Voiceprint model training method and device, user identity detection method and device and intelligent household electrical appliance

Abstract

The invention relates to the technical field of voice recognition, and discloses a voiceprint model training method, a user identity detection method and device, and an intelligent household electrical appliance, and the voiceprint model training method comprises the steps: carrying out the preprocessing of a voice signal, and extracting a mixed feature vector containing a static frequency spectrum feature and a time difference feature; training a Gaussian mixture model to be converged by adopting an expectation maximization algorithm based on the mixed feature vector; and outputting the special voiceprint model of the speaker. The voiceprint model obtained by training through the training method is used for user identity detection, and high robustness, high real-time performance and high accuracy of voiceprint recognition in a complex noise environment are achieved.

Description

Technical Field

[0001] The present application relates to the field of speech recognition technology, for example, to a voiceprint model training method, a user identity detection method, a device and a smart home appliance. Background Art

[0002] Speech recognition-based identity detection technology has important applications in numerous fields, such as home appliance control, security, finance, and mobile device authentication. As application scenarios expand, higher requirements are placed on the accuracy, robustness, real-time performance, and versatility of speaker identification.

[0003] In order to improve the accuracy of user identity detection, the relevant technology discloses a training method and device for a speech recognition system, and a speech recognition method and device. The acoustic model is trained by extracting acoustic and voiceprint features from the audio data of the target speaker in the training sample of the speech recognition system and merging them as convolutional neural network input.

[0004] Although the relevant technology has achieved a higher recognition accuracy, its solution is based on silence detection with a fixed threshold to identify the noise and silence parts in the audio. It cannot adapt to the noise parameters of different environments. When the intensity of the environmental noise changes, it is easy to misjudge the silence segment or retain the noise segment, which increases the system deployment and maintenance costs and affects the accuracy and robustness of the voiceprint recognition system.

[0005] It should be noted that the information disclosed in the above background technology section is only used to enhance the understanding of the background of this application, and therefore may include information that does not constitute prior art known to ordinary technicians in this field. Summary of the Invention

[0006] In order to provide a basic understanding of some aspects of the disclosed embodiments, a brief summary is given below. The summary is not an extensive review, nor is it intended to identify key / critical elements or delineate the scope of protection of these embodiments, but rather serves as a prelude to the detailed description that follows.

[0007] The embodiments of the present disclosure provide a voiceprint model training method, a user identity detection method, a device, and a smart home appliance to improve the accuracy and robustness of speaker identity recognition in noisy environments.

[0008] In some embodiments, the voiceprint model training method includes: The speech signal is preprocessed to extract a mixed feature vector including static spectrum features and time difference features; based on the mixed feature vector, a Gaussian mixture model is trained using an expectation maximization algorithm until convergence; and a speaker-specific voiceprint model is output.

[0009] Optionally, the preprocessing of the speech signal to extract a mixed feature vector containing static spectrum features and their time difference features includes: using a high-pass filter to pre-emphasize the speech signal to suppress low-frequency noise and enhance high-frequency speech features; framing and windowing the pre-emphasized speech signal; performing frequency domain conversion and Mel filtering on the windowed speech signal; performing feature extraction on the energy spectrum after Mel filtering to generate 20-dimensional MFCC coefficients representing static spectrum features; calculating the time difference of each frame of MFCC coefficients to generate a 20-dimensional incremental feature representing the time difference feature; and obtaining a 40-dimensional mixed feature vector based on the 20-dimensional MFCC coefficients and the 20-dimensional incremental features.

[0010] Optionally, based on the mixed eigenvector, an expectation-maximization algorithm is used to train a Gaussian mixture model until convergence, including: obtaining an initialized Gaussian mixture model, the training parameters of the Gaussian mixture model including the number of Gaussian components, the mean of each component, the covariance matrix and the weight coefficient; iteratively performing the following steps: calculating the probability that each eigenvector belongs to the kth Gaussian component; updating the mean, covariance matrix and weight coefficient of each component according to the probability that each eigenvector belongs to the kth Gaussian component; and determining convergence when the rate of change of the log-likelihood function value of adjacent iterations is less than a convergence threshold; the convergence threshold is 0.001.

[0011] Optionally, the calculation of the probability that each eigenvector belongs to the kth Gaussian component includes: using a simplified probability density function to calculate the likelihood that the eigenvector belongs to the kth Gaussian component; combining the calculated likelihood and the prior weights of each component, and calculating the probability that each eigenvector belongs to the kth Gaussian component through the Bayesian formula.

[0012] In this way, by extracting a mixed feature vector containing static spectral features and their time difference features, the representation of transient changes in speech is enhanced and the interference of steady-state background noise is suppressed; the computational efficiency of the model is optimized and the training time is compressed to achieve a simultaneous improvement in the accuracy, robustness and computational efficiency of voiceprint recognition.

[0013] In some embodiments, the user identity detection method based on the voiceprint model includes: The user voice to be detected is preprocessed to extract a mixed feature vector of the user voice; the voice frames and noise frames of the user voice are distinguished through a two-stage decision strategy; the mixed feature vectors of the voice frames are input into a pre-trained voiceprint model to generate a user identity recognition result, the result including a matched speaker identity or a match failure judgment; wherein the pre-trained voiceprint model is trained using the above-mentioned training method; when a preset number of noise frames are continuously detected, the noise parameter value in the two-stage decision strategy is updated.

[0014] Optionally, preprocess the user voice to be detected and extract the mixed feature vector of the user voice, including: performing pre-emphasis, framing and windowing preprocessing on the user voice to be detected; extracting a 40-dimensional mixed feature vector, which includes 20-dimensional MFCC coefficients representing static spectral features and 20-dimensional delta features representing time difference features.

[0015] Optionally, distinguish speech frames and noise frames of the user voice through a two-stage decision strategy, including: calculating the short-time energy E of the frame to be detected after the user voice is framed; if E < Emin, determine the frame to be detected as a noise frame; if E ≥ Emin, perform frequency-domain sub-band division on the frame to be detected; Emin is the minimum energy threshold; calculate the speech-noise likelihood log ratio Li of each sub-band and the global likelihood log ratio sum Lt; if there is any sub-band satisfying Li > T τ or the global likelihood log ratio sum Lt > T a , determine the frame to be detected as a speech frame; otherwise, determine it as a noise frame; T τ is the local threshold, and T a is the global threshold.

[0016] Optionally, when a preset number of noise frames are continuously detected, update the noise parameter value in the two-stage judgment strategy, including: when a preset number of noise frames are continuously detected, perform a moving average update on the noise parameter value in the two-stage judgment strategy; apply the updated noise parameter to the discrimination of speech frames and noise frames of the subsequent frames of the user.

[0017] Optionally, input the mixed feature vector of the speech frame into the pre-trained voiceprint model to generate the user identity recognition result, including: calculating the likelihood log ratio LLR of the mixed feature vector of the speech frame and each voiceprint model in the database; if the highest LLR value is greater than the recognition threshold, determine that the corresponding speaker identity is matched as the user identity recognition result; otherwise, return a match failure as the user identity recognition result.

[0018] In this way, through the preprocessing of the speech signal and the extraction of mixed features, the cooperation of noise suppression and feature enhancement is realized. At the same time, through the two-stage decision strategy, low-energy noise frames are quickly filtered, reducing invalid calculations, forming a cascaded mechanism of "coarse screening plus refined judgment", and reducing the error rate of the system in a non-stationary noise environment; in addition, through the update of noise parameters, the environmental noise adaptability is optimized, and the speaker features are prevented from drifting due to noise changes. The high robustness, high real-time performance and high accuracy of voiceprint recognition in a complex noise environment are achieved.

