Speech synthesis method and device based on optimization strategy algorithm, equipment and medium

Abstract

The invention relates to the technical field of intelligent decision making, can be applied to business system platforms of financial science and technology, medical health and the like, and discloses a speech synthesis method, device, equipment and medium based on an optimization strategy algorithm, comprising: extracting a time sequence processing network unit and a data sampling scheduling unit in a speech synthesis model; mapping a denoising function in the time sequence processing network unit into a multi-step Markov decision function, and converting an ordinary differential equation in the data sampling scheduling unit into a multi-source stochastic differential equation; sampling multiple groups of independent audio tracks corresponding to the input text based on a multi-source stochastic differential equation; calculating a strategy gradient modulation factor by using an optimization strategy algorithm and a multi-step Markov decision function; optimizing strategy parameters of the speech synthesis model according to the strategy gradient modulation factor to obtain an optimized speech synthesis model; and obtaining a to-be-converted text, and synthesizing voice corresponding to the to-be-converted text by using the optimized voice synthesis model. And the speech synthesis accuracy is improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of intelligent decision-making, and in particular to a speech synthesis method and device based on an optimization strategy algorithm, equipment and a medium. BACKGROUND

[0002] Text-to-speech technology is a technology that uses a speech synthesis system to convert input text information into audible speech output. It simulates human pronunciation characteristics, intonation and rhythm, and accurately converts text into continuous, natural and human-recognizable speech signals based on text content. However, existing technologies still have deficiencies in the naturalness, intelligibility and emotional expression of generated speech, especially when dealing with complex contexts and diverse user needs.

[0003] In the field of financial technology, text-to-speech technology can be used to develop intelligent customer service systems to provide automated customer service. However, due to the lack of naturalness and emotional expression of speech, the customer experience is poor, affecting customer satisfaction.

[0004] In the field of medical health, text-to-speech technology can be used to assist doctors in communicating with patients. However, due to the lack of intelligibility of speech, the accuracy and efficiency of information transmission are affected.

[0005] The inventors realized that the synthesis effect of existing text-to-speech technology is heavily dependent on the performance of the speech generation model. However, due to the lack of adaptability of the generation mechanism to the reinforcement learning requirements, the model optimization capability is limited, which in turn makes the text-to-speech accuracy low in complex scenarios. SUMMARY

[0006] The present application provides a speech synthesis method and device based on an optimization strategy algorithm to solve the technical problem of low speech accuracy for speech synthesis.

[0007] In a first aspect, a speech synthesis method based on an optimization strategy algorithm is provided, comprising:

[0008] extracting a time series processing network unit and a data sampling scheduling unit in a preset speech synthesis model;

[0009] mapping a denoising function in the time series processing network unit to a multi-step Markov decision function, and converting an ordinary differential equation in the data sampling scheduling unit to a multi-source stochastic differential equation;

[0010] sampling a plurality of groups of independent audio tracks corresponding to a preset input text based on the multi-source stochastic differential equation;

[0011] calculating a strategy gradient modulation factor of the plurality of groups of independent audio tracks using a preset optimization strategy algorithm and the multi-step Markov decision function;

[0012] According to the policy gradient modulation factor, policy parameters of the speech synthesis model are optimized to obtain an optimized speech synthesis model.

[0013] Text to be converted is acquired, and the optimized speech synthesis model is used to synthesize speech corresponding to the text to be converted.

[0014] In a second aspect, a speech synthesis device based on an optimized policy algorithm is provided, and the device includes:

[0015] A unit extraction module is configured to extract a time sequence processing network unit and a data sampling scheduling unit in a preset speech synthesis model.

[0016] A unit mapping module is configured to map a denoising function in the time sequence processing network unit into a multi-step Markov decision function, and convert a common differential equation in the data sampling scheduling unit into a multi-source stochastic differential equation.

[0017] An audio track sampling module is configured to sample a plurality of groups of independent audio tracks corresponding to a preset input text based on the multi-source stochastic differential equation.

[0018] A policy gradient modulation factor calculation module is configured to calculate policy gradient modulation factors of the plurality of groups of independent audio tracks by using a preset optimization policy algorithm and the multi-step Markov decision function.

[0019] A speech synthesis model optimization module is configured to optimize policy parameters of the speech synthesis model according to the policy gradient modulation factors, to obtain an optimized speech synthesis model.

[0020] A speech synthesis module is configured to acquire text to be converted, and use the optimized speech synthesis model to synthesize speech corresponding to the text to be converted.

[0021] In a third aspect, a computer device is provided, which includes a memory, a processor, and a computer program stored in the memory and executable on the processor, and the processor implements steps of the speech synthesis method based on the optimized policy algorithm when executing the computer program.

[0022] In a fourth aspect, a computer readable storage medium is provided, which stores a computer program, and the computer program implements steps of the speech synthesis method based on the optimized policy algorithm when executed by a processor.

[0023] The scheme implemented by the speech synthesis method, device, equipment and medium based on the optimization strategy algorithm can extract a time sequence processing network unit and a data sampling scheduling unit in a preset speech synthesis model through a client; map a denoising function in the time sequence processing network unit into a multi-step Markov decision function, and convert an ordinary differential equation in the data sampling scheduling unit into a multi-source stochastic differential equation; sample multiple groups of independent audio tracks corresponding to a preset input text based on the multi-source stochastic differential equation; calculate strategy gradient modulation factors of the multiple groups of independent audio tracks by using a preset optimization strategy algorithm and the multi-step Markov decision function; optimize strategy parameters of the speech synthesis model according to the strategy gradient modulation factors, to obtain an optimized speech synthesis model; obtain text to be converted, and synthesize speech corresponding to the text to be converted by using the optimized speech synthesis model. The synthesized speech is fed back to the client. In the application, the core unit is extracted to realize accurate disassembly of the model structure, laying a foundation for subsequent optimization. The denoising function is mapped into the multi-step Markov decision function, which can realize step-by-step fine control of the denoising process, improve the accuracy of speech feature processing, and convert the ordinary differential equation into the multi-source stochastic differential equation, which can introduce diversified random disturbances and enhance the flexibility and diversity of audio generation. Sampling multiple groups of independent audio tracks based on the multi-source stochastic differential equation can provide rich sample references and provide more comprehensive basis for model optimization. The optimization strategy algorithm and the multi-step Markov decision function are used to calculate the strategy gradient modulation factors, which can accurately quantify the quality contribution of different tracks and indicate the direction of parameter adjustment. The strategy parameters are optimized according to the modulation factors, which can enable the model to focus on high-quality feature learning, accelerate convergence and improve the naturalness and adaptability of synthesized speech. Finally, the optimized model is used to synthesize speech, which can efficiently complete the text-to-speech conversion and improve the accuracy of the speech synthesis system. BRIEF DESCRIPTION OF DRAWINGS

[0024] In order to more clearly illustrate the technical solutions of the embodiments of the present application, the following will briefly introduce the drawings needed to be used in the description of the embodiments of the present application. Obviously, the drawings in the following description are only some embodiments of the present application, and other drawings can also be obtained by those skilled in the art without creative labor.

