Voiceprint recognition score normalization method, system and terminal fusing environmental channel features

Abstract

The present application belongs to the field of biometric identification technology, and relates to a voiceprint recognition score normalization method and system fusing environmental channel features and a terminal. The method comprises the following steps: presetting multiple acoustic environments, constructing a false claimant subqueue for each acoustic environment, and obtaining multiple false claimant subqueues; obtaining an environment feature vector and a voiceprint feature vector of a to-be-identified voice; obtaining a registered voiceprint template of a user, calculating the similarity between the voiceprint feature vector and the registered voiceprint template, and obtaining an original score; selecting or fusing at least one false claimant subqueue from the multiple false claimant subqueues based on the environment feature vector, and calculating a normalization parameter; and performing normalization processing on the original score by using the normalization parameter, and obtaining a final score. The present application constructs a subqueue library containing multiple environmental features, and can adaptively adjust the normalization parameter according to the environmental features of the input voice, thereby greatly improving the voiceprint recognition accuracy in cross-domain and complex noise scenarios.

Description

Technical Field

[0001] This invention relates to the field of biometric recognition technology, and in particular to a method, system, and terminal for normalizing voiceprint recognition scores by integrating environmental channel features. Background Technology

[0002] Speaker verification is a biometric technology that identifies a speaker through their voice signals. In practical applications, score normalization is typically required to map the output score of the verification system to a uniform range and determine whether verification is accepted. Current mainstream normalization methods (such as Z-Norm, T-Norm, and AS-Norm) all rely on a pre-built "imposter cohort," which is used to calculate the mean and variance of the scores and correct the original scores.

[0003] However, these methods have the following shortcomings: First, there is a domain mismatch. Existing imposter queues are usually built based on high-quality, noise-free "clean" speech data. However, in real-world application scenarios (such as in-vehicle environments, noisy streets, and far-field home environments), the user's input speech often carries strong background noise or special channel characteristics (such as reverberation). Second, the normalization parameters are distorted. When there is a significant difference between the acoustic environment of the input speech and the acoustic environment of the imposter queue, the calculated statistical parameters (mean and variance) cannot accurately reflect the score distribution in the current environment. Third, the recognition capability decreases. This mismatch in environmental characteristics leads to deviations in the normalized scores, resulting in a significant increase in the false recognition rate (FAR) and false rejection rate (FRR) in complex scenarios, which seriously affects the robustness of the voiceprint recognition system.

[0004] Therefore, there is an urgent need to propose a voiceprint recognition score normalization scheme with high recognition accuracy in complex environments. Summary of the Invention

[0005] The purpose of this invention is to address the shortcomings of existing technologies by proposing a voiceprint recognition score normalization method, system, and terminal that integrates environmental channel features. This invention proposes a dynamic normalization library matching mechanism based on "multi-modal / multi-scenario" characteristics. It breaks away from the traditional limitation of a single, static, pure imposter queue, constructs a sub-queue library containing multiple environmental features, and adaptively adjusts the normalization parameters based on the environmental features of the input speech during the inference phase. This significantly improves the accuracy of voiceprint recognition in cross-domain and complex noise scenarios.

[0006] To achieve the above objectives, the present invention adopts the following technical solution: In a first aspect, the present invention provides a method for normalizing voiceprint recognition scores by incorporating environmental channel features, comprising the following steps: S1. Preset multiple acoustic environments, construct an imposter subqueue for each acoustic environment, and obtain multiple imposter subqueues; S2. Obtain the speech to be recognized, extract the environmental features and voiceprint features of the speech to be recognized, and obtain the environmental feature vector and the voiceprint feature vector to be tested; S3. Obtain the user's registered voiceprint template, calculate the similarity between the voiceprint feature vector to be tested and the registered voiceprint template, and obtain the original score; S4. Based on the feature vector of the environment under test, select or merge at least one imposter subqueue from multiple imposter subqueues and calculate the normalization parameter; S5. Normalize the original score using the normalization parameter to obtain the final score.