[0019] In some embodiments, the voiceprint model training device includes a processor and a memory storing program instructions, and the processor is configured to execute the voiceprint model training method as described above when running the program instructions.

[0020] In some embodiments, the user identity detection device based on the voiceprint model includes a processor and a memory storing program instructions, and the processor is configured to execute the user identity detection method based on the voiceprint model as described above when running the program instructions.

[0021] In some embodiments, the smart home appliance includes a product body, which is equipped with the voiceprint model training device as described above and / or the user identity detection device based on the voiceprint model as described above.

[0022] The voiceprint model training method, user identity detection method, device, and smart home appliance provided by the embodiments of the present disclosure can achieve the following technical effects: The introduction of a silence detection algorithm that combines hybrid features with sub-band spectrum segmentation solves the problem of traditional methods being sensitive to dynamic noise. A hybrid feature vector with time-difference features is used to capture the temporal changes in speech and improve feature discrimination; at the same time, the sub-band dynamic parameter update adapts to the noise spectrum distribution in real time, effectively improving recognition accuracy. In addition, the disclosed embodiment significantly reduces computational complexity and further shortens decision time by simplifying Gaussian distribution calculation and sub-band parallel processing architecture, meeting the needs of real-time voice authentication. Finally, by updating noise parameters, the ability to adapt to environmental noise is optimized to avoid drift of speaker features due to noise changes. This achieves high robustness, high real-time performance, and high accuracy in voiceprint recognition in complex noise environments.

[0023] The above general description and the following description are exemplary and explanatory only and are not intended to limit the present application. BRIEF DESCRIPTION OF THE DRAWINGS

[0024] One or more embodiments are exemplarily described by corresponding drawings. These exemplary descriptions and drawings do not limit the embodiments. Elements with the same reference numerals in the drawings are shown as similar elements. The drawings do not constitute a scale limitation. In addition, Figure 1 is an environmental diagram of an implementation environment of an embodiment of the present disclosure; Figure 2 is a schematic diagram of a voiceprint model training method provided by an embodiment of the present disclosure; Figure 3 is a schematic diagram of a user identity detection method based on a voiceprint model provided by an embodiment of the present disclosure; Figure 4 is a schematic diagram of a first user identity detection device based on a voiceprint model according to an embodiment of the present disclosure; Figure 5 is a schematic diagram of a voiceprint model training device according to an embodiment of the present disclosure; Figure 62 is a schematic diagram of a second user identity detection device based on a voiceprint model provided in an embodiment of the present disclosure.

[0025] Reference numerals: 100, smart home appliances; 200, routers; 300, servers; 41. Sound acquisition module; 42. Feature extraction module; 43. Model training module; 44. Model matching and verification module; 45. Dynamic parameter adaptation module; 500, first processor; 501, first memory; 502, first communication interface; 503, first bus; 600, second processor; 601, second memory; 602, second communication interface; 603, second bus. DETAILED DESCRIPTION

[0026] In order to be able to understand the features and technical content of the embodiments of the present disclosure in more detail, the implementation of the embodiments of the present disclosure is described in detail below in conjunction with the accompanying drawings. The accompanying drawings are for reference only and are not used to limit the embodiments of the present disclosure. In the following technical description, for the sake of convenience of explanation, a full understanding of the disclosed embodiments is provided through multiple details. However, one or more embodiments can still be implemented without these details. In other cases, to simplify the drawings, well-known structures and devices can be simplified for display.

[0027] In the description and claims of the embodiments of the present disclosure, as well as in the accompanying drawings, the terms "first," "second," and the like are used to distinguish similar items and are not necessarily used to describe a particular order or precedence. It should be understood that the terms used in this manner are interchangeable where appropriate to describe the embodiments of the present disclosure herein. In addition, the terms "including," "having," and any variations thereof are intended to cover non-exclusive inclusions.

[0028] Unless otherwise stated, the term "plurality" means two or more.

[0029] In the embodiment of the present disclosure, the character " / " indicates that the preceding and following objects are in an "or" relationship. For example, A / B means: A or B.

[0030] The term "and / or" describes an association between objects, indicating that three relationships can exist. For example, A and / or B means: A or B, or A and B.

[0031] The term "correspondence" may refer to an association relationship or a binding relationship. The correspondence between A and B means that there is an association relationship or a binding relationship between A and B.

[0032] In the embodiments of the present disclosure, smart home appliances refer to home appliance products that are formed by introducing microprocessors, sensor technology, and network communication technology into home appliances. They have the characteristics of intelligent control, intelligent perception, and intelligent application. The operation process of smart home appliances often relies on the application and processing of modern technologies such as the Internet of Things, the Internet, and electronic chips. For example, smart home appliances can realize remote control and management of smart home appliances by users by connecting to electronic devices.

[0033] In the disclosed embodiments, a terminal device refers to an electronic device with a wireless connection function. The terminal device can communicate with the above-mentioned smart home appliances by connecting to the Internet, or can communicate with the above-mentioned smart home appliances directly through Bluetooth, Wi-Fi, etc. In some embodiments, the terminal device is, for example, a mobile device, a computer, or a vehicle-mounted device built into a hover car, or any combination thereof. Mobile devices may include, for example, mobile phones, smart home devices, wearable devices, smart mobile devices, virtual reality devices, etc., or any combination thereof, wherein wearable devices include, for example, smart watches, smart bracelets, pedometers, etc.

[0034] Figure 1 Schematic diagram of the implementation environment of the embodiment of the present disclosure. Figure 1 As shown, the implementation environment may include a smart home appliance 100 , a router 200 and a server 300 .

[0035] Smart home appliances 100 can connect to the Internet using a Wi-Fi (Wireless Fidelity) module. Smart home appliances 100 include smart home appliances, such as smart speakers, smart refrigerators, and smart TVs. These devices typically have built-in or external microphones for capturing user voice input and built-in or external speakers for outputting synthesized voice responses. They can also include mobile devices, such as smartphones and tablets, which implement voice interaction capabilities through built-in microphones and speakers. They can also be computer devices equipped with microphones and speakers.

[0036] The router 200 is used to connect all devices in the home and transmit the voice data of these devices to the Internet or a local server.

[0037] Server 300 can be a single server, a server cluster consisting of multiple servers, or a cloud computing service center, which is not limited in the present embodiment. It is used to process user voice interaction information, including complex computing tasks such as voiceprint recognition, natural language processing, and speech synthesis. Cloud servers typically have powerful processing capabilities and a large amount of storage space.

[0038] It should be understood that Figure 1The number of smart home appliances, routers and servers in the figure is only for reference. Any number of smart home appliances, routers and servers may be provided according to actual needs. For example, one mobile terminal may correspond to multiple smart home appliances.

[0039] In this way, when users use smart home appliances such as smart speakers and smart air conditioners at home, they can wake up the devices and issue commands or questions through voice. After the smart appliances capture the user's voice, they use the locally deployed voiceprint model to verify the user's identity, or transmit the voice data to the server through the home router and use the voiceprint model deployed in the cloud to verify the user's identity.