[0025] Figure 1 is an application environment schematic diagram of the speech synthesis method based on the optimization strategy algorithm in an embodiment of the present application;

[0026] Figure 2 is a flowchart of the speech synthesis method based on the optimization strategy algorithm in an embodiment of the present application;

[0027] Figure 3 is Figure 2is a flow chart of a specific embodiment of step S2 in the method;

[0028] Figure 4 is Figure 2 is a flow chart of a specific embodiment of step S4 in the method;

[0029] Figure 5 is a structural schematic diagram of a speech synthesis device based on an optimization strategy algorithm in an embodiment of the present application;

[0030] Figure 6 is a structural schematic diagram of a computer device in an embodiment of the present application;

[0031] Figure 7 is another structural schematic diagram of a computer device in an embodiment of the present application. DETAILED DESCRIPTION

[0032] The technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are some of the embodiments of the present application, rather than all the embodiments of the present application. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative effort belong to the scope of protection of the present application.

[0033] The speech synthesis method based on an optimization strategy algorithm provided by the embodiments of the present application can be applied in, for example, a computer device. Figure 1In an application environment of the application, the client communicates with the server through a network. The server can extract a time sequence processing network unit and a data sampling scheduling unit in a preset voice synthesis model through the client; map a denoising function in the time sequence processing network unit as a multi-step Markov decision function, and convert an ordinary differential equation in the data sampling scheduling unit as a multi-source stochastic differential equation; sample multiple groups of independent audio tracks corresponding to a preset input text based on the multi-source stochastic differential equation; calculate a strategy gradient modulation factor of the multiple groups of independent audio tracks by using a preset optimization strategy algorithm and the multi-step Markov decision function; optimize a strategy parameter of the voice synthesis model according to the strategy gradient modulation factor, to obtain an optimized voice synthesis model; obtain a text to be converted, and synthesize a voice corresponding to the text to be converted by using the optimized voice synthesis model. The synthesized voice is fed back to the client. In the application, the core unit is extracted to realize accurate disassembly of the model structure, to lay a foundation for subsequent optimization; the denoising function is mapped as the multi-step Markov decision function, to realize step-by-step fine control of the denoising process, to improve the accuracy of voice feature processing, and to convert the ordinary differential equation as the multi-source stochastic differential equation, to introduce diversified random disturbance, to enhance the flexibility and diversity of audio generation; the multiple groups of independent audio tracks are sampled based on the multi-source stochastic differential equation, to provide rich sample references, to provide more comprehensive basis for model optimization; the strategy gradient modulation factor is calculated by using the optimization strategy algorithm and the multi-step Markov decision function, to accurately quantify the quality contribution of different tracks, to indicate the direction for parameter adjustment; the strategy parameter is optimized according to the modulation factor, to enable the model to focus on high-quality feature learning, to accelerate convergence and to improve the naturalness and adaptability of synthesized voice; finally, the voice is synthesized by using the optimized model, to efficiently complete the conversion from text to voice, and to improve the accuracy of the voice synthesis system. The client can be, but is not limited to, various personal computers, notebook computers, smart phones, tablet computers and portable wearable devices. The server can be implemented by an independent server or a server cluster composed of multiple servers. The application will be described in detail through specific embodiments.

[0034] Please refer to Figure 2 as shown, Figure 2 A flowchart of a voice synthesis method based on an optimization strategy algorithm provided by an embodiment of the application is shown, including the following steps:

[0035] S1, extract a time sequence processing network unit and a data sampling scheduling unit in a preset voice synthesis model.

[0036] In the embodiment of the application, the voice synthesis model is a system based on a flow matching model, which relies on an ordinary differential equation to generate voice, and can generate voice output with high fidelity and naturalness according to text input.

[0037] In detail, the temporal processing network unit is the core structure responsible for denoising and feature conversion in stream matching-based speech synthesis models (such as F5-TTS). The denoising function is integrated into the diffusion Transformer module by stacking multiple layers of attention mechanism and convolution operations.

[0038] Specifically, the data sampling scheduling unit is a module that embeds ordinary differential equations (ODEs) in the model and is used to control the generation process. It describes the continuous evolution trajectory from random noise to target speech features and schedules the iterative update of feature states according to preset time steps.

[0039] Next, we first load the structure and parameters of the entire speech synthesis model. Following the module division logic defined during model design, we identify the substructures responsible for temporal feature processing (such as multi-layer attention and convolutional stacking for denoising) within the model's network hierarchy. These substructures and parameters are then separated out to form the temporal processing network unit. Next, we identify the submodules that embed ordinary differential equations and govern the feature generation process (evolving from noise to target speech). We extract their equation logic, scheduling rules, and other information to serve as data sampling and scheduling units.

[0040] In an embodiment of the present invention, the timing processing network unit and the data sampling scheduling unit in the preset speech synthesis model are extracted to realize modular design, which is convenient for optimizing the timing feature processing accuracy and sampling generation efficiency separately, improving the model reusability and iterative flexibility, while reducing the complexity of improving specific functions (such as denoising logic and generation rhythm).

[0041] In the field of financial technology, by extracting these two units, the timing processing unit can be optimized to improve the pronunciation accuracy of financial terms, and the scheduling unit can be adjusted to speed up the response speed of market reports, ensuring that users obtain accurate and timely financial information.

[0042] In the medical and health field, after the electronic medical record voice conversion tool extracts the unit, it can optimize the timing processing unit to improve the clarity of medical terminology, adapt to the doctor's oral rhythm through the scheduling unit, and reduce the risk of mishearing during medical record conversion.

[0043] S2. Mapping the denoising function in the time series processing network unit into a multi-step Markov decision function, and converting the ordinary differential equation in the data sampling scheduling unit into a multi-source stochastic differential equation.

