Method and device for training prosody prediction model, and method and device for human-computer interaction

Abstract

The embodiment of the present specification provides a training method and device of a prosody prediction model, and a man-machine interaction method and device. The training method comprises: obtaining text corpus in a target service scenario; performing normalization processing on the text corpus; obtaining sample corpus; wherein the sample corpus is obtained by marking the prosody label on the normalized text corpus; the prosody label is used to indicate the pause duration; the model structure and parameters of the punctuation symbol prediction model are used to initialize the model structure and parameters of the prosody prediction model; and the sample corpus with the prosody label is input into the initialized prosody prediction model to train the prosody prediction model. The method of the embodiment of the present specification can make the voice played by the machine device more natural and more convenient for the user to understand.

Description

TECHNICAL FIELD

[0001] The one or more embodiments of the present specification relate to network communication technology, and particularly relate to a method and apparatus for training a prosody prediction model, and a method and apparatus for human-computer interaction. BACKGROUND

[0002] At present, there are more and more human-computer interaction services. A machine device can broadcast corresponding dialogues according to the dialogue content of a user. In a real intelligent dialogue, it is impossible to synthesize a voice in advance in the machine device, and the machine device needs to generate and broadcast a voice in real time according to the dialogue content sent by the user in real time.

[0003] When a human being has a dialogue, the voice spoken is prosodic, such as where to pause after a word and how long the pause is. In human-computer interaction, in order to make the voice broadcast by the machine device more natural and easier to understand, the machine device also needs to process the prosody of the voice content generated in real time, and play the voice according to the prosody after processing.

[0004] In the existing technology, the prosody processing method in human-computer interaction mainly includes: after the machine device generates a voice content to be broadcast, the machine device processes the voice content by word segmentation, and sets a prosody label after each word according to the word segmentation result. When broadcasting the voice content, a pause is made when playing to the position of the prosody label.

[0005] The prosody processing method in the prior art will cause the voice broadcast by the machine device to be unnatural and not easy for the user to understand, and reduce the practicability of human-computer interaction. SUMMARY

[0006] The one or more embodiments of the present specification describe a method and apparatus for training a prosody prediction model, and a method and apparatus for human-computer interaction, which can make the voice broadcast by the machine device more natural and easier for the user to understand.

[0007] According to a first aspect, a method for training a prosody prediction model is provided, wherein the method comprises:

[0008] obtaining text corpus in a target service scenario;

[0009] performing normalization processing on the text corpus;

[0010] obtaining sample corpus; wherein the sample corpus is obtained by marking the normalized text corpus with a prosody label; the prosody label is used to indicate the pause duration;

[0011] initializing the model structure and parameters of the prosody prediction model by using the model structure and parameters of the punctuation symbol prediction model trained;

[0012] inputting the sample corpus with the prosody labels into the initialized prosody prediction model to train the prosody prediction model.

[0013] The normalization processing includes at least one of the following: removing non-Chinese characters in the text corpus that do not affect semantic understanding, and converting non-Chinese characters in the text corpus that affect semantic understanding into corresponding Chinese characters.

[0014] The prosody labels are used to indicate tones of the speech.

[0015] The prosody labels include at least one of the following: word segmentation, small pause, large pause, and end of sentence.

[0016] The initialization of the model structure and parameters of the prosody prediction model by using the model structure and parameters of the trained punctuation prediction model further includes:

[0017] removing a full connection layer in a network structure of the punctuation prediction model, and setting categories output by the model as category labels included in the prosody labels.

[0018] The prosody prediction model includes a text preprocessing module, a Word2Vec module, and a BiLSTM module.

[0019] The text preprocessing module is configured to perform normalization processing on the text corpus to obtain a sample corpus.

[0020] The Word2Vec module is configured to obtain embedding values of each character in the sample corpus.

[0021] The BiLSTM module is configured to output a prediction result according to the embedding values of each character in the sample corpus.

[0022] According to a second aspect, a human-computer interaction method is provided, and the method includes:

[0023] obtaining interaction content to be played to a user;

[0024] inputting the interaction content into a pre-trained prosody prediction model to obtain the interaction content with prosody labels output by the prosody prediction model;

[0025] playing the interaction content to the user in a prosody-carrying speech mode according to the prosody labels carried in the interaction content.

