Voiceprint recognition method and device, computer readable storage medium, and terminal

Abstract

A voiceprint recognition method and device, a computer readable storage medium and a terminal are provided. The voiceprint recognition method comprises: obtaining a to-be-recognized voiceprint feature vector of a test object and registered voiceprint feature vectors of N registered objects in a voiceprint library; performing feature separation on the to-be-recognized voiceprint feature vector and the N registered voiceprint feature vectors respectively to obtain an identity feature vector and a non-identity feature vector of the test object and identity feature vectors and non-identity feature vectors of the N registered objects; calculating the similarity between the identity feature vector of the test object and the identity feature vectors of the N registered objects to obtain N original similarities; calculating a compensation coefficient according to the similarity between the non-identity feature vector of the test object and the non-identity feature vectors of the registered objects; compensating the original similarities using the compensation coefficient to obtain compensated similarities; and obtaining a voiceprint recognition result of the test object based on the compensated similarities. The above scheme can improve the accuracy of voiceprint recognition.

Description

Technical Field

[0001] The present invention relates to the field of voiceprint recognition technology, and in particular to a voiceprint recognition method and apparatus, a computer-readable storage medium, and a terminal. Background Technology

[0002] Voiceprint recognition is a technology that extracts speaker identity features from a speaker's speech and performs identity verification. It has been applied in various fields such as finance and securities, autonomous driving, and smart homes, bringing greater convenience and security to people's lives and work. Voiceprint recognition typically uses machine learning and deep learning-based feature extraction methods to extract voiceprint features. These extracted features are then compared with the voiceprint features of registered individuals in a voiceprint database. Based on the comparison result and a set decision threshold, the voiceprint recognition result is obtained.

[0003] However, due to environmental differences in real-world scenarios, the accuracy of voiceprint recognition results is relatively low. Summary of the Invention

[0004] The technical problem solved by the embodiments of the present invention is how to improve the accuracy of voiceprint recognition.

[0005] To address the aforementioned technical problems, this invention provides a voiceprint recognition method, comprising: acquiring a voiceprint feature vector to be identified from a test object and registered voiceprint feature vectors from N registered objects in a voiceprint database, where N is a positive integer; performing feature separation on the voiceprint feature vector to be identified and the N registered voiceprint feature vectors respectively to obtain the identity feature vector and non-identity feature vector of the test object, and the identity feature vector and non-identity feature vector of the N registered objects; calculating the similarity between the identity feature vector of the test object and the identity feature vectors of the N registered objects to obtain N original similarities; calculating a compensation coefficient based on the similarity between the non-identity feature vector of the test object and the non-identity feature vector of the registered objects; compensating the original similarities using the compensation coefficient to obtain compensated similarities; and obtaining the voiceprint recognition result of the test object based on the compensated similarities.

[0006] Optionally, the step of calculating the compensation coefficient based on the similarity between the non-identity feature vector of the test object and the non-identity feature vector of the registered object includes: selecting original similarities that satisfy the compensation conditions from N original similarities; for each original similarity that satisfies the compensation conditions, calculating the compensation coefficient corresponding to each original similarity that satisfies the compensation conditions based on the similarity between the non-identity feature vector of the registered object corresponding to the original similarity that satisfies the compensation conditions and the non-identity feature vector of the test object; wherein, the compensation conditions include any of the following conditions: the original similarity is within a preset confidence interval; the number of original similarities greater than or equal to the confidence threshold is multiple, and the difference between the multiple original similarities is within a preset difference range.

[0007] Optionally, the compensation coefficient may be calculated using any of the following similarity calculation algorithms: cosine similarity calculation algorithm; Euclidean distance-based similarity calculation algorithm; or neural network model-based similarity calculation algorithm.

[0008] Optionally, the step of using the compensation coefficient to compensate the original similarity to obtain the compensated similarity includes: calculating the upward fluctuation of the upper limit of the confidence interval relative to the confidence threshold, and the downward fluctuation of the lower limit of the confidence interval relative to the confidence threshold; calculating the compensation amount based on the compensation coefficient, the upward fluctuation, and the downward fluctuation; and using the compensation amount to compensate the original similarity to obtain the compensated similarity.

[0009] Optionally, the compensated similarity can be calculated using the following formula: score final =score original +Δscore;Δscore=μ*(lower+upper); where, score final The score represents the compensated similarity. original The original similarity is denoted as Δscore, the compensation amount is Δscore, μ is the compensation coefficient, lower is the downward floating amount, and upper is the upward floating amount.

[0010] Optionally, the step of performing feature separation on the voiceprint feature vector to be identified and the N registered voiceprint feature vectors includes: based on the probabilistic linear discriminant analysis algorithm, using the probabilistic linear discriminant analysis transformation matrix to perform feature separation on the voiceprint feature vector to be identified and the N registered voiceprint feature vectors respectively.

[0011] Optionally, obtaining the voiceprint recognition result of the test object based on the compensated similarity includes: obtaining the voiceprint recognition result based on the relationship between the largest compensated similarity and the confidence threshold; or, if there are multiple largest compensated similarities, obtaining the voiceprint recognition result based on the relationship between the maximum value among the original similarities corresponding to the multiple largest compensated similarities and the confidence threshold.

[0012] This invention also provides a voiceprint recognition device, comprising: an acquisition unit for acquiring a voiceprint feature vector to be identified of a test object and registered voiceprint feature vectors of N registered objects in a voiceprint database, where N is a positive integer; a feature separation unit for performing feature separation on the voiceprint feature vector to be identified and the N registered voiceprint feature vectors respectively, to obtain an identity feature vector and a non-identity feature vector of the test object, and an identity feature vector and a non-identity feature vector of the N registered objects; a similarity calculation unit for calculating the similarity between the identity feature vector of the test object and the identity feature vectors of the N registered objects, to obtain N original similarities; a compensation coefficient calculation unit for calculating a compensation coefficient based on the similarity between the non-identity feature vector of the test object and the non-identity feature vector of the registered objects; a compensation unit for compensating the original similarities using the compensation coefficient, to obtain a compensated similarity; and a voiceprint recognition result determination unit for obtaining the voiceprint recognition result of the test object based on the compensated similarity.

