Artificial intelligence-based speech synthesis method and device, computer equipment and medium

Abstract

The application belongs to the field of artificial intelligence and the field of financial technology, and relates to a voice synthesis method based on artificial intelligence, which comprises the following steps: obtaining long sentence text; obtaining historical context sentences corresponding to the long sentence text; obtaining voice information of a historical moment corresponding to the long sentence text; performing style prediction on the long sentence text based on a style predictor and the historical context sentences to obtain first style embedding; performing style extraction on the voice information based on a style extractor to obtain second style embedding; generating target style embedding based on the first style embedding and the second style embedding; and performing voice synthesis on the long sentence text based on the target style embedding and a voice synthesis model to generate target voice. The application also provides a voice synthesis device based on artificial intelligence, a computer device and a storage medium. The application uses the target style embedding and the voice synthesis model to adjust the voice synthesis process of the long sentence text, and realizes the target voice which is coherent in rhythm and connection, good in hearing sense and high in expressiveness.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of artificial intelligence development and the field of financial technology, and in particular relates to a voice synthesis method and device based on artificial intelligence, a computer device and a storage medium. BACKGROUND

[0002] Voice synthesis (Text-to-Speech, TTS) technology is a technology that converts text into natural speech. It can make computers generate realistic voices and provide users with high-quality audio reading experiences. Voice synthesis technology is also widely used in the financial insurance field. For example, in the business scenario of promoting insurance products in the financial insurance field, insurance product promotion texts or insurance business popularization texts with long sentences often need to be processed by voice synthesis to promote audio books of related insurance texts.

[0003] With the continuous progress of neural network technology, especially the end-to-end neural network architecture, the quality and coherence of voice synthesis have been significantly improved. However, the existing voice synthesis method is usually for single sentences with relatively short texts. When encountering a long sentence voice synthesis task, the long sentence can only be split into single sentences for voice synthesis, and the synthesized voices are spliced. Such a voice synthesis method will result in insufficient coherence of the prosody of the two sentences, poor listening experience, and cannot guarantee the high expressiveness of the synthesized voice. SUMMARY

[0004] The purpose of the embodiments of the present application is to propose a voice synthesis method and device based on artificial intelligence, a computer device and a storage medium to solve the technical problem that the existing voice synthesis method is usually for single sentences with relatively short texts. When encountering a long sentence voice synthesis task, the long sentence can only be split into single sentences for voice synthesis, and the synthesized voices are spliced. Such a voice synthesis method will result in insufficient coherence of the prosody of the two sentences, poor listening experience, and cannot guarantee the high expressiveness of the synthesized voice.

[0005] To solve the above technical problems, the embodiments of the present application provide a voice synthesis method based on artificial intelligence, which adopts the following technical solutions:

[0006] Obtain a long sentence text to be processed;

[0007] Obtain a historical context sentence corresponding to the long sentence text;

[0008] Obtain voice information of a historical moment corresponding to the long sentence text;

[0009] Perform style prediction processing on the long sentence text based on a preset style predictor and the historical context sentence to obtain a corresponding first style embedding;

[0010] The speech information is processed by a preset style extractor to obtain the corresponding second style embedding.

[0011] A target style embedding is generated based on the first style embedding and the second style embedding;

[0012] The long sentence text is processed by speech synthesis based on the target style embedding and the preset speech synthesis model to generate the target speech corresponding to the long sentence text; wherein, the speech synthesis model includes a phoneme extractor, a first position encoder, a phoneme encoder, a multi-class predictor, a second position encoder, a decoder, and a vocoder.

[0013] Furthermore, the step of performing speech synthesis processing on the long sentence text based on the target style embedding and a preset speech synthesis model to generate target speech corresponding to the long sentence text specifically includes:

[0014] The long sentence text is input into the speech synthesis model, and the long sentence text is extracted and processed by the phoneme extractor to obtain a phoneme sequence. The phoneme sequence is then processed to generate a phoneme embedding sequence.

[0015] Based on the first position encoder and the phoneme encoder, feature extraction processing is performed on the phoneme embedding sequence to obtain the corresponding phoneme features;

[0016] The phoneme features are fused with the target style embedding to obtain the corresponding embedding feature vector;

[0017] Based on the multi-class predictor, the embedded feature vector is predicted to obtain the corresponding first output result;

[0018] The first output result is processed based on the second position encoder to obtain the corresponding second output result;

[0019] Based on the decoder, the second output result is decoded to generate the corresponding specified Mel spectrogram;

[0020] The specified Mel spectrogram is converted using the vocoder to obtain the corresponding specified long speech, and the specified long speech is used as the target speech.

[0021] The further step of performing style prediction processing on the long sentence text based on a preset style predictor and the historical context statements to obtain the corresponding first style embedding specifically includes:

[0022] Data extraction is performed on the long sentence text based on a pre-trained language model to obtain the corresponding word-level text embedding;

[0023] The style predictor generates a predicted style embedding sequence corresponding to the word-level text embedding;

[0024] Obtain the first style embedding sequence corresponding to the historical context statement;

[0025] The predicted style embedding sequence is fused with the first style embedding sequence to obtain the corresponding second style embedding sequence;

[0026] The first style embedding is generated based on the second style embedding sequence.

[0027] Furthermore, the step of performing style extraction processing on the speech information based on a preset style extractor to obtain the corresponding second style embedding specifically includes:

[0028] Obtain the Mel spectrogram corresponding to the speech information;

[0029] Invoke the style extractor;

[0030] The Mel spectrogram is input into the style extractor, and the style extractor performs style extraction processing on the Mel spectrogram to obtain the corresponding specified style embedding.

[0031] The specified style embedding is used as the second style embedding.

[0032] Furthermore, the step of generating the target style embedding based on the first style embedding and the second style embedding specifically includes:

[0033] Obtain the preset splicing strategy;

[0034] Based on the splicing strategy, the first style embedding and the second style embedding are spliced ​​together to obtain the corresponding spliced ​​style embedding;

[0035] The splicing style is embedded as the target style.

[0036] Furthermore, after the step of performing speech synthesis processing on the long sentence text based on the target style embedding and a preset speech synthesis model to generate target speech corresponding to the long sentence text, the method further includes:

[0037] The target speech is optimized to obtain the corresponding optimized speech;

[0038] Determine the target storage method corresponding to the optimized speech;

[0039] The optimized speech is stored and processed based on the target storage method.