[0044] In an embodiment of the present invention, the multi-step Markov decision function is a mathematical model that formalizes the sequence generation process as a sequential decision, which defines the noisy data of each step as a state, the denoising operation predicted by the model as an action, and the policy function maps the state to the probability distribution of the action, and describes the complete generation trajectory through deterministic state transitions and a sparse reward mechanism (output quality is evaluated only in the final step).

[0045] In the embodiments of the present application, referring to Figure 3 As shown in the figure, the mapping of the denoising function in the time series processing network unit to a multi-step Markov decision function comprises:

[0046] S31, discretize the denoising process of the denoising function into multiple time steps, and map the noisy speech data in the denoising function to Markov decision state data in each time step;

[0047] S32, generate a speech estimate value according to the state data, and map the speech estimate value to action data of Markov decision;

[0048] S33, construct a multi-step Markov decision function based on the action data and the state data.

[0049] In detail, first, the total iteration step T is preset, and the continuous generation time interval [0, 1] is uniformly discretized into T ordered time points {t_T, t_{T-1},..., t_0}. At the beginning of each iteration step, the program reads the noisy speech data vector x_t of the target step from the memory buffer. Then, the target time step index t is converted into an embedding vector, which is spliced with the encoded input text feature vector to form a conditional context vector. The noisy speech data x_t, the target time step index t and the conditional context vector are jointly encapsulated into a structured data object s_t=(x_t, t, c), which is defined as the state data of the Markov decision process.

[0050] Specifically, the noisy speech data vector x_t, the time step index t and the conditional text vector c are deconstructed from the state data s_t. Then, the program calls the forward propagation interface of the loaded time series processing network with the noisy speech data vector, the time step index and the conditional text vector as input parameters. In the network, the data is first down-sampled and feature-extracted through the encoder layer, then fused with the conditional text vector through the attention mechanism layer, and finally up-sampled and reconstructed features through the decoder layer. Finally, a vector x_0 with the same dimension as the input noisy speech data vector is output, which is the clean speech estimate value corresponding to the target step predicted by the model. The output vector x_0 is encapsulated as an action data object a_t.

[0051] Then, according to the rules of ordinary differential equations of flow matching, the numerical solver (such as Euler method) is used to calculate the determined state s_{t-1} of the target time from the target state data s_t and the action data, and the calculation process is defined as the state transition function P(s_{t-1}|s_t, a_t)

[0052] Further, the state data, action data, policy function (i.e., the neural network itself), state transition function, and reward function R(s_t, a_t) defined above are combined to form a complete multi-step Markov decision function (S, A, p0, P, R) that can be used in a reinforcement learning framework.

[0053] In the embodiment of the present application, the multi-source stochastic differential equation is a dynamic equation introducing multiple independent random noise sources on the basis of ordinary differential equation, and by explicitly adding a drift compensation term and a time-controlled noise intensity term, the deterministic generating path is converted into a stochastic process, and under the premise of ensuring the consistency of data distribution, controllable randomness is injected for sampling to support the exploration required by reinforcement learning.

[0054] In the embodiment of the present application, the conversion of the ordinary differential equation in the data sampling scheduling unit into a multi-source stochastic differential equation comprises:

[0055] The velocity field in the ordinary differential equation is obtained, the score function of the velocity field is calculated, and the time step in the velocity field is extracted;

[0056] The total noise intensity and the weight angle corresponding to the time step are calculated according to a predefined noise intensity function, and the intensity coefficients corresponding to the total noise intensity and the weight angle are calculated respectively;

[0057] The velocity field, the score function, and the intensity coefficient are combined into a multi-source stochastic differential equation.

[0058] In detail, the dynamics rule dx_t=v_θ(x_t,t,c) of the ordinary differential equation is first analyzed and the deterministic velocity field function v_θ therein is located, and the time step in the velocity field is extracted.

[0059] Specifically, the noisy data vector x_t of the target time step, the time step scalar t, and the calculated velocity field vector v_θ(x_t,t,c) are read. Then, a division operation x_t / t is performed on each element of the noisy data vector, and a scaling factor (1-t) / t is calculated and multiplied with the velocity field vector v_θ element by element to obtain an intermediate result vector. Then, the program calls a vector adder to add the two result vectors element by element, but takes the negative sign, i.e., performs the calculation score=-(x_t / t)-((1-t) / t)*v_θ. Finally, the program performs numerical stability check (such as preventing division by zero error) on the output vector, and finally obtains the score function logp_t(x_t).

[0060] Further, take the target time step index t as an input parameter, call the noise scheduling function sigma(t) and the angle scheduling function omega(t) preset in the memory (the two functions are usually implemented in a lookup table or an analytical expression). Then, calculate the target total noise intensity scalar value sigma_current=sigma(t) and the weight angle radian value omega_current=omega(t) through the floating-point operation unit. Then, call the trigonometric function calculator in the mathematical library to perform cos(omega_current) and sin(omega_current) operations respectively to obtain two weight coefficients. Finally, perform scalar multiplication operation on the total noise intensity and the weight coefficients to calculate the final intensity coefficients of the two noise channels sigma1=sigma_current*cos(omega_current) and sigma2=sigma_current*sin(omega_current).

[0061] Then, linearly combine the velocity field v_ theta(x_t, t, c) with the score function term to form the drift term [V-X / t-(1-t)V / t]dt, and multiply the two noise intensity coefficients with the corresponding random increments to form the diffusion term cos(omega(t))*sigma(t)*dW0(t)+sin(omega(t))*sigma(t)*dW1(t). Add the drift term and the diffusion term through the floating-point operation unit to obtain the complete multi-source stochastic differential equation dx_t=[v_t(x_t)-(x_t / t)-((1-t) / t)v_t(x_t)]*dt+cos(omega(t))*sigma(t)*dW1(t)+sin(omega(t))*sigma(t)*dW2(t).

[0062] In the embodiment of the application, the denoising function in the time series processing network unit is mapped to a multi-step Markov decision function, and the ordinary differential equation in the data sampling scheduling unit is converted into a multi-source stochastic differential equation. The core benefits are that the traditional deterministic generation process is converted into an optimized and exploratory random decision framework, and the fitting ability of the model to complex data distribution and the generation quality are significantly improved through the introduction of a reinforcement learning mechanism. Mapping to a multi-step Markov decision function makes each step of the generation process guided and optimized by a reward function, and converting into a multi-source stochastic differential equation enhances exploration efficiency by injecting controllable randomness, avoids local optimum, and finally realizes higher fidelity and diversity of data generation.

