Children voiceprint feature extraction method and system

Abstract

The invention discloses a child voiceprint feature extraction method and system. The method comprises the following steps: acquiring an original voice signal; identifying a voiced segment and a multi-peak energy structure of the original voice signal so as to judge whether a plurality of sound sources are overlapped or not; when a plurality of sound sources are overlapped, analyzing a current frame frequency spectrum of the original voice signal, and comparing the current frame frequency spectrum with a pre-established child voice feature mask, marking and suppressing non-target interference frequency components to identify a child sound source so as to obtain a time-frequency mask track; performing weighted suppression on a frequency spectrum in the original voice signal by using the time-frequency mask track, eliminating noise by using an adaptive Wiener filter, and reconstructing a target child voice signal; multi-dimensional voiceprint features are extracted from the target child voice signal, normalization processing is carried out, a high-dimensional voiceprint feature vector is constructed, and the multi-dimensional voiceprint features comprise MFCC, fundamental frequency and formant. By implementing the method provided by the invention, the accuracy and stability of child voiceprint recognition are improved.

Description

Technical Field

[0001] This invention relates to voiceprint processing methods, and more specifically to methods and systems for extracting voiceprint features from children. Background Technology

[0002] In the home environment, speech recognition faces complex challenges from multiple sound sources. Conversations between parents in different rooms generate complex reflections and reverberation through walls and door cracks. Combined with continuous background noise from devices like televisions, tablets, and stereos, these sound bands highly overlap with children's speech, significantly increasing recognition difficulty. Furthermore, the significant noise generated by children playing—such as slapping, dragging, and collisions—further reduces the signal-to-noise ratio of the speech signal. Traditional voiceprint systems based on single-channel feature extraction and template matching typically assume the target speaker is dominant, making it difficult to handle simultaneous sound sources, leading to poor performance in overlapping and alternating processing. Common noise reduction methods tend towards global filtering, which can easily weaken key details of children's speech in rapidly changing home noise environments, affecting the final recognition results. When multiple sound sources overlap, the formant structures of other speakers may contaminate the child's voiceprint features, causing an abnormally increased matching distance and reducing recognition accuracy.

[0003] The performance of existing children's voiceprint recognition systems is influenced by the design benchmarks of adult voicebanks. Because feature extraction parameters, frequency band divisions, and model training data are primarily geared towards adults, the higher proportion of high-frequency energy and shorter vocal tract length in children are ignored. Directly applying adult thresholds and decision boundaries to children's scenarios often leads to higher false rejection rates. Furthermore, the lack of specific modeling for children's non-standard vocalization patterns makes the system unstable when processing children's unique vocalization patterns. Some systems treat noise reduction, sound source separation, and voiceprint recognition as independent processes, lacking effective end-to-end error feedback mechanisms and collaborative optimization. Most noise reduction algorithms focus on improving speech clarity rather than optimizing "voiceprint separability," potentially leading to the loss of key identifiable features. Current voiceprint systems generally lack a mechanism to prioritize the recognition of target children's speech. Under interference from other, louder sound sources, they tend to favor non-target sound sources, limiting the application and development of children's voiceprint recognition technology.

[0004] Therefore, it is necessary to design a new method to improve the accuracy and stability of children's voiceprint recognition, in order to solve the problems faced by existing technologies in children's voiceprint recognition, such as interference from adults and other sound sources, complex noise backgrounds, and the incompatibility of traditional algorithms, which lead to a decline in recognition accuracy and stability. Summary of the Invention

[0005] The purpose of this invention is to overcome the shortcomings of the prior art and provide a method and system for extracting children's voiceprint features.

[0006] To achieve the above objectives, the present invention adopts the following technical solution: a method for extracting voiceprint features in children, comprising:

[0007] Acquire the raw speech signal;

[0008] The original speech signal is identified by identifying vocal segments and multi-peak energy structure to determine whether multiple sound sources overlap.

[0009] When multiple sound sources overlap, the current frame spectrum of the original speech signal is analyzed, and non-target interference frequency components are marked and suppressed by comparing it with a pre-established child speech feature mask to identify the child sound source and obtain the time-frequency mask trajectory.

[0010] The frequency spectrum of the original speech signal is weighted and suppressed using the time-frequency mask trajectory, and noise is eliminated using an adaptive Wiener filter to reconstruct the target child's speech signal.

[0011] Multi-dimensional voiceprint features are extracted from the target child's speech signal and normalized to construct a high-dimensional voiceprint feature vector. The multi-dimensional voiceprint features include MFCC, fundamental frequency, and formants.

[0012] The further technical solution is as follows: the identification of vocal segments and multi-peak energy structure in the original speech signal to determine whether multiple sound sources overlap includes:

[0013] The original speech signal is preprocessed, framed, and its short-time energy and zero-crossing rate are calculated to identify sound segments and multi-peak energy structures in order to determine whether multiple sound sources overlap.

[0014] The further technical solution is as follows: Preprocessing the original speech signal, framing, calculating short-time energy and zero-crossing rate, identifying audible segments and multi-peak energy structures to determine whether multiple sound sources overlap, includes:

[0015] The original speech signal is processed using a pre-emphasis filter to enhance the high-frequency components and balance the spectrum to obtain a pre-processed signal.

[0016] The preprocessed signal is divided into several sampling points to form multiple time window signals, so as to obtain the signal of each frame;

[0017] The short-time energy of each frame of signal is calculated to identify sound segments. By detecting the extreme values ​​of the energy envelope and combining them with a set energy threshold, when the short-time energy of several consecutive frames exceeds the energy threshold, it is confirmed that the signal has a multi-peak energy structure.

[0018] Calculate the average zero-crossing rate of each frame of signal to distinguish different types of phonemes and noise types;

[0019] By combining short-time energy and average zero-crossing rate, when high energy is found to be accompanied by an abnormally high zero-crossing rate, it can be determined that there is multi-source aliasing or strong noise interference in the current environment.

[0020] The further technical solution is as follows: when multiple sound sources overlap, the current frame spectrum of the original speech signal is analyzed, and by comparing it with a pre-established child speech feature mask, non-target interference frequency components are marked and suppressed to identify the child's sound source, thereby obtaining the time-frequency mask trajectory, including:

[0021] Based on the data from the registration phase, the key frequency bands of the target children's speech were determined to obtain the children's characteristics;

[0022] Perform a Fast Fourier Transform on the original speech signal to obtain the spectrum, and calculate the cosine similarity between the spectrum and the mask of the child's features;

[0023] When the cosine similarity is lower than a preset threshold, the spectrum is marked as non-target interference, and a binarized frequency mask is generated.

[0024] The binarized frequency mask is smoothed by applying medium-frequency filtering to remove isolated noise points and form a continuous time-frequency mask trajectory.