[0040] Furthermore, the step of optimizing the target speech to obtain the corresponding optimized speech specifically includes:

[0041] The target speech is denoised to obtain the corresponding first target speech;

[0042] The first target speech is subjected to volume normalization processing to obtain the corresponding second target speech;

[0043] The second target speech is used as the optimized speech.

[0044] To address the aforementioned technical problems, this application also provides an artificial intelligence-based speech synthesis device, employing the following technical solution:

[0045] The first acquisition module is used to acquire the long sentence text to be processed;

[0046] The second acquisition module is used to acquire the historical context statements corresponding to the long sentence text;

[0047] The third acquisition module is used to acquire the voice information of the historical moment corresponding to the long sentence text;

[0048] The prediction module is used to perform style prediction processing on the long sentence text based on a preset style predictor and the historical context statements to obtain the corresponding first style embedding.

[0049] The extraction module is used to perform style extraction processing on the speech information based on a preset style extractor to obtain the corresponding second style embedding;

[0050] The generation module is used to generate a target style embedding based on the first style embedding and the second style embedding;

[0051] The processing module is used to perform speech synthesis processing on the long sentence text based on the target style embedding and a preset speech synthesis model to generate target speech corresponding to the long sentence text; wherein, the speech synthesis model includes a phoneme extractor, a first position encoder, a phoneme encoder, a multi-class predictor, a second position encoder, a decoder, and a vocoder.

[0052] To address the aforementioned technical problems, this application also provides a computer device that employs the following technical solution:

[0053] Obtain the long sentence text to be processed;

[0054] Obtain the historical context statements corresponding to the long sentence text;

[0055] Obtain the audio information of the historical moment corresponding to the long sentence text;

[0056] Based on a preset style predictor and the historical context statements, the long sentence text is processed for style prediction to obtain the corresponding first style embedding.

[0057] The speech information is processed by a preset style extractor to obtain the corresponding second style embedding.

[0058] A target style embedding is generated based on the first style embedding and the second style embedding;

[0059] The long sentence text is processed by speech synthesis based on the target style embedding and the preset speech synthesis model to generate the target speech corresponding to the long sentence text; wherein, the speech synthesis model includes a phoneme extractor, a first position encoder, a phoneme encoder, a multi-class predictor, a second position encoder, a decoder, and a vocoder.

[0060] To address the aforementioned technical problems, this application also provides a computer-readable storage medium, employing the technical solution described below:

[0061] Obtain the long sentence text to be processed;

[0062] Obtain the historical context statements corresponding to the long sentence text;

[0063] Obtain the audio information of the historical moment corresponding to the long sentence text;

[0064] Based on a preset style predictor and the historical context statements, the long sentence text is processed for style prediction to obtain the corresponding first style embedding.

[0065] The speech information is processed by a preset style extractor to obtain the corresponding second style embedding.

[0066] A target style embedding is generated based on the first style embedding and the second style embedding;

[0067] The long sentence text is processed by speech synthesis based on the target style embedding and the preset speech synthesis model to generate the target speech corresponding to the long sentence text; wherein, the speech synthesis model includes a phoneme extractor, a first position encoder, a phoneme encoder, a multi-class predictor, a second position encoder, a decoder, and a vocoder.

[0068] Compared with the prior art, the embodiments of this application have the following main advantages:

[0069] This application first obtains a long sentence text to be processed; then obtains the historical context statements corresponding to the long sentence text; and obtains the speech information of the historical moment corresponding to the long sentence text; then performs style prediction processing on the long sentence text based on a preset style predictor and the historical context statements to obtain a corresponding first style embedding; subsequently, performs style extraction processing on the speech information based on a preset style extractor to obtain a corresponding second style embedding; further, generates a target style embedding based on the first style embedding and the second style embedding; finally, performs speech synthesis processing on the long sentence text based on the target style embedding and a preset speech synthesis model to generate the target speech corresponding to the long sentence text. This application, after obtaining the long sentence text to be processed, the historical context statements corresponding to the long sentence text, and the speech information of the historical moment corresponding to the long sentence text, integrates the target style embedding generated by combining the first style embedding corresponding to the long sentence text predicted by the style predictor based on the historical context statements and the second style embedding obtained by the style extractor through style extraction processing of the speech information of the historical moment corresponding to the long sentence text into the speech synthesis model. The target style embedding is then used to adjust the speech synthesis process corresponding to the long sentence text, so as to generate high-quality target speech with coherent prosody and good listening experience that conforms to the target style embedding. This is beneficial to improving the listening quality of the target speech that matches the long sentence text, ensuring the high expressiveness of the generated target speech, and thus helping to improve the user experience of the listener. Attached Figure Description

[0070] To more clearly illustrate the solutions in this application, the accompanying drawings used in the description of the embodiments of this application will be briefly introduced below. Obviously, the accompanying drawings described below are some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0071] Figure 1 This is an exemplary system architecture diagram to which this application can be applied;

[0072] Figure 2 A flowchart of an embodiment of the artificial intelligence-based speech synthesis method according to this application;

[0073] Figure 3 This is a schematic diagram of the structure of an embodiment of the artificial intelligence-based speech synthesis device according to this application;

[0074] Figure 4 This is a schematic diagram of the structure of one embodiment of the computer device according to this application. Detailed Implementation

[0075] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application pertains; the terminology used herein in the specification of the application is for the purpose of describing particular embodiments only and is not intended to be limiting of the application; the terms "comprising" and "having," and any variations thereof, in the specification, claims, and foregoing drawings of this application, are intended to cover non-exclusive inclusion. The terms "first," "second," etc., in the specification, claims, or foregoing drawings of this application are used to distinguish different objects, not to describe a particular order.

[0076] In this document, the term "embodiment" means that a particular feature, structure, or characteristic described in connection with an embodiment may be included in at least one embodiment of this application. The appearance of this phrase in various places throughout the specification does not necessarily refer to the same embodiment, nor is it a separate or alternative embodiment mutually exclusive with other embodiments. It will be explicitly and implicitly understood by those skilled in the art that the embodiments described herein can be combined with other embodiments.

[0077] To enable those skilled in the art to better understand the present application, the technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the accompanying drawings.

[0078] like Figure 1 As shown, system architecture 100 may include terminal devices 101, 102, and 103, a network 104, and a server 105. Network 104 serves as the medium for providing communication links between terminal devices 101, 102, and 103 and server 105. Network 104 may include various connection types, such as wired or wireless communication links, or fiber optic cables, etc.

[0079] Users can use terminal devices 101, 102, and 103 to interact with server 105 via network 104 to receive or send messages, etc. Various communication client applications can be installed on terminal devices 101, 102, and 103, such as web browser applications, shopping applications, search applications, instant messaging tools, email clients, social media platform software, etc.

