Age estimation method and device, electronic equipment and storage medium

Abstract

The invention provides an age estimation method and device, electronic equipment and a computer readable storage medium, and relates to the technical field of age estimation. The age estimation method comprises the following steps: acquiring basic speech features of a target object; based on a Titanet neural model and a residual block structure, performing feature optimization on the basic speech features to obtain a first embedded feature vector of the target object; and estimating the age of the target object based on the age segmentation point set, the recursive binary tree structure and the first embedded feature vector of the target object. The problem that the age recognition effect is poor due to dependence on basic acoustic features in the related technology is at least solved. The method is suitable for age estimation scenes.

Description

Technical Field

[0001] The present invention relates to the technical field of age estimation, and particularly to an age estimation method, device, electronic device and computer-readable storage medium. Background Art

[0002] In the prior art, the method for age recognition based on voice usually optimizes parameters through a traditional k-Nearest Neighbors (KNN) algorithm regression model and cross-entropy loss. The basic acoustic feature expression ability is insufficient, and the traditional k-Nearest Neighbors algorithm regression model has limitations in age estimation, and the estimation accuracy is not high, resulting in poor age recognition effect.

[0003] In addition, the traditional k-Nearest Neighbors algorithm regression model has limitations in age estimation, and the estimation accuracy is not high, resulting in poor age recognition effect. Summary of the Invention

[0004] The technical problem to be solved by the present invention is to provide an age estimation method, device, electronic device and computer-readable storage medium for the above-mentioned deficiencies of the prior art. This method can enhance the extracted voice features and the accuracy of the age estimation model, thereby improving the accuracy and efficiency of age estimation, and realizing reliable and rapid age estimation.

[0005] In a first aspect, the present invention provides an age estimation method, including: obtaining the basic voice features of a target object; optimizing the basic voice features based on the Titanet neural model and the residual block structure to obtain the first embedded feature vector of the target object; estimating the age of the target object based on the age segmentation point set, the recursive binary tree structure and the first embedded feature vector of the target object.

[0006] Preferably, the obtaining the basic voice features of the target object specifically includes: obtaining the original voice of the target object; preprocessing the original voice to obtain the acoustic features of the original voice, where the preprocessing includes one or more of the following: noise reduction, framing, windowing; inputting the acoustic features into a Mel filter bank and outputting the basic voice features of the target object.

[0007] Preferably, the optimizing the basic voice features based on the Titanet neural model and the residual block structure to obtain the first embedded feature vector of the target object specifically includes: extracting the global features and local features of the target object from the basic voice features based on the Titanet neural model to generate a second embedded feature vector of a preset dimension, where the global features include at least one of the following: fundamental frequency, timbre, vowel transition, vocal tract length, noise robustness, and the local features include at least one of the following: local plosive frequency, local consonant intensity; enhancing age-related features and suppressing age-unrelated noise for the second embedded feature vector of the preset dimension based on the residual block structure to obtain the first embedded feature vector of the target object.

[0008] Preferably, the residual block structure includes the activation function of the first fully-connected layer, the activation function of the second fully-connected layer, and the regularization function of the skip connection layer. Based on the residual block structure, age-related feature enhancement and age-unrelated noise suppression are performed on the second embedding feature vector of a preset dimension to obtain the first embedding feature vector. Specifically, according to formula (1), age-related feature enhancement and age-unrelated noise suppression are performed on the second embedding feature vector of a preset dimension:

[0009] f optimized = Dropout(LeakyReLU(W2·ReLU(W1·f Titanet + b1)+ b2)) (1),

[0010] where f optimized represents the first embedding feature vector of the target object, Dropout(·) represents the regularization function of the skip connection layer, LeakyReLU(·) represents the activation function of the second fully-connected layer, W2 represents the weight of the second fully-connected layer, ReLU(·) represents the activation function of the first fully-connected layer, W1 represents the weight of the first fully-connected layer, f Titanet represents the second embedding feature vector of a preset dimension, b1 represents the bias term of the first fully-connected layer, and b2 represents the bias term of the second fully-connected layer.

[0011] Preferably, before estimating the age of the target object based on the age segmentation point set, the recursive binary tree structure, and the first embedding feature vector of the target object, the age estimation method further includes: obtaining the first embedding feature vectors of different objects at all ages; recursively classifying the first embedding feature vectors of different objects at all ages based on the age segmentation point set to obtain M age ranges and the corresponding first embedding feature vectors of different objects, where M is a positive integer; determining the M age ranges as the M nodes of the recursive binary tree structure respectively; fitting the first embedding feature vectors of different objects corresponding to the M age ranges based on the least squares method to obtain the linear regression coefficients and linear regression intercept terms corresponding to the M age ranges; and determining the linear regression coefficients and linear regression intercept terms corresponding to the M third embedding feature vectors as the node attributes of the M nodes respectively.

[0012] Preferably, estimating the age of the target object based on the age segmentation point set, the recursive binary tree structure, and the first embedded feature vector of the target object specifically includes: recursively classifying the first embedded feature vector of the target object based on the age segmentation point set to obtain N age ranges of the first embedded feature of the target object, where N is a positive integer less than M; matching the linear regression coefficient and the linear regression intercept term corresponding to the shortest age range among the N age ranges from the recursive binary tree structure; and estimating the age of the target object based on the first embedded feature vector of the target object, the linear regression coefficient corresponding to the shortest age range, and the linear regression intercept term.

[0013] Preferably, estimating the age of the target object based on the first embedded feature vector of the target object, the linear regression coefficient corresponding to the shortest age range, and the linear regression intercept term specifically includes: estimating the age of the target object according to formula (2):

[0014]

[0015] where, f x represents the age range of the first embedded feature vector x, represents the node corresponding to the age range f x corresponding node, represents the linear regression coefficient of the node in the recursive binary tree structure, represents the linear regression intercept term of the node in the recursive binary tree structure, represents the age of the target object.