[0025] The further technical solution is as follows: The method of using the time-frequency mask trajectory to perform weighted suppression on the spectrum of the original speech signal, using an adaptive Wiener filter to eliminate noise, and reconstructing the target child's speech signal includes:

[0026] The time-frequency mask trajectory is used to perform a dot product operation on the spectrum of the original speech signal to suppress non-children's voice source components in the frequency domain, so as to obtain the remaining signal components;

[0027] The remaining signal components are then enhanced a second time to obtain the enhanced spectrum;

[0028] The enhanced spectrum is subjected to an inverse fast Fourier transform, and the time-domain continuous waveform is recovered using the overlap-addition method to obtain the target child's speech signal.

[0029] A further technical solution is as follows: the secondary enhancement of the remaining signal components to obtain the enhanced spectrum includes:

[0030] The remaining signal components are then enhanced a second time using a Wiener filter based on prior signal-to-noise ratio estimation to obtain the enhanced spectrum.

[0031] The further technical solution is as follows: the remaining signal components are enhanced a second time using a Wiener filter based on prior signal-to-noise ratio estimation to obtain the enhanced spectrum, including:

[0032] The remaining signal components are subjected to a Wiener filter based on prior signal-to-noise ratio estimation. The gain function of the Wiener filter is dynamically adjusted according to the local signal-to-noise ratio. While preserving the formant structure of the child's speech, residual stationary background noise is filtered out to obtain the enhanced spectrum.

[0033] The further technical solution is as follows: Extracting multi-dimensional voiceprint features from the target child's speech signal and performing normalization processing to construct a high-dimensional voiceprint feature vector includes:

[0034] The target child's speech signal is processed by Mel filter bank, logarithmic energy conversion and discrete cosine transform to generate static MFCC features, and first-order and second-order differences are calculated to capture the temporal variation characteristics of the speech, forming a multidimensional voiceprint feature sub-vector.

[0035] The fundamental frequency of the target child's speech signal is extracted within a specific frequency range using the autocorrelation function method combined with center clipping technology. The first three formant frequencies are then solved using the linear predictive coding analysis method. These frequencies, along with the normalized statistical values, are added to the multidimensional voiceprint feature sub-vector to obtain the feature vector.

[0036] The feature vector is normalized by cepstral mean variance to eliminate channel effects and individual volume differences, and then defined as the current input and registered template feature vector to obtain a high-dimensional voiceprint feature vector.

[0037] The further technical solution is as follows: The autocorrelation function method combined with center clipping technology is used to extract the fundamental frequency of the target child's speech signal within a specific frequency range. The first three formant frequencies are then solved using linear predictive coding analysis. These formsant frequencies, along with normalized statistical values, are appended to the multidimensional voiceprint feature sub-vector to obtain the feature vector, which includes:

[0038] The fundamental frequency of the target child's speech signal is extracted in the frequency range of 200Hz to 800Hz by using the autocorrelation function method combined with center clipping technology.

[0039] Using the linear predictive coding analysis method, with the prediction order set to 12, the frequencies of the first three formants are solved to obtain the formants.

[0040] The normalized fundamental frequency mean, variance, and center frequency of the formant are determined and added to the multidimensional voiceprint feature sub-vector to obtain the feature vector.

[0041] This invention also provides a system for extracting voiceprint features in children, comprising:

[0042] The acquisition unit is used to acquire the raw speech signal;

[0043] The judgment unit is used to identify the sound segments and multi-peak energy structure of the original speech signal in order to determine whether there are multiple overlapping sound sources.

[0044] The trajectory determination unit is used to analyze the current frame spectrum of the original speech signal when multiple sound sources overlap, and to mark and suppress non-target interference frequency components by comparing with a pre-established child speech feature mask to identify the child sound source and obtain the time-frequency mask trajectory.

[0045] The suppression unit is used to weighted suppress the spectrum in the original speech signal using the time-frequency mask trajectory, eliminate noise using an adaptive Wiener filter, and reconstruct the target child's speech signal;

[0046] The feature extraction unit is used to extract multi-dimensional voiceprint features from the target child's speech signal, and perform normalization processing to construct a high-dimensional voiceprint feature vector. The multi-dimensional voiceprint features include MFCC, fundamental frequency, and formants.

[0047] The advantages of this invention compared to existing technologies are as follows: This invention first identifies and processes the phonological segments and multi-peak energy structure in the original speech signal to determine whether multiple sound sources overlap. When multiple sound sources are detected, a method comparing with a pre-established child speech feature mask is used to mark and suppress non-target interference frequency components, thereby effectively identifying the child's voice source. Subsequently, the obtained time-frequency mask trajectory is used to perform spectral weighted suppression on the original speech signal, and an adaptive Wiener filter is used to eliminate noise, reconstructing a clear target child's speech signal. Finally, multi-dimensional voiceprint features, including MFCC, fundamental frequency, and formants, are extracted from the reconstructed child's speech signal. After normalization processing, a high-dimensional voiceprint feature vector is constructed to improve the accuracy and stability of child voiceprint recognition, thereby overcoming the inadequacy of traditional algorithms in the face of interference from adults and other sound sources, and complex noise backgrounds. This method significantly improves the performance of child voiceprint recognition in complex environments.

[0048] The present invention will be further described below with reference to the accompanying drawings and specific embodiments. Attached Figure Description

[0049] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0050] Figure 1 A flowchart illustrating the method for extracting voiceprint features of children provided in an embodiment of the present invention;

[0051] Figure 2 A schematic block diagram of a children's voiceprint feature extraction system provided in an embodiment of the present invention;

[0052] Figure 3 A schematic block diagram of a computer device provided for an embodiment of the present invention. Detailed Implementation

[0053] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0054] It should be understood that, when used in this specification and the appended claims, the terms "comprising" and "including" indicate the presence of the described features, integrals, steps, operations, elements and / or components, but do not exclude the presence or addition of one or more other features, integrals, steps, operations, elements, components and / or collections thereof.

[0055] It should also be understood that the terminology used in this specification is for the purpose of describing particular embodiments only and is not intended to limit the invention. As used in this specification and the appended claims, the singular forms “a,” “an,” and “the” are intended to include the plural forms unless the context clearly indicates otherwise.

[0056] It should also be further understood that the term "and / or" as used in this specification and the appended claims refers to any combination of one or more of the associated listed items and all possible combinations, and includes such combinations.