[0080] Terminal devices 101, 102, and 103 can be various electronic devices with displays and support web browsing, including but not limited to smartphones, tablets, e-book readers, MP3 players (Moving Picture Experts Group Audio Layer III), MP4 players (Moving Picture Experts Group Audio Layer IV), laptops, and desktop computers, etc.

[0081] Server 105 can be a server that provides various services, such as a backend server that supports the pages displayed on terminal devices 101, 102, and 103.

[0082] It should be noted that the AI-based speech synthesis method provided in this application is generally executed by a server / terminal device, and correspondingly, the AI-based speech synthesis device is generally located in the server / terminal device.

[0083] It should be understood that Figure 1 The number of terminal devices, networks, and servers shown is merely illustrative. Depending on implementation needs, any number of terminal devices, networks, and servers can be included.

[0084] Continue to refer to Figure 2 This document illustrates a flowchart of an embodiment of the AI-based speech synthesis method according to this application. The order of steps in the flowchart can be changed, and some steps can be omitted, depending on different requirements. The AI-based speech synthesis method provided in this application can be applied to any scenario requiring speech synthesis of long sentences, and thus can be applied to products in these scenarios, such as speech synthesis in the financial insurance field. The AI-based speech synthesis method includes the following steps:

[0085] Step S201: Obtain the long sentence text to be processed.

[0086] In this embodiment, the artificial intelligence-based speech synthesis method runs on an electronic device (e.g., Figure 1The server / terminal device shown can acquire long text sentences to be processed via wired or wireless connections. It should be noted that the aforementioned wireless connection methods may include, but are not limited to, 3G / 4G / 5G connections, Wi-Fi connections, Bluetooth connections, Wi-Fi connections, Zigbee connections, UWB (ultra-Wi-Fi) connections, and other currently known or future-developed wireless connection methods. In the business scenario of insurance product promotion in the financial insurance industry, the aforementioned long text sentences can be insurance product promotion texts or insurance business popularization texts that require speech synthesis and contain long sentence segments.

[0087] Step S202: Obtain the historical context statements corresponding to the long sentence text.

[0088] In this embodiment, the aforementioned historical context statements can specifically be context statements associated with the long sentence text at a preset historical target time. The selection of this historical target time is not specifically limited and can be chosen according to actual business needs.

[0089] Step S203: Obtain the audio information of the historical moment corresponding to the long sentence text.

[0090] In this embodiment, the context text corresponding to the historical moment of the long sentence text can be obtained first, and then the speech information that has undergone speech synthesis processing corresponding to the context text can be obtained. The selection of the aforementioned historical moment is not specifically limited and can be chosen according to actual business needs.

[0091] Step S204: Perform style prediction processing on the long sentence text based on the preset style predictor and the historical context statement to obtain the corresponding first style embedding.

[0092] In this embodiment, the specific implementation process of performing style prediction processing on the long sentence text based on the preset style predictor and the historical context statement to obtain the corresponding first style embedding will be further described in detail in subsequent specific embodiments of this application, and will not be elaborated on here.

[0093] Step S205: Perform style extraction processing on the speech information based on a preset style extractor to obtain the corresponding second style embedding.

[0094] In this embodiment, the specific implementation process of performing style extraction processing on the speech information based on the preset style extractor to obtain the corresponding second style embedding will be further described in detail in subsequent specific embodiments of this application, and will not be elaborated on here.

[0095] Step S206: Generate a target style embedding based on the first style embedding and the second style embedding.

[0096] In this embodiment, the specific implementation process of generating the target style embedding based on the first style embedding and the second style embedding will be further described in detail in subsequent specific embodiments of this application, and will not be elaborated on here.

[0097] Step S207: Based on the target style embedding and the preset speech synthesis model, the long sentence text is processed by speech synthesis to generate target speech corresponding to the long sentence text; wherein, the speech synthesis model includes a phoneme extractor, a first position encoder, a phoneme encoder, a multi-class predictor, a second position encoder, a decoder, and a vocoder.

[0098] In this embodiment, the specific implementation process of performing speech synthesis processing on the long sentence text based on the target style embedding and the preset speech synthesis model to generate the target speech corresponding to the long sentence text will be further described in detail in subsequent specific embodiments of this application, and will not be elaborated on here.

[0099] This application first obtains a long sentence text to be processed; then obtains the historical context statements corresponding to the long sentence text; and obtains the speech information of the historical moment corresponding to the long sentence text; then performs style prediction processing on the long sentence text based on a preset style predictor and the historical context statements to obtain a corresponding first style embedding; subsequently, performs style extraction processing on the speech information based on a preset style extractor to obtain a corresponding second style embedding; further, generates a target style embedding based on the first style embedding and the second style embedding; finally, performs speech synthesis processing on the long sentence text based on the target style embedding and a preset speech synthesis model to generate the target speech corresponding to the long sentence text. This application, after obtaining the long sentence text to be processed, the historical context statements corresponding to the long sentence text, and the speech information of the historical moment corresponding to the long sentence text, integrates the target style embedding generated by combining the first style embedding corresponding to the long sentence text predicted by the style predictor based on the historical context statements and the second style embedding obtained by the style extractor through style extraction processing of the speech information of the historical moment corresponding to the long sentence text into the speech synthesis model. The target style embedding is then used to adjust the speech synthesis process corresponding to the long sentence text, so as to generate high-quality target speech with coherent prosody and good listening experience that conforms to the target style embedding. This is beneficial to improving the listening quality of the target speech that matches the long sentence text, ensuring the high expressiveness of the generated target speech, and thus helping to improve the user experience of the listener.

[0100] In some alternative implementations, step S207 includes the following steps:

[0101] The long sentence text is input into the speech synthesis model, and the long sentence text is extracted and processed by the phoneme extractor to obtain a phoneme sequence. The phoneme sequence is then processed to generate a phoneme embedding sequence.

[0102] In this embodiment, the main architecture of the speech synthesis model can follow the structure of the FastSpeech2 speech synthesis model, which is a speech synthesis model that adopts the Transformer structure. After the long sentence text is input into the phoneme extractor in the speech synthesis model, the phoneme extractor will extract phoneme sequences from the long sentence text and process the extracted phoneme sequences to generate corresponding phoneme embedding sequences.

[0103] Based on the first position encoder and the phoneme encoder, feature extraction processing is performed on the phoneme embedding sequence to obtain the corresponding phoneme features.