[0026] According to a third aspect, a training device of a prosody prediction model is provided, and the device includes:

[0027] The sample obtaining module is configured to obtain text corpus in a target service scenario, perform normalization processing on the text corpus, and obtain sample corpus, wherein the sample corpus is obtained by marking the text corpus after the normalization processing with prosody labels, and the prosody labels are used to indicate pause duration.

[0028] The transfer learning module is configured to initialize model structure and parameters of the prosody prediction model by using model structure and parameters of the punctuation prediction model after the training.

[0029] The training execution module is configured to input the sample corpus with the prosody labels into the initialized prosody prediction model to train the prosody prediction model.

[0030] According to a fourth aspect, a human-computer interaction device is provided, wherein the device comprises:

[0031] The interaction content obtaining module is configured to obtain interaction content that needs to be played to a user.

[0032] The prosody processing module is configured to input the interaction content into the pre-trained prosody prediction model to obtain interaction content with prosody labels output by the prosody prediction model, wherein the prosody prediction model is trained by using the device for training the prosody prediction model in the embodiments of the present specification.

[0033] The playing module is configured to play the interaction content to the user in a prosody-enabled voice mode according to the prosody labels carried in the interaction content.

[0034] According to a fifth aspect, a computing device is provided, comprising a memory and a processor, wherein the memory stores executable code, and the processor executes the executable code to implement the method according to any of the embodiments of the present specification.

[0035] The method and device for training a prosody prediction model, the human-computer interaction method and device provided in the embodiments of the present specification use a punctuation prediction model to train a prosody prediction model. Considering that the punctuation prediction model has the function of adding punctuation to a piece of text content, and punctuation actually has strong correlation with prosody, that is, where there is punctuation, it is necessary to pause in prosody. Therefore, the ability of the punctuation prediction model is used to train the prosody prediction model, so that the trained prosody prediction model can mark the positions where various punctuation should be added in a piece of text content in subsequent actual prediction services, thereby having the ability of prosody prediction. BRIEF DESCRIPTION OF DRAWINGS

[0036] In order to more clearly illustrate the technical solutions in the embodiments of the present specification or the prior art, the accompanying drawings needed to be used in the embodiments or prior art description will be briefly introduced as follows. Obviously, the accompanying drawings in the following description are some embodiments of the present specification, and for those skilled in the art, other drawings can also be obtained without creative labor based on these drawings.

[0037] Figure 1 is a schematic diagram of a system architecture to which an embodiment of the present specification is applied.

[0038] Figure 2 is a flowchart of a training method of a prosody prediction model in an embodiment of the present specification.

[0039] Figure 3 is a flowchart of a human-computer interaction method in an embodiment of the present specification.

[0040] Figure 4 is a structural schematic diagram of a training device of a prosody prediction model in an embodiment of the present specification.

[0041] Figure 5 is a structural schematic diagram of a human-computer interaction device in an embodiment of the present specification. DETAILED DESCRIPTION

[0042] As described above, in the prior art, the machine device sets the prosody label according to the result of word segmentation, so as to meet the prosody requirement of the voice broadcast by the machine device. However, in human conversation, the pause is not only according to the word. Especially when there is no punctuation mark to separate the words in the sentence to be broadcast and the text is long, if the pause is only according to the word, the broadcast voice will be unnatural and mechanical, and in serious cases, it will also affect the understanding of the listener. For example, “today the weather is particularly good”, if the machine device pauses between the four words “today”, “weather”, “particularly” and “good” respectively, it will lead to the broadcast voice being mechanical and unnatural. The longer the voice to be broadcast by the machine device, the more serious the problem will be. For example, the machine device should pause according to the way of “today the weather” and “particularly good” according to the language habits of human beings.

[0043] Therefore, the prosody processing method of the prior art cannot meet the prosody requirement of voice broadcast in human-computer interaction, greatly reducing the satisfaction of human-computer interaction.

[0044] The solutions provided by the present specification will be described below with reference to the accompanying drawings.

[0045] It should be noted that the terms used in the embodiments of the present application are merely for the purpose of describing specific embodiments and are not intended to limit the present application. The singular forms "a", "said" and "the" used in the embodiments of the present application and the appended claims are also intended to include the plural forms unless the context clearly indicates otherwise.

[0046] It should be understood that the term "and / or" used herein is merely a description of the association relationship of the associated objects, which means that there can be three relationships, for example, A and / or B, which can represent the three cases of A alone, A and B together, and B alone. In addition, the character " / " herein generally represents an "or" relationship between the front and rear associated objects.

