METHOD, ELECTRONIC DEVICE, COMPUTER PROGRAM PRODUCT FOR GENERATING A Cross-MODAL ENCODER

By generating a cross-modal encoder, combining video, audio and text features, the problem that traditional speech cloning technology is difficult to capture speech emotions and context is solved, and the generation of speech with specific emotions and styles is realized, and the multimodality of speech cloning is enhanced Expression ability.

CN118918876APending Publication Date: 2024-11-08DELL PROD LP
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202310506665.9
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2023-05-06
Publication Date
2024-11-08

AI Technical Summary

Technical Problem

Traditional speech cloning technology is difficult to effectively solve the Video to Text-to-Clip task, and fails to fully consider the emotions and contexts contained in the new speech, resulting in a lack of emotional and stylistic diversity in generated speech.

Method used

By generating a cross-modal encoder, features of video, audio, and text are extracted, combined and scrambled operations, cross-modal sample features are generated, and tuned by the encoder to generate speech with specific emotions and styles.

Benefits of technology

The conversion of text into speech with specified audio and video emotions is implemented, and the generated speech has the same sound style and emotion as reference video and audio, suitable for the understanding and expression of multimodal data.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN118918876A_ABST
    Figure CN118918876A_ABST
Patent Text Reader

Abstract

The embodiment of the invention relates to a method for generating a cross-modal encoder. The method comprises the following steps: extracting sample video features of a sample video, sample audio features of a sample audio and sample text features of a sample text; the method further includes combining, via a cross-modal encoder, the sample video feature, the sample audio feature, and the sample text feature to generate a first cross-modal sample feature, perturbing the cross-modal feature to generate a perturbed cross-modal sample feature, and outputting the perturbed cross-modal sample feature. The cross-modal encoder encodes the disturbed cross-modal sample feature to generate a second cross-modal sample feature, and adjusts the cross-modal encoder based on the first cross-modal sample feature and the second cross-modal sample feature. By means of the method, the text can be converted into the voice with the needed emotion specified by the reference audio and the reference video, the cross-modal data serves as input, and the generated voice has the same sound but has the emotion different from that of the reference video.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] Embodiments of the present disclosure relate to the field of cross-modal coding, and more particularly, to a method, an electronic device, and a computer program product for generating a cross-modal encoder. Background Art

[0002] Cross-modal voice cloning technology is a technology that combines visual information with voice information. It can combine voice features with those in the video to generate voice corresponding to the video content. Using cross-modal voice cloning technology, artificial intelligence algorithms can be used to learn the relationship between human voice and facial expressions, so that the processing model can automatically capture the correlation between voice and facial expressions. Summary of the invention

[0003] Embodiments of the present disclosure relate to methods, electronic devices, and computer program products for generating a cross-modal encoder.

[0004] According to a first aspect of the present disclosure, a method for generating a cross-modal encoder is provided. The method includes extracting sample video features of a sample video, sample audio features of a sample audio, and sample text features of a sample text, wherein the contents of the sample video, the sample audio, and the sample text are associated. And via a cross-modal encoder, the sample video features, the sample audio features, and the sample text features are combined to generate a first cross-modal sample feature, and the sample video features, the sample audio features, and the sample text features are disturbed to generate a disturbed cross-modal sample feature. The disturbed cross-modal sample features are encoded via a cross-modal encoder to generate a second cross-modal sample feature, and the cross-modal encoder is adjusted based on the first cross-modal sample feature and the second cross-modal sample feature.

[0005] According to a second aspect of the present disclosure, an electronic device for generating a cross-modal encoder is provided, comprising at least one processor, and a memory, coupled to the at least one processor and having instructions stored thereon, the instructions causing the electronic device to perform actions when executed by the at least one processor, the actions comprising extracting sample video features of a sample video, sample audio features of a sample audio, and sample text features of a sample text, wherein the contents of the sample video, the sample audio, and the sample text are associated. And via a cross-modal encoder, the sample video features, the sample audio features, and the sample text features are combined to generate a first cross-modal sample feature, and the sample video features, the sample audio features, and the sample text features are disturbed to generate a disturbed cross-modal sample feature. The disturbed cross-modal sample features are encoded via the cross-modal encoder to generate a second cross-modal sample feature, and the cross-modal encoder is adjusted based on the first cross-modal sample feature and the second cross-modal sample feature.

[0006] According to a third aspect of the present disclosure, there is provided a computer program product, which is tangibly stored on a non-volatile computer-readable medium and includes machine-executable instructions, which when executed cause a machine to perform the steps of the method implemented in the first aspect of the present disclosure. BRIEF DESCRIPTION OF THE DRAWINGS

[0007] The above and other objects, features and advantages of the present disclosure will become more apparent through a more detailed description of exemplary embodiments of the present disclosure in conjunction with the accompanying drawings, wherein like reference numerals generally represent like elements throughout the exemplary embodiments of the present disclosure.

[0008] Figure 1 FIG. 1 illustrates an architecture diagram of a video-assisted voice cloning system according to an embodiment of the present disclosure;

[0009] Figure 2 A flowchart of a method for generating a cross-modal encoder according to an embodiment of the present disclosure is illustrated;

[0010] Figure 3 A schematic diagram for generating a cross-modal encoder according to an embodiment of the present disclosure is illustrated;

[0011] Figure 4 A schematic diagram of performing an XOR operation on sample feature data is illustrated;

[0012] Figure 5 A schematic diagram illustrating another method for generating a cross-modal encoder according to an embodiment of the present disclosure is illustrated;

[0013] Figure 6 A schematic diagram of a model architecture for cross-modal understanding using a cross-modal encoder according to an embodiment of the present disclosure is illustrated;

[0014] Figure 7 A schematic diagram of providing virtual customer service using a cross-modal encoder according to an embodiment of the present disclosure is illustrated;

[0015] Figure 8 A schematic diagram illustrating a virtual image reconstruction process associated with a cross-modal encoder according to an embodiment of the present disclosure is illustrated;

[0016] Fig. 9A A schematic diagram of a process of using a universal voice cloning according to an embodiment of the present disclosure is illustrated;

[0017] Fig. 9B A schematic diagram of a video-assisted voice cloning process according to an embodiment of the present disclosure is illustrated;

[0018] Fig.10 FIG. 1 illustrates an experimental result diagram of speaker classification using video-assisted voice cloning according to an embodiment of the present disclosure;

[0019] Fig.11 FIG. 1 illustrates an experimental result diagram of voice quality generated using video-assisted voice cloning according to an embodiment of the present disclosure;

[0020] Fig.12 A schematic block diagram of an example device suitable for implementing embodiments of the present disclosure is illustrated. DETAILED DESCRIPTION

[0021] Embodiments of the present disclosure will be described in more detail below with reference to the accompanying drawings. Although certain embodiments of the present disclosure are shown in the accompanying drawings, it should be understood that the present disclosure can be implemented in various forms and should not be construed as being limited to the embodiments described herein, which are instead provided for a more thorough and complete understanding of the present disclosure. It should be understood that the drawings and embodiments of the present disclosure are only for exemplary purposes and are not intended to limit the scope of protection of the present disclosure.