[0007] As one possible implementation, S4 includes: The acoustic environment to which the speech to be identified belongs is determined based on the feature vector of the environment to be tested, and the imposter subqueue corresponding to the acoustic environment is selected. Calculate the similarity value between the voiceprint feature vector to be tested and each sample in the imposter subqueue; Sort the similarity values ​​from high to low, select the top N similarity scores, and calculate their mean and standard deviation, which are the normalization parameters.

[0008] As one possible implementation, S4 includes: Obtain the environment center vector for each imposter subqueue; Calculate the similarity value between the feature vector of the environment to be tested and the center vector of each environment, and normalize each similarity value to the weight coefficient of its corresponding imposter subqueue; For each imposter sub-queue, the similarity value between the voiceprint feature vector to be tested and each sample is calculated to obtain the local mean and local standard deviation of each imposter sub-queue. The fusion mean is obtained by weighting and summing the local means of each imposter sub-queue based on the weighting coefficients; the fusion standard deviation is obtained by weighting and summing the local standard deviations of each imposter sub-queue based on the weighting coefficients; the fusion mean and fusion standard deviation are the normalization parameters.

[0009] As one possible implementation, the environment center vector of each imposter subqueue is obtained using the following method: Extract the environmental features of all speech samples in each imposter subqueue, obtain the environmental feature vector of each speech sample, and calculate the average value of each environmental feature vector, which is the environmental center vector.

[0010] As one possible implementation, an imposter subqueue is constructed for each acoustic environment using the following method: S10. Obtain a baseline set of imposter voices, wherein the baseline set of imposter voices includes clean voices from a preset number of people; S11. For each acoustic environment, perform convolutional superposition on the baseline imposter speech set based on its noise data to obtain the imposter subqueue corresponding to that acoustic environment.

[0011] As one possible approach, a pre-trained signal-to-noise ratio estimator or a pre-trained scene classification neural network can be used to extract environmental features of the speech to be recognized.

[0012] As one possible implementation, S5 is as follows: in, This indicates the final score. This represents the raw score. , Here, are the normalization parameters, representing the mean and standard deviation, respectively.

[0013] Secondly, embodiments of the present invention provide a voiceprint recognition score normalization system that integrates environmental channel features, comprising: The multi-scenario queue construction unit presets multiple acoustic environments and constructs an imposter sub-queue for each acoustic environment, resulting in multiple imposter sub-queues. The feature extraction unit acquires the speech to be recognized, extracts the environmental features and voiceprint features of the speech to be recognized, and obtains the environmental feature vector and the voiceprint feature vector to be tested. The raw score calculation unit obtains the user's registered voiceprint template, calculates the similarity between the voiceprint feature vector to be tested and the registered voiceprint template, and obtains the raw score. The normalization parameter generation unit selects or merges at least one imposter subqueue from multiple imposter subqueues based on the feature vector of the environment under test, and calculates the normalization parameter. And a normalization unit, which uses normalization parameters to normalize the original score to obtain the final score.

[0014] As one possible implementation, the normalization parameter generation unit includes: The hard matching subunit determines the acoustic environment to which the speech to be recognized belongs based on the feature vector of the environment to be tested. It selects the imposter subqueue corresponding to the acoustic environment, calculates the similarity value between the feature vector of the voiceprint to be tested and each sample in the imposter subqueue, sorts the similarity values ​​from high to low, selects the top N similarity scores, and calculates their mean and standard deviation to obtain the normalization parameters. And / or, the soft interpolation subunit obtains the environment center vector of each imposter sub-queue, calculates the similarity value between the test environment feature vector and each environment center vector, normalizes each similarity value to the weight coefficient of its corresponding imposter sub-queue, calculates the similarity value between the test voiceprint feature vector and each sample for each imposter sub-queue, obtains the local mean and local standard deviation of each imposter sub-queue, weights and sums the local means of each imposter sub-queue based on the weight coefficient to obtain the fusion mean; weights and sums the local standard deviations of each imposter sub-queue based on the weight coefficient to obtain the fusion standard deviation; and uses the fusion mean and fusion standard deviation as normalization parameters.

[0015] Thirdly, the present invention provides a terminal including a processor and a communication interface coupled to the processor, the processor being used to run computer programs or instructions to implement the voiceprint recognition score normalization method based on fused environmental channel features provided in the first aspect.