[0063] In the field of financial technology, multi-step Markov decision can be applied to high-frequency trading strategy generation, taking the market noise state as the multi-step Markov decision state and the trading instruction as the action, exploring diversified strategy paths through the multi-source stochastic differential equation, and using the yield index as the reward to optimize the decision model to generate a high-return and risk-diversified trading strategy.

[0064] In the field of financial technology, the voice interaction of intelligent customer service can be optimized to ensure accurate pronunciation of financial terms through multi-step decision-making, and to improve service experience by generating voice styles suitable for different customers from multiple sources.

[0065] In the field of medical health, multi-step Markov decision can be used for medical image synthesis, taking low-resolution or noisy images as states and denoising reconstruction operations as actions to generate multiple possible high-definition pathological image candidates through stochastic differential equations, and then optimizing the generation process by taking image consistency as the reward function.

[0066] S3, sampling a plurality of groups of independent audio tracks corresponding to the preset input text based on the multi-source stochastic differential equation.

[0067] In the embodiment of the application, the input text refers to the original character sequence provided by the user and to be converted into voice, containing language information (such as vocabulary, sentence or paragraph) that needs to be synthesized, which is the core semantic and linguistic feature of the condition input control generation process of the speech synthesis model.

[0068] In the embodiment of the application, the independent audio track refers to the acoustic waveform data sequence generated by sampling based on the same input text through the multi-source stochastic differential equation. Each track corresponds to a solution path of the stochastic differential equation in mathematics, and shows randomness differences in prosody, emotion and other paralinguistic features in acoustic characteristics. Each track takes time as the axis and is composed of numerical arrays of parameters such as amplitude and frequency of audio signal changing with time.

[0069] In the embodiment of the application, the sampling of a plurality of groups of independent audio tracks corresponding to the preset input text based on the multi-source stochastic differential equation comprises:

[0070] Encoding the preset input text to obtain a text vector;

[0071] Randomly sampling an initial state of each preset random seed from a standard Gaussian distribution;

[0072] Iterating the target state in the multi-source stochastic differential equation from the initial state according to a preset time step, and in each iteration, calculating a target velocity field of the input text according to the target state, the time step and the text vector;

[0073] Updating the target state in the multi-source stochastic differential equation according to the target velocity field;

[0074] When the iteration reaches the time step of zero, aggregating the initial state and all target states into an audio track of the input text;

[0075] Combining the audio track into a plurality of groups of audio tracks corresponding to the input text.

[0076] In detail, the text is segmented into words or characters, and then each word is converted into a fixed-length vector using a pre-trained word embedding model. All word vectors are then concatenated to form a long vector, i.e., a text vector.

[0077] Next, K random seed values that do not overlap are generated according to a preset number of trajectories K, random scalar arrays conforming to the standard normal distribution N(0, 1) are randomly generated, and these arrays are reshaped into a floating-point matrix matching the dimension of the target audio vector (such as [K, 1, 80, 400]), and the generated random vectors are marked as the initial state set {x_T^(1), x_T^(2),..., x_T^(K)}.

[0078] Specifically, a countdown loop from t=T to t=0 is performed, and the target noise state vector is read at the beginning of each step iteration. Next, the time step t is converted into a sinusoidal position encoding vector and concatenated with the text vector. Then, the fused conditional vector and the initial state are extracted through a 4-layer down-sampling convolution block to extract multi-scale features, the text semantic information is aligned using a cross-attention mechanism, and the features are reconstructed through a 4-layer up-sampling transpose convolution block. Finally, a velocity field vector with the same dimension as the target noise state vector is output.

[0079] Next, the target velocity field vector is read, and the intensity coefficient and compensation coefficient corresponding to the target time step are obtained by querying the preset noise schedule table. Two independent Gaussian random number generators are called to generate random variables conforming to the N(0, 1) distribution. The score function is calculated according to the flow matching formula logp_t(x_t)=-x_t / t-((1-t) / t)v_θ, and the vector calculation is performed through the floating-point operation unit. First, the deterministic drift term Then, the random diffusion term σ1(t)ε1√Δt+σ2(t)ε2√Δt is calculated, and the target state is obtained by adding the drift term and the diffusion term to the target state.

[0080] Further, when the program detects that the time step counter t is zero, all state vectors (including the initial state x_T and the iteratively generated x_{T-1} to x_0) are extracted, and then the audio reconstruction module is called. The vector sequence containing different noise levels is input into the inverse short-time Fourier transform processing unit, the phase information is gradually restored through the overlap-add algorithm, and the frequency domain features are converted into time domain waveforms. Then, the pre-defined post-processing (such as amplitude calibration, resampling and normalization, etc.) is applied to the generated original waveform, the processed floating-point audio data is packaged, and an audio track is obtained.

[0081] Then, a floating-point vector buffer with dimensions [K, L] is created (K is the number of random seeds, and L is the length of the audio sampling points), and each random seed corresponding to the audio data memory address is traversed to combine multiple audio tracks corresponding to multiple random seeds into multiple groups of audio tracks, forming a structured data set for reinforcement learning training.

[0082] In the embodiments of the present application, multiple groups of independent audio tracks are sampled based on multi-source stochastic differential equations, and the core benefit is to provide rich voice generation options, ensure trajectory independence by introducing controllable randomness, meet the differentiated needs of voice style and rhythm in different scenarios, and also filter the optimal results through multi-version comparison to improve the flexibility of voice synthesis.

[0083] In the field of financial technology, this can be used to create more natural virtual assistants or automated customer service systems, making customer interaction more human and improving user satisfaction.

[0084] In the field of financial technology, it can be used to generate multiple sets of voice for introducing financial products for intelligent customer service, adapting to different customers' preferences for speech speed and professional term explanation.

[0085] In the field of medical health, multiple sets of voice versions can be generated for the same electronic medical record to allow doctors to choose the clearest pronunciation to reduce medical record checking errors

[0086] S4, calculating the strategy gradient modulation factor of the multiple groups of independent audio tracks by using a preset optimization strategy algorithm and the multi-step Markov decision function.

[0087] In the embodiments of the present application, the optimization strategy algorithm is an algorithm used in the field of reinforcement learning (relative policy optimization (GRPO) algorithm), which aims to improve the decision-making strategy of an agent in a given environment through an iterative process. This algorithm calculates the gradient of the policy and uses this gradient information to update the policy parameters in order to maximize the cumulative reward in the long term.

[0088] In the embodiments of the present application, the strategy gradient modulation factor is a key component of the optimization strategy algorithm, which represents the relative advantage or value of taking a certain action in a specific state. This factor is calculated by comparing the difference between the reward generated by the action and the average reward, and is used to guide the adjustment direction of the policy parameters.

