包括：
步骤S31，获取目标用户的目标语音，并判断所述目标语音的语音时长是否大于预设时长阈值；
当目标语音持续较长时间时，若直接对目标语音进行年龄识别的处理，可能会导致模型处理过程的计算量过大，对此本实施例中可以是将时长较长的语音进行切割处理，得到多段语音，并分别对各段语音进行年龄识别，从而减少计算量，并且还有利于提高年龄识别的准确性。具体的，服务器在获取到目标用户的目标语音时，首先可判断该目标语音的语音时长是否大于预设时长阈值，该预设时长阈值可以是根据实际需要进行设置。
步骤S32，若所述语音时长大于所述预设时长阈值，则对所述目标语音进行语音切割， 得到两段以上的语音片段，并分别将各语音片段转换为对应的片段频谱图；
本实施例中，若目标语音的语音时长大于预设时长阈值，则服务器会对目标语音进行语音切割，得到两段以上的语音片段，然后分别将各语音片段转换为对应的片段频谱图。在进行语音切割时，对于各段语音片段时长，可以根据实际情况定义不同的规则进行确定。例如，当目标语音的语音时长大于预设时长阈值时，对于不同的语音时长，则可以对应不同语音片段数量，如预设时长阈值为3秒，则大于3秒且不大于4秒的语音时长对应2个语音片段，大于4秒的语音时长对应3个语音片段，然后可根据确定的语音片段对目标语音进行均值切割，以使各语音片段的片段时长相同；又例如，当目标语音的语音时长大于预设时长阈值时，可以是每隔一预设片长进行一次切割，如预设时长阈值为3秒，则可以是每隔3秒进行一次切割，对于5秒的语音时长，则切割为两个语音片段，分别对应时长为3秒和2秒。当然在实际中还可以是其它切割方式。值得说明的是，若目标语音的语音时长小于或等于预设时长阈值，则可直接将整段目标语音转换为对应的输入频谱图并执行上述步骤S40的年龄识别处理。
所述步骤S40包括：
步骤S41，通过所述年龄识别网络模型分别提取各片段频谱图的深度特征，并根据各片段频谱图的深度特征分别确定各片段频谱图对应的片段年龄段；
本实施例中，在得到各片段频谱图时，即可通过年龄识别网络模型分别提取各片段频谱图的深度特征，并根据各片段频谱图的深度特征分别确定各片段频谱图对应的片段年龄段。而对于每个片段频谱图的特征提取过程和片段年龄段确定过程可参照上述步骤S40，此处不再赘述。
步骤S42，根据各片段频谱图对应的片段年龄段确定所述目标用户所属的目标年龄段。
本实施例中，在确定各片段频谱图对应的片段年龄段时，即可确定所述目标用户所属的目标年龄段。其中，若各片段频谱图对应的片段年龄段相同，则将该相同的片段年龄段确定为目标用户所属的目标年龄段；而若各片段频谱图对应的片段年龄段不相同，则可根据实际情况定义投票决策规则，以根据该规则和多个确定的片段年龄段确定目标用户所属的目标年龄段。例如，对于该定义投票决策规则，可以是中值平均的方式，如目标语音对应三个片段频谱图，其中一个对应年龄段为22到24岁，一个对应年龄段为26到28岁，一个对应年龄段为28到30岁，则可根据分别取三个年龄段的中值23、27、29，然后取三个中值的均值26.3作为目标用户的目标年龄，进而确定目标用户所属的目标年龄段；对于该定义投票决策规则，也可以是多数确定的方式，如目标语音对应三个片段频谱图，其中两个对应年龄段为22到24岁，一个对应年龄段为26到28岁，则可将对应片段频谱图数量最多的年龄段22到24岁确定为目标用户所属的目标年龄段。当然除了上述举例外，还可以采用其它形式的投票决策规则，例如根据各片段语音的语音时长加权计算各年龄段的可信度、并将可信度最高的年龄段作为目标用户所属的目标年龄段等。
本实施例中，通过对时长较长的语音进行切割处理，得到多段语音，在分别对各段语音进行年龄识别，从而可减少运算量，提高年龄识别效率，并且由于对各段语音识别后对再根据各识别结果确定目标用户所属的目标年龄段，还有利于减少识别过程偶然因素导致的识别误差，提高了年龄识别的准确性。
基于上述图2所示实施例，提出本申请年龄识别方法第三实施例。
参照图4，图4为本申请年龄识别方法第三实施例的流程示意图。
本实施例中，所述步骤S30之后，还包括：
步骤S50，在接收到催收指令时，根据所述催收指令拨打催收电话，并在电话接通后获取对应的接通语音；
本实施例中，当服务器接收到催收指令时，即可根据该催收指令获取对应的催收项目 信息，例如某一笔贷款的贷款用户(目标用户)、联系方式等，然后根据催收项目信息拨打催收电话，并在电话接通后获取对方的接通语音。其中，该催收指令可以是管理人员通过某一终端触发，还可以是服务器中存放有相关的催收计划，当时间到达该催收计划所设置的催收时间时，自动触发催收指令。
