Methods for identifying user identity based on voiceprint in voice IVR interaction
By acquiring hybrid voiceprint features during IVR interaction and using a background channel model and frequency domain complex-valued ICA algorithm for voiceprint feature separation, the problem of insufficient voiceprint recognition capability is solved, and efficient and accurate user authentication is achieved.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- 北京微呼科技有限公司
- Filing Date
- 2023-07-06
- Publication Date
- 2026-07-17
Smart Images
Figure CN116863943B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of voiceprint recognition, and more particularly to a method for identifying user identity based on voiceprint in voice IVR interaction. Background Technology
[0002] IVR stands for Interactive Voice Response, a powerful automated telephone service system that uses pre-recorded or TTS (Text-to-Speech) technology to synthesize speech for automatic responses. Voiceprint, on the other hand, is the sound wave spectrum carrying speech information displayed using electroacoustic instruments. Modern scientific research shows that voiceprints are not only specific but also relatively stable. After adulthood, a person's voice can remain relatively stable for a long period. Experiments have shown that regardless of whether a speaker deliberately imitates another person's voice and tone, or whispers, even if the imitation is very accurate, the voiceprint will always be different.
[0003] Prior art CN110619880A discloses a voiceprint processing system and user identification method, including: acquiring a user's first voice audio data, and converting the first voice audio data into first voiceprint data; acquiring a user's second voice audio data, and converting the second voice audio data into second voiceprint data, wherein the second voice audio data is the user's voice audio data responding to a mobile terminal; comparing the first voiceprint data and the second voiceprint data, and identifying the user corresponding to the second voiceprint data based on the comparison result;
[0004] The following problems exist:
[0005] (1) The separation between mixed voiceprint features is not obvious, and the ability to distinguish voiceprint features is poor. Summary of the Invention
[0006] To address the aforementioned problems, this invention proposes a method for identifying user identity based on voiceprint in voice IVR interaction, in order to more accurately resolve the problems described above.
[0007] This invention is achieved through the following technical solution:
[0008] This invention proposes a method for identifying user identity based on voiceprint in voice IVR interaction, including:
[0009] S1: Receive mixed voice data in IVR interaction, obtain the mixed voiceprint features of the mixed voiceprint features, and construct the corresponding mixed voiceprint feature vector based on the voiceprint features;
[0010] S2: Input the hybrid voiceprint feature vector into the pre-trained background channel model to construct the current voiceprint identification vector in the hybrid speech data;
[0011] S3: Based on the similarity matching between the current voiceprint identification vector and the voiceprint feature vector in the interactive voiceprint database, a matching score is obtained, and a judgment is made based on the matching score. If the matching score is greater than or equal to the threshold, the identity corresponding to the voiceprint feature in the interactive voiceprint database is queried according to the mapping relationship to identify the user's identity. If it is less than the threshold, the matching continues until the preset matching limit is reached and then stops.
[0012] Furthermore, in the method for identifying user identity based on voiceprint in the aforementioned voice IVR interaction, the step of receiving mixed voice data in the IVR interaction includes:
[0013] Pre-emphasis is performed based on mixed voice data received during IVR interaction;
[0014] The mixed speech data is framed according to the framing method using overlapping segments;
[0015] Apply Hamming window and distinguish between speech frames and non-speech frames, discarding the non-speech frames.
[0016] Furthermore, in the method for identifying user identity based on voiceprint in the aforementioned voice IVR interaction, the voice frame includes current voiceprint identification voice information and background sound information.
[0017] Furthermore, in the method for identifying user identity based on voiceprint in the aforementioned voice IVR interaction, the background channel model step includes:
[0018] Step 1: First, preprocess the speech frame vector X(t) in the obtained mixed voiceprint feature vector, and then divide the preprocessed speech frame X(t) into l segments on average;
[0019] Step 2: For each preprocessed speech frame vector X(t) = (x1, x2, ..., x... n Perform a Fast Fourier Transform; after the Fourier Transform, the speech frame vector is converted into a complex-valued speaker feature vector, which can be expressed by the following formula:
[0020]
[0021] in, Represented as the decomposition matrix of the background channel model, S = (s1, s2, ..., s...). m ) represents the current voiceprint identification vector matrix; the number of voice frame vector matrices is set to be the same as the current voiceprint identification vector matrix, i.e., n = m.
