Speech synthesis method and system, electronic equipment and storage medium

Abstract

The invention provides a speech synthesis method and system, electronic equipment and a storage medium, and the method comprises the steps: inputting the speech features of the source speech of a target object into an encoder in a speech synthesis model, and obtaining the first coding feature and the second coding feature of the source speech; inputting the first coding feature of the source voice into an age sensing module in a voice synthesis model to obtain a target age feature of first age data, and obtaining a target age feature of second age data according to the target age feature of the first age data; and inputting the target age feature of the second age data, the second coding feature of the source speech and the to-be-synthesized target text into a speech synthesis module in a speech synthesis model to obtain a target speech of the target object under the second age data. According to the invention, through age feature decoupling and age feature stretching, the cost and complexity of data acquisition are effectively reduced, and meanwhile, high-precision synthesis of voice of a specific age is improved.

Description

Technical Field

[0001] The present invention relates to the field of speech synthesis technology, and in particular to a speech synthesis method, system, electronic equipment and storage medium. Background Art

[0002] In recent years, with the rapid progress of artificial intelligence technology, speech synthesis technology has gradually become one of the indispensable core components of human-computer interaction. Therefore, how to perform high-precision speech synthesis to improve the quality and naturalness of speech synthesis is an important topic that urgently needs to be studied.

[0003] In the existing technology, in order to achieve high-precision synthesis of the speech of a specific speaker at a specific age, it is often necessary to have a large-scale speech dataset covering all stages of the speaker's life cycle. However, such datasets have problems such as difficulty in obtaining and insufficient sample representativeness. Therefore, they lack the ability to accurately capture the sound characteristics of a specific age, making it difficult to achieve high-precision synthesis of speech of a specific age. Summary of the invention

[0004] The present invention provides a speech synthesis method, system, electronic device and storage medium, which are used to solve the defects of the prior art that the data set is insufficient and it is difficult to achieve high-precision synthesis of speech for a specific age, and to improve the high-precision synthesis of speech for a specific age.

[0005] The present invention provides a speech synthesis method, comprising: Acquire speech features of the source speech of the target object; Inputting the speech feature of the source speech into an encoder in a speech synthesis model to obtain a first coding feature and a second coding feature of the source speech; the first coding feature is a feature associated with age, and the second coding feature is a feature unrelated to age; Inputting the first encoded feature of the source speech into the age perception module in the speech synthesis model to obtain a target age feature of first age data, and obtaining a target age feature of second age data according to the target age feature of the first age data; the first age data is the age data corresponding to the source speech, and the second age data is the age data corresponding to the speech synthesis requirement of the target object; Inputting the target age feature of the second age data, the second encoding feature of the source speech, and the target text to be synthesized into a speech synthesis module in the speech synthesis model to obtain the target speech of the target object under the second age data; The speech synthesis model is trained based on the sample speech of the sample object, and the age label and sample text corresponding to the sample speech.

[0006] According to the present invention, there is provided a speech synthesis method, wherein the step of obtaining the target age feature of the second age data according to the target age feature of the first age data comprises: searching, in an age dictionary, a standard age feature corresponding to the first age data and a standard age feature corresponding to the second age data; calculating a deviation between a standard age feature corresponding to the first age data and a standard age feature corresponding to the second age data to obtain a stretching parameter; Performing feature conversion according to the stretching parameter and the target age feature of the first age data to obtain an age conversion feature corresponding to the source speech; A target age feature of the second age data is obtained according to the age conversion feature corresponding to the source voice.

[0007] According to the present invention, there is provided a speech synthesis method, wherein the step of obtaining a target age feature of the second age data according to an age conversion feature corresponding to the source speech comprises: When the number of the source voice is one, using the age conversion feature corresponding to the source voice as the target age feature of the second age data; When there are multiple source voices, an average value of age conversion features corresponding to the multiple source voices is calculated to obtain a target age feature of the second age data.

[0008] According to the present invention, a speech synthesis method is provided, wherein the step of constructing the age dictionary comprises: In a database, obtaining each of the sample voices and an age label corresponding to each of the sample voices; Inputting the speech feature of each of the sample speech into the encoder to obtain a first encoding feature of each of the sample speech; Inputting the first coding feature of each of the sample voices into the age perception module to obtain a target age feature of the age label corresponding to each of the sample voices; Calculating the average of the target age features corresponding to all sample voices under each age label to obtain the standard age features corresponding to each age label; Each age tag and the standard age feature corresponding to each age tag are mapped and stored to obtain the age dictionary.

[0009] According to the present invention, a speech synthesis method is provided, wherein the target age feature of the second age data, the second encoding feature of the source speech, and the target text to be synthesized are input into a speech synthesis module in the speech synthesis model to obtain the target speech of the target object under the second age data, including: Inputting the target age feature of the second age data and the second encoding feature of the source speech into a decoder in the speech synthesis module to obtain a decoding feature; Inputting the decoded features and the target text into the acoustic model in the speech synthesis module to obtain reconstructed acoustic features; The reconstructed acoustic features are input into a vocoder in the speech synthesis module to obtain the target speech.

[0010] According to the present invention, a speech synthesis method is provided, wherein the speech synthesis model is constructed based on the following steps: Inputting the speech feature of the sample speech into an initial encoder to obtain a first encoding feature and a second encoding feature of the sample speech; Inputting the first coding feature and the second coding feature of the sample speech into an initial decoder to obtain a sample decoding feature; Inputting the sample decoding feature and the sample text into an initial acoustic model to obtain a first sample reconstructed acoustic feature; Training the initial encoder, the initial decoder, and the initial acoustic model according to the age category corresponding to the first encoding feature of the sample speech, the age category corresponding to the second encoding feature of the sample speech, the age label, the original acoustic features of the sample speech, and the first sample reconstructed acoustic features; Constructing a speech model to be trained according to the trained initial encoder, the trained initial decoder, the trained initial acoustic model, and the initial age perception module; Training the speech model to be trained according to the sample speech, the original acoustic features and the sample text to obtain a trained speech model; The speech synthesis model is constructed according to the trained speech model.

[0011] According to the present invention, a speech synthesis method is provided, wherein the speech model to be trained is trained according to the sample speech, the original acoustic features and the sample text to obtain the trained speech model, comprising: Inputting the speech features of the sample speech and the sample text into the speech model to be trained to obtain a second sample reconstructed acoustic feature; According to the original acoustic features and the second sample reconstructed acoustic features, the to-be-trained age perception module, the to-be-trained decoder and the to-be-trained acoustic model in the to-be-trained speech model are iteratively trained to obtain the trained speech model.