[0040] Figure 2 It is a schematic diagram of a voiceprint model training method provided by an embodiment of the present disclosure.

[0041] Combine Figure 2 As shown, the method includes: Step S201 : pre-processing the speech signal to extract a mixed feature vector including static spectrum features and time difference features thereof.

[0042] By preprocessing the speech signal, performing pre-emphasis and frame windowing, low-frequency noise is effectively suppressed, spectrum leakage is reduced, and the speech signal input to the subsequent model is purer. The noise contamination rate of the mixed feature vector is reduced by 30%, laying the foundation for accurate extraction of speech features.

[0043] Step S202: Based on the mixture feature vector, the Gaussian mixture model is trained using the expectation maximization algorithm until convergence.

[0044] By iteratively optimizing model parameters through the expectation-maximization algorithm, the model can adapt to the distribution of training data and accurately model the probability density of the speaker's mixed feature vector. The log-likelihood function value is improved by 30%. Each Gaussian component corresponds to a typical pattern of speaker characteristics, such as the characteristics of different speech units such as high-frequency fricatives and low-frequency voiced sounds, improving the model's ability to express speaker characteristics.

[0045] Step S203: outputting a speaker-specific voiceprint model.

[0046] The Gaussian mixture model after training convergence is output as the speaker-specific voiceprint model, and the initial parameters of the noise model are calculated based on the speech frames screened out in the training phase and output as the initial values of the noise parameters.

[0047] In this way, by extracting a mixed feature vector containing static spectral features and their time difference features, the representation of transient changes in speech is enhanced and the interference of steady-state background noise is suppressed; the computational efficiency of the model is optimized and the training time is compressed to achieve a simultaneous improvement in the accuracy, robustness and computational efficiency of voiceprint recognition.

[0048] Optionally, the speech signal is preprocessed to extract a mixed feature vector containing static spectrum features and their time difference features, including: using a high-pass filter to pre-emphasize the speech signal to suppress low-frequency noise and enhance high-frequency speech features; framing and windowing the pre-emphasized speech signal; performing frequency domain conversion and Mel filtering on the windowed speech signal; performing feature extraction on the energy spectrum after Mel filtering to generate 20-dimensional MFCC (Mel-Frequency Cepstral Coefficients) coefficients representing static spectrum features; calculating the time difference of the MFCC coefficients of each frame to generate a 20-dimensional incremental feature representing the time difference feature; and obtaining a 40-dimensional mixed feature vector based on the 20-dimensional MFCC coefficients and the 20-dimensional incremental features.

[0049] Furthermore, a high-pass filter is used to pre-emphasize the speech signal, including: Input the original speech signal x(n) and use a high-pass filter to remove low-frequency noise interference according to the following formula:

[0050] Where X(n) represents the sample value of the original speech signal at discrete time point n; Y(n) represents the sample value of the speech signal after pre-emphasis processing at discrete time point n; μ is the pre-emphasis coefficient; In the embodiment of the present disclosure, the value of μ is 0.97. is the pre-emphasis function for continuous signals.

[0051] This operation increases the energy of the high-frequency components in the original signal by subtracting 0.97 times the delay of one sampling point from the original signal, suppressing low-frequency noise interference and making the high-frequency components of the speech signal more prominent. This enhances the intelligibility of the speech signal and provides a higher-quality signal foundation for subsequent feature extraction.

[0052] Furthermore, the pre-emphasized speech signal is framed and windowed, including: The pre-emphasized signal is divided into consecutive subframes with a frame length of 25ms and a 10ms overlap; Each frame of speech data x(n) is truncated using the Hamming window function according to the following formula:

[0053] in, a is the adjustment coefficient, N is the frame length, W(n, a ) is the windowed speech signal, which is used to eliminate the spectrum leakage of FFT. In the embodiment of the present disclosure, a The value of is 0.46.

[0054] Through framing, the continuous speech signal is divided into short time segments suitable for processing; through windowing operation, the spectrum leakage caused by framing is reduced, making the signal more accurate in the frequency domain.

[0055] Furthermore, the windowed speech signal is subjected to frequency domain conversion and Mel filtering, including: The windowed speech signal x(n) is subjected to a fast Fourier transform to obtain the frequency domain spectrum data Xa(k), where k is the frequency index, thus realizing the conversion of the speech signal from the time domain to the frequency domain. The spectrum is smoothed using a Mel filter bank consisting of 6 triangular bandpass filters. The frequency range covers 80Hz to 4000Hz. The center frequency of each subband m is f(m) After mapping using the Mel-scale formula, the signals are evenly spaced, and a triangular bandpass filter Hm(k) is defined accordingly. This filter bank simulates the characteristics of human hearing, compressing the high-frequency band through non-uniform frequency division, thereby reducing the energy of high-frequency harmonic components. Simultaneously, its filtering characteristics match the frequency distribution of speech formants, effectively preserving the formant characteristics and highlighting the essential information of the speech signal. Xa(k) is then filtered and smoothed to produce the Mel-filtered spectrum data. Here, the FFT and Mel-filter bank calculations for the six subbands are distributed to a multi-core processor to improve real-time performance.

[0056] Furthermore, feature extraction is performed on the energy spectrum after Mel filtering to generate 20-dimensional MFCC coefficients that represent static spectrum features, including: Perform discrete cosine transform (DCT) on the energy spectrum after Mel filtering to obtain the 20-dimensional MFCC coefficient C(n):

[0057] Where n is the MFCC coefficient order (1≤n≤20), m is the bandpass filter order, s(m) is the energy value after filtering, and M is the number of bandpass filters, which corresponds to the number of subbands.

[0058] The DCT transform can convert spectral information into cepstral domain features, further compressing the data dimension while focusing on the spectral envelope characteristics of the speech signal, effectively extracting key information that can characterize the static spectral characteristics of speech, such as the speaker's vocal tract shape, resonance peak position, etc.

[0059] Furthermore, the temporal difference of the MFCC coefficients of each frame is calculated to generate a 20-dimensional incremental feature representing the temporal difference feature, including: The temporal difference of the MFCC coefficients of each frame is calculated according to the following formula to obtain the 20-dimensional incremental feature (ΔMFCC):

[0060] Where dt represents the incremental MFCC feature value of the current frame, which is used to characterize the dynamic change rate of MFCC features over time; C t+n Represents the MFCC coefficient of the nth frame after t frame, C t-n represents the MFCC features of the nth frame before the tth frame; N represents the upper limit of the offset. In the embodiment disclosed in the present disclosure, N=2, which means that the features of the current frame ±1 frame and ±2 frames are used; is a normalization coefficient used to balance the contribution of different offsets to the differential result.

[0061] This operation calculates the difference between the MFCCs of the current frame and the preceding and following frames, weights them with an offset n, and finally normalizes them to obtain the incremental feature dt. This approach focuses on capturing temporal dynamic features, capturing the changes in the MFCC coefficients of adjacent frames to characterize the temporal dynamics of speech signals, such as pitch variations, speech rate, and other prosodic information.

[0062] The 20-dimensional MFCC coefficients are combined with the 20-dimensional incremental features to obtain a 40-dimensional hybrid feature vector. This vector integrates the static spectral features and time-difference features of the speech signal, comprehensively and richly describing the characteristic information of the speech signal, providing high-quality input data for subsequent voiceprint model training.