[0022] In the description of the embodiments of the present disclosure, the term "including" and similar terms should be understood as open inclusion, that is, "including but not limited to". The term "based on" should be understood as "based at least in part on". The term "one embodiment" or "the embodiment" should be understood as "at least one embodiment". The terms "first", "second", etc. may refer to different or the same objects. Other explicit and implicit definitions may also be included below.

[0023] Traditional voice cloning (VC) methods cannot solve the V2C (Video to Text-to-Clip) task well because they only focus on how to convert the input text into speech with the voice / tone shown in the reference audio, which converts text to speech only with the help of the reference audio. Existing VC tasks aim to convert paragraph text into speech using the expected audio specified by the reference audio, without considering the emotions and context contained in the new speech.

[0024] At least to solve the above and other potential problems, an embodiment of the present disclosure provides a method for generating a cross-modal encoder, the method comprising extracting sample video features of a sample video, sample audio features of a sample audio, and sample text features of a sample text, wherein the contents of the sample video, the sample audio, and the sample text are associated. And via a cross-modal encoder, the sample video features, the sample audio features, and the sample text features are combined to generate a first cross-modal sample feature, and the sample video features, the sample audio features, and the sample text features are disturbed to generate a disturbed cross-modal sample feature. The disturbed cross-modal sample features are encoded via a cross-modal encoder to generate a second cross-modal sample feature, and the cross-modal encoder is adjusted based on the first cross-modal sample feature and the second cross-modal sample feature. By using this method, text can be converted into speech with the desired emotion specified by the reference audio and the desired emotion specified by the reference video, and a triple (text / subtitle, reference audio, reference video) is used as input, and it is expected that the generated speech has the same sound, but has a different emotion from the reference video.

[0025] The basic principles and several exemplary embodiments of the present disclosure will be described in detail below with reference to the accompanying drawings. Figure 1 FIG. 1 illustrates an architecture diagram of a video-aided voice cloning (VaVC) system 100 according to an embodiment of the present disclosure. It should be understood that Figure 1 The numbers and arrangements of components, elements, and systems shown are merely examples, and the architecture diagram may include different numbers and arrangements of components, elements, and systems.

[0026] like Figure 1 As shown, the video-assisted speech cloning system 100 may include elements and / or components such as a cross-modal encoder 101, a duration predictor 102, a pitch predictor 103, an energy predictor 104, a mel-spectrogram decoder 105, a transposed convolutional network 107, a multi-receptive field fusion (MRF) module 108, and a generated speech 109.

[0027] The video-assisted voice cloning system 100 can generate the desired voice 109 based on the reference video 110, the reference audio 115, and the text 120. Specifically, the cross-modal encoder (converter) 101 in the video-assisted voice cloning system 100 can convert the triple data formed by the reference video 110, the reference audio 115, and the text 120 into a desired voice 109. Encode to generate a cross-modal data set The reference video 110 and the reference audio 115 can be used to guide or define the video-assisted voice cloning system 100 to generate voices associated with its content description and / or context. The text 120 can correspond to the generated voice 109 .

[0028] For example, in some embodiments, after receiving the reference video 110 and the reference audio 115 and processing them, the video-assisted voice cloning system 100 can finally generate a voice 109 that is associated with or similar to the style, emotion, mood, etc. contained in the reference video 110 and the reference audio 115. For example, when the content of the reference video 110 and the reference audio 115 is a very sad person speaking, the voice finally generated can also have a very sad tone and intonation, even if the context of the text 120 corresponding to the generated voice 109 is a scene where a happy person is giving an award speech. In some embodiments, when the reference video 110 and the reference audio 115 are a person skiing, the generated voice 109 can also include the corresponding sound of sliding across the snow and the sound of the wind. The above examples are for reference only, and the present disclosure does not impose any limitation on this.

[0029] Cross-modal data collection It can be expressed as Where L represents the length of the input sentence in the reference video 110 and the reference audio 115 (i.e., the length of the phoneme), and f(·) represents the processing by the cross-modal encoder 101. Cross-modal data set It may include various elements or features from data such as reference video 110, reference audio 115, text 120, etc., such as phoneme features for characterizing the pronunciation position, pronunciation method, vocal cord vibration, airflow direction, etc., pitch features for characterizing the basic frequency, basic period, zero crossing rate, timbre, intonation, emotion, etc. of the sound, and volume features and prosody features for characterizing the size, loudness, rhythm, melody, time domain distribution, stress, and tone of the speech. It should be understood that the cross-modal data set in the present disclosure is The features are only used for exemplary purposes and are not used to limit the protection scope of the present disclosure.

[0030] Cross-modal data sets or conditional phoneme embeddings The mel-spectrogram 106 may then be fed into a synthesizer (not shown) in the video-assisted voice cloning system 100, which may then process the cross-modal data set to generate a mel-spectrogram 106 associated with the cross-modal data set. The mel-spectrogram 106 may include a plurality of mel-spectrogram frames. Wherein T represents the number of Mel-spectrogram frames. As an example, in the present disclosure, the synthesizer may include a duration predictor 102, a pitch predictor 103, an energy predictor 104, and a Mel-spectrogram decoder 105. It should be understood that the elements in the synthesizer in the present disclosure are only used for exemplary purposes and are not intended to limit the scope of protection of the present disclosure. The synthesizer may include any known or unknown elements or components.

[0031] In an embodiment of the present disclosure, the duration predictor 102 is used to solve the cross-modal data set. Frame with Mel Spectrogram The problem of length mismatch between (ie, L≠T). Duration predictor 102S d Cross-modal data can be collected As input, and predict the duration of each phoneme feature The duration d of the i-th phoneme i Indicates the i-th phoneme feature x i Then the length controller (LengthRegulator, )103 You can use the predicted duration To adjust the length of the Mel-spectrogram 106 to be generated, the process can be expressed as

[0032]

[0033] in is an extended phoneme feature with length T.

[0034] Cross-modal data collection Predicted duration Extended phoneme features Such data can then be fed to the pitch predictor 103S p The pitch spectrum can represent the trajectory of the fundamental frequency change of the sound on the time axis to reflect the contour and change of the audio melody. The data generated above can also be fed to the energy predictor 104S. e And calculate the L2 norm of the amplitude of each short-time Fourier transform (STFT) frame in the data as the predicted energy value, thereby predicting the prosody and volume of the generated speech.

[0035] The predicted pitch spectrum and predicted energy value can then be passed through the feature layers φ and The predicted pitch spectrum feature and the predicted energy value feature can then be added to the extended phoneme feature by adding operations 111-a and 111-b. The final Mel spectrum It can be generated by the following formula:

[0036]

[0037] Wherein g(·) represents a mel-spectrogram decoder 105 , which is composed of 6 fast Fourier transform (FFT) blocks and is used to convert the input feature data into a mel-spectrogram 106 .