[0022] Furthermore, in the method for identifying user identity based on voiceprint in the aforementioned voice IVR interaction, after step two, the following steps are included:
[0023] Step 3: For each complex-valued speaker feature vector, perform linear separation of the speech frames using the frequency domain complex-valued ICA algorithm;
[0024] The algorithm steps are as follows:
[0025] The update iterative formula for W can be expressed by the following formula:
[0026] W(n+1)=E{zG[W T [(n)z]-G[W(n)z]}×W(n),
[0027] Where E{.} denotes the iterative expectation operation, G{.} denotes iterating over any non-quadratic function, W(n+1) denotes the iterative decomposition matrix at time n+1, and W(n) denotes the iterative decomposition matrix at time n, E{zG[W T [(n)z]-G[W(n)z]} represents the difference between the expected operation of the iteration at time n+1 and time n, z represents the iteration convergence step size, and T represents the number of iterations;
[0028] If W(n+1) = W(n), then the iteration ends; otherwise, it continues.
[0029] Step 4: After iteration, the decomposition matrix W of the background channel model is transformed back to the time domain using IFPT:
[0030] W′←IFFT(W),
[0031] Where W′ represents the time-domain decomposition matrix;
[0032] Step 5: Perform time-domain solution:
[0033] S = WX(t);
[0034] The time-domain solution process involves blindly separating each segment of the preprocessed speech frame vector X(t), adjusting the order of the output vectors to keep the signals of each output channel consistent, and then connecting them sequentially to obtain the final time-domain voiceprint output signal, i.e., the current voiceprint identification vector matrix.
[0035] Furthermore, in the method for identifying user identity based on voiceprint in the aforementioned voice IVR interaction, the step of obtaining a matching score by performing similarity matching between the current voiceprint identification vector and the voiceprint feature vector in the interactive voiceprint database includes:
[0036] The cosine similarity function is used to calculate the similarity between the current voiceprint identification vector and the sample voiceprint feature vectors of users in the interactive voiceprint database to obtain an evaluation score.
[0037] The evaluation score vector for each speech segment is obtained based on the evaluation score.
[0038] The score vector is then normalized using a softmax layer.
[0039] Furthermore, in the method for identifying user identity based on voiceprint in the aforementioned voice IVR interaction, after step S3, the method includes:
[0040] If the preset limit for matching attempts is reached and no matching of the voiceprint feature vector of a registered person is found, a voice prompt will be broadcast.
[0041] Furthermore, in the aforementioned method for identifying user identity based on voiceprint in voice IVR interaction, the interactive voiceprint database is obtained by learning from the corpus of daily conversations collected from multiple devices, and the feature value of each collected voice data; the interactive voiceprint database contains the user identity corresponding to the voiceprint features.
[0042] Furthermore, in the aforementioned method for identifying user identity based on voiceprint in voice IVR interaction, the feature values include feature values on the spectrum, cepstral, formants, pitch, reflection coefficient, prosody, rhythm, speed, intonation, and volume feature vectors.
[0043] The beneficial effects of this invention are:
[0044] (1) The present invention proposes to input the hybrid voiceprint feature vector into a pre-trained background channel model to construct the current voiceprint identification vector in the hybrid speech data; the obtained hybrid voiceprint feature vector X(t) is first preprocessed to eliminate the second-order correlation of the hybrid voiceprint feature vector, and the covariance matrix with zero mean is made into the identity matrix, and the preprocessed hybrid voiceprint feature vector X(t) is divided into l segments; a fast Fourier transform is performed on each segment of the preprocessed hybrid voiceprint feature vector; after the Fourier transform, the hybrid voiceprint feature vector is converted into a complex-valued voiceprint feature vector; for each segment of the complex-valued voiceprint feature vector, according to the frequency domain... The complex-valued ICA algorithm iterates through linear separation of hybrid voiceprint features. After iteration, the decomposition matrix W of the background channel model is transformed back to the time domain using IFPT. Finally, the current voiceprint discrimination vector matrix is obtained by solving in the time domain. The pre-trained background channel model is obtained through mining and comparison training of a large amount of speech data. This model can accurately depict the background voiceprint features of the user while preserving the user's voiceprint features to the maximum extent. It can also remove this feature during recognition and extract the inherent features of the user's voice, which can greatly improve the accuracy and efficiency of user authentication.