[0016] Compared with the prior art, the beneficial effects of the present invention are as follows: 1. The present invention proposes a speaker recognition score normalization method, system, and terminal that integrates environmental channel features. This method pre-constructs imposter sub-queues covering multiple acoustic environments and, during the inference phase, perceives the environmental features of the input speech in real time, adaptively selecting or fusing the most matching normalization parameters. Compared to traditional methods that use a single "clean" queue, this invention effectively eliminates the training-test environment mismatch problem.

[0017] 2. The voiceprint recognition score normalization method, system, and terminal proposed in this invention, which integrates environmental channel features, is a back-end score processing method, independent of the front-end voiceprint feature extraction. When expanding to new application scenarios, there is no need to retrain or adjust the main voiceprint extraction network; only offline supplementation of the imposter sub-queue for the corresponding scenario is required to adapt the system to the new environment, demonstrating good scalability.

[0018] 3. This invention proposes a voiceprint recognition score normalization method, system, and terminal that integrates environmental channel features. Through a soft interpolation weighted fusion mechanism, it can generate smooth normalized parameters for continuously changing or mixed environments (such as a vehicle gradually accelerating or a window partially open), avoiding parameter jumps that may be caused by hard matching. This allows the system to output stable scores even during gradual environmental changes, reducing the number of repeated verifications required for legitimate users in complex environments and improving the smoothness of interaction.

[0019] 4. This invention proposes a voiceprint recognition score normalization method, system, and terminal that integrates environmental channel features. While ensuring a reduced user rejection rate under high noise conditions, it does not sacrifice the ability to defend against real attacks, thus improving the security and convenience of the application. Attached Figure Description

[0020] The accompanying drawings, which are included to provide a further understanding of the invention and form part of this invention, illustrate exemplary embodiments of the invention and are used to explain the invention, but do not constitute an undue limitation of the invention. In the drawings: Figure 1 , Figure 2 A flowchart of the voiceprint recognition score normalization method based on the fusion of environmental channel features provided in this embodiment of the invention; Figure 3 , Figure 4 This is a flowchart for calculating the normalized parameters in an embodiment of the present invention. Detailed Implementation

[0021] To facilitate a clear description of the technical solutions in the embodiments of the present invention, the terms "first" and "second" are used to distinguish identical or similar items with essentially the same function and effect. For example, the first threshold and the second threshold are merely used to distinguish different thresholds and do not limit their order. Those skilled in the art will understand that the terms "first" and "second" do not limit the quantity or execution order, and that the terms "first" and "second" are not necessarily different.

[0022] It should be noted that in this invention, the terms "exemplary" or "for example" are used to indicate examples, illustrations, or descriptions. Any embodiment or design described as "exemplary" or "for example" in this invention should not be construed as being more preferred or advantageous than other embodiments or designs. Specifically, the use of terms such as "exemplary" or "for example" is intended to present the relevant concepts in a concrete manner.

[0023] In this invention, "at least one" refers to one or more, and "more than one" refers to two or more. "And / or" describes the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A alone, A and B simultaneously, or B alone, where A and B can be singular or plural. The character " / " generally indicates that the preceding and following related objects are in an "or" relationship. "At least one" or similar expressions refer to any combination of these items, including any combination of singular or plural items. For example, "at least one of a, b, or c" can represent: a, b, c, a combination of a and b, a combination of a and c, a combination of b and c, or a, b, and c, where a, b, and c can be single or multiple.

[0024] The present invention aims to provide a method, system, and terminal for normalizing voiceprint recognition scores by integrating environmental channel features. It breaks through the limitations of the traditional single, static pure imposter queue, constructs a sub-queue library containing multiple environmental features, and adaptively adjusts the normalization parameters according to the environmental features of the input speech during the inference stage, which greatly improves the accuracy of voiceprint recognition in cross-domain and complex noise scenarios.