[0038] In known solutions, the mel-spectrogram cannot be accurately generated, so in some embodiments, the duration predictor 102, the pitch predictor 103, and the energy predictor 104 can also be trained and adjusted to optimize the generated mel-spectrogram 106. In some embodiments, in order to optimize the duration predictor 102, the Montreal forced alignment (MFA) method can be used to obtain the reference truth phoneme duration sequence of the cross-modal training sample data, and then the predicted phoneme duration sequence generated by the duration predictor 102 is calculated. Phoneme duration sequence with ground truth The mean square error (MSE) loss between the two. If the MSE loss between the two is lower than the desired threshold, the duration predictor 102 can be further optimized and trained to minimize the MSE loss. The desired threshold can be set to any value according to actual needs. The process can be expressed by the following formula:

[0039]

[0040] In some embodiments, the pitch predictor 103 can also be trained, for example, a continuous wavelet transform (CWT) can be used to convert the continuous pitch of the cross-modal sample data into a pitch spectrum graph, and the pitch spectrum is used as a reference truth pitch spectrum to compare with the pitch value in the predicted pitch spectrum, and the mean variance loss between the two is calculated. If the MSE loss between the two is lower than the desired threshold, the pitch predictor 103 can be further optimized and trained to minimize the MSE loss. The desired threshold can be set to an arbitrary value according to actual needs. The process can be expressed by the following formula:

[0041]

[0042] in represents the tth benchmark truth value, p t =S p (x t ) represents the predicted pitch value.

[0043] Similarly as above, the energy predictor 104 can be trained, and the energy value of the cross-modal sample data can be used as a reference truth energy value to compare with the predicted energy value, and the mean variance loss between the two is calculated. If the MSE loss between the two is lower than the desired threshold, the energy predictor 104 can be further optimized and trained to minimize the MSE loss. The desired threshold can be set to any value according to actual needs. The process can be expressed by the following formula:

[0044]

[0045] in represents the tth benchmark truth value, e t =S e (x t ) represents the predicted energy value.

[0046] Similarly as above, the mel-spectrogram decoder 105 can also be trained, and the mel-spectrogram of the cross-modal sample data can be used as a reference truth value to compare with the predicted mel-spectrogram generated by the mel-spectrogram decoder, and the mean variance loss between the two can be calculated. If the MSE loss between the two is lower than the desired threshold, the mel-spectrogram decoder 105 can be further optimized and trained to minimize the MSE loss. The desired threshold can be set to any value according to actual needs. The process can be expressed by the following formula:

[0047]

[0048] in represents the t-th frame reference truth Mel spectrum graph, and Represents the predicted Mel-spectrogram of the t-th frame.

[0049] The loss function of the final synthesizer can be expressed by the following formula:

[0050]

[0051] in and denote the prediction losses of the above-mentioned Mel-spectrogram decoder 105, duration predictor 102, pitch predictor 103, and energy predictor 104, respectively. λ1, λ2, and λ3 are hyperparameters, and in practical applications can be set to λ1=λ2=λ3=1. The minimization is performed to make the generated mel-spectrogram 106 more accurate.

[0052] Generated Mel-spectrogram 106 It can then be fed into a speech encoder (vocoder, not shown) to be converted into a time domain waveform In the present disclosure, as an example, the speech encoder can be an audio processing model based on HiFi-GAN, which mainly focuses on generating original waveforms from Mel-spectrogram spectrograms through generative adversarial networks (GANs), which utilizes periodic patterns in speech signals to improve sample quality, and can perform conditional generation based on acoustic features to achieve studio-quality speech enhancement. It should be understood that the HiFi-GAN-based speech encoder in the present disclosure is only used for exemplary purposes and is not intended to limit the scope of protection of the present disclosure.

[0053] The speech encoder may include a transposed convolution (ConvTranspose) network module 107 and a multi-receptive field fusion (MRF) module 108. The transposed convolution network module 107 may first perform a mel-spectrogram 106 on the input. Upsampling is performed, which can align the length of the output feature with the time resolution of the original waveform. By convolving the input feature with a reversed convolution kernel, a smaller input feature can be expanded to a larger output feature, restoring it to its original size, and padding is used to control the size of the output feature.

[0054] The upsampled feature data can then be fed into the MRF module 108, which includes multiple residual blocks, and the sum of the outputs of these residual blocks is used as the final predicted waveform. In the present disclosure, different receptive fields can be ensured by using settings of residual blocks with different kernel sizes and expansion rates. In the present disclosure, the speech encoder can also be optimized by including an objective function based on LSGAN loss, Mel spectrogram loss, and feature matching loss. In actual application, the "universal" checkpoint provided in the HiFi-GAN official code can also be used without fine-tuning any synthesized Mel spectrogram 106. The speech 109 generated by the above process can have contexts such as voice, tone, emotion, etc. associated with the reference video 110 and the reference audio and has content corresponding to the text 120. It should be understood that the method for generating speech 109 described above is merely exemplary and is not intended to limit the scope of the present disclosure. If the reference video 101 is lost, the cross-modal encoder of the present disclosure can naturally encode more emotion-related information, with only text as input, because the modality mask process is in the pre-training stage. The method implemented according to the present disclosure can better realize the data properties of the encoder part, it can fully capture the connection information between different modalities, and introduces many metrics to make the generation process more file-oriented.

[0055] The above describes the generation method in combination with the duration predictor 102, pitch predictor 103, energy predictor 104 and their associated training processes of the present disclosure. However, if the cross-modal encoder 101 cannot be adjusted and trained, more accurate speech cannot be generated. Therefore, it is necessary to pre-train the cross-modal encoder to implement the method described in the present disclosure. Figure 2 hereinafter describes a process 200 for generating a cross-modal encoder.

[0056] At box 201, sample video features of the sample video, sample audio features of the sample audio, and sample text features of the sample text are extracted, wherein the contents of the sample video, the sample audio, and the sample text are associated. According to an embodiment of the present disclosure, a unimodal encoder such as a video encoder, an audio encoder, and a text encoder can respectively extract corresponding sample video features, sample audio features, and sample text features for sample videos, sample audio, and sample text corresponding to the same content. As an example, in an embodiment of the present disclosure, the video encoder can be an encoder based on Faster R_CNN, the audio encoder can be an audio encoder based on wave2vec 2.0, and the text encoder can be a text encoder based on BERT, and the present disclosure does not impose any restrictions on this. The sample to be extracted can be a movie clip containing subtitles, and the video encoder, audio encoder, and text encoder encode the video clip, audio clip, and subtitles therein, and extract corresponding sample video features, sample audio features, and sample text features.