步骤S60，判断所述接通语音中是否存在两个以上的用户语音；
在得到接通语音时，服务器将判断该接通语音中是否存在两个以上的用户语音。
步骤S70，若所述接通语音中存在两个以上的用户语音，则根据各用户语音的语音持续时长和/或语音音量在各用户语音中确定目标用户的目标语音。
本实施例中，目标用户在接听电话时，可能是在嘈杂的环境或是正在与人交谈，此时服务器所获取到的接通语音会存在两个以上的用户语音，对此，若接通语音中存在两个以上的用户语音，服务器需要从中确定出目标用户的目标语音，以准确识别目标用户的年龄。具体的，服务器可根据频率将各用户语音区分，然后获取各用户语音的语音属性，该语音属性包括语音频率、语音持续时长、语音音量等，然后根据该语音属性确定出目标用户的目标语音。例如，在实际中，目标用户是一直手持电话进行通话的，因此可以是将语音持续时长最长的用户语音确定为目标用户的目标语音；又例如，目标用户距离电话最近的用户，因此对电话目标用户的声音相对来说是最大的，因此可以是将语音音量最大的用户语音确定为目标用户的目标语音；当然还可以将上述两种因素结合，对于一个用户语音，根据语音持续时长和语音音量可分别获得对应的时长分和音量分，然后将时长分和音量分进行加和得到综合分，并将综合分最高的用户语音确定为目标用户的目标语音。值得说说明的是，若接通语音仅存在一个的用户语音，则可直接将该用户语音作为目标语音。在确定目标用户的目标语音时，即可执行上述步骤S30和S40的相关年龄识别处理。
通过以上方式，本实施例的服务器在进行电话催收时，若接通语音中存在两个以上的用户语音，则可先从中确定出目标语音，再进行后续年龄识别处理，避免对多个用户语音进行年龄识别所可能导致的识别错误问题，提高年龄识别的准确性。
基于上述图2所示实施例，提出本申请年龄识别方法第四实施例。
本实施例中，所述步骤S40之后，还包括：
获取预设数量或预设周期的历史目标用户的历史目标年龄段，并根据所述历史目标年龄段获得催收年龄分布；
本实施例中，服务器在得到当前的目标用户所属的目标年龄段时，还可将其进行存储。当收集到预设数量的历史目标用户或某一预设周期内的历史目标用户的历史目标年龄段时，可以对这些历史目标年龄段进行汇总和统计，得到对应的催收年龄分布。例如，在100位历史目标用户中，有30位属于26到28岁年龄段，有70位属于30到32岁年龄段；又例如在上个月的催收周期中，共有100位历史目标用户，其中有30位属于26到28岁年龄段，有70位属于30到32岁年龄段。
根据所述催收年龄分布判断是否存在用户数量异常的年龄段，若存在，则向对应管理终端发送对应的异常提示信息。
本实施例中，服务器在得到催收年龄分布时，可根据该催收年龄分布判断是否存在用户数量异常的年龄段。对于该异常判断，可以是预先设置对应的异常规则，然后基于该规则进行异常判断，例如年龄段对应的异常人数阈值或比例，当某一年龄段对应的用户数量超过该异常人数阈值或比例时，即可认为该年龄段用户数量异常。本实施例中，若存在用户数量异常的年龄段，则有可能是当前贷款业务在该年龄段存在一定风险，也可能是年龄识别网络模型的识别能力降低、导致将过多的用户识别为该年龄段，此时服务器可向对应的管理终端发送对应的异常提示信息，以提示相关管理人员及时进行检视和处理。
通过以上方式，本实施例的服务器可根据分析历史目标用户的催收年龄分布，并根据该催收年龄分布分析是否存在异常情况，有利于及时发现异常情况，降低业务风险性和年龄识别网络模型的稳定性。
此外，本申请实施例还提供一种年龄识别装置。
本实施例中，所述年龄识别装置包括：
样本扩充模块，用于从预设数据库中获取真实语音样本，并基于生成式对抗网络GAN对所述真实语音样本进行样本扩充，得到扩充语音样本；
模型训练模块，用于通过所述扩充语音样本训练得到年龄识别网络模型；
语音转换模块，用于获取目标用户的目标语音，并将所述目标语音转换为对应的输入频谱图；
年龄确定模块，用于通过所述年龄识别网络模型提取所述输入频谱图的深度特征，并根据所述深度特征确定所述目标用户所属的目标年龄段。
其中，上述年龄识别装置的各虚拟功能模块存储于图1所示年龄识别设备的存储器1005中，用于实现计算机程序的所有功能；各模块被处理器1001执行时，可实现年龄识别的功能。