[0089] In the embodiments of the present application, referring to Figure 4 As shown in the figure, the strategy gradient modulation factor of the multiple groups of independent audio tracks is calculated by using a preset optimization strategy algorithm and the multi-step Markov decision function, which includes:

[0090] S41, calculating the reward value of the multiple groups of independent audio tracks by using the reward function in the multi-step Markov decision function;

[0091] S42, calculate the average value of the reward value;

[0092] S43, calculate the strategy gradient modulation factor of each audio track according to the preset optimization strategy algorithm and the average value.

[0093] In detail, a plurality of generated independent audio tracks and their corresponding complete state-action sequences {s_t, a_t} are obtained, wherein each track corresponds to a final output audio x_0, a reward function R(x_0, c) in a multi-step Markov decision function is called to evaluate the quality of the final output audio x_0 of each track based on the original input text, and the reward value of each track is calculated.

[0094] Specifically, according to the core idea of the group relative policy optimization (GRPO) algorithm, the average reward value of the tracks corresponding to the training batch is first calculated. The advantage function value A of each track k is calculated, which is the reward value minus the average reward value. The advantage function value A is the factor used to modulate the policy gradient.

[0095] In the embodiment of the application, the strategy gradient modulation factor of the plurality of independent audio tracks is calculated by using the preset optimization strategy algorithm and the multi-step Markov decision function, which can significantly improve the performance and adaptability of the speech synthesis model. This method simulates different decision paths and evaluates their results, enabling the model to learn which actions to take in a particular state to produce high-quality speech output.

[0096] In the field of financial technology, this optimized speech synthesis technology can be used to develop more natural and credible virtual assistants, improve the efficiency of customer service and user satisfaction, for example, by providing more natural voice interaction to enhance the user experience of online transaction platforms.

[0097] In the field of medical health, optimized speech synthesis can help create more friendly and empathetic virtual nursing assistants to improve patient interaction with the medical system, such as providing medical advice and emotional support through natural language interaction.

[0098] S5, optimize the policy parameters of the speech synthesis model according to the strategy gradient modulation factor to obtain an optimized speech synthesis model.

[0099] In the embodiment of the application, the policy parameters are parameters used to define the probability distribution of the agent taking a specific action in a specific state in the process of reinforcement learning and Markov decision. In the context of the speech synthesis model, these parameters determine the probability of the model generating a specific audio output given the input text and the target audio state.

[0100] In the embodiment of the present application, the policy parameters of the speech synthesis model are optimized according to the policy gradient modulation factor to obtain an optimized speech synthesis model, comprising:

[0101] The historical policy parameters and the target policy parameters of the speech synthesis model are obtained, and the policy probability ratio of the historical policy parameters and the target policy parameters is calculated;

[0102] The target function value of the target policy parameters is calculated based on the policy gradient modulation factor and the policy probability ratio, and the target policy parameters are updated based on the target function value;

[0103] The updated target policy parameters are taken as the target policy parameters, and returned to the step of calculating the policy probability ratio of the historical policy parameters and the target policy parameters;

[0104] When the update times corresponding to the updated target policy parameters reach a preset number of times, the speech synthesis model corresponding to the updated target policy parameters is taken as the optimized speech synthesis model.

[0105] In detail, the historical policy parameters θ_old saved by the last training and the target policy parameters θ in the memory are obtained, and the following operations are performed on each state-action pair (s_t, a_t) in the batch data: forward propagation is performed by θ_old and θ respectively through model forward propagation, the state s_t is input into the policy network, the probability values p_old(a_t|s_t) and p_θ(a_t|s_t) of selecting the action a_t under the old parameters and the new parameters are calculated, and finally the policy probability ratio r_i^t(θ)=p_θ(a_t|s_t) / p_old(a_t|s_t) is directly calculated by division operation, which quantifies the probability change degree of the new and old policies to produce the action.

[0106] Specifically, the policy probability ratio is multiplied by the corresponding policy gradient modulation factor, and the r_i(θ) is clipped to the interval [1-ε, 1+ε] to obtain the clipped value, and min(r_i^t(θ)A_i^t, clipped_r_i^tA_i^t) is taken as the policy target item, and the KL divergence of the target policy and the historical policy is calculated as a regularization term. Then, the total target function value J(θ)=mean(min-βKL) is obtained by averaging G trajectories and T time steps. Finally, the gradient of J(θ) with respect to θ is calculated by automatic differentiation, and the Adam optimizer is used to update θ according to The parameters are updated by gradient ascent, where α is the learning rate.

[0107] Further, the updated target policy parameter is taken as a target policy parameter, and the step of calculating the policy probability ratio of the historical policy parameter and the target policy parameter is returned to, a gradient ascent update target parameter is performed, when the update number corresponding to the updated target policy parameter reaches a preset number, the cycle is immediately terminated, the target parameter theta in the memory is serialized and saved to a model file, and finally the file containing the updated parameter generated is the optimized speech synthesis model.

[0108] In the embodiment of the present application, the policy gradient modulation factor is used to optimize the policy parameter of the speech synthesis model, which can focus the model on the generation rule of high-quality audio tracks, enhance the learning of key features (such as professional term clarity and rhythm adaptability), and improve the accuracy and stability of synthesized speech. At the same time, the gradient modulation can accelerate the convergence of the model.

[0109] In the field of financial technology, the optimized speech synthesis model can be used in customer service systems to provide more natural and professional voice interaction, enhance customer trust, and improve service efficiency. For example, virtual assistants can provide financial consulting and transaction guidance with more natural speech.

[0110] In the field of financial technology, the speech broadcast of intelligent consultants can be optimized to make policy interpretation and income broadcast more accurate and reduce information transmission errors.

[0111] In the field of medical health, electronic medical record voice conversion can be made more in line with medical specifications, reduce term misreading, and improve the efficiency of doctor-patient communication.

[0112] S6, obtaining the text to be converted, and synthesizing the corresponding speech of the text to be converted by using the optimized speech synthesis model.

[0113] In the embodiment of the present application, the text to be converted is any form of written language, including sentences, paragraphs, articles or dialogues, and contains specific content to be converted into speech.

[0114] In the embodiment of the present application, the synthesis of the corresponding speech of the text to be converted by using the optimized speech synthesis model comprises:

[0115] Encoding the text to be converted to obtain a text feature vector;

[0116] Generating audio data of the text to be converted according to the text feature vector and a data sampling scheduling unit of the optimized speech synthesis model;

[0117] Packaging the audio data to obtain the corresponding speech of the text to be converted.

