Speech synthesis method and device, readable medium and electronic equipment

By using mutual information to determine the target neural network model for training, the decoupling of timbre features and emotional features is achieved, solving the problem of timbre feature transfer during emotional transfer in speech synthesis and improving the sound quality of speech synthesis.

CN119028313APending Publication Date: 2024-11-26BEIJING YOUZHUJU NETWORK TECH CO LTD
View PDF 0 Cites 2 Cited by

Patent Information

Application Number
CN202310605902.7
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2023-05-25
Publication Date
2024-11-26

AI Technical Summary

Technical Problem

In existing technologies, when emotion transfer occurs during speech synthesis, other speech-related feature information is transferred along with the emotion transfer, resulting in poor emotion transfer performance.

Method used

By using mutual information to determine the training of the target neural network model, the timbre features and emotional features are decoupled, and only the emotional features in the source speech are extracted to generate the target speech.

Benefits of technology

It improves the effect of emotion transfer, so that the emotional features in the generated target speech are not mixed with timbre information, resulting in better sound quality.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN119028313A_ABST
    Figure CN119028313A_ABST
Patent Text Reader

Abstract

The invention relates to a speech synthesis method and device, a readable medium and electronic equipment. The method comprises the following steps: acquiring a target text and a target tone identifier; determining a target emotion identifier of the source voice; the target text, the target timbre identifier and the target emotion identifier are input into a speech synthesis model to obtain a target Mel spectrum feature output by the speech synthesis model, and the speech synthesis model is obtained by training a target neural network model through a mutual information determination model and a plurality of sample sets; and generating a target voice according to the target Mel spectrum feature. Decoupling of the timbre features and the emotional features is realized through the mutual information determination model, so that only the emotional features in the source speech are extracted in the process of determining the target Mel spectrum features according to the speech synthesis model, the timbre information of the source speech is not mixed in the emotional features, the effect of emotional migration is improved, and the efficiency of emotional migration is improved. And the sound quality of the generated target voice is better.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present disclosure relates to the field of speech processing technology, and in particular, to a speech synthesis method, device, readable medium and electronic device. Background Art

[0002] With the development of artificial intelligence technology, speech synthesis is increasingly widely used. Emotion transfer is a common application scenario in speech synthesis. The goal of emotion transfer is to transfer the emotion of one timbre to another timbre, and to synthesize speech even without emotion data. For example, in the speech synthesis process of audio novels, speech synthesis can be performed for different characters and styles of speech through emotion transfer, thus improving the expressiveness of audio novels.

[0003] In the related art, emotion transfer is achieved by training an emotion transfer model. However, this method transfers other feature information related to speech while transferring emotions, resulting in poor effect of speech obtained by emotion transfer. Summary of the invention

[0004] This section is provided to introduce the concepts in a brief form, which will be described in detail in the detailed implementation section below. This section is not intended to identify the key features or essential features of the technical solution claimed for protection, nor is it intended to be used to limit the scope of the technical solution claimed for protection.

[0005] In a first aspect, the present disclosure provides a speech synthesis method, comprising:

[0006] Obtaining target text and target timbre identification;

[0007] Determine the target emotion label of the source speech;

[0008] Inputting the target text, the target timbre identifier and the target emotion identifier into a speech synthesis model to obtain target mel-spectrogram features output by the speech synthesis model, wherein the speech synthesis model is obtained by training a target neural network model through a mutual information determination model and multiple sample sets;

[0009] Generate target speech according to the target mel-spectrogram features.

[0010] In a second aspect, the present disclosure provides a speech synthesis device, comprising:

[0011] A first acquisition module acquires a target text and a target timbre identifier;

[0012] A determination module, used to determine the target emotion identifier of the source speech;

[0013] The second acquisition module is used to input the target text, the target timbre identifier and the target emotion identifier into a speech synthesis model to obtain the target mel-spectrogram features output by the speech synthesis model, wherein the speech synthesis model is obtained by training a target neural network model through a mutual information determination model and multiple sample sets.

[0014] In a third aspect, the present disclosure provides a computer-readable medium having a computer program stored thereon, which, when executed by a processing device, implements the steps of the method described in the first aspect of the present disclosure.

[0015] In a fourth aspect, the present disclosure provides an electronic device, including:

[0016] a storage device having at least one computer program stored thereon;

[0017] At least one processing device is used to execute the at least one computer program in the storage device to implement the steps of the method described in the first aspect of the present disclosure.

[0018] Through the above technical solution, the target text and the target timbre identifier are obtained; the target emotion identifier of the source speech is determined; the target text, the target timbre identifier and the target emotion identifier are input into the speech synthesis model to obtain the target Mel-spectrogram feature output by the speech synthesis model, wherein the speech synthesis model is obtained by training the target neural network model through the mutual information determination model and multiple sample sets; and the target speech is generated according to the target Mel-spectrogram feature. That is to say, in the process of training the speech synthesis model, the decoupling of the timbre feature and the emotion feature is achieved through the mutual information determination model, so that in the process of determining the target Mel-spectrogram feature according to the speech synthesis model, only the emotion feature in the source speech is extracted, and the timbre feature in the source speech is not extracted. In this way, the target speech generated according to the target Mel-spectrogram feature transfers the emotion feature in the source speech, and the timbre information of the source speech is not mixed in the emotion feature, which improves the effect of emotion transfer, thereby making the sound quality of the generated target speech better.

[0019] Other features and advantages of the present disclosure will be described in detail in the following detailed description. BRIEF DESCRIPTION OF THE DRAWINGS

[0020] The above and other features, advantages and aspects of the embodiments of the present disclosure will become more apparent with reference to the following detailed description in conjunction with the accompanying drawings. Throughout the drawings, the same or similar reference numerals represent the same or similar elements. It should be understood that the drawings are schematic and that components and elements are not necessarily drawn to scale. In the drawings:

[0021] Figure 1 is a flow chart of a speech synthesis method according to an exemplary embodiment of the present disclosure;

[0022] Figure 2 is a flowchart of a model training method according to an exemplary embodiment of the present disclosure;

[0023] Figure 3 is a flowchart showing a model training step according to an exemplary embodiment of the present disclosure;

[0024] Figure 4 is a flow chart of a method for determining estimated mel-spectrogram features according to an exemplary embodiment of the present disclosure;

[0025] Figure 5 is a schematic diagram of a model training method according to an exemplary embodiment of the present disclosure;

[0026] Figure 6 is a block diagram of a speech synthesis device according to an exemplary embodiment of the present disclosure;

[0027] Figure 7 is a block diagram of another speech synthesis device according to an exemplary embodiment of the present disclosure;

[0028] Figure 8 It is a schematic structural diagram of an electronic device according to an exemplary embodiment of the present disclosure. DETAILED DESCRIPTION

[0029] 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.

[0030] It should be understood that the various steps described in the method embodiments of the present disclosure may be performed in different orders and / or in parallel. In addition, the method embodiments may include additional steps and / or omit the steps shown. The scope of the present disclosure is not limited in this respect.