进一步的，所述年龄识别网络模型包括中间特征层和特征优化层，所述年龄确定模块包括：
特征提取单元，用于通过所述年龄识别网络模型的中间特征层对所述输入频谱图进行原始特征提取，得到对应的原始特征；
特征优化单元，用于通过所述年龄识别网络模型的特征优化层、基于注意力机制对所述原始特征进行特征优化，得到对应的优化特征，并将所述优化特征确定为所述输入频谱图的深度特征。
进一步的，所述原始特征包括原始特征图F，所述优化特征包括优化特征图F”，所述特征优化单元，还用于通过所述通过所述年龄识别网络模型的特征优化层获取所述原始特征的原始特征图F；计算所述F对应的一维通道注意力图channel attention map；对所述F和所述channel attention map进行元素相乘element-wise multiplication，得到对应的中间特征图F'；计算所述F'对应的二维通道注意力图spatial attention map；对所述F'和所述spatial attention map进行element-wise multiplication，得到对应的优化特征图F”。
进一步的，所述语音转换模块包括：
时长判断单元，用于获取目标用户的目标语音，并判断所述目标语音的语音时长是否大于预设时长阈值；
语音分割单元，用于若所述语音时长大于所述预设时长阈值，则对所述目标语音进行语音切割，得到两段以上的语音片段，并分别将各语音片段转换为对应的片段频谱图；
所述年龄确定模块，还用于通过所述年龄识别网络模型分别提取各片段频谱图的深度特征，并根据各片段频谱图的深度特征分别确定各片段频谱图对应的片段年龄段；根据各片段频谱图对应的片段年龄段确定所述目标用户所属的目标年龄段。
进一步的，所述年龄识别装置还包括：
语音获取模块，用于在接收到催收指令时，根据所述催收指令拨打催收电话，并在电话接通后获取对应的接通语音；
语音判断模块，用于判断所述接通语音中是否存在两个以上的用户语音；
语音确定模块，用于若所述接通语音中存在两个以上的用户语音，则根据各用户语音的语音持续时长和/或语音音量在各用户语音中确定目标用户的目标语音。
进一步的，所述年龄识别装置还包括：
语音催收模块，用于根据所述目标年龄段获取对应的目标话术模板，并根据所述目标 话术模板对所述目标用户进行语音催收。
进一步的，所述年龄识别装置还包括：
分布获取模块，用于获取预设数量或预设周期的历史目标用户的历史目标年龄段，并根据所述历史目标年龄段获得催收年龄分布；
异常判断模块，用于根据所述催收年龄分布判断是否存在用户数量异常的年龄段，若存在，则向对应管理终端发送对应的异常提示信息。
其中，上述年龄识别装置中各个模块的功能实现与上述年龄识别方法实施例中各步骤相对应，其功能和实现过程在此处不再一一赘述。
此外，本申请实施例还提供一种计算机可读存储介质。
本申请计算机可读存储介质上存储有计算机程序，其中所述计算机程序被处理器执行时，实现如上述的年龄识别方法的步骤。
其中，计算机程序被执行时所实现的方法可参照本申请年龄识别方法的各个实施例，此处不再赘述。
可选的，本申请涉及的存储介质如计算机可读存储介质可以是非易失性的，也可以是易失性的。
需要说明的是，在本文中，术语“包括”、“包含”或者其任何其他变体意在涵盖非排他性的包含，从而使得包括一系列要素的过程、方法、物品或者系统不仅包括那些要素，而且还包括没有明确列出的其他要素，或者是还包括为这种过程、方法、物品或者系统所固有的要素。在没有更多限制的情况下，由语句“包括一个……”限定的要素，并不排除在包括该要素的过程、方法、物品或者系统中还存在另外的相同要素。
上述本申请实施例序号仅仅为了描述，不代表实施例的优劣。
通过以上的实施方式的描述，本领域的技术人员可以清楚地了解到上述实施例方法可借助软件加必需的通用硬件平台的方式来实现，当然也可以通过硬件，但很多情况下前者是更佳的实施方式。基于这样的理解，本申请的技术方案本质上或者说对现有技术做出贡献的部分可以以软件产品的形式体现出来，该计算机软件产品存储在如上所述的一个存储介质(如ROM/RAM、磁碟、光盘)中，包括若干指令用以使得一台终端设备(可以是手机，计算机，服务器，空调器，或者网络设备等)执行本申请各个实施例所述的方法。
以上仅为本申请的优选实施例，并非因此限制本申请的专利范围，凡是利用本申请说明书及附图内容所作的等效结构或等效流程变换，或直接或间接运用在其他相关的技术领域，均同理包括在本申请的专利保护范围内。