[0016] Preferably, after estimating the age of the target object based on the age segmentation point set, the recursive binary tree structure, and the first embedded feature vector of the target object, it further includes: freezing the Titanet neural model and the residual block structure, optimizing the age segmentation point set and the recursive binary tree structure to minimize the prediction error, where the prediction error refers to the mean square error between the true value and the predicted value of age estimation; freezing the age segmentation point set and the recursive binary tree structure, optimizing the Titanet neural model and the residual block structure to minimize the prediction error.

[0017] In a second aspect, the present invention further provides an age estimation device, including a first acquisition module, a feature optimization module, and an age estimation module. The first acquisition module is used to acquire the basic voice features of the target object. The feature optimization module is connected to the first acquisition module and is used to perform feature optimization on the basic voice features based on the Titanet neural model and the residual block structure to obtain the first embedded feature vector of the target object. The age estimation module is connected to the feature optimization module and is used to estimate the age of the target object based on the age segmentation point set, the recursive binary tree structure, and the first embedded feature vector of the target object.

[0018] In a third aspect, the present invention further provides an electronic device, including a memory and a processor. A computer program is stored in the memory, and the processor is configured to run the computer program to implement the age estimation method provided in the first aspect above.

[0019] In a fourth aspect, the present invention further provides a computer-readable storage medium, on which a computer program is stored. When the computer program is executed by a processor, the age estimation method provided in the first aspect above is implemented.

[0020] An age estimation method, device, electronic device and computer-readable storage medium provided by the present invention optimize the extraction process of voice features through the Titanet neural model and the residual block structure, thereby improving the accuracy of voice features and providing more accurate data support for subsequent age estimation; through a dynamically changing age segmentation point set and a recursive binary tree structure, the age estimation process is further optimized, and the accuracy and efficiency of age estimation are improved. Therefore, the present invention can enhance the accuracy of the extracted voice features and the age estimation model, thereby improving the accuracy and efficiency of age estimation and achieving reliable and rapid age estimation. BRIEF DESCRIPTION OF THE DRAWINGS

[0021] Figure 1 is a flowchart of an age estimation method according to Embodiment 1 of the present invention;

[0022] Figure 2 is an example diagram of a recursive binary tree structure in Embodiment 1 of the present invention;

[0023] Figure 3 is an example diagram of fitting the first embedding feature vectors of different objects corresponding to M age ranges in Embodiment 1 of the present invention;

[0024] Figure 4 is a flowchart of an age estimation method according to Embodiment 2 of the present invention;

[0025] Figure 5 is a schematic structural diagram of an age estimation device according to Embodiment 3 of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0026] To enable those skilled in the art to better understand the technical solutions of the present invention, the embodiments of the present invention will be further described in detail below with reference to the accompanying drawings.

[0027] It can be understood that the specific embodiments and drawings described herein are only for explaining the present invention and are not intended to limit the present invention.

[0028] It can be understood that, without conflict, the embodiments of the present invention and the features in the embodiments can be combined with each other.

[0029] It can be understood that, for the convenience of description, only the parts related to the present invention are shown in the drawings of the present invention, while the parts unrelated to the present invention are not shown in the drawings.

[0030] It can be understood that each unit and module involved in the embodiments of the present invention may correspond to only one entity structure, or may be composed of multiple entity structures. Alternatively, multiple units and modules may also be integrated into one entity structure.

[0031] It can be understood that, without conflict, the functions and steps marked in the flowcharts and block diagrams of the present invention may occur in a different order from that marked in the drawings.

[0032] It can be understood that in the flowcharts and block diagrams of the present invention, the possible architectures, functions, and operations of the systems, devices, equipment, and methods according to the embodiments of the present invention are shown. Among them, each block in the flowchart or block diagram may represent a unit, module, program segment, or code, which contains executable instructions for implementing the specified function. Moreover, each block or combination of blocks in the block diagram and flowchart may be implemented by a hardware-based system for implementing the specified function, or may be implemented by a combination of hardware and computer instructions.

[0033] It can be understood that the units and modules involved in the embodiments of the present invention may be implemented in software or in hardware. For example, the units and modules may be located in the processor.

[0034] Embodiment 1:

[0035] As Figure 1 shown, this embodiment provides an age estimation method. The age estimation method includes:

[0036] S101, obtain the basic voice features of the target object.

[0037] In this embodiment, the target object refers to the speaker to whom the original voice to be analyzed or processed belongs, and the basic voice features refer to the features extracted from the voice signal that can reflect the basic attributes of the voice. The basic voice features include, but are not limited to: MFCC (Mel Frequency Cepstral Coefficient), fundamental frequency (F0), formants, energy or loudness, duration information.

[0038] Specifically, S101: Obtain the basic voice features of the target object, including steps S1011 - S1013:

[0039] S1011, obtain the original voice of the target object.

[0040] In this embodiment, the original speech of a preset duration is sampled to output an array of a preset length. The preset duration includes, but is not limited to, 3 seconds, the sampling rate includes, but is not limited to, 16 kHz, and the preset length includes, but is not limited to, 48,000. For example, sampling the original speech sample of 3 seconds can obtain 48,000 sampling points and output an array of length 48,000: [0.023, -0.045, 0.012,..., 0.034].

[0041] S1012. Preprocess the original speech to obtain the acoustic features of the original speech. The preprocessing includes one or more of the following: noise reduction, framing, and windowing.

[0042] In this embodiment, the acoustic features refer to the features that reflect the physical and perceptual attributes of sound. The acoustic features include, but are not limited to: Short-time Energy, Zero-crossing Rate, Mel Frequency Cepstral Coefficients, Filter Bank Energies, Linear Prediction Cepstral Coefficients (LPCC), Spectrogram, and Fundamental Frequency (F0).