[0013] This invention also provides a computer-readable storage medium storing a computer program thereon, wherein the computer program, when run by a processor, executes the steps of any of the above-described voiceprint recognition methods.

[0014] This invention also provides a terminal, including a memory and a processor, wherein the memory stores a computer program that can run on the processor, and the processor executes the steps of any of the above-described voiceprint recognition methods when running the computer program.

[0015] Compared with the prior art, the technical solution of the embodiments of the present invention has the following beneficial effects:

[0016] In this embodiment of the invention, feature separation is performed on the voiceprint vector to be identified of the test object and the registered voiceprint feature vectors of N registered objects to obtain the identity feature vector and non-identity feature vector of the test object, and the identity feature vector and non-identity feature vector of the N registered objects. Based on the similarity between the identity feature vector of the test object and the identity feature vectors of the N registered objects, N original similarity scores are obtained. Since the compensation coefficient is calculated based on the similarity between the non-identity feature vector of the test object and the non-identity feature vector of the registered objects, the compensation coefficient can characterize the difference between the voiceprint feature vector to be identified of the test object and the registered voiceprint feature vectors of the N registered objects in non-identity features. Therefore, by using the compensation coefficient to compensate for the original similarity scores, the influence of non-identity feature vectors can be effectively reduced, the confidence of the compensated similarity scores can be improved, and thus the accuracy of voiceprint recognition can be improved.

[0017] Furthermore, from N original similarities, original similarities that meet the compensation conditions are selected. For each original similarity that meets the compensation conditions, a compensation coefficient is calculated based on the similarity between the non-identity feature vector of the registered object and the non-identity feature vector of the test object. The compensation conditions include any of the following: the original similarity is within a preset confidence interval; the number of original similarities greater than or equal to a confidence threshold is multiple, and the difference between multiple original similarities is within a preset difference range. Compared to the existing method of directly determining the recognition result using a confidence threshold, this embodiment of the invention can compensate for some original similarities with uncertain results by setting compensation conditions. This can improve the confidence of the compensated similarity, allowing the compensated similarity to better differentiate itself and effectively eliminate the influence of factors such as channel information and environmental differences. This ensures that the compensated similarity can better reflect the true similarity between the test object and the registered object. Consequently, when determining the voiceprint recognition result based on the compensated similarity, the rationality of the decision can be improved, and the accuracy of voiceprint recognition of the test object can be increased. Attached Figure Description

[0018] Figure 1 This is a flowchart of a voiceprint recognition method according to an embodiment of the present invention;

[0019] Figure 2 yes Figure 1 A flowchart of a specific implementation of step 15 in the process;

[0020] Figure 3 This is a flowchart of another voiceprint recognition method in an embodiment of the present invention;

[0021] Figure 4 This is a schematic diagram of the structure of a voiceprint recognition device according to an embodiment of the present invention. Detailed Implementation

[0022] As mentioned above, voiceprint recognition involves two key steps: voiceprint feature extraction and voiceprint feature recognition. The goal of voiceprint feature recognition is to match the voiceprint feature vector or embedding vector of the test speech with the registered voiceprint feature vectors or embedding vectors in the voiceprint database to obtain a similarity score. Based on the relationship between the obtained similarity score and a set decision threshold, the voiceprint recognition result is obtained. However, due to the influence of environmental differences in real-world scenarios, the accuracy of voiceprint recognition obtained using this method is relatively low.

[0023] In current methods, to improve the accuracy of voiceprint recognition, voiceprint decision thresholds are set based on application scenarios. By determining the application scenario of the test speech, the corresponding voiceprint threshold for that scenario is selected. Generally, thresholds for different application scenarios are obtained during the training phase. However, voiceprint recognition is a problem involving unknown categories, making it difficult to cover all application scenarios during training. Furthermore, different standards exist for classifying application scenarios. Therefore, in practical applications, the influence of application scenarios obtained from different standards can affect the accuracy of voiceprint recognition. Consequently, the problem of relatively low accuracy in voiceprint recognition persists.

[0024] To address the aforementioned issues, in this embodiment of the invention, feature separation is performed on the voiceprint vector to be identified of the test object and the registered voiceprint feature vectors of N registered objects, respectively, to obtain the identity feature vector and non-identity feature vector of the test object, and the identity feature vector and non-identity feature vector of the N registered objects. Based on the similarity between the identity feature vector of the test object and the identity feature vectors of the N registered objects, N original similarity scores are obtained. Since the compensation coefficient is calculated based on the similarity between the non-identity feature vector of the test object and the non-identity feature vector of the registered objects, the compensation coefficient can characterize the difference between the voiceprint feature vector to be identified of the test object and the registered voiceprint feature vectors of the N registered objects in non-identity features. Therefore, by using the compensation coefficient to compensate for the original similarity scores, the influence of non-identity feature vectors can be effectively reduced, the confidence of the compensated similarity scores can be improved, and thus the accuracy of voiceprint recognition can be improved.

[0025] To make the above-mentioned objectives, features and beneficial effects of the embodiments of the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0026] This invention provides a voiceprint recognition method, which can be executed by a terminal, a chip or chip module in the terminal with voiceprint recognition function, a chip or chip module in the terminal with data processing function, or a baseband chip in the terminal.

[0027] Reference Figure 1 The present invention provides a flowchart of a voiceprint recognition method according to an embodiment of the invention. The voiceprint recognition method may specifically include the following steps:

[0028] Step 11: Obtain the voiceprint feature vector to be identified for the test object and the registered voiceprint feature vectors of N registered objects in the voiceprint database;

[0029] Step 12: Perform feature separation on the voiceprint feature vector to be identified and the N registered voiceprint feature vectors to obtain the identity feature vector and non-identity feature vector of the test object, and the identity feature vector and non-identity feature vector of the N registered objects.

[0030] Step 13: Calculate the similarity between the identity feature vector of the test object and the identity feature vectors of N registered objects to obtain N original similarity scores;

[0031] Step 14: Calculate the compensation coefficient based on the similarity between the non-identity feature vector of the test object and the non-identity feature vector of the registered object;

[0032] Step 15: Compensate the original similarity using the compensation coefficient to obtain the compensated similarity;

[0033] Step 16: Obtain the voiceprint recognition result of the test object based on the compensated similarity.