[0012] The present invention also provides a speech synthesis system, characterized in that it comprises: A feature extraction unit, used to obtain speech features of the source speech of the target object; An encoding unit, configured to input the speech feature of the source speech into an encoder in a speech synthesis model to obtain a first encoding feature and a second encoding feature of the source speech; the first encoding feature is a feature associated with age, and the second encoding feature is a feature unrelated to age; a processing unit, configured to input the first encoding feature of the source speech into the age perception module in the speech synthesis model to obtain a target age feature of the first age data, and obtain a target age feature of the second age data according to the target age feature of the first age data; the first age data is the age data corresponding to the source speech, and the second age data is the age data corresponding to the speech synthesis requirement of the target object; a synthesis unit, configured to input the target age feature of the second age data, the second encoding feature of the source speech, and the target text to be synthesized into a speech synthesis module in the speech synthesis model to obtain the target speech of the target object under the second age data; The speech synthesis model is trained based on the sample speech of the sample object, and the age label and sample text corresponding to the sample speech.

[0013] The present invention also provides an electronic device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein when the processor executes the program, any of the above-mentioned speech synthesis methods is implemented.

[0014] The present invention also provides a non-transitory computer-readable storage medium on which a computer program is stored. When the computer program is executed by a processor, the speech synthesis method described in any one of the above is implemented.

[0015] The present invention also provides a computer program product, comprising a computer program, wherein when the computer program is executed by a processor, the computer program implements any of the above-mentioned speech synthesis methods.

[0016] The speech synthesis method, system, electronic device and storage medium provided by the present invention separate the age-related information and age-independent information of the target object from the inherent speech features of the source speech of the target object through an encoder, so as to flexibly control the age attribute and the speaker attribute, and extract and stretch the age-related information through an age perception module to obtain the target age feature of the age data corresponding to the speech synthesis demand of the target object, and synthesize the target speech of the age data corresponding to the speech synthesis demand of the target object accordingly. The whole process only requires a small amount of the source speech of the target object, and can realize the ability to accurately capture the sound features of the specific age of the target object, so as to carefully learn the age features and other features unrelated to the age of the target object, so that the synthesized speech can more realistically reflect the sound characteristics of the target object at different ages, thereby significantly improving the quality and naturalness of speech synthesis, thereby reducing the cost and complexity of data collection while improving the synthesis accuracy of speech of specific age, thereby providing users with highly customized, natural and fluent speech synthesis services with a strong sense of reality, and improving user experience. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] In order to more clearly illustrate the technical solutions in the present invention or the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0018] Figure 1 This is one of the flow charts of the speech synthesis method provided by the present invention.

[0019] Figure 2 This is the second flow chart of the speech synthesis method provided by the present invention.

[0020] Figure 3 It is a flowchart of the speech synthesis model training provided by the present invention.

[0021] Figure 4 It is a structural schematic diagram of the speech synthesis system provided by the present invention.

[0022] Figure 5 It is a structural schematic diagram of the electronic device provided by the present invention. DETAILED DESCRIPTION

[0023] In order to make the purpose, technical solution and advantages of the present invention clearer, the technical solution of the present invention will be clearly and completely described below in conjunction with the drawings of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0024] In recent years, with the rapid progress of artificial intelligence technology, speech synthesis technology has gradually become one of the indispensable core components of human-computer interaction. Its application scope covers many application scenarios from automatic telephone reply systems, intelligent navigation assistance to virtual personal assistants, and has shown broad application prospects in many fields such as education and entertainment. Traditional speech synthesis technology mainly focuses on converting input text into natural and smooth voice output. In this process, factors such as sound quality, speaking speed and intonation are the key to determining the user experience.

[0025] In order to achieve high-precision synthesis of the speech of a specific speaker at a specific age, in theory, a large-scale speech dataset covering all stages of the speaker's life cycle is needed. The construction of such a dataset faces many challenges in reality, such as the difficulty in obtaining data and insufficient sample representativeness. In addition, when simulating the sound characteristics of different age groups, existing speech synthesis technology can often only roughly divide them into coarse age groups such as youth, middle-aged, and elderly. There is a lack of large-scale speech datasets at various stages of the life cycle, and thus the ability to accurately capture the sound characteristics of specific ages. This has, to a certain extent, restricted the personalized performance and accuracy level of speech synthesis technology, making it difficult to achieve high-precision synthesis of speech of specific ages.

[0026] In this regard, the present application provides a speech synthesis method to perform speech synthesis for different ages through age decoupling and adaptation, thereby improving the high-precision synthesis of speech for a specific age.

[0027] It should be noted that the execution subject of the method may be a speech synthesis system, which may be a smart speaker, a speech synthesizer, etc., and this implementation does not specifically limit this. The system is like a sound time machine, which allows users to freely travel through their own sound history and experience a voice journey across the years.

[0028] Figure 1 It is one of the flowcharts of the speech synthesis method provided by the present invention; Figure 1 As shown, the method includes step 110 , step 120 , step 130 and step 140 .

[0029] Step 110: Acquire speech features of the source speech of the target object.

[0030] The target object here may be a speaker (or speaker) for whom speech synthesis of a specific age is required. The source speech here is a registered speech with an age data label input by a user when speech synthesis of the target object is required. The source speech is hereinafter referred to as the registered speech. The number of the registered speech may be one or more, which is not specifically limited in this embodiment.

[0031] The so-called user input may be information input through a command line interface, a graphical interface, a voice input, etc., and this embodiment does not specifically limit this.

[0032] Optionally, after the source speech of the target object is acquired, feature extraction may be performed on the source speech to acquire speech features of the source speech.

[0033] The feature extraction here can be to first extract acoustic features from the source speech to obtain the original acoustic features, and then further extract features from the original acoustic features to obtain speech features that can represent the speaker's speaking features. For example, first extract the Mel spectrum features from the original waveform of the source speech as acoustic features, and then input the Mel spectrum features into a deep learning network, such as an integrated deep neural network (IDNN), to further extract features from the original acoustic features, extract an embedded vector X-vector that can represent the speaker's information, which can also be recorded as X, and thus obtain speech features that can represent the speaker's speaking features.

[0034] Among them, the steps of feature extraction by IDNN include: after processing the extracted features through a series of deep neural network (DNN) layers, the network usually includes a special layer to aggregate time information to ensure that the length of the output feature vector is fixed. This aggregation layer can be a statistical pooling layer, which extracts statistical information such as mean and standard deviation from a series of frame-level features to generate a fixed-length vector, regardless of whether the input speech is long or short. The aggregated features are sent to a deeper network layer to eventually form a fixed-length embedding vector X-vector.