[0043] It should be noted that noise reduction is used to remove the background noise and noise of the original speech and improve the quality of the original speech. The common methods of noise reduction include, but are not limited to: Spectral Subtraction, Wiener filtering, and wavelet transform. Framing is used to divide the continuous speech signal into short-time frames (generally 20 - 40 milliseconds). For example, for a signal with a sampling frequency of 16 kHz, if the length of each frame is 25 ms, then each frame has 400 sampling points. Windowing refers to applying a window function (such as Hamming window, Hanning window) to each frame of the original speech to reduce the frame end boundary effect and avoid spectral leakage.

[0044] S1013. Input the acoustic features into the Mel filter bank and output the basic speech features of the target object.

[0045] In this embodiment, taking the spectral energy feature as the basic speech feature as an example, the acoustic features are input into the Mel filter bank, and the basic speech features of the target object are output, which specifically includes: after preprocessing, a fast Fourier transform (FFT) is performed on each frame of acoustic features to obtain the spectrum (amplitude spectrum or power spectrum) of each frame of acoustic features, where the spectra of each frame of acoustic features are summarized into a logarithmic Mel spectrogram; the Mel filter bank is a set of triangular band-pass filters distributed on the Mel frequency scale (simulating the non-linear frequency perception of the human ear), and each filter extracts the energy of the spectrum in the corresponding frequency band and outputs the spectral energy feature of the target object, where the spectral energy feature of the target object is output in the form of a time-frequency two-dimensional matrix (number of frames × number of Mel bands). For example, 3 seconds of original speech corresponds to 300 frames, with a dimension of 300×40, and a 300×40 matrix is output: [[0.12, 0.45, 0.78,..., 0.23], # 40-dimensional Mel energy of the first frame

[0046] [0.09, 0.51, 0.81,..., 0.18], # The second frame

[0047] ……,

[0048] [0.15, 0.38, 0.72,..., 0.29], # The 300th frame]. In this embodiment, by performing noise reduction, framing, windowing on the original speech and inputting it into the Mel filter bank, the extracted basic speech features are closer to the characteristics of human auditory perception, enhancing the expression ability of speech features, effectively suppressing noise and interference, maintaining the key spectral information of the speech, making the speech features more stable and reliable, facilitating the training and inference of machine learning models, and improving the accuracy and robustness of subsequent age estimation.

[0049] S102. Based on the Titanet neural model and the residual block structure, optimize the basic speech features to obtain the first embedded feature vector of the target object.

[0050] Specifically, S102: Based on the Titanet neural model and the residual block structure, optimize the basic speech features to obtain the first embedded feature vector of the target object, including steps S1021 - S1022:

[0051] S1021. Based on the Titanet neural model, extract the global features and local features of the target object from the basic speech features, and generate a second embedded feature vector with a preset dimension, where the global features include at least one of the following: fundamental frequency, timbre, vowel transition, vocal tract length, noise robustness, and the local features include at least one of the following: local plosive frequency, local consonant intensity.

[0052] In this embodiment, taking the preset dimension as 192 dimensions as an example, the Titanet neural model is used to extract the global features and local features of the target object from the basic speech features, and the output is a 192-dimensional second embedded feature vector. Speaker recognition relies more on global features (such as dimensions 0 and 4); voice age prediction requires combining global (vocal tract length) and local (vocal cord micro-vibration) features. Global features include but are not limited to: fundamental frequency, timbre, vowel transition, vocal tract length, noise robustness, and local features include but are not limited to: local plosive frequency, local consonant intensity. Therefore, the second embedded feature vector encodes the speaker's voiceprint features (such as timbre, pronunciation habits) and abstract information irrelevant to the speech content. For example: [0.12, # Dimension 0 → Global fundamental frequency (female)

[0053] -0.05, # Dimension 1 → Mix (timbre + vowel transition)

[0054] 0.87, # Dimension 2 → Local / s / sound intensity

[0055] -0.23, # Dimension 3 → Local plosive frequency

[0056] 0.45, # Dimension 4 → Global vocal tract length (age)

[0057] ……,

[0058] 0.34 # Dimension 191 → Mix (noise robustness)], where global features refer to features that reflect the overall statistical information of the speech segment (such as timbre, average fundamental frequency), typical dimensions: 0, 4, 25, which are strongly correlated with long-term trends; local features refer to features that reflect short-term events or details (such as phonemes, instantaneous energy changes), typical dimensions: 2, 3, 150, which are strongly correlated with local patterns. In this embodiment, through the Titanet neural model, the global features and local features of the target object are extracted, making full use of the rich and multi-dimensional speech representation ability. By combining global and local features, subsequent age estimation can simultaneously focus on the long-term stable attributes and short-term change characteristics of the speech, improving the discrimination ability of the speaker's age.

[0059] S1022. Based on the residual block structure, perform age-related feature enhancement and age-independent noise suppression on the second embedded feature vector of the preset dimension to obtain the first embedded feature vector of the target object.

[0060] Specifically, the residual block structure includes the activation function of the first fully connected layer, the activation function of the second fully connected layer, and the regularization function of the skip connection layer.

[0061] In this embodiment, the residual block structure includes two fully connected layers: fully connected layer 1 (FC1) and fully connected layer 2 (FC2). The fully connected layer 1 and the fully connected layer 2 respectively include the activation function ReLU(·) and the activation function LeakyReLU(·). The activation function ReLU(·) is used to enhance the non-linear expression ability. The negative slope of the activation function LeakyReLU(·) = 0.2, which alleviates the problem of neuron death. The residual block structure also includes a skip connection layer, and the skip connection layer includes the regularization function Dropout(·), which is used to add the second embedded feature vector to the outputs of FC1 and FC2 (i.e., perform residual learning). The ratio of Dropout(·) is 0.2 (i.e., randomly discard 20% of the neurons), which prevents overfitting. In this embodiment, the residual block structure improves the adaptability of the second embedded feature vector to the age prediction task, suppresses irrelevant voiceprint information, and improves the accuracy of age prediction.

