Speech synthesis method and device, computer device and storage medium

Abstract

The application relates to the field of audio conversion and discloses a speech synthesis method and device, computer equipment and a storage medium, the method comprising the following steps: obtaining vocoder input data and a previous moment hidden state, wherein the vocoder input data comprises a mel spectrum of previous moment speech and current speech; processing the previous moment hidden state and the vocoder input data through a self-recurrence encoder to generate a current hidden state of the current speech; processing the current hidden state through a full connection layer to generate a connection layer function; processing the connection layer function through a binary tree classifier to generate a category of the current speech, so as to generate a current synthesized speech according to the category of the current speech. The binary tree classifier is used to generate the category of the speech, the number of calculation times of the speech category can be greatly reduced, the dimension of a weight matrix can be greatly reduced, and therefore the calculation complexity of the vocoder can be greatly reduced.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of audio conversion, and in particular to a speech synthesis method and device, a computer device and a storage medium. BACKGROUND

[0002] With the development of deep learning technology, it is feasible to synthesize high-quality speech from low-bit-rate audio using a speech synthesis vocoder algorithm based on deep learning technology. However, this vocoder algorithm relies on the floating-point operation of a high-performance processor at hundreds of billions of times per second. That is, the computational complexity of the vocoder algorithm is high.

[0003] Due to factors such as battery performance, mobile terminals cannot be equipped with high-performance processors. Therefore, the high computational complexity limits the application of the speech synthesis vocoder algorithm in mobile terminals. SUMMARY

[0004] Therefore, it is necessary to provide a speech synthesis method, device, computer device and storage medium to reduce the computational complexity of the vocoder algorithm in order to solve the above technical problems.

[0005] A speech synthesis method comprises:

[0006] obtaining vocoder input data and a previous hidden state, the vocoder input data comprising a mel-spectrogram of a previous speech and a current speech;

[0007] processing the previous hidden state and the vocoder input data through an autoregressive encoder to generate a current hidden state of the current speech;

[0008] processing the current hidden state through a fully connected layer to generate a connection layer function;

[0009] processing the connection layer function through a binary tree classifier to generate a class of the current speech, so as to generate a current synthesized speech according to the class of the current speech.

[0010] A speech synthesis device comprises:

[0011] a data acquisition module configured to obtain vocoder input data and a previous hidden state, the vocoder input data comprising a mel-spectrogram of a previous speech and a current speech;

[0012] a hidden state module configured to process the previous hidden state and the vocoder input data through an autoregressive encoder to generate a current hidden state of the current speech;

[0013] a connection layer module configured to process the current hidden state through a fully connected layer to generate a connection layer function;

[0014] a voice category module, configured to process the connection layer function by a binary tree classifier to generate a category of the current voice, and to generate a current synthesized voice according to the category of the current voice.

[0015] A computer device includes a memory, a processor, and computer readable instructions stored in the memory and executable on the processor, and the processor implements the voice synthesis method when executing the computer readable instructions.

[0016] One or more readable storage media storing computer readable instructions, and the computer readable instructions are executed by one or more processors to make the one or more processors execute the voice synthesis method.

[0017] The voice synthesis method, device, computer device and storage medium, first obtain the current hidden state of the current voice through the autoregressive encoder, and then process through two full connection layers and a binary tree classifier to obtain the category of the current voice. Since the binary tree classifier used can greatly reduce the number of voice category calculations and greatly reduce the dimension of the weight matrix, the computational complexity of the vocoder can be greatly reduced. And through subjective listening test, the synthesized voice by the present application does not reduce the quality of the synthesized voice. BRIEF DESCRIPTION OF DRAWINGS

[0018] 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 for those skilled in the art, other drawings can also be obtained without creative labor.

[0019] Figure 1 is a flowchart of a voice synthesis method in an embodiment of the present application;

[0020] Figure 2 is a structural schematic diagram of a voice synthesis device in an embodiment of the present application;

[0021] Figure 3 is a schematic diagram of a computer device in an embodiment of the present application. DETAILED DESCRIPTION

[0022] The technical solutions in the embodiments of the present application will be described clearly and completely in the following with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only some embodiments of the present application, not all embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor are within the scope of protection of the present application.