[0031] The term "including" and its variations used herein are open inclusions, i.e., "including but not limited to". The term "based on" means "based at least in part on". The term "one embodiment" means "at least one embodiment"; the term "another embodiment" means "at least one additional embodiment"; the term "some embodiments" means "at least some embodiments". The relevant definitions of other terms will be given in the following description.

[0032] It should be noted that the concepts such as "first" and "second" mentioned in the present disclosure are only used to distinguish different devices, modules or units, and are not used to limit the order or interdependence of the functions performed by these devices, modules or units.

[0033] It should be noted that the modifications of "one" and "plurality" mentioned in the present disclosure are illustrative rather than restrictive, and those skilled in the art should understand that unless otherwise clearly indicated in the context, it should be understood as "one or more".

[0034] The names of the messages or information exchanged between multiple devices in the embodiments of the present disclosure are only used for illustrative purposes and are not used to limit the scope of these messages or information.

[0035] All actions of acquiring signals, information or data in the present disclosure are carried out in compliance with the relevant data protection laws and policies of the country where they are located, and with the authorization given by the owner of the corresponding device.

[0036] It is understandable that before using the technical solutions disclosed in the embodiments of the present disclosure, the types, scope of use, usage scenarios, etc. of the personal information involved in the present disclosure should be informed to the user and the user's authorization should be obtained in an appropriate manner in accordance with relevant laws and regulations.

[0037] For example, in response to receiving an active request from a user, a prompt message is sent to the user to clearly prompt the user that the operation requested to be performed will require obtaining and using the user's personal information. Thus, the user can autonomously choose whether to provide personal information to software or hardware such as an electronic device, application, server, or storage medium that performs the operation of the technical solution of the present disclosure according to the prompt message.

[0038] As an optional but non-limiting implementation, in response to receiving an active request from the user, the prompt information may be sent to the user in the form of a pop-up window, in which the prompt information may be presented in text form. In addition, the pop-up window may also carry a selection control for the user to choose "agree" or "disagree" to provide personal information to the electronic device.

[0039] It is understandable that the above notification and the process of obtaining user authorization are merely illustrative and do not constitute a limitation on the implementation of the present disclosure. Other methods that meet the relevant laws and regulations may also be applied to the implementation of the present disclosure.

[0040] At the same time, it is understandable that the data involved in this technical solution (including but not limited to the data itself, the acquisition or use of the data) shall comply with the requirements of relevant laws, regulations and relevant provisions.

[0041] The present disclosure is described below in conjunction with specific embodiments.

[0042] Figure 1 is a flowchart of a speech synthesis method according to an exemplary embodiment of the present disclosure. Figure 1 As shown, the method may include:

[0043] S101, obtaining a target text and a target timbre identifier.

[0044] Among them, the target text can be the text corresponding to the voice that needs to be emotionally transferred. For example, if the timbre of the first voice is A1, the first voice has no emotion, the timbre of the second voice is A2, and the timbre of the second voice is B, then the emotion B of the second voice can be transferred to the first voice so that the timbre A1 has emotion B. The text corresponding to the first voice is the target text.

[0045] In this step, the second speech that needs to be emotionally transferred can be obtained first, the target text and the target timbre can be determined from the second speech, and the target timbre identification corresponding to the target timbre can be determined according to the preset timbre identification association relationship. The timbre identification association relationship may include the correspondence between different timbres and timbre identifications. By way of example, the timbre may include a first timbre and a second timbre. The timbre identification corresponding to the first timbre may be "1", and the timbre identification corresponding to the second timbre may be "2". It should be noted that the above-mentioned timbre identification association relationship is only an example, and the present disclosure does not limit this. For example, if the target timbre of the second speech is determined to be the second timbre, the target timbre identification is "2".

[0046] S102: Determine the target emotion identifier of the source speech.

[0047] The source voice may be a voice with transferred emotion. Taking the example in step S101 as an example, the second voice is the source voice.

[0048] In this step, the target emotion of the source speech can be determined, and the target emotion identifier corresponding to the target emotion can be determined according to the preset emotion identifier association relationship. The emotion identifier association relationship can include the correspondence between different emotions and emotion identifiers. For example, the emotions can include sadness, happiness, and anger. The emotion identifier corresponding to sadness can be "1", the emotion identifier corresponding to happiness can be "2", and the emotion identifier corresponding to anger can be "3". It should be noted that the above-mentioned emotion identifier association relationship is only an example for illustration, and the present disclosure does not limit this. For example, if it is determined that the target emotion of the source speech is happiness, then the target emotion identifier is "2".

[0049] S103: input the target text, the target timbre identifier, and the target emotion identifier into a speech synthesis model to obtain target mel-spectrogram features output by the speech synthesis model.

[0050] The speech synthesis model is obtained by training a target neural network model through a mutual information determination model and multiple sample sets.

[0051] In a possible implementation, the sample set includes a sample text and sample speech information of a sample speech corresponding to the sample text, the sample speech information includes a sample timbre identifier, a sample emotion identifier, and a sample mel spectrum feature, the mutual information determination model is used by the target neural network model to limit the mutual information between the sample timbre feature and the sample emotion feature, the sample timbre feature is the timbre feature corresponding to the sample timbre identifier, and the sample emotion feature is the emotion feature corresponding to the sample emotion identifier. The mutual information determination model can be a model trained based on the CLUB (Contrastive Log-ratio UpperBound) method.

[0052] In this step, after determining the target timbre identifier, the target emotion identifier and the target text, the target timbre identifier, the target emotion identifier and the target text can be input into the speech synthesis model, and the target timbre feature corresponding to the target timbre identifier and the target emotion feature corresponding to the target emotion identifier can be determined by the speech synthesis model, and the target mel-spectrogram feature can be determined according to the target timbre feature, the target emotion feature and the target text. By way of example, the speech synthesis model can include a mel-spectrogram feature acquisition submodel, and the target timbre feature, the target emotion feature and the target text are input into a pre-trained mel-spectrogram feature acquisition submodel to acquire the target mel-spectrogram feature output by the mel-spectrogram feature acquisition submodel.

[0053] S104: Generate target speech according to the target Mel-spectrogram features.

[0054] In this step, after the target Mel-spectrogram feature is determined, the target speech can be generated by a vocoder in the prior art according to the target Mel-spectrogram feature, which will not be described in detail here.

[0055] By adopting the above method, in the process of training the speech synthesis model, the decoupling of timbre features and emotional features is achieved through the mutual information determination model, so that in the process of determining the target Mel-spectrogram features according to the speech synthesis model, only the emotional features in the source speech are extracted, and the timbre features in the source speech are not extracted. In this way, the emotional features in the source speech are transferred to the target speech generated according to the target Mel-spectrogram features, and the timbre information of the source speech is not mixed in the emotional features, which improves the effect of emotion transfer, thereby making the sound quality of the generated target speech better.

[0056] Figure 2 is a flowchart of a model training method according to an exemplary embodiment of the present disclosure. Figure 2As shown, the method may include:

[0057] S21. Obtain multiple sample sets.