[0025] In a first aspect, embodiments of the present invention provide a method for normalizing voiceprint recognition scores by fusing environmental channel features, see [link to previous document]. Figure 1 and Figure 2 It includes the following steps: S1. Preset multiple acoustic environments, construct an imposter subqueue for each acoustic environment, and obtain multiple imposter subqueues; Traditional methods typically use imposter queues built from high-quality, noise-free "clean" speech data. However, when the actual input speech contains complex environmental features such as noise and reverberation, the statistical parameters (mean and variance) calculated from this "clean" imposter queue will mismatch with the true score distribution of the current environment, leading to normalization failure. This invention breaks away from this single, static imposter queue by constructing multiple imposter sub-queues corresponding to different acoustic environments (e.g., in-vehicle noise environment, street noise environment, quiet indoor environment, etc.). This allows for the selection of the most suitable imposter sub-queue for normalization based on the actual environment of the current input speech, thus solving the environment mismatch problem.

[0026] As one possible implementation, an imposter subqueue is constructed for each acoustic environment using the following method: S10. Obtain a baseline set of imposter voices, wherein the baseline set of imposter voices includes clean voices from a preset number of people; As an example, noise-free, high-quality, clean audio recordings of 1000 non-target speakers were selected, with each speaker recording approximately 5 seconds long, as the baseline imposter audio.

[0027] S11. For each acoustic environment, perform convolutional superposition on the baseline imposter speech set based on its noise data to obtain the imposter subqueue corresponding to that acoustic environment.

[0028] As an example, K typical acoustic environments are preset, such as a quiet indoor environment, a high-speed car ride, and a noisy street. The baseline imposter speech set is convolved and superimposed under the three different acoustic environments, as shown in Table 1 below: Table 1. Imposter subqueues corresponding to different acoustic environments After processing, each imposter subqueue still contains the voices of 1,000 people, but with different environmental noises superimposed on them.

[0029] S2. Obtain the speech to be recognized, extract the environmental features and voiceprint features of the speech to be recognized, and obtain the environmental feature vector and the voiceprint feature vector to be tested; As one possible approach, a pre-trained signal-to-noise ratio estimator or a pre-trained scene classification neural network can be used to extract environmental features of the speech to be recognized.

[0030] As an example, residual networks (ResNet), such as ResNet34 and ResNet50, can be used to extract voiceprint features. Scene classification neural networks, such as audio scene classifiers based on convolutional neural networks (CNNs), can also be used. Taking the three scenarios mentioned above as examples, assume the pre-trained scene classification neural network outputs a three-dimensional probability vector representing the probability of belonging to three acoustic environments: [0.1, 0.8, 0.1]. That is, the probability of belonging to a quiet indoor environment is 10%, the probability of belonging to a high-speed driving vehicle environment is 80%, and the probability of belonging to a noisy street environment is 10%.

[0031] S3. Obtain the user's registered voiceprint template, calculate the similarity between the voiceprint feature vector to be tested and the registered voiceprint template, and obtain the original score; As an example, a user speaks a fixed or free text in a relatively quiet environment, typically lasting a few seconds to a dozen seconds. The residual network (such as ResNet34) is then used to analyze this speech, generating a fixed-dimensional voiceprint feature vector. This feature vector is then used as the user's registered voiceprint template. For example, the cosine similarity between the test voiceprint feature vector and the registered voiceprint template is calculated. Assume the obtained original score is... =0.75.

[0032] S4. Based on the feature vector of the environment under test, select or merge at least one imposter subqueue from multiple imposter subqueues and calculate the normalization parameter; As one possible implementation, see Figure 3 S4 includes: The acoustic environment to which the speech to be identified belongs is determined based on the feature vector of the environment to be tested, and the imposter subqueue corresponding to the acoustic environment is selected. Calculate the similarity value between the voiceprint feature vector to be tested and each sample in the imposter subqueue; Sort the similarity values ​​from high to low, select the top N similarity scores, and calculate their mean and standard deviation, which are the normalization parameters.

[0033] As an example, since the aforementioned feature vector of the test environment indicates the highest probability (80%) of belonging to a high-speed vehicle driving environment, the imposter sub-queue corresponding to the high-speed vehicle driving environment is directly selected, namely the second imposter sub-queue in Table 1. The cosine similarity value between the voiceprint feature vector of the test and each sample in the second imposter sub-queue is calculated, resulting in 1000 scores. These 1000 scores are sorted from highest to lowest, and the top 100 similarity scores are selected. Their mean and standard deviation are calculated, assuming the results are 0.25 and 0.12, respectively.