[0063] Optionally, based on the mixed eigenvector, an expectation-maximization algorithm is used to train a Gaussian mixture model until convergence, including: obtaining an initialized Gaussian mixture model, the training parameters of the Gaussian mixture model include the number of Gaussian components, the mean of each component, the covariance matrix and the weight coefficient; iteratively performing the following steps: calculating the probability that each eigenvector belongs to the kth Gaussian component; updating the mean, covariance matrix and weight coefficient of each component according to the probability that each eigenvector belongs to the kth Gaussian component; converging is determined when the rate of change of the log-likelihood function value of adjacent iterations is less than a convergence threshold; the convergence threshold is 0.001.

[0064] Furthermore, the training parameters of the Gaussian mixture model include the number of Gaussian components K, the mean value of each component μ k , covariance matrix Σk and weight coefficient ω k .

[0065] The number of Gaussian components K is set to 8, and the speaker feature space is divided into 8 sub-distributions. Each component corresponds to a speech feature pattern (such as high-frequency fricative sounds, low-frequency voiced sounds, etc.).

[0066] Randomly sample K feature vectors from the training data as the initial mean, μ k Assign values to ensure that the model covers different regions of the feature space.

[0067] Set the covariance matrix Σk to the identity matrix or the global sample variance.

[0068] The weight coefficient ω k Set to 1 / K, assuming that each component contributes equally to the overall distribution.

[0069] In this way, random sampling of the mean ensures that the initial model covers the feature space and avoids local optimality.

[0070] Furthermore, the probability of each eigenvector belonging to the kth Gaussian component is calculated, including: Use the simplified probability density function to calculate the likelihood that the eigenvector belongs to the kth Gaussian component; Combining the calculated likelihood and the prior weights of each component, the probability that each eigenvector belongs to the kth Gaussian component is calculated using the Bayesian formula.

[0071] in, The simplified probability density function is used for calculation and obtained through the following formula:

[0072] in, It represents the probability density value of the eigenvector x under the condition of given Gaussian component parameters. In this way, the probability of belonging to each component is calculated.

[0073] This simplified form is obtained by omitting the constant term in the standard Gaussian formula. , retaining only exponential operations, can reduce 30% of floating-point operations.

[0074] Furthermore, the calculated likelihood and the prior weights of each component are combined to calculate the probability that each eigenvector belongs to the kth Gaussian component using the Bayesian formula, including: The probability of each speech sample belonging to the kth Gaussian component is calculated as follows:

[0075] in, represents the probability that the i-th sample belongs to the k-th Gaussian component, ω k represents the weight coefficient of the kth Gaussian component, P k （x i ） represents the probability density function of the kth component, x i represents the mixed feature vector of the i-th sample, μ k represents the mean of the k-th Gaussian component, and Σk represents the covariance matrix of the k-th Gaussian component.

[0076] Furthermore, the mean, covariance matrix, and weight coefficient of each component are updated according to the probability that each eigenvector belongs to the kth Gaussian component, including: Update the parameters of each component according to the following formula.

[0077]

[0078]

[0079]

[0080] In this way, the above steps are repeated iteratively and the process stops when the likelihood change rate of two consecutive iterations is less than the convergence threshold to ensure that the model parameters are stable.

[0081] Furthermore, when the rate of change of the log-likelihood function value of adjacent iterations is less than a convergence threshold, convergence is determined, including: The log-likelihood function value of all training samples after the tth iteration is calculated according to the following formula:

[0082] Where N is the number of speech frames, K is the Gaussian component value, The updated model training parameters for the tth iteration.

[0083] The rate of change of the log-likelihood function value between adjacent iterations is calculated as follows:

[0084] in, is the log-likelihood function value at the t-1th iteration.

[0085] when When , stop iteration; When the convergence threshold is 0.001, it means that the log-likelihood value between two adjacent iterations changes by less than 0.1%. At this time, the model parameters have been fully optimized and the benefits of further iterations are minimal.

[0086] In this way, through the voiceprint model training method provided by the embodiment of the present disclosure, by extracting a mixed feature vector containing static spectral features and their time difference features, the representation of transient changes in speech is enhanced and the interference of steady-state background noise is suppressed; and the computational efficiency of the model is optimized, the training time is compressed, and the voiceprint recognition accuracy, robustness and computational efficiency are simultaneously improved.

[0087] Figure 3 This is a schematic diagram of a user identity detection method based on a voiceprint model provided by an embodiment of the present disclosure. Figure 1 In the environment shown.

[0088] Combine Figure 3As shown, the method includes: Step S301 : pre-processing the user voice to be detected, and extracting a mixed feature vector of the user voice.

[0089] In response to the user's voice information, it is pre-processed as the user's voice to be detected to determine the speaker's identity. After the sound is collected, silence detection is performed on the collected voice stream data in real time. When the speaker finishes speaking, a feature model is extracted from the speaker's voice sample for this sentence, and in subsequent steps, it is matched with the feature models already in the database.

[0090] The preprocessing process includes framing, windowing, and pre-emphasis. By preprocessing the voice signal, pre-emphasis and framing and windowing are performed to effectively suppress low-frequency noise and compensate for the high-frequency attenuation of the voice signal during transmission.

[0091] A hybrid feature vector is an extraction scheme that incorporates multiple speech features. In this disclosed embodiment, the hybrid feature vector includes 20-dimensional MFCC coefficients representing static spectral features and 20-dimensional incremental features representing temporal difference features. The 20-dimensional MFCC coefficients and 20-dimensional incremental features are combined to create a 40-dimensional hybrid feature vector. Hybrid features integrate both time-domain and frequency-domain information, enabling a more comprehensive representation of speech characteristics than single features, improving noise immunity and speaker discrimination.

[0092] Step S302: Distinguish the speech frames and noise frames of the user's speech through a two-stage decision strategy.

[0093] This two-stage strategy combines the speed of a simple threshold with the robustness of a statistical model, reducing false positives caused by a single threshold. In the first stage, a coarse judgment filters out most noise frames, leaving only the suspected speech frames for detailed judgment, reducing computational complexity. In the second stage, the model adapts to real-time noise changes by updating parameters, improving VAD performance in complex environments.

[0094] In step S303, the mixed feature vector of the speech frame is input into the pre-trained voiceprint model to generate a user identity recognition result, which includes a matched speaker identity or a match failure judgment; wherein the pre-trained voiceprint model is trained by the above-mentioned training method; when a preset number of noise frames are continuously detected, the noise parameter value is updated.

[0095] The pre-trained voiceprint model is based on a Gaussian mixture model, whose model parameters are trained using the mixed feature vectors of speech samples during the training phase. A matching result is obtained by calculating the log-likelihood ratio of the mixed feature vectors of the input speech frames to the pre-trained voiceprint model. This disclosed embodiment also triggers the update of noise parameters by continuously detecting a preset number of noise frames, enabling the system to track background noise changes in real time, reducing noise interference with feature extraction and model matching. This is particularly effective in non-stationary noise environments.

[0096] Thus, through the preprocessing of the voice signal and the extraction of mixed features, the collaboration between noise suppression and feature enhancement is achieved. At the same time, through the two-stage decision-making strategy, low-energy noise frames are quickly filtered to reduce ineffective calculations, forming a cascaded mechanism of "coarse screening plus refined judgment", which reduces the error rate of the system in a non-stationary noise environment. In addition, through the update of noise parameters, the environmental noise adaptability is optimized to prevent the speaker features from drifting due to noise changes. The high robustness, high real-time performance and high accuracy of voiceprint recognition in a complex noise environment are realized.

