Speaker verification based adaptive margin optimization method, system, and electronic device

Abstract

This invention provides an adaptive margin optimization method, system, and electronic device based on speaker verification. The method includes: inputting speech training data including various speech durations into a speaker verification model; determining the loss function of the speaker verification model; adaptively optimizing the margin parameters of the loss function based on the speech durations in the speech training data and a preset target margin for each speech duration; and training the speaker verification model using the margin parameters of the adaptively optimized loss function to determine the acceptable training difficulty of the speaker verification model. This invention utilizes training speech of varying lengths to better simulate real-life scenarios. Through adaptive optimization and fine-tuning of the margin, adjusting the margin according to the duration and similarity of each speech, this method achieves good speaker verification performance for speech of different durations in real-world scenarios.

Description

Technical Field

[0001] This invention relates to the field of intelligent speech, and more particularly to an adaptive margin optimization method, system, and electronic device based on speaker verification. Background Technology

[0002] In recent years, deep learning has made progress in the field of ASV (automatic speaker verification). ASV models based on neural networks typically consist of three modules: frame-level speaker feature extraction, pooling layers for statistical extraction, and a loss function for optimization. The loss function is based on the softmax function; current techniques optimize speaker embeddings in the hypersphere space using angular softmax. Furthermore, a margin is added to the loss function. The purpose of adding the margin is to minimize intra-class distance and maximize inter-class distance.

[0003] Existing techniques typically use margin-based loss, AAM (additive angular margin) softmax. Theoretically, increasing the margin within a reasonable range can make speaker embeddings more discriminative. However, the training segments commonly used in ASV are very short, such as 2 seconds. Excessive margin can make the optimization task very challenging and may even lead to poor training results. Existing techniques often use LMFT (large margin fine-tuning), which uses longer speech segments to adapt to a larger margin in the second stage of ASV model training.

[0004] In the process of realizing this invention, the inventors discovered at least the following problems in the related technology:

[0005] The margin in the loss function plays a crucial role in the ASV task, helping the neural network extract more discernible speaker embeddings. However, simply increasing the margin value can significantly increase the training difficulty and lead to a degrade in ASV performance. While the LMFT technique alleviates this problem, in real-world validation scenarios, people may speak very short words or phrases for validation. In such cases, the LMFT model may encounter duration mismatches, resulting in an insufficiently precise margin determination and affecting the validation results. Summary of the Invention

[0006] To address the issue that existing LMFT models within the current ASV framework may encounter duration mismatches when processing speech of varying durations in real-world scenarios, thus affecting verification results, this invention provides, in a first aspect, an adaptive margin optimization method based on speaker verification, comprising:

[0007] Speech training data including various speech durations are input into the speaker verification model, and the loss function of the speaker verification model is determined.

[0008] Based on the speech of each speech duration in the speech training data, and the target margin preset for each speech duration, the margin parameter of the loss function is adaptively optimized, wherein the adaptive optimization includes: duration-based adaptive optimization and cosine similarity-based adaptive optimization.

[0009] The speaker verification model is trained by determining the acceptable training difficulty of the speaker verification model using the margin parameter of the adaptively optimized loss function, wherein the training difficulty is related to the speech duration and cosine similarity.

[0010] Secondly, embodiments of the present invention provide an adaptive margin optimization system based on speaker verification, comprising:

[0011] The loss function determination module is used to input speech training data including multiple speech durations into the speaker verification model and determine the loss function of the speaker verification model.

[0012] An adaptive optimization program module is used to adaptively optimize the margin parameters of the loss function based on the speech of each speech duration in the speech training data and the preset target margin for each speech duration. The adaptive optimization includes: duration-based adaptive optimization and cosine similarity-based adaptive optimization.

[0013] The training module is used to train the speaker verification model by determining the acceptable training difficulty of the speaker verification model using the margin parameter of the adaptively optimized loss function, wherein the training difficulty is related to the speech duration and cosine similarity.