[0062] Specifically, S1022: Based on the residual block structure, perform age-related feature enhancement and age-unrelated noise suppression on the second embedded feature vector of a preset dimension to obtain the first embedded feature vector of the target object, including: According to formula (1), perform age-related feature enhancement and age-unrelated noise suppression on the second embedded feature vector of a preset dimension:

[0063] f optimized = Dropout(LeakyReLU(W2·ReLU(W1·f Titanet + b1)+ b2)) (1),

[0064] where, f optimized represents the first embedded feature vector of the target object, Dropout(·) represents the regularization function of the skip connection layer, LeakyReLU(·) represents the activation function of the second fully connected layer, W2 represents the weight of the second fully connected layer, ReLU(·) represents the activation function of the first fully connected layer, W1 represents the weight of the first fully connected layer, f Titanet represents the second embedded feature vector of a preset dimension, b1 represents the bias term of the first fully connected layer, and b2 represents the bias term of the second fully connected layer.

[0065] In this embodiment, W1 and W2 are in the shape of 256×192 arrays, b1 is a 256-dimensional vector, and b2 is a 192-dimensional vector. For example, W1 is [[0.12, -0.03,..., 0.05],...], W2 is [[-0.08, 0.15,..., 0.02],...], b1 is [0.1, -0.2,..., 0.3], b2 is [0.05, 0.1,...,-0.1], and the first embedded feature vector of the output target object is a 192-dimensional feature vector, that is, [0.15, -0.02, 0.91,..., 0.29].

[0066] S103. Estimate the age of the target object based on the age segmentation point set, the recursive binary tree structure, and the first embedded feature vector of the target object.

[0067] In this embodiment, the age segmentation point refers to the age threshold t for classifying the target object * , for example: 30, 27, 33, 26, 29, 34, 25, 28, 36. The recursive binary tree structure is a binary tree, and each node of the binary tree corresponds to a binary classification problem, that is, to judge which side of the two parts divided by the current node the age of the target object belongs to. The training process of the recursive binary tree structure is as follows: starting from the age range of the entire population, select an age segmentation point (such as the midpoint of the age interval or the boundary determined according to the data distribution), and divide the whole into two age groups (left and right child nodes); for each child node, recursively execute the same "binary classification" strategy, and continue to divide the age interval into smaller intervals (until the stop condition is met, such as the age range of the leaf node is narrow enough, or the classification performance reaches the requirement). The prediction process of the recursive binary tree structure is as follows: using the first embedded feature vector of the target object, input it layer by layer into the tree structure for binary classification judgment, starting from the root node, and recursively "walk" to the leaf node according to the binary classification result. The leaf node corresponds to an age interval or a specific age value, which is used as the estimated result. In this embodiment, through the age segmentation point set and the recursive binary tree structure, the continuous age prediction problem is transformed into a series of simple "biased towards older age or younger age" binary classification problems, reducing the regression complexity, decomposing the task using the hierarchical structure, making it more convenient for training and interpretation, and combining the expression ability of the feature vector to gradually narrow the age prediction range and improve the accuracy.

[0068] Optionally, before S103: Estimate the age of the target object based on the age segmentation point set, the recursive binary tree structure, and the first embedded feature vector of the target object, the age estimation method further includes:

[0069] S104. Obtain the first embedded feature vectors of different objects at all ages.

[0070] In this embodiment, similar to obtaining the first embedding feature vector of the target object, the original speech of different objects at all ages is obtained from the TIMIT corpus (The DARPA TIMIT Acoustic-Phonetic Continuous Speech Corpus, an acoustic-phonetic continuous speech corpus jointly constructed by Texas Instruments, the Massachusetts Institute of Technology, and Stanford Research Institute), for example: the original speech of different speakers in the age range of 20 to 58 years old. Then, the original speech of different objects at all ages is preprocessed, input into the Mel filter bank, the global features and local features of different objects at all ages are extracted, age-related feature enhancement and age-independent noise suppression are performed to obtain the first embedding feature vector of different objects at all ages.

[0071] It should be noted that in this embodiment, the first embedding feature vectors of different objects at all ages can be divided into 4610 training samples and 1680 test samples for subsequent construction and optimization of the age segmentation point set and the recursive binary tree structure.

[0072] This embodiment can construct and optimize the age segmentation point set based on the quantiles (10%, 20%,..., 90%) of the actual ages (i.e., the true values of ages) in 4610 training samples and 1680 test samples. For example, if the quantiles of the actual ages (0 - 20, 20 - 35,..., 60 - 100) in 4610 training samples and 1680 test samples are 10%, 20%,..., 90% respectively, then the subsequent age segmentation point set can be determined as [20, 35,..., 60].

[0073] S105, Based on the age segmentation point set, recursively classify the first embedding feature vectors of different objects at all ages to obtain M age ranges and the corresponding first embedding feature vectors of different objects, where M is a positive integer.