[0034] As another possible implementation, see Figure 4 S4 includes: Obtain the environment center vector for each imposter subqueue; As one possible implementation, the environment center vector of each imposter subqueue is obtained using the following method: Extract the environmental features of all speech samples in each imposter subqueue, obtain the environmental feature vector of each speech sample, and calculate the average value of each environmental feature vector, which is the environmental center vector.

[0035] As an example, referring to Table 1, the third imposter subqueue contains 1000 audio segments. For each audio segment, an environmental feature vector is extracted using the aforementioned environmental feature extraction network, exemplarily represented as follows: , This represents the first of 1000 speech segments. This yielded a segment of speech. Consequently, 1000 environmental feature vectors were obtained: , ,…., The average of these 1000 environmental feature vectors is used to obtain the environmental center vector of the third imposter sub-queue. : Similarly, we can obtain the environment center vectors of the other two imposter subqueues.

[0036] Calculate the similarity value between the feature vector of the environment to be tested and the center vector of each environment, and normalize each similarity value to the weight coefficient of its corresponding imposter subqueue; For each imposter sub-queue, the similarity value between the voiceprint feature vector to be tested and each sample is calculated to obtain the local mean and local standard deviation of each imposter sub-queue. The fusion mean is obtained by weighting and summing the local means of each imposter sub-queue based on the weighting coefficients; the fusion standard deviation is obtained by weighting and summing the local standard deviations of each imposter sub-queue based on the weighting coefficients; the fusion mean and fusion standard deviation are the normalization parameters.

[0037] As an example, the similarity value (e.g., cosine similarity) between the feature vector of the environment to be tested and each environment center vector is calculated, and the similarity is normalized using Softmax to obtain the weights. Assume the weight coefficients calculated with the three environment center vectors are 0.2, 0.7, and 0.1, respectively. Taking the second imposter sub-queue as an example again, the similarity value between the voiceprint feature vector to be tested and each sample is calculated, resulting in 1000 similarity values. The mean and standard deviation of these 1000 similarity values ​​are calculated, which are their local mean and local standard deviation, denoted as […]. , Similarly, the local mean and local standard deviation of the first imposter subqueue can be obtained. , And the local mean and local standard deviation of the third imposter subqueue. , The fusion mean is calculated based on the weighting coefficients: ; Calculate the fusion standard deviation: .

[0038] This invention utilizes a soft interpolation weighted fusion mechanism to generate smooth, normalized parameters for continuously changing or mixed environments (such as a vehicle gradually accelerating or a window partially open), avoiding parameter jumps that may result from hard matching. This allows the system to output stable scores even during gradual environmental changes, reducing the number of times legitimate users need to repeatedly verify their credentials in complex environments and improving the smoothness of interaction.

[0039] S5. Normalize the original score using the normalization parameter to obtain the final score.

[0040] As another possible implementation, S5 is as follows: in, This indicates the final score. This represents the raw score. , Here, are the normalization parameters, representing the mean and standard deviation, respectively.

[0041] Taking the aforementioned mean of 0.25 and standard deviation of 0.12 as an example, the raw score =0.75, then the final score For example, if a recognition threshold is preset, say 3, then 4.17 is greater than 3, and the speech to be recognized can be determined to belong to the user.

[0042] Secondly, embodiments of the present invention provide a voiceprint recognition score normalization system that integrates environmental channel features, comprising: The multi-scenario queue construction unit presets multiple acoustic environments and constructs an imposter sub-queue for each acoustic environment, resulting in multiple imposter sub-queues. The feature extraction unit acquires the speech to be recognized, extracts the environmental features and voiceprint features of the speech to be recognized, and obtains the environmental feature vector and the voiceprint feature vector to be tested. The raw score calculation unit obtains the user's registered voiceprint template, calculates the similarity between the voiceprint feature vector to be tested and the registered voiceprint template, and obtains the raw score. The normalization parameter generation unit selects or merges at least one imposter subqueue from multiple imposter subqueues based on the feature vector of the environment under test, and calculates the normalization parameter. And a normalization unit, which uses normalization parameters to normalize the original score to obtain the final score.