[0057] Please see Figure 1 , Figure 1This is a flowchart illustrating the child voiceprint feature extraction method provided in an embodiment of the present invention. This method is applied in a server. The method identifies and processes the phonological segments and multi-peak energy structure in the original speech signal to determine if multiple sound sources overlap. It uses a method of comparing with a pre-established child speech feature mask to mark and suppress non-target interference frequency components to identify the child's voice source, and utilizes an adaptive Wiener filter to eliminate noise and reconstruct the target child's speech signal. Subsequently, multi-dimensional voiceprint features, including MFCC, fundamental frequency, and formants, are extracted from the target child's speech signal, and normalization processing is performed to construct a high-dimensional voiceprint feature vector. This method effectively solves the problems of interference from adults and other sound sources, complex noise backgrounds, and the inadequacy of traditional algorithms in existing technologies for child voiceprint recognition, thereby significantly improving the accuracy and stability of child voiceprint recognition in complex home environments or situations with multiple interference sources.

[0058] Figure 1 This is a schematic flowchart of the child voiceprint feature extraction method provided in an embodiment of the present invention. Figure 1 As shown, the method includes the following steps S110 to S150.

[0059] S110, Acquire the original speech signal.

[0060] In this embodiment, the raw speech signal refers to unprocessed sound data directly captured from the environment using a microphone or other audio acquisition device. This signal may contain information from multiple sound sources, such as a child's voice, a parent's voice, television, or background music. To ensure coverage of the frequency bands of the child's voice and its environmental noise, the sampling frequency is typically set to 16000Hz. This step is fundamental to all subsequent processing, and its quality directly affects the final recognition and separation results.

[0061] S120. Identify the sound segments and multi-peak energy structure of the original speech signal to determine whether there are multiple overlapping sound sources.

[0062] In this embodiment, the original speech signal is preprocessed, framed, and its short-time energy and zero-crossing rate are calculated to identify sound segments and multi-peak energy structures in order to determine whether multiple sound sources overlap.

[0063] In one embodiment, step S120 described above may include steps S121 to S125.

[0064] S121. The original speech signal is processed using a pre-emphasis filter to enhance the high-frequency components and balance the spectrum to obtain a pre-processed signal.

[0065] In this embodiment, a pre-emphasis filter is first used to process the original speech signal, and its transfer function is H(z) = 1 - 0.97z. -1 The aim is to enhance the high-frequency components of the signal, thereby balancing the entire spectrum. This is because high-frequency components are relatively important in children's voices, while natural speech signals tend to decay faster in the high-frequency range.

[0066] S122. Divide the preprocessed signal into several sampling points to form multiple time window signals, so as to obtain the signal of each frame.

[0067] Next, the signal is divided into multiple time windows (frames) using an overlapping segmentation method. Each frame has a length of 512 sampling points (corresponding to 32ms) and a frame shift of 256 sampling points (corresponding to 16ms). A Hamming window is chosen as the window function, with the expression w(n) = 0.54 - 0.46cos(2πn / (N-1)), where 0 ≤ n ≤ N-1, to ensure short-time stationarity.

[0068] S123. Calculate the short-time energy of each frame of signal to identify the sound segment, and by detecting the extreme value of the energy envelope and combining it with the set energy threshold, when the short-time energy of several consecutive frames exceeds the energy threshold, it is confirmed that the signal has a multi-peak energy structure.

[0069] In this embodiment, short-time energy is defined as the sum of the energy of each frame of signal, and the calculation formula is as follows: The system sets an energy threshold Eth. This threshold is set when the energy of K consecutive frames (e.g., K=5) is low for a short period. When the value is greater than Eth, it is considered that the sound segment has been entered. Furthermore, within the sound segment, extreme value detection is performed on the energy envelope. If multiple energy peaks are detected within a unit time window (e.g., 500ms) and the depth of the trough between these peaks is less than 30% of the energy of the main peak, then a multi-peak energy structure is considered to exist, suggesting that multiple sound sources may overlap.

[0070] S124. Calculate the average zero-crossing rate of each frame of signal to distinguish different types of phonemes and noise types.

[0071] In this embodiment, the average zero-crossing rate refers to the number of times the signal crosses the zero level, and the calculation formula is as follows: , where sgn[.] represents the sign function. This index helps distinguish between unvoiced, voiced, and noise types.

[0072] S125. Combining short-time energy and average zero-crossing rate, when high energy is found to be accompanied by an abnormally high zero-crossing rate, it is determined that there is multi-source aliasing or strong noise interference in the current environment.

[0073] Finally, by combining the two indicators of short-time energy and average zero-crossing rate, if high energy is detected along with an abnormally high zero-crossing rate (exceeding the statistical range of single-person speech), it can be preliminarily determined that there is multi-source aliasing or strong background noise interference in the current environment. This analysis provides a basis for further sound source separation.

[0074] S130. When multiple sound sources overlap, analyze the current frame spectrum of the original speech signal, and by comparing it with a pre-established child speech feature mask, mark and suppress non-target interference frequency components to identify the child sound source and obtain the time-frequency mask trajectory.

[0075] In this embodiment, the time-frequency mask trajectory refers to a binary matrix on the time-frequency plane used to distinguish the target child's voice source from other interfering voice sources. Each element in this matrix corresponds to a signal component within a specific time point and frequency range, where "1" indicates preservation (i.e., considered to belong to the target child's voice source) and "0" indicates suppression (i.e., considered not to belong to the target child's voice source). Through this step, a pure child's voice signal can be extracted from a complex multi-source environment, providing higher quality input for subsequent voiceprint recognition.

[0076] In one embodiment, step S130 described above may include steps S131 to S134.

[0077] S131. Based on the data from the registration phase, determine the key frequency bands of the target child's speech to obtain the child's characteristics.

[0078] In this embodiment, child characteristics refer to the key frequency bands and spectral patterns obtained from the target child's speech data and used to identify their unique vocal features.

[0079] First, during the registration phase, speech data of the target children is collected, and their long-term average power spectral density (PSD) in the 0-8000Hz range is calculated to identify key frequency bands with concentrated energy. Typically, the fundamental frequency F0 of children's speech fluctuates between 250Hz and 600Hz, while the first and second formants are located relatively high. This information is used to construct a feature mask, serving as the basis for subsequent steps to determine whether a voice source belongs to the target child.

[0080] S132. Perform a Fast Fourier Transform on the original speech signal to obtain the spectrum, and calculate the cosine similarity between the spectrum and the mask of the child's features.

[0081] In this embodiment, the original speech signal is then subjected to a Fast Fourier Transform (FFT) to convert it to the frequency domain, obtaining the spectrum Xi(k) of the current frame. This spectrum is then compared with the previously established target child speech feature mask by calculating the cosine similarity between the two. Cosine similarity is an index that measures the consistency of the directions of two vectors, with a value ranging from -1 to 1; a larger value indicates a higher similarity.

[0082] S133. When the cosine similarity is lower than a preset threshold, the spectrum is marked as non-target interference, and a binarized frequency mask is generated.

[0083] In this embodiment, the binarized frequency mask refers to a two-dimensional matrix that marks whether a sound source belongs to the target child by comparing the current frame spectrum with the child's features, where 1 represents retention (belonging to the target child sound source) and 0 represents suppression (not belonging to the target child sound source).