[0047] In order to facilitate the understanding of the method provided in the present specification, first, the system architecture related to and applicable to the present specification is described. As shown in Figure 1 The system architecture mainly includes a user and a machine device for human-computer interaction.

[0048] The machine device for human-computer interaction with the user can include but is not limited to, such as: smart mobile terminal, smart home device, network device, wearable device, smart medical device, PC (personal computer) and the like. The smart mobile device can include, such as: mobile phone, tablet computer, notebook computer, PDA (personal digital assistant), Internet car and the like. The smart home device can include smart home appliance devices, such as: smart TV, smart air conditioner, smart water heater, smart refrigerator, smart air purifier and the like, the smart home device can also include smart door lock, smart socket, smart lamp, smart camera and the like. The network device can include, such as: switch, wireless AP, server and the like. The wearable device can include, such as: smart watch, smart glasses, smart bracelet, virtual reality device, augmented reality device, mixed reality device (i.e. a device that can support virtual reality and augmented reality) and the like. The smart medical device can include, such as: smart thermometer, smart sphygmomanometer, smart blood glucose meter and the like.

[0049] It should be understood that Figure 1 The number of users and machine devices in

[0050] In the embodiments of the present specification, first, the prosody prediction model is trained, and then the prosody prediction model is used for human-computer interaction.

[0051] First, the training method of the prosody prediction model in the embodiments of the present specification is described.

[0052] Figure 2is a flowchart of a training method of a prosody prediction model in an embodiment of the present specification. The execution subject of the method is a training device of the prosody prediction model. It can be understood that the method can also be executed by any device, equipment, platform, cluster of equipment with computing and processing capabilities. Referring to Figure 2 , the method comprises:

[0053] Step 201: Obtain text corpus in a target service scenario.

[0054] Step 203: Perform normalization processing on the text corpus.

[0055] Step 205: Obtain sample corpus; wherein the sample corpus is obtained by marking the prosody label on the normalized text corpus; and the prosody label is used to indicate the pause duration.

[0056] Step 207: Initialize the model structure and parameters of the prosody prediction model by using the model structure and parameters of the punctuation prediction model.

[0057] Step 209: Input the sample corpus with the prosody label into the initialized prosody prediction model to train the prosody prediction model.

[0058] It can be seen that, in the training method of the prosody prediction model of the embodiment of the present specification shown in Figure 2 , the punctuation prediction model is used to train the prosody prediction model. Considering that the punctuation prediction model has the function of adding punctuation to a piece of text content, and the punctuation actually has strong correlation with the prosody, that is, the place with punctuation is the place where the prosody needs to be paused, therefore, the ability of the punctuation prediction model is used to train the prosody prediction model, so that the trained prosody prediction model can mark the prosody label on the position where the punctuation is needed in a piece of text in the subsequent actual prediction service, indicating that the prosody needs to be paused, thereby having the ability of prosody prediction.

[0059] If the artificial way is used to perform the artificial marking processing on the prosody of the massive text corpus, thereby obtaining the massive training samples, it can be understood that the process efficiency is extremely low and the cost is extremely large. Referring to Figure 2 the method of the embodiment of the present specification, because the model structure and parameters of the punctuation prediction model are used, when the prosody prediction model is trained, the massive training samples after the artificial marking on the prosody are not needed, that is, the small sample can be used to train the prosody prediction model meeting the service requirements.

[0060] In Figure 2In the method of the embodiment of the present specification, the prosody prediction model is not trained in the manner of words, and the prosody prediction model determines which words or characters are followed by punctuation marks in history according to the training process, which is actually which words or characters should be followed by prosodic pauses in history. Therefore, even if a sentence to be broadcast does not have punctuation marks and the text is long, the prosody prediction model can accurately predict the prosodic pauses in the long text according to the trained ability, so that the played voice is natural and does not affect the understanding of the listener.

[0061] The following describes Figure 2 Each step of the process is described separately.

[0062] First, for step 201: obtaining text corpus in the target business scenario.

[0063] For example, the target business scenario is a customer service scenario, such as handling customer complaints, so in this step 201, customer service-related text corpus can be obtained for subsequent generation of training samples.

[0064] Next, for step 203: normalizing the text corpus.