[0057] At box 202, sample video features, sample audio features, and sample text features are combined via a cross-modal encoder to generate a first cross-modal sample feature. According to an embodiment of the present disclosure, as an example, a cross-modal encoder can be a deep learning model that can process information of multiple modalities (such as text, image, audio, etc.), which can use the correlation between different modalities to improve the representation ability and generalization ability. The cross-modal encoder is used to interact and fuse features of different modalities. The cross-modal encoder can adapt to different downstream tasks such as visual question answering, image retrieval, machine translation, etc. through pre-training and fine-tuning. Sample video features, sample audio features, and sample text features extracted by single-modal encoders such as video encoders, audio encoders, and text encoders can then be fed into the cross-modal encoder for fusion, thereby generating a first (original) cross-modal sample feature. Cross-modal features include features extracted from different data modalities (for example, images, text, audio, etc.) that can express the same or associated semantic information. Cross-modal features can be used to achieve matching, retrieval, generation, etc. between different modal data.

[0058] At block 203, a perturbation operation is performed on the sample video features, the sample audio features, and the sample text features to generate perturbed cross-modal sample features. According to an embodiment of the present disclosure, the video-assisted speech cloning system may perform a perturbation operation on the extracted sample video features, the sample audio features, and the sample text features, and the perturbation operation may include a random masking operation and an XOR operation, and finally generate perturbed cross-modal sample features.

[0059] The video-assisted speech cloning system may divide one or more of the sample video features, sample audio features, and sample text features into multiple parts. The random masking operation may include performing reasoning by randomly masking or covering one or more of the divided parts, so that the cross-modal encoder can predict the target output from incomplete information and ultimately improve the generalization ability of the cross-modal encoder.

[0060] The XOR operation may include randomly cross-joining multiple parts divided by the sample video features, the sample audio features, and the sample text features to form a new spliced ​​sample feature. For example, the first part may be selected from the sample video features, the second part may be selected from the sample audio features, and the third part may be selected from the sample text features, and these parts may be spliced ​​together to form a new spliced ​​sample feature, wherein the parts of the new spliced ​​sample feature are decoupled from each other, which helps the cross-modal encoder learn to better adapt to different input changes. For example, in the context of voice cloning, the target output after the cross-modal encoder processes the new spliced ​​sample feature may be the result of the voices of two speakers mixed together. The similarity or distance between features may be calculated by using the XOR operation. Additionally or alternatively, the random masking operation and the XOR operation may be performed simultaneously or sequentially, and the present disclosure does not impose any restrictions on this.

[0061] At block 204, the perturbed cross-modal sample features are encoded via a cross-modal encoder to generate second cross-modal sample features. The cross-modal encoder may then encode the perturbed cross-modal sample features to generate second (perturbed) modality sample features, and the various parts in the second modality sample features may be decoupled.

[0062] At box 205, based on the first cross-modal sample feature and the second cross-modal sample feature, the cross-modal encoder is adjusted. The cross-modal encoder can compare the generated first cross-modal sample feature with the second cross-modal sample feature. As an example, the cross-modal encoder can compare the KL (Kullback-Leible) divergence between the first cross-modal sample feature and the second cross-modal sample feature, wherein the KL divergence can be used to describe the degree of difference between the two features. If the KL divergence between the two is very small, it means that the difference between the first cross-modal sample feature and the second cross-modal sample feature is very small and belongs to the same description. For example, both the first cross-modal sample feature and the second cross-modal sample feature are video, audio, and text descriptions of the same movie content. On the contrary, if the KL divergence between the two is larger, it means that the difference between the first cross-modal sample feature and the second cross-modal sample feature is very large and belongs to different descriptions. For example, both the first cross-modal sample feature and the second cross-modal sample feature are video, audio, and text descriptions of completely different movie contents.

[0063] If the KL divergence between the first cross-modal sample feature and the second cross-modal sample feature is higher than a predetermined threshold, it means that the cross-modal encoder needs to continue training and adjusting the parameters in the cross-modal encoder, and the cross-modal encoder continues to be trained and adjusted. On the contrary, if the KL divergence between the first cross-modal sample feature and the second cross-modal sample feature is lower than a predetermined threshold, it means that the cross-modal encoder has met the needs of the user or the usage scenario, and the training and adjustment of the cross-modal encoder can be stopped.

[0064] Additionally or alternatively, training and adjusting the cross-modal encoder may include optimizing model parameters using a gradient descent method, such as continuously updating the model parameters along the negative gradient direction of the objective function until a minimum value or convergence is reached, including but not limited to a batch gradient descent method, a stochastic gradient descent method, a mini-batch gradient descent method, etc. As an example, in the present disclosure, the predetermined threshold value may be set according to user needs or the needs of the usage scenario, and the present disclosure does not impose any limitation on this.

[0065] Figure 3 FIG. 4 illustrates a schematic diagram for generating a cross-modal encoder 300 according to an embodiment of the present disclosure. Figure 3 As shown, sample video 301, sample audio 302, and sample text 303 corresponding to the same content or associated with the same content can be respectively encoded by a video encoder 304, an audio encoder 305, and a text encoder 306, and features can be extracted to obtain sample video features 307, sample audio features 308, and sample text features 309.

[0066] The extracted sample video features 307, sample audio features 308, and sample text features 309 can be input into a cross-modal encoder 310-a implemented according to an embodiment of the present disclosure to generate a first (original) cross-modal sample feature 312. The extracted sample video features 307, sample audio features 308, and sample text features 309 can also be perturbed 311 by the video-assisted speech cloning system, and finally perturbed cross-modal sample features 313 are generated, wherein the perturbation 311 can include a random masking operation and an XOR operation.

[0067] The perturbed cross-modal sample features 313 may be input into the cross-modal encoder 310-b to generate the second (perturbed) modality sample features 315, wherein there is no necessary correlation between the various parts in the second modality sample features 315. Additionally or alternatively, in some embodiments, the cross-modal encoder 310-a and the cross-modal encoder 310-b may be the same or different encoders, and the present disclosure does not impose any limitation on this.

[0068] Then the cross-modal encoder 310-a or the cross-modal encoder 310-b can calculate 314 and compare the KL divergence between the first cross-modal sample feature 312 and the second modality sample feature 315. If the KL divergence between the two is higher than a predetermined threshold, it means that the current cross-modal encoder 310-a or the cross-modal encoder 310-b does not meet the needs of the user or the usage scenario, and it is still necessary to adjust and optimize the various parameters and indicators in the cross-modal encoder 316, and further training is required. If the KL divergence between the two is lower than a predetermined threshold, it means that the cross-modal encoder has met the needs of the user or the usage scenario, and the adjustment and optimization of the cross-modal encoder can be stopped 316.

[0069] Figure 4 FIG. 4 is a schematic diagram of performing an XOR operation 400 on sample feature data. Figure 4 As shown, the sample video features 401, sample audio features 402, and sample text features 403 extracted via a single-modal encoder can be divided into multiple sample video feature parts 404, multiple sample audio feature parts 405, and multiple sample text feature parts 406 via a video-assisted speech cloning system.