[0074] In this embodiment, as Figure 2 shown, based on the age segmentation points (i.e., Figure 2 30, 27, 33, 26, 29, 34, 25, 28, 36) and the pre-trained linear classifier divide 4610 training samples and 1680 test samples into a left subset (y(x) ≤ t * ) and a right subset (y(x) > t * ), and the segmentation process is recursively repeated until the stop condition is reached (such as the maximum depth D max= 10 or the sample size is less than 50) or cannot be further effectively segmented (when constructing a tree structure using recursive binary partitioning, continuing to segment the current sample no longer significantly improves the age prediction performance, or overfitting occurs to the linear regression coefficient and linear regression intercept term after segmentation, resulting in a decline in generalization ability). Therefore, 19 age groups and their corresponding training samples and test samples can be obtained. Among them, the longest age group corresponding to 4610 training samples and 1680 test samples is included in the 19 age groups. For example, if the age group corresponding to 4610 training samples and 1680 test samples is 0 - 100, then the age group of 0 - 100 is included in the 19 age groups.

[0075] The training samples and test samples corresponding to the 19 age groups are regarded as the feature space F extracted from the voice signals of different objects at all ages. The first embedded feature vector of the subsequent target object corresponds to at least one feature f in the feature space F x (that is, the training samples and test samples corresponding to at least one age group among the 19 age groups), f x ∈F.

[0076] S106. Determine the M age groups as the M nodes of the recursive binary tree structure respectively.

[0077] In this embodiment, the 19 features in the feature space F are divided into 19 non - overlapping convex units Determine the 19 non - overlapping convex units as the 19 nodes of the recursive binary tree structure, that is, determine the training samples and test samples corresponding to the 19 age groups as the 19 nodes of the recursive binary tree structure.

[0078] The longest age group corresponding to 4610 training samples and 1680 test samples is used as the root node of the recursive binary tree structure (that is, Figure 2 node 1 in), and the age groups corresponding to the training samples and test samples that cannot be segmented among the 19 age groups are used as the leaf nodes of the recursive binary tree structure (that is, Figure 2 node 14, node 15, node 9, node 16, node in). In this embodiment, through recursive classification based on age segmentation points, the features of objects in different age groups are distinguished, the subtle differences in the influence of age on voice features are captured, and the feature vector nodes corresponding to age groups are organized into a recursive binary tree, which is beneficial to quickly locate and retrieve the feature information of different age layers. The tree structure is convenient for hierarchical processing and subsequent incremental learning, and supports dynamic addition or adjustment of age groups.

[0079] It should be noted that in the feature space F, if a group of convex units C i and C j (that is, Figure 2 node 2 and node 3 in) completely cover all samples in the feature space and do not overlap with each other, that is, Then C i and C j must be polyhedrons bounded by hyperplanes, that is, the segmentation nodes corresponding to the age segmentation point H, and H→{C i , C j} can be defined. Each segmentation node mapped by H is responsible for dividing the current region (i.e., its own training samples and test samples) into two sub-regions. For example: Nodes 2 and 3 divide their own training samples and test samples into Nodes 4, 5, 6, and 7.

[0080] S107. Based on the least squares method, fit the first embedding feature vectors of different objects corresponding to M age groups to obtain the linear regression coefficients and linear regression intercept terms corresponding to M age groups.

[0081] In this embodiment, as Figure 3 shown, within the convex cell C n (i.e., the nth node among the 19 nodes of the recursive binary tree structure), fit a linear model through the least squares method where f represents the training samples and test samples corresponding to the nth age group, represents the linear regression coefficient corresponding to the nth age group, and b n represents the regression intercept term corresponding to the nth age group. For example: is a 192-dimensional regression coefficient vector [0.02, -0.15, 0.07,..., 0.12], and b n is 22.5 years old. The linear regression coefficients and intercepts obtained by fitting in this embodiment accurately describe the change trend of the first embedding feature vectors within each age group, which helps to capture the law of speech feature changes with age and supports age perception and speech analysis under age conditions.

[0082] S108. Determine the linear regression coefficients and linear regression intercept terms corresponding to the M third embedding feature vectors as the node attributes of the M nodes respectively.

[0083] Specifically, S103: Estimate the age of the target object based on the age segmentation point set, the recursive binary tree structure, and the first embedding feature vector of the target object, including steps S1031 - S1033:

[0084] S1031. Based on the age segmentation point set, recursively classify the first embedding feature vector of the target object to obtain N age groups of the first embedding feature of the target object, where N is a positive integer less than M.

[0085] In this embodiment, taking the actual age of the target object as 32 years old as an example, similar to recursively classifying the first embedded feature vectors of different objects at all ages, the first embedded feature vector of the target object is recursively classified, and 3 age ranges of the first embedded feature of the target object can be obtained, namely 1-100, 30-100, and 30-33.

[0086] S1032: From the recursive binary tree structure, match the linear regression coefficient and the linear regression intercept term corresponding to the shortest age range among the N age ranges.

[0087] In this embodiment, from the recursive binary tree structure, match the linear regression coefficient and the linear regression intercept term corresponding to the age range of 30-33 (that is, Figure 2 the of the middle node 6

[0088] S1033: Based on the first embedded feature vector of the target object, the linear regression coefficient and the linear regression intercept term corresponding to the shortest age range, estimate the age of the target object.

[0089] Specifically, S1033: Based on the first embedded feature vector of the target object, the linear regression coefficient and the linear regression intercept term corresponding to the shortest age range, estimate the age of the target object, including: According to formula (2), estimate the age of the target object:

[0090]

[0091] where, f x represents the age range of the first embedded feature vector x, represents the node corresponding to the age range f x corresponding to, represents the node in the recursive binary tree structure of the linear regression coefficient, represents the node in the recursive binary tree structure of the linear regression intercept term, represents the age of the target object.

[0092] In this embodiment, the actual age of the target object is 32 years old, and the estimated age of the target object is 31.7 years old; the actual age of the target object is 44 years old, and the estimated age of the target object is 43.9 years old.

[0093] Optionally, after S103: Based on the age segmentation point set, the recursive binary tree structure and the first embedded feature vector of the target object, estimate the age of the target object, the age estimation method further includes:

[0094] S109. Freeze the Titanet neural model and the residual block structure, and optimize the age segmentation point set and the recursive binary tree structure to minimize the prediction error, where the prediction error refers to the mean squared error between the true value and the predicted value of age estimation.