Claims (20)

  1. 一种年龄识别方法，其中，所述年龄识别方法包括：
    从预设数据库中获取真实语音样本，并基于生成式对抗网络GAN对所述真实语音样本进行样本扩充，得到扩充语音样本；
    通过所述扩充语音样本训练得到年龄识别网络模型；
    获取目标用户的目标语音，并将所述目标语音转换为对应的输入频谱图；
    通过所述年龄识别网络模型提取所述输入频谱图的深度特征，并根据所述深度特征确定所述目标用户所属的目标年龄段。
  2. 如权利要求1所述的年龄识别方法，其中，所述年龄识别网络模型包括中间特征层和特征优化层，所述通过所述年龄识别网络模型提取所述输入频谱图的深度特征的步骤包括：
    通过所述年龄识别网络模型的中间特征层对所述输入频谱图进行原始特征提取，得到对应的原始特征；
    通过所述年龄识别网络模型的特征优化层、基于注意力机制对所述原始特征进行特征优化，得到对应的优化特征，并将所述优化特征确定为所述输入频谱图的深度特征。
  3. 如权利要求2所述的年龄识别方法，其中，所述原始特征包括原始特征图F，所述优化特征包括优化特征图F”，
    所述通过所述年龄识别网络模型的特征优化层、基于注意力机制对所述原始特征进行特征优化，得到对应的优化特征的步骤包括：
    通过所述年龄识别网络模型的特征优化层获取所述原始特征的原始特征图F；
    计算所述F对应的一维通道注意力图channel attention map；
    对所述F和所述channel attention map进行元素相乘element-wise multiplication，得到对应的中间特征图F'；
    计算所述F'对应的二维通道注意力图spatial attention map；
    对所述F'和所述spatial attention map进行element-wise multiplication，得到对应的优化特征图F”。
  4. 如权利要求1所述的年龄识别方法，其中，所述获取目标用户的目标语音，并将所述目标语音转换为对应的目标频谱图的步骤包括：
    获取目标用户的目标语音，并判断所述目标语音的语音时长是否大于预设时长阈值；
    若所述语音时长大于所述预设时长阈值，则对所述目标语音进行语音切割，得到两段以上的语音片段，并分别将各语音片段转换为对应的片段频谱图；
    所述通过所述年龄识别网络模型提取所述输入频谱图的深度特征，并根据所述深度特征确定所述目标用户所属的目标年龄段的步骤包括：
    通过所述年龄识别网络模型分别提取各片段频谱图的深度特征，并根据各片段频谱图的深度特征分别确定各片段频谱图对应的片段年龄段；
    根据各片段频谱图对应的片段年龄段确定所述目标用户所属的目标年龄段。
  5. 如权利要求1所述的年龄识别方法，其中，所述获取目标用户的目标语音，并将所述目标语音转换为对应的输入频谱图的步骤之前，还包括：
    在接收到催收指令时，根据所述催收指令拨打催收电话，并在电话接通后获取对应的接通语音；
    判断所述接通语音中是否存在两个以上的用户语音；
    若所述接通语音中存在两个以上的用户语音，则根据各用户语音的语音持续时长和/或语音音量在各用户语音中确定目标用户的目标语音。
  6. 如权利要求1所述的年龄识别方法，其中，所述通过所述年龄识别网络模型提取所述输入频谱图的深度特征，并根据所述深度特征确定所述目标用户所属的目标年龄段的步骤之后，还包括：
    根据所述目标年龄段获取对应的目标话术模板，并根据所述目标话术模板对所述目标用户进行语音催收。
  7. 如权利要求1至6中任一项所述的年龄识别方法，其中，所述通过所述年龄识别网络模型提取所述输入频谱图的深度特征，并根据所述深度特征确定所述目标用户所属的目标年龄段的步骤之后，还包括：
    获取预设数量或预设周期的历史目标用户的历史目标年龄段，并根据所述历史目标年龄段获得催收年龄分布；
    根据所述催收年龄分布判断是否存在用户数量异常的年龄段，若存在，则向对应管理终端发送对应的异常提示信息。
  8. 一种年龄识别装置，其中，所述年龄识别装置包括：
    样本扩充模块，用于从预设数据库中获取真实语音样本，并基于生成式对抗网络GAN对所述真实语音样本进行样本扩充，得到扩充语音样本；