[0070] The video-assisted speech cloning system can then randomly cross-splice or combine any one or more of the multiple sample video feature parts 404, the multiple sample audio feature parts 405, and the multiple sample text feature parts 406 to form a spliced ​​cross-modal sample feature. The spliced ​​cross-modal sample features can also be processed by a random masking operation and finally generate a disturbed cross-modal sample feature. Through the XOR operation, the features of different modalities can be fused into a common feature space, thereby reducing the differences between the modalities and avoiding information loss in traditional feature fusion methods.

[0071] Figure 5 FIG. 5 is a schematic diagram of another method 500 for generating a cross-modal encoder according to an embodiment of the present disclosure. Figure 5 As shown, the sample text features (or embeddings) T501 for the text channel, the sample video features (or embeddings) V502 for the video channel, and the sample audio features (or embeddings) A503 for the audio channel output from the unimodal encoder can be combined to form a link tensor for each channel (Sample-level modeling). The entries in this tensor are integers from 1 to 5. For example, if T 501, V 502, and A 503 are all linked together, then the entry (1,1,1) will be 1.

[0072] Then the tensor 504 can use CP decomposition (CANDECOMP / PARAFAC decomposition

[0073] Solution) method is decomposed into three vectors. The first vector The third vector Where I and J are any predetermined hyperparameters, and the core matrix 508, noise data 509. These three vectors 505, 506, 507 contain the denoised link information stored in the original tensor 504, and are the main vectors for text, audio, and video. The rank of vectors 505, 506, 507 is 1 and has different directions and is orthogonal to each other. The CP decomposition process is regarded as a transformation from the original space to the frequency space, which is also one of the basic functions of CP decomposition.

[0074] The core matrix 508 is a three-dimensional tensor that is formed by the outer product of three factor matrices. The factor matrix is ​​a matrix formed by merging the vectors of the same dimension that make up the single-rank tensor. Then, the noise data 509 can be discarded and the vectors 505, 506, 507 and the core matrix 508 can be used to reconstruct the original unimodal encoder output. The final reconstructed features can contain the denoised link information, which means it is a good embedding space for these three modalities. Specifically, the optimization process can be expressed by the following formula:

[0075]

[0076] This process can be understood as minimizing the characteristics of the reconstruction The difference between the sample features T, V, and A. A , μ T , μ V are the parameters of three MLP (Multi-layer Perceptron) networks whose outputs are the reconstructed embeddings, and A de is the decomposition vector for audio in the tensor decomposition part. Then the other channels are exactly the same as before. In the process of reconstructing the features to pre-train the cross-modal encoder, the existing token-level modeling strategy is used, but a new tensor-based modal link modeling method is proposed to replace the original modal-level modeling and sample-level modeling.

[0077] Figure 6 A schematic diagram of a model architecture 600 for cross-modal understanding using a cross-modal encoder according to an embodiment of the present disclosure is illustrated.

[0078] In previous techniques, the full-sensing pre-trainer simply connects three single-modal encoders with three outputs, and then inputs them into the cross-encoder to perform the task. This process is completely implicit and loses a lot of useful information. The tensor-based method disclosed in this disclosure makes up for the information loss and efficiency problems that occurred in previous techniques.

[0079] like Figure 6 As shown, the Omni-perception PreTrainer (OPT) 601 may include a unimodal encoder such as a text encoder 602, a video encoder 603, and an audio encoder 604, and a cross-modal encoder 605 trained according to the method disclosed herein. The Omni-perception PreTrainer 601 uses the cross-modal encoder 605 network structure to perform machine learning (NLP) on multimodal data, wherein a self-supervisory target is generated by applying a masking (“mask”) mechanism on the perceptual modality. For example, for text data 606, a portion of the text can be masked, and then components such as the cross-modal encoder 605 are trained to predict the masked language modeling 607.

[0080] Similar self-supervised learning strategies can also be applied to audio and video data. For example, for video data 609, a portion of the video can be masked, and then components such as the cross-modal encoder 605 can be trained to predict the masked video modeling 611. For example, for audio data 612, a portion of the audio can be masked, and then components such as the cross-modal encoder 605 can be trained to predict the masked audio modeling 614.

[0081] Then the associated decoder can generate corresponding data based on the predicted modal modeling. For example, the text encoder 615 can generate a complete sentence 616 "a person is skiing on a snowboard" based on the masked speech modeling 607. The video decoder 617 can generate a complete video 618 of a person skiing on a snowboard based on the masked video modeling 611. The audio decoder 619 can generate a corresponding complete audio 620 of a person skiing on a snowboard based on the masked video modeling 614.

[0082] The full-perceptual pre-trainer 601 includes token-level modeling, including masked language modeling (MLM) 607, masked video modeling (MVM) 611, and masked audio modeling (MAM) 614; (2) modality-level modeling, including denoised text reconstruction 608, denoised video reconstruction 610, and denoised audio reconstruction 613; (3) sample-level modeling, where "√" indicates that the corresponding modality is matched. For token-level modeling, a token-level masking mechanism is introduced, and for modality-level modeling and enabling any number of input modalities, a modality-level masking mechanism is introduced. In this way, the full-perceptual pre-trainer 601 can learn the associations between multimodal data and improve the understanding and expression capabilities of multimodal data.

[0083] Figure 7The diagram shows a schematic diagram of providing virtual customer service 700 using a cross-modal encoder according to an embodiment of the present disclosure. In the customer service system of the information service provider, the customer 701 can submit a voice service request to the customer service to request a tracking and resolution module when encountering problems when using the information system. The customer 701 often prefers manual service rather than intelligent service provided by a computer, but manual service is usually costly and time-consuming.

[0084] Therefore, in the virtual customer service terminal 702 that can use the video-assisted voice cloning system based on the embodiment of the present disclosure, it can clone and generate the voice of the artificial customer service. For example, the cross-modal encoder implemented according to the present disclosure can capture the artificial voice more clearly, and clone and generate the original artificial customer service voice as much as possible based on the captured voice data and the emotions, expressions and other data of the artificial customer service.

[0085] Therefore, when the manual customer service is busy or encounters a problem that cannot be solved at the moment, the manual customer service can seamlessly switch to the virtual customer service terminal 702, so that the customer experience of the customer 701 can be improved, which can also increase production efficiency and save costs.

[0086] Figure 8 FIG. 8 is a schematic diagram of a virtual image reconstruction process 800 associated with a cross-modal encoder according to an embodiment of the present disclosure. Figure 8 As shown, the reconstruction process 800 can use a given portrait video 801 and an image without people (background image 802) as input, and these data can be converted into a learnable code 804. The dynamics of the human face, i.e., facial expressions 803, are captured using a facial tracking method. The low-dimensional expression parameters 808 of the deformable model generated based on the facial expression are used as conditions for the neural scene representation network 806. The neural scene representation network 806 stores the dynamic neural radiance field 805 used during volume rendering. Data along the line of sight, such as the low-dimensional expression parameters 808, the learnable code 802, and the sampling data can be input to the dynamic neural radiance field 805. Since the background is static, the color of the last sampling point of each ray can be set to the corresponding value of the background image. Finally, based on the estimated posture and expression, volume rendering is used to synthesize the facial image 807.