[0034] In step 11, the speech data of the test subject can be acquired, and feature extraction can be performed on the speech data to obtain the voiceprint feature vector to be identified. Machine learning or deep learning algorithms can be used to extract features from the speech data to obtain the voiceprint feature vector to be identified. For example, ivector, xvector, etc. can be used to obtain the voiceprint feature vector to be identified. Alternatively, embedded vectors can be extracted from the speech to be identified, and the extracted embedded vectors can be used as the feature vector of the voiceprint to be identified.

[0035] In practice, the voiceprint database is pre-registered. The database stores the registered voiceprint feature vectors of multiple registered objects; that is, each registered object can have its own registered voiceprint feature vector. The registered voiceprint feature vector for each registered user can be obtained based on the user's speech data. The N registered voiceprint feature vectors can be all registered voiceprint feature vectors in the database, or a subset of them, where N is a positive integer.

[0036] That is, in some embodiments, when performing voiceprint recognition on a test object, the voiceprint feature vector of the test object can be compared with the registered voiceprint feature vectors of all registered objects in the voiceprint database.

[0037] In other embodiments, if the characteristics of the test subject are already known, then registered voiceprint feature vectors matching the characteristics of the test subject can be selected from the voiceprint database for comparison, without comparing with all registered voiceprint feature vectors. Accordingly, during the voiceprint database creation process, each registered voiceprint feature vector can be classified according to the characteristics of the registered subject. Different categories of registered voiceprint feature vectors can be distinguished by storing them in different sub-databases or by labeling. The classification of each registered voiceprint feature vector can be based on factors such as the registered subject's region, gender, and language. Thus, by selecting a subset of registered voiceprint feature vectors matching the characteristics of the test subject from the voiceprint database for comparison, instead of comparing with all registered voiceprint feature vectors, the efficiency of voiceprint recognition can be improved.

[0038] Understandably, if comparing a subset of registered voiceprint feature vectors from the voiceprint database that match the characteristics of the test subject fails to yield a voiceprint recognition result, then the remaining registered voiceprint feature vectors from the voiceprint database can be selected and compared with the test subject again. This approach improves voiceprint recognition efficiency while ensuring both effectiveness and accuracy.

[0039] For example, machine learning or deep learning algorithms can be used to extract features from the voice data of the registered object to obtain the registered voiceprint feature vector. Alternatively, embedded vectors can be extracted from the voice data of the registered object, and the extracted embedded vectors can be used as the registered voiceprint feature vector.

[0040] In one specific implementation of step 12, the voiceprint feature vector to be identified and the N registered voiceprint feature vectors can be decomposed by dimensionality reduction to obtain the identity feature vector and non-identity feature vector of the test object, and the identity feature vector and non-identity feature vector of the N registered objects.

[0041] Typically, voiceprint feature vectors (such as the voiceprint feature vector to be identified or the registered voiceprint feature vector) contain information such as speaker differences, noise pollution, and channel differences. By performing feature separation on the voiceprint feature vectors, the resulting identity feature vector is used to characterize the characteristics of the speaker and to distinguish different speakers. The resulting non-identity feature vector is used to characterize channel differences, environmental noise, etc.

[0042] Furthermore, the Probabilistic Linear Discriminant Analysis (PLDA) algorithm can be used to perform dimensionality reduction decomposition on the voiceprint feature vector to be identified and the N registered voiceprint feature vectors respectively, in order to achieve feature separation. For example, the standard PLDA algorithm can be used to perform dimensionality reduction decomposition on the voiceprint feature vector to be identified and the N registered voiceprint feature vectors respectively.

[0043] Specifically, based on the probabilistic linear discriminant analysis algorithm, the probabilistic linear discriminant analysis (PLDA) transformation matrix is ​​used to perform feature separation on the voiceprint feature vector to be identified, so as to obtain the identity feature vector and non-identity feature vector of the voiceprint feature vector to be identified.

[0044] Based on the probabilistic linear discriminant analysis algorithm, the probabilistic linear discriminant analysis (PLDA) transformation matrix is ​​used to separate the features of N registered voiceprint feature vectors, thereby obtaining the identity feature vector and non-identity feature vector of each registered voiceprint feature vector.

[0045] In some embodiments, the PLDA transformation matrix may include an identity feature transformation matrix, which is used to separate the identity features of the voiceprint feature vector to be identified from the voiceprint feature vector to be identified, thereby obtaining the identity feature vector of the test object and the identity feature vector of the registered object.

[0046] In other embodiments, the PLDA transformation matrix may include a non-identity feature transformation matrix, which is used to separate the non-identity features of the voiceprint feature vector to be identified from the voiceprint feature vector to be identified, thereby obtaining the identity feature vector of the test object and the non-identity feature vector of the registered object.

[0047] In some embodiments, the N registered voiceprint feature vectors can be pre-separated and stored in a voiceprint database. When performing similarity calculation in step 13, the identity feature vectors of the N objects are retrieved from the voiceprint database. Similarly, when calculating the compensation coefficient in step 14, the non-identity feature vectors of the registered objects are retrieved from the voiceprint database. This eliminates the need to perform feature separation on the N registered voiceprint feature vectors each time voiceprint recognition is performed, thus improving voiceprint recognition efficiency while reducing computational overhead.

[0048] In other embodiments, feature separation can be performed on the N registered voiceprint feature vectors each time voiceprint recognition is performed.

[0049] In one specific implementation of step 13, a cosine similarity algorithm can be used to calculate the cosine value of the identity feature vector of the test object and the identity feature vectors of N registered objects, and N original similarities can be obtained based on the obtained cosine values. For example, the cosine value of the angle between the identity feature vector of the test object and the identity feature vectors of each registered object, that is, the cosine value, can be used as the original similarity with the identity feature vector of each registered voiceprint feature vector.

[0050] In another specific implementation of step 13, when the PLDA algorithm is used to perform feature separation on the voiceprint feature vector to be identified and the N registered voiceprint feature vectors respectively, the similarity between the identity feature vector of the test object and the identity feature vectors of the N registered objects can be calculated based on the PLDA algorithm.

[0051] It should be noted that in step 13, other appropriate existing similarity calculation algorithms can also be used to calculate N original similarities, and this embodiment of the invention does not limit this.