[0035] Step 120, inputting the speech features of the source speech into an encoder in a speech synthesis model to obtain a first coding feature and a second coding feature of the source speech; the first coding feature is a feature associated with age, and the second coding feature is a feature unrelated to age; wherein the speech synthesis model is trained based on sample speech of a sample object, and an age label and sample text corresponding to the sample speech.

[0036] Optionally, before executing step 120, a speech synthesis model may be first obtained through training. Specifically, an initialization model may be first constructed, and the initialization model here may include an initial encoder, an initial age perception module, and an initial speech synthesis module; the initial encoder may be a model prepared for age and timbre decoupling coding after parameter initialization, or a pre-trained model with age and timbre decoupling coding function; the initial age perception module may be a model prepared for age feature perception of a specific age after parameter initialization, or a pre-trained model with age feature perception function of a specific age; the initial speech synthesis module may be a model prepared for speech feature decoding and acoustic feature reconstruction after parameter initialization, or a pre-trained model with speech feature decoding and acoustic feature reconstruction function, and the embodiment of the present invention does not specifically limit this.

[0037] In addition, a large number of sample voices of sample speakers (also called sample subjects) at different ages and their age labels and sample texts may be collected to construct a data set; for example, sample voices may be collected and the age labels corresponding to the sample voices may be obtained through manual annotation, or age labels may be collected first and then the sample voices corresponding to the age labels may be obtained through manual recording or speech synthesis. The age labels of the sample voices provided by each sample speaker may be the same or different, and this embodiment does not specifically limit this. The sample text here is the content description text of the sample voice.

[0038] Subsequently, the initialization model can be iteratively trained based on the sample speech, the corresponding age label, and the sample text, so as to construct a speech synthesis model that can accurately synthesize speech of a specific age through training. The training can be an overall training of the initialization model based on the sample speech, the corresponding age label, and the sample text, or a multi-task training of freezing the initialization model in stages based on the sample speech, the corresponding age label, and the sample text, etc., which is not specifically limited in this embodiment.

[0039] In practical applications, the encoder in the trained speech synthesis model can be used to apply the speech features of the source speech to perform age decoupling, that is, decoupling to obtain a first coding feature associated with age and a second coding feature unrelated to age. The encoder here at least includes bottleneck information for age association (that is, the first coding feature) Extracted age feature encoder and bottleneck information for age-independent processing (i.e., the second encoding feature) Extracted age-independent feature encoder ; Among them, the age feature encoder Age-independent feature encoder The same model structure can be used.

[0040] Step 130: Input the first encoded feature of the source speech into the age perception module in the speech synthesis model to obtain the target age feature of the first age data, and obtain the target age feature of the second age data based on the target age feature of the first age data; the first age data is the age data corresponding to the source speech, and the second age data is the age data corresponding to the speech synthesis demand of the target object.

[0041] Optionally, after the first coding feature and the second coding feature are extracted by the encoder, the first coding feature The age-aware module in the speech synthesis model is input to apply the first encoding feature by the age-aware module The age attribute vector is precisely controlled and stretched to achieve vector extraction of the sound features of the first age data corresponding to the source voice, so as to obtain the first age data Target age characteristics .

[0042] The age perception module here may be a linear perception module, which obtains the corresponding target age feature by performing linearization processing on the first coding feature.

[0043] It should be noted that, since the generative flow model of reversible convolution (Glow model for short) has been proven in computer vision that after unsupervised training, the potential representation obtained by the model can stretch multiple attributes in the image, realize flexible control of the attributes in the image, and generate controllable and diverse images. Therefore, considering the powerful ability of the Glow model in attribute control, the age perception module in this embodiment can be constructed and implemented based on the Glow model, so as to be able to accurately control and stretch the age attribute vector, thereby realizing the vector extraction of sound features under different age data.

[0044] The Glow model consists of multiple flow modules; each single flow module consists of the following three parts: Step (a), Activation Normalization (Actnorm) layer. An affine transformation of the activations is performed using the scale and bias parameters of each channel, which are initialized so that the post-normalized activations of each channel have zero mean and unit variance given the initial data. After initialization, the scale and bias are treated as regular trainable parameters that are independent of the data.

[0045] Step (b), reversible 1 1 convolutional layer. Output of the age feature encoder Convolution is performed. Among them, is the inverse of the weight matrix.

[0046] ; in, It is the feature after the convolution layer processing; is the convolution kernel of the convolution layer.

[0047] Step (c), affine coupling layer. The affine coupling layer consists of the following three steps: Step (c1), zero initialization. Initializing the last convolutional layer of each nonlinear mapping neural network with zeros, and each affine coupling layer initially performs an identity function, helps to train very deep networks.

[0048] Step (c2), split and merge. Copy the channel dimension to form a new input, split and concatenate along the channel dimension.

[0049] Step (c3), permutation. Add reversible 1 before the first two steps. 1 convolution ensures that each dimension can affect other dimensions.

[0050] After obtaining the target age feature of the first age data through the age perception module, the age stretching can be performed according to the target age feature of the first age data, so as to stretch the target age feature of the first age data to form the second age data corresponding to the speech synthesis requirement of the target object. Target age characteristics .

[0051] Step 140: input the target age feature of the second age data, the second encoding feature of the source speech, and the target text to be synthesized into a speech synthesis module in the speech synthesis model to obtain the target speech of the target object under the second age data.

[0052] Optionally, after obtaining the target age feature of the second age data corresponding to the speech synthesis demand of the target object, the target age feature of the second age data can be fused with the second coding feature, so that the fusion result and the target text are input into the speech synthesis module in the speech synthesis model, so that the speech synthesis module applies the fusion feature of the target age feature and the second coding feature to simulate the target speech of the target object under the second age data, thereby realizing that only a small amount of source speech of the target object is needed to synthesize the target speech accurately for different age stages of the target object, so as to provide users with highly customized, natural, smooth and realistic speech synthesis services, meet the needs of different scenarios, and improve the user experience. The fusion here can be splicing, linear combination, etc., which is not specifically limited in this embodiment.