    模型训练模块，用于通过所述扩充语音样本训练得到年龄识别网络模型；
    语音转换模块，用于获取目标用户的目标语音，并将所述目标语音转换为对应的输入频谱图；
    年龄确定模块，用于通过所述年龄识别网络模型提取所述输入频谱图的深度特征，并根据所述深度特征确定所述目标用户所属的目标年龄段。
  9. 一种年龄识别设备，其中，所述年龄识别设备包括处理器、存储器、以及存储在所述存储器上并可被所述处理器执行的计算机程序，其中所述计算机程序被所述处理器执行时，实现年龄识别方法，所述年龄识别方法包括以下步骤：
    从预设数据库中获取真实语音样本，并基于生成式对抗网络GAN对所述真实语音样本进行样本扩充，得到扩充语音样本；
    通过所述扩充语音样本训练得到年龄识别网络模型；
    获取目标用户的目标语音，并将所述目标语音转换为对应的输入频谱图；
    通过所述年龄识别网络模型提取所述输入频谱图的深度特征，并根据所述深度特征确定所述目标用户所属的目标年龄段。
  10. 如权利要求9所述的年龄识别设备，其中，所述年龄识别网络模型包括中间特征层和特征优化层，执行所述通过所述年龄识别网络模型提取所述输入频谱图的深度特征的步骤，包括：
    通过所述年龄识别网络模型的中间特征层对所述输入频谱图进行原始特征提取，得到对应的原始特征；
    通过所述年龄识别网络模型的特征优化层、基于注意力机制对所述原始特征进行特征优化，得到对应的优化特征，并将所述优化特征确定为所述输入频谱图的深度特征。
  11. 如权利要求10所述的年龄识别设备，其中，所述原始特征包括原始特征图F，所述优化特征包括优化特征图F”，
    执行所述通过所述年龄识别网络模型的特征优化层、基于注意力机制对所述原始特征进行特征优化，得到对应的优化特征的步骤，包括：
    通过所述年龄识别网络模型的特征优化层获取所述原始特征的原始特征图F；
    计算所述F对应的一维通道注意力图channel attention map；
    对所述F和所述channel attention map进行元素相乘element-wise multiplication，得到 对应的中间特征图F'；
    计算所述F'对应的二维通道注意力图spatial attention map；
    对所述F'和所述spatial attention map进行element-wise multiplication，得到对应的优化特征图F”。
  12. 如权利要求9所述的年龄识别设备，其中，执行所述获取目标用户的目标语音，并将所述目标语音转换为对应的目标频谱图的步骤，包括：
    获取目标用户的目标语音，并判断所述目标语音的语音时长是否大于预设时长阈值；
    若所述语音时长大于所述预设时长阈值，则对所述目标语音进行语音切割，得到两段以上的语音片段，并分别将各语音片段转换为对应的片段频谱图；
    执行所述通过所述年龄识别网络模型提取所述输入频谱图的深度特征，并根据所述深度特征确定所述目标用户所属的目标年龄段的步骤，包括：
    通过所述年龄识别网络模型分别提取各片段频谱图的深度特征，并根据各片段频谱图的深度特征分别确定各片段频谱图对应的片段年龄段；
    根据各片段频谱图对应的片段年龄段确定所述目标用户所属的目标年龄段。
  13. 如权利要求9所述的年龄识别设备，其中，执行所述获取目标用户的目标语音，并将所述目标语音转换为对应的输入频谱图的步骤之前，还包括：
    在接收到催收指令时，根据所述催收指令拨打催收电话，并在电话接通后获取对应的接通语音；
    判断所述接通语音中是否存在两个以上的用户语音；
    若所述接通语音中存在两个以上的用户语音，则根据各用户语音的语音持续时长和/或语音音量在各用户语音中确定目标用户的目标语音。
  14. 如权利要求9至13中任一项所述的年龄识别设备，其中，执行所述通过所述年龄识别网络模型提取所述输入频谱图的深度特征，并根据所述深度特征确定所述目标用户所属的目标年龄段的步骤之后，还包括：
    获取预设数量或预设周期的历史目标用户的历史目标年龄段，并根据所述历史目标年龄段获得催收年龄分布；