[0097] Optionally, preprocess the voice of the user to be detected and extract the mixed feature vector of the user's voice, including: Use a high-pass filter to pre-emphasize the voice signal to suppress low-frequency noise and enhance high-frequency voice features; frame and window the pre-emphasized voice signal; perform frequency-domain conversion and Mel filtering on the windowed voice signal; extract features from the energy spectrum after Mel filtering to generate 20-dimensional MFCC coefficients representing static spectral features; calculate the time difference for each frame of MFCC coefficients to generate 20-dimensional delta features representing time-difference features; obtain a 40-dimensional mixed feature vector based on the 20-dimensional MFCC coefficients and the 20-dimensional delta features.

[0098] It should be understood that the specific implementation methods of the preprocessing and mixed feature extraction of the user's voice are similar to the steps of the preprocessing and mixed feature extraction of the voice signal in the above voiceprint model training method, and will not be elaborated here.

[0099] The voice activity detection algorithm of the present disclosure adopts a two-stage decision-making strategy, combining sub-band local judgment and the sum of global likelihood log ratios to distinguish the voice frames and noise frames of the user's voice.

[0100] Optionally, distinguish the voice frames and noise frames of the user's voice through a two-stage decision-making strategy, including: Calculate the short-time energy E of the frame to be detected after the user's voice is framed; If E≥Emin, perform frequency-domain sub-band division on the frame to be detected; Emin is the minimum energy threshold.

[0101] Calculate the voice-noise likelihood log ratio Li of each sub-band and the sum of global likelihood log ratios Lt; If there exists any sub-band satisfying Li>T τ or the sum of global likelihood log ratios Lt>T a , then determine the frame to be detected as a voice frame; Otherwise, determine it as a noise frame; T τ is the local threshold, and T a is the global threshold; If E<Emin, then determine the frame to be detected as a noise frame.

[0102] Optionally, the local threshold T is adjusted according to the quietness of the environment. τ .

[0103] For example, the local threshold T is adjusted by the following formula: τ Size:

[0104] Where M is the number of sub-band divisions; is the environmental coefficient, 0< <2.

[0105] The quieter the environment, the better the environmental coefficient The higher the value, the noisier the environment. For example, in a quiet environment, =1.2; in a noisy environment, =0.8.

[0106] Short-term energy reflects signal amplitude strength, and speech frames typically have higher energy than noise frames. Here, a coarse judgment is performed to filter out most noise frames, and only the suspected speech frames are subjected to a detailed judgment, reducing computational complexity.

[0107] The short-time energy E can be calculated by calculating the sum of squares of the amplitudes after windowing the speech frame signal, reflecting the instantaneous strength of the signal. In the embodiment of the present disclosure, the window length and frame shift are set to 25ms and 10ms to ensure time domain resolution.

[0108] Furthermore, the frame to be detected is divided into frequency domain sub-bands, including: The frequency domain is divided into M overlapping or non-overlapping sub-bands through a Mel filter bank, each sub-band corresponds to a specific frequency range, simulating the human ear's perception of frequency. Specifically, in the embodiment of the present disclosure, the value of M is 6, and the frequency ranges corresponding to the sub-bands are 80 to 250 Hz, 250 to 500 Hz, 500 to 1000 Hz, 1000 to 2000 Hz, 2000 to 3000 Hz, and 3000 to 4000 Hz, respectively. The frequency domain is decomposed into multiple sub-bands, and the noise characteristics of different frequency bands (such as high-frequency noise and low-frequency noise) are processed separately to avoid interference of broadband noise on the overall judgment.

[0109] The speech-to-noise likelihood logarithm ratio Li of each subband is calculated by the following formula:

[0110] Wherein, x and y are two characteristic parameters of sub-band i, which are energy and spectrum envelope respectively in the embodiment of the present disclosure; P s is the probability that the subband belongs to a speech frame, P n is the probability that the subband belongs to a noisy frame.

[0111] Specifically, the probability that subband i belongs to speech or noise is calculated according to the following formula:

[0112]

[0113] in, is the mean energy of the speech class in subband i, is the energy variance of the speech class in subband i; is the mean energy of the noise class in subband i, is the energy variance of the noise class in subband i.

[0114]

[0115]

[0116] in, is the frequency envelope mean of the speech class in subband i, is the frequency envelope variance of the speech class in subband i; is the frequency envelope mean of the noise class in subband i, is the frequency envelope variance of the noise class in subband i.

[0117] Furthermore, the global likelihood log ratio sum Lt is calculated by the following formula:

[0118] Among them, K i is the subband weight coefficient, 1≤K i ≤3.

[0119] Optionally, the Ki coefficient can be dynamically adjusted according to the sub-band signal-to-noise ratio, and the sub-band with a high signal-to-noise ratio is given a higher weight, such as K i =3; the weight of the subband with low signal-to-noise ratio is reduced, such as K i =1.

[0120] Thus, if there exists any subband satisfying Li>T τ Or the total log-likelihood ratio Lt>T a , then the frame is determined to be speech; otherwise it is determined to be noise. τ =3.0, T a =15.0.

[0121] Here, the likelihood logarithm ratio is calculated based on the probability model and can more accurately characterize the statistical difference between speech and noise than the simple energy threshold method. The total likelihood logarithm ratio reflects the degree of match between the overall frequency domain features and the speech and noise models. If the subband satisfies Li>Tτ , indicating that the subband feature is highly matched with the speech model, so it can be determined as a speech frame; if the total logarithmic ratio of the global likelihood Lt>T a , indicating that the overall frequency domain features conform to the speech model distribution, and therefore can be identified as a speech frame. Here, dual thresholding is used for judgment: the local threshold captures the speech characteristics of a specific frequency band, while the global threshold suppresses the cumulative impact of cross-band noise. This avoids single-condition misjudgments such as those caused by localized burst noise exceeding the limit, thus reducing the misjudgment rate in noisy environments.

[0122] Optionally, when a preset number of noise frames are detected continuously, the noise parameter value is updated, including: when a preset number of noise frames are detected continuously, performing a sliding average update on the noise parameter value; and applying the updated noise parameter to the speech frame and noise frame discrimination of subsequent frames.

[0123] Here, a timer can be used to record the number of noise frame detections. For example, a sliding window counter can be maintained. When a frame is determined to be a noise frame, the counter is incremented by 1; if it is determined to be a speech frame, the counter is reset. When the counter reaches a preset number, a noise parameter update is triggered, ensuring that the parameters are updated only in a persistent noise environment to avoid transient noise interference. In the disclosed embodiment, the preset number is 10 frames, that is, the update operation is triggered when 10 consecutive noise frames are detected.

[0124] Specifically, the noise parameters are updated using the sliding average method in the noise frame in the following way:

[0125]

[0126] in, is the mean energy of the noise class in subband i, is the energy variance of the noise class in subband i, a is the adaptive coefficient, 0.9≤ a ≤0.95.

[0127] By setting the adaptive coefficient, the parameters are updated slowly to avoid overfitting.

[0128] The current noise mean is the mean parameter of the kth Gaussian component, reflecting the statistical center of the current noise feature.

[0129] The current noise mean is the instantaneous statistical mean of the characteristic parameters of the current noise frame, which is used to represent the average intensity of the current noise frame in the six subbands. The current noise variance is the instantaneous statistical dispersion of the characteristic parameters of the current noise frame, which is used to quantify the uniformity of the current noise distribution in the frequency domain.