[0052] It should be noted that the original similarity can be represented in various ways, such as scores, percentages, or values ​​within a preset range. The specific method chosen depends on the similarity calculation algorithm used.

[0053] In step 14, the following algorithms can be used: cosine similarity calculation algorithm; Euclidean distance-based similarity calculation algorithm; and neural network model-based similarity calculation algorithm.

[0054] Based on the non-identity feature vectors of the registered object (derived from the non-identity feature vector separated from the registered voiceprint feature vector) and the test object (derived from the non-identity feature vector separated from the voiceprint feature vector to be identified), the difference between the non-identity vectors of the registered object and the test object is calculated. When the difference is less than a certain range, it can help improve the credibility of the corresponding original similarity; this is called a positive difference. Conversely, when the difference exceeds this range, the larger the difference, the lower the credibility of the original similarity; this is called a negative difference. The calculation of positive and negative differences can be characterized by the cosine value of the angle between the non-identity feature vectors of the registered object and the test object. When the angle between the non-identity feature vectors of the registered object and the test object is between -90° and 90°, this type of difference is a positive difference. When the angle between the non-identity feature vectors of the registered object and the test object is between 90° and 270°, this type of difference is a negative difference.

[0055] In this embodiment of the invention, a compensation coefficient is calculated based on the difference in non-identity feature vectors between the registered object and the test object. More compensation is given for positive differences to increase confidence; less compensation is given for negative differences to reduce confidence and bring it back to a normal confidence range. Therefore, the compensated similarity can more reliably and realistically reflect the similarity to the voiceprint of the registered object, thereby improving the voiceprint recognition results for the test object.

[0056] In some embodiments, the compensation coefficient can be obtained based on the cosine of the angle between the non-identity feature vectors of the test object and the registered object.

[0057] The cosine of the angle ranges from -1 to 1. The size of the angle can be used to characterize the similarity between the non-identity feature vectors of the test object and the registered object. Specifically, the smaller the angle between the non-identity feature vectors of the test object and the registered object, and the larger the cosine value of the angle, the greater the probability that the non-identity feature vectors of the test object and the registered object are in the same direction. The closer the channels and environments represented by their non-identity feature vectors are, the higher the reliability of the obtained original similarity. When the similarity between the non-identity feature vectors of the test object and the registered object is high, it indicates that the channel difference between them is small. Therefore, there is reason to believe that the confidence of the original similarity calculated from the identity feature vectors of the test object and the registered object is high, and thus a high compensation can be given to the original similarity between the test object and the registered object. Conversely, if the similarity between the non-identity feature vectors of the test object and the registered object is small, it indicates that the channel difference between them is large. In this case, even if the original similarity falls within the confidence interval, the confidence of the original similarity is not high, so a smaller compensation should be given.

[0058] As described above, by compensating for the original similarity falling within the confidence interval based on the similarity between the non-identity feature vector of the test object and the non-identity feature vector of the registered object, it helps to widen the gap in the compensated similarity and make the compensated similarity more credible and realistic in reflecting the similarity with the voiceprint of the registered object, thereby improving the voiceprint recognition results of the test object.

[0059] In some embodiments, the cosine of the angle between the non-identity feature vector of the test object and the non-identity feature vector of the registered object can be calculated using the following formula (1).

[0060]

[0061] Where cos(e1,e2) is the cosine of the angle between the non-identity feature vector of the test object and the non-identity feature vector of the registered object; e1 is the non-identity feature vector of the test object; e2 is the non-identity feature vector of the registered object; ||e1||2 is the norm of the non-identity feature vector of the test object, used to characterize the length of e1; ||e2||2 is the length of the non-identity feature vector of the registered object, used to characterize the length of e2.

[0062] The compensation coefficient is determined based on the cosine of the angle between the non-identity feature vector of the test object and the non-identity feature vector of the registered object. The compensation coefficient is positively correlated with the cosine of the angle.

[0063] For example, the cosine of the angle between the non-identity feature vector of the test object and the non-identity feature vector of the registered object can be used as a compensation coefficient. Alternatively, the cosine of the angle between the non-identity feature vector of the test object and the non-identity feature vector of the registered object can be calculated, and the result can be used as a compensation coefficient.

[0064] In other embodiments, the similarity between the non-identity feature vector of the test object and the non-identity feature vector of the registered object can be calculated based on Euclidean distance.

[0065] For example, the Euclidean distance between the non-identity feature vector of the test object and the non-identity feature vector of the registered object can be calculated using the following formula (2).

[0066] d(e1,e2)=||e1-e2||2; (2)

[0067] Where d(e1,e2) is the Euclidean distance between the non-identity feature vector of the test object and the non-identity feature vector of the registered object; e1 is the non-identity feature vector of the test object; and e2 is the non-identity feature vector of the registered object.

[0068] Since the magnitude of the Euclidean distance is inversely correlated with the similarity between the non-identity feature vectors of the test object and the registered object, the Euclidean distance needs to be processed so that the processed Euclidean distance is positively correlated with the similarity between the non-identity feature vectors of the test object and the registered object.

[0069] For example, the reciprocal of the obtained Euclidean distance can be taken, and the reciprocal of the Euclidean distance between the non-identity feature vector of the test object and the non-identity feature vector of the registered object can be used as the compensation coefficient. It is understandable that other monotonically decreasing transformations can be applied to the Euclidean distance between the non-identity feature vectors of the test object and the non-identity feature vectors of the registered object, and the result of the monotonically decreasing transformation can be used as the compensation coefficient.

[0070] In some other embodiments, a neural network model can be used to calculate the compensation coefficient. It is understood that a more complex similarity matching function can also be implemented using a machine learning model, and the compensation coefficient can be obtained based on the similarity matching function.

[0071] In some non-limiting embodiments, compensation coefficients may be calculated separately for each original similarity.

[0072] In some other non-limiting embodiments, a compensation coefficient can be calculated for the original similarities that meet the compensation conditions. Specifically, original similarities that meet the compensation conditions are selected from N original similarities. For each original similarity that meets the compensation conditions, a compensation coefficient is calculated based on the similarity between the non-identity feature vector of the registered object corresponding to the original similarity that meets the compensation conditions and the non-identity feature vector of the test object.