[0045] (2) The interactive voiceprint database proposed in this invention is obtained by learning from the corpus of daily calls collected from multiple devices, and the feature value of each voice data collected; the interactive voiceprint database contains the user identity corresponding to the voiceprint feature; users do not need to read a large amount of training text in advance to input voiceprint features and thus build a voiceprint feature library, but can use the corpus of daily calls and the voice data in instant messaging messages as training samples for the user's voiceprint features, thereby saving the user's time to input voiceprint features and improving the convenience of operation. Attached Figure Description
[0046] Figure 1 This is a flowchart illustrating the method for identifying user identity based on voiceprint in voice IVR interaction according to the present invention. Detailed Implementation
[0047] To more clearly and completely illustrate the technical solution of the present invention, the present invention will be further described below with reference to the accompanying drawings.
[0048] Please refer to Figure 1 This invention proposes a method for identifying user identity based on voiceprint in voice IVR interaction, including:
[0049] S1: Receive mixed voice data in IVR interaction, obtain the mixed voiceprint features of the mixed voiceprint features, and construct the corresponding mixed voiceprint feature vector based on the voiceprint features;
[0050] In this embodiment, the voice data is acquired by a voice acquisition device. During the call center inbound process, in the IVR interaction system, existing toolkits such as Kaldi, ESPNet, or Liborosa are used to receive mixed voice data from the training audio dataset. In the step of receiving mixed voice data in the IVR interaction, pre-emphasis is first applied based on the mixed voice data in the IVR interaction; then, the mixed voice data to be received is framed according to the overlapping segmentation framing method; a Hamming window is applied, and voice frames and non-voice frames are distinguished and discarded. Non-voice frames include silence segments, music, ringtones, caller ID tones, fax tones, etc., which can seriously affect the effect and efficiency of voiceprint recognition; the pre-emphasis processing is actually a high-pass filtering process to filter out… Low-frequency data makes the high-frequency characteristics in speech data more prominent. Specifically, the transfer function of the high-pass filter is: H(Z) = 1 - bZ⁻¹, where Z is the speech data, α is a constant coefficient, and b is 0.97. Since the sound signal is stationary only for a short period of time, a segment of the sound signal is divided into N short-time signals (i.e., N frames). To avoid losing the continuity of the sound, there is a repeating region between adjacent frames, which is generally half the length of each frame. After framing the speech data, each frame is treated as a stationary signal. However, due to the Gibbs effect, the start and end frames of the speech data are discontinuous. After framing, the data deviates further from the original speech. Therefore, windowing processing is required for the speech data. The speech frame includes the current voiceprint identification speech information and background sound information. The current voiceprint identification vector is extracted by first calculating the first and second order coefficients. The first order coefficients can be obtained by summing the probability matrix columns. Where Gamma is the i-th element of the first-order coefficient vector, Loglike ji Let represent the j-th row and i-th element of the probability matrix, where m represents the total number of mixed voiceprint feature vectors. The second-order coefficients can be obtained by multiplying the transpose of the probability matrix by the data matrix. X = Loglike T ×feats, where X represents the second-order coefficient matrix, Loglike is the probability matrix, feats represents the feature data matrix, and T represents the number of iterations;
[0051] S2: Input the hybrid voiceprint feature vector into the pre-trained background channel model to construct the current voiceprint identification vector in the hybrid speech data;
[0052] In this embodiment, the specific steps for inputting the hybrid voiceprint feature vector into the pre-trained background channel model are as follows:
[0053] Step 1: The speech frame vector X(t) in the acquired mixed voiceprint feature vector is first preprocessed. Preprocessing is implemented using a digital filter. The relationship between the output of the preprocessing network and the input speech signal s(n) is as follows: Where 'a' represents the pre-emphasis coefficient, typically a = 0.9375; the second-order correlation of the mixed voiceprint feature vectors is eliminated, and the covariance matrix with a mean of zero is used as the identity matrix. The preprocessed speech frame X(t) is divided into l segments on average. To improve the robustness of the system, noise reduction processing can also be performed on the speech signal. For example, by analyzing the short-time energy and short-time zero-crossing rate of the speech signal, the continuous speech signal is segmented into independent speech segments and non-speech segments. Then, front-end noise reduction processing is used to reduce interference from channel noise and background noise, improve the speech signal-to-noise ratio, and provide a clean signal for subsequent system processing.