[0084] If the calculated cosine similarity is lower than a preset threshold gamma (e.g., 0.6), the spectrum of the current frame is considered to not match the target child's speech, and is therefore marked as "non-target interference." The generated binarized frequency mask M(i,k) is a two-dimensional matrix, where 1 represents preservation (i.e., considered to belong to the target child's voice source) and 0 represents suppression (i.e., considered not to belong to the target child's voice source). This process effectively achieves preliminary filtering of non-child voice source components in the mixed speech signal.

[0085] S134. Apply medium-frequency filtering to smooth the binarized frequency mask, remove isolated noise points, and form a continuous time-frequency mask trajectory.

[0086] Finally, to further improve the mask quality, median filtering was used to smooth the generated binarized frequency mask. Median filtering is a commonly used image processing technique that can effectively remove isolated noise points while maintaining sharp edges. After this step, the time-frequency mask trajectory is more continuous and smoother, which helps to extract the target child's voice signal more accurately from complex backgrounds.

[0087] Through the above steps, the system can accurately separate the target child's voice signal in a multi-sound source environment, providing a reliable foundation for subsequent voiceprint recognition.

[0088] S140. The spectrum in the original speech signal is weighted and suppressed using the time-frequency mask trajectory, and noise is eliminated using an adaptive Wiener filter to reconstruct the target child's speech signal.

[0089] In this embodiment, the target child's voice signal refers to the sound signal that is separated and enhanced from the mixed voice signal after processing, and that is specifically belonging to the target child. This process ensures that even if multiple sound sources (such as parents, television, etc.) exist in the same space, the voice features of the target child can be clearly extracted for subsequent voiceprint recognition or other interactive applications.

[0090] In one embodiment, step S140 described above may include steps S141 to S143.

[0091] S141. The frequency spectrum in the original speech signal is multiplied using the time-frequency mask trajectory to suppress non-children's voice source components in the frequency domain, so as to obtain the remaining signal components.

[0092] First, the spectrum Y(i,k) of the original speech signal is multiplied by the previously constructed binarized time-frequency mask M(i,k), i.e., S_hat(i,k)=Y(i,k)M(i,k). This step aims to directly suppress frequency components that do not belong to the target child's voice source in the frequency domain, thereby preserving those components that match the voice characteristics of the target child.

[0093] S142. Perform secondary enhancement on the remaining signal components to obtain the enhanced spectrum.

[0094] In this embodiment, the remaining signal components are enhanced a second time using a Wiener filter based on prior signal-to-noise ratio estimation to obtain the enhanced spectrum.

[0095] Specifically, a Wiener filter based on prior signal-to-noise ratio estimation is applied to the remaining signal components. The gain function of the Wiener filter is dynamically adjusted according to the local signal-to-noise ratio. While preserving the formant structure of the child's speech, residual stationary background noise is filtered out to obtain the enhanced spectrum.

[0096] Next, for the remaining signal components S_hat(i,k) obtained in the above steps, a secondary enhancement is performed using an adaptive Wiener filter based on prior signal-to-noise ratio estimation. The Wiener filter dynamically adjusts its gain function G(k) according to the local signal-to-noise ratio to effectively filter out residual stationary background noise, while preserving the key formant structure of the child's speech as much as possible, thus obtaining a clearer spectrum of the target child's speech.

[0097] S143. Perform an inverse fast Fourier transform on the enhanced spectrum and use the overlap-addition method to recover the time-domain continuous waveform to obtain the target child's speech signal.

[0098] Finally, an inverse fast Fourier transform (IFFT) is performed on the enhanced spectrum S_hat(i,k) to convert it back to the time domain signal. To ensure the continuous and uninterrupted time-domain waveform, an overlap-addition method is used to smoothly connect the signals of each frame, ultimately forming a high-quality target child's speech signal. This process not only improves the intelligibility and naturalness of the target child's speech but also provides more accurate and reliable input data for subsequent voiceprint recognition.

[0099] In this embodiment, the technical fields involved in the method include multi-source speech separation technology and child voiceprint recognition technology. Firstly, regarding multi-source speech separation technology, this technology aims to model and distinguish multiple speech signals from parents, children, television, music, etc., within the same space. It employs source separation and enhancement algorithms based on the power spectrum distribution, spatial location, and temporal structure characteristics of different sound sources, and achieves real-time processing and online adaptation of aliased speech on a low-computing-power embedded platform. Secondly, the child voiceprint recognition technology focuses on building a robust acoustic feature extraction model within the range of children's voice frequencies, considering the high-frequency energy concentration and rapid frequency fluctuations of children's voices. This technology also includes matching strategy design based on multi-dimensional voiceprint vectors (such as distance measurement, threshold setting, and adaptive updates), and dynamic correction of the stability of children's voiceprints to ensure that the system maintains reliable recognition even under conditions of changes in children's pitch, emotional fluctuations, and environmental changes.

[0100] In terms of application scenarios, the method in this embodiment can be used in children's smart toys, supporting personalized addressing and dialogue in multi-sound source scenarios such as family living rooms and bedrooms. It prioritizes the identification and locking of the target child's voice source to avoid confusion in dialogue, and executes differentiated interaction strategies based on voiceprint results. In addition, it is also suitable for multi-terminal collaborative scenarios in the home, realizing "toy-first response" for children through voiceprint recognition, supporting linkage with parents' mobile apps to record voice interaction logs or learning progress, and performing role differentiation and permission control when multiple devices pick up voices simultaneously.

[0101] The core issues to be addressed include separating multiple sound sources to enhance the quality of children's voice sources, ensuring the stability of children's voiceprints, and improving voiceprint matching accuracy. Specific measures include determining the presence of multiple speakers under single-microphone or limited-microphone conditions; distinguishing the target child's voice source from other interfering sound sources when multiple sound sources are present and enhancing it accordingly; and improving the signal-to-noise ratio of children's speech in the target frequency band. Simultaneously, the system automatically models the natural fluctuation range of children's voiceprints, using an adaptive template update mechanism based on historical records to ensure correct recognition even as children's voices change slowly with age, and employing specific strategies in extreme tone cases. Furthermore, by constructing a multi-dimensional voiceprint vector including MFCC, F0, formant positions, and other robust time-frequency features, a distance metric and decision threshold suitable for children are designed, and the template and threshold are dynamically adjusted using an adaptive correction mechanism.