[0053] The method provided in this embodiment separates the age-related information and age-independent information of the target object from the inherent speech features of the source speech of the target object through an encoder, so as to flexibly control the age attribute and the speaker attribute, and extracts and stretches the age-related information through an age perception module to obtain the target age feature of the age data corresponding to the speech synthesis demand of the target object, and synthesizes the target speech of the age data corresponding to the speech synthesis demand of the target object accordingly. The whole process only requires a small amount of the source speech of the target object to achieve the ability to accurately capture the sound features of the specific age of the target object, and carefully learns the age features and other features that are independent of the age of the target object, so that the synthesized speech can more realistically reflect the sound characteristics of the target object at different ages, thereby significantly improving the quality and naturalness of speech synthesis, thereby reducing the cost and complexity of data collection while improving the synthesis accuracy of speech of a specific age, thereby providing users with a highly customized, natural, fluent and realistic speech synthesis service, and improving the user experience.

[0054] In some embodiments, step 130 specifically includes: searching, in an age dictionary, a standard age feature corresponding to the first age data and a standard age feature corresponding to the second age data; calculating a deviation between a standard age feature corresponding to the first age data and a standard age feature corresponding to the second age data to obtain a stretching parameter; Performing feature conversion according to the stretching parameter and the target age feature of the first age data to obtain an age conversion feature corresponding to the source speech; A target age feature of the second age data is obtained according to the age conversion feature corresponding to the source voice.

[0055] Figure 2 FIG. 2 is a flow chart of the speech synthesis method provided by the present invention; Figure 2 As shown, the step of obtaining the target age feature of the second age data by age stretching in step 130 specifically includes: According to the mapping relationship between different age data and standard age features in the age dictionary, the first age data is obtained by association Corresponding standard age characteristics and second age data Corresponding standard age characteristics Among them, the age dictionary uses age data as the key and the corresponding standard age feature as the value.

[0056] Then, calculate the standard age feature corresponding to the first age data Standard age characteristics corresponding to the second age data The deviation between them gives the stretching parameter , the specific calculation formula is as follows: ; Then, the stretch parameter and the target age characteristics of the first age data Perform linear stretching to obtain the age conversion features corresponding to the source speech.

[0057] After obtaining the age conversion feature corresponding to the source voice, it can be directly used as the target age feature of the second age data, or the age conversion feature corresponding to the source voice can be processed to obtain the target age feature of the second age data. The specific age feature can be adaptively determined according to the number of source voices.

[0058] Exemplarily, in some embodiments, the step of obtaining the target age feature of the second age data specifically includes: When the number of the source voice is one, using the age conversion feature corresponding to the source voice as the target age feature of the second age data; When there are multiple source voices, an average value of age conversion features corresponding to the multiple source voices is calculated to obtain a target age feature of the second age data.

[0059] Optionally, when the target object only provides one source voice, the stretching step is performed on the source voice to obtain the corresponding age conversion feature obtained by stretching the source voice, and then the corresponding age conversion feature obtained by stretching the source voice is used as the target age feature of the second age data.

[0060] When the target object provides two or more source voices, the above stretching step is iteratively performed on each source voice to obtain the corresponding age conversion features obtained by stretching each source voice, and then the age conversion features corresponding to all source voices are averaged as the target age features of the second age data, so as to more fully restore the age features of the second age data through multiple source voices, thereby further improving the synthesis accuracy, so as to provide users with highly customized, natural, fluent and realistic speech synthesis services to meet the needs of different scenarios.

[0061] The method provided in this embodiment can adaptively generate age features of a target age space from age features of a specific age space through age stretching and feature conversion, and perform target speech in the target age space based on the age features of the target age space obtained by stretching, so that the synthesized speech can more realistically reflect the sound characteristics of the target age space of the target object, and can achieve high customization, natural fluency and strong sense of reality.

[0062] In some embodiments, the step of constructing the age dictionary includes: In a database, obtaining each of the sample voices and an age label corresponding to each of the sample voices; Inputting the speech feature of each of the sample speech into the encoder to obtain a first encoding feature of each of the sample speech; Inputting the first coding feature of each of the sample voices into the age perception module to obtain a target age feature of the age label corresponding to each of the sample voices; Calculating the average of the target age features corresponding to all sample voices under each age label to obtain the standard age features corresponding to each age label; Each age tag and the standard age feature corresponding to each age tag are mapped and stored to obtain the age dictionary.

[0063] Optionally, the steps of constructing the age dictionary specifically include: The database data set The sample speech of multiple speakers corresponding to each age label in is used as a sub-dataset , , is the number of age labels.

[0064] For each sub-dataset Each sample speech in the batch is input into the speech synthesis model to obtain the sub-dataset The corresponding standard age features associated with the age label are implemented as follows: The original acoustic features (such as Mel-spectrogram features) of the sample speech are extracted, and the original acoustic features are input into the IDNN network to further extract the X-vector vector of the sample speech, that is, the speech features of the sample speech.

[0065] The speech feature of the sample speech is input into the age feature encoder to obtain the first encoding feature of the sample speech.

[0066] The first coding feature of the sample speech is input into the age perception module to linearize the first coding feature of the sample speech, and obtain the target age feature of the age label corresponding to the sample speech.

[0067] Save the subdataset The target age features of the age labels corresponding to all sample voices in the sub-dataset are obtained by taking the average value. The standard age feature corresponding to the age label is the key of this sub-dataset. The value corresponding to the corresponding age label.

[0068] Follow the above steps to loop through each sub-dataset , get each age label and its corresponding standard age feature, and save each age label as a key and its corresponding standard age feature as a value to form an age dictionary.

[0069] It should be noted that after the target speech under the second age data is obtained based on steps 110-140, the target speech and the corresponding second age data can also be associated and stored in the database to expand the speech segments of each speaker at different age stages, thereby alleviating the problem of sparse data of each speaker in different age dimensions, and assisting in improving the performance of various speech processing tasks that require model training or task processing of speech segments of each speaker at different age stages.

[0070] In addition, according to the second age data, the target speech can be expanded to the sub-dataset corresponding to the corresponding age label to update the age dictionary, thereby improving the accuracy of model reasoning synthesis.

[0071] In some embodiments, step 140 specifically includes: Inputting the target age feature of the second age data and the second encoding feature of the source speech into a decoder in the speech synthesis module to obtain a decoding feature; Inputting the decoded features and the target text into the acoustic model in the speech synthesis module to obtain reconstructed acoustic features; The reconstructed acoustic features are input into a vocoder in the speech synthesis module to obtain the target speech.

[0072] like Figure 2 As shown, the speech synthesis module includes a decoder, an acoustic model and a vocoder connected in sequence. The decoder is used for feature decoding; the acoustic model is used for acoustic feature reconstruction; and the vocoder is used for speech synthesis.