[0014] Thirdly, an electronic device is provided, comprising: at least one processor, and a memory communicatively connected to the at least one processor, wherein the memory stores instructions executable by the at least one processor, the instructions being executed by the at least one processor to enable the at least one processor to perform the steps of the speaker verification-based adaptive margin optimization method of any embodiment of the present invention.

[0015] Fourthly, embodiments of the present invention provide a storage medium storing a computer program thereon, characterized in that, when the program is executed by a processor, it implements the steps of the adaptive margin optimization method based on speaker verification according to any embodiment of the present invention.

[0016] The beneficial effects of this invention are as follows: by using training speech of varying lengths to better simulate real-life scenarios, and by adaptively optimizing and fine-tuning the margin, the margin is adjusted according to the duration and similarity of each speech, so that this method can have good speaker verification performance when processing speech of different durations in real-life scenarios. Attached Figure Description

[0017] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0018] Figure 1 This is a flowchart of an adaptive margin optimization method based on speaker verification provided in an embodiment of the present invention;

[0019] Figure 2 This is a schematic diagram of adaptive fitting parameters for an adaptive margin optimization method based on speaker verification provided in an embodiment of the present invention;

[0020] Figure 3 This is a schematic diagram comparing the error rates of system performance, etc., of an adaptive margin optimization method based on speaker verification provided in an embodiment of the present invention.

[0021] Figure 4 This is a schematic diagram illustrating the target score change of an adaptive margin optimization method based on speaker verification provided in an embodiment of the present invention;

[0022] Figure 5 This is a schematic diagram of the standard deviation of an adaptive margin optimization method based on speaker verification provided in an embodiment of the present invention;

[0023] Figure 6 This is a schematic diagram of the structure of an adaptive margin optimization system based on speaker verification provided in an embodiment of the present invention;

[0024] Figure 7 This is a schematic diagram of an embodiment of an electronic device based on speaker verification and adaptive margin optimization, provided as an embodiment of the present invention. Detailed Implementation

[0025] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0026] like Figure 1 The diagram shows a flowchart of an adaptive margin optimization method based on speaker verification according to an embodiment of the present invention, which includes the following steps:

[0027] S11: Input speech training data including various speech durations into the speaker verification model, and determine the loss function of the speaker verification model;

[0028] S12: Based on the speech of each speech duration in the speech training data, and the target margin preset for each speech duration, the margin parameter of the loss function is adaptively optimized, wherein the adaptive optimization includes: duration-based adaptive optimization and cosine similarity-based adaptive optimization.

[0029] S13: The speaker verification model is trained by determining the acceptable training difficulty of the speaker verification model using the margin parameter of the adaptively optimized loss function, wherein the training difficulty is related to the speech duration and cosine similarity.

[0030] In this implementation, by testing existing techniques using the VoxCeleb1 dataset, our method found that performance degradation occurs in scenarios with mismatched durations, necessitating the use of dynamic and varying speech lengths to match more scenarios. We also discovered that a higher margin helps the model extract stronger speaker features; selecting a reasonable margin is crucial. Considering the shortcomings of existing techniques, our method enables the speaker verification model to perceive speech inputs of different durations and adjusts the margin during the fine-tuning phase based on data and the corresponding training difficulty.

[0031] For step S11, regarding the selection of training data, this method selects speech training data of various durations. This is because, in order to match more realistic test scenarios, randomly selecting speech of various lengths to train the speaker verification model enriches the variety of training samples. In this way, the model can cope with speech of various durations, thereby improving the scope of application scenarios of this method.

[0032] For example, the speech training data includes speech samples of varying lengths to be verified by the speaker, along with the corresponding target speaker embeddings. The speaker verification samples are input into the speaker verification model to obtain predicted speaker embeddings. The loss function is then determined using the predicted speaker embeddings and the target speaker embeddings. The commonly used softmax classification loss in existing technologies is expressed as follows:

[0033]

[0034] Where N is the number of training samples, C is the number of speakers in the training samples, and x i It is the speaker embedding of the i-th sample, y i is the corresponding label index, w is the parameter of the last fully connected layer, and b is the bias.