[0087] However, a large number of current studies have shown that audio can also play an important role in expression capture, and text related to the reconstruction process can also contain information in expression reconstruction. Therefore, when combined with a cross-modal encoder implemented according to the present disclosure, the face tracking reconstruction process can be made more accurate by capturing more information from the real world.

[0088] Fig. 9A FIG. 9 is a schematic diagram of a general voice cloning process 900a according to an embodiment of the present disclosure. Fig. 9A As shown, the content of the text 901 corresponding to the speech to be generated and the sound of the reference audio 902 are input into the speech cloning system 903, so as to generate the target speech 904.

[0089] The current voice cloning system 903 is based on neural text-to-speech (TTS) technology, including WaveNet, Tacotron1 and Tacotron2. Most of these models use autoregression to generate mel-spectrograms. Generally, these models only support single-speaker systems. With the increasing demand for personalized speech generation, these TTS can be modeled and a feature learning and projection component is proposed to generate various voices of multiple speakers. The adaptability of TTS models to new speakers has been widely studied. The current method is to train the model on a large multi-speaker dataset and then fine-tune the entire model or only part of the model. However, they usually achieve low adaptive performance for unseen speakers, especially when the reference speech length is short.

[0090] Fig. 9B FIG. 9 is a schematic diagram of a video-assisted voice cloning process 900b according to an embodiment of the present disclosure. Fig. 9B As shown, the content of the corresponding text 905 of the desired speech, the sound of the reference audio 906, and the facial expressions of the characters in the reference video 907 are input into the video-assisted speech cloning system 908, thereby generating the target speech 909.

[0091] Given an input triple (i.e., subtitles / text, reference audio, and reference video), the video-assisted speech cloning system 908 can text Generating speech in 905 (ie, a waveform in the time domain), speech 909 contains the sound of reference audio 906 and the emotion of reference video 907. Reference audio 906 may only provide the expected speech, and its content may be irrelevant.

[0092] exist Fig. 9B In the example, movie dubbing can be used as an example. Given a movie clip (i.e., reference video 907), subtitles (i.e., text 905), and reference audio 906, a video-assisted voice cloning system 908 can synthesize the voice in the subtitles 909 based on the emotions of the characters in the movie and the sounds in the reference audio. This method can flexibly synthesize voices with learned speaker characteristics, and improve the adaptation performance of the video-assisted voice cloning system 908 to unseen speakers based on the improved linear projection.

[0093] Fig.10FIG1000 illustrates experimental results of speaker classification using video-assisted voice cloning according to an embodiment of the present disclosure. The voice cloning experiment was performed on the VCTK dataset using a video-assisted voice cloning system. VCTK includes audio samples of 108 native English speakers with different accents sampled at 48KHz. Fig.10 As shown, the ordinate represents the accuracy of speaker classification, the abscissa represents the number of adaptation iterations of the video-assisted voice cloning system, "whole model" represents the accuracy of speaker classification using the video-assisted voice cloning system, and "embedding" represents the accuracy of speaker classification using only the cross-modal encoder. The accuracy of speaker classification using the video-assisted voice cloning system is significantly higher than that of speaker classification using the cross-modal encoder. It can be found that the classification accuracy increases significantly with the increase in the number of samples, and the number of samples can reach up to 100. In the case of low sample numbers, adjusting speaker embeddings (features) is less likely to overfit the samples than adjusting the entire video-assisted voice cloning system. The two methods also require different numbers of iterations to converge. Compared with the overall video-assisted voice cloning system adaptation iterations (even 100 cloned audio samples will converge in about 1000 iterations), embedding adaptation requires significantly more iterations to converge, so the cloning time is much longer.

[0094] Fig.11 The experimental results 1100 of the use of video-assisted voice cloning to generate voice quality according to an embodiment of the present disclosure are illustrated. The text of the experiment is "If they knew I knew, they wouldn't let him stay", the speaker is Tim, and the video emotion is sad. Wherein graph 1101 represents the Mel spectrogram generated by FastSpeech2, graph 1102 represents the Mel spectrogram generated by the present disclosure implementation (V2C-Net), and graph 1103 represents the reference truth Mel spectrogram. It can be seen that compared with graph 1101, the energy (volume) curve and the fundamental frequency curve (i.e., F0 curve) of the Mel spectrogram 1102 generated by the present disclosure implementation are more similar to the reference truth Mel spectrogram. It is worth noting that since the duration of the audio should also be predicted, the length of the generated audio and the reference truth audio may be different. The ordinate represents the audio contour and audio energy, and the horizontal axis is the duration of the audio. The circles in the figure represent the main differences.

[0095] Fig.12 A schematic block diagram of an example device 1200 is shown that may be used to implement embodiments of the present disclosure. Figure 1The computing device in the device 1200 can be implemented by using the device 1200. As shown in the figure, the device 1200 includes a central processing unit (CPU) 1201, which can perform various appropriate actions and processes according to computer program instructions stored in a read-only memory (ROM) 1202 or computer program instructions loaded from a storage unit 1208 to a random access memory (RAM) 1203. In the RAM 1203, various programs and data required for the operation of the storage device 1200 can also be included. The CPU 1201, the ROM 1202, and the RAM 1203 are connected to each other via a bus 1204. An input / output (I / O) interface 1205 is also connected to the bus 1204.

[0096] A number of components in the device 1200 are connected to the I / O interface 1205, including: an input unit 1206, such as a keyboard, a mouse, etc.; an output unit 1207, such as various types of displays, speakers, etc.; a storage page 1208, such as a disk, an optical disk, etc.; and a communication unit 1209, such as a network card, a modem, a wireless communication transceiver, etc. The communication unit 1209 allows the device 1200 to exchange information / data with other devices through a computer network such as the Internet and / or various telecommunication networks.

[0097] The various processes and processing described above, such as method 200, may be performed by processing unit 1201. For example, in some embodiments, method 200 may be implemented as a computer software program, which is tangibly contained in a machine-readable medium, such as storage unit 1208. In some embodiments, part or all of the computer program may be loaded and / or installed on device 1200 via ROM 1202 and / or communication unit 1209. When the computer program is loaded into RAM 1203 and executed by CPU 1201, one or more actions of method 200 described above may be performed.

[0098] The present disclosure may be a method, an apparatus, a system and / or a computer program product. The computer program product may include a computer-readable storage medium carrying computer-readable program instructions for executing various aspects of the present disclosure.