[0073] Wherein, step 140 specifically includes: The target age feature of the second age data and the second encoding feature of the source speech are fused and input into the decoder in the speech synthesis module, so that the decoder decodes and outputs a decoded feature, which contains all the information of the speech feature of the source speech. The calculation formula of the decoder is as follows: ; in, To decode features; is the model function of the decoder; is the target age feature of the second age data; is the second encoding feature of the source speech.

[0074] After obtaining the decoded features output by the decoder, the decoded features and the target text may be jointly input into the acoustic model in the speech synthesis module, so that the acoustic model, the decoded features and the target text are applied to reconstruct and generate corresponding reconstructed acoustic features. The target text here may be a synthesized speech description text for describing the speech content of the second age data to be synthesized, or may be a text containing tags or annotations with feature information related to the second age data, etc., which is not specifically limited in this embodiment.

[0075] After obtaining the reconstructed acoustic features output by the acoustic model, the reconstructed acoustic features can be input into a vocoder in a speech synthesis module, so that the vocoder applies the reconstructed acoustic features to perform speech synthesis to synthesize the target speech of the target object under the second age data.

[0076] The method provided in this embodiment utilizes the collaboration of a decoder, an acoustic model, and a vocoder to accurately fuse the target age characteristics of the second age data and the second coding characteristics of the source speech, so as to reconstruct and synthesize a target speech that is highly realistic and conforms to the age characteristics of the target object under the second age data.

[0077] In some embodiments, the speech synthesis model is constructed based on the following steps: Inputting the speech feature of the sample speech into an initial encoder to obtain a first encoding feature and a second encoding feature of the sample speech; Inputting the first coding feature and the second coding feature of the sample speech into an initial decoder to obtain a sample decoding feature; Inputting the sample decoding feature and the sample text into an initial acoustic model to obtain a first sample reconstructed acoustic feature; Training the initial encoder, the initial decoder, and the initial acoustic model according to the age category corresponding to the first encoding feature of the sample speech, the age category corresponding to the second encoding feature of the sample speech, the age label, the original acoustic features of the sample speech, and the first sample reconstructed acoustic features; Constructing a speech model to be trained according to the trained initial encoder, the trained initial decoder, the trained initial acoustic model, and the initial age perception module; Training the speech model to be trained according to the sample speech, the original acoustic features and the sample text to obtain a trained speech model; The speech synthesis model is constructed according to the trained speech model.

[0078] Figure 3 is a flow chart of the speech synthesis model training provided by the present invention; Figure 3As shown, the initialization model specifically includes an initial encoder, an initial age perception module, an initial decoder and an initial acoustic model.

[0079] The speech synthesis model is trained and constructed based on the initialization model. To ensure the stability of the model training, the training of the speech synthesis model can be divided into two stages, including the first stage training and the second stage training.

[0080] Optionally, the first stage of training specifically includes the following steps: First, a dataset is constructed, which contains a large number of sample voices of sample subjects at different ages, as well as the age labels corresponding to the sample voices.

[0081] Subsequently, acoustic feature extraction is performed on the sample speech to obtain the original acoustic features of the sample speech, and the embedding vector of the speaker information is extracted from the original acoustic features of the sample speech to obtain the speech features of the sample speech. The specific implementation steps can refer to the speech feature extraction steps of the source speech, which will not be repeated here.

[0082] Subsequently, the speech features of the sample speech are input into an initial encoder, which performs age decoupling of the speech features of the sample speech, so as to decouple age-related features and age-independent features from the speech features, thereby obtaining first coding features and second coding features of the sample speech.

[0083] The specific implementation steps of obtaining the coding features include: The speech features of the sample speech The first coding feature of the sample speech is input into the initial age feature encoder of the initial encoder to obtain the first coding feature of the sample speech output by the initial age feature encoder; the speech feature of the sample speech is input into the initial age-independent feature encoder of the initial encoder to obtain the second coding feature of the sample speech output by the initial age-independent feature encoder. The calculation formula is as follows: ; in, is the model function of the initial age feature encoder.

[0084] Second coding feature The calculation formula is as follows: ; in, is the model function of the initial age-independent feature encoder.

[0085] In addition, the first coding feature of the sample speech can be input into the age classifier, so that the age classifier can use the first coding feature of the sample speech to perform an age classification task and obtain an age category corresponding to the first coding feature of the sample speech. And the second coding feature of the sample speech can be input into the age classifier, so that the age classifier can use the second coding feature of the sample speech to perform an age classification task and obtain an age category corresponding to the second coding feature of the sample speech.

[0086] In addition, the first coding feature and the second coding feature of the sample speech can be combined and input into the initial decoder, so that the initial decoder applies the combined features of the first coding feature and the second coding feature to output all information containing the speech features of the sample speech, that is, the sample decoding feature. , the specific calculation formula is as follows: ; in, is the model function of the initial decoder.

[0087] Then, the sample decoding features and sample text input to the initial acoustic model, so that the initial acoustic model applies sample decoding features The acoustic features are reconstructed with the sample text to obtain the first sample reconstructed acoustic features.

[0088] Then, the first classification loss function is calculated based on the error between the age category corresponding to the first coding feature and the age label ; Gradient flipping is performed based on the error between the age category and the age label corresponding to the second encoding feature to obtain the second classification loss function ; According to the deviation between the original acoustic features of the sample speech and the reconstructed acoustic features of the first sample, the reconstruction loss function is obtained .

[0089] Then, the first classification loss function is integrated , the second classification loss function And the reconstruction loss function , calculate the total loss function of the first stage training The fusion here can be direct addition or weighted addition, which is not specifically limited in this embodiment. For example, the total loss function of the first stage training is The calculation formula can be: .

[0090] Then, according to the total loss function of the first stage training , the model parameters of the initial encoder, initial decoder and initial acoustic model in the initialization model are iteratively optimized to optimize the model performance, so that the trained initial encoder can effectively realize age decoupling to reduce the mutual information between age features and age-independent features, that is, the trained initial age feature encoder tends to retain age-related speaker features in the encoding features (that is, bottleneck information) and remove redundant features, and the trained initial age-independent feature encoder tends to retain age-independent speaker features in the encoding features (that is, bottleneck information), and the trained initial decoder and initial acoustic model can effectively reconstruct the acoustic features.

[0091] Among them, the mutual information calculation formula between age features and age-independent features can be expressed as follows: ; in, is the mutual information between the age feature M and the age-independent feature N.