[0035] After normalizing the weights of each parameter in the loss function, zeroing the bias, and introducing a margin, the margin-based softmax loss function L is obtained. margin_S :

[0036]

[0037] Where s is a scaling factor, θ j It is w j and x i The angle between them Rewritten as s·cos(θ) j ) or s·f(θ j ).

[0038] The following angle function illustrates the form of the maximum loss based on the margin. In the adaptive fine-tuning optimization of this method, only the case of AAM (additive angular margin) softmax is considered, that is, the case where m1 = 1 and m3 = 0.

[0039]

[0040] For step S12, the optimal margin for a specific training time has been pre-identified through this method. For example, when using 2s of training speech, the margin in the AAM loss function is best set to 0.2, and the margin for 6s of training speech is best set to 0.5. By adaptively optimizing the margin parameter of the loss function using these speech durations and corresponding target margins, the adaptation of this method can be called ALMFT (adaptive large margin fine-tuning).

[0041] Specifically, when the adaptive optimization is duration-based adaptive optimization, the linear expression of the increasing relationship between the target margin and the speech duration is as follows:

[0042] Margin = A * Duration + B

[0043] Wherein, Margin is the target margin of the loss function, Duration is the speech duration, and A and B are margin parameters for adaptive optimization.

[0044] In this implementation, the duration-based adaptive optimization can be called D-ALMFT (duration-based adaptive large margin fine-tuning). A larger margin helps in learning to identify speaker embeddings, but it also increases the training difficulty. Longer training utterances contain more speaker information, are easier to classify, and can help the model adapt to a larger margin, as well as to commonly used long-duration evaluation scenarios. Therefore, this method introduces a dynamic function that allows the margin to automatically adapt to the duration of the training utterances. Here, a linear transformation is used to describe the increasing relationship between the margin and the utterance duration. By setting the range of the margin and the range of the utterance duration, a straight line is fitted to obtain the values ​​of A and B in Margin = A*Duration + B.

[0045] As another implementation, when the adaptive optimization is based on cosine similarity, the exponential function of the relationship between the target margin and cosine similarity is:

[0046] Margin=min(αexp(β*Similarity),γ),

[0047] Wherein, Margin is the target margin of the loss function, Similarity is the cosine similarity, γ is the preset maximum margin value, and α and β are adaptively optimized margin parameters.

[0048] In this embodiment, the adaptive optimization based on cosine similarity can be called: S-ALMFT (similarity-based adaptive large margin fine-tuning).

[0049] Psychological research demonstrates that people should set challenging goals, but it's necessary to consider the complexity of the task to avoid excessive stress. Neural networks function much like a mimicry of the human brain; therefore, the margin in the loss function can be seen as a mapping of the model's challenge, and it should also satisfy a negative correlation with training difficulty. In the ASV system, training difficulty can be reflected in two ways. On one hand, using longer speech segments provides more information about the speaker, reducing the difficulty of speaker classification. On the other hand, the cosine function in AAM-softmax calculates the similarity between the current embedding and the target center, which is also a measure of classification difficulty. Higher similarity indicates the ability to extract more robust embeddings, so the margin should be increased to enhance the difficulty, and vice versa. To represent the positive correlation between margin and cosine similarity, α and β in the formula Margin = min(αexp(β*Similarity), γ) can be obtained by fitting the margin range and speech duration. When the cosine similarity approaches 1, the margin becomes very high. To avoid this, this method sets a maximum margin γ to keep the margin within a reasonable range. The duration of training speech is used as a bridge to find the correspondence between margin and cosine similarity.

[0050] For step S13, as Figure 2 As shown, sample 2s or 6s long speech samples are fed into the speaker verification model, and the mean cosine similarity between all embeddings and target classification weights is calculated. The obtained mean similarity corresponds to the duration of the input speech, thus determining the corresponding optimal margin. The speaker verification model is trained using the margin of the loss function determined by existing mathematical software.

[0051] As can be seen from this implementation method, by using training speech of varying lengths to better simulate real-life scenarios, and by adaptively optimizing and fine-tuning the margin, adjusting the margin according to the duration and similarity of each speech, this method can achieve good speaker verification performance for speech of different durations in real-world scenarios.