[0095] S110. Freeze the age segmentation point set and the recursive binary tree structure, and optimize the Titanet neural model and the residual block structure to minimize the prediction error.

[0096] In this embodiment, freeze the parameters of the Titanet and the residual block structure (i.e., W1, W2, b1, and b2), and only optimize the age segmentation point set (i.e., the age threshold t * ) and the recursive binary tree structure (i.e., the nodes of the recursive binary tree structure and their node attributes: and ), and minimize the mean squared error (MSE) between the true value and the predicted value of age estimation. Freeze the age segmentation point set (i.e., the age threshold t * ) and the recursive binary tree structure (i.e., the nodes of the recursive binary tree structure and their node attributes: and ), and update the parameters of the Titanet and the residual block structure (i.e., W1, W2, b1, and b2) in the reverse direction through the mean squared error between the true value and the predicted value of age estimation to optimize the parameters of the Titanet and the residual block structure (i.e., W1, W2, b1, and b2).

[0097] Optimize the age segmentation point set (i.e., the age threshold t * ), and the recursive binary tree structure, specifically including: Let According to the formula and the formula Optimize the recursive binary tree structure, where y represents the true value of age estimation, represents the predicted value of age estimation; According to the formula Optimize the age segmentation point set (i.e., the age threshold t * ), where represents the predicted value of age estimation based on the left subset, represents the predicted value of age estimation based on the right subset. In this embodiment, by first dynamically optimizing the recursive binary tree structure to make the recursive binary tree structure change dynamically, then calculating the error between the true value and the predicted value of age estimation for the left and right subsets in the dynamically optimized recursive binary tree structure, and based on the error between the true value and the predicted value of age estimation for the left and right subsets, dynamically optimize the age segmentation point set to make the age segmentation point set change dynamically, thereby improving the accuracy of the age estimation model and greatly improving the age estimation effect.

[0098] Optimize the parameters of the Titanet and residual block structures, specifically including: according to the formula Optimize the parameters of the Titanet and residual block structures. In this embodiment, by adopting the alternating freezing and optimization strategy, the potential of the age segmentation point set, the recursive binary tree structure, the Titanet neural model, and the residual block structure can be fully explored, achieving the best effect of age prediction and improving the accuracy and adaptability of the model.

[0099] An age estimation method provided in this embodiment optimizes the process of extracting speech features through the Titanet neural model and the residual block structure, thereby improving the accuracy of speech features and providing more accurate data support for subsequent age estimation; through the dynamically changing age segmentation point set and the recursive binary tree structure, the age estimation process is further optimized, improving the accuracy and efficiency of age estimation, enhancing the accuracy of the extracted speech features and the age estimation model, and further improving the accuracy and efficiency of age estimation, realizing reliable and fast age estimation.

[0100] Embodiment 2:

[0101] As Figure 4 shown, this embodiment provides an age estimation method. The age estimation method includes:

[0102] S201, obtain the original speech of the target object.

[0103] S202, preprocess the original speech to obtain the acoustic features of the original speech, where the preprocessing includes one or more of the following: noise reduction, framing, windowing; input the acoustic features into the Mel filter bank and output the basic speech features of the target object.

[0104] S203, based on the Titanet neural model, extract the global features and local features of the target object from the basic speech features to generate a second embedding feature vector with a preset dimension, where the global features include at least one of the following: fundamental frequency, timbre, vowel transition, vocal tract length, noise robustness, and the local features include at least one of the following: local plosive frequency, local consonant intensity.

[0105] In this embodiment, the Titanet neural model is the Figure 4 Titanet large model in Figure 4 and the second embedding feature vector with a preset dimension is the

[0106] 192-dimensional embedding vector in

[0107] S205. Based on the set of age segmentation points, recursively classify the first embedded feature vectors of the target object to obtain N age ranges of the first embedded features of the target object, where N is a positive integer less than M.

[0108] In this embodiment, the nodes corresponding to the N age ranges of the first embedded features of the target object in the recursive binary tree structure are Figure 4 the convex cells in

[0109] S206. From the recursive binary tree structure, match the linear regression coefficient and the linear regression intercept term corresponding to the shortest age range among the N age ranges.

[0110] S207. Based on the first embedded feature vector of the target object, the linear regression coefficient corresponding to the shortest age range, and the linear regression intercept term, estimate the age of the target object.

[0111] S208. Freeze the Titanet neural model and the residual block structure, and optimize the set of age segmentation points and the recursive binary tree structure to minimize the prediction error, where the prediction error refers to the mean square error between the true value and the predicted value of age estimation; freeze the set of age segmentation points and the recursive binary tree structure, and optimize the Titanet neural model and the residual block structure to minimize the prediction error.

[0112] In this embodiment, the Titanet neural model and the residual block structure are Figure 4 the feature extractor in Figure 4 and the set of age segmentation points and the recursive binary tree structure are

[0113] An age estimation method provided in this embodiment optimizes the extraction process of speech features through the Titanet neural model and the residual block structure, thereby improving the accuracy of speech features and providing more accurate data support for subsequent age estimation; through the dynamically changing set of age segmentation points and the recursive binary tree structure, it further optimizes the age estimation process, improves the accuracy and efficiency of age estimation, enhances the accuracy of the extracted speech features and the age estimation model, and further improves the accuracy and efficiency of age estimation, realizing reliable and fast age estimation.