[0104] In this embodiment, positional encoding is added to the input phoneme embedding sequence using the first position encoder to provide the order information of the elements in the phoneme embedding sequence, thus obtaining a first phoneme embedding sequence. Then, the phoneme encoder is used to extract features from the encoder-processed first phoneme embedding sequence, resulting in a second phoneme embedding sequence containing phoneme features, which can also be referred to as phoneme features.

[0105] The phoneme features are fused with the target style embedding to obtain the corresponding embedding feature vector.

[0106] In this embodiment, the fusion process refers to concatenating the phoneme features with each element of the target style embedding to form an extended phoneme embedding sequence, i.e., the aforementioned embedding feature vector, so that each factor is associated with style information. Alternatively, a conditional layer can be added after the phoneme encoder in the speech synthesis model. This conditional layer accepts the style embedding as input and generates a style adjustment matrix with the same shape as the phoneme embedding sequence. Then, this style adjustment matrix is ​​multiplied or added element-wise with the phoneme embedding sequence to adjust the phoneme embedding and obtain the corresponding embedding feature vector.

[0107] The embedded feature vector is predicted based on the multi-class predictor to obtain the corresponding first output result.

[0108] In this embodiment, the aforementioned multiple predictors include a Transformer encoder, a length predictor, an energy predictor, and a pitch predictor. Specifically, the embedded feature vector is first input into the Transformer encoder, which consists of multiple self-attention layers and feedforward neural network layers. These layers capture long-term dependencies in the sequence and generate a hidden state sequence containing rich contextual information. If style embedding is integrated into the model, the Transformer encoder will consider style information when processing phoneme embeddings, thereby embedding style features into the generated hidden states. The length predictor is used to predict the phoneme duration of each phoneme. The length adjuster predicts the duration of each phoneme based on the phoneme embedding and the hidden state sequence. Additionally, the energy predictor is used to predict energy; the energy prediction task can be generated based on the hidden states output by the Transformer encoder and the embedded feature vector. The pitch predictor is used to predict pitch; the pitch prediction task can be generated based on the hidden states output by the Transformer encoder and the embedded feature vector. In practice, additional fully connected or convolutional layers can be added after the Transformer encoder to perform capability prediction and pitch prediction tasks.

[0109] The second position encoder is used to process the first output result to obtain the corresponding second output result.

[0110] In this embodiment, the second position encoder performs position encoding processing on the input first output result and obtains the corresponding second output result.

[0111] The second output result is decoded based on the decoder to generate the corresponding specified Mel spectrogram.

[0112] In this embodiment, by decoding the second output result based on the processing output of the multi-class predictor and the second position encoder using the decoder, a corresponding specified Mel spectrogram is generated. While there may not be a direct mechanism to explicitly embed style as input during decoding, the generated Mel spectrogram naturally reflects the specific speaking style because the decoder relies on the output of the Transformer encoder (which already contains style information). Furthermore, the decoder uses position-sensitive convolutional layers or autoregressive mechanisms to generate the spectral frame sequence.

[0113] The specified Mel spectrogram is converted using the vocoder to obtain the corresponding specified long speech, and the specified long speech is used as the target speech.

[0114] In this embodiment, the vocoder can specifically be a trained WaveGlobe, HiFi-GAN, or similar vocoder. By inputting the specified Mel spectrogram into the vocoder, the vocoder converts the specified Mel spectrogram into a time-domain signal, i.e., a speech waveform, and obtains the specified long speech. During the conversion process, the vocoder preserves the tone quality and stylistic features in the Mel spectrogram.

[0115] This application inputs the long sentence text into the speech synthesis model, extracts phoneme sequences from the long sentence text using the phoneme extractor, and processes the phoneme sequences to generate phoneme embedding sequences. Then, based on the first position encoder and the phoneme encoder, feature extraction processing is performed on the phoneme embedding sequences to obtain corresponding phoneme features. Subsequently, the phoneme features are fused with the target style embedding to obtain corresponding embedding feature vectors. Subsequently, the embedding feature vectors are predicted based on the multi-class predictor to obtain a corresponding first output result. The first output result is then processed based on the second position encoder to obtain a corresponding second output result. The second output result is further decoded based on the decoder to generate a corresponding specified Mel spectrogram. Finally, the specified Mel spectrogram is converted based on the vocoder to obtain a corresponding specified long speech, and the specified long speech is used as the target speech. This application integrates a target style embedding generated by combining a first style embedding corresponding to a long sentence text predicted by a style predictor based on historical context statements, and a second style embedding obtained by a style extractor through style extraction processing of speech information at historical moments corresponding to the long sentence text, into a speech synthesis model. The target style embedding is then used to adjust the speech generation process corresponding to the long sentence text, so as to generate high-quality target speech that conforms to the target style embedding, is stylistically coherent and rhythmically smooth. This is beneficial to improving the listening quality of the target speech that matches the long sentence text, thereby enhancing the user experience.

[0116] In some optional implementations of this embodiment, step S204 includes the following steps:

[0117] Data extraction is performed on the long sentence text based on a pre-trained language model to obtain the corresponding word-level text embedding.

[0118] In this embodiment, the pre-trained language model can specifically be XLNet. XLNet is a Transformer-based pre-trained language model. XLNet combines the advantages of autoregressive and autoencoder models, aiming to fully utilize contextual information for pre-training through autoregression. It solves the problem of data inconsistency between the pre-training and fine-tuning stages caused by the use of masking in the pre-training stage of models such as BERT, and achieves bidirectional language modeling through Permutation Language Modeling (PLM). By using XLNet to extract data from the long sentence text, word-level text embeddings are extracted. Various embeddings can be used for computation, including category embeddings, position embeddings, and fragment embeddings. Category embeddings are used to distinguish between text-end markers and speech-end markers. Position embeddings provide the relative position information of sentences. Fragment embeddings provide the position of sentences within their paragraphs.

[0119] The style predictor generates a predicted style embedding sequence corresponding to the word-level text embedding.

[0120] In this embodiment, the style predictor is a context-aware style predictor that considers both text-side context information and speech-side style information from previous speeches. This style predictor can predict the speech style at time t based on the input text-side context information and the speech-side style information prior to time t. During style predictor training, since the model can see speech information before and after time t, the context information consists of a fixed number of past and future sentences excluding the current sentence. During inference, the model can only see past sentences, so the input style information is represented as the style embedding of past sentences. Here, N is the number of sentences considered in the past or future. By concatenating all 2N+1 sentences, a long text sequence is formed as the context. Specifically, the style predictor includes a sentence encoder, a context encoder, and a fused context encoder. The sentence encoder can be used to encode the word-level text embeddings extracted from the pre-trained language model, thereby exploring the meaning of low-level words in the sentence, converting the input into a hidden state sequence, and obtaining the corresponding sentence encoding. The resulting sentence encoding is then fed into the next step of the fusion context encoder to learn the high-level contextual semantics between sentences and use it for style prediction to generate the corresponding predicted style embedding sequence.