[0073] The compensation conditions may include any of the following: Condition 1: The original similarity is within a preset confidence interval. The number of original similarities within the preset confidence interval can be one or more. Condition 2: The number of original similarities greater than or equal to the confidence threshold is multiple, and the difference between the multiple original similarities is within a preset difference range. That is, as long as the original similarity satisfies either Condition 1 or Condition 2, the compensation coefficient corresponding to the original similarity is calculated, and compensation is applied.

[0074] Research has revealed that current technologies, whether using cosine similarity or probabilistic linear discriminant analysis to calculate similarity, directly rely on the relationship between the obtained similarity and the voiceprint decision threshold to provide voiceprint recognition results. On one hand, the voiceprint decision threshold is generally given during the training phase and is determined by the training samples. Therefore, the optimal decision threshold determined based on the training samples inevitably contains a certain degree of error. The optimal threshold distribution corresponding to different sample sets will fall within a certain range. In this case, simply using the absolute threshold corresponding to a single sample as the voiceprint decision threshold, and relying solely on the voiceprint decision threshold as the basis for judgment, is not reasonable. On the other hand, since voiceprint feature vectors often simultaneously contain information such as speaker differences, noise contamination, and channel differences, when scores are similar or within a certain range near the threshold, these similar scores cannot accurately reflect the similarity between the test object and each registered object. Simply selecting the maximum value among the matching scores for threshold decision, and directly relying on the feature vector or features, is prone to misjudgment.

[0075] In this embodiment of the invention, when there are multiple original similarities within a preset confidence interval, or multiple original similarities greater than or equal to a confidence threshold, multiple original similarities will satisfy the compensation conditions. In this case, for each original similarity satisfying the compensation conditions, a compensation coefficient is calculated based on the non-identity feature vector of the registered object and the non-identity feature vector of the test object corresponding to the original similarity. That is, each original similarity satisfying the compensation conditions corresponds to a compensation coefficient. In this embodiment of the invention, by setting compensation conditions and compensating for the original similarities satisfying the compensation conditions, the compensated similarity can effectively reduce the impact of channel and environmental differences on the voiceprint test results of the test object. Especially for some relatively close original similarities, compensation helps to widen the gap between the original similarities, thereby improving the accuracy of the voiceprint recognition results of the test object obtained based on the compensated similarity. Furthermore, while improving the accuracy of the voiceprint recognition results, it can also reduce computational overhead and improve recognition efficiency.

[0076] For original similarity scores that do not meet the compensation conditions, there are generally two scenarios. The first scenario is where the original similarity score is below the confidence threshold and below the minimum value of the confidence interval. Therefore, this type of original similarity score has low acceptance and a low probability of matching the registered object in the voiceprint database. The second scenario is where the original similarity score is above the confidence threshold and above the maximum value of the confidence interval. This type of original similarity score has high acceptance, and the original similarity score is sufficient to accurately determine the match with the registered object. Therefore, regardless of whether it is the first or second scenario, no compensation is needed to determine the voiceprint recognition result, that is, to accurately confirm the match with the registered object in the voiceprint database.

[0077] In one specific embodiment of step 15, the compensated similarity can be calculated in the following manner, referring to... Figure 2 It gave Figure 1 A flowchart of a specific implementation of step 15 may include the following steps:

[0078] Step 151: Calculate the upward and downward fluctuations of the upper and lower limits of the confidence interval relative to the confidence threshold, respectively.

[0079] Step 152: Calculate the compensation amount based on the compensation coefficient, the upward fluctuation amount, and the downward fluctuation amount.

[0080] Step 153: The original similarity is compensated using the compensation amount to obtain the compensated similarity.

[0081] In practice, the upward fluctuation is the difference between the upper limit of the confidence interval and the confidence threshold. The downward fluctuation is the difference between the confidence threshold and the lower limit of the confidence interval. The confidence interval can be [confidence threshold - downward fluctuation, confidence threshold + upward fluctuation].

[0082] In a non-limiting embodiment of step 152, the compensation coefficient can be multiplied by the sum of the upward and downward floating amounts, and the result of the multiplication can be used as the compensation amount.

[0083] For example, the compensation amount can be calculated using the following formula (3).

[0084] Δscore=μ*(lower+upper); (3)

[0085] Where Δscore is the compensation amount; μ is the compensation coefficient; lower is the downward fluctuation amount; and upper is the upward fluctuation amount.

[0086] In a non-limiting embodiment of step 153, the sum of the compensation amount and the original similarity is used as the compensated similarity.

[0087] For example, the compensated similarity can be calculated using the following formula (4).

[0088] score final =score original +Δscore; (4)

[0089] Among them, score final The score represents the compensated similarity. original The original similarity is denoted as Δscore, and the compensation amount is Δscore.

[0090] In specific implementation, the confidence interval and confidence threshold in the above embodiments can be obtained in the following way. The determination of the confidence interval and confidence threshold will be explained below using the feature separation model obtained by the PLDA algorithm to perform feature separation on the voiceprint feature vector to be identified and the N registered voiceprint feature vectors respectively.

[0091] Specifically, a large number of different training sample data are used. These training sample data can be voiceprint feature vector samples and their corresponding labels, or embedding vector sample sets and their corresponding labels. A portion of the training sample data is used as the training set for the feature separation model, and another portion is used as the test set. The feature separation model is trained using the training set. The trained feature separation model is then tested using the test set. Histograms can be used to statistically analyze the calculated scores, determining the score center point and its upper and lower fluctuation ranges. The score center point can be used as the confidence threshold, also known as the absolute threshold. The upper and lower fluctuation ranges are denoted as the upper fluctuation amount and the lower fluctuation amount, respectively. The difference between the confidence threshold and the lower fluctuation amount is the lower limit of the confidence interval, and the sum of the confidence threshold and the upper fluctuation amount is the upper limit of the confidence interval. For example, if the score center point is 90, then the confidence threshold is 90. The upper fluctuation amount is denoted as "upper," and the lower fluctuation amount as "lower," so the confidence interval is [90-lower, 90+upper].