[0023] In an embodiment, as shown in FIG. 1, a speech synthesis method is provided, including steps S10-S40. Figure 1

[0024] S10, obtain vocoder input data and last moment hidden state, the vocoder input data including the mel spectrum of the last moment speech and the current speech.

[0025] Understandably, in the speech synthesis, the vocoder can be selected according to actual needs, such as an autoregressive vocoder. Speech refers to various audio. In an example, the autoregressive vocoder can be WaveRNN (a neural network-based vocoder). Vocoder input data refers to data used to input the vocoder. Here, the vocoder input data includes the mel spectrum of the last moment speech and the current speech. f The last moment hidden state refers to the hidden state of the last moment speech. Hidden state refers to the output result of GRU (Gated Recurrent Unit). The vocoder input data can be expressed as: Mel spectrum, i.e., mel frequency spectrum. The mel spectrum can be obtained by processing the spectrogram of the current speech by a mel scale filter bank of a specified specification.

[0026] S20, process the last moment hidden state and the vocoder input data by an autoregressive encoder to generate the current hidden state of the current speech.

[0027] S30, process the current hidden state by a fully connected layer to generate a connection layer function.

[0028] S40, process the connection layer function by a binary tree classifier to generate the category of the current speech, so as to generate the current synthesized speech according to the category of the current speech.

[0029] Understandably, the vocoder includes an autoregressive encoder, a fully connected layer, and a binary tree classifier. Here, the autoregressive encoder can adopt GRU (Gated Recurrent Unit). The autoregressive encoder can output the current hidden state of the current speech. The current hidden state is the output parameter of GRU and is the input data of the fully connected layer.

[0030] The current hidden state is processed by two layers of fully connected layers and a binary tree classifier to obtain the category of the current speech. The binary tree classifier can be an 8-layer binary tree classifier constructed based on a sigmoid function.

[0031] Speech has a value range of ​In computers, 16 bits are usually used for quantization, i.e. 16-bit binary representation. Then the speech In There are 65536 (2^16) categories in total. Usually, the speech is compressed into 256 categories by using the u-law method, specifically:

[0032]

[0033] wherein, is the compressed result of the speech ;

[0034] is an operator of a sign function;

[0035] u = 255.

[0036] That is, the speech can be represented by 256 categories.

[0037] The sigmoid function can be used to construct an 8-layer binary tree classifier. The connection layer function is input into the binary tree classifier, and hierarchical calculation is performed. Each node uses the sigmoid function to calculate the category belonging to the next layer. When the node output value is less than 0.5, the category is the left node of the next layer, and vice versa. Through the binary tree classifier, only 8 times of calculation is required, i.e. 256 categories of the speech are calculated. Moreover, when the binary tree classifier is used, the dimension of the weight matrix in the connection layer function is greatly reduced from the original 256 dimensions to 8 dimensions, greatly reducing the calculation complexity of the vocoder.

[0038] After all categories of the current speech are calculated, i.e. 256-bit compressed speech is obtained, then the u-law method is used to decompress the compressed speech, i.e. 16-bit current synthesized speech is obtained.

[0039] In this embodiment, the current hidden state of the current speech is obtained through the autoregressive encoder, and then processed through two fully connected layers and a binary tree classifier to obtain the category of the current speech. Since the binary tree classifier used can greatly reduce the number of speech category calculations and greatly reduce the dimension of the weight matrix, the calculation complexity of the vocoder can be greatly reduced. Moreover, through subjective listening tests, the synthesized speech obtained by this embodiment does not reduce the quality of the synthesized speech.

[0040] Optionally, step S20, i.e. processing the previous hidden state and the vocoder input data through the autoregressive encoder to generate the current hidden state of the current speech, comprises:

[0041] S201, processing the last moment hidden state and the current vocoder input data by a first encoder to generate a first intermediate quantity; processing the last moment hidden state and the current vocoder input data by a second encoder to generate a second intermediate quantity.

[0042] In some examples, the first encoder comprises:

[0043]

[0044] wherein, is the first intermediate quantity;

[0045] is an operator of a sigmoid function;

[0046] is a weight matrix in the first encoder for processing the last moment hidden state;

[0047] is the last moment hidden state;

[0048] is a weight matrix in the first encoder for processing the vocoder input data;

[0049] is the vocoder input data.