[0114] Embodiment 3:

[0115] As Figure 5As shown in the figure, this embodiment provides an age estimation device, including a first acquisition module 31, a feature optimization module 32, and an age estimation module 33. The first acquisition module 31 is used to acquire the basic voice features of the target object. The feature optimization module 32 is connected to the first acquisition module 31 and is used to optimize the basic voice features based on the Titanet neural model and the residual block structure to obtain the first embedded feature vector of the target object. The age estimation module 33 is connected to the feature optimization module 32 and is used to estimate the age of the target object based on the age segmentation point set, the recursive binary tree structure, and the first embedded feature vector of the target object.

[0116] Specifically, the first acquisition module 31 includes: an acquisition unit 311, a preprocessing unit 312, and a Mel filter unit 313. The acquisition unit 311 is used to acquire the original voice of the target object. The preprocessing unit 312 is used to preprocess the original voice to obtain the acoustic features of the original voice. Among them, the preprocessing includes one or more of the following: noise reduction, framing, windowing. The Mel filter unit 313 is used to input the acoustic features into the Mel filter bank and output the basic voice features of the target object.

[0117] Specifically, the feature optimization module 32 includes: a feature extraction unit 321 and a feature enhancement unit 322. The feature extraction unit 321 is used to extract the global features and local features of the target object from the basic voice features based on the Titanet neural model and generate a second embedded feature vector with a preset dimension. Among them, the global features include at least one of the following: fundamental frequency, timbre, vowel transition, vocal tract length, noise robustness. The local features include at least one of the following: local plosive frequency, local consonant intensity. The feature enhancement unit 322 is used to perform age-related feature enhancement and age-independent noise suppression on the second embedded feature vector with a preset dimension based on the residual block structure to obtain the first embedded feature vector of the target object.

[0118] Specifically, the feature enhancement unit 322 includes: a feature enhancement subunit, which is used to perform age-related feature enhancement and age-independent noise suppression on the second embedded feature vector with a preset dimension according to formula (1):

[0119] f optimized = Dropout(LeakyReLU(W2·ReLU(W1·f Titanet +b1)+b2)) (1),

[0120] where f optimizeddenotes the first embedded feature vector of the target object, Dropout(·) denotes the regularization function of the skip connection layer, LeakyReLU(·) denotes the activation function of the second fully connected layer, W2 denotes the weight of the second fully connected layer, ReLU(·) denotes the activation function of the first fully connected layer, W1 denotes the weight of the first fully connected layer, f Titanet denotes the second embedded feature vector of a preset dimension, b1 denotes the bias term of the first fully connected layer, and b2 denotes the bias term of the second fully connected layer.

[0121] Optionally, the age estimation device further includes: a second acquisition module 34, a recursive classification module 35, a first determination module 36, a fitting module 37, and a second determination module 38. The second acquisition module 34 is configured to acquire the first embedded feature vectors of different objects at all ages. The recursive classification module 35 is configured to recursively classify the first embedded feature vectors of different objects at all ages based on the set of age segmentation points, to obtain M age ranges and the corresponding first embedded feature vectors of different objects, where M is a positive integer. The first determination module 36 is configured to respectively determine the M age ranges as M nodes of a recursive binary tree structure. The fitting module 37 is configured to fit the first embedded feature vectors of different objects corresponding to the M age ranges based on the least squares method, to obtain the linear regression coefficients and linear regression intercept terms corresponding to the M age ranges. The second determination module 38 is configured to respectively determine the linear regression coefficients and linear regression intercept terms corresponding to the M third embedded feature vectors as the node attributes of the M nodes.

[0122] Specifically, the age estimation module 33 includes: a recursive classification unit 331, a matching unit 332, and an estimation unit 333. The recursive classification unit 331 is configured to recursively classify the first embedded feature vector of the target object based on the set of age segmentation points, to obtain N age ranges of the first embedded features of the target object, where N is a positive integer less than M. The matching unit 332 is configured to match the linear regression coefficients and linear regression intercept terms corresponding to the shortest age range among the N age ranges from the recursive binary tree structure. The estimation unit 333 is configured to estimate the age of the target object based on the first embedded feature vector of the target object, the linear regression coefficients, and the linear regression intercept terms corresponding to the shortest age range.

[0123] Specifically, the estimation unit 333 includes: an estimation subunit, configured to estimate the age of the target object according to formula (2):

[0124]

[0125] where f x denotes the age range of the first embedded feature vector x, denotes the age range f x corresponding node, denotes the node in the recursive binary tree structure The linear regression coefficient of represents the node in the recursive binary tree structure The linear regression intercept term of represents the age of the target object.

[0126] Optionally, the age estimation device further includes: a first optimization module 39 and a second optimization module 40. The first optimization module 39 is configured to freeze the Titanet neural model and the residual block structure, and optimize the age segmentation point set and the recursive binary tree structure to minimize the prediction error, where the prediction error refers to the mean square error between the true value and the predicted value of age estimation. The second optimization module 40 is configured to freeze the age segmentation point set and the recursive binary tree structure, and optimize the Titanet neural model and the residual block structure to minimize the prediction error.

[0127] It can be understood that the above-provided age estimation device executes the age estimation method corresponding to Embodiment 1 provided above. Therefore, the beneficial effects it can achieve can refer to the beneficial effects of the solution corresponding to the age estimation method in Embodiment 1 above, which will not be elaborated here.

[0128] Embodiment 4:

[0129] This embodiment provides an electronic device, including a memory and a processor. A computer program is stored in the memory, and the processor is configured to run the computer program to implement the age estimation method in Embodiment 1 or Embodiment 2 above.

[0130] Embodiment 5:

[0131] This embodiment provides a computer-readable storage medium, on which a computer program is stored. When the computer program is executed by a processor, it implements the age estimation method in Embodiment 1 or Embodiment 2 above.

[0132] It can be understood that the above embodiments are merely exemplary embodiments adopted to illustrate the principles of the present invention. However, the present invention is not limited thereto. For those of ordinary skill in the art, various modifications and improvements can be made without departing from the spirit and essence of the present invention, and these modifications and improvements are also regarded as the protection scope of the present invention.