[0052] Experiments were conducted using the VoxCeleb1 dataset, a large-scale audiovisual speaker recognition dataset extracted from YouTube videos. The VoxCeleb1 development set was used for training in all experiments. In addition to VoxCeleb1, three other datasets were built upon it for evaluation.

[0053] 1. The VoxCeleb1 dataset is a large-scale audiovisual dataset containing at least 3 seconds of speech, with an average of 8 seconds.

[0054] 2. The fixed-length VoxCeleb1 dataset consists of six subsets. Each subset contains fixed-length speech samples, ranging from 1 second to 6 seconds, to analyze the system's performance under different speech lengths.

[0055] 3. The variable-length VoxCeleb1 dataset was used to validate the model's ability to extract embeddings and its robustness to speech duration. The duration of the registration and test speech was randomly selected from 1 second to 6 seconds.

[0056] 4. Asymmetric datasets reflect real-world scenarios. Registration uses the full-length speech, but testing involves truncating 1 or 2 seconds of audio. Users tend to prefer recording longer audio clips. However, in real-world verification, convenience is crucial, requiring speaker verification systems to react quickly.

[0057] To investigate the effectiveness of this method in adaptive margin optimization fine-tuning, such as Figure 3 As shown, experiments were conducted using LMFT and the proposed D-ALMFT and S-ALMFT strategies. Cosine distance was applied and equal error rate (EER) was used to evaluate performance. It can be observed that the D-ALMFT and S-ALMFT strategies helped the baseline system achieve improvements of approximately 26.2%, 17.3%, and 17.6% in the O, E, and H trials of VoxCeleb1, respectively, which is comparable to the LMFT system, demonstrating that this method is also very practical for ASV models in the VoxSRC2020 competition.

[0058] VoxCeleb is a large-scale audiovisual dataset, but it is not an ideal evaluation dataset for real-world ASV scenarios for the following reasons: First, the average speech length in VoxCeleb is 8 seconds, and it is difficult to encounter such long test data in real-life scenarios, such as smart wake-up and identity verification. Second, the lengths of registration and test speech are often asymmetrical. Due to environmental and time constraints, test utterances are usually relatively short.

[0059] To validate the method in more real-world scenarios, three additional datasets were generated. To clearly observe system performance over different durations, a fixed-length dataset based on VoxCeleb1 was constructed. This fixed-length dataset demonstrates system performance over different durations. Figure 3As shown, the system employing the LMFT method suffers a 33% and 16% performance degradation in 1-second and 2-second scenarios, respectively, despite its success on the VoxCeleb1 dataset by leveraging the advantage of data matching. Similar to LMFT, the ALMFT method improves model performance as data duration increases. This benefits from duration matching and enhanced intra-speaker compactness, providing greater margin for longer speech. Meanwhile, ALMFT avoids severe system degradation in short-duration scenarios, although performance decreases slightly at 1 second. The longer training data (longer than 1 second) inevitably impacts performance at 1 second, but other test environments remain unaffected.

[0060] Using fixed-length datasets to measure the performance of real-world ASV applications is incomplete. Therefore, this method also designs variable-length datasets to investigate the robustness of the speaker validation model to data of different lengths. Figure * Table 1 shows the limitations of the LMFT method in the variable-length case, with its performance decreasing by nearly 12%. However, both the D-ALMFT and S-ALMFT methods overcome this problem and even achieve slight improvements.

[0061] If the ASV system is robust to speech of different lengths, then for the same person, the embeddings extracted from short speech should be very similar to those extracted from long speech. Figure 4 The leftmost section (a) compares the average scores of speaker embeddings extracted from speech of different lengths for target pairs (registered and test speech from the same speaker) based on the variable-length dataset E. The scores decrease as the duration of the registered or test speech decreases, which is reflected in the color change of the diagonal lines. However, all scores are less than 0.61, indicating that there is still significant room for improvement in the system's robustness to duration. Figure 4 The rightmost section (b) shows the relative score changes between the baseline system and either the LMFT or S-ALMFT system. The red portion in the upper left corner of the LMFT system represents its poor ability to extract short speech embeddings. The green portion of the S-ALMFT system shows the improvement on variable-length datasets. This further demonstrates that our method addresses the inherent limitations of LMFT and achieves a performance improvement in robustness to duration when extracting embeddings.