[0050] Here, the first encoder is used to generate the first intermediate quantity . The first intermediate quantity is equivalent to the output value of the reset gate in the GRU. The weight matrix , is the result of the training of the vocoder. The weight matrix is used to input the last moment hidden state. The weight matrix is used to input the vocoder input data.

[0051] In some examples, the second encoder comprises:

[0052]

[0053] wherein, is the second intermediate quantity;

[0054] is an operator of a sigmoid function;

[0055] is a weight matrix in the second encoder for processing the last moment hidden state;

[0056] is the last moment hidden state;

[0057] is a weight matrix for processing the vocoder input data in the second encoder;

[0058] is the vocoder input data.

[0059] Here, the second encoder is used to generate a second intermediate quantity . The second intermediate quantity is equivalent to the output value of the update gate in the GRU. The weight matrix , is the result of the training of the vocoder. The weight matrix is used to input the previous time hidden state. The weight matrix is used to input the vocoder input data.

[0060] S202, processing the second intermediate quantity, the previous time hidden state and the current vocoder input data by a third encoder to generate a third intermediate quantity.

[0061] In some examples, the third encoder comprises:

[0062]

[0063] wherein, is the third intermediate quantity;

[0064] is an operator of the hyperbolic tangent function;

[0065] is the second intermediate quantity;

[0066] is a matrix point multiplication operator;

[0067] is a weight matrix for processing the previous time hidden state in the third encoder;

[0068] is the previous time hidden state;

[0069] is a weight matrix for processing the vocoder input data in the third encoder;

[0070] is the vocoder input data.

[0071] Here, the third encoder is used to generate a third intermediate quantity . The third intermediate quantity can be used to calculate the hidden state. The weight matrix , is the result of the training of the vocoder. The weight matrix is used to input the previous hidden state. The weight matrix is used to input the vocoder input data.

[0072] S203, processing the first intermediate quantity, the third intermediate quantity and the previous hidden state by the hidden state encoder to generate the current hidden state.

[0073] In some examples, the hidden state encoder comprises:

[0074]

[0075] wherein, is the current hidden state of the current speech;

[0076] is the first intermediate quantity;

[0077] is a matrix point multiplication calculation symbol;

[0078] is the previous hidden state;

[0079] is the third intermediate quantity.

[0080] Here, the hidden state encoder is used to generate the current hidden state . The current hidden state may be input data of a fully connected layer. The hidden state is affected by the previous hidden state .

[0081] Optionally, the connection layer function comprises:

[0082]

[0083] wherein, is the connection layer function;

[0084] is the current speech;

[0085] is an operator of an activation function;

[0086] is a first weight matrix of the fully connected layer;

[0087] is a second weight matrix of the fully connected layer;

[0088] a current hidden state of the current speech.

[0089] It can be understood that the connection layer function may be input data of the binary tree classifier. The full connection layer here includes two layers, each having a respective weight matrix. When the number of layers of the binary tree classifier is 8, the dimensions of the first weight matrix and the second weight matrix can be 8.

[0090] 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.

[0091] In an embodiment, a speech synthesis device is provided, which corresponds to the speech synthesis method in the above embodiments. As shown in Figure 2 , the speech synthesis device includes a data acquisition module 10, a hidden state module 20, a connection layer module 30, and a speech category module 40. The functions of each module are described in detail as follows:

[0092] The data acquisition module 10 is configured to acquire vocoder input data and a previous hidden state, wherein the vocoder input data includes a mel spectrum of a previous speech and a current speech.

[0093] The hidden state module 20 is configured to generate a current hidden state of the current speech by processing the previous hidden state and the vocoder input data through an autoregressive encoder.

[0094] The connection layer module 30 is configured to generate a connection layer function by processing the current hidden state through a full connection layer.

[0095] The speech category module 40 is configured to generate a category of the current speech by processing the connection layer function through a binary tree classifier, and to generate a current synthesized speech according to the category of the current speech.