[0058] The sample set may include sample text and sample speech information of the sample speech corresponding to the sample text, the sample speech information includes a sample timbre identifier, a sample emotion identifier and a sample mel-spectrogram feature, and the sample speech may be a speech signal for the target text collected by an electronic device. The method for determining the sample timbre identifier may refer to the method for determining the target timbre identifier in step S101, and the method for determining the sample emotion identifier may refer to the method for determining the target emotion identifier in step S102, which will not be repeated here. The sample mel-spectrogram feature can be extracted from the sample speech by the method of the prior art, which will not be repeated here.

[0059] S22. Loop through the model training steps according to the multiple sample sets until the trained target neural network model meets the preset stop iteration condition, and use the trained target neural network model as the speech synthesis model.

[0060] Among them, the target neural network model may include a timbre feature extraction sub-model and an emotion feature extraction sub-model.

[0061] After obtaining multiple sample sets, the model training step can be executed cyclically according to the multiple sample sets, and the estimated mel spectrum features can be determined through the target neural network model. According to the estimated mel spectrum features and the sample mel spectrum features of the sample set, it can be determined whether the target neural network model meets the preset stop iteration condition. If it is determined that the target neural network model meets the preset stop iteration condition, the target neural network model can be used as the speech synthesis model; if it is determined that the target neural network model does not meet the preset stop iteration condition, the model training step is continued until it is determined that the trained target neural network model meets the preset stop iteration condition.

[0062] Figure 3 is a flowchart of a model training step according to an exemplary embodiment of the present disclosure. Figure 3 As shown, the model training steps may include:

[0063] S31. Acquire the sample timbre feature corresponding to the sample timbre identifier through the timbre feature extraction sub-model.

[0064] For example, the sample timbre identification may be input into the timbre feature extraction sub-model, and the timbre feature extraction sub-model may be used to extract features from the sample timbre identification to obtain the sample timbre features corresponding to the sample timbre identification.

[0065] S32. Obtain the sample emotion feature corresponding to the sample emotion identifier through the emotion feature extraction sub-model.

[0066] For example, the sample emotion identifier may be input into the emotion feature extraction sub-model, and the emotion feature extraction sub-model may be used to perform feature extraction on the sample emotion identifier to obtain a sample emotion feature corresponding to the sample emotion identifier.

[0067] S33. According to the sample timbre feature and the sample emotional feature, the parameters of the timbre feature extraction sub-model and the emotional feature extraction sub-model are updated through the mutual information determination model to obtain a new timbre feature extraction sub-model and a new emotional feature extraction sub-model.

[0068] In one possible implementation, the sample timbre feature and the sample emotion feature can be input into the mutual information determination model to obtain the sample mutual information between the sample timbre feature and the sample emotion feature; a second loss value is determined based on the sample mutual information; and the parameters of the timbre feature extraction sub-model and the emotion feature extraction sub-model are updated based on the second loss value.

[0069] For example, the second loss value can be calculated by the following formula:

[0070]

[0071] Among them, L2 is the second loss value, N is the number of sample sets for each execution of the model training step, emotion i is the sample sentiment feature corresponding to the i-th sample set in N sample sets, speaker i is the sample timbre feature corresponding to the i-th sample set in N sample sets, emotion j is the sample sentiment feature corresponding to the jth sample set in N sample sets, speaker j is the sample timbre feature corresponding to the jth sample set in N sample sets, q θ (emotion i |speaker i ) is the conditional distribution between the sample timbre features and sample emotion features corresponding to the i-th sample set in the N sample sets, q θ (emotion j |speaker j ) is the conditional distribution between the sample timbre features and sample emotion features corresponding to the jth sample set among the N sample sets.

[0072] After determining the second loss value, the parameters of the timbre feature extraction sub-model and the emotion feature extraction sub-model can be updated according to the second loss value using the method of the prior art.

[0073] Before determining the second loss value according to the sample mutual information, it can be determined whether the mutual information determination model satisfies the preset model stability condition. In a possible implementation, it can be determined whether the mutual information determination model satisfies the preset model stability condition according to the sample mutual information; when it is determined that the mutual information determination model satisfies the preset model stability condition, the second loss value is determined according to the sample mutual information. For example, if the sample mutual information is within the preset mutual information range, it means that the mutual information determination model satisfies the preset model stability condition. The preset mutual information range can be preset based on experience. For example, the preset mutual information range can be 0 to 10; if the sample mutual information is outside the preset mutual information range, for example, the sample mutual information is less than 0, it means that the mutual information determination model does not satisfy the preset model stability condition.

[0074] In another possible implementation, the number of iterations of the mutual information determination model may be used to determine whether the mutual information determination model satisfies the preset model stability condition. For example, if the number of iterations is greater than or equal to a preset number threshold, it indicates that the mutual information determination model satisfies the preset model stability condition; if the number of iterations is less than the preset number threshold, it indicates that the mutual information determination model does not satisfy the preset model stability condition. For example, the preset number threshold may be preset based on experience, for example, the preset number threshold may be 5.

[0075] When it is determined that the mutual information determination model satisfies the preset model stability condition, the second loss value is determined based on the sample mutual information. In this way, since the accuracy of the sample mutual information determined by the mutual information determination model is relatively high, the accuracy of the second loss value determined based on the sample mutual information is relatively high, thereby improving the efficiency of model training.

[0076] In the case where it is determined that the mutual information determination model does not satisfy the preset model stability condition, a third loss value is determined according to the sample mutual information; and the parameters of the mutual information determination model are updated according to the third loss value to obtain a new mutual information determination model. For example, if it is determined that the mutual information determination model does not satisfy the preset model stability condition, for example, the sample mutual information is not within the preset mutual information range or the number of iterations of the mutual information determination model is less than the preset number threshold, it means that the mutual information determination model is unstable, and the mutual information determination model needs to be trained first.

[0077] For example, the third loss value can be calculated by the following formula:

[0078]

[0079] Among them, L3 is the third loss value, sg() represents gradient truncation, and when training the mutual information determination model, the timbre feature extraction sub-model and the emotion feature extraction sub-model are not trained.

[0080] S34, determining an estimated mel spectrum feature according to the sample timbre feature, the sample emotion feature and the sample text.

[0081] After determining the sample timbre feature, the sample emotion feature and the sample text, the sample timbre feature, the sample emotion feature and the sample text may be used through a mel spectrum acquisition submodel in the target neural network model to determine the estimated mel spectrum feature.

[0082] Figure 4 is a flow chart of a method for determining estimated mel spectrum features according to an exemplary embodiment of the present disclosure. Figure 4 As shown, the method may include:

[0083] S1. Determine sample text features corresponding to the sample text.

[0084] S2. Determine a first sample splicing feature according to the sample timbre feature and the sample emotion feature.

[0085] S3. Concatenate the sample sentiment feature and the sample text feature to obtain a second sample concatenated feature.

[0086] S4. Determine an estimated Mel spectrum feature according to the first sample splicing feature and the second sample splicing feature.

[0087] For example, the target neural network model may include a text encoder, and after the sample text is input into the text encoder, a sample text feature corresponding to the sample text may be obtained. After determining the sample timbre feature and the sample emotion feature, the sample timbre feature and the sample emotion feature may be concatenated in a concat manner to obtain the first sample concatenated feature, and the sample emotion feature and the sample text feature may be concatenated in a concat manner to obtain the second sample concatenated feature.