[0043] As one possible implementation, the normalization parameter generation unit includes: The hard matching subunit determines the acoustic environment to which the speech to be recognized belongs based on the feature vector of the environment to be tested. It selects the imposter subqueue corresponding to the acoustic environment, calculates the similarity value between the feature vector of the voiceprint to be tested and each sample in the imposter subqueue, sorts the similarity values ​​from high to low, selects the top N similarity scores, and calculates their mean and standard deviation to obtain the normalization parameters. And / or, the soft interpolation subunit obtains the environment center vector of each imposter sub-queue, calculates the similarity value between the test environment feature vector and each environment center vector, normalizes each similarity value to the weight coefficient of its corresponding imposter sub-queue, calculates the similarity value between the test voiceprint feature vector and each sample for each imposter sub-queue, obtains the local mean and local standard deviation of each imposter sub-queue, weights and sums the local means of each imposter sub-queue based on the weight coefficient to obtain the fusion mean; weights and sums the local standard deviations of each imposter sub-queue based on the weight coefficient to obtain the fusion standard deviation; and uses the fusion mean and fusion standard deviation as normalization parameters.

[0044] Thirdly, the present invention provides a terminal including a processor and a communication interface coupled to the processor, the processor being used to run computer programs or instructions to implement the voiceprint recognition score normalization method based on fused environmental channel features provided in the first aspect.

[0045] This invention focuses on backend score processing, independent of frontend voiceprint feature extraction. When expanding to new application scenarios, there is no need to retrain or adjust the main voiceprint extraction network; simply supplementing the imposter subqueue for the corresponding scenario offline allows the system to adapt to the new environment, demonstrating excellent scalability.

[0046] Although the invention has been described herein in conjunction with various embodiments, those skilled in the art will understand and implement other variations of the disclosed embodiments by reviewing the accompanying drawings, the disclosure, and the description of the drawings, in carrying out the claimed invention. In this specification, the word "comprising" does not exclude other components or steps, and "a" or "an" does not exclude multiple components. A single processor or other unit can implement several of the functions listed in the specification. While certain measures are described in different embodiments, this does not mean that these measures cannot be combined to produce good results.

[0047] Although the invention has been described in conjunction with specific features and embodiments, it is obvious that various modifications and combinations can be made therein without departing from the spirit and scope of the invention. Accordingly, this specification and drawings are merely illustrative of the invention and are considered to cover any and all modifications, variations, combinations, or equivalents within the scope of the invention. Clearly, those skilled in the art can make various alterations and modifications to the invention without departing from its spirit and scope. Thus, if such modifications and modifications fall within the scope of the invention and its equivalents, the invention is also intended to include such modifications and modifications.

Claims

1. A method for normalizing voiceprint recognition scores by integrating environmental channel features, characterized in that, Includes the following steps: S1. Preset multiple acoustic environments, construct an imposter subqueue for each acoustic environment, and obtain multiple imposter subqueues; S2. Obtain the speech to be recognized, extract the environmental features and voiceprint features of the speech to be recognized, and obtain the environmental feature vector and the voiceprint feature vector to be tested; S3. Obtain the user's registered voiceprint template, calculate the similarity between the voiceprint feature vector to be tested and the registered voiceprint template, and obtain the original score; S4. Based on the feature vector of the environment under test, select or merge at least one imposter subqueue from multiple imposter subqueues and calculate the normalization parameter; S5. Normalize the original score using the normalization parameter to obtain the final score.

2. The voiceprint recognition score normalization method based on fusion of environmental channel features according to claim 1, characterized in that, S4 includes: The acoustic environment to which the speech to be identified belongs is determined based on the feature vector of the environment to be tested, and the imposter subqueue corresponding to the acoustic environment is selected. Calculate the similarity value between the voiceprint feature vector to be tested and each sample in the imposter subqueue; Sort the similarity values ​​from high to low, select the top N similarity scores, and calculate their mean and standard deviation, which are the normalization parameters.