[0096] It can be understood that when performing speech synthesis, a vocoder can be selected according to actual needs, such as an autoregressive vocoder. Speech refers to various audio. In an example, the autoregressive vocoder can be WaveRNN (a neural network-based vocoder). Vocoder input data refers to data used to input a vocoder. Here, the vocoder input data includes a mel spectrum of a previous speech and a current speech. f The previous hidden state refers to a hidden state of a previous speech. The current hidden state refers to the output result of a GRU (Gated Recurrent Unit). The vocoder input data may be represented as: Mel-spectrogram, i.e., Mel-spectrum. The Mel-spectrum can be obtained by processing the spectrogram of the current speech with a Mel-scale filter bank of specified specification.

[0097] The vocoder includes an autoregressive encoder, a fully connected layer, and a binary tree classifier. Here, the autoregressive encoder can employ a GRU (Gated Recurrent Unit). The autoregressive encoder can output a current hidden state of the current speech. The current hidden state is an output parameter of the GRU and is input data of the fully connected layer.

[0098] The current hidden state is processed by two layers of the fully connected layer and the binary tree classifier to obtain a class of the current speech. The binary tree classifier can be an 8-layer binary tree classifier constructed based on a sigmoid function.

[0099] Speech The range is In a computer, 16 bits are often used for quantization, i.e., 16-bit binary representation. Then the speech has 65536 (2^16) classes between . Generally, the speech is compressed to 256 classes by using a u-law method, specifically:

[0100]

[0101] wherein, is the compression result of the speech ;

[0102] is an operator of a sign function;

[0103] u = 255.

[0104] That is, the speech can be represented by 256 classes.

[0105] An 8-layer binary tree classifier can be constructed using a sigmoid function. The connection layer function is input into the binary tree classifier, and hierarchical calculation is performed. Each node calculates the class belonging to the next layer using a sigmoid function. When the node output value is less than 0.5, the class is the left node of the next layer, and vice versa. Through the binary tree classifier, only 8 times of calculation is required, i.e., 256 classes of the speech are calculated. Moreover, when the binary tree classifier is used, the dimension of the weight matrix in the connection layer function is greatly reduced from the original 256 dimensions to 8 dimensions, greatly reducing the calculation complexity of the vocoder.

[0106] After all the categories of the current speech are calculated, 256-bit compressed speech is obtained, and then the compressed speech is decompressed by using the u-law method, so that 16-bit current synthesized speech is obtained.

[0107] In this embodiment, the current hidden state of the current speech is obtained through the autoregressive encoder, and then processed through two full connection layers and a binary tree classifier to obtain the category of the current speech. Since the binary tree classifier used can greatly reduce the number of calculations of the speech category and greatly reduce the dimension of the weight matrix, the computational complexity of the vocoder can be greatly reduced. Moreover, through subjective listening tests, the quality of the synthesized speech obtained by this embodiment has not been reduced.

[0108] Optionally, the hidden state module 20 comprises:

[0109] The first intermediate quantity unit is configured to generate a first intermediate quantity by processing the previous hidden state and the current vocoder input data through a first encoder.

[0110] The second intermediate quantity unit is configured to generate a second intermediate quantity by processing the previous hidden state and the current vocoder input data through a second encoder.

[0111] The third intermediate quantity unit is configured to generate a third intermediate quantity by processing the second intermediate quantity, the previous hidden state and the current vocoder input data through a third encoder.

[0112] The hidden state generating unit is configured to generate the current hidden state by processing the first intermediate quantity, the third intermediate quantity and the previous hidden state through a hidden state encoder.

[0113] Optionally, the first encoder comprises:

[0114]

[0115] wherein, is the first intermediate quantity;

[0116] is an operator of a sigmoid function;

[0117] is a weight matrix in the first encoder for processing the previous hidden state;

[0118] is the previous hidden state;

[0119] is a weight matrix in the first encoder for processing the vocoder input data;

[0120] input data for the vocoder.

[0121] Optionally, the second encoder comprises:

[0122]

[0123] wherein, is the second intermediate quantity;

[0124] is an operator of sigmoid function;

[0125] is a weight matrix in the second encoder for processing the previous time step hidden state;

[0126] is the previous time step hidden state;

[0127] is a weight matrix in the second encoder for processing the input data for the vocoder;

[0128] is the input data for the vocoder.