[0062] In real-world scenarios, the speaker verification registration process requires speakers to input relatively long speech entries, while speakers tend to use shorter utterances for verification to ensure convenience. The information asymmetry provided by the registration and testing speech samples poses a challenge for ASV tasks. To simulate this asymmetric situation, our method is evaluated on an asymmetric dataset. Figure 3The performance of the LMFT method shown is unsatisfactory, with a 13.8% reduction in test E. However, the ALMFT method achieved performance improvements of 4.5%, 4.1%, and 6.7% over the baseline system in tests O, E, and H, respectively.

[0063] Figure 5 The score distributions of target pairs and non-target pairs and their corresponding standard deviations were compared among different models. Figure 5 In the figure, (a) represents the difference between the average scores of target pairs and non-target pairs, which also represents the inter-class distance for different speakers. It can be observed that D-ALMFT increases this difference, while LMFT and S-ALMFT slightly reduce the inter-class distance for speakers. Figure 5 (b) shows the standard deviation of the target pair, indicating the compactness within the speaking human. Both D-ALMFT and S-ALMFT enhance this compactness, while LMFT reduces it. Figure 5 (c) in the table compares the standard deviations of non-target pairs, and all three algorithms improve the ability to identify non-target speakers. Although S-ALMFT does not further expand the separability between speaker groups, it can be seen that the method still outperforms the baseline system due to the reduction in intra-speaker differences.

[0064] In summary, while addressing the shortcomings of existing LMFT methods, this approach emphasizes the necessity of co-variation between training speech and training penalties. We propose a fine-tuning method for adaptive large margin model optimization, adjusting the margin based on the duration of each speech sample and its similarity to its class center. In addition to VoxCeleb1, we construct fixed-length, variable-length, and asymmetric datasets to better simulate real-life scenarios. Finally, our ALMFT achieves comparable performance to LMFT on VoxCeleb1, without performance degradation in other scenarios, and even outperforms them.

[0065] like Figure 6 The diagram shown is a schematic diagram of an adaptive margin optimization system based on speaker verification provided in an embodiment of the present invention. The system can execute the adaptive margin optimization method based on speaker verification described in any of the above embodiments and is configured in the terminal.

[0066] This embodiment provides an adaptive margin optimization system 10 based on speaker verification, which includes: a loss function determination module 11, an adaptive optimization module 12, and a training module 13.

[0067] The loss function determination module 11 is used to input speech training data including multiple speech durations into the speaker verification model and determine the loss function of the speaker verification model; the adaptive optimization module 12 is used to adaptively optimize the margin parameter of the loss function based on the speech of each speech duration in the speech training data and the preset target margin for each speech duration, wherein the adaptive optimization includes: duration-based adaptive optimization and cosine similarity-based adaptive optimization; the training module 13 is used to train the speaker verification model by determining the acceptable training difficulty of the speaker verification model using the margin parameter of the adaptively optimized loss function, wherein the training difficulty is related to the speech duration and cosine similarity.

[0068] This invention also provides a non-volatile computer storage medium storing computer-executable instructions that can execute the speaker verification-based adaptive margin optimization method in any of the above method embodiments.

[0069] In one embodiment, the non-volatile computer storage medium of the present invention stores computer-executable instructions, which are configured as follows:

[0070] Speech training data including various speech durations are input into the speaker verification model, and the loss function of the speaker verification model is determined.

[0071] Based on the speech of each speech duration in the speech training data, and the target margin preset for each speech duration, the margin parameter of the loss function is adaptively optimized, wherein the adaptive optimization includes: duration-based adaptive optimization and cosine similarity-based adaptive optimization.