Claims

1. An age estimation method, characterized in that, Including: Obtain the basic voice features of the target object; Based on the Titanet neural model and the residual block structure, optimize the basic voice features to obtain the first embedded feature vector of the target object; Estimate the age of the target object based on the age segmentation point set, the recursive binary tree structure, and the first embedded feature vector of the target object.

2. The age estimation method according to claim 1, characterized in that The obtaining of the basic voice features of the target object specifically includes: Obtain the original voice of the target object; Preprocess the original voice to obtain the acoustic features of the original voice, where the preprocessing includes one or more of the following: noise reduction, frame segmentation, windowing; Input the acoustic features into the Mel filter bank and output the basic voice features of the target object.

3. The age estimation method according to claim 1, wherein The optimizing of the basic voice features based on the Titanet neural model and the residual block structure to obtain the first embedded feature vector of the target object specifically includes: Based on the Titanet neural model, extract the global features and local features of the target object from the basic voice features to generate a second embedded feature vector of a preset dimension, where the global features include at least one of the following: fundamental frequency, timbre, vowel transition, vocal tract length, noise robustness, and the local features include at least one of the following: local plosive frequency, local consonant intensity; Based on the residual block structure, perform age-related feature enhancement and age-independent noise suppression on the second embedded feature vector of the preset dimension to obtain the first embedded feature vector of the target object.

4. The age estimation method according to claim 3, characterized in that The residual block structure includes the activation function of the first fully connected layer, the activation function of the second fully connected layer, and the regularization function of the skip connection layer. The performing of age-related feature enhancement and age-independent noise suppression on the second embedded feature vector of the preset dimension based on the residual block structure to obtain the first embedded feature vector specifically includes: According to formula (1), perform age-related feature enhancement and age-independent noise suppression on the second embedded feature vector of the preset dimension: f optimized = Dropout(LeakyReLU(W2·ReLU(W1·f Titanet + b1)+ b2)) (1), Among them, f optimized represents the first embedded feature vector of the target object, Dropout(·) represents the regularization function of the skip connection layer, LeakyReLU(·) represents the activation function of the second fully connected layer, W2 represents the weight of the second fully connected layer, ReLU(·) represents the activation function of the first fully connected layer, W1 represents the weight of the first fully connected layer, f Titanet represents the second embedded feature vector of the preset dimension, b1 represents the bias term of the first fully connected layer, and b2 represents the bias term of the second fully connected layer.

5. The age estimation method according to claim 1, wherein Before estimating the age of the target object based on the age segmentation point set, the recursive binary tree structure, and the first embedded feature vector of the target object, it further includes: Obtain the first embedded feature vectors of different objects at all ages; Based on the age segmentation point set, recursively classify the first embedded feature vectors of different objects at all ages to obtain M age ranges and the corresponding first embedded feature vectors of different objects, where M is a positive integer; Determine the M age ranges as the M nodes of the recursive binary tree structure respectively; Based on the least squares method, fit the first embedded feature vectors of different objects corresponding to the M age ranges to obtain the linear regression coefficients and linear regression intercept terms corresponding to the M age ranges; Determine the linear regression coefficients and linear regression intercept terms corresponding to the M third embedded feature vectors as the node attributes of the M nodes respectively.

6. The age estimation method according to claim 5, wherein The estimating of the age of the target object based on the age segmentation point set, the recursive binary tree structure, and the first embedded feature vector of the target object specifically includes: Based on the age segmentation point set, recursively classify the first embedded feature vector of the target object to obtain N age ranges of the first embedded feature of the target object, where N is a positive integer less than M; Match the linear regression coefficient and the linear regression intercept term corresponding to the shortest age range among N age ranges from the recursive binary tree structure; Estimate the age of the target object based on the first embedded feature vector of the target object, the linear regression coefficient and the linear regression intercept term corresponding to the shortest age range.

7. The age estimation method according to claim 6, characterized in that, The estimating the age of the target object based on the first embedded feature vector of the target object, the linear regression coefficient and the linear regression intercept term corresponding to the shortest age range specifically includes: Estimate the age of the target object according to formula (2): Among them, f x represents the age range of the first embedded feature vector x, represents the node corresponding to the age range f x ; represents the linear regression coefficient of the node in the recursive binary tree structure, represents the linear regression intercept term of the node in the recursive binary tree structure, represents the age of the target object.

8. The age estimation method according to claim 1, wherein After estimating the age of the target object based on the age segmentation point set, the recursive binary tree structure and the first embedded feature vector of the target object, it further includes: Freeze the Titanet neural model and the residual block structure, and optimize the age segmentation point set and the recursive binary tree structure to minimize the prediction error, where the prediction error refers to the mean square error between the true value and the predicted value of age estimation; Freeze the age segmentation point set and the recursive binary tree structure, and optimize the Titanet neural model and the residual block structure to minimize the prediction error.

9. An age estimation device, characterized in that, It includes a first acquisition module, a feature optimization module and an age estimation module, The first acquisition module is used to acquire the basic voice features of the target object, The feature optimization module is connected to the first acquisition module and is used to perform feature optimization on the basic voice features based on the Titanet neural model and the residual block structure to obtain the first embedded feature vector of the target object, The age estimation module is connected to the feature optimization module and is used to estimate the age of the target object based on the age segmentation point set, the recursive binary tree structure and the first embedded feature vector of the target object.

10. An electronic device, characterized in that, It includes a memory and a processor, and a computer program is stored in the memory. The processor is configured to run the computer program to implement an age estimation method according to any one of claims 1 to 8.

11. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements an age estimation method according to any one of claims 1 to 8.

Images