[0092] To improve the robustness of the trained feature separation model, training samples from different genders, age groups, and scenarios are provided during the training phase to enhance the transformation accuracy of the PLDA transformation matrix. Improving the transformation accuracy of the PLDA transformation matrix enhances the precision of separating identity and non-identity feature vectors during feature separation. Furthermore, it improves the reasonableness and accuracy of the obtained confidence thresholds and confidence intervals.

[0093] The PLDA transformation matrix can include identity feature transformation distance and non-identity feature transformation matrix (also known as non-identity difference space transformation matrix).

[0094] In some implementations, in step 16 above, the maximum compensated similarity refers to the maximum value among the compensated similarities. The voiceprint recognition result can be obtained based on the relationship between the maximum compensated similarity and the confidence threshold.

[0095] For example, the registered object corresponding to the maximum value among all compensated similarities that exceed the confidence threshold is taken as the recognition result of the test object. For instance, if the registered object corresponding to the maximum value among all compensated similarities that exceed the confidence threshold is registered object A, then the voiceprint recognition result indicates that the test object is registered object A.

[0096] In other embodiments, if there are multiple largest numbers of compensated similarities, the voiceprint recognition result is obtained based on the relationship between the maximum value among the original similarities corresponding to the multiple largest compensated similarities and the confidence threshold. For example, the compensated similarity B and compensated similarity C have equal values, and both are maximum values. If the maximum value of compensated similarity B among the original similarities before compensation and the original similarities of compensated similarity C is greater than the confidence threshold, the registered object corresponding to the maximum value of compensated similarity B among the original similarities before compensation and the original similarities of compensated similarity C is used as the test object.

[0097] The voiceprint recognition method provided in this invention can be applied to voiceprint recognition tasks for text-related, text-independent, long, and short speech.

[0098] As described above, by performing feature separation on the voiceprint vector to be identified of the test object and the registered voiceprint feature vectors of N registered objects, we obtain the identity feature vector and non-identity feature vector of the test object, and the identity feature vector and non-identity feature vector of the N registered objects. Based on the similarity between the identity feature vector of the test object and the identity feature vectors of the N registered objects, we obtain N original similarity scores. Since the compensation coefficient is calculated based on the similarity between the non-identity feature vector of the test object and the non-identity feature vector of the registered objects, the compensation coefficient can characterize the difference between the voiceprint feature vector to be identified of the test object and the registered voiceprint feature vectors of the N registered objects in non-identity features. Therefore, by using the compensation coefficient to compensate for the original similarity scores, the compensated similarity scores can effectively reduce the influence of non-identity feature vectors, improve the confidence of the compensated similarity scores, and thus improve the accuracy of voiceprint recognition.

[0099] To facilitate a better understanding and implementation of the embodiments of the present invention by those skilled in the art, the following description is provided in conjunction with... Figure 3 The present invention provides a flowchart of another voiceprint recognition method in an embodiment of the invention, using the PLDA algorithm (e.g., the Standard PLDA algorithm) as an example to illustrate the specific process of the voiceprint recognition method.

[0100] The voiceprint recognition method may include a training phase, a registration phase, and an application phase. The training phase is used to train the feature separation model, including steps 301 to 304 below, to obtain the identity feature vector transformation matrix, the non-identity feature vector matrix, the confidence interval, and the confidence threshold. The registration phase is used to build a voiceprint database, specifically including steps 305 to 307 below. The application phase is used to perform voiceprint recognition on the speech of the test subject, and to determine the identity of the test subject based on the voiceprint recognition results, including steps 308 to 314 below.

[0101] Step 301: Input training sample data.

[0102] Step 302, Training the Stardard PLDA transformation matrix.

[0103] Step 303: Train to obtain the identity feature vector transformation matrix and the non-identity feature vector transformation matrix.

[0104] The Stardard PLDA transformation matrix is ​​obtained by training the Stardard PLDA transformation matrix in step 302. The Stardard PLDA transformation matrix includes an identity feature vector transformation matrix and a non-identity feature vector transformation matrix. The identity feature vector transformation matrix is ​​used to separate the voiceprint feature vector to obtain the identity feature vector. The non-identity feature vector transformation matrix is ​​used to separate the voiceprint feature vector to obtain the non-identity feature vector.

[0105] Step 304 yields the confidence interval and the confidence threshold.

[0106] Step 305: Enter the registration voice data.

[0107] In practice, the registration voice data of each registered object can be input.

[0108] Step 306, voiceprint feature vector extraction.

[0109] Specifically, voiceprint feature vectors are extracted from the voice data of the registered objects to obtain the voiceprint feature vectors of each registered object.

[0110] Step 307: Construct a voiceprint library.

[0111] Based on the obtained registration voiceprint feature vectors of each registered object, a voiceprint database is constructed. The voiceprint database includes the registration voiceprint feature vectors of multiple registered objects.

[0112] Step 308: Input the voice data of the test subject.

[0113] Step 309, voiceprint feature vector extraction.

[0114] Specifically, the voiceprint feature vector of the test subject is extracted from the speech data to obtain the voiceprint feature vector of the test subject, which is then used as the voiceprint feature vector to be identified.

[0115] Step 310: Perform dimensionality reduction decomposition on the voiceprint feature vector to obtain the identity feature vector and the non-identity feature vector.

[0116] In practice, an identity feature vector transformation matrix is ​​used to perform dimensionality reduction decomposition on both the voiceprint feature vector to be identified and the registered voiceprint feature vector, respectively, to obtain the identity feature vector of the test object and the identity feature vector of the registered object. Similarly, a non-identity feature vector transformation matrix is ​​used to perform dimensionality reduction decomposition on both the voiceprint feature vector to be identified and the registered voiceprint feature vector, respectively, to obtain the non-identity feature vector of the test object and the non-identity feature vector of the registered object.

[0117] Step 311: Calculate the original similarity based on the identity feature vector.

[0118] Specifically, the similarity between the identity feature vector of the test object and the identity feature vector of each registered object is calculated to obtain the original similarity.

[0119] Step 312: Calculate the compensation coefficient based on the non-identity feature vector.

[0120] In some embodiments, the original similarity that meets the compensation conditions can be selected. For the original similarity that meets the compensation conditions, the compensation coefficient of each original similarity that meets the compensation conditions is calculated based on the similarity between the non-identity feature vector of the corresponding registered object and the non-identity feature vector of the test object.