[0130] It should be understood that the noise parameters 、 The update method of is the same as the update method of the above energy mean and energy variance, which will not be described here.

[0131] This sliding average update allows the system to adapt to non-stationary noise environments in real time. Adaptive parameters improve speech frame recognition accuracy even in fluctuating noise energy scenarios. Furthermore, sliding averages require only constant time complexity, reducing computational overhead by over 95% compared to retraining the noise model. Furthermore, there's no need to store historical noise frames; only current parameters and counters are maintained. This reduces memory usage from megabytes to kilobytes, making it suitable for deployment on embedded devices.

[0132] Optionally, the mixed feature vector of the speech frame is input into a pre-trained voiceprint model to generate a user identification result, including: Calculate the likelihood logarithm ratio (LLR) between the mixed feature vector of the speech frame and each voiceprint model in the database; If the highest LLR value is greater than the recognition threshold, the corresponding speaker identity is determined to be matched as the user identity recognition result; otherwise, a match failure is returned as the user identity recognition result.

[0133] Here, the preprocessed speech frame's mixed feature vector is input into the pretrained voiceprint model. The model uses forward propagation to calculate the degree of match between the feature vector and each speaker model in the database. Because the mixed feature vector incorporates both time-domain and frequency-domain speech features, it more comprehensively represents individual characteristics such as the speaker's timbre and intonation, improving robustness against noise and individual differences. For each voiceprint model in the database, the likelihood ratio (LLR) is calculated between the probability that the input feature vector belongs to that model and the probability that it belongs to the background noise model. By quantifying the degree of match between the features and the target model, the probability comparison is converted to a difference calculation in the logarithmic domain to avoid underflow for small probabilities. Logarithmic operations are also used to convert multiplication into addition, improving computational efficiency. The LLR values of all speaker models are compared, and the maximum value is compared to the recognition threshold. If the highest LLR exceeds Ta, the corresponding speaker identity is determined to be matched; otherwise, a match failure is returned.

[0134] The recognition threshold is used to balance the false positive rate (misidentifying non-target speakers as targets) and the false negative rate (misidentifying target speakers as non-targets). It can be adjusted dynamically based on scenario requirements. In this disclosed embodiment, the recognition threshold is set to 3.0. A match is considered successful if the LLR is greater than 3.0, and the speaker identity corresponding to the model with the highest LLR is output as the user identity detection result.

[0135] Furthermore, the recognition threshold Ta is adjusted according to the scenario category to dynamically balance system security and user experience: In high-security scenarios, the recognition threshold is increased, the false recognition rate is reduced, and the false recognition rate is controlled below 0.1%; In high-convenience scenarios, lower the recognition threshold, reduce the rejection rate, and control the rejection rate below 5%.

[0136] In this way, through the technical chain of "multi-feature fusion - probability quantization - threshold determination", a balance between computational efficiency and environmental adaptability is achieved while ensuring recognition accuracy. The hybrid feature vector enhances the model's ability to capture the speaker's individual characteristics, LLR calculation provides a standardized matching metric, and the threshold strategy offers flexible parameter adjustment for different application scenarios.

[0137] Furthermore, the likelihood logarithm ratio (LLR) between the mixed feature vector of the speech frame and each voiceprint model in the database is calculated according to the following formula:

[0138] Among them, P(x|spk) refers to the probability that the input mixed feature vector belongs to the target speaker model; P(x|noise) refers to the probability that the input mixed feature vector belongs to the background noise.

[0139] The probability that the mixed feature vector belongs to the target speaker model is calculated by the following formula:

[0140] Where K is the number of Gaussian components, represents the weight of the kth component, The aforementioned simplified probability density function is used for calculation and will not be described in detail here.

[0141] The probability that the mixed eigenvector belongs to the background noise is calculated by the following formula:

[0142] Where M is the number of Gaussian components of the noise model, represents the weight of the mth component, The aforementioned simplified probability density function is used for calculation, which will not be described in detail here, wherein each noise parameter can be updated by the aforementioned sliding average method.

[0143] In this way, by comparing the probability distribution of the target speaker and the background noise, the characteristic differences between different speakers are magnified, and the system's discrimination ability is improved.

[0144] The user identity detection device based on the voiceprint model provided by the embodiment of the present disclosure solves the problem of traditional methods being sensitive to dynamic noise by introducing a silence detection algorithm that combines mixed features with sub-band spectrum segmentation. The problem of traditional methods being sensitive to dynamic noise is solved by introducing a silence detection algorithm that combines 40-dimensional mixed features with 6-sub-band spectrum segmentation. Incremental MFCC captures temporal changes in speech, improves feature discrimination, and increases recognition accuracy from 92% to 97%. The sliding average mechanism dynamically updates parameters, adapting to the noise spectrum distribution in real time, avoiding misjudgments caused by fixed thresholds, and maintaining an accuracy of over 95% in a noisy environment with a signal-to-noise ratio of 5dB.

[0145] Secondly, the disclosed embodiments significantly reduce computational complexity and decision time by simplifying Gaussian distribution calculations and subband parallel processing architecture, thus meeting the requirements of real-time voice authentication. By simplifying Gaussian distribution calculations and removing constant terms, computational complexity is significantly reduced, reducing floating-point operations by 30%. Combined with the parallel task allocation of multi-core processors, system latency is compressed from 70ms in traditional methods to 45ms. Furthermore, multi-level threshold linkage, such as environmental coefficient adjustment, avoids high-dimensional spatial mapping, further shortening decision time and meeting the requirements of real-time voice authentication.

[0146] Thirdly, dynamic parameter adaptive updates are triggered by noise frames, and noise parameters are updated through sliding average. The adaptation time is less than 0.5 seconds, which optimizes the environmental noise adaptation capability and prevents speaker characteristics from drifting due to noise changes. The fixed speech parameters ensure the stability of speaker characteristics.

[0147] The overall solution achieves high robustness, high real-time performance and high accuracy in voiceprint recognition in complex noise environments. In cross-environment (alternating quiet / noisy) tests, the recognition rate fluctuation is reduced by 13% (the traditional method drops by 15%, while the present invention only drops by 2%).

[0148] Combine Figure 4 As shown, the embodiment of the present disclosure provides a user identity detection device based on a voiceprint model, including a sound collection module 41, a feature extraction module 42, a model training module 43, a model matching and verification module 44, and a dynamic parameter adaptation module 45. The modules are connected through data streams and control signals to form a closed-loop real-time processing architecture.

[0149] The output of the sound collection module 41 is connected to the feature extraction module 42, which simultaneously outputs a silence detection result and a voiceprint feature vector. The silence detection result is connected to the dynamic parameter adaptation module 45 via a feedback path, which returns the updated parameters to the silence detection submodule of the feature extraction module 42. The model training module 43 is directly associated with the sound collection module 41 during the registration phase and stores the trained voiceprint model in the database. During the verification phase, the model matching and verification module 44 receives the voiceprint feature vector output by the feature extraction module 42, calls the trained voiceprint recognition module from the database for calculation, and ultimately outputs the identity recognition result.

[0150] Specifically, the sound collection module 41 is configured to collect original voice signals through a microphone or a sound card and output digital voice stream data.