    根据所述催收年龄分布判断是否存在用户数量异常的年龄段，若存在，则向对应管理终端发送对应的异常提示信息。
  15. 一种计算机可读存储介质，其中，所述计算机可读存储介质上存储有计算机程序，其中所述计算机程序被处理器执行时，实现年龄识别方法，所述年龄识别方法包括以下步骤：
    从预设数据库中获取真实语音样本，并基于生成式对抗网络GAN对所述真实语音样本进行样本扩充，得到扩充语音样本；
    通过所述扩充语音样本训练得到年龄识别网络模型；
    获取目标用户的目标语音，并将所述目标语音转换为对应的输入频谱图；
    通过所述年龄识别网络模型提取所述输入频谱图的深度特征，并根据所述深度特征确定所述目标用户所属的目标年龄段。
  16. 如权利要求15所述的计算机可读存储介质，其中，所述年龄识别网络模型包括中间特征层和特征优化层，执行所述通过所述年龄识别网络模型提取所述输入频谱图的深度特征的步骤，包括：
    通过所述年龄识别网络模型的中间特征层对所述输入频谱图进行原始特征提取，得到对应的原始特征；
    通过所述年龄识别网络模型的特征优化层、基于注意力机制对所述原始特征进行特征优化，得到对应的优化特征，并将所述优化特征确定为所述输入频谱图的深度特征。
  17. 如权利要求16所述的计算机可读存储介质，其中，所述原始特征包括原始特征图F，所述优化特征包括优化特征图F”，
    执行所述通过所述年龄识别网络模型的特征优化层、基于注意力机制对所述原始特征进行特征优化，得到对应的优化特征的步骤，包括：
    通过所述年龄识别网络模型的特征优化层获取所述原始特征的原始特征图F；
    计算所述F对应的一维通道注意力图channel attention map；
    对所述F和所述channel attention map进行元素相乘element-wise multiplication，得到对应的中间特征图F'；
    计算所述F'对应的二维通道注意力图spatial attention map；
    对所述F'和所述spatial attention map进行element-wise multiplication，得到对应的优化特征图F”。
  18. 如权利要求15所述的计算机可读存储介质，其中，执行所述获取目标用户的目标语音，并将所述目标语音转换为对应的目标频谱图的步骤，包括：
    获取目标用户的目标语音，并判断所述目标语音的语音时长是否大于预设时长阈值；
    若所述语音时长大于所述预设时长阈值，则对所述目标语音进行语音切割，得到两段以上的语音片段，并分别将各语音片段转换为对应的片段频谱图；
    执行所述通过所述年龄识别网络模型提取所述输入频谱图的深度特征，并根据所述深度特征确定所述目标用户所属的目标年龄段的步骤，包括：
    通过所述年龄识别网络模型分别提取各片段频谱图的深度特征，并根据各片段频谱图的深度特征分别确定各片段频谱图对应的片段年龄段；
    根据各片段频谱图对应的片段年龄段确定所述目标用户所属的目标年龄段。
  19. 如权利要求15所述的计算机可读存储介质，其中，执行所述获取目标用户的目标语音，并将所述目标语音转换为对应的输入频谱图的步骤之前，还包括：
    在接收到催收指令时，根据所述催收指令拨打催收电话，并在电话接通后获取对应的接通语音；
    判断所述接通语音中是否存在两个以上的用户语音；
    若所述接通语音中存在两个以上的用户语音，则根据各用户语音的语音持续时长和/或语音音量在各用户语音中确定目标用户的目标语音。
  20. 如权利要求15至19中任一项所述的计算机可读存储介质，其中，执行所述通过所述年龄识别网络模型提取所述输入频谱图的深度特征，并根据所述深度特征确定所述目标用户所属的目标年龄段的步骤之后，还包括：
    获取预设数量或预设周期的历史目标用户的历史目标年龄段，并根据所述历史目标年龄段获得催收年龄分布；
    根据所述催收年龄分布判断是否存在用户数量异常的年龄段，若存在，则向对应管理终端发送对应的异常提示信息。
PCT/CN2021/071262 2020-02-12 2021-01-12 年龄识别方法、装置、设备及计算机可读存储介质 Ceased WO2021159902A1 (zh)