[0121] Obtain the first style embedding sequence corresponding to the historical context statement.

[0122] In this embodiment, for historical context statements, historical speech information that matches the historical context statement can be obtained, and then a style extractor can be used to extract the style of the historical speech information to generate a first style embedding sequence corresponding to the historical context statement.

[0123] The predicted style embedding sequence and the first style embedding sequence are fused to obtain the corresponding second style embedding sequence.

[0124] In this embodiment, the predicted style embedding sequence and the first style embedding sequence can be input into the aforementioned fusion context encoder. The fusion context encoder then merges the text-side predicted style embedding sequence corresponding to the long sentence text and the speech-side first style embedding sequence corresponding to the historical context sentences to form a new tag sequence, namely the aforementioned second style embedding sequence. The aforementioned merging process can refer to a splicing process.

[0125] The first style embedding is generated based on the second style embedding sequence.

[0126] In this embodiment, the second style embedding sequence can be directly used as the first style embedding.

[0127] This application extracts data from long sentence text based on a pre-trained language model to obtain corresponding word-level text embeddings. Then, a style predictor generates a predicted style embedding sequence corresponding to the word-level text embeddings. Next, a first style embedding sequence corresponding to the historical context is obtained. Subsequently, the predicted style embedding sequence and the first style embedding sequence are fused to obtain a corresponding second style embedding sequence. Finally, the first style embedding is generated based on the second style embedding sequence. This application, by using a style predictor, predicts the word-level text embeddings of long sentence text extracted by the pre-trained language model to generate a predicted style embedding sequence. Then, the first style embedding sequence corresponding to the historical context is obtained, and the predicted style embedding sequence and the first style embedding sequence are fused to generate the final first style embedding. This achieves intelligent and accurate style prediction processing of long sentence text based on the style predictor and the historical context. This facilitates subsequent speech synthesis processing of the long sentence text using the generated first style embedding and a speech synthesis model, transferring the first style embedding into the target speech so that the target speech contains the style information corresponding to the first style embedding, thereby improving the adaptability of the target speech and enhancing the user experience.

[0128] In some alternative implementations, step S205 includes the following steps:

[0129] Obtain the Mel spectrogram corresponding to the speech information.

[0130] In this embodiment, a Mel spectrogram related to the speech information can be extracted by performing feature extraction on the speech information.

[0131] Invoke the style extractor.

[0132] In this embodiment, the style extractor described above is a pre-built model capable of unsupervised learning of style features from large amounts of speech data. This style extractor does not require pre-labeled style tags but automatically discovers patterns in speech and associates these patterns with specific styles. Specifically, the style extractor includes a reference encoder and a style tagging layer. During training, the style extractor is trained unsupervised by passing the Mel spectrogram of the utterance to it and using its output as the style embedding of that utterance. This results in a style extractor capable of obtaining utterance-level style embeddings from speech. The functions of the style extractor include: Feature extraction: Through analysis of the speech signal (Mel spectrogram), the style extractor can capture key style-related features. These features can be low-level acoustic features (such as frequency, energy, duration, etc.) or higher-level style representations relatively independent of language content. Style embedding: The extracted style features are converted into a compact, continuous representation, i.e., style embedding. This embedding form is convenient for use in speech synthesis systems because it allows the model to flexibly adjust style parameters when generating speech to match specific speaking styles. Style transfer: Style extractors enable style transfer, which transfers the style of one speech to another without altering the language content. By introducing style extractors and combining them with speech synthesis models, more natural and expressive speech can be generated. Style embedding provides additional contextual information to speech synthesis models, enabling them to more accurately simulate style variations in human speech.

[0133] The Mel spectrogram is input into the style extractor, and the style extractor performs style extraction processing on the Mel spectrogram to obtain the corresponding specified style embedding.

[0134] In this embodiment, after the Mel spectrogram is input into the style extractor, the style extractor first analyzes the Mel spectrogram and extracts key features related to style, and then converts the key features into the corresponding style embedding, thus obtaining the specified style embedding.

[0135] The specified style embedding is used as the second style embedding.

[0136] This application obtains the Mel spectrogram corresponding to the speech information; then calls the style extractor; subsequently, the Mel spectrogram is input into the style extractor, which performs style extraction processing on the Mel spectrogram to obtain the corresponding specified style embedding; and finally, the specified style embedding is used as the second style embedding. By obtaining the Mel spectrogram corresponding to the speech information and then using the style extractor to perform style extraction processing on the Mel spectrogram, this application can quickly and accurately predict the discourse style embedding of the speech information and generate a second style embedding. This facilitates subsequent speech synthesis processing of long sentences using the generated second style embedding and a speech synthesis model, enabling the transfer of the aforementioned speech information into the target speech, thereby improving the adaptability of the target speech and enhancing the user experience.

[0137] In some alternative implementations, step S206 includes the following steps:

[0138] Obtain the preset splicing strategy.

[0139] In this embodiment, the above-mentioned concatenation strategy refers to the strategy of concatenating different style embeddings together to form an extended style embedding vector.

[0140] Based on the splicing strategy, the first style embedding and the second style embedding are spliced ​​together to obtain the corresponding spliced ​​style embedding.

[0141] In this embodiment, an extended style embedding vector can be formed by concatenating each element of the first style embedding and the second style embedding, namely the concatenated style embedding.

[0142] The splicing style is embedded as the target style.

[0143] This application obtains a preset splicing strategy; then, based on the splicing strategy, it splices the first style embedding and the second style embedding to obtain a corresponding spliced ​​style embedding; subsequently, it uses the spliced ​​style embedding as the target style embedding. This application, by using a splicing strategy to splice the first style embedding and the second style embedding, can quickly and accurately generate the required target style embedding, improving the generation efficiency and accuracy of the target style embedding.

[0144] In some optional implementations of this embodiment, after step S207, the electronic device may further perform the following steps:

[0145] The target speech is optimized to obtain the corresponding optimized speech.

[0146] In this embodiment, the specific implementation process of optimizing the target speech to obtain the corresponding optimized speech will be further described in detail in subsequent specific embodiments of this application, and will not be elaborated on here.

[0147] Determine the target storage method corresponding to the optimized speech.