[0088] In a possible implementation, after determining the sample timbre feature and the sample emotion feature, the sample timbre feature and the sample emotion feature may be spliced ​​to obtain a pending sample splicing feature; and the pending sample splicing feature may be subjected to CLN (Conditional Layer Normalization) processing to obtain the first sample splicing feature. It should be noted that the CLN processing method may refer to the processing method of the prior art and will not be described in detail here.

[0089] After determining the first sample splicing feature and the second sample splicing feature, the sample fundamental frequency and sample energy can be determined according to the first sample splicing feature and the second sample splicing feature; the second sample splicing feature, the sample fundamental frequency and the sample energy are spliced ​​to obtain a third sample splicing feature; the estimated Mel spectrum feature is determined according to the first sample splicing feature and the third sample splicing feature. By way of example, the first sample splicing feature and the second sample splicing feature can be input into a predictor in the target neural network model, and the sample fundamental frequency and the sample energy are determined by the predictor. The second sample splicing feature, the sample fundamental frequency and the sample energy are spliced ​​together in a concat manner to obtain the third sample splicing feature, the first sample feature and the third sample feature are input into the duration prediction and upsampling sub-model of the target neural network model to obtain the target sample feature, the target sample feature and the first sample splicing feature are input into the Mel spectrum acquisition sub-model to obtain the estimated Mel spectrum feature.

[0090] It should be noted that the model structure of the text encoder, the predictor, the duration prediction and upsampling sub-model and the mel-spectrogram acquisition sub-model can be the structure of an existing TTS (Text To Speech) model, and the present disclosure does not limit this.

[0091] S35. Determine a first loss value according to the estimated mel spectrum feature and the sample mel spectrum feature.

[0092] S36. When it is determined according to the first loss value that the target neural network model does not satisfy the preset stop iteration condition, update the parameters of the target neural network model according to the first loss value to obtain a trained target neural network model, and use the trained target neural network model as a new target neural network model.

[0093] After determining the first loss value, if it is determined that the first loss value is less than or equal to the preset loss value threshold, then it is determined that the target neural network model meets the preset stop iteration condition, and the target neural network model is used as the speech synthesis model; if it is determined that the first loss value is greater than the preset loss value threshold, then it is determined that the target neural network model does not meet the preset stop iteration condition, and the parameters of the target neural network model are updated according to the first loss value to obtain the trained target neural network model, and the trained target neural network model is used as the new target neural network model to continue to execute the above model training steps.

[0094] Figure 5 is a schematic diagram of a model training method according to an exemplary embodiment of the present disclosure, such as Figure 5As shown, the target neural network model includes a timbre feature extraction submodel, an emotional feature extraction submodel, a text encoder, a predictor, a duration prediction and upsampling submodel, and a mel spectrum acquisition submodel, and the outputs of the timbre feature extraction submodel and the emotional feature extraction submodel are coupled with the input of the mutual information determination model. The sample text is input into the text encoder to obtain the sample text feature, the sample timbre identifier is input into the timbre feature extraction submodel to obtain the sample timbre feature, and the sample emotional identifier is input into the emotional feature extraction submodel to obtain the sample emotional feature. The sample timbre feature and the sample emotional feature are input into the mutual information determination model, and the third loss value is determined by the mutual information determination model. The parameters of the mutual information determination model are updated according to the third loss value to obtain a new mutual information determination model. The sample timbre feature and the sample emotional feature are spliced ​​by concat to obtain the pending sample splicing feature, the pending sample splicing feature is CLN processed to obtain the first sample splicing feature, and the sample emotional feature and the sample text feature are spliced ​​by concat to obtain the second sample splicing feature. The first sample splicing feature and the second sample splicing feature are input into the predictor to obtain the sample fundamental frequency and sample energy, and the second sample splicing feature, the sample fundamental frequency and the sample energy are spliced ​​by concat to obtain the third sample splicing feature. The first sample splicing feature and the third sample splicing feature are input into the duration prediction and upsampling sub-model to obtain the target sample feature, and the first sample splicing feature and the target sample feature are input into the Mel spectrum acquisition sub-model to obtain the estimated Mel spectrum feature. A first loss value is determined based on the estimated Mel spectrum feature and the sample Mel spectrum feature, and the parameters of the timbre feature extraction sub-model, the emotion feature extraction sub-model, the text encoder, the predictor, the duration prediction and upsampling sub-model and the Mel spectrum acquisition sub-model in the target neural network model are updated based on the first loss value to obtain a new target neural network model.

[0095] It should be noted that the above model training process is the first iterative process, and the parameters of the mutual information determination model are updated. When the number of iterations is greater than or equal to the preset number threshold, the mutual information determination model may not be updated. The mutual information determination model can determine the second loss value based on the sample timbre feature and the sample emotional feature, and update the parameters of the timbre feature extraction submodel and the emotional feature extraction submodel based on the second loss value, and train the timbre feature extraction submodel and the emotional feature extraction submodel. For example, if the preset number threshold is 5, the mutual information determination model is trained in the first 5 iterations, and the timbre feature extraction submodel and the emotional feature extraction submodel are trained in the 6th to 10th iterations. The mutual information determination model is trained again in the 11th to 15th iterations, and the timbre feature extraction submodel and the emotional feature extraction submodel are trained again in the 16th to 20th iterations. And so on, until the target neural network model meets the preset stop iteration condition. The above training process is just an example, and the present disclosure does not limit the specific iteration method.

[0096] By adopting the above-mentioned model training method, in the process of training the target neural network model, the timbre feature extraction submodel and the emotion feature extraction submodel are trained through the mutual information determination model, so that the mutual information between the sample timbre features extracted by the timbre feature extraction submodel and the sample emotion features extracted by the emotion feature extraction submodel is minimized, and the decoupling between the sample timbre features and the sample emotion features is achieved, so that the trained speech synthesis model will not be mixed with timbre information when performing emotion transfer, thereby improving the speech effect after emotion transfer. In addition, the sample timbre features and the sample emotion features are combined with other features in the form of CLN, so as to improve the adaptability of the target neural network model to the sample timbre features and the sample emotion features, so that the robustness of the speech synthesis model is better.

[0097] Figure 6 is a block diagram of a speech synthesis device according to an exemplary embodiment of the present disclosure. Figure 6 As shown, including:

[0098] The first acquisition module 601 acquires the target text and the target timbre identifier;

[0099] A determination module 602 is used to determine a target emotion identifier of a source speech;

[0100] A second acquisition module 603 is used to input the target text, the target timbre identifier and the target emotion identifier into a speech synthesis model to obtain a target mel-spectrogram feature output by the speech synthesis model, wherein the speech synthesis model is obtained by training a target neural network model through a mutual information determination model and multiple sample sets;

[0101] The generating module 604 is used to generate the target speech according to the target mel-spectrogram feature.

[0102] Optionally, the sample set includes sample text and sample speech information of a sample speech corresponding to the sample text, the sample speech information includes a sample timbre identifier, a sample emotion identifier and a sample mel-spectrogram feature, and the mutual information determination model is used by the target neural network model to limit the mutual information between the sample timbre feature and the sample emotion feature, the sample timbre feature is the timbre feature corresponding to the sample timbre identifier, and the sample emotion feature is the emotion feature corresponding to the sample emotion identifier.