[0054] Step 2: For each preprocessed speech frame vector X(t) = (x1, x2, ..., x... n Perform a Fast Fourier Transform; after the Fourier Transform, the speech frame vector is converted into a complex-valued speaker feature vector, which can be expressed by the following formula:
[0055] in, Represented as the decomposition matrix of the background channel model, S = (s1, s2, ..., s...). m () represents the current voiceprint identification vector matrix; to simplify the model, the number of speech frame vector matrices is set to be the same as the current voiceprint identification vector matrix, i.e., n = m;
[0056] Step 3: For each complex-valued speaker feature vector, perform linear separation of the speech frames using the frequency domain complex-valued ICA algorithm;
[0057] The algorithm steps are as follows:
[0058] The update iterative formula for W can be expressed by the following formula:
[0059] W(n+1)=E{zG[W T (n)z]-G[W(n)z]} T ×W(n),
[0060] Where E{.} denotes the iterative expectation operation, G{.} denotes iterating over any non-quadratic function, W(n+1) denotes the iterative decomposition matrix at time n+1, and W(n) denotes the iterative decomposition matrix at time n, E{zG[W T [(n)z]-G′[W(n)z]} represents the difference between the expected operation of the iteration at time n+1 and time n, z represents the iteration convergence step size, and T represents the number of iterations;
[0061] If W(n+1) = W(n), then the iteration ends; otherwise, it continues.
[0062] Step 4: After iteration, the decomposition matrix W of the background channel model is transformed back to the time domain using IFPT:
[0063] W′←IFFT(W),
[0064] Where W′ represents the time-domain decomposition matrix;
[0065] Step 5: Perform time-domain solution:
[0066] S = W′X(t);
[0067] The time-domain solution process involves blindly separating each segment of the preprocessed speech frame vector X(t), adjusting the order of the output vectors to keep the signals of each output channel consistent, and then connecting them sequentially to obtain the final time-domain voiceprint output signal, i.e., the current voiceprint identification vector matrix.
[0068] In this experiment, the selected signals were voice signals simultaneously recorded with a microphone in a typical office room as the received sound during the call, music signals played from a tape recorder as background, and multiple voices chatting. The voice signals, music signals, and microphones were positioned at the four vertices of a square, each 60cm apart. The sampling frequency of the signals was 16kHz, with a total of 110,000 sampling points. The voice frame vector X(t) in the first obtained mixed voiceprint feature vector was preprocessed using a digital filter to eliminate the second-order correlation of the mixed voiceprint feature vector and make its covariance matrix with zero mean the identity matrix. The preprocessed voice frame X(t) was then divided into l segments. Each preprocessed voice frame vector segment was further processed. Perform a Fast Fourier Transform, given the speech frame X(t) and the decomposition matrices for each time segment, because... If the decomposition matrices for different time periods are inconsistent, the decomposition matrix is iterated until W(n+1) = W(n) is obtained.
[0069] W(n+1)=E{zG[W T (n)z]-G[W(n)z]}×W(n), where E{.} denotes the iterative expectation operation, G{.} denotes iterating any non-quadratic function, W(n+1) denotes the iterative decomposition matrix at time n+1, and W(n) denotes the iterative decomposition matrix at time n, E{zG[W T [W(n)z]-G[W(n)z]} represents the difference between the expected operation of the iteration at time n+1 and time n, z represents the iteration convergence step size, and T represents the number of iterations. When the expected operation of the iteration is equal to 1, we can get W(n+1)=W(n).
[0070] After each iteration, if W(n+1) = W(n), the iteration ends; otherwise, it continues.
[0071] Step 4: After iteration, the decomposition matrix W of the background channel model is transformed back to the time domain using IFPT:
[0072] W′←IFFT(W),
[0073] Where W′ represents the time-domain decomposition matrix;
[0074] Step 5: Perform time-domain solution:
[0075] S = W′X(t);
[0076] In the time-domain solution process, the preprocessed speech frame vector X(t) is blindly separated segment by segment. The order of the output vectors is adjusted to keep the signals of each output channel consistent, and then they are sequentially connected to obtain the final time-domain voiceprint output signal, i.e., the current voiceprint identification vector matrix. This method can effectively separate the user's voiceprint signal and background music signal, and the iteration speed is relatively fast. Compared with the iteration time of the time-domain convolution system for mixed speech separation, it is better and greatly improves the separation speed. The background channel model pre-trained in this embodiment is obtained by mining and comparing a large amount of speech data. This model can accurately depict the background voiceprint features of the user while preserving the user's voiceprint features to the maximum extent. It can also remove these features during recognition and extract the inherent features of the user's voice, which can greatly improve the accuracy and efficiency of user authentication. In addition, this embodiment makes full use of the voiceprint features related to the vocal tract in the human voice. These voiceprint features do not need to be restricted by the text, so they have greater flexibility in the recognition and verification process.