[0129] Optionally, the third encoder comprises:

[0130]

[0131] wherein, is the third intermediate quantity;

[0132] is an operator of hyperbolic tangent function;

[0133] is the second intermediate quantity;

[0134] is a matrix dot product operator;

[0135] is a weight matrix in the third encoder for processing the previous time step hidden state;

[0136] is the previous time step hidden state;

[0137] is a weight matrix in the third encoder for processing the input data for the vocoder;

[0138] is the input data for the vocoder.

[0139] Optionally, the hidden state encoder comprises:

[0140]

[0141] wherein, is a current hidden state of the current speech;

[0142] is the first intermediate quantity;

[0143] is a matrix point multiplication calculation symbol;

[0144] is the last hidden state;

[0145] is the third intermediate quantity.

[0146] Optionally, the connection layer function comprises:

[0147]

[0148] wherein, is the connection layer function;

[0149] is the current speech;

[0150] is an operator of an activation function;

[0151] is a first weight matrix of the full connection layer;

[0152] is a second weight matrix of the full connection layer;

[0153] is a current hidden state of the current speech.

[0154] The specific limitations on the speech synthesis device can refer to the limitations on the speech synthesis method in the foregoing, which will not be described herein. Each module in the speech synthesis device described above can be realized by software, hardware and a combination thereof in whole or in part. Each module described above can be embedded in or independent of a processor in a computer device in hardware form, or can be stored in a memory in a computer device in software form, so as to be called and executed by a processor to perform the operations corresponding to each module.

[0155] In an embodiment, a computer device is provided, which can be a terminal, and an internal structure diagram of the computer device can be as shown in FIG. 6. Figure 3As shown in the figure. The computer device includes a processor, a memory, a network interface, a display screen and an input device connected through a system bus. Among them, the processor of the computer device is used to provide computing and control capabilities. The memory of the computer device includes a readable storage medium, an internal memory. The non-volatile storage medium stores an operating system and computer readable instructions. The internal memory provides an environment for the operation of the operating system and computer readable instructions in the readable storage medium. The network interface of the computer device is used to communicate with the external server through the network connection. The computer readable instructions are executed by the processor to implement a speech synthesis method. The readable storage medium provided by the embodiment includes a non-volatile readable storage medium and a volatile readable storage medium.

[0156] In one embodiment, a computer device is provided, comprising a memory, a processor and computer readable instructions stored on the memory and executable on the processor, the processor executing the computer readable instructions to implement the following steps:

[0157] Obtaining vocoder input data and a previous hidden state, the vocoder input data comprising a mel spectrum of a previous speech and a current speech;

[0158] Processing the previous hidden state and the vocoder input data through an autoregressive encoder to generate a current hidden state of the current speech;

[0159] Processing the current hidden state through a fully connected layer to generate a connection layer function;

[0160] Processing the connection layer function through a binary tree classifier to generate a category of the current speech, to generate a current synthesized speech according to the category of the current speech.

[0161] In one embodiment, one or more computer readable storage media having computer readable instructions stored thereon are provided. The readable storage medium provided by the embodiment includes a non-volatile readable storage medium and a volatile readable storage medium. The computer readable instructions are stored on the readable storage medium, and the computer readable instructions are executed by one or more processors to implement the following steps:

[0162] Obtaining vocoder input data and a previous hidden state, the vocoder input data comprising a mel spectrum of a previous speech and a current speech;

[0163] Processing the previous hidden state and the vocoder input data through an autoregressive encoder to generate a current hidden state of the current speech;

[0164] Processing the current hidden state through a fully connected layer to generate a connection layer function;

[0165] The connection layer function is processed by a binary tree classifier to generate a category of the current speech, so as to generate a current synthesized speech according to the category of the current speech.

[0166] 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 relevant hardware through computer readable instructions, and the computer readable instructions can be stored in a non-volatile readable storage medium or a volatile readable storage medium. When the computer readable instructions are 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 each embodiment provided by 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).

[0167] Those skilled in the art can clearly understand that, for the convenience and brevity of description, only the division of the above-mentioned 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.