Applications Claiming Priority (2)

Application Number Priority Date Filing Date Title
CN202010094834.9A CN111312286A (zh) 2020-02-12 2020-02-12 年龄识别方法、装置、设备及计算机可读存储介质
CN202010094834.9 2020-02-12

Publications (1)

Publication Number Publication Date
WO2021159902A1 true WO2021159902A1 (zh) 2021-08-19

Family

ID=71150902

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/CN2021/071262 Ceased WO2021159902A1 (zh) 2020-02-12 2021-01-12 年龄识别方法、装置、设备及计算机可读存储介质

Country Status (2)

Country Link
CN (1) CN111312286A (zh)
WO (1) WO2021159902A1 (zh)

Cited By (7)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN113963687A (zh) * 2021-10-20 2022-01-21 百度在线网络技术（北京）有限公司 语音交互方法、装置、设备和存储介质
CN113989100A (zh) * 2021-09-18 2022-01-28 西安电子科技大学 一种基于样式生成对抗网络的红外纹理样本扩充方法
CN114067780A (zh) * 2021-11-04 2022-02-18 国家工业信息安全发展研究中心 一种车载语音识别仿真测试方法、系统及存储介质
CN114694685A (zh) * 2022-04-12 2022-07-01 北京小米移动软件有限公司 语音质量评估方法、装置及存储介质
CN114760523A (zh) * 2022-03-30 2022-07-15 咪咕数字传媒有限公司 音视频处理方法、装置、设备及存储介质
CN116403557A (zh) * 2022-12-14 2023-07-07 西安工业大学 一种基于生成对抗网络的语音隐私保护方法
CN117690431A (zh) * 2023-12-25 2024-03-12 杭州恒芯微电子技术有限公司 一种基于语音识别的麦克风系统

Families Citing this family (6)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN111312286A (zh) * 2020-02-12 2020-06-19 深圳壹账通智能科技有限公司 年龄识别方法、装置、设备及计算机可读存储介质
CN114861746B (zh) * 2021-12-15 2024-10-18 平安科技（深圳）有限公司 基于大数据的反欺诈识别方法、装置及相关设备
CN114358209A (zh) * 2022-01-14 2022-04-15 平安科技（深圳）有限公司 基于标签分布的年龄预测方法、装置、设备及存储介质
CN114596864B (zh) * 2022-02-28 2025-07-15 腾讯音乐娱乐科技（深圳）有限公司 音频处理方法、设备和计算机程序产品
CN114708872B (zh) * 2022-03-22 2024-10-22 青岛海尔科技有限公司 语音指令的响应方法及装置、存储介质及电子装置
CN119107955A (zh) * 2023-06-09 2024-12-10 Oppo广东移动通信有限公司 语音识别方法，终端设备及存储介质

Citations (8)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20080249774A1 (en) * 2007-04-03 2008-10-09 Samsung Electronics Co., Ltd. Method and apparatus for speech speaker recognition
CN103310788A (zh) * 2013-05-23 2013-09-18 北京云知声信息技术有限公司 一种语音信息识别方法及系统
US20170018270A1 (en) * 2015-07-16 2017-01-19 Samsung Electronics Co., Ltd. Speech recognition apparatus and method
CN108922518A (zh) * 2018-07-18 2018-11-30 苏州思必驰信息科技有限公司 语音数据扩增方法和系统
CN109147810A (zh) * 2018-09-30 2019-01-04 百度在线网络技术（北京）有限公司 建立语音增强网络的方法、装置、设备和计算机存储介质
CN109299701A (zh) * 2018-10-15 2019-02-01 南京信息工程大学 基于gan扩充多人种特征协同选择的人脸年龄估计方法
CN109559736A (zh) * 2018-12-05 2019-04-02 中国计量大学 一种基于对抗网络的电影演员自动配音方法
CN111312286A (zh) * 2020-02-12 2020-06-19 深圳壹账通智能科技有限公司 年龄识别方法、装置、设备及计算机可读存储介质

Family Cites Families (8)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US10685262B2 (en) * 2015-03-20 2020-06-16 Intel Corporation Object recognition based on boosting binary convolutional neural network features
KR101809511B1 (ko) * 2016-08-04 2017-12-15 세종대학교산학협력단 발화자의 연령대 인식 장치 및 방법
WO2019225801A1 (ko) * 2018-05-23 2019-11-28 한국과학기술원 사용자의 음성 신호를 기반으로 감정, 나이 및 성별을 동시에 인식하는 방법 및 시스템
WO2019222996A1 (en) * 2018-05-25 2019-11-28 Beijing Didi Infinity Technology And Development Co., Ltd. Systems and methods for voice recognition
CN108924218B (zh) * 2018-06-29 2020-02-18 百度在线网络技术（北京）有限公司 用于推送信息的方法和装置
CN110136726A (zh) * 2019-06-20 2019-08-16 厦门市美亚柏科信息股份有限公司 一种语音性别的估计方法、装置、系统及存储介质
CN110556129B (zh) * 2019-09-09 2022-04-19 北京大学深圳研究生院 双模态情感识别模型训练方法及双模态情感识别方法
CN110619889B (zh) * 2019-09-19 2022-03-15 Oppo广东移动通信有限公司 体征数据识别方法、装置、电子设备和存储介质