[0077] S3: Based on the similarity matching between the current voiceprint identification vector and the voiceprint feature vector in the interactive voiceprint database, a matching score is obtained, and a judgment is made based on the matching score. If the matching score is greater than or equal to the threshold, the identity corresponding to the voiceprint feature in the interactive voiceprint database is queried according to the mapping relationship to identify the user's identity. If it is less than the threshold, the matching continues until the preset matching limit is reached and then stops.
[0078] In this embodiment, the interactive voiceprint database is obtained by learning from the corpus of daily conversations collected from multiple devices, and the feature values of each collected voice data are included. These feature values include spectral density, cepstral density, formants, pitch, reflection coefficient, prosody, rhythm, tempo, pitch, and volume feature vectors. Users do not need to read large amounts of training text beforehand to input voiceprint features and build a voiceprint feature database. Instead, the terminal or server can collect voice data from the instant messaging messages sent by the user daily as training samples for the user's voiceprint features, thus saving the user's time in inputting voiceprint features and improving operational convenience. The interactive voiceprint database contains the user identity corresponding to the voiceprint features. It uses the Milvus database as a vector engine for similarity matching to obtain matching scores. The Milvus database supports data partitioning and sharding, data persistence, incremental data ingestion, scalar vector hybrid query, time travel, and other functions, while significantly optimizing vector retrieval performance to meet the application needs of any vector retrieval scenario. Importantly, using the Milvus database, it can achieve performance of hundreds of millions of vectors and return results in seconds, enabling rapid voiceprint retrieval. The judgment is made based on the matching score. If the matching score is greater than or equal to the threshold, the threshold can be set to 89%. If the threshold is set too low, the security will decrease; if it is set too high, the error tolerance will be low, resulting in a low matching rate. The identity corresponding to the voiceprint features in the interactive voiceprint database is queried according to the mapping relationship to identify the user's identity. If it is less than the threshold, the matching continues until the preset maximum number of matching attempts is reached, at which point it stops. The cosine similarity function is used to calculate the similarity between the current voiceprint identification vector and the sample voiceprint feature vector of the user in the interactive voiceprint database to obtain the evaluation score. The cosine similarity function can be expressed by the following formula: Where S represents the current voiceprint identification vector, and S1 represents the sample voiceprint feature vector of the user in the interactive voiceprint database; cosine distance is a measure of the difference between two individuals using the cosine of the angle between two vectors in the vector space. An evaluation score vector is obtained for each speech segment based on the evaluation score; then, the score vector is normalized through a softmax layer; and then organized into similarity probabilities. If the preset matching limit is reached and no registered user's voiceprint feature vector is found, a voice prompt is played. The preset matching limit is set to 4 times. If no registered user's voiceprint feature vector is found during the matching process, the voice IVR interaction system plays a voice prompt, stating that there is no voiceprint information of a registered user on this machine, and asking if the user is a relative of this machine. If the speaker says yes, a question is asked to confirm whether the user's information belongs to this machine. If the question is answered correctly, the speaker's information is recorded and automatically registered. If a user fails identity verification multiple times consecutively via voice, the user's account can be locked, preventing further login until the user unlocks using other identity verification methods. Alternatively, the target user account can be locked for a certain period of time, and then unlocked after the lockout period expires, allowing the user to log in to the system. This prevents criminals from repeatedly attempting to authenticate the user by imitating their voice, thus improving security.
[0079] It should be noted that, in this document, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, apparatus, article, or method that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, apparatus, article, or method. Unless otherwise specified, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, apparatus, article, or method that includes that element.
[0080] Of course, the present invention may have many other embodiments. Based on this embodiment, other embodiments obtained by those skilled in the art without any creative effort are all within the scope of protection of the present invention.