[0151] The feature extraction module 42 is configured to pre-process the voice stream data, including framing, windowing, and spectrum analysis, and extract 40-dimensional mixed features (MFCC + incremental MFCC) and 6 sub-band spectrum parameters. At the same time, a two-stage decision strategy is implemented to distinguish between the voice frames and noise frames of the user's voice to filter out background noise.

[0152] The model training module 43 is configured to perform Gaussian mixture model training on the speech features in the registration phase, generate a speaker-specific voiceprint model and store it in a database.

[0153] The model matching and verification module 44 is configured to input the feature vector of the speech to be recognized into the trained voiceprint model, and calculate the likelihood logarithm ratio (LLR) between the feature vector and each voiceprint model in the database to determine the matching result.

[0154] The dynamic parameter adaptation module 45 is configured to dynamically adjust the noise parameters used to distinguish between speech frames and noise frames of the user's speech in the two-stage decision strategy stage according to the real-time environmental noise characteristics, so as to achieve environmental adaptation capability.

[0155] The sound acquisition module 41 uses a condenser microphone (such as Knowles SPM0638) and a sound card (such as Creative Sound Blaster), with a sampling rate of 44.1kHz and a quantization bit of 16 bits. Software implementation: Developed in Python, the core algorithm library is optimized based on NumPy and SciPy; Use TensorFlow or PyTorch framework to achieve fast training and inference of GMM models; System operating environment: Linux Ubuntu 20.04 system, supporting multi-threaded parallel processing.

[0156] Combine Figure 5 As shown, an embodiment of the present disclosure provides a voiceprint model training device, including a first processor 500 and a first memory 501. Optionally, the device may also include a first communication interface 502 and a first bus 503. The first processor 500, the first communication interface 502, and the first memory 501 can communicate with each other through the first bus 503. The first communication interface 502 can be used for information transmission. The first processor 500 can call the logic instructions in the first memory 501 to execute the voiceprint model training method of the above embodiment.

[0157] In addition, the logic instructions in the first memory 501 can be implemented in the form of software functional units and can be stored in a computer-readable storage medium when sold or used as an independent product.

[0158] The first memory 501, as a computer-readable storage medium, can be used to store software programs and computer-executable programs, such as the program instructions / modules corresponding to the methods in the embodiments of the present disclosure. The first processor 500 executes the program instructions / modules stored in the first memory 501 to perform functional applications and data processing, thereby implementing the voiceprint model training method in the above-mentioned embodiments.

[0159] The first memory 501 may include a program storage area and a data storage area. The program storage area may store an operating system and at least one application required for a function; the data storage area may store data generated based on the use of the terminal device. Furthermore, the first memory 501 may include a high-speed random access memory and a non-volatile memory.

[0160] The first processor 500 uses a multi-core embedded processor (such as Intel Core i7 or ARM Cortex-A76) and is responsible for feature extraction and model calculation; The first memory 501 uses an SSD hard disk to store the GMM model database, supporting fast reading and writing (writing speed ≥ 300MB / s).

[0161] Combine Figure 6 As shown, an embodiment of the present disclosure provides a user identity detection device based on a voiceprint model, comprising a second processor 600 and a second memory 601. Optionally, the device may further comprise a second communication interface 602 and a second bus 603. The second processor 600, the second communication interface 602, and the second memory 601 may communicate with each other via the second bus 603. The second communication interface 602 may be used for information transmission. The second processor 600 may invoke logic instructions in the second memory 601 to execute the user identity detection method based on a voiceprint model according to the above embodiment.

[0162] In addition, the logic instructions in the second memory 601 can be implemented in the form of software functional units and can be stored in a computer-readable storage medium when sold or used as an independent product.

[0163] The second memory 601, as a computer-readable storage medium, can be used to store software programs and computer-executable programs, such as the program instructions / modules corresponding to the methods in the embodiments of the present disclosure. The second processor 600 executes the program instructions / modules stored in the second memory 601 to perform functional applications and data processing, thereby implementing the user identity detection method based on the voiceprint model in the above-mentioned embodiments.

[0164] The second memory 601 may include a program storage area and a data storage area. The program storage area may store an operating system and at least one application required for a function; the data storage area may store data generated based on the use of the terminal device. Furthermore, the second memory 601 may include a high-speed random access memory and a non-volatile memory.

[0165] The embodiment of the present disclosure also provides a smart home appliance 100, including: a product body, equipped with the above-mentioned voiceprint model training device and / or the above-mentioned user identity detection device based on the voiceprint model. The installation relationship described here is not limited to placement inside the product body, but also includes installation connections with other components of the smart home appliance 100, including but not limited to physical connections, electrical connections or signal transmission connections, etc. Those skilled in the art will understand that the voiceprint model training device and the user identity detection device based on the voiceprint model can be adapted to a feasible product body, thereby realizing other feasible embodiments. Optionally, the smart home appliance 100 is an air conditioner. In this way, through the user identity detection method and user identity detection device provided by the embodiment of the present disclosure, it is possible to detect the identity of the user who issues the air conditioner operation instruction, thereby executing the operation parameters related to the user identity and improving the environmental comfort.

[0166] An embodiment of the present disclosure provides a computer-readable storage medium storing computer-executable instructions, wherein the computer-executable instructions are configured to execute the above-mentioned voiceprint model training method and / or the user identity detection method based on the voiceprint model.

[0167] The technical solutions of the embodiments of the present disclosure may be embodied in the form of a software product, which is stored in a storage medium and includes one or more instructions for causing a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the method described in the embodiments of the present disclosure. The aforementioned storage medium may be a non-transitory storage medium, such as a USB flash drive, a mobile hard drive, a read-only memory (ROM), a random access memory (RAM), a magnetic disk, or an optical disk, among other media capable of storing program code.

[0168] The above description and the accompanying drawings sufficiently illustrate the embodiments of the present disclosure to enable those skilled in the art to practice them. Other embodiments may include structural, logical, electrical, process and other changes. The embodiments represent only possible variations. Unless expressly required, individual components and functions are optional, and the order of operations may vary. Portions and features of some embodiments may be included in or replace portions and features of other embodiments. Moreover, the terms used in this application are only used to describe the embodiments and are not used to limit the claims. As used in the description of the embodiments and claims, the singular forms "a", "an" and "the" are intended to also include the plural forms unless the context clearly indicates otherwise. Similarly, the term "and / or" as used in this application means any and all possible combinations of one or more of the associated listings. In addition, when used in this application, the term "comprise" and its variations "comprises" and / or comprising refer to the presence of stated features, wholes, steps, operations, elements, and / or components, but do not exclude the presence or addition of one or more other features, wholes, steps, operations, elements, components and / or groups thereof. In the absence of further restrictions, an element defined by the sentence "comprising a..." does not exclude the presence of other identical elements in the process, method or device that includes the element. In this article, each embodiment may focus on the differences from other embodiments, and the same and similar parts between the various embodiments can be referenced to each other. For the methods, products, etc. disclosed in the embodiments, if they correspond to the method part disclosed in the embodiments, then the relevant parts can be referred to the description of the method part.

[0169] Those skilled in the art will appreciate that the units and algorithm steps of each example described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are performed in hardware or software may depend on the specific application and design constraints of the technical solution. The technicians may use different methods to implement the described functions for each specific application, but such implementation should not be considered to be beyond the scope of the embodiments of the present disclosure. The technicians will clearly understand that, for the convenience and brevity of description, the specific working processes of the systems, devices and units described above can refer to the corresponding processes in the aforementioned method embodiments and will not be repeated here.