[0092] Optionally, the second stage of training is fine-tuning training of the initialized model after the first stage of training as the base model, with the purpose of enabling the model to learn the linear representation of age characteristics to facilitate flexible age stretching. The training steps may include: based on sample speech, the original acoustic features of the sample speech and sample text, with the goal of the model being able to accurately learn the linear representation of age characteristics, fine-tuning training is performed on the trained initial encoder, the trained initial decoder and the trained initial acoustic model in the initialized model after the first stage of training, and the speech model to be trained formed by the initial age perception module to obtain the trained speech model, thereby constructing a speech synthesis model based on the trained speech model.

[0093] The fine-tuning training here may be fine-tuning training of the overall parameters of the speech model to be trained, or fine-tuning training of freezing part of the parameters of the speech model to be trained, which is not specifically limited in this embodiment.

[0094] Exemplarily, in some embodiments, the second stage training specifically includes: Inputting the speech features of the sample speech and the sample text into the speech model to be trained to obtain a second sample reconstructed acoustic feature; According to the original acoustic features and the second sample reconstructed acoustic features, the to-be-trained age perception module, the to-be-trained decoder and the to-be-trained acoustic model in the to-be-trained speech model are iteratively trained to obtain the trained speech model.

[0095] Optionally, in the second stage of training, the speech features of the sample speech can be sequentially passed through the speech model to be trained, so that the encoder to be trained (that is, the initial encoder after training) in the speech model to be trained can extract the first coding features and the second coding features of the sample speech, and the age perception module to be trained (that is, the initial age perception module) in the speech model to be trained can linearly map the first coding features of the sample speech to obtain the target age features corresponding to the age label of the sample speech, and the decoder to be trained (that is, the initial decoder after training) in the speech model to be trained can decode the sample speech using the first coding features and the target age features corresponding to the age label of the sample speech to obtain the sample decoding features of the sample speech, and the acoustic model to be trained (that is, the initial acoustic model after training) in the speech model to be trained can reconstruct the sample text using the sample decoding features of the sample speech to obtain the second sample reconstructed acoustic features.

[0096] Subsequently, according to the error between the reconstructed acoustic features of the second sample and the original acoustic features of the sample speech, the loss function of the second stage training is constructed, and on the basis of freezing the encoder to be trained, the loss function of the second stage training is used to iteratively train the age perception module to be trained, the decoder to be trained and the acoustic model to be trained in the speech model to be trained to obtain the trained speech model. Based on the trained speech model, a speech synthesis model that can accurately synthesize speech for a specific age is constructed. That is, by adding a vocoder on the basis of the trained speech model, a speech synthesis model that can accurately synthesize speech for a specific age can be constructed.

[0097] The method provided in this embodiment uses an initialization model to perform age classification and acoustic feature reconstruction on sample speech in the first stage of training, so as to construct a composite loss function including classification loss and reconstruction loss, and iteratively optimize the model parameters so that the encoder can effectively distinguish and encode age-related and irrelevant features respectively, while the decoder and acoustic model can accurately reconstruct acoustic features, thereby ensuring that the model can accurately capture the speaker's age characteristics and lay a solid foundation for subsequent age stretching and speech synthesis. In the second stage of training, the model trained in the first stage is used as the base, and the initial age perception module is introduced to construct the speech model to be trained. Through fine-tuning training, the model further learns the linear representation of age characteristics, which is convenient for achieving flexible age stretching effects, which not only improves the generalization ability of the model, but also ensures the accuracy and naturalness of the synthesized speech at a specific age stage. Therefore, through a sophisticated two-stage training strategy, a speech synthesis model that can accurately synthesize speech at a specific age stage can be effectively constructed.

[0098] The speech synthesis system provided by the present invention is described below. The speech synthesis system described below and the speech synthesis method described above can be referred to each other.

[0099] Figure 4 Schematic diagram of the structure of the speech synthesis system provided by the present invention; Figure 4 As shown, the system includes: The feature extraction unit 410 is used to obtain the speech features of the source speech of the target object; The encoding unit 420 is used to input the speech feature of the source speech into the encoder in the speech synthesis model to obtain a first encoding feature and a second encoding feature of the source speech; the first encoding feature is a feature associated with age, and the second encoding feature is a feature unrelated to age; The processing unit 430 is used to input the first encoding feature of the source speech into the age perception module in the speech synthesis model to obtain the target age feature of the first age data, and obtain the target age feature of the second age data according to the target age feature of the first age data; the first age data is the age data corresponding to the source speech, and the second age data is the age data corresponding to the speech synthesis demand of the target object; The synthesis unit 440 is used to input the target age feature of the second age data, the second encoding feature of the source speech and the target text to be synthesized into the speech synthesis module in the speech synthesis model to obtain the target speech of the target object under the second age data; The speech synthesis model is trained based on the sample speech of the sample object, and the age label and sample text corresponding to the sample speech.

[0100] The system provided in this embodiment separates the target object's age-related information and age-independent information from the inherent speech features of the target object's source speech through an encoder, so as to flexibly control the age attribute and the speaker attribute, and extracts and stretches the age-related information through an age perception module to obtain the target age feature of the age data corresponding to the speech synthesis demand of the target object, and synthesizes the target speech of the age data corresponding to the speech synthesis demand of the target object accordingly. The entire process only requires a small amount of the target object's source speech to achieve the ability to accurately capture the sound features of the target object's specific age, and to carefully learn the age features and other features that are independent of the target object's age, so that the synthesized speech can more realistically reflect the sound characteristics of the target object at different ages, thereby significantly improving the quality and naturalness of speech synthesis, thereby reducing the cost and complexity of data collection while improving the synthesis accuracy of speech of a specific age, thereby providing users with a highly customized, natural, fluent and realistic speech synthesis service, and improving the user experience.

[0101] In some embodiments, the processing unit is specifically configured to: searching, in an age dictionary, a standard age feature corresponding to the first age data and a standard age feature corresponding to the second age data; calculating a deviation between a standard age feature corresponding to the first age data and a standard age feature corresponding to the second age data to obtain a stretching parameter; Performing feature conversion according to the stretching parameter and the target age feature of the first age data to obtain an age conversion feature corresponding to the source speech; A target age feature of the second age data is obtained according to the age conversion feature corresponding to the source voice.

[0102] In some embodiments, the processing unit is further configured to: When the number of the source voice is one, using the age conversion feature corresponding to the source voice as the target age feature of the second age data; When there are multiple source voices, an average value of age conversion features corresponding to the multiple source voices is calculated to obtain a target age feature of the second age data.