3. The voiceprint recognition score normalization method based on fusion of environmental channel features according to claim 1, characterized in that, S4 includes: Obtain the environment center vector for each imposter subqueue; Calculate the similarity value between the feature vector of the environment to be tested and the center vector of each environment, and normalize each similarity value to the weight coefficient of its corresponding imposter subqueue; For each imposter sub-queue, the similarity value between the voiceprint feature vector to be tested and each sample is calculated to obtain the local mean and local standard deviation of each imposter sub-queue. The fusion mean is obtained by weighting and summing the local means of each imposter sub-queue based on the weighting coefficients; the fusion standard deviation is obtained by weighting and summing the local standard deviations of each imposter sub-queue based on the weighting coefficients; the fusion mean and fusion standard deviation are the normalization parameters.

4. The voiceprint recognition score normalization method based on fusion of environmental channel features according to claim 3, characterized in that, The environment center vector for each imposter subqueue is obtained using the following method: Extract the environmental features of all speech samples in each imposter subqueue, obtain the environmental feature vector of each speech sample, and calculate the average value of each environmental feature vector, which is the environmental center vector.

5. The voiceprint recognition score normalization method based on fusion of environmental channel features according to claim 1, characterized in that, An imposter subqueue is constructed for each acoustic environment using the following method: S10. Obtain a baseline set of imposter voices, wherein the baseline set of imposter voices includes clean voices from a preset number of people; S11. For each acoustic environment, perform convolutional superposition on the baseline imposter speech set based on its noise data to obtain the imposter subqueue corresponding to that acoustic environment.

6. The voiceprint recognition score normalization method based on fusion of environmental channel features according to claim 1, characterized in that, Environmental features of the speech to be recognized are extracted using a pre-trained signal-to-noise ratio estimator or a pre-trained scene classification neural network.

7. The voiceprint recognition score normalization method based on fusion of environmental channel features according to claim 1, characterized in that, Specifically, S6 is: in, This indicates the final score. This represents the raw score. , Here, are the normalization parameters, representing the mean and standard deviation, respectively.

8. A voiceprint recognition score normalization system that integrates environmental channel features, characterized in that, include: The multi-scenario queue construction unit presets multiple acoustic environments and constructs an imposter sub-queue for each acoustic environment, resulting in multiple imposter sub-queues. The feature extraction unit acquires the speech to be recognized, extracts the environmental features and voiceprint features of the speech to be recognized, and obtains the environmental feature vector and the voiceprint feature vector to be tested. The raw score calculation unit obtains the user's registered voiceprint template, calculates the similarity between the voiceprint feature vector to be tested and the registered voiceprint template, and obtains the raw score. The normalization parameter generation unit selects or merges at least one imposter subqueue from multiple imposter subqueues based on the feature vector of the environment under test, and calculates the normalization parameter. And a normalization unit, which uses normalization parameters to normalize the original score to obtain the final score.

9. The voiceprint recognition score normalization system integrating environmental channel features according to claim 8, characterized in that, The normalization parameter generation unit includes: The hard matching subunit determines the acoustic environment to which the speech to be identified belongs based on the feature vector of the environment to be tested. It selects the imposter subqueue corresponding to the acoustic environment, calculates the similarity value between the feature vector of the voiceprint to be tested and each sample in the imposter subqueue, sorts the similarity values ​​from high to low, selects the top N similarity scores, and calculates their mean and standard deviation to obtain the normalization parameters. And / or, the soft interpolation subunit obtains the environment center vector of each imposter sub-queue, calculates the similarity value between the test environment feature vector and each environment center vector, normalizes each similarity value to the weight coefficient of its corresponding imposter sub-queue, calculates the similarity value between the test voiceprint feature vector and each sample for each imposter sub-queue, obtains the local mean and local standard deviation of each imposter sub-queue, weights and sums the local means of each imposter sub-queue based on the weight coefficient to obtain the fusion mean; weights and sums the local standard deviations of each imposter sub-queue based on the weight coefficient to obtain the fusion standard deviation; and uses the fusion mean and fusion standard deviation as normalization parameters.

10. A terminal, characterized in that, The device includes a processor and a communication interface coupled to the processor, the processor being used to run computer programs or instructions to implement the voiceprint recognition score normalization method based on the fusion of environmental channel features as described in any one of claims 1 to 7.

Images