[0102] The multi-source analysis module determines the number of sound sources based on energy and temporal structure analysis. First, the original speech signal x(n) is acquired via a microphone, with a sampling frequency of fs=16000Hz to cover children's speech and most environmental noise frequencies. Pre-emphasis is applied with an alpha coefficient of 0.97. The signal is then framed using an overlapping segmentation method, with a Hamming window as the window function. The frame length N is set to 512 sampling points (corresponding to 32ms), and the frame shift is set to 256 sampling points (corresponding to 16ms). Next, the short-time energy Ei of each frame is calculated. When the energy of K consecutive frames exceeds the threshold Eth, it is considered a sound segment. Extreme value detection is performed on the energy envelope to identify potential multi-peak energy structures indicating multi-source overlap. Furthermore, the short-time average zero-crossing rate ZCR (Zero Crossing Rate) is calculated to distinguish between unvoiced, voiced, and noise levels. An abnormally high zero-crossing rate is preliminarily identified as multi-source aliasing or a noisy environment. Based on spectral differences, the target child's voice source is identified. A feature mask is established to statistically analyze the key frequency bands of the target child's speech. The similarity between the current frame's spectrum and the child's feature mask is calculated. Non-target interference is labeled to generate a binarized time-frequency mask M(i,k), which is then smoothed using median filtering to form a continuous time-frequency mask trajectory. Finally, the target signal is enhanced using filtering and masking methods. The generated time-frequency mask is used to perform a dot product operation on the mixed speech spectrum to suppress non-child voice source components. An adaptive Wiener filter is used for secondary enhancement, and an IFFT is performed on the enhanced spectrum to recover the target child's speech signal.

[0103] The method in this embodiment significantly improves the recognizability of children's voiceprints in multi-source aliasing scenarios. Through multi-source analysis and target child voice source enhancement, it maintains a high recognition success rate even in the presence of television, parental conversations, and environmental noise. It is specifically optimized to address the rich high-frequency components of children's voices, reducing the energy encroachment of interfering sound sources on key frequency bands. In the sound source separation stage, the mixed speech is input and undergoes frame-level preprocessing, including pre-emphasis, framing, and windowing. Short-time energy, ZCR, and simplified spectrum are calculated for each frame. A multi-source detection algorithm is used to determine whether multiple speakers or strong background speech are present. If multiple sound sources are detected, the multi-source analysis module is invoked to enhance the target sound source based on the child's feature mask, outputting the child's enhanced speech result. If a single sound source is detected, the speech is directly sent to the next stage.

[0104] S150. Extract multi-dimensional voiceprint features from the target child's speech signal and perform normalization processing to construct a high-dimensional voiceprint feature vector, wherein the multi-dimensional voiceprint features include MFCC, fundamental frequency, and formants.

[0105] Step S150 involves extracting multi-dimensional voiceprint features from the target child's speech signal and normalizing them to construct a high-dimensional voiceprint feature vector. Here, "multi-dimensional voiceprint features" refers to various acoustic features extracted from the target child's speech signal that can characterize individual differences, including but not limited to MFCC (Mel-frequency cepstral coefficients), fundamental frequency (F0), and formants. These features collectively constitute a high-dimensional voiceprint feature vector that can be used to identify and distinguish different speakers.

[0106] In this embodiment, multi-dimensional voiceprint features refer to a high-dimensional feature vector that can comprehensively describe the unique voice characteristics of a speaker by integrating multiple acoustic features (such as MFCC, fundamental frequency, formants, etc.).

[0107] A high-dimensional voiceprint feature vector is a comprehensive feature representation formed by combining various voiceprint features (such as MFCC, fundamental frequency, formants, etc.) and performing necessary preprocessing (such as normalization). This vector can comprehensively reflect the voice characteristics of a speaker, thus providing a basis for subsequent matching and recognition.

[0108] In one embodiment, step S150 described above may include steps S151 to S153.

[0109] S151. Static MFCC features are generated from the target child's speech signal through Mel filter bank, logarithmic energy conversion and discrete cosine transform, and first-order and second-order differences are calculated to capture the temporal variation characteristics of the speech, forming a multidimensional voiceprint feature sub-vector.

[0110] In this embodiment, the multidimensional voiceprint feature sub-vector refers to a specific set of features containing static and dynamic information extracted and combined from these acoustic features, which is used to enhance the speech recognition system's ability to distinguish individual voices.

[0111] First, an FFT transform is performed on the enhanced time-domain signal to obtain the power spectrum. Then, the power spectrum is passed through a set of triangular filters distributed according to the Mel scale. For the characteristics of children's voices, 26 filters were selected to cover the frequency range from 0Hz to 8000Hz.

[0112] Next, the logarithm of the energy output of each filter is taken, and a Discrete Cosine Transform (DCT) is applied. This step removes the correlation between the feature dimensions, producing 12-dimensional MFCC coefficients. Combined with an additional logarithmic energy term, a 13-dimensional static eigenvector is formed.

[0113] To capture the temporal dynamics of speech, first-order difference (Delta) and second-order difference (Delta-Delta) were calculated. In this way, the original 13-dimensional static feature vector was expanded into a 39-dimensional feature sub-vector that includes temporal variation information.

[0114] S152. The autocorrelation function method combined with center clipping technology is used to extract the fundamental frequency of the target child's speech signal within a specific frequency range, and the first three formant frequencies are solved using the linear predictive coding analysis method. These frequencies, along with the normalized statistical values, are added to the multidimensional voiceprint feature subvector to obtain multidimensional voiceprint features.

[0115] In one embodiment, step S152 described above may include steps S1521 to S1523.

[0116] S1521. Using the autocorrelation function method combined with center clipping technology, the fundamental frequency of the target child's speech signal is extracted within a set frequency range;

[0117] S1522. Using the linear predictive coding analysis method, set the prediction order and solve for the frequencies of the first three formants to obtain the formants.

[0118] S1523. Determine the normalized fundamental frequency mean, variance, and center frequency of the formant, and append them to the multidimensional voiceprint feature sub-vector to obtain multidimensional voiceprint features.

[0119] Specifically, the autocorrelation function (ACF) method combined with center clipping technique is used to accurately extract the fundamental frequency of children's speech signals within a specific frequency range (e.g., 200Hz to 800Hz). This method is particularly suitable for analyzing children's voices with high fundamental frequencies and rapid changes. Linear predictive coding (LPC) analysis is used, with an appropriate prediction order (e.g., 12th order) to solve for the frequencies of the first three formants. LPC is an effective model capable of accurately estimating the positions of formants, which is crucial for identifying the acoustic characteristics of different individuals. The mean and variance of the fundamental frequency and the center frequencies of the first three formants are determined and standardized, then added to the previously obtained 39-dimensional feature sub-vectors to form the final multi-dimensional voiceprint features.

[0120] S153. After the multi-dimensional voiceprint features are normalized by cepstral mean variance to eliminate channel effects and individual volume differences, they are defined as the current input and registered template feature vectors to obtain high-dimensional voiceprint feature vectors.

[0121] Cepstral mean-variance normalization (CMVN) is applied to all extracted feature vectors to eliminate the effects of channel effects and individual volume differences. This process ensures that the system maintains high recognition performance even under different environmental or device conditions.