[0072] The speaker verification model is trained by determining the acceptable training difficulty of the speaker verification model using the margin parameter of the adaptively optimized loss function, wherein the training difficulty is related to the speech duration and cosine similarity.

[0073] As a non-volatile computer-readable storage medium, it can be used to store non-volatile software programs, non-volatile computer-executable programs, and modules, such as the program instructions / modules corresponding to the methods in the embodiments of the present invention. One or more program instructions are stored in the non-volatile computer-readable storage medium, and when executed by a processor, the adaptive margin optimization method based on speaker verification in any of the above method embodiments is executed.

[0074] Figure 7 This is a schematic diagram of the hardware structure of an electronic device based on an adaptive margin optimization method for speaker verification, provided in another embodiment of this application. Figure 7 As shown, the device includes:

[0075] One or more processors 710 and memory 720, Figure 7 Taking a processor 710 as an example, the device based on the adaptive margin optimization method for speaker verification may further include an input device 730 and an output device 740.

[0076] The processor 710, memory 720, input device 730, and output device 740 can be connected via a bus or other means. Figure 7 Taking the example of a connection between China and Israel via a bus.

[0077] The memory 720, as a non-volatile computer-readable storage medium, can be used to store non-volatile software programs, non-volatile computer-executable programs, and modules, such as the program instructions / modules corresponding to the speaker verification-based adaptive margin optimization method in the embodiments of this application. The processor 710 executes various server functions and data processing by running the non-volatile software programs, instructions, and modules stored in the memory 720, thereby implementing the speaker verification-based adaptive margin optimization method described in the above embodiments.

[0078] The memory 720 may include a program storage area and a data storage area, wherein the program storage area may store the operating system and applications required for at least one function; the data storage area may store data, etc. Furthermore, the memory 720 may include high-speed random access memory, and may also include non-volatile memory, such as at least one disk storage device, flash memory device, or other non-volatile solid-state storage device. In some embodiments, the memory 720 may optionally include memory remotely located relative to the processor 710, and these remote memories can be connected to the mobile device via a network. Examples of such networks include, but are not limited to, the Internet, corporate intranets, local area networks, mobile communication networks, and combinations thereof.

[0079] Input device 730 can receive input numerical or character information. Output device 740 may include display devices such as a display screen.

[0080] The one or more modules are stored in the memory 720, and when executed by the one or more processors 710, they execute the adaptive margin optimization method based on speaker verification in any of the above method embodiments.

[0081] The above-described product can perform the methods provided in the embodiments of this application, and has the corresponding functional modules and beneficial effects for performing the methods. Technical details not described in detail in this embodiment can be found in the methods provided in the embodiments of this application.

[0082] Non-volatile computer-readable storage media may include a stored program area and a stored data area, wherein the stored program area may store an operating system and an application program required for at least one function; the stored data area may store data created based on the use of the device, etc. Furthermore, the non-volatile computer-readable storage medium may include high-speed random access memory, and may also include non-volatile memory, such as at least one disk storage device, flash memory device, or other non-volatile solid-state storage device. In some embodiments, the non-volatile computer-readable storage medium may optionally include memory remotely located relative to the processor, and these remote memories may be connected to the device via a network. Examples of such networks include, but are not limited to, the Internet, corporate intranets, local area networks, mobile communication networks, and combinations thereof.

[0083] This invention also provides an electronic device comprising: at least one processor, and a memory communicatively connected to the at least one processor, wherein the memory stores instructions executable by the at least one processor, the instructions being executed by the at least one processor to enable the at least one processor to perform the steps of the speaker verification-based adaptive margin optimization method of any embodiment of this invention.

[0084] The electronic devices described in this application exist in various forms, including but not limited to:

[0085] (1) Mobile communication devices: These devices are characterized by their mobile communication capabilities and primarily aim to provide voice and data communication. These terminals include smartphones, multimedia phones, feature phones, and low-end phones.

[0086] (2) Ultra-mobile personal computer devices: These devices fall under the category of personal computers, possessing computing and processing capabilities, and generally also have mobile internet access features. These terminals include PDAs, MIDs, and UMPCs, such as tablet computers.