[0103] In some embodiments, the system further includes a dictionary building unit, specifically configured to: In a database, obtaining each of the sample voices and an age label corresponding to each of the sample voices; Inputting the speech feature of each of the sample speech into the encoder to obtain a first encoding feature of each of the sample speech; Inputting the first coding feature of each of the sample voices into the age perception module to obtain a target age feature of the age label corresponding to each of the sample voices; Calculating the average of the target age features corresponding to all sample voices under each age label to obtain the standard age features corresponding to each age label; Each age tag and the standard age feature corresponding to each age tag are mapped and stored to obtain the age dictionary.

[0104] In some embodiments, the synthesis unit is specifically used for: Inputting the target age feature of the second age data and the second encoding feature of the source speech into a decoder in the speech synthesis module to obtain a decoding feature; Inputting the decoded features and the target text into the acoustic model in the speech synthesis module to obtain reconstructed acoustic features; The reconstructed acoustic features are input into a vocoder in the speech synthesis module to obtain the target speech.

[0105] In some embodiments, the system further includes a model training unit, specifically configured to: Inputting the speech feature of the sample speech into an initial encoder to obtain a first encoding feature and a second encoding feature of the sample speech; Inputting the first coding feature and the second coding feature of the sample speech into an initial decoder to obtain a sample decoding feature; Inputting the sample decoding feature and the sample text into an initial acoustic model to obtain a first sample reconstructed acoustic feature; Training the initial encoder, the initial decoder, and the initial acoustic model according to the age category corresponding to the first encoding feature of the sample speech, the age category corresponding to the second encoding feature of the sample speech, the age label, the original acoustic features of the sample speech, and the first sample reconstructed acoustic features; Constructing a speech model to be trained according to the trained initial encoder, the trained initial decoder, the trained initial acoustic model, and the initial age perception module; Training the speech model to be trained according to the sample speech, the original acoustic features and the sample text to obtain a trained speech model; The speech synthesis model is constructed according to the trained speech model.

[0106] In some embodiments, the model training unit is further used to: Inputting the speech features of the sample speech and the sample text into the speech model to be trained to obtain a second sample reconstructed acoustic feature; According to the original acoustic features and the second sample reconstructed acoustic features, the to-be-trained age perception module, the to-be-trained decoder and the to-be-trained acoustic model in the to-be-trained speech model are iteratively trained to obtain the trained speech model.

[0107] The system provided by the present invention is used to execute the above-mentioned method embodiments. Please refer to the above-mentioned embodiments for the specific processes and detailed contents, which will not be repeated here.

[0108] Figure 5 An example of a physical structure diagram of an electronic device is shown in FIG. Figure 5As shown, the electronic device may include: a processor (processor) 510, a communication interface (Communications Interface) 520, a memory (memory) 530 and a communication bus 540, wherein the processor 510, the communication interface 520, and the memory 530 communicate with each other through the communication bus 540. The processor 510 can call the logic instructions in the memory 530 to execute the speech synthesis method, which includes: obtaining the speech features of the source speech of the target object; inputting the speech features of the source speech into the encoder in the speech synthesis model to obtain the first encoding feature and the second encoding feature of the source speech; the first encoding feature is a feature associated with age, and the second encoding feature is a feature unrelated to age; inputting the first encoding feature of the source speech into the age perception module in the speech synthesis model to obtain the target age feature of the first age data, and obtaining the target age feature of the second age data according to the target age feature of the first age data; the first age data is the age data corresponding to the source speech, and the second age data is the age data corresponding to the speech synthesis demand of the target object; inputting the target age feature of the second age data, the second encoding feature of the source speech and the target text to be synthesized into the speech synthesis module in the speech synthesis model to obtain the target speech of the target object under the second age data; wherein the speech synthesis model is trained based on the sample speech of the sample object, and the age label and sample text corresponding to the sample speech.

[0109] In addition, the logic instructions in the above-mentioned memory 530 can be implemented in the form of a software functional unit and can be stored in a computer-readable storage medium when it is sold or used as an independent product. Based on such an understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art or the part of the technical solution, can be embodied in the form of a software product, and the computer software product is stored in a storage medium, including a number of instructions for a computer device (which can be a personal computer, a server, or a network device, etc.) to perform all or part of the steps of the method described in each embodiment of the present invention. The aforementioned storage medium includes: U disk, mobile hard disk, read-only memory (ROM, Read-Only Memory), random access memory (RAM, Random Access Memory), disk or optical disk, etc. Various media that can store program codes.

[0110] On the other hand, the present invention also provides a computer program product, which includes a computer program, which can be stored on a non-transitory computer-readable storage medium. When the computer program is executed by a processor, the computer can execute the speech synthesis method provided by the above methods, which includes: obtaining speech features of a source speech of a target object; inputting the speech features of the source speech into an encoder in a speech synthesis model to obtain a first encoding feature and a second encoding feature of the source speech; the first encoding feature is a feature associated with age, and the second encoding feature is a feature unrelated to age; inputting the first encoding feature of the source speech into an age perception module in the speech synthesis model to obtain a target age feature of first age data, and obtaining a target age feature of second age data based on the target age feature of the first age data; the first age data is age data corresponding to the source speech, and the second age data is age data corresponding to the speech synthesis requirement of the target object; inputting the target age feature of the second age data, the second encoding feature of the source speech, and the target text to be synthesized into the speech synthesis module in the speech synthesis model to obtain the target speech of the target object under the second age data; wherein the speech synthesis model is trained based on sample speech of a sample object, and the age label and sample text corresponding to the sample speech.

[0111] On the other hand, the present invention also provides a non-transitory computer-readable storage medium having a computer program stored thereon, which is implemented by a processor to execute the speech synthesis method provided by the above methods when executed, the method comprising: obtaining speech features of a source speech of a target object; inputting the speech features of the source speech into an encoder in a speech synthesis model to obtain a first encoding feature and a second encoding feature of the source speech; the first encoding feature is a feature associated with age, and the second encoding feature is a feature unrelated to age; inputting the first encoding feature of the source speech into an age perception module in the speech synthesis model to obtain a target age feature of first age data, and obtaining a target age feature of second age data based on the target age feature of the first age data; the first age data is age data corresponding to the source speech, and the second age data is age data corresponding to the speech synthesis requirement of the target object; inputting the target age feature of the second age data, the second encoding feature of the source speech, and the target text to be synthesized into a speech synthesis module in the speech synthesis model to obtain a target speech of the target object under the second age data; wherein the speech synthesis model is trained based on sample speech of a sample object, and the age label and sample text corresponding to the sample speech.