[0118] In detail, the to-be-converted text is preprocessed first, split by words, and each word is mapped to a unique digital identifier. Then, a pre-trained embedding layer is called to convert each word into a short numerical array according to the digital identifier query. The arrays are connected in the order of the text to integrate them into a fixed-length one-dimensional numerical array. This array is the text feature vector that can represent the semantic of the text.

[0119] Specifically, the computer first loads the optimized model parameters, randomly samples an initial noise vector from a standard Gaussian distribution, and then iterates for a preset number of time steps T. At each step t, the program inputs the target noise, time step encoding, and text feature vector into the velocity field network of the model to calculate the denoising direction v_θ(x_t,t,c). Then, the state is updated using the Euler method according to the formula x_{t-1}=x_t+v_θ(x_t,t,c)*Δt (where Δt=1 / T) to gradually convert the noise into speech data. Finally, the waveform data obtained at t=0 is the audio data. Then, the waveform data is PCM encoded and packaged into a standard WAV file header to output the final speech file.

[0120] In the embodiments of the present application, the to-be-converted text is obtained and synthesized into speech by using an optimized speech synthesis model. With the model optimization characteristics, the written content can be accurately converted into natural and clear speech, efficiently delivering information and improving the interaction experience and information reach efficiency.

[0121] In the field of financial technology, it can be used for intelligent customer service to synthesize real-time voice of financial product introduction, transaction reminders, etc., so that users can conveniently receive professional financial information.

[0122] In the field of medical health, the optimized speech synthesis model is used to convert electronic medical records and health popular science texts into speech, which can assist medical staff in quickly communicating medical conditions, help patients understand medical information, and optimize the medical service process and patient experience.

[0123] In the field of medical health, the optimized speech synthesis technology can help create a more friendly and easy-to-understand electronic health record system, allowing patients to easily access medical information or providing emotional support and health guidance through a virtual nursing assistant with natural language interaction, thereby improving patient satisfaction.

[0124] As can be seen, in the above scheme, by modularly splitting the model core unit and reconstructing it into a more flexible decision and generation mechanism, the accuracy, diversity, and controllability of speech synthesis can be significantly improved. The denoising logic is optimized by multi-step Markov decision to enhance speech clarity, the audio track diversity is expanded by multi-source stochastic differential equation to adapt to different scenarios, and the model parameter directional optimization is realized through policy gradient modulation, thereby improving the accuracy of the speech synthesis model in synthesizing speech.

[0125] It should be understood that the size of the serial number of each step in the above embodiments does not mean the order of execution, and the execution order of each process should be determined according to its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of the present application.

[0126] In an embodiment, a speech synthesis device based on an optimization strategy algorithm is provided, which corresponds to the speech synthesis method based on the optimization strategy algorithm in the above embodiment. As shown in the figure, the speech synthesis device based on the optimization strategy algorithm includes a unit extraction module 101, a unit mapping module 102, an audio track sampling module 103, a strategy gradient modulation factor calculation module 104, a speech synthesis model optimization module 105, and a speech synthesis module 106. The functions of each module are described in detail as follows: Figure 5

[0127] The unit extraction module 101 is configured to extract a time sequence processing network unit and a data sampling scheduling unit in a preset speech synthesis model.

[0128] The unit mapping module 102 is configured to map a denoising function in the time sequence processing network unit into a multi-step Markov decision function, and convert a common differential equation in the data sampling scheduling unit into a multi-source stochastic differential equation.

[0129] The audio track sampling module 103 is configured to sample a plurality of groups of independent audio tracks corresponding to a preset input text based on the multi-source stochastic differential equation.

[0130] The strategy gradient modulation factor calculation module 104 is configured to calculate strategy gradient modulation factors of the plurality of groups of independent audio tracks by using a preset optimization strategy algorithm and the multi-step Markov decision function.

[0131] The speech synthesis model optimization module 105 is configured to optimize strategy parameters of the speech synthesis model according to the strategy gradient modulation factors to obtain an optimized speech synthesis model.

[0132] The speech synthesis module 106 is configured to obtain a text to be converted, and synthesize speech corresponding to the text to be converted by using the optimized speech synthesis model.

[0133] In an embodiment, the unit mapping module 102, when mapping the denoising function in the time sequence processing network unit into the multi-step Markov decision function, is configured to:

[0134] discretize a denoising process of the denoising function into a plurality of time steps, and map noisy speech data in the denoising function into state data of Markov decision at each time step;

[0135] ​generate a speech estimation value according to the state data, and map the speech estimation value as action data of a Markov decision;

[0136] construct a multi-step Markov decision function based on the action data and the state data.

[0137] In an embodiment, the unit mapping module 102, when performing the conversion of the ordinary differential equation in the data sampling scheduling unit into a multi-source stochastic differential equation, is configured to:

[0138] obtain a velocity field in the ordinary differential equation, calculate a score function of the velocity field, and extract a time step in the velocity field;

[0139] calculate a total noise intensity and a weight angle corresponding to the time step according to a predefined noise intensity function, and respectively calculate intensity coefficients corresponding to the total noise intensity and the weight angle;

[0140] combine the velocity field, the score function, and the intensity coefficients into a multi-source stochastic differential equation.

[0141] In an embodiment, the audio track sampling module 103, when performing sampling of a plurality of groups of independent audio tracks corresponding to a preset input text based on the multi-source stochastic differential equation, is configured to:

[0142] encode the preset input text to obtain a text vector;

[0143] randomly sample an initial state of each preset random seed from a standard Gaussian distribution;

[0144] iterate a target state in the multi-source stochastic differential equation from the initial state according to a preset time step, and in each iteration, calculate a target velocity field of the input text according to the target state, the time step, and the text vector;

[0145] update the target state in the multi-source stochastic differential equation according to the target velocity field;

[0146] when the iteration reaches the time step of zero, aggregate the initial state and all target states into an audio track of the input text;

[0147] combine the audio track into a plurality of groups of audio tracks corresponding to the input text.

[0148] In an embodiment, the policy gradient modulation factor calculation module 104, when performing calculation of a policy gradient modulation factor of the plurality of groups of independent audio tracks by using a preset optimization policy algorithm and the multi-step Markov decision function, is configured to:

[0149] calculating a reward value of the multiple sets of independent audio tracks by using a reward function in the multiple-step Markov decision function;

[0150] calculating an average value of the reward values;

[0151] calculating a policy gradient modulation factor of each audio track according to a preset optimization strategy algorithm and the average value.