[0087] (3) Portable entertainment devices: These devices can display and play multimedia content. This category includes audio and video players, handheld game consoles, e-book readers, as well as smart toys and portable car navigation devices.

[0088] (4) Other electronic devices with data processing functions.

[0089] In this document, relational terms such as "first" and "second" are used merely to distinguish one entity or operation from another, without necessarily requiring or implying any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising" or "including" include not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Unless otherwise specified, an element defined by the phrase "comprising..." does not exclude the presence of additional identical elements in the process, method, article, or apparatus that includes said element.

[0090] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate. 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 modules can be selected to achieve the purpose of this embodiment according to actual needs. Those skilled in the art can understand and implement this without any creative effort.

[0091] Through the above description of the embodiments, those skilled in the art can clearly understand that each embodiment can be implemented by means of software plus necessary general-purpose hardware platforms, and of course, it can also be implemented by hardware. Based on this understanding, the above technical solutions, in essence or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods described in the various embodiments or some parts of the embodiments.

[0092] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. An adaptive margin optimization method based on speaker verification, comprising: Speech training data including various speech durations are input into the speaker verification model, and the loss function of the speaker verification model is determined. Based on the speech samples of each duration in the speech training data, and the preset target margin for each speech sample duration, the margin parameters of the loss function are adaptively optimized. The adaptive optimization includes: duration-based adaptive optimization and cosine similarity-based adaptive optimization, including: When the adaptive optimization is duration-based, the linear relationship between the target margin and the speech duration is expressed as: Margin = A * Duration + B, where Margin is the target margin of the loss function, Duration is the speech duration, and A and B are the margin parameters for adaptive optimization. When the adaptive optimization is based on cosine similarity, the exponential function of the relationship between the target margin and cosine similarity is: Margin = min(αexp(β* Similarity),γ), where Margin is the target margin of the loss function, Similarity is the cosine similarity, γ is the preset maximum margin value, and α and β are the margin parameters of the adaptive optimization. The speaker verification model is trained by determining the acceptable training difficulty of the speaker verification model using the margin parameter of the adaptively optimized loss function, wherein the training difficulty is related to the speech duration and cosine similarity.

2. The method according to claim 1, wherein, The cosine similarity is determined by the speaker embedding predicted by the speaker verification model and the preset target speaker embedding.

3. An adaptive margin optimization system based on speaker verification, comprising: The loss function determination module is used to input speech training data including multiple speech durations into the speaker verification model and determine the loss function of the speaker verification model. An adaptive optimization module is used to adaptively optimize the margin parameters of the loss function based on the speech durations in the speech training data and the preset target margins for the speech durations. The adaptive optimization includes: duration-based adaptive optimization and cosine similarity-based adaptive optimization. When the adaptive optimization is duration-based, the linear relationship between the target margin and the speech duration is expressed as: Margin = A * Duration + B, where Margin is the target margin of the loss function, Duration is the speech duration, and A and B are the margin parameters for adaptive optimization. When the adaptive optimization is based on cosine similarity, the exponential function of the relationship between the target margin and cosine similarity is: Margin = min(αexp(β* Similarity),γ), where Margin is the target margin of the loss function, Similarity is the cosine similarity, γ is the preset maximum margin value, and α and β are the margin parameters of the adaptive optimization. The training module is used to train the speaker verification model by determining the acceptable training difficulty of the speaker verification model using the margin parameter of the adaptively optimized loss function, wherein the training difficulty is related to the speech duration and cosine similarity.

4. The system according to claim 3, wherein, The cosine similarity is determined by the speaker embedding predicted by the speaker verification model and the preset target speaker embedding.

5. An electronic device comprising: At least one processor, and a memory communicatively connected to the at least one processor, wherein the memory stores instructions executable by the at least one processor to enable the at least one processor to perform the steps of the method according to any one of claims 1-2.

6. A storage medium having a computer program stored thereon, characterized in that, When executed by a processor, the program implements the steps of the method described in any one of claims 1-2.

Images