[0065] In this step 203, the normalization is to enable subsequent machine devices to recognize the text corpus. The normalization can include at least one of the following:

[0066] Removing non-Chinese characters in the text corpus that do not affect semantic understanding;

[0067] Converting non-Chinese characters in the text corpus that affect semantic understanding to corresponding Chinese characters, such as converting “￥” to the Chinese character “Renminbi”.

[0068] Next, for step 205: obtaining sample corpus; wherein the sample corpus is obtained by labeling the normalized text corpus with prosodic labels; the prosodic labels are used to indicate pause duration.

[0069] In the embodiment of the present specification, the prosodic labels are used to indicate pause duration, such as prosodic label “#1” representing word segmentation (word segmentation can correspond to the smallest pause duration), prosodic label “#2” representing small pause (small pause can correspond to a smaller pause duration such as 0.2 seconds), prosodic label “#3” representing large pause (large pause can correspond to a longer pause duration such as 0.5 seconds), and prosodic label “#4” representing end of sentence (end of sentence can correspond to the longest pause duration such as 1 second).

[0070] In the embodiment of the present specification, the prosody label can be further used to indicate the tone of the voice. For example, the prosody label "#5" represents playing with a rising tone, and the prosody label "#6" represents playing with a falling tone.

[0071] In step 205, artificial marking is performed, that is, artificial prosody labels are marked on the normalized text corpus. For example, if a long text does not contain punctuation marks, an artificial mark is made to indicate which word in the long text should be followed by a prosody label. For example, which word in the long text should be followed by a prosody label "#1", which word in the long text should be followed by a prosody label "#2", and which word in the long text should be followed by a prosody label "#5".

[0072] As described above, because the prosody prediction model is trained on the basis of the punctuation prediction model in the embodiment of the present specification, a large number of training samples are not required, that is, only a small amount of text corpus needs to be artificially marked to generate a small number of training samples, such as several hundred training samples.

[0073] Next, step 207 is performed: the model structure and parameters of the punctuation prediction model are used to initialize the model structure and parameters of the prosody prediction model.

[0074] Like the model structure of the punctuation prediction model, the prosody prediction model includes a text preprocessing module, a Word2Vec module, and a BiLSTM module.

[0075] In the prosody prediction model, the functions of the text preprocessing module include: performing normalization processing on the text corpus to obtain a sample corpus;

[0076] The functions of the Word2Vec module include: obtaining the embedding value of each word in the sample corpus;

[0077] The functions of the BiLSTM module include: outputting a prediction result according to the embedding value of each word in the sample corpus. The categories of the prosody labels output by the BiLSTM module can be determined according to the artificial labels.

[0078] In one embodiment of the present specification, during step 207, the prosody prediction model can be fine-tuned in the initialization stage according to the difference between the functions of the prosody prediction model and the functions of the punctuation prediction model. For example, the number of categories of the punctuation label output by the punctuation prediction model is 4, and the number of categories of the prosody label output by the prosody prediction model is 5. Therefore, the initialization stage further includes: removing the fully connected layer in the network structure of the punctuation prediction model; and setting various categories output by the model to various categories included in the prosody label.

[0079] Next, for step 209: input the sample corpus with prosody labels into the initialized prosody prediction model to train the prosody prediction model.

[0080] The initialized prosody prediction model has the ability to predict punctuation marks. After inputting the sample corpus with artificial prosody labels into the initialized prosody prediction model, the initialized prosody prediction model can learn the correspondence between punctuation marks and prosody labels. For example, because the initialized prosody prediction model has the ability to predict punctuation marks, for the input sample corpus, if it is determined that the Chinese character 1 in the sample corpus should be followed by a comma, and the artificial prosody label is “#2”, then the prosody prediction model can determine that the prosody label “#2” is output for the position of the character that should add a comma. At the same time, the prosody prediction model can also learn how to prosody mark the text according to the artificial prosody label.

[0081] It can be seen that in the embodiments of the present specification, the prosody prediction model has the ability to first predict punctuation marks and then predict prosody labels according to the punctuation marks; and the ability to directly predict prosody labels according to the text content.

[0082] After multiple rounds of training, the prosody prediction model is trained.

[0083] After the prosody prediction model is trained, in subsequent human-computer interaction, the prosody prediction model can be used to prosody process the speech to be played, that is, to mark the prosody label for the speech to be played, so that the machine device can pause according to the marked prosody label during playing the speech. Referring to Figure 3 , the human-computer interaction method comprises:

[0084] Step 301: obtain interactive content to be played to a user.