[0168] 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 they can modify the technical solutions recorded in the foregoing embodiments, or make equivalent replacement for part of the technical features; 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 characterized by, The method comprises: obtaining vocoder input data and a previous moment hidden state, wherein the vocoder input data comprises a mel spectrum of a previous moment speech and a current speech; processing the previous moment hidden state and the vocoder input data through an autoregressive encoder to generate a current hidden state of the current speech; processing the current hidden state through a fully connected layer to generate a connection layer function; processing the connection layer function through a binary tree classifier to generate a category of the current speech, so as to generate a current synthesized speech according to the category of the current speech; wherein the category of the current speech refers to a category obtained by compressing the current speech by using a u-law method, and specifically comprises: wherein is the current speech, speech has a value range of and is represented using 16 bits in binary. for speech compression results; is the operator of the sign function; =255。 2. The speech synthesis method of claim 1, wherein, the processing of the previous moment hidden state and the vocoder input data through the autoregressive encoder to generate the current hidden state of the current speech comprises: processing the previous moment hidden state and the vocoder input data through a first encoder to generate a first intermediate quantity, and processing the previous moment hidden state and the vocoder input data through a second encoder to generate a second intermediate quantity; processing the second intermediate quantity, the previous moment hidden state and the vocoder input data through a third encoder to generate a third intermediate quantity; processing the first intermediate quantity, the third intermediate quantity and the previous moment hidden state through a hidden state encoder to generate the current hidden state.

3. The speech synthesis method of claim 2, wherein, the first encoder comprises: wherein is the first intermediate quantity; sigmoid is an operator that is a sigmoid function; a weight matrix used in the first encoder for processing the previous time hidden state; the last moment hidden state; a weight matrix for use in the first encoder for processing the vocoder input data; Input data for the vocoder.

4. The speech synthesis method of claim 2, wherein, the second encoder comprises: wherein is the second intermediate quantity; sigmoid is an operator for sigmoid function; a weight matrix for the second encoder for processing the previous time hidden state; the last moment hidden state; a weight matrix for use in the second encoder for processing the vocoder input data; Input data for the vocoder.

5. The speech synthesis method of claim 2, wherein, the third encoder comprises: wherein is the third intermediate quantity; is the operator of the hyperbolic tangent function; is the second intermediate quantity; Compute the sign of the matrix dot product; a weight matrix for the third encoder for processing the previous time hidden state; the last moment hidden state; a weight matrix for the third encoder for processing the vocoder input data; Input data for the vocoder.

6. The speech synthesis method of claim 2, wherein, the hidden state encoder comprises: wherein, is the current latent state for the current speech; is the first intermediate quantity; Compute the sign of the matrix dot product; the last moment hidden state; is the third intermediate quantity.

7. The speech synthesis method of claim 1, wherein, the connection layer function comprises: wherein is the connection layer function; for the current speech; ) an operator that is an activation function; is a first weight matrix for a fully connected layer; a second weight matrix for a fully connected layer; is the current latent state for the current speech.

8. A speech synthesis apparatus characterized by comprising: The method comprises: an acquisition data module configured to acquire vocoder input data and a previous moment hidden state, wherein the vocoder input data comprises a mel spectrum of a previous moment speech and a current speech; a hidden state module configured to process the previous moment hidden state and the vocoder input data through an autoregressive encoder to generate a current hidden state of the current speech; a connection layer module configured to process the current hidden state through a fully connected layer to generate a connection layer function; a speech category module configured to process the connection layer function through a binary tree classifier to generate a category of the current speech, so as to generate a current synthesized speech according to the category of the current speech; wherein the category of the current speech refers to a category obtained by compressing the current speech by using a u-law method, and specifically comprises: wherein is the current speech, speech has a value range and is represented using 16 bits in binary. for speech compression results; is the operator of the sign function; =255。 9.A computer device, comprising a memory, a processor, and computer readable instructions stored in the memory and executable on the processor, wherein, the processor implements the speech synthesis method according to any one of claims 1 to 7 when executing the computer readable instructions.

10. One or more readable storage media storing computer readable instructions, which are executed by one or more processors to cause the one or more processors to perform the speech synthesis method according to any one of claims 1 to 7.

Images