[0148] In this embodiment, the selection of the target storage method is not specifically limited and can be determined according to actual business needs. For example, the target storage method can be any one of local database storage, cloud storage, blockchain storage, etc. Preferably, blockchain storage is used as the target storage method. By using blockchain storage to store and manage the optimized voice, the security and immutability of the optimized voice can be effectively guaranteed.

[0149] The optimized speech is stored and processed based on the target storage method.

[0150] In this embodiment, the optimized speech can be stored in the target storage medium according to the location of the target storage medium by determining the target storage medium that matches the target storage method described above, thereby completing the storage of the optimized speech.

[0151] This application optimizes the target speech to obtain optimized speech; then determines the target storage method corresponding to the optimized speech; and subsequently stores the optimized speech based on the target storage method. After generating target speech corresponding to long sentence text based on target style embedding and speech synthesis model, this application intelligently optimizes the target speech to obtain optimized speech, thereby improving the listening quality of the final speech and enhancing the user experience. Furthermore, it further utilizes the target storage method corresponding to the optimized speech for storage processing, improving the standardization and intelligence of optimized speech storage and effectively ensuring the data security of the optimized speech.

[0152] In some optional implementations of this embodiment, optimizing the target speech to obtain the corresponding optimized speech includes the following steps:

[0153] The target speech is denoised to obtain the corresponding first target speech.

[0154] In this embodiment, the target speech can be denoised using a mean filter or block matching and 3D filtering to obtain the first target speech.

[0155] The first target speech is subjected to volume normalization processing to obtain the corresponding second target speech.

[0156] In this embodiment, a preset standard volume can be obtained, and the current volume of the first target speech can be adjusted to that standard volume to obtain the corresponding second target speech. The volume value of the aforementioned standard volume can be determined based on actual volume testing experiments.

[0157] The second target speech is used as the optimized speech.

[0158] This application denoises the target speech to obtain a first target speech; then it normalizes the volume of the first target speech to obtain a second target speech; subsequently, the second target speech is used as the optimized speech. This application achieves rapid and accurate optimization of the target speech by performing denoising and volume normalization, effectively improving the listening quality of the optimized speech and thus enhancing the user experience.

[0159] It should be understood that the sequence number of each step in the above embodiments does not imply the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of the present invention.

[0160] It should be emphasized that, to further ensure the privacy and security of the aforementioned target voice, the target voice can also be stored in a node of a blockchain.

[0161] The blockchain referred to in this application is a novel application model of computer technologies such as distributed data storage, peer-to-peer transmission, consensus mechanisms, and encryption algorithms. Essentially, a blockchain is a decentralized database, a chain of data blocks linked together using cryptographic methods. Each data block contains information about a batch of network transactions, used to verify the validity of the information (anti-counterfeiting) and generate the next block. A blockchain can include an underlying blockchain platform, a platform product service layer, and an application service layer.

[0162] The embodiments of this application can acquire and process relevant data based on artificial intelligence technology. Artificial intelligence (AI) is the theory, method, technology, and application system that uses digital computers or machines controlled by digital computers to simulate, extend, and expand human intelligence, perceive the environment, acquire knowledge, and use that knowledge to obtain optimal results.

[0163] Foundational technologies for artificial intelligence generally include sensors, dedicated AI chips, cloud computing, distributed storage, big data processing, operating / interactive systems, and mechatronics. AI software technologies mainly encompass computer vision, robotics, biometrics, speech processing, natural language processing, and machine learning / deep learning.

[0164] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by instructing related hardware through computer-readable instructions. These computer-readable instructions can be stored in a computer-readable storage medium. When the program is executed, it can include the processes of the embodiments of the methods described above. The aforementioned storage medium can be a non-volatile storage medium such as a magnetic disk, optical disk, or read-only memory (ROM), or random access memory (RAM).

[0165] It should be understood that although the steps in the flowcharts of the accompanying figures are shown sequentially as indicated by the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless explicitly stated herein, there is no strict order restriction on the execution of these steps, and they can be executed in other orders. Moreover, at least some steps in the flowcharts of the accompanying figures may include multiple sub-steps or multiple stages. These sub-steps or stages are not necessarily completed at the same time, but can be executed at different times, and their execution order is not necessarily sequential, but can be performed alternately or in turn with other steps or at least some of the sub-steps or stages of other steps.

[0166] Further reference Figure 3 As a response to the above Figure 2 The implementation of the method shown in this application provides an embodiment of an artificial intelligence-based speech synthesis device, which is similar to... Figure 2 Corresponding to the method embodiments shown, this device can be specifically applied to various electronic devices.

[0167] like Figure 3 As shown, the artificial intelligence-based speech synthesis device 300 described in this embodiment includes: a first acquisition module 301, a second acquisition module 302, a third acquisition module 303, a prediction module 304, an extraction module 305, a generation module 306, and a processing module 307. Wherein:

[0168] The first acquisition module 301 is used to acquire the long sentence text to be processed;

[0169] The second acquisition module 302 is used to acquire the historical context statements corresponding to the long sentence text;

[0170] The third acquisition module 303 is used to acquire the voice information of the historical moment corresponding to the long sentence text;

[0171] The prediction module 304 is used to perform style prediction processing on the long sentence text based on a preset style predictor and the historical context statement to obtain the corresponding first style embedding.

[0172] Extraction module 305 is used to perform style extraction processing on the speech information based on a preset style extractor to obtain the corresponding second style embedding;

[0173] Generation module 306 is used to generate a target style embedding based on the first style embedding and the second style embedding;

[0174] The processing module 307 is used to perform speech synthesis processing on the long sentence text based on the target style embedding and a preset speech synthesis model to generate target speech corresponding to the long sentence text; wherein, the speech synthesis model includes a phoneme extractor, a first position encoder, a phoneme encoder, a multi-class predictor, a second position encoder, a decoder, and a vocoder.

[0175] In this embodiment, the operations performed by the above modules or units correspond one-to-one with the steps of the artificial intelligence-based speech synthesis method in the aforementioned embodiments, and will not be repeated here.

[0176] In some optional implementations of this embodiment, the processing module 307 includes:

[0177] The first processing submodule is used to input the long sentence text into the speech synthesis model, extract and process the long sentence text through the phoneme extractor to obtain a phoneme sequence, and process the phoneme sequence to generate a phoneme embedding sequence.

[0178] The second processing submodule is used to perform feature extraction processing on the phoneme embedding sequence based on the first position encoder and the phoneme encoder to obtain the corresponding phoneme features.