[0103] Optionally, Figure 7 is a block diagram of another speech synthesis device according to an exemplary embodiment of the present disclosure, such as Figure 7 As shown, the device also includes:

[0104] The model training module 605 is used to:

[0105] Obtain multiple sample sets;

[0106] Execute the model training step cyclically according to the multiple sample sets until the trained target neural network model meets the preset stop iteration condition, and use the trained target neural network model as the speech synthesis model, wherein the target neural network model includes a timbre feature extraction sub-model and an emotion feature extraction sub-model;

[0107] The model training steps include:

[0108] Acquire the sample timbre feature corresponding to the sample timbre identifier of the sample set through the timbre feature extraction sub-model;

[0109] Obtaining the sample emotion feature corresponding to the sample emotion identifier of the sample set through the emotion feature extraction sub-model;

[0110] According to the sample timbre feature and the sample emotional feature, the parameters of the timbre feature extraction sub-model and the emotional feature extraction sub-model are updated through the mutual information determination model to obtain a new timbre feature extraction sub-model and a new emotional feature extraction sub-model;

[0111] Determine an estimated mel spectrum feature according to the sample timbre feature, the sample emotion feature and the sample text;

[0112] Determine a first loss value according to the estimated Mel spectrum feature and the sample Mel spectrum feature;

[0113] When it is determined according to the first loss value that the target neural network model does not satisfy the preset stop iteration condition, the parameters of the target neural network model are updated according to the first loss value to obtain a trained target neural network model, and the trained target neural network model is used as a new target neural network model.

[0114] Optionally, the model training module 605 is further used to:

[0115] Inputting the sample timbre feature and the sample emotion feature into the mutual information determination model to obtain sample mutual information between the sample timbre feature and the sample emotion feature;

[0116] Determine a second loss value according to the sample mutual information;

[0117] The parameters of the timbre feature extraction sub-model and the emotion feature extraction sub-model are updated according to the second loss value.

[0118] Optionally, the model training module 605 is further used to:

[0119] Determining whether the mutual information determination model satisfies a preset model stability condition according to the sample mutual information;

[0120] The determining the second loss value according to the sample mutual information includes:

[0121] When it is determined that the mutual information determination model satisfies the preset model stability condition, a second loss value is determined according to the sample mutual information.

[0122] Optionally, the model training module 605 is further used to:

[0123] When it is determined that the mutual information determination model does not satisfy the preset model stability condition, determining a third loss value according to the sample mutual information;

[0124] The parameters of the mutual information determination model are updated according to the third loss value to obtain a new mutual information determination model.

[0125] Optionally, the model training module 605 is further used to:

[0126] Determine a sample text feature corresponding to the sample text;

[0127] Determining a first sample splicing feature according to the sample timbre feature and the sample emotion feature;

[0128] The sample sentiment feature and the sample text feature are concatenated to obtain a second sample concatenated feature;

[0129] The estimated Mel spectrum feature is determined according to the first sample splicing feature and the second sample splicing feature.

[0130] Optionally, the model training module 605 is further used to:

[0131] Determine a sample fundamental frequency and a sample energy according to the first sample splicing feature and the second sample splicing feature;

[0132] Performing splicing processing on the second sample splicing feature, the sample fundamental frequency and the sample energy to obtain a third sample splicing feature;

[0133] The estimated Mel spectrum feature is determined according to the first sample splicing feature and the third sample splicing feature.

[0134] Optionally, the model training module 605 is further used to:

[0135] The sample timbre feature and the sample emotion feature are spliced ​​together to obtain a spliced ​​feature of the pending sample;

[0136] The undetermined sample splicing feature is subjected to CLN conditional layer normalization processing to obtain the first sample splicing feature.

[0137] Through the above-mentioned device, in the process of training the speech synthesis model, the mutual information is used to determine the model to achieve the decoupling of timbre features and emotional features, so that in the process of determining the target Mel-spectrogram features according to the speech synthesis model, only the emotional features in the source speech are extracted, and the timbre features in the source speech are not extracted. In this way, the emotional features in the source speech are transferred to the target speech generated according to the target Mel-spectrogram features, and the timbre information of the source speech is not mixed in the emotional features, which improves the effect of emotional transfer, thereby making the sound quality of the generated target speech better.

[0138] Reference below Figure 8 , which shows a schematic diagram of the structure of an electronic device 800 suitable for implementing the embodiment of the present disclosure. The terminal device in the embodiment of the present disclosure may include but is not limited to mobile terminals such as mobile phones, laptop computers, digital broadcast receivers, PDAs (personal digital assistants), PADs (tablet computers), PMPs (portable multimedia players), vehicle-mounted terminals (such as vehicle-mounted navigation terminals), etc., and fixed terminals such as digital TVs, desktop computers, etc. Figure 8 The electronic device shown is only an example and should not bring any limitation to the functions and scope of use of the embodiments of the present disclosure.

[0139] like Figure 8As shown, the electronic device 800 may include a processing device (e.g., a central processing unit, a graphics processing unit, etc.) 801, which can perform various appropriate actions and processes according to a program stored in a read-only memory (ROM) 802 or a program loaded from a storage device 808 into a random access memory (RAM) 803. In the RAM 803, various programs and data required for the operation of the electronic device 800 are also stored. The processing device 801, the ROM 802, and the RAM 803 are connected to each other via a bus 804. An input / output (I / O) interface 805 is also connected to the bus 804.

[0140] Typically, the following devices may be connected to the I / O interface 805: input devices 806 including, for example, a touch screen, a touch pad, a keyboard, a mouse, a camera, a microphone, an accelerometer, a gyroscope, etc.; output devices 807 including, for example, a liquid crystal display (LCD), a speaker, a vibrator, etc.; storage devices 808 including, for example, a magnetic tape, a hard disk, etc.; and communication devices 809. The communication device 809 may allow the electronic device 800 to communicate with other devices wirelessly or by wire to exchange data. Although Figure 8 The electronic device 800 is shown with various devices, but it should be understood that it is not required to implement or possess all the devices shown. More or fewer devices may be implemented or possessed instead.

[0141] In particular, according to an embodiment of the present disclosure, the process described above with reference to the flowchart can be implemented as a computer software program. For example, an embodiment of the present disclosure includes a computer program product, which includes a computer program carried on a non-transitory computer-readable medium, and the computer program contains program code for executing the method shown in the flowchart. In such an embodiment, the computer program can be downloaded and installed from a network through a communication device 809, or installed from a storage device 808, or installed from a ROM 802. When the computer program is executed by the processing device 801, the above-mentioned functions defined in the method of the embodiment of the present disclosure are executed.