[0152] In an embodiment, the speech synthesis model optimization module 105, when performing optimization of policy parameters of the speech synthesis model according to the policy gradient modulation factor to obtain an optimized speech synthesis model, is configured to:

[0153] obtaining historical policy parameters and target policy parameters of the speech synthesis model, and calculating a policy probability ratio of the historical policy parameters and the target policy parameters;

[0154] calculating a target function value of the target policy parameters based on the policy gradient modulation factor and the policy probability ratio, and updating the target policy parameters based on the target function value;

[0155] taking the updated target policy parameters as target policy parameters, and returning to the step of calculating the policy probability ratio of the historical policy parameters and the target policy parameters;

[0156] when the number of updates corresponding to the updated target policy parameters reaches a preset number of times, taking the speech synthesis model corresponding to the updated target policy parameters as the optimized speech synthesis model.

[0157] In an embodiment, the speech synthesis module 106, when performing synthesis of speech corresponding to the text to be converted by using the optimized speech synthesis model, is configured to:

[0158] encoding the text to be converted to obtain a text feature vector;

[0159] generating audio data of the text to be converted according to the text feature vector and a data sampling scheduling unit of the optimized speech synthesis model;

[0160] packaging the audio data to obtain the speech corresponding to the text to be converted.

[0161] The application provides a speech synthesis device based on an optimization strategy algorithm, which can significantly improve the accuracy, diversity and controllability of speech synthesis by modularizing and splitting the model core unit and reconstructing a more flexible decision and generation mechanism. The speech synthesis device enhances speech clarity by using a multi-step Markov decision optimization denoising logic, expands audio track diversity by using a multi-source stochastic differential equation to adapt to different scenes, and realizes directional optimization of model parameters through policy gradient modulation, thereby improving the accuracy of synthesized speech of the speech synthesis model.

[0162] For the specific definition of the speech synthesis device based on the optimization strategy algorithm, please refer to the definition of the speech synthesis method based on the optimization strategy algorithm above, which will not be repeated here. The various modules in the above-mentioned speech synthesis device based on the optimization strategy algorithm can be implemented in whole or in part by software, hardware and a combination thereof. The above-mentioned modules can be embedded in or independent of the processor in the computer device in the form of hardware, or can be stored in the memory of the computer device in the form of software, so that the processor can call and execute the operations corresponding to the above modules.

[0163] In one embodiment, a computer device is provided. The computer device may be a server, and its internal structure diagram may be as follows: Figure 6 As shown. The computer device includes a processor, a memory, a network interface and a database connected via a system bus. The processor of the computer device is used to provide computing and control capabilities. The memory of the computer device includes a non-volatile and / or volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system, a computer program and a database. The internal memory provides an environment for the operation of the operating system and the computer program in the non-volatile storage medium. The network interface of the computer device is used to communicate with an external client via a network connection. When the computer program is executed by the processor, it implements the functions or steps on the server side of a speech synthesis method based on an optimization strategy algorithm.

[0164] In one embodiment, a computer device is provided. The computer device may be a client, and its internal structure diagram may be as follows: Figure 7 As shown. The computer device includes a processor, memory, network interface, display screen, and input device connected via a system bus. The processor of the computer device is used to provide computing and control capabilities. The memory of the computer device includes a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system and a computer program. The internal memory provides an environment for the operation of the operating system and computer program in the non-volatile storage medium. The network interface of the computer device is used to communicate with an external server via a network connection. When the computer program is executed by the processor, it implements the functions or steps on the client side of a speech synthesis method based on an optimization strategy algorithm.

[0165] In one embodiment, a computer device is provided, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, the following steps are performed:

[0166] Extracting the timing processing network unit and the data sampling scheduling unit in the preset speech synthesis model;

[0167] mapping a de-noising function in the time series processing network unit as a multi-step Markov decision function, and converting a common differential equation in the data sampling scheduling unit as a multi-source stochastic differential equation;

[0168] sampling a plurality of groups of independent audio tracks corresponding to a preset input text based on the multi-source stochastic differential equation;

[0169] calculating a strategy gradient modulation factor of the plurality of groups of independent audio tracks by using a preset optimization strategy algorithm and the multi-step Markov decision function;

[0170] optimizing a strategy parameter of the speech synthesis model according to the strategy gradient modulation factor to obtain an optimized speech synthesis model;

[0171] obtaining a text to be converted, and synthesizing a speech corresponding to the text to be converted by using the optimized speech synthesis model.

[0172] In one embodiment, a computer readable storage medium is provided, and a computer program is stored on the computer readable storage medium. The computer program is executed by a processor to implement the following steps:

[0173] extracting a time series processing network unit and a data sampling scheduling unit in a preset speech synthesis model;

[0174] mapping a de-noising function in the time series processing network unit as a multi-step Markov decision function, and converting a common differential equation in the data sampling scheduling unit as a multi-source stochastic differential equation;

[0175] sampling a plurality of groups of independent audio tracks corresponding to a preset input text based on the multi-source stochastic differential equation;

[0176] calculating a strategy gradient modulation factor of the plurality of groups of independent audio tracks by using a preset optimization strategy algorithm and the multi-step Markov decision function;

[0177] optimizing a strategy parameter of the speech synthesis model according to the strategy gradient modulation factor to obtain an optimized speech synthesis model;

[0178] obtaining a text to be converted, and synthesizing a speech corresponding to the text to be converted by using the optimized speech synthesis model.

[0179] It should be noted that the functions or steps that the computer readable storage medium or the computer device can implement are described above with reference to the related descriptions of the server side and the client side in the foregoing method embodiments. To avoid repetition, they will not be described again here.

[0180] Those skilled in the art can understand that all or part of the processes in the above-mentioned embodiment methods can be completed by instructing the relevant hardware through a computer program. The computer program can be stored in a non-volatile computer readable storage medium, and when the computer program is executed, the processes of the above-mentioned embodiments of the methods can be included. Any reference to memory, storage, database or other medium used in the embodiments provided in the present application can include non-volatile and / or volatile memory. Non-volatile memory can include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM) or flash memory. Volatile memory can include random access memory (RAM) or external cache memory. As an illustration but not limitation, RAM is available in various forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), double data rate SDRAM (DDR SDRAM), enhanced SDRAM (ESDRAM), synchronous link (Synchlink) DRAM (SLDRAM), memory bus (Rambus) direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM), etc.