[0179] The third processing submodule is used to fuse the phoneme features with the target style embedding to obtain the corresponding embedding feature vector.

[0180] The fourth processing submodule is used to perform prediction processing on the embedded feature vector based on the multi-class predictor to obtain the corresponding first output result;

[0181] The fifth processing submodule is used to process the first output result based on the second position encoder to obtain the corresponding second output result;

[0182] The sixth processing submodule is used to decode the second output result based on the decoder to generate the corresponding specified Mel spectrogram;

[0183] The first determining submodule is used to perform conversion processing on the specified Mel spectrogram based on the vocoder to obtain the corresponding specified long speech, and to use the specified long speech as the target speech.

[0184] In this embodiment, the operations performed by the above modules or units correspond one-to-one with the steps of the artificial intelligence-based speech synthesis method in the aforementioned embodiments, and will not be repeated here.

[0185] In some optional implementations of this embodiment, the prediction module 304 includes:

[0186] The first extraction submodule is used to extract data from the long sentence text based on a preset pre-trained language model to obtain the corresponding word-level text embedding;

[0187] The first generation submodule is used to generate a predicted style embedding sequence corresponding to the word-level text embedding through the style predictor;

[0188] The first acquisition submodule is used to acquire the first style embedding sequence corresponding to the historical context statement;

[0189] The fusion submodule is used to fuse the predicted style embedding sequence and the first style embedding sequence to obtain the corresponding second style embedding sequence.

[0190] The second generation submodule is used to generate the first style embedding based on the second style embedding sequence.

[0191] In this embodiment, the operations performed by the above modules or units correspond one-to-one with the steps of the artificial intelligence-based speech synthesis method in the aforementioned implementation method, and will not be repeated here.

[0192] In some optional implementations of this embodiment, the extraction module 305 includes:

[0193] The second acquisition submodule is used to acquire the Mel spectrogram corresponding to the voice information;

[0194] Call the submodule to invoke the style extractor;

[0195] The second extraction submodule is used to input the Mel spectrogram into the style extractor, and to perform style extraction processing on the Mel spectrogram to obtain the corresponding specified style embedding.

[0196] The second determining submodule is used to embed the specified style as the second style embedding.

[0197] In this embodiment, the operations performed by the above modules or units correspond one-to-one with the steps of the artificial intelligence-based speech synthesis method in the aforementioned embodiments, and will not be repeated here.

[0198] In some optional implementations of this embodiment, the generation module 306 includes:

[0199] The third acquisition submodule is used to acquire the preset splicing strategy;

[0200] The splicing submodule is used to splice the first style embedding and the second style embedding based on the splicing strategy to obtain the corresponding splicing style embedding.

[0201] The third determining submodule is used to embed the splicing style as the target style embedding.

[0202] In this embodiment, the operations performed by the above modules or units correspond one-to-one with the steps of the artificial intelligence-based speech synthesis method in the aforementioned embodiments, and will not be repeated here.

[0203] In some optional implementations of this embodiment, the artificial intelligence-based speech synthesis device further includes:

[0204] The optimization module is used to optimize the target speech to obtain the corresponding optimized speech;

[0205] The determining module is used to determine the target storage method corresponding to the optimized speech;

[0206] A storage module is used to store and process the optimized voice based on the target storage method.

[0207] In this embodiment, the operations performed by the above modules or units correspond one-to-one with the steps of the artificial intelligence-based speech synthesis method in the aforementioned embodiments, and will not be repeated here.

[0208] In some optional implementations of this embodiment, the optimization module includes:

[0209] The seventh processing submodule is used to perform noise reduction processing on the target speech to obtain the corresponding first target speech;

[0210] The eighth processing submodule is used to perform volume normalization processing on the first target speech to obtain the corresponding second target speech;

[0211] The fourth determining submodule is used to use the second target speech as the optimized speech.

[0212] In this embodiment, the operations performed by the above modules or units correspond one-to-one with the steps of the artificial intelligence-based speech synthesis method in the aforementioned embodiments, and will not be repeated here.

[0213] To address the aforementioned technical problems, embodiments of this application also provide a computer device. Please refer to [link / reference needed]. Figure 4 , Figure 4 This is a basic structural block diagram of the computer device in this embodiment.

[0214] The computer device 4 includes a memory 41, a processor 42, and a network interface 43 that are interconnected via a system bus. It should be noted that only the computer device 4 with components 41-43 is shown in the figure; however, it should be understood that it is not required to implement all the shown components, and more or fewer components can be implemented alternatively. Those skilled in the art will understand that the computer device described here is a device capable of automatically performing numerical calculations and / or information processing according to pre-set or stored instructions. Its hardware includes, but is not limited to, microprocessors, application-specific integrated circuits (ASICs), programmable gate arrays (FPGAs), digital digital processors (DSPs), embedded devices, etc.

[0215] The computer device can be a desktop computer, laptop, handheld computer, or cloud server, etc. The computer device can interact with the user via a keyboard, mouse, remote control, touchpad, or voice control.

[0216] The memory 41 includes at least one type of readable storage medium, including flash memory, hard disk, multimedia card, card-type memory (e.g., SD or DX memory), random access memory (RAM), static random access memory (SRAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), programmable read-only memory (PROM), magnetic memory, magnetic disk, optical disk, etc. In some embodiments, the memory 41 may be an internal storage unit of the computer device 4, such as the hard disk or memory of the computer device 4. In other embodiments, the memory 41 may also be an external storage device of the computer device 4, such as a plug-in hard disk, smart memory card (SMC), secure digital card (SD) card, flash memory card, etc. of the computer device 4. Of course, the memory 41 may also include both the internal storage unit and the external storage device of the computer device 4. In this embodiment, the memory 41 is typically used to store the operating system and various application software installed on the computer device 4, such as computer-readable instructions based on artificial intelligence speech synthesis methods. In addition, the memory 41 can also be used to temporarily store various types of data that have been output or will be output.

[0217] In some embodiments, the processor 42 may be a central processing unit (CPU), a controller, a microcontroller, a microprocessor, or other data processing chip. The processor 42 is typically used to control the overall operation of the computer device 4. In this embodiment, the processor 42 is used to execute computer-readable instructions stored in the memory 41 or to process data, for example, to execute computer-readable instructions of the artificial intelligence-based speech synthesis method.

[0218] The network interface 43 may include a wireless network interface or a wired network interface, which is typically used to establish communication connections between the computer device 4 and other electronic devices.

[0219] This application also provides another embodiment, namely, providing a computer-readable storage medium storing computer-readable instructions that can be executed by at least one processor to cause the at least one processor to perform the steps of the artificial intelligence-based speech synthesis method described above.