[0142] It should be noted that the computer-readable medium disclosed above may be a computer-readable signal medium or a computer-readable storage medium or any combination of the above two. The computer-readable storage medium may be, for example, but not limited to, an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor system, device or device, or any combination of the above. More specific examples of computer-readable storage media may include, but are not limited to: an electrical connection with one or more wires, a portable computer disk, a hard disk, a random access memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM or flash memory), an optical fiber, a portable compact disk read-only memory (CD-ROM), an optical storage device, a magnetic storage device, or any suitable combination of the above. In the present disclosure, a computer-readable storage medium may be any tangible medium containing or storing a program that may be used by or in combination with an instruction execution system, device or device. In the present disclosure, a computer-readable signal medium may include a data signal propagated in a baseband or as part of a carrier wave, in which a computer-readable program code is carried. This propagated data signal may take a variety of forms, including but not limited to an electromagnetic signal, an optical signal, or any suitable combination of the above. The computer readable signal medium may also be any computer readable medium other than a computer readable storage medium, which may send, propagate or transmit a program for use by or in conjunction with an instruction execution system, apparatus or device. The program code contained on the computer readable medium may be transmitted using any suitable medium, including but not limited to: wires, optical cables, RF (radio frequency), etc., or any suitable combination of the above.

[0143] In some embodiments, the client and the server may communicate using any currently known or future developed network protocol such as HTTP (HyperText Transfer Protocol), and may be interconnected with any form or medium of digital data communication (e.g., a communication network). Examples of communication networks include a local area network ("LAN"), a wide area network ("WAN"), an internet (e.g., the Internet), and a peer-to-peer network (e.g., an ad hoc peer-to-peer network), as well as any currently known or future developed network.

[0144] The computer-readable medium may be included in the electronic device, or may exist independently without being incorporated into the electronic device.

[0145] The computer-readable medium carries one or more programs. When the one or more programs are executed by the electronic device, the electronic device: obtains a target text and a target timbre identifier; determines a target emotion identifier of a source speech; inputs the target text, the target timbre identifier and the target emotion identifier into a speech synthesis model to obtain a target mel-spectrogram feature output by the speech synthesis model, wherein the speech synthesis model is obtained by training a target neural network model using a mutual information determination model and multiple sample sets; and generates a target speech based on the target mel-spectrogram feature.

[0146] Computer program code for performing the operations of the present disclosure may be written in one or more programming languages ​​or a combination thereof, including, but not limited to, object-oriented programming languages, such as Java, Smalltalk, C++, and conventional procedural programming languages, such as "C" or similar programming languages. The program code may be executed entirely on the user's computer, partially on the user's computer, as a separate software package, partially on the user's computer and partially on a remote computer, or entirely on a remote computer or server. In cases involving a remote computer, the remote computer may be connected to the user's computer through 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., via the Internet using an Internet service provider).

[0147] The flow chart and block diagram in the accompanying drawings illustrate the possible architecture, function and operation of the system, method and computer program product according to various embodiments of the present disclosure. In this regard, each square box in the flow chart or block diagram can represent a module, a program segment or a part of a code, and the module, the program segment or a part of the code contains one or more executable instructions for realizing the specified logical function. It should also be noted that in some implementations as replacements, the functions marked in the square box can also occur in a sequence different from that marked in the accompanying drawings. For example, two square boxes represented in succession can actually be executed substantially in parallel, and they can sometimes be executed in the opposite 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 a specified function or operation, or can be implemented with a combination of dedicated hardware and computer instructions.

[0148] The modules involved in the embodiments described in the present disclosure may be implemented by software or hardware. The name of the module does not limit the module itself in some cases. For example, the first acquisition module may also be described as a "module for acquiring target text and target timbre identification".

[0149] The functions described above herein may be performed at least in part by one or more hardware logic components. For example, without limitation, exemplary types of hardware logic components that may be used include: field programmable gate arrays (FPGAs), application specific integrated circuits (ASICs), application specific standard products (ASSPs), systems on chips (SOCs), complex programmable logic devices (CPLDs), and the like.

[0150] In the context of the present disclosure, a machine-readable medium may be a tangible medium that may contain or store a program for use by or in conjunction with an instruction execution system, device, or equipment. A machine-readable medium may be a machine-readable signal medium or a machine-readable storage medium. A machine-readable medium may include, but is not limited to, an electronic, magnetic, optical, electromagnetic, infrared, or semiconductor system, device, or equipment, or any suitable combination of the foregoing. A more specific example of a machine-readable storage medium may include an electrical connection based on one or more lines, a portable computer disk, a hard disk, a random access memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM or flash memory), an optical fiber, a portable compact disk read-only memory (CD-ROM), an optical storage device, a magnetic storage device, or any suitable combination of the foregoing.

[0151] According to one or more embodiments of the present disclosure, Example 1 provides a speech synthesis method, including: obtaining a target text and a target timbre identifier; determining a target emotion identifier of a source speech; inputting the target text, the target timbre identifier, and the target emotion identifier into a speech synthesis model to obtain a target mel-spectrogram feature output by the speech synthesis model, wherein the speech synthesis model is obtained by training a target neural network model through a mutual information determination model and multiple sample sets; and generating a target speech based on the target mel-spectrogram feature.

[0152] According to one or more embodiments of the present disclosure, Example 2 provides the method of Example 1, wherein the sample set includes sample text and sample speech information of a sample speech corresponding to the sample text, the sample speech information includes a sample timbre identifier, a sample emotion identifier, and a sample mel-spectrogram feature, and the mutual information determination model is used by the target neural network model to limit the mutual information between the sample timbre feature and the sample emotion feature, the sample timbre feature is the timbre feature corresponding to the sample timbre identifier, and the sample emotion feature is the emotion feature corresponding to the sample emotion identifier.

[0153] According to one or more embodiments of the present disclosure, Example 3 provides the method of Example 2, wherein the speech synthesis model is generated in the following manner: obtaining multiple sample sets; cyclically executing a model training step according to the multiple sample sets until the trained target neural network model meets a preset stop iteration condition, and using the trained target neural network model as the speech synthesis model, wherein the target neural network model includes a timbre feature extraction submodel and an emotion feature extraction submodel; the model training step includes: obtaining sample timbre features corresponding to the sample timbre identifier of the sample set through the timbre feature extraction submodel; obtaining sample emotion features corresponding to the sample emotion identifier of the sample set through the emotion feature extraction submodel; and obtaining sample emotion features corresponding to the sample timbre features according to the sample timbre features. and the sample emotional features, update the parameters of the timbre feature extraction submodel and the emotional feature extraction submodel through the mutual information determination model to obtain a new timbre feature extraction submodel and a new emotional feature extraction submodel; determine the estimated Mel-spectrogram features according to the sample timbre features, the sample emotional features and the sample text; determine a first loss value according to the estimated Mel-spectrogram features and the sample Mel-spectrogram features; when it is determined according to the first loss value that the target neural network model does not meet the preset stop iteration condition, update the parameters of the target neural network model according to the first loss value to obtain a trained target neural network model, and use the trained target neural network model as a new target neural network model.

[0154] According to one or more embodiments of the present disclosure, Example 4 provides the method of Example 3, wherein the method of updating the parameters of the timbre feature extraction sub-model and the emotion feature extraction sub-model through the mutual information determination model based on the sample timbre features and the sample emotion features includes: inputting the sample timbre features and the sample emotion features into the mutual information determination model to obtain the sample mutual information between the sample timbre features and the sample emotion features; determining a second loss value based on the sample mutual information; and updating the parameters of the timbre feature extraction sub-model and the emotion feature extraction sub-model based on the second loss value.