[0122] Through the detailed steps described above, this embodiment effectively extracts pure target child speech signals from complex environments and constructs a highly representative and discriminative high-dimensional voiceprint feature vector, providing a solid foundation for subsequent voiceprint recognition. This meticulous approach not only improves recognition accuracy but also enhances the system's robustness and adaptability.

[0123] In this embodiment, the target child's speech signal is subjected to FFT transformation again to obtain the power spectrum. The power spectrum is then passed through a set of Mel-scale triangular filters. The number of filters M is set to 26, covering a frequency range of 0-8000Hz. The conversion formula between Mel frequency and linear frequency f is: The logarithm of the energy output of each filter is taken, and a Discrete Cosine Transform (DCT) is performed to remove correlations across dimensions, resulting in 12-dimensional MFCC coefficients C(m). These are combined with the logarithmic energy term of a frame to form a 13-dimensional static feature vector. To capture the temporal dynamics of speech, the first-order difference (Delta) and second-order difference (Delta-Delta) are calculated. The formula for calculating the first-order difference is: The final combination yields a basic voiceprint feature vector of 39 dimensions: 13 (static) + 13 (first order) + 13 (second order).

[0124] The fundamental frequency was extracted using the autocorrelation function (ACF) method combined with center clipping. For children's speech, the fundamental frequency search range was set to 200Hz~800Hz. Linear predictive coding (LPC) analysis was used, with a prediction order P=12 (for a 16kHz sampling rate), to solve the LPC equation roots to estimate the frequencies of the first three formants (F1, F2, F3). The normalized fundamental frequency mean, variance, and center frequencies of the first three formants were appended to the MFCC vector to construct a high-dimensional voiceprint feature vector.

[0125] The extracted feature vectors are subjected to cepstral mean and variance normalization (CMVN). For feature dimension j, the normalized features... for: ;in This is the mean of that dimension within the current sliding window. The standard deviation is given. This step eliminates the effects of channel effects and varying speech volumes. The processed feature vectors are denoted as V_cur (current input) and V_ref (registered template).

[0126] In summary, MFCC, fundamental frequency, and formant features are extracted from the child's voice output in the previous stage and normalized to generate the current voiceprint vector V_cur. The template vector V_ref corresponding to the child bound to the current device is read from storage, the vector distance d between the two is calculated, and compared with the current threshold to make a preliminary identity determination. If d is less than the threshold and the results are consistent across multiple consecutive frames, the voice is considered to originate from the bound child, and a "pass" result is fed back to the upper-layer application logic; otherwise, a "non-target child or uncertain" status is output.

[0127] The aforementioned method for extracting children's voiceprint features first identifies and processes the phonological segments and multi-peak energy structure in the original speech signal to determine if multiple sound sources overlap. When multiple sound sources are detected, non-target interference frequency components are marked and suppressed by comparing them with a pre-established children's speech feature mask, thus effectively identifying the children's voice source. Subsequently, the obtained time-frequency mask trajectory is used to perform spectral weighted suppression on the original speech signal, and an adaptive Wiener filter is used to eliminate noise, reconstructing a clear target children's speech signal. Finally, multi-dimensional voiceprint features, including MFCC, fundamental frequency, and formants, are extracted from the reconstructed children's speech signal. After normalization processing, a high-dimensional voiceprint feature vector is constructed to improve the accuracy and stability of children's voiceprint recognition, thereby overcoming the inadequacy of traditional algorithms in the face of interference from adults and other sound sources, and complex noise backgrounds. This method significantly improves the performance of children's voiceprint recognition in complex environments.

[0128] Figure 2 This is a schematic block diagram of a child voiceprint feature extraction system 300 provided in an embodiment of the present invention. Figure 2 As shown, corresponding to the above-described method for extracting children's voiceprint features, the present invention also provides a children's voiceprint feature extraction system 300. This children's voiceprint feature extraction system 300 includes a unit for performing the above-described children's voiceprint feature extraction method, and the system can be configured in a server. Specifically, please refer to... Figure 2 The child voiceprint feature extraction system 300 includes an acquisition unit 301, a judgment unit 302, a trajectory determination unit 303, an inhibition unit 304, and a feature extraction unit 305.

[0129] The acquisition unit 301 is used to acquire the original speech signal; the judgment unit 302 is used to identify the vocal segments and multi-peak energy structure of the original speech signal to determine whether multiple sound sources overlap; the trajectory determination unit 303 is used to analyze the current frame spectrum of the original speech signal when multiple sound sources overlap, and to mark and suppress non-target interference frequency components by comparing with a pre-established child speech feature mask to identify the child sound source and obtain the time-frequency mask trajectory; the suppression unit 304 is used to use the time-frequency mask trajectory to perform weighted suppression on the spectrum in the original speech signal, use an adaptive Wiener filter to eliminate noise, and reconstruct the target child speech signal; the feature extraction unit 305 is used to extract multi-dimensional voiceprint features from the target child speech signal, perform normalization processing, and construct a high-dimensional voiceprint feature vector, wherein the multi-dimensional voiceprint features include MFCC, fundamental frequency, and formants.

[0130] In one embodiment, the judgment unit 302 is used to preprocess, frame, calculate short-time energy and zero-crossing rate of the original speech signal, identify sound segments and multi-peak energy structure, so as to determine whether there are multiple overlapping sound sources.

[0131] In one embodiment, the determination unit 302 includes:

[0132] The preprocessing subunit processes the original speech signal using a pre-emphasis filter to enhance the high-frequency components and balance the spectrum to obtain a preprocessed signal. The segmentation subunit divides the preprocessed signal into several sampling points, forming multiple time window signals to obtain each frame signal. The short-time energy calculation subunit calculates the short-time energy of each frame signal to identify sound segments and, by detecting the extreme values ​​of the energy envelope and combining them with a set energy threshold, confirms the presence of a multi-peak energy structure when the short-time energy of several consecutive frames exceeds the energy threshold. The average zero-crossing rate calculation subunit calculates the average zero-crossing rate of each frame signal to distinguish different types of phonemes and noise types. The determination subunit combines short-time energy and average zero-crossing rate; when high energy is found accompanied by an abnormally high zero-crossing rate, it determines that there is multi-source aliasing or strong noise interference in the current environment.

[0133] In one embodiment, the trajectory determination unit 303 includes:

[0134] The key frequency band determination subunit is used to determine the key frequency bands of the target child's speech based on the data from the registration stage, so as to obtain the child's features; the calculation subunit is used to perform a fast Fourier transform on the original speech signal to obtain the spectrum and calculate the cosine similarity between the spectrum and the mask of the child's features; the frequency mask generation subunit is used to mark the spectrum as non-target interference when the cosine similarity is lower than a preset threshold and generate a binarized frequency mask; the removal subunit is used to apply medium-level filtering to smooth the binarized frequency mask, remove isolated noise points, and form a continuous time-frequency mask trajectory.