Claims
1. A method for identifying user identity based on voiceprint in voice IVR interaction, characterized in that, include: S1: Receive mixed voice data in IVR interaction, obtain the mixed voiceprint features of the mixed voice data, and construct a corresponding mixed voiceprint feature vector based on the mixed voiceprint features; S2: Input the mixed voiceprint feature vector into a pre-trained background channel model to construct the current voiceprint identification vector in the mixed speech data; wherein, the step of inputting the mixed voiceprint feature vector into a pre-trained background channel model to construct the current voiceprint identification vector summarizing the mixed speech data includes: Step 1: Analyze the speech frame vectors in the obtained mixed speakerprint feature vector. First, preprocessing is performed, and then the preprocessed audio frames are... Average score part; Step 2: For each preprocessed speech frame vector Perform a Fast Fourier Transform; after the Fourier Transform, the speech frame vector is converted into a complex-valued speaker feature vector, expressed by the following formula: ， in, Represented as the decomposition matrix of the background channel model, This represents the current voiceprint identification vector matrix; the number of speech frame vector matrices is set to be the same as the current voiceprint identification vector matrix, i.e. ; Step 3: For each complex-valued speaker feature vector, perform linear separation of the speech frames using the frequency domain complex-valued ICA algorithm; The algorithm steps are as follows: The update iteration formula is expressed by the following formula: ， in, This represents the iterative expectation operation. To represent iterating over any non-quadratic function, express Iteratively decompose the matrix at each time step. express Iteratively decompose the matrix at each time step. express Time and The expected difference in iterations at time step. Indicates the iteration convergence step size. Indicates the number of iterations; like If the condition is met, the iteration ends; otherwise, it continues. Step 4: After iteration, decompose the background channel model into a matrix. Returning to the time domain via IFPT transformation: ， in, Represents the time-domain decomposition matrix; Step 5: Perform time-domain solution: ； In the time-domain solution process, including... preprocessed speech frame vectors By performing blind separation segment by segment, adjusting the order of the output vectors to keep the signals of each output channel consistent, and connecting them sequentially, the final time-domain voiceprint output signal, i.e., the current voiceprint identification vector matrix, can be obtained. S3: Based on the similarity matching between the current voiceprint identification vector and the voiceprint feature vector in the interactive voiceprint database, a matching score is obtained, and a judgment is made based on the matching score. If the matching score is greater than or equal to the threshold, the identity corresponding to the voiceprint feature in the interactive voiceprint database is queried according to the mapping relationship to identify the user's identity. If it is less than the threshold, the matching continues until the preset matching limit is reached and then stops.
2. The method for identifying user identity based on voiceprint in voice IVR interaction according to claim 1, characterized in that, The step of receiving mixed voice data in IVR interaction includes: Pre-emphasis is performed based on mixed voice data received during IVR interaction; The mixed speech data is framed according to the framing method using overlapping segments; Apply Hamming window and distinguish between speech frames and non-speech frames, discarding the non-speech frames.
3. The method for identifying user identity based on voiceprint in voice IVR interaction according to claim 2, characterized in that, The voice frame includes current voiceprint identification voice information and background sound information.
4. The method for identifying user identity based on voiceprint in voice IVR interaction according to claim 1, characterized in that, The step of obtaining a matching score by performing similarity matching between the current voiceprint identification vector and the voiceprint feature vector in the interactive voiceprint database includes: The cosine similarity function is used to calculate the similarity between the current voiceprint identification vector and the sample voiceprint feature vectors of users in the interactive voiceprint database to obtain an evaluation score. The evaluation score vector for each speech segment is obtained based on the evaluation score. The score vector is then normalized using a softmax layer.
5. The method for identifying user identity based on voiceprint in voice IVR interaction according to claim 1, characterized in that, After step S3, the following is included: If the preset limit for matching attempts is reached and no matching of the voiceprint feature vector of a registered person is found, a voice prompt will be broadcast.
6. The method for identifying user identity based on voiceprint in voice IVR interaction according to claim 1, characterized in that, The interactive voiceprint database is obtained by learning from the speech data of daily calls collected from multiple devices, and the feature value of each voice data collected; The interactive voiceprint database contains user identities corresponding to voiceprint features.
7. The method for identifying user identity based on voiceprint in voice IVR interaction according to claim 6, characterized in that, The eigenvalues include eigenvalues on the spectral, cepstral, formant, fundamental, reflection coefficient, prosody, rhythm, tempo, pitch, and volume eigenvectors.