[0220] Through the above description of the embodiments, those skilled in the art can clearly understand that the methods of the above embodiments can be implemented by means of software plus necessary general-purpose hardware platforms. Of course, they can also be implemented by hardware, but in many cases the former is a better implementation method. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product is stored in a storage medium (such as ROM / RAM, magnetic disk, optical disk), and includes several instructions to cause a terminal device (which may be a mobile phone, computer, server, air conditioner, or network device, etc.) to execute the methods described in the various embodiments of this application.

[0221] Obviously, the embodiments described above are only some embodiments of this application, not all embodiments. The accompanying drawings show preferred embodiments of this application, but do not limit the patent scope of this application. This application can be implemented in many different forms; rather, the purpose of providing these embodiments is to provide a more thorough and comprehensive understanding of the disclosure of this application. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing specific embodiments, or make equivalent substitutions for some of the technical features. Any equivalent structures made using the content of this application's specification and drawings, directly or indirectly applied to other related technical fields, are similarly within the scope of patent protection of this application.

Claims

1. An artificial intelligence-based speech synthesis method, characterized by, The method comprises the following steps: acquiring a long sentence text to be processed; acquiring a historical context sentence corresponding to the long sentence text; acquiring speech information of a historical moment corresponding to the long sentence text; performing style prediction processing on the long sentence text based on a preset style predictor and the historical context sentence, to obtain a corresponding first style embedding; performing style extraction processing on the speech information based on a preset style extractor, to obtain a corresponding second style embedding; generating a target style embedding based on the first style embedding and the second style embedding; performing speech synthesis processing on the long sentence text based on the target style embedding and a preset speech synthesis model, to generate target speech corresponding to the long sentence text; wherein the speech synthesis model comprises a phoneme extractor, a first position encoder, a phoneme encoder, a multi-class predictor, a second position encoder, a decoder, and a vocoder. 2.The AI-based speech synthesis method of claim 1, wherein, The step of performing speech synthesis processing on the long sentence text based on the target style embedding and a preset speech synthesis model, to generate target speech corresponding to the long sentence text, specifically comprises: inputting the long sentence text into the speech synthesis model, performing extraction processing on the long sentence text by the phoneme extractor to obtain a phoneme sequence, and performing processing on the phoneme sequence to generate a phoneme embedding sequence; performing feature extraction processing on the phoneme embedding sequence based on the first position encoder and the phoneme encoder, to obtain corresponding phoneme features; performing fusion processing on the phoneme features and the target style embedding, to obtain a corresponding embedding feature vector; performing prediction processing on the embedding feature vector based on the multi-class predictor, to obtain a corresponding first output result; performing processing on the first output result based on the second position encoder, to obtain a corresponding second output result; performing decoding processing on the second output result based on the decoder, to generate a corresponding specified mel-spectrogram; performing conversion processing on the specified mel-spectrogram based on the vocoder to obtain a corresponding specified long speech, and taking the specified long speech as the target speech. 3.The AI-based voice synthesis method of claim 1, wherein, The step of performing style prediction processing on the long sentence text based on a preset style predictor and the historical context sentence, to obtain a corresponding first style embedding, specifically comprises: performing data extraction on the long sentence text based on a preset pre-trained language model, to obtain a corresponding word-level text embedding; generating a prediction style embedding sequence corresponding to the word-level text embedding by the style predictor; acquiring a first style embedding sequence corresponding to the historical context sentence; performing fusion processing on the prediction style embedding sequence and the first style embedding sequence, to obtain a corresponding second style embedding sequence; generating the first style embedding based on the second style embedding sequence. 4.The AI-based voice synthesis method of claim 1, wherein, The step of performing style extraction processing on the speech information based on a preset style extractor, to obtain a corresponding second style embedding, specifically comprises: acquiring a mel-spectrogram corresponding to the speech information; calling the style extractor; inputting the mel-spectrogram into the style extractor, and performing style extraction processing on the mel-spectrogram by the style extractor to obtain a corresponding specified style embedding; taking the specified style embedding as the second style embedding. 5.The AI-based voice synthesis method of claim 1, wherein, The step of generating a target style embedding based on the first style embedding and the second style embedding specifically includes: obtaining a preset splicing strategy; performing splicing processing on the first style embedding and the second style embedding based on the splicing strategy to obtain a corresponding spliced style embedding; taking the spliced style embedding as the target style embedding. 6.The AI-based voice synthesis method of claim 1, wherein, After the step of performing speech synthesis processing on the long sentence text based on the target style embedding and a preset speech synthesis model to generate a target speech corresponding to the long sentence text, the method further includes: performing optimization processing on the target speech to obtain a corresponding optimized speech; determining a target storage manner corresponding to the optimized speech; performing storage processing on the optimized speech based on the target storage manner. 7.The artificial intelligence-based speech synthesis method of claim 6, wherein, The step of performing optimization processing on the target speech to obtain a corresponding optimized speech specifically includes: performing denoising processing on the target speech to obtain a corresponding first target speech; performing volume normalization processing on the first target speech to obtain a corresponding second target speech; taking the second target speech as the optimized speech.

8. An artificial intelligence-based speech synthesis device, characterized by, The method includes: a first obtaining module configured to obtain a long sentence text to be processed; a second obtaining module configured to obtain a historical context sentence corresponding to the long sentence text; a third obtaining module configured to obtain speech information of a historical time corresponding to the long sentence text; a prediction module configured to perform style prediction processing on the long sentence text based on a preset style predictor and the historical context sentence to obtain a corresponding first style embedding; an extraction module configured to perform style extraction processing on the speech information based on a preset style extractor to obtain a corresponding second style embedding; a generation module configured to generate a target style embedding based on the first style embedding and the second style embedding; a processing module configured to perform speech synthesis processing on the long sentence text based on the target style embedding and a preset speech synthesis model to generate a target speech corresponding to the long sentence text; wherein the speech synthesis model includes a phoneme extractor, a first position encoder, a phoneme encoder, a multi-class predictor, a second position encoder, a decoder, and a vocoder.

9. A computer device including a memory and a processor, the memory having stored therein computer readable instructions, the processor implementing the steps of the artificial intelligence-based speech synthesis method of any one of claims 1 to 7 when executing the computer readable instructions.

10. A computer-readable storage medium, characterized in that, The computer readable storage medium has stored thereon computer readable instructions, which are executed by the processor to implement the steps of the artificial intelligence-based speech synthesis method of any one of claims 1 to 7.

Images