[0099] Computer readable storage medium can be a tangible device that can hold and store instructions used by an instruction execution device. Computer readable storage medium can be, for example, but not limited to, an electrical storage device, a magnetic storage device, an optical storage device, an electromagnetic storage device, a semiconductor storage device, or any suitable combination of the above. More specific examples (non-exhaustive list) of computer readable storage medium include: portable computer disk, hard disk, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), static random access memory (SRAM), portable compact disk read-only memory (CD-ROM), digital versatile disk (DVD), memory stick, floppy disk, mechanical encoding device, such as a punch card or a convex structure in a groove on which instructions are stored, and any suitable combination of the above. The computer readable storage medium used here is not interpreted as a transient signal itself, such as a radio wave or other freely propagating electromagnetic wave, an electromagnetic wave propagated by a waveguide or other transmission medium (e.g., a light pulse by an optical fiber cable), or an electrical signal transmitted by a wire.

[0100] The computer-readable program instructions described herein can be downloaded from a computer-readable storage medium to each computing / processing device, or downloaded to an external computer or external storage device via a network, such as the Internet, a local area network, a wide area network, and / or a wireless network. The network can include copper transmission cables, optical fiber transmissions, wireless transmissions, routers, firewalls, switches, gateway computers, and / or edge servers. The network adapter card or network interface in each computing / processing device receives the computer-readable program instructions from the network and forwards the computer-readable program instructions for storage in the computer-readable storage medium in each computing / processing device.

[0101] The computer program instructions for performing the operation of the present disclosure may be assembly instructions, instruction set architecture (ISA) instructions, machine instructions, machine-related instructions, microcode, firmware instructions, state setting data, or source code or object code written in any combination of one or more programming languages, including object-oriented programming languages, such as Smalltalk, C++, etc., and conventional procedural programming languages, such as "C" language or similar programming languages. Computer-readable program instructions may be executed entirely on a user's computer, partially on a user's computer, as an independent software package, partially on a user's computer, partially on a remote computer, or completely on a remote computer or server. In the case of a remote computer, the remote computer may be connected to the user's computer via any type of network, including a local area network (LAN) or a wide area network (WAN), or may be connected to an external computer (e.g., using an Internet service provider to connect via the Internet). In some embodiments, an electronic circuit, such as a programmable logic circuit, a field programmable gate array (FPGA), or a programmable logic array (PLA), may be customized by utilizing the state information of the computer-readable program instructions, and the electronic circuit may execute the computer-readable program instructions, thereby realizing various aspects of the present disclosure.

[0102] Various aspects of the present disclosure are described herein with reference to the flowcharts and / or block diagrams of the methods, devices (systems) and computer program products according to the embodiments of the present disclosure. It should be understood that each box in the flowchart and / or block diagram and the combination of each box in the flowchart and / or block diagram can be implemented by computer-readable program instructions.

[0103] These computer-readable program instructions can be provided to a processing unit of a general-purpose computer, a special-purpose computer, or other programmable data processing device, thereby producing a machine, so that when these instructions are executed by the processing unit of the computer or other programmable data processing device, a device that implements the functions / actions specified in one or more boxes in the flowchart and / or block diagram is generated. These computer-readable program instructions can also be stored in a computer-readable storage medium, and these instructions cause the computer, programmable data processing device, and / or other equipment to work in a specific manner, so that the computer-readable medium storing the instructions includes a manufactured product, which includes instructions for implementing various aspects of the functions / actions specified in one or more boxes in the flowchart and / or block diagram.

[0104] Computer-readable program instructions may also be loaded onto a computer, other programmable data processing apparatus, or other device so that a series of operating steps are performed on the computer, other programmable data processing apparatus, or other device to produce a computer-implemented process, thereby causing the instructions executed on the computer, other programmable data processing apparatus, or other device to implement the functions / actions specified in one or more boxes in the flowchart and / or block diagram.

[0105] The flow chart and block diagram in the accompanying drawings show the possible architecture, function and operation of the system, method and computer program product according to multiple embodiments of the present disclosure. In this regard, each square box in the flow chart or block diagram can represent a part of a module, program segment or instruction, and a part of a module, program segment or instruction includes one or more executable instructions for realizing the specified logical function. In some alternative implementations, the functions marked in the square box can also occur in a sequence different from that marked in the accompanying drawings. For example, two continuous square boxes can actually be executed substantially in parallel, and they can sometimes be executed in reverse order, depending on the functions involved. It should also be noted that each square box in the block diagram and / or flow chart, and the combination of the square boxes in the block diagram and / or flow chart can be implemented with a dedicated hardware-based system that performs the specified function or action, or can be implemented with a combination of special hardware and computer instructions.

[0106] The embodiments of the present disclosure have been described above, and the above description is exemplary, not exhaustive, and is not limited to the disclosed embodiments. Many modifications and variations will be apparent to those of ordinary skill in the art without departing from the scope and spirit of the described embodiments. The choice of terms used herein is intended to best explain the principles of the embodiments, practical applications, or technical improvements to the technology in the market, or to enable other persons of ordinary skill in the art to understand the embodiments disclosed herein.

Claims

1. A method for generating a cross-modal encoder, include: Extracting sample video features of a sample video, sample audio features of a sample audio, and sample text features of a sample text, wherein the contents of the sample video, the sample audio, and the sample text are associated with each other; The sample video feature, the sample audio feature, and the sample text feature are combined via the cross-modal encoder to generate a first cross-modal sample feature; Performing a perturbation operation on the sample video feature, the sample audio feature, and the sample text feature to generate a perturbed cross-modal sample feature; encoding the perturbed cross-modal sample features via the cross-modal encoder to generate second cross-modal sample features; and The cross-modal encoder is adjusted based on the first cross-modal sample feature and the second cross-modal sample feature.

2. According to the method of claim 1, the sample video feature, the sample audio feature, and the sample text feature are disturbed to generate a disturbed cross-modal sample feature include: Dividing one or more of the sample video features, the sample audio features, and the sample text features into multiple parts; randomly masking at least one of the plurality of portions to generate a plurality of masked portions; as well as An XOR operation is performed on the masked portions.

3. The method according to claim 2, wherein the XOR operation is performed on the masked multiple parts include: The masked parts are randomly concatenated to generate the disturbed cross-modal sample features.

4. The method according to claim 1, wherein the cross-modal encoder is adjusted based on the first cross-modal sample feature and the second cross-modal sample feature. include: Comparing the generated first cross-modal sample features with the second cross-modal sample features; In response to a KL (Kullback-Leible) divergence between the first cross-modal sample feature and the second cross-modal sample feature being higher than a predetermined threshold, training the cross-modal encoder by adjusting parameters in the cross-modal encoder; and In response to a KL divergence between the first cross-modal sample feature and the second cross-modal sample feature being lower than the predetermined threshold, adjusting parameters in the cross-modal encoder is stopped.

5. The method according to claim 1, wherein the cross-modal encoder is further adjusted based on reconstructed sample video features, reconstructed sample audio features, and reconstructed sample text features.

6. The method according to claim 5, wherein the sample video features, the sample audio features, and the sample text features are reconstructed include: Combining the sample video features, the sample audio features, and the sample text features into a tensor; Decomposing the tensor into three vectors using CP (CANDECOMP / PARAFAC) decomposition, wherein the three vectors are orthogonal to each other; and The sample video features, the sample audio features, and the sample text features are reconstructed based on the three vectors.