[0121] Step 313: Compensate the original similarity based on the confidence interval and the confidence threshold.

[0122] Step 314: Obtain the voiceprint recognition result based on the compensated similarity.

[0123] This invention also provides a voiceprint recognition device, as described in the embodiments of the present invention. Figure 4 A schematic diagram of a voiceprint recognition device according to an embodiment of the present invention is provided. The voiceprint recognition device 40 may specifically include:

[0124] The acquisition unit 41 is used to acquire the voiceprint feature vector to be identified of the test object and the registered voiceprint feature vectors of N registered objects in the voiceprint database, where N is a positive integer.

[0125] The feature separation unit 42 is used to perform feature separation on the voiceprint feature vector to be identified and the N registered voiceprint feature vectors respectively, to obtain the identity feature vector and non-identity feature vector of the test object, and the identity feature vector and non-identity feature vector of the N registered objects.

[0126] The similarity calculation unit 43 is used to calculate the similarity between the identity feature vector of the test object and the identity feature vectors of N registered objects, and obtain N original similarity scores.

[0127] The compensation coefficient calculation unit 44 is used to calculate the compensation coefficient based on the similarity between the non-identity feature vector of the test object and the non-identity feature vector of the registered object.

[0128] Compensation unit 45 is used to compensate the original similarity using the compensation coefficient to obtain the compensated similarity;

[0129] Voiceprint recognition result obtaining unit 46 is used to obtain the voiceprint recognition result of the test object based on the compensated similarity.

[0130] In specific implementations, the voiceprint recognition device 40 described above can be used to implement the voiceprint recognition method provided in any of the above embodiments. For the specific working principle and workflow of the voiceprint recognition device 40, please refer to the description in the voiceprint recognition method provided in any of the above embodiments; it will not be repeated here.

[0131] In specific implementations, the aforementioned voiceprint recognition device 40 may correspond to a chip in the terminal that has voiceprint recognition function, such as a SOC (System-On-a-Chip), a baseband chip, etc.; or to a chip module in the terminal that includes a voiceprint recognition function; or to a chip module that has a data processing function chip; or to the terminal itself.

[0132] This invention also provides a computer-readable storage medium storing a computer program thereon, which, when executed by a processor, performs the steps of the voiceprint recognition method provided in any of the above embodiments of this invention.

[0133] The computer-readable storage medium may include non-volatile or non-transitory memory, and may also include optical discs, hard disk drives, solid-state drives, etc.

[0134] Specifically, in this embodiment of the invention, the processor can be a central processing unit (CPU), or it can be other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. The general-purpose processor can be a microprocessor or any conventional processor.

[0135] It should also be understood that the memory in the embodiments of this application can be volatile memory or non-volatile memory, or may include both volatile and non-volatile memory. The non-volatile memory can be read-only memory (ROM), programmable read-only memory (PROM), erasable programmable read-only memory (EPROM), electrically erasable programmable read-only memory (EEPROM), or flash memory. The volatile memory can be random access memory (RAM), which is used as an external cache. By way of example, but not limitation, many forms of random access memory (RAM) are available, such as static random access memory (SRAM), dynamic random access memory (DRAM), synchronous dynamic random access memory (SDRAM), double data rate synchronous dynamic random access memory (DDR SDRAM), enhanced synchronous dynamic random access memory (ESDRAM), synchronous linked dynamic random access memory (Synchlink DRAM, SLDRAM), and direct memory bus RAM (DR RAM).

[0136] This invention also provides a terminal, including a memory and a processor. The memory stores a computer program that can run on the processor. When the processor runs the computer program, it executes the steps of the voiceprint recognition method provided in any of the above embodiments.

[0137] The memory and the processor are coupled, and the memory can be located inside or outside the terminal. The memory and the processor can be connected via a communication bus.

[0138] Terminals can include, but are not limited to, mobile phones, computers, tablets, and other terminal devices, as well as servers, cloud platforms, etc.

[0139] The above embodiments can be implemented, in whole or in part, by software, hardware, firmware, or any other combination thereof. When implemented using software, the above embodiments can be implemented, in whole or in part, as a computer program product. The computer program product includes one or more computer instructions or computer programs. When the computer instructions or computer program are loaded or executed on a computer, all or part of the processes or functions described in the embodiments of this application are generated. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer program can be stored in a computer-readable storage medium or transferred from one computer-readable storage medium to another. For example, the computer program can be transferred from one website, computer, server, or data center to another website, computer, server, or data center via wired or wireless means.

[0140] In the embodiments provided in this application, it should be understood that the disclosed methods, apparatuses, and systems can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative; for example, the division of units is merely a logical functional division, and other division methods may exist in actual implementation; for example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. 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, that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs. For example, the embodiments provided above can be implemented by software code running on the processor of a platform chip, embedded chip, or terminal chip or stored in a memory chip; they can also be implemented by hardware circuitry and embedded in a chip or chip module.

[0141] Furthermore, the functional units in the various embodiments of the present invention can be integrated into one processing unit, or each unit can be physically included separately, or two or more units can be integrated into one unit. The integrated units described above can be implemented in hardware or in a combination of hardware and software functional units. For example, for various devices or products applied to or integrated into a chip, each module / unit can be implemented using hardware such as circuits, or at least some modules / units can be implemented using software programs running on a processor integrated within the chip, while the remaining (if any) modules / units can be implemented using hardware such as circuits; for various devices or products applied to or integrated into a chip module, each module / unit can be implemented using hardware such as circuits, and different modules / units can be located in the same component (e.g., chip, circuit module, etc.) or different components of the chip module, or at least some modules / units can be implemented using hardware such as circuits. The components can be implemented using software programs that run on the processor integrated within the chip module. The remaining (if any) modules / units can be implemented using hardware methods such as circuits. For various devices and products applied to or integrated into the terminal, each of its components / units can be implemented using hardware methods such as circuits. Different modules / units can be located in the same component (e.g., chip, circuit module, etc.) or in different components within the terminal. Alternatively, at least some modules / units can be implemented using software programs that run on the processor integrated within the terminal, while the remaining (if any) modules / units can be implemented using hardware methods such as circuits.