[0155] According to one or more embodiments of the present disclosure, Example 5 provides the method of Example 4, and before determining the second loss value based on the sample mutual information, the method also includes: determining whether the mutual information determination model satisfies a preset model stability condition based on the sample mutual information; determining the second loss value based on the sample mutual information includes: determining the second loss value based on the sample mutual information when it is determined that the mutual information determination model satisfies the preset model stability condition.

[0156] According to one or more embodiments of the present disclosure, Example 6 provides the method of Example 5, which also includes: when it is determined that the mutual information determination model does not satisfy the preset model stability condition, determining a third loss value based on the sample mutual information; updating the parameters of the mutual information determination model based on the third loss value to obtain a new mutual information determination model.

[0157] According to one or more embodiments of the present disclosure, Example 7 provides the method of Example 3, wherein determining the estimated Mel-spectrogram feature based on the sample timbre feature, the sample emotion feature and the sample text includes: determining a sample text feature corresponding to the sample text; determining a first sample splicing feature based on the sample timbre feature and the sample emotion feature; splicing the sample emotion feature and the sample text feature to obtain a second sample splicing feature; and determining the estimated Mel-spectrogram feature based on the first sample splicing feature and the second sample splicing feature.

[0158] According to one or more embodiments of the present disclosure, Example 8 provides the method of Example 7, wherein determining the estimated Mel-spectrogram feature based on the first sample splicing feature and the second sample splicing feature includes: determining the sample fundamental frequency and the sample energy based on the first sample splicing feature and the second sample splicing feature; performing splicing processing on the second sample splicing feature, the sample fundamental frequency, and the sample energy to obtain a third sample splicing feature; and determining the estimated Mel-spectrogram feature based on the first sample splicing feature and the third sample splicing feature.

[0159] According to one or more embodiments of the present disclosure, Example 9 provides the method of Example 7, wherein determining the first sample splicing feature based on the sample timbre feature and the sample emotion feature includes: performing splicing processing on the sample timbre feature and the sample emotion feature to obtain a pending sample splicing feature; performing CLN conditional layer normalization processing on the pending sample splicing feature to obtain the first sample splicing feature.

[0160] According to one or more embodiments of the present disclosure, Example 10 provides a speech synthesis device, comprising: a first acquisition module, for acquiring a target text and a target timbre identifier; a determination module, for determining a target emotion identifier of a source speech; a second acquisition module, for inputting the target text, the target timbre identifier and the target emotion identifier into a speech synthesis model to obtain a target mel-spectrogram feature output by the speech synthesis model, wherein the speech synthesis model is obtained by training a target neural network model through a mutual information determination model and multiple sample sets; and a generation module, for generating a target speech based on the target mel-spectrogram feature.

[0161] According to one or more embodiments of the present disclosure, Example 11 provides the device of Example 10, wherein the sample set includes sample text and sample speech information of a sample speech corresponding to the sample text, the sample speech information includes a sample timbre identifier, a sample emotion identifier, and a sample mel-spectrogram feature, and the mutual information determination model is used by the target neural network model to limit the mutual information between the sample timbre feature and the sample emotion feature, the sample timbre feature is the timbre feature corresponding to the sample timbre identifier, and the sample emotion feature is the emotion feature corresponding to the sample emotion identifier.

[0162] According to one or more embodiments of the present disclosure, Example 12 provides the apparatus of Example 11, wherein the apparatus further comprises: a model training module, which is used to: obtain multiple sample sets; cyclically execute the model training steps according to the multiple sample sets until the trained target neural network model meets the preset stop iteration condition, and use the trained target neural network model as the speech synthesis model, wherein the target neural network model comprises a timbre feature extraction submodel and an emotion feature extraction submodel; the model training steps comprise: obtaining sample timbre features corresponding to the sample timbre identification of the sample set through the timbre feature extraction submodel; obtaining sample emotion features corresponding to the sample emotion identification of the sample set through the emotion feature extraction submodel; and obtaining sample emotion features corresponding to the sample emotion identification of the sample set according to the sample timbre feature extraction submodel. The mutual information determines the parameters of the timbre feature extraction sub-model and the emotion feature extraction sub-model according to the sample timbre feature, the sample emotion feature and the sample text, and determines the estimated Mel-spectrogram feature according to the estimated Mel-spectrogram feature and the sample Mel-spectrogram feature. When it is determined according to the first loss value that the target neural network model does not meet the preset stop iteration condition, the parameters of the target neural network model are updated according to the first loss value to obtain the trained target neural network model, and the trained target neural network model is used as the new target neural network model.

[0163] According to one or more embodiments of the present disclosure, Example 13 provides the device of Example 12, and the model training module is further used to: input the sample timbre features and the sample emotion features into the mutual information determination model to obtain the sample mutual information between the sample timbre features and the sample emotion features; determine a second loss value based on the sample mutual information; and update the parameters of the timbre feature extraction submodel and the emotion feature extraction submodel based on the second loss value.

[0164] According to one or more embodiments of the present disclosure, Example 14 provides the device of Example 13, wherein the model training module is further used to: determine whether the mutual information determination model satisfies a preset model stability condition based on the sample mutual information; determining the second loss value based on the sample mutual information includes: determining the second loss value based on the sample mutual information when it is determined that the mutual information determination model satisfies the preset model stability condition.

[0165] According to one or more embodiments of the present disclosure, Example 15 provides the device of Example 14, and the model training module is also used to: determine a third loss value based on the sample mutual information when it is determined that the mutual information determination model does not meet the preset model stability condition; update the parameters of the mutual information determination model based on the third loss value to obtain a new mutual information determination model.

[0166] According to one or more embodiments of the present disclosure, Example 16 provides the device of Example 12, and the model training module is further used to: determine the sample text features corresponding to the sample text; determine the first sample splicing feature based on the sample timbre features and the sample emotion features; splice the sample emotion features and the sample text features to obtain the second sample splicing feature; determine the estimated Mel-spectrogram feature based on the first sample splicing feature and the second sample splicing feature.

[0167] According to one or more embodiments of the present disclosure, Example 17 provides the device of Example 16, and the model training module is further used to: determine the sample fundamental frequency and the sample energy according to the first sample splicing feature and the second sample splicing feature; perform splicing processing on the second sample splicing feature, the sample fundamental frequency and the sample energy to obtain a third sample splicing feature; determine the estimated Mel spectrum feature according to the first sample splicing feature and the third sample splicing feature.

[0168] According to one or more embodiments of the present disclosure, Example 18 provides the device of Example 16, and the model training module is further used to: perform splicing processing on the sample timbre features and the sample emotion features to obtain the pending sample splicing features; perform CLN conditional layer normalization processing on the pending sample splicing features to obtain the first sample splicing features.

[0169] According to one or more embodiments of the present disclosure, Example 19 provides a computer-readable medium having a computer program stored thereon, which, when executed by a processing device, implements the steps of the method described in any one of Examples 1-9.