[0085] Step 303: input the interactive content into the pre-trained prosody prediction model to obtain the interactive content with prosody labels output by the prosody prediction model.

[0086] Step 305: play the interactive content to the user in a prosody speech manner according to the prosody labels carried in the interactive content.

[0087] For example, the interactive content with prosody labels includes “XX#1XXXX#2XXXXX#2XXXX#3……”, so that the machine device pauses for 0.1 seconds after playing the text before “#1” and then continues to play, and then pauses for 0.2 seconds after playing the text before the first “#2” and then continues to play. In this way, the machine device can play the speech with prosody similar to that of human speech.

[0088] In an embodiment of the present specification, a device for training a prosody prediction model is also provided, referring to Figure 4 The device comprises:

[0089] The sample obtaining module 401 is configured to obtain text corpus in a target service scenario; perform normalization processing on the text corpus; obtain sample corpus; wherein the sample corpus is obtained by marking the text corpus after normalization processing with a prosody label; the prosody label is used to indicate the duration of the pause;

[0090] The transfer learning module 402 is configured to initialize the model structure and parameters of the prosody prediction model by using the model structure and parameters of the punctuation prediction model after training.

[0091] The training execution module 403 is configured to input the sample corpus with the prosody label into the initialized prosody prediction model to train the prosody prediction model.

[0092] In the embodiment of the device of the present specification shown in Figure 4 In the embodiment of the device of the present specification shown in

[0093] In the embodiment of the device of the present specification shown in Figure 4 In the embodiment of the device of the present specification shown in

[0094] In the embodiment of the device of the present specification shown in Figure 4 In the embodiment of the device of the present specification shown in

[0095] In the embodiment of the device of the present specification shown in Figure 4 In the embodiment of the device of the present specification shown in

[0096] Remove the full connection layer in the network structure of the punctuation prediction model; set the class of the model output as the class label included in the prosody label.

[0097] In the embodiment of the device of the present specification shown in Figure 4 In the embodiment of the device of the present specification shown in

[0098] The text preprocessing module is used to perform normalization processing on the text corpus to obtain the sample corpus;

[0099] The Word2Vec module is used to obtain the embedding value of each word in the sample corpus;

[0100] The BiLSTM module is configured to output a prediction result according to an embedding value of each word in a sample corpus.

[0101] One embodiment of the present specification also provides a human-computer interaction device, which is arranged in a machine device for human-computer interaction with a user. Referring to Figure 5 The device comprises:

[0102] An interaction content obtaining module 501 is configured to obtain interaction content to be played to a user;

[0103] A prosody processing module 502 is configured to input the interaction content into a pre-trained prosody prediction model to obtain interaction content with prosody labels output by the prosody prediction model; wherein the prosody prediction model is trained by the prosody prediction model training device in the embodiment of the present specification;

[0104] A playing module 503 is configured to play the interaction content to the user in a prosody carrying manner according to the prosody labels carried in the interaction content.

[0105] It should be noted that the above devices are usually implemented on a server side, and can be arranged on independent servers, or some or all of the devices can be combined and arranged on the same server. The server can be a single server or a server cluster composed of multiple servers, and the server can be a cloud server, also known as a cloud computing server or a cloud host, which is a host product in a cloud computing service system. The above devices can also be implemented on a computer terminal with strong computing power.

[0106] One embodiment of the present specification provides a computer readable storage medium having a computer program stored thereon, when the computer program is executed in a computer, the computer program causes the computer to execute the method in any one of the embodiments of the present specification.

[0107] One embodiment of the present specification provides a computing device comprising a memory and a processor, the memory having executable code stored therein, and the processor executes the executable code to implement the method in any one of the embodiments of the present specification.

[0108] It can be understood that the structure illustrated in the embodiments of the present specification does not constitute a specific limitation on the device of the embodiments of the present specification. In other embodiments of the present specification, the above device can include more or fewer components than the illustration, or combine certain components, or split certain components, or different arrangement of components. The illustrated components can be implemented in hardware, software, or a combination of software and hardware.

[0109] The various embodiments described in this specification are presented by way of example, and embodiments can be implemented in any combination of hardware and / or software. As such, the application is not limited to any one implementation. In addition, the various embodiments described in this specification are not necessarily mutually exclusive, as some embodiments can be combined with one or more other embodiments to form new embodiments.