[0181] Those skilled in the art can clearly understand that, for the convenience and brevity of description, only the above-mentioned division of functional units and modules is exemplified, and in actual application, the above-mentioned functions can be completed by different functional units and modules according to needs, that is, the internal structure of the device is divided into different functional units or modules to complete all or part of the functions described above.

[0182] It should be noted that if non-company software tools or components appear in the embodiments of the present application, they are only used for example introduction and do not represent actual use.

[0183] The above-mentioned embodiments are only used to illustrate the technical solutions of the present application, and not to limit them; although the present application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that the technical solutions recorded in the foregoing embodiments can be modified, or some technical features can be replaced by equivalents; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the spirit and scope of the technical solutions of the embodiments of the present application, and should be included in the protection scope of the present application.

Claims

1. A speech synthesis method based on an optimization strategy algorithm, characterized in that: include: Extracting the timing processing network unit and the data sampling scheduling unit in the preset speech synthesis model; Mapping the denoising function in the time series processing network unit into a multi-step Markov decision function, and converting the ordinary differential equation in the data sampling scheduling unit into a multi-source stochastic differential equation; Sampling a plurality of independent audio tracks corresponding to preset input texts based on the multi-source stochastic differential equation; Calculating the policy gradient modulation factors of the multiple groups of independent audio tracks using a preset optimization strategy algorithm and the multi-step Markov decision function; Optimizing the strategy parameters of the speech synthesis model according to the strategy gradient modulation factor to obtain an optimized speech synthesis model; The text to be converted is obtained, and the speech corresponding to the text to be converted is synthesized using the optimized speech synthesis model.

2. The speech synthesis method based on the optimization strategy algorithm according to claim 1, characterized in that: Mapping the denoising function in the time series processing network unit to a multi-step Markov decision function includes: Discretizing the denoising process of the denoising function into a plurality of time steps, and mapping the noisy speech data in the denoising function into state data of a Markov decision in each time step; generating a speech estimation value according to the state data, and mapping the speech estimation value into action data of a Markov decision; A multi-step Markov decision function is constructed based on the action data and the state data.

3. The speech synthesis method based on the optimization strategy algorithm according to claim 1, characterized in that: The converting the ordinary differential equation in the data sampling scheduling unit into a multi-source stochastic differential equation includes: Obtaining a velocity field in the ordinary differential equation, calculating a score function of the velocity field, and extracting a time step in the velocity field; Calculating the total noise intensity and the weight angle corresponding to the time step according to a predefined noise intensity function, and respectively calculating the intensity coefficients corresponding to the total noise intensity and the weight angle; The velocity field, the score function, and the intensity coefficient are combined as a multi-source stochastic differential equation.

4. The speech synthesis method based on the optimization strategy algorithm according to claim 1, characterized in that: The plurality of independent audio tracks corresponding to the preset input texts based on the multi-source stochastic differential equation sampling include: Encode the preset input text to obtain a text vector; Generate the initial state of each preset random seed by random sampling from a standard Gaussian distribution; Iterating the target state in the multi-source stochastic differential equation from the initial state according to a preset time step, and in each iteration step, calculating the target velocity field of the input text according to the target state, the time step and the text vector; updating a target state in the multi-source stochastic differential equation according to the target velocity field; When the iteration reaches the time step zero, aggregating the initial state and all target states into an audio track of the input text; The audio tracks are combined into a plurality of groups of audio tracks corresponding to the input text.

5. The speech synthesis method based on the optimization strategy algorithm according to claim 1, wherein: The calculating of the policy gradient modulation factors of the multiple groups of independent audio tracks by using a preset optimization strategy algorithm and the multi-step Markov decision function includes: Calculating reward values ​​for the multiple groups of independent audio tracks using a reward function in the multi-step Markov decision function; Calculating an average of the reward values; The strategy gradient modulation factor of each audio track is calculated according to the preset optimization strategy algorithm and the average value.

6. The speech synthesis method based on the optimization strategy algorithm according to claim 1, characterized in that: Optimizing the strategy parameters of the speech synthesis model according to the strategy gradient modulation factor to obtain an optimized speech synthesis model includes: Obtaining historical strategy parameters and target strategy parameters of the speech synthesis model, and calculating a strategy probability ratio between the historical strategy parameters and the target strategy parameters; Calculating an objective function value of a target policy parameter based on the policy gradient modulation factor and the policy probability ratio, and updating the target policy parameter based on the objective function value; Using the updated target policy parameters as the target policy parameters, and returning to the step of calculating the policy probability ratio between the historical policy parameters and the target policy parameters; When the number of updates corresponding to the updated target strategy parameters reaches a preset number, the speech synthesis model corresponding to the updated target strategy parameters is used as the optimized speech synthesis model.

7. The speech synthesis method based on the optimization strategy algorithm according to claim 1, characterized in that: The step of synthesizing speech corresponding to the text to be converted by using the optimized speech synthesis model includes: Encoding the text to be converted to obtain a text feature vector; Generate audio data of the text to be converted according to the text feature vector and the data sampling scheduling unit of the optimized speech synthesis model; The audio data is encapsulated to obtain the speech corresponding to the text to be converted.

8. A speech synthesis device based on an optimization strategy algorithm, characterized in that: include: A unit extraction module is used to extract the timing processing network unit and the data sampling scheduling unit in the preset speech synthesis model; A unit mapping module, configured to map the denoising function in the time series processing network unit into a multi-step Markov decision function, and to convert the ordinary differential equation in the data sampling scheduling unit into a multi-source stochastic differential equation; An audio track sampling module, configured to sample a plurality of independent audio tracks corresponding to preset input texts based on the multi-source stochastic differential equation; A policy gradient modulation factor calculation module, configured to calculate the policy gradient modulation factors of the plurality of independent audio tracks using a preset optimization strategy algorithm and the multi-step Markov decision function; A speech synthesis model optimization module, configured to optimize the strategy parameters of the speech synthesis model according to the strategy gradient modulation factor to obtain an optimized speech synthesis model; The speech synthesis module is used to obtain the text to be converted and synthesize the speech corresponding to the text to be converted using the optimized speech synthesis model.

9. A computer 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 computer program, the speech synthesis method based on the optimization strategy algorithm according to any one of claims 1 to 7 is implemented.

10. A computer-readable storage medium storing a computer program, characterized in that: When the computer program is executed by a processor, the speech synthesis method based on the optimization strategy algorithm according to any one of claims 1 to 7 is implemented.

Images