[0170] According to one or more embodiments of the present disclosure, Example 20 provides an electronic device, comprising: a storage device on which a computer program is stored; and a processing device for executing the computer program in the storage device to implement the steps of the method described in any one of Examples 1-9.

[0171] The above description is only a preferred embodiment of the present disclosure and an explanation of the technical principles used. Those skilled in the art should understand that the scope of disclosure involved in the present disclosure is not limited to the technical solutions formed by a specific combination of the above technical features, but should also cover other technical solutions formed by any combination of the above technical features or their equivalent features without departing from the above disclosed concept. For example, the above features are replaced with the technical features with similar functions disclosed in the present disclosure (but not limited to) by each other to form a technical solution.

[0172] In addition, although each operation is described in a specific order, this should not be understood as requiring these operations to be performed in the specific order shown or in a sequential order. Under certain circumstances, multitasking and parallel processing may be advantageous. Similarly, although some specific implementation details are included in the above discussion, these should not be interpreted as limiting the scope of the present disclosure. Some features described in the context of a separate embodiment can also be implemented in a single embodiment in combination. On the contrary, the various features described in the context of a single embodiment can also be implemented in multiple embodiments individually or in any suitable sub-combination mode.

[0173] Although the subject matter has been described in language specific to structural features and / or method logic actions, it should be understood that the subject matter defined in the appended claims is not necessarily limited to the specific features or actions described above. On the contrary, the specific features and actions described above are merely example forms of implementing the claims. Regarding the device in the above embodiment, the specific manner in which each module performs the operation has been described in detail in the embodiment related to the method, and will not be elaborated here.

Claims

1. A speech synthesis method, characterized in that: include: Obtaining target text and target timbre identification; Determine the target emotion label of the source speech; Inputting the target text, the target timbre identifier and the target emotion identifier into a speech synthesis model to obtain target mel-spectrogram features output by the speech synthesis model, wherein the speech synthesis model is obtained by training a target neural network model through a mutual information determination model and multiple sample sets; Generate target speech according to the target mel-spectrogram features.

2. The method according to claim 1, characterized in that The sample set includes sample text and sample speech information of the sample speech corresponding to the sample text, the sample speech information includes a sample timbre identifier, a sample emotion identifier and a sample mel spectrum feature, the mutual information determination model is used by the target neural network model to limit the mutual information between the sample timbre feature and the sample emotion feature, the sample timbre feature is the timbre feature corresponding to the sample timbre identifier, and the sample emotion feature is the emotion feature corresponding to the sample emotion identifier.

3. The method according to claim 2, characterized in that The speech synthesis model is generated in the following way: Obtain multiple sample sets; Execute the model training step cyclically according to the plurality of sample sets until the trained target neural network model meets the preset stop iteration condition, and use the trained target neural network model as the speech synthesis model, wherein the target neural network model includes a timbre feature extraction sub-model and an emotion feature extraction sub-model; The model training step includes: Acquire the sample timbre features corresponding to the sample timbre identifiers of the sample set through the timbre feature extraction sub-model; Acquire the sample emotion feature corresponding to the sample emotion identifier of the sample set through the emotion feature extraction sub-model; According to the sample timbre features and the sample emotional features, the parameters of the timbre feature extraction submodel and the emotional feature extraction submodel are updated through the mutual information determination model to obtain a new timbre feature extraction submodel and a new emotional feature extraction submodel; Determining an estimated mel spectrum feature according to the sample timbre feature, the sample emotion feature and the sample text; Determining a first loss value according to the estimated Mel spectrum feature and the sample Mel spectrum feature; When it is determined according to the first loss value that the target neural network model does not satisfy the preset stop iteration condition, the parameters of the target neural network model are updated according to the first loss value to obtain a trained target neural network model, and the trained target neural network model is used as a new target neural network model.

4. The method according to claim 3, characterized in that The updating of the parameters of the timbre feature extraction sub-model and the emotion feature extraction sub-model by the mutual information determination model according to the sample timbre feature and the sample emotion feature comprises: Inputting the sample timbre feature and the sample emotion feature into the mutual information determination model to obtain sample mutual information between the sample timbre feature and the sample emotion feature; Determine a second loss value according to the sample mutual information; Update the parameters of the timbre feature extraction sub-model and the emotion feature extraction sub-model according to the second loss value.

5. The method according to claim 4, characterized in that Before determining the second loss value according to the sample mutual information, the method further includes: Determining whether the mutual information determination model satisfies a preset model stability condition according to the sample mutual information; Determining the second loss value according to the sample mutual information includes: When it is determined that the mutual information determination model satisfies the preset model stability condition, a second loss value is determined according to the sample mutual information.

6. The method according to claim 5, characterized in that The method further comprises: In the case where it is determined that the mutual information determination model does not satisfy the preset model stability condition, determining a third loss value according to the sample mutual information; The parameters of the mutual information determination model are updated according to the third loss value to obtain a new mutual information determination model.

7. The method according to claim 3, characterized in that The determining of the estimated Mel spectrum feature according to the sample timbre feature, the sample emotion feature and the sample text comprises: Determining a sample text feature corresponding to the sample text; Determining a first sample splicing feature according to the sample timbre feature and the sample emotion feature; Performing splicing processing on the sample sentiment feature and the sample text feature to obtain a second sample splicing feature; The estimated Mel spectrum feature is determined according to the first sample splicing feature and the second sample splicing feature.

8. The method according to claim 7, characterized in that The determining the estimated Mel spectrum feature according to the first sample splicing feature and the second sample splicing feature includes: Determine a sample fundamental frequency and a sample energy according to the first sample splicing feature and the second sample splicing feature; Performing splicing processing on the second sample splicing feature, the sample fundamental frequency, and the sample energy to obtain a third sample splicing feature; The estimated Mel spectrum feature is determined according to the first sample splicing feature and the third sample splicing feature.

9. The method according to claim 7, characterized in that: Determining the first sample splicing feature according to the sample timbre feature and the sample emotion feature includes: The sample timbre feature and the sample emotion feature are spliced ​​to obtain a pending sample splicing feature; The undetermined sample splicing feature is subjected to CLN conditional layer normalization processing to obtain the first sample splicing feature.

10. A speech synthesis device, characterized in that: include: A first acquisition module acquires a target text and a target timbre identifier; A determination module, used to determine the target emotion identifier of the source speech; A second acquisition module is used to input the target text, the target timbre identifier and the target emotion identifier into a speech synthesis model to obtain a target mel spectrum feature output by the speech synthesis model, wherein the speech synthesis model is obtained by training a target neural network model through a mutual information determination model and multiple sample sets; A generation module is used to generate a target speech according to the target Mel spectrum features.

11. A computer readable medium having a computer program stored thereon, characterized in that: When the program is executed by a processing device, the steps of the method described in any one of claims 1 to 9 are implemented.

12. An electronic device, characterized in that: include: a storage device having at least one computer program stored thereon; At least one processing device, configured to execute the at least one computer program in the storage device to implement the steps of the method according to any one of claims 1 to 9.

Citation Information

Cited By

  • Speech synthesis method and device, electronic equipment and storage medium

    CN119763545A

  • Speech synthesis method, device, electronic device and storage medium

    CN119763545B