[0135] In one embodiment, the suppression unit 304 includes:

[0136] The dot product subunit is used to perform a dot product operation on the spectrum of the original speech signal using the time-frequency mask trajectory to suppress non-child voice source components in the frequency domain to obtain the remaining signal components; the secondary enhancement subunit is used to perform secondary enhancement on the remaining signal components to obtain the enhanced spectrum; the transform subunit is used to perform an inverse fast Fourier transform on the enhanced spectrum and use the overlap-add method to recover the time-domain continuous waveform to obtain the target child speech signal.

[0137] In one embodiment, the secondary enhancement subunit is used to perform secondary enhancement on the remaining signal components using a Wiener filter based on prior signal-to-noise ratio estimation to obtain the enhanced spectrum.

[0138] In one embodiment, the secondary enhancement subunit is used to apply a Wiener filter based on prior signal-to-noise ratio estimation to the remaining signal components. The gain function of the Wiener filter is dynamically adjusted according to the local signal-to-noise ratio to filter out residual stationary background noise while preserving the formant structure of the child's speech, so as to obtain the enhanced spectrum.

[0139] In one embodiment, the feature extraction unit 305 includes:

[0140] The sub-vector extraction sub-unit is used to generate static MFCC features from the target child's speech signal through Mel filter bank, logarithmic energy conversion, and discrete cosine transform, and to calculate the first and second order differences to capture the temporal variation characteristics of the speech, forming a multidimensional voiceprint feature sub-vector. The appending sub-unit is used to extract the fundamental frequency of the target child's speech signal within a specific frequency range using the autocorrelation function method combined with center clipping technology, and to solve for the first three formant frequencies using the linear predictive coding analysis method, and append them together with the normalized statistical values ​​to the multidimensional voiceprint feature sub-vector to obtain multidimensional voiceprint features. The vector processing sub-unit is used to normalize the multidimensional voiceprint features through cepstral mean variance to eliminate channel effects and individual volume differences, and then define them as the current input and registered template feature vectors to obtain a high-dimensional voiceprint feature vector.

[0141] In one embodiment, the additional subunit includes:

[0142] The fundamental frequency extraction module is used to extract the fundamental frequency of the target child's speech signal within a set frequency range using the autocorrelation function method combined with center clipping technology; the formant extraction module is used to use the linear predictive coding analysis method, set the prediction order, and solve for the frequencies of the first three formants to obtain the formants; the feature addition module is used to determine the normalized fundamental frequency mean, variance, and center frequency of the formants, and add them to the multidimensional voiceprint feature subvector to obtain multidimensional voiceprint features.

[0143] It should be noted that those skilled in the art can clearly understand that the specific implementation process of the above-mentioned child voiceprint feature extraction system 300 and each unit can be referred to the corresponding description in the foregoing method embodiments. For the sake of convenience and brevity, it will not be repeated here.

[0144] The aforementioned child voiceprint feature extraction system 300 can be implemented as a computer program, which can, for example... Figure 3 It runs on the computer device shown.

[0145] Please see Figure 3 , Figure 3 This is a schematic block diagram of a computer device provided in an embodiment of this application. The computer device 500 can be a server, wherein the server can be a standalone server or a server cluster composed of multiple servers.

[0146] See Figure 3 The computer device 500 includes a processor 502, a memory, and a network interface 505 connected via a system bus 501. The memory may include a non-volatile storage medium 503 and internal memory 504.

[0147] The non-volatile storage medium 503 may store an operating system 5031 and a computer program 5032. The computer program 5032 includes program instructions that, when executed, cause the processor 502 to perform a method for extracting a child's voiceprint features.

[0148] The processor 502 provides computing and control capabilities to support the operation of the entire computer device 500.

[0149] The internal memory 504 provides an environment for the operation of the computer program 5032 in the non-volatile storage medium 503. When the computer program 5032 is executed by the processor 502, the processor 502 can perform a method for extracting a child's voiceprint features.

[0150] This network interface 505 is used for network communication with other devices. Those skilled in the art will understand that... Figure 3 The structure shown is merely a block diagram of a portion of the structure related to the present application and does not constitute a limitation on the computer device 500 to which the present application is applied. The specific computer device 500 may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements.

[0151] The processor 502 is used to run a computer program 5032 stored in a memory to implement all the steps of the child voiceprint feature extraction method.

[0152] It should be understood that in the embodiments of this application, the processor 502 may be a central processing unit (CPU), or it may be other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. The general-purpose processor may be a microprocessor or any conventional processor.

[0153] It will be understood by those skilled in the art that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program includes program instructions and can be stored in a storage medium, which is a computer-readable storage medium. The program instructions are executed by at least one processor in the computer system to implement the process steps of the embodiments of the above methods.

[0154] Therefore, the present invention also provides a storage medium. This storage medium can be a computer-readable storage medium. The storage medium stores a computer program, wherein when executed by a processor, the computer program causes the processor to perform all the steps of the child voiceprint feature extraction method.

[0155] The storage medium can be any computer-readable storage medium capable of storing program code, such as a USB flash drive, portable hard drive, read-only memory (ROM), magnetic disk, or optical disk.

[0156] Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, computer software, or a combination of both. To clearly illustrate the interchangeability of hardware and software, the components and steps of the various examples have been generally described in terms of functionality in the foregoing description. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementations should not be considered beyond the scope of this invention.

[0157] In the embodiments provided by this invention, it should be understood that the disclosed systems and methods can be implemented in other ways. For example, the system embodiments described above are merely illustrative. For example, the division of each unit is only a logical functional division, and there may be other division methods in actual implementation. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed.

[0158] The steps in the method of this invention can be adjusted, merged, or reduced in order according to actual needs. The units in the system of this invention can be merged, divided, or reduced according to actual needs. Furthermore, the functional units in the various embodiments of this invention can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit.

[0159] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a storage medium. Based on this understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, a terminal, or a network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention.

[0160] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any person skilled in the art can easily conceive of various equivalent modifications or substitutions within the technical scope disclosed in the present invention, and these modifications or substitutions should all be covered within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A method for extracting voiceprint features from children, characterized in that, include: Acquire the raw speech signal; The original speech signal is identified by identifying vocal segments and multi-peak energy structure to determine whether multiple sound sources overlap. When multiple sound sources overlap, the current frame spectrum of the original speech signal is analyzed, and non-target interference frequency components are marked and suppressed by comparing it with a pre-established child speech feature mask to identify the child sound source and obtain the time-frequency mask trajectory. The frequency spectrum of the original speech signal is weighted and suppressed using the time-frequency mask trajectory, and noise is eliminated using an adaptive Wiener filter to reconstruct the target child's speech signal. Multi-dimensional voiceprint features are extracted from the target child's speech signal and normalized to construct a high-dimensional voiceprint feature vector. The multi-dimensional voiceprint features include MFCC, fundamental frequency, and formants.