[0142] It should be understood that the term "and / or" in this article is merely a description of the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A existing alone, A and B existing simultaneously, or B existing alone. Additionally, the character " / " in this article indicates that the preceding and following related objects have an "or" relationship.

[0143] In the embodiments of this application, "multiple" refers to two or more.

[0144] The descriptions of "first," "second," "third," etc., appearing in the embodiments of this application are for illustrative purposes and to distinguish the objects being described. They have no order and do not indicate any special limitation on the number of devices in the embodiments of this application, nor do they constitute any limitation on the embodiments of this application.

[0145] It should be noted that the sequence number of each step in this embodiment does not represent a limitation on the execution order of each step.

[0146] While the present invention has been disclosed above, it is not limited thereto. Any person skilled in the art can make various modifications and alterations without departing from the spirit and scope of the invention; therefore, the scope of protection of the present invention should be determined by the scope defined in the claims.

Claims

1. A voiceprint recognition method, characterized in that, include: Obtain the voiceprint feature vector of the test object to be identified and the registered voiceprint feature vectors of N registered objects in the voiceprint database, where N is a positive integer; Feature separation is performed on the voiceprint feature vector to be identified and the N registered voiceprint feature vectors to obtain the identity feature vector and non-identity feature vector of the test object, and the identity feature vector and non-identity feature vector of the N registered objects. Calculate the similarity between the identity feature vector of the test object and the identity feature vectors of N registered objects to obtain N original similarity scores; The compensation coefficient is calculated based on the similarity between the non-identity feature vector of the test object and the non-identity feature vector of the registered object; The original similarity is compensated using the compensation coefficient to obtain the compensated similarity. The voiceprint recognition result of the test object is obtained based on the compensated similarity. The step of calculating the compensation coefficient based on the similarity between the non-identity feature vector of the test object and the non-identity feature vector of the registered object includes: Select the original similarity that satisfies the compensation condition from N original similarity values; For each original similarity that meets the compensation conditions, the compensation coefficient corresponding to each original similarity that meets the compensation conditions is calculated based on the similarity between the non-identity feature vector of the registered object corresponding to the original similarity that meets the compensation conditions and the non-identity feature vector of the test object. The compensation conditions include any of the following: The original similarity is within a preset confidence interval; There are multiple original similarities that are greater than or equal to the confidence threshold, and the difference between the multiple original similarities is within a preset difference range.

2. The voiceprint recognition method as described in claim 1, characterized in that, The compensation coefficient is calculated using any of the following similarity calculation algorithms: Cosine similarity calculation algorithm; Similarity calculation algorithm based on Euclidean distance; A similarity calculation algorithm based on a neural network model.

3. The voiceprint recognition method as described in claim 1, characterized in that, The step of compensating the original similarity using the compensation coefficient to obtain the compensated similarity includes: Calculate the upward fluctuation of the upper limit of the confidence interval relative to the confidence threshold, and the downward fluctuation of the lower limit of the confidence interval relative to the confidence threshold. The compensation amount is calculated based on the compensation coefficient, the upward fluctuation amount, and the downward fluctuation amount. The original similarity is compensated using the compensation amount to obtain the compensated similarity.

4. The voiceprint recognition method as described in claim 3, characterized in that, The compensated similarity is calculated using the following formula: ； ； in, The compensated similarity is... The original similarity, The compensation amount is... The compensation coefficient is... The downward fluctuation amount, This refers to the upward floating amount.

5. The voiceprint recognition method as described in claim 1, characterized in that, The step of performing feature separation on the voiceprint feature vector to be identified and the N registered voiceprint feature vectors includes: Based on the probabilistic linear discriminant analysis algorithm, the probabilistic linear discriminant analysis transformation matrix is ​​used to perform feature separation on the voiceprint feature vector to be identified and the N registered voiceprint feature vectors respectively.

6. The voiceprint recognition method as described in claim 1, characterized in that, The process of obtaining the voiceprint recognition result of the test object based on the compensated similarity includes: The voiceprint recognition result is obtained based on the relationship between the maximum compensated similarity and the confidence threshold. Alternatively, if there are multiple largest numbers of compensated similarities, the voiceprint recognition result is obtained based on the relationship between the maximum value among the original similarities corresponding to the multiple largest compensated similarities and the confidence threshold.

7. A voiceprint recognition device, characterized in that, include: The acquisition unit is used to acquire the voiceprint feature vector to be identified of the test object and the registered voiceprint feature vectors of N registered objects in the voiceprint database, where N is a positive integer. The feature separation unit is used to perform feature separation on the voiceprint feature vector to be identified and the N registered voiceprint feature vectors respectively, to obtain the identity feature vector and non-identity feature vector of the test object, and the identity feature vector and non-identity feature vector of the N registered objects. The similarity calculation unit is used to calculate the similarity between the identity feature vector of the test object and the identity feature vectors of N registered objects, and obtain N original similarity scores. The compensation coefficient calculation unit is used to calculate the compensation coefficient based on the similarity between the non-identity feature vector of the test object and the non-identity feature vector of the registered object. The compensation unit is used to compensate the original similarity using the compensation coefficient to obtain the compensated similarity. A voiceprint recognition result determination unit is used to obtain the voiceprint recognition result of the test object based on the compensated similarity. The compensation coefficient calculation unit is used to select original similarities that satisfy the compensation conditions from N original similarities; for each original similarity that satisfies the compensation conditions, the compensation coefficient corresponding to each original similarity that satisfies the compensation conditions is calculated based on the similarity between the non-identity feature vector of the registered object corresponding to the original similarity that satisfies the compensation conditions and the non-identity feature vector of the test object; wherein, the compensation conditions include any of the following conditions: The original similarity is within a preset confidence interval; There are multiple original similarities that are greater than or equal to the confidence threshold, and the difference between the multiple original similarities is within a preset difference range.

8. A computer-readable storage medium having a computer program stored thereon, characterized in that, The computer program, when run by a processor, performs the steps of the voiceprint recognition method according to any one of claims 1 to 6.

9. A terminal comprising a memory and a processor, wherein the memory stores a computer program capable of running on the processor, characterized in that, When the processor runs the computer program, it performs the steps of the voiceprint recognition method according to any one of claims 1 to 6.

Images