[0170] In the embodiments disclosed herein, the disclosed methods and products (including but not limited to devices and equipment) can be implemented in other ways. For example, the device embodiments described above are merely illustrative. For example, the division of units may be merely a logical functional division. In actual implementation, other divisions may be used, such as combining or integrating multiple units or components into another system, or omitting or disabling some features. Furthermore, the couplings or direct couplings or communication connections shown or discussed may be through interfaces, indirect couplings or communication connections between devices or units, and may be electrical, mechanical, or other forms. The units described as separate components may or may not be physically separate, and the components shown as units may or may not be physical units, i.e., they may be located in one location or distributed across multiple network units. Some or all of these units may be selected to implement the embodiments according to actual needs. Furthermore, the functional units in the disclosed embodiments may be integrated into a single processing unit, each unit may exist physically separately, or two or more units may be integrated into a single unit.

[0171] The flowcharts and block diagrams in the accompanying drawings show the possible implementation architectures, functions and operations of the systems, methods and computer program products according to the embodiments of the present disclosure. In this regard, each box in the flowchart or block diagram can represent a module, program segment or part of the code, and the module, program segment or part of the code contains one or more executable instructions for implementing the specified logical functions. In some alternative implementations, the functions marked in the box can also occur in an order different from that marked in the accompanying drawings. For example, two consecutive boxes can actually be executed substantially in parallel, or they can sometimes be executed in the opposite order, which can depend on the functions involved. In the descriptions corresponding to the flowcharts and block diagrams in the accompanying drawings, the operations or steps corresponding to different boxes can also occur in an order different from that disclosed in the description, and sometimes there is no specific order between different operations or steps. For example, two consecutive operations or steps can actually be executed substantially in parallel, or they can sometimes be executed in the opposite order, which can depend on the functions involved. Each block in the block diagrams and / or flowcharts, and combinations of blocks in the block diagrams and / or flowcharts, may be implemented by a dedicated hardware-based system that performs the specified function or action, or may be implemented by a combination of dedicated hardware and computer instructions.

Claims

1. A voiceprint model training method, characterized in that: It includes: Preprocess the speech signal to extract a mixed feature vector containing static spectral features and their time difference features; Based on the mixed feature vector, use the Expectation-Maximization algorithm to train the Gaussian mixture model until convergence; Output the speaker-specific voiceprint model.

2. The voiceprint model training method according to claim 1, characterized in that: The preprocessing of the speech signal to extract the mixed feature vector containing static spectral features and their time difference features includes: Use a high-pass filter to pre-emphasize the speech signal to suppress low-frequency noise and enhance high-frequency speech features; Frame and window the pre-emphasized speech signal; Perform frequency-domain conversion and Mel filtering on the windowed speech signal; Extract features from the energy spectrum after Mel filtering to generate 20-dimensional MFCC coefficients representing static spectral features; Calculate the time difference for each frame of MFCC coefficients to generate 20-dimensional delta features representing time difference features; Obtain a 40-dimensional mixed feature vector based on the 20-dimensional MFCC coefficients and the 20-dimensional delta features.

3. The voiceprint model training method according to claim 1, characterized in that: Based on the mixed feature vector, using the Expectation-Maximization algorithm to train the Gaussian mixture model until convergence includes: Obtain an initialized Gaussian mixture model, and the training parameters of the Gaussian mixture model include the number of Gaussian components, the mean of each component, the covariance matrix, and the weight coefficient; Iteratively execute the following steps: Calculate the probability that each feature vector belongs to the k-th Gaussian component; Update the mean, covariance matrix, and weight coefficient of each component according to the probability that each feature vector belongs to the k-th Gaussian component; Determine convergence when the change rate of the log-likelihood function value between adjacent iterations is less than the convergence threshold; the convergence threshold is 0.

001.

4. The voiceprint model training method according to claim 3, characterized in that: The calculation of the probability that each feature vector belongs to the k-th Gaussian component includes: Use a simplified probability density function to calculate the likelihood that the feature vector belongs to the k-th Gaussian component; Combine the calculated likelihood and the prior weights of each component, and calculate the probability that each feature vector belongs to the k-th Gaussian component through Bayes' formula.

5. A user identity detection method based on a voiceprint model, characterized in that: It includes: Preprocess the speech of the user to be detected to extract the mixed feature vector of the user's speech; Distinguish the speech frames and noise frames of the user's speech through a two-stage decision strategy; Input the mixed feature vector of the speech frame into the pre-trained voiceprint model to generate a user identity recognition result, and the result includes the matching speaker identity or a determination of matching failure; Among them, the pre-trained voiceprint model is trained by the training method described in any one of claims 1 to 4; When a preset number of noise frames are continuously detected, update the noise parameter value in the two-stage decision strategy.

6. The user identity detection method according to claim 5, characterized in that: The preprocessing of the speech of the user to be detected to extract the mixed feature vector of the user's speech includes: Perform pre-emphasis, framing, and windowing preprocessing on the speech of the user to be detected; Extract a 40-dimensional mixed feature vector, which contains 20-dimensional MFCC coefficients representing static spectral features and 20-dimensional delta features representing time difference features.

7. The user identity detection method according to claim 5, characterized in that: The distinguishing of the speech frames and noise frames of the user's speech through the two-stage decision strategy includes: Calculate the short-time energy E of the frame to be detected after the user's speech is framed; If E < Emin, then determine the frame to be detected as a noise frame; If E≥Emin, the frame to be detected is divided into frequency domain sub-bands; Emin is the minimum energy threshold; Calculate the speech-to-noise likelihood logarithm ratio Li of each subband and the total global likelihood logarithm ratio Lt; If there exists any subband satisfying Li>T τ Or the total log-likelihood ratio Lt>T a , then the frame to be detected is determined to be a speech frame; otherwise it is determined to be a noise frame; T τ is the local threshold, T a is the global threshold.

8. The user identity detection method according to claim 5, characterized in that: When a preset number of noise frames are continuously detected, updating the noise parameter value in the two-stage judgment strategy includes: When a preset number of noise frames are detected continuously, a sliding average update is performed on the noise parameter value in the two-stage judgment strategy; The updated noise parameters are applied to the speech frame and noise frame discrimination of the user's subsequent frames.

9. The user identity detection method according to claim 5, characterized in that: The mixed feature vector of the speech frame is input into the pre-trained voiceprint model to generate the user identity recognition results, including: Calculate the likelihood logarithm ratio (LLR) between the mixed feature vector of the speech frame and each voiceprint model in the database; If the highest LLR value is greater than the recognition threshold, the corresponding speaker identity is determined to be matched as the user identity recognition result; Otherwise, a match failure is returned as the user identification result.

10. A voiceprint model training device, comprising a processor and a memory storing program instructions, characterized in that: The processor is configured to execute the voiceprint model training method according to any one of claims 1 to 4 when running the program instructions.

11. A user identity detection device based on a voiceprint model, comprising a processor and a memory storing program instructions, characterized in that: The processor is configured to execute the user identity detection method based on the voiceprint model according to any one of claims 5 to 9 when running the program instructions.

12. A smart home appliance, characterized in that: include: The product body is equipped with the voiceprint model training device as claimed in claim 10 and / or the user identity detection device based on the voiceprint model as claimed in claim 11.

Images