[0110] Those skilled in the art should understand that, in one or more examples described above, the functions described in the specification can be implemented in hardware, software, a plug-in, or any combination thereof. When implemented in software, these functions can be stored in a computer readable medium or transmitted as one or more instructions or codes on a computer readable medium.

[0111] The above detailed description has been presented for the purposes of clarity and understanding. It is not intended to be exhaustive or to limit the application to the precise form described. Many modifications and variations are possible in the light of the above teachings. It is intended that the scope of the application be limited not with this detailed description.

Claims

1. A method for training a prosody prediction model, the method comprising: obtaining a text corpus in a target business scenario; normalizing the text corpus; obtaining a sample corpus, wherein the sample corpus is obtained by labeling the normalized text corpus with prosody labels, and the prosody labels are used to indicate pause durations; initializing a model structure and parameters of the prosody prediction model using a model structure and parameters of a punctuation prediction model that has been trained; inputting the sample corpus with the prosody labels into the initialized prosody prediction model to train the prosody prediction model; wherein the initializing the model structure and parameters of the prosody prediction model using the model structure and parameters of the punctuation prediction model further comprises: removing a fully connected layer in a network structure of the punctuation prediction model; and setting categories output by the model to categories included in the prosody labels; the prosody prediction model obtained by the training has the abilities of predicting punctuation, predicting prosody labels based on the punctuation, and predicting the prosody labels directly based on text content. The normalizing comprises at least one of the following: removing non-Chinese characters in the text corpus that do not affect semantic understanding, and converting non-Chinese characters in the text corpus that affect semantic understanding into Chinese characters corresponding to the semantic understanding. The prosody labels are also used to indicate tones of speech. The prosody labels comprise at least one of the following: word segmentation, small pause, large pause, and end of a sentence. The prosody prediction model comprises a text preprocessing module, a Word2Vec module, and a BiLSTM module, wherein the text preprocessing module is configured to perform the normalizing on the text corpus to obtain the sample corpus; the Word2Vec module is configured to obtain embedding values of each character in the sample corpus; and the BiLSTM module is configured to output a prediction result based on the embedding values of each character in the sample corpus. 6.A human-computer interaction method, the method comprising: obtaining interaction content to be played to a user; inputting the interaction content into a prosody prediction model that has been pre-trained to obtain the interaction content with prosody labels output by the prosody prediction model, wherein the prosody prediction model is trained using the method of any one of claims 1 to 5; and playing the interaction content to the user in a prosody-carrying speech manner according to the prosody labels carried in the interaction content. 7.An apparatus for training a prosody prediction model, the apparatus comprising: a sample obtaining module configured to obtain a text corpus in a target business scenario; a normalizing module configured to normalize the text corpus; a sample corpus obtaining module configured to obtain a sample corpus, wherein the sample corpus is obtained by labeling the normalized text corpus with prosody labels, and the prosody labels are used to indicate pause durations; a transfer learning module configured to initialize a model structure and parameters of the prosody prediction model using a model structure and parameters of a punctuation prediction model that has been trained; and a training execution module configured to input the sample corpus with the prosody labels into the initialized prosody prediction model to train the prosody prediction model. ​ 2. The method of claim 1, wherein, ​ 3. The method of claim 1, wherein, ​ 4. The method of claim 1, wherein, ​ 5. The method of claim 1, wherein, ​ ​ ​ ​ ​ ​ ​ ​ ​ ​ ​ ​ ​ ​ The migration learning module is further configured to perform: removing a full connection layer in a network structure of the punctuation prediction model; and setting a class of a model output to a class label included in the prosody label. The prosody prediction model obtained by training has the abilities of predicting punctuation, predicting a prosody label according to punctuation, and directly predicting a prosody label according to text content.

8. An interactive device, wherein the device comprises: an interactive content obtaining module configured to obtain interactive content to be played to a user; a prosody processing module configured to input the interactive content into a pre-trained prosody prediction model to obtain interactive content with a prosody label output by the prosody prediction model, wherein the prosody prediction model is trained by the device of claim 7; a playing module configured to play the interactive content to the user in a prosody-enabled voice manner according to the prosody label carried in the interactive content.

9. A computing device comprising a memory and a processor, wherein the memory stores executable code, and the processor executes the executable code to implement the method of any one of claims 1-6.

Images