2. The method for extracting children's voiceprint features according to claim 1, characterized in that, The step of identifying vocal segments and multi-peak energy structures in the original speech signal to determine whether multiple sound sources overlap includes: The original speech signal is preprocessed, framed, and its short-time energy and zero-crossing rate are calculated to identify sound segments and multi-peak energy structures in order to determine whether multiple sound sources overlap.

3. The method for extracting children's voiceprint features according to claim 2, characterized in that, The preprocessing, framing, short-time energy and zero-crossing rate calculation of the original speech signal, and the identification of vocal segments and multi-peak energy structures to determine whether multiple sound sources overlap, include: The original speech signal is processed using a pre-emphasis filter to enhance the high-frequency components and balance the spectrum to obtain a pre-processed signal. The preprocessed signal is divided into several sampling points to form multiple time window signals, so as to obtain the signal of each frame; The short-time energy of each frame of signal is calculated to identify sound segments. By detecting the extreme values ​​of the energy envelope and combining them with a set energy threshold, when the short-time energy of several consecutive frames exceeds the energy threshold, it is confirmed that the signal has a multi-peak energy structure. Calculate the average zero-crossing rate of each frame of signal to distinguish different types of phonemes and noise types; By combining short-time energy and average zero-crossing rate, when high energy is found to be accompanied by an abnormally high zero-crossing rate, it can be determined that there is multi-source aliasing or strong noise interference in the current environment.

4. The method for extracting children's voiceprint features according to claim 1, characterized in that, When multiple sound sources overlap, the current frame spectrum of the original speech signal is analyzed, and non-target interference frequency components are marked and suppressed by comparing it with a pre-established child speech feature mask to identify the child sound source and obtain the time-frequency mask trajectory, including: Based on the data from the registration phase, the key frequency bands of the target children's speech were determined to obtain the children's characteristics; Perform a Fast Fourier Transform on the original speech signal to obtain the spectrum, and calculate the cosine similarity between the spectrum and the mask of the child's features; When the cosine similarity is lower than a preset threshold, the spectrum is marked as non-target interference, and a binarized frequency mask is generated. The binarized frequency mask is smoothed by applying medium-frequency filtering to remove isolated noise points and form a continuous time-frequency mask trajectory.

5. The method for extracting children's voiceprint features according to claim 1, characterized in that, The process of using the time-frequency mask trajectory to perform weighted suppression on the spectrum of the original speech signal, using an adaptive Wiener filter to eliminate noise, and reconstructing the target child's speech signal includes: The time-frequency mask trajectory is used to perform a dot product operation on the spectrum of the original speech signal to suppress non-children's voice source components in the frequency domain, so as to obtain the remaining signal components; The remaining signal components are then enhanced a second time to obtain the enhanced spectrum; The enhanced spectrum is subjected to an inverse fast Fourier transform, and the time-domain continuous waveform is recovered using the overlap-addition method to obtain the target child's speech signal.

6. The method for extracting children's voiceprint features according to claim 5, characterized in that, The secondary enhancement of the remaining signal components to obtain the enhanced spectrum includes: The remaining signal components are then enhanced a second time using a Wiener filter based on prior signal-to-noise ratio estimation to obtain the enhanced spectrum.

7. The method for extracting children's voiceprint features according to claim 1, characterized in that, The step of performing secondary enhancement on the remaining signal components using a Wiener filter based on prior signal-to-noise ratio estimation to obtain the enhanced spectrum includes: The remaining signal components are subjected to a Wiener filter based on prior signal-to-noise ratio estimation. The gain function of the Wiener filter is dynamically adjusted according to the local signal-to-noise ratio. While preserving the formant structure of the child's speech, residual stationary background noise is filtered out to obtain the enhanced spectrum.

8. The method for extracting children's voiceprint features according to claim 7, characterized in that, The step of extracting multi-dimensional voiceprint features from the target child's speech signal, performing normalization processing, and constructing a high-dimensional voiceprint feature vector includes: The target child's speech signal is processed by Mel filter bank, logarithmic energy conversion and discrete cosine transform to generate static MFCC features, and first-order and second-order differences are calculated to capture the temporal variation characteristics of the speech, forming a multidimensional voiceprint feature sub-vector. The autocorrelation function method combined with center clipping technology is used to extract the fundamental frequency of the target child's speech signal within a specific frequency range. The first three formant frequencies are solved using the linear predictive coding analysis method. These frequencies, along with the normalized statistical values, are added to the multidimensional voiceprint feature subvector to obtain multidimensional voiceprint features. The multi-dimensional voiceprint features are normalized by cepstral mean variance to eliminate channel effects and individual volume differences, and then defined as the current input and registered template feature vectors to obtain high-dimensional voiceprint feature vectors.

9. The method for extracting children's voiceprint features according to claim 8, characterized in that, The method employs the autocorrelation function method combined with center clipping technology to extract the fundamental frequency of the target child's speech signal within a specific frequency range, and uses the linear predictive coding analysis method to solve for the first three formant frequencies. These frequencies, along with the normalized statistical values, are then appended to the multidimensional voiceprint feature sub-vector to obtain multidimensional voiceprint features, including: The fundamental frequency of the target child's speech signal is extracted within a set frequency range by using the autocorrelation function method combined with center clipping technology. Using the linear predictive coding analysis method, the prediction order is set, and the frequencies of the first three resonant peaks are solved to obtain the resonant peaks; The normalized fundamental frequency mean, variance, and center frequency of the formant are determined and added to the multidimensional voiceprint feature subvector to obtain multidimensional voiceprint features.

10. A system for extracting voiceprint features from children, characterized in that, include: The acquisition unit is used to acquire the raw speech signal; The judgment unit is used to identify the sound segments and multi-peak energy structure of the original speech signal in order to determine whether there are multiple overlapping sound sources. The trajectory determination unit is used to analyze the current frame spectrum of the original speech signal when multiple sound sources overlap, and to mark and suppress non-target interference frequency components by comparing with a pre-established child speech feature mask to identify the child sound source and obtain the time-frequency mask trajectory. The suppression unit is used to weighted suppress the spectrum in the original speech signal using the time-frequency mask trajectory, eliminate noise using an adaptive Wiener filter, and reconstruct the target child's speech signal; The feature extraction unit is used to extract multi-dimensional voiceprint features from the target child's speech signal, and perform normalization processing to construct a high-dimensional voiceprint feature vector. The multi-dimensional voiceprint features include MFCC, fundamental frequency, and formants.

Images