[0112] The device embodiments described above are merely illustrative, wherein the units described as separate components may or may not be physically separated, and the components displayed as units may or may not be physical units, that is, they may be located in one place, or they may be distributed on multiple network units. Some or all of the modules may be selected according to actual needs to achieve the purpose of the scheme of this embodiment. Ordinary technicians in this field can understand and implement it without paying creative labor.

[0113] Through the description of the above implementation methods, those skilled in the art can clearly understand that each implementation method can be implemented by means of software plus a necessary general hardware platform, and of course, can also be implemented by hardware. Based on this understanding, the above technical solution is essentially or the part that contributes to the prior art can be embodied in the form of a software product, and the computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, a disk, an optical disk, etc., including a number of instructions for a computer device (which can be a personal computer, a server, or a network device, etc.) to execute the methods described in each embodiment or some parts of the embodiments.

[0114] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some of the technical features therein. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A speech synthesis method, characterized in that: include: Acquire speech features of the source speech of the target object; Inputting the speech feature of the source speech into an encoder in a speech synthesis model to obtain a first coding feature and a second coding feature of the source speech; the first coding feature is a feature associated with age, and the second coding feature is a feature unrelated to age; Inputting the first encoded feature of the source speech into the age perception module in the speech synthesis model to obtain the target age feature of the first age data, and obtaining the target age feature of the second age data according to the target age feature of the first age data; The first age data is age data corresponding to the source voice, and the second age data is age data corresponding to the speech synthesis requirement of the target object; Inputting the target age feature of the second age data, the second encoding feature of the source speech, and the target text to be synthesized into a speech synthesis module in the speech synthesis model to obtain the target speech of the target object under the second age data; The speech synthesis model is trained based on the sample speech of the sample object, and the age label and sample text corresponding to the sample speech.

2. The speech synthesis method according to claim 1, characterized in that: The acquiring the target age feature of the second age data according to the target age feature of the first age data includes: searching, in an age dictionary, a standard age feature corresponding to the first age data and a standard age feature corresponding to the second age data; calculating a deviation between a standard age feature corresponding to the first age data and a standard age feature corresponding to the second age data to obtain a stretching parameter; Performing feature conversion according to the stretching parameter and the target age feature of the first age data to obtain an age conversion feature corresponding to the source speech; A target age feature of the second age data is obtained according to the age conversion feature corresponding to the source voice.

3. The speech synthesis method according to claim 2, characterized in that: The acquiring, according to the age conversion feature corresponding to the source voice, a target age feature of the second age data includes: When the number of the source voice is one, using the age conversion feature corresponding to the source voice as the target age feature of the second age data; When there are multiple source voices, an average value of age conversion features corresponding to the multiple source voices is calculated to obtain a target age feature of the second age data.

4. The speech synthesis method according to claim 2, characterized in that: The steps of constructing the age dictionary include: In a database, obtaining each of the sample voices and an age label corresponding to each of the sample voices; Inputting the speech feature of each of the sample speech into the encoder to obtain a first encoding feature of each of the sample speech; Inputting the first coding feature of each of the sample voices into the age perception module to obtain a target age feature of the age label corresponding to each of the sample voices; Calculating the average of the target age features corresponding to all sample voices under each age label to obtain the standard age features corresponding to each age label; Each age tag and the standard age feature corresponding to each age tag are mapped and stored to obtain the age dictionary.

5. The speech synthesis method according to any one of claims 1 to 4, characterized in that: The step of inputting the target age feature of the second age data, the second encoding feature of the source speech, and the target text to be synthesized into a speech synthesis module in the speech synthesis model to obtain the target speech of the target object under the second age data includes: Inputting the target age feature of the second age data and the second encoding feature of the source speech into a decoder in the speech synthesis module to obtain a decoding feature; Inputting the decoded features and the target text into the acoustic model in the speech synthesis module to obtain reconstructed acoustic features; The reconstructed acoustic features are input into a vocoder in the speech synthesis module to obtain the target speech.

6. The speech synthesis method according to any one of claims 1 to 4, characterized in that: The speech synthesis model is constructed based on the following steps: Inputting the speech feature of the sample speech into an initial encoder to obtain a first encoding feature and a second encoding feature of the sample speech; Inputting the first coding feature and the second coding feature of the sample speech into an initial decoder to obtain a sample decoding feature; Inputting the sample decoding feature and the sample text into an initial acoustic model to obtain a first sample reconstructed acoustic feature; Training the initial encoder, the initial decoder, and the initial acoustic model according to the age category corresponding to the first encoding feature of the sample speech, the age category corresponding to the second encoding feature of the sample speech, the age label, the original acoustic features of the sample speech, and the first sample reconstructed acoustic features; Constructing a speech model to be trained according to the trained initial encoder, the trained initial decoder, the trained initial acoustic model, and the initial age perception module; Training the speech model to be trained according to the sample speech, the original acoustic features and the sample text to obtain a trained speech model; The speech synthesis model is constructed according to the trained speech model.

7. The speech synthesis method according to claim 6, characterized in that: The step of training the speech model to be trained according to the sample speech, the original acoustic features and the sample text to obtain a trained speech model comprises: Inputting the speech features of the sample speech and the sample text into the speech model to be trained to obtain a second sample reconstructed acoustic feature; According to the original acoustic features and the second sample reconstructed acoustic features, the to-be-trained age perception module, the to-be-trained decoder and the to-be-trained acoustic model in the to-be-trained speech model are iteratively trained to obtain the trained speech model.

8. A speech synthesis system, characterized in that: include: A feature extraction unit, used to obtain speech features of the source speech of the target object; an encoding unit, configured to input the speech feature of the source speech into an encoder in a speech synthesis model to obtain a first encoding feature and a second encoding feature of the source speech; the first encoding feature is a feature associated with age, and the second encoding feature is a feature unrelated to age; a processing unit, configured to input the first encoding feature of the source speech into an age perception module in the speech synthesis model to obtain a target age feature of the first age data, and obtain a target age feature of the second age data according to the target age feature of the first age data; The first age data is age data corresponding to the source voice, and the second age data is age data corresponding to the speech synthesis requirement of the target object; a synthesis unit, configured to input the target age feature of the second age data, the second encoding feature of the source speech, and the target text to be synthesized into a speech synthesis module in the speech synthesis model to obtain the target speech of the target object under the second age data; The speech synthesis model is trained based on the sample speech of the sample object, and the age label and sample text corresponding to the sample speech.

9. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein: When the processor executes the program, the speech synthesis method according to any one of claims 1 to 7 is implemented.

10. A non-transitory computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the speech synthesis method according to any one of claims 1 to 7 is implemented.

Images