Patent Citations (8)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20080249774A1 (en) * 2007-04-03 2008-10-09 Samsung Electronics Co., Ltd. Method and apparatus for speech speaker recognition
CN103310788A (zh) * 2013-05-23 2013-09-18 北京云知声信息技术有限公司 一种语音信息识别方法及系统
US20170018270A1 (en) * 2015-07-16 2017-01-19 Samsung Electronics Co., Ltd. Speech recognition apparatus and method
CN108922518A (zh) * 2018-07-18 2018-11-30 苏州思必驰信息科技有限公司 语音数据扩增方法和系统
CN109147810A (zh) * 2018-09-30 2019-01-04 百度在线网络技术（北京）有限公司 建立语音增强网络的方法、装置、设备和计算机存储介质
CN109299701A (zh) * 2018-10-15 2019-02-01 南京信息工程大学 基于gan扩充多人种特征协同选择的人脸年龄估计方法
CN109559736A (zh) * 2018-12-05 2019-04-02 中国计量大学 一种基于对抗网络的电影演员自动配音方法
CN111312286A (zh) * 2020-02-12 2020-06-19 深圳壹账通智能科技有限公司 年龄识别方法、装置、设备及计算机可读存储介质

Cited By (7)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN113989100A (zh) * 2021-09-18 2022-01-28 西安电子科技大学 一种基于样式生成对抗网络的红外纹理样本扩充方法
CN113963687A (zh) * 2021-10-20 2022-01-21 百度在线网络技术（北京）有限公司 语音交互方法、装置、设备和存储介质
CN114067780A (zh) * 2021-11-04 2022-02-18 国家工业信息安全发展研究中心 一种车载语音识别仿真测试方法、系统及存储介质
CN114760523A (zh) * 2022-03-30 2022-07-15 咪咕数字传媒有限公司 音视频处理方法、装置、设备及存储介质
CN114694685A (zh) * 2022-04-12 2022-07-01 北京小米移动软件有限公司 语音质量评估方法、装置及存储介质
CN116403557A (zh) * 2022-12-14 2023-07-07 西安工业大学 一种基于生成对抗网络的语音隐私保护方法
CN117690431A (zh) * 2023-12-25 2024-03-12 杭州恒芯微电子技术有限公司 一种基于语音识别的麦克风系统

Also Published As

Publication number Publication date
CN111312286A (zh) 2020-06-19

Similar Documents

Publication Publication Date Title
WO2021159902A1 (zh) 年龄识别方法、装置、设备及计算机可读存储介质
CN111694940B (zh) 一种用户报告的生成方法及终端设备
WO2021000408A1 (zh) 面试评分方法、装置、设备及存储介质
CN108922543B (zh) 模型库建立方法、语音识别方法、装置、设备及介质
CN113903361B (zh) 基于人工智能的语音质检方法、装置、设备及存储介质
KR20190022432A (ko) 전자장치, 신분 검증 방법, 시스템 및 컴퓨터 판독 가능한 저장매체
CN114065720B (zh) 会议纪要生成方法、装置、存储介质及电子设备
US9947323B2 (en) Synthetic oversampling to enhance speaker identification or verification
CN112562736B (zh) 一种语音数据集质量评估方法和装置
WO2021072893A1 (zh) 一种声纹聚类方法、装置、处理设备以及计算机存储介质
US20210287682A1 (en) Information processing apparatus, control method, and program
CN113889081A (zh) 语音识别方法、介质、装置和计算设备
CN107316635A (zh) 语音识别方法及装置、存储介质、电子设备
CN117690436A (zh) 基于情绪识别的智能外呼方法、装置、设备及存储介质
WO2025161664A1 (zh) 语音通信质量的评估方法、装置、服务器及存储介质
CN118865970A (zh) 基于人工智能的语音识别方法、装置及机器人设备
CN117726984A (zh) 目标对象识别方法、装置、计算机设备和存储介质
WO2021077333A1 (zh) 同声传译方法及装置、存储介质
CN116486789A (zh) 语音识别模型的生成方法、语音识别方法、装置及设备
CN112712792B (zh) 一种方言识别模型的训练方法、可读存储介质及终端设备
CN119629636A (zh) 垃圾电话的识别方法、装置、计算机设备及存储介质
CN119271793A (zh) 一种笔录矛盾检测方法、装置、存储介质及设备
CN118609536A (zh) 音频生成方法、装置、设备和存储介质
CN116756348A (zh) 数据处理方法、装置、电子设备及存储介质
JP7287442B2 (ja) 情報処理装置、制御方法、及びプログラム

Legal Events

Date Code Title Description
121 Ep: the epo has been informed by wipo that ep was designated in this application

Ref document number: 21753357

Country of ref document: EP

Kind code of ref document: A1

NENP Non-entry into the national phase

Ref country code: DE

32PN Ep: public notification in the ep bulletin as address of the adressee cannot be established

Free format text: NOTING OF LOSS OF RIGHTS PURSUANT TO RULE 112(1) EPC (EPO FORM 1205A DATED 12.12.2022)

122 Ep: pct application non-entry in european phase

Ref document number: 21753357

Country of ref document: EP

Kind code of ref document: A1