7. The method of claim 1 , further comprising using the cross-modal encoder to: receiving a reference video, a reference audio, and a text; Encoding the reference video, the reference audio, and the text to generate a first cross-modal data set; Predicting the duration of speech based on each phoneme feature in the first cross-modal data set; Predicting a pitch spectrum of the speech by converting pitch features in the first cross-modal data set into a pitch spectrum; Predicting the energy value of the speech by calculating the volume feature and the prosody feature in the first cross-modal data set; generating a second cross-modal data set based on the duration, the pitch spectrum, and the energy value; Decoding the second cross-modal data set to generate a mel-spectrogram; as well as The speech associated with the text is generated based on the mel-spectrogram.

8. The method according to claim 7, wherein the speech is generated based on the mel-spectrogram include: Sampling the mel-spectrogram by performing a transposed convolution on the mel-spectrogram to generate sampled data; The speech is generated by performing multi-receptive field fusion on the mel-spectrogram.

9. The method according to claim 7, further comprising: include: Predicting, by a duration predictor, the duration of the speech based on each phoneme feature in the first cross-modal data set; adjusting the duration predictor by comparing the predicted duration to a reference truth phoneme duration; Predicting, by a pitch predictor, a pitch spectrum of the speech by converting pitch features in the first cross-modal data set into a pitch spectrum; adjusting the pitch predictor by comparing the predicted pitch spectrum with a reference truth pitch spectrum; An energy predictor predicts an energy value of the speech by calculating a volume feature and a prosody feature in the first cross-modal data set; as well as The energy predictor is adjusted by comparing the difference between the predicted energy value and a reference truth energy value.

10. An electronic device, include: at least one processor; as well as A memory coupled to at least one processor and having instructions stored thereon, the instructions causing the electronic device to perform actions when executed by the at least one processor, the actions comprising: Extracting sample video features of a sample video, sample audio features of a sample audio, and sample text features of a sample text, wherein the contents of the sample video, the sample audio, and the sample text are associated with each other; Combining the sample video feature, the sample audio feature, and the sample text feature via a cross-modal encoder to generate a first cross-modal sample feature; Performing a perturbation operation on the sample video feature, the sample audio feature, and the sample text feature to generate a perturbed cross-modal sample feature; encoding the perturbed cross-modal sample features via the cross-modal encoder to generate second cross-modal sample features; and The cross-modal encoder is adjusted based on the first cross-modal sample feature and the second cross-modal sample feature.

11. The electronic device according to claim 10, wherein the sample video feature, the sample audio feature, and the sample text feature are disturbed to generate a disturbed cross-modal sample feature include: Dividing one or more of the sample video features, the sample audio features, and the sample text features into multiple parts; randomly masking at least one of the plurality of portions to generate a plurality of masked portions; as well as An XOR operation is performed on the masked portions.

12. The electronic device according to claim 11, wherein the XOR operation is performed on the masked portions. include: The masked parts are randomly concatenated to generate the disturbed cross-modal sample features.

13. The electronic device according to claim 10, wherein the cross-modal encoder is adjusted based on the first cross-modal sample feature and the second cross-modal sample feature. include: Comparing the generated first cross-modal sample features with the second cross-modal sample features; In response to a KL (Kullback-Leible) divergence between the first cross-modal sample feature and the second cross-modal sample feature being higher than a predetermined threshold, training the cross-modal encoder by adjusting parameters in the cross-modal encoder; as well as In response to a KL divergence between the first cross-modal sample feature and the second cross-modal sample feature being lower than the predetermined threshold, adjusting the parameters in the cross-modal encoder is stopped.

14. The electronic device of claim 10, wherein the cross-modal encoder is further adjusted based on reconstructed sample video features, reconstructed sample audio features, and reconstructed sample text features.

15. The electronic device according to claim 14, wherein the sample video feature, the sample audio feature, and the sample text feature are reconstructed include: Combining the sample video features, the sample audio features, and the sample text features into a tensor; Decomposing the tensor into three vectors using CP (CANDECOMP / PARAFAC) decomposition, wherein the three vectors are orthogonal to each other; and The sample video features, the sample audio features, and the sample text features are reconstructed based on the three vectors.

16. The electronic device of claim 10, further comprising using the cross-modal encoder for: receiving a reference video, a reference audio, and a text; Encoding the reference video, the reference audio, and the text to generate a first cross-modal data set; Predicting the duration of speech based on each phoneme feature in the first cross-modal data set; Predicting a pitch spectrum of the speech by converting pitch features in the first cross-modal data set into a pitch spectrum; Predicting the energy value of the speech by calculating the volume feature and the prosody feature in the first cross-modal data set; generating a second cross-modal data set based on the duration, the pitch spectrum, and the energy value; Decoding the second cross-modal data set to generate a mel-spectrogram; as well as The speech associated with the text is generated based on the mel-spectrogram.

17. The electronic device according to claim 16, wherein the speech is generated based on the Mel-spectrogram include: Sampling the mel-spectrogram by performing a transposed convolution on the mel-spectrogram to generate sampled data; The speech is generated by performing multi-receptive field fusion on the mel-spectrogram.

18. The electronic device according to claim 16, further comprising: include: Predicting, by a duration predictor, the duration of the speech based on each phoneme feature in the first cross-modal data set; adjusting the duration predictor by comparing the predicted duration to a reference truth phoneme duration; Predicting, by a pitch predictor, a pitch spectrum of the speech by converting pitch features in the first cross-modal data set into a pitch spectrum; adjusting the pitch predictor by comparing the predicted pitch spectrum with a reference truth pitch spectrum; An energy predictor predicts an energy value of the speech by calculating a volume feature and a prosody feature in the first cross-modal data set; as well as The energy predictor is adjusted by comparing the difference between the predicted energy value and a reference truth energy value.

19. A computer program product tangibly stored on a non-transitory computer readable storage medium and comprising computer executable instructions that, when executed, cause a computer to perform: Extracting sample video features of a sample video, sample audio features of a sample audio, and sample text features of a sample text, wherein the contents of the sample video, the sample audio, and the sample text are associated with each other; Combining the sample video feature, the sample audio feature, and the sample text feature via a cross-modal encoder to generate a first cross-modal sample feature; Performing a perturbation operation on the sample video feature, the sample audio feature, and the sample text feature to generate a perturbed cross-modal sample feature; encoding the perturbed cross-modal sample features via the cross-modal encoder to generate second cross-modal sample features; and The cross-modal encoder is adjusted based on the first cross-modal sample feature and the second cross-modal sample feature.

20. The computer program product according to claim 19, wherein the sample video feature, the sample audio feature, and the sample text feature are disturbed to generate a disturbed cross-modal sample feature include: Dividing one or more of the sample video features, the sample audio features, and the sample text features into multiple parts; randomly masking at least one of the plurality of portions to generate a plurality of masked portions; as well as An XOR operation is performed on the masked portions.