KR102311239B1 - Deep neural network based non-autoregressive speech synthesizer method and system - Google Patents

Abstract

A deep neural network-based non-autoregressive speech synthesis method and system are presented. According to an embodiment, a non-autoregressive speech synthesis system based on a deep neural network includes: a sentence data analyzer configured to analyze sentence data and output refined sentence data; a speech feature vector sequence synthesizing unit for generating a template input and adding the refined sentence data to the generated template using an attention mechanism to generate a speech feature vector sequence; and a voice reconstruction unit that converts the voice feature vector sequence into voice data.

Description

DEEP NEURAL NETWORK BASED NON-AUTOREGRESSIVE SPEECH SYNTHESIZER METHOD AND SYSTEM

Embodiments of the present invention relate to a method and system for non-autoregressive speech synthesis based on a deep neural network, and more particularly, a deep neural network for generating a speech vector non-recursively by estimating the length of a speech feature vector and generating an empty input. It relates to a non-autoregressive speech synthesis method and system based on the present invention. This invention is a research conducted with the support of the Information and Communication Technology Promotion Center with funding from the government (Ministry of Science and ICT) in 2017 (No.2017-0-00474, Development of Intelligent Voice Signal Processing Technology for AI Speaker Voice Assistant) .

Deep Neural Network (DNN)-based speech synthesis technology is a technology that generates speech data from sentence data using a deep neural network. The speech feature vector sequence synthesizing unit is configured as one network.

The first stage, the sentence data analysis unit, separates the sentence data into letters and converts the separated letters into a valid vector sequence as a neural network input. The network consisting of

The second step, the speech feature vector sequence synthesizing unit, also consists of two steps. In the first step, information suitable for speech data is selectively collected from the sentence data vector sequence refined by the attention mechanism, and then the Based on the information, a recurrent neural network generates a Mel-Filterbank negative feature vector. At this time, the input of the recurrent neural network consists of the output Mel filter bank voice data generated in the previous step as the input according to the autoregressive method. In the second step, the Mel filter bank voice data is mapped to a log-power spectrogram.

In general, the quality of speech synthesis is measured using M.O.S (Mean Opinion Score), which is a subjective evaluation method. In the case of some vocoder models, the degree of speech distortion is measured, but only M.O.S is used in the case of an end-to-end model based on a deep neural network (DNN).

Deep neural network (DNN)-based end-to-end speech synthesis technology refers to a technology in which one deep neural network (DNN) model analyzes sentence data to generate spectrogram-based speech features or speech signals. Since the sampling rates of sentence data and speech data are different, a seq2seq (sequence-to-sequence) network and an attention mechanism are used to solve this problem. The seq2seq network consists of an encoder and a decoder. The encoder plays a role in refining the sentence data, and the decoder generates spectrogram-based speech features based on the information refined by the encoder. The decoder sequentially generates outputs based on an autoregressive flow in which the generated output becomes an input of the next time. However, the autoregressive flow enables efficient information transfer, but has a disadvantage in that it is slow because outputs must be sequentially generated according to time order.

As such, the existing neural network-based speech synthesis technology synthesizes speech feature vectors by using the Mel filter bank speech feature vector generated in the previous step as an input of the current step again in an autoregressive method. In the learning phase, there is no need to wait for the output of the previous step because the input is configured in a way that shifts and uses the already given target speech feature vector sequence. .

In the case of an autoregressive speech synthesis model, since the output goes back into the input, the sound volume of the speech frequently decreases as it goes backward due to the characteristics of the recurrent neural network. In addition, a structure for separately finding the end of a voice is required when generating a voice, and if the end is not found properly, the voice generation stops in the middle.

Wang Y, Skerry-Ryan RJ, Stanton D, Wu Y, Weiss RJ, Jaitly N, Yang Z, Xiao Y, Chen Z, Bengio S, Le Q. Tacotron: A fully end-to-end text-to-speech synthesis model. arXiv preprint arXiv:1703.10135. 2017 Mar 29.

Embodiments of the present invention describe a non-autoregressive speech synthesis method and system based on a deep neural network, and more specifically, speech synthesis that generates a speech vector non-recursively by estimating the length of a speech feature vector and generating an empty input Provides model creation technology.

Embodiments of the present invention remove the autoregressive flow of the existing autoregressive speech synthesis technology in order to solve the problem of the existing autoregressive speech synthesis technology, and use a template It is possible to provide a deep neural network-based non-autoregressive speech synthesis method and system that constructs a new input called

An object of the present invention is to provide a method for solving the problem that a speech synthesis model has to generate speech feature vectors step by step using a new input called a template. In addition, it is an object of the present invention to provide a method for solving a problem commonly encountered when synthesizing a long speech in an existing autoregressive model.

According to an embodiment, a non-autoregressive speech synthesis system based on a deep neural network includes: a sentence data analyzer configured to analyze sentence data and output refined sentence data; a speech feature vector sequence synthesizing unit for generating a template input and adding the refined sentence data to the generated template using an attention mechanism to generate a speech feature vector sequence; and a voice reconstruction unit that converts the voice feature vector sequence into voice data.

The sentence data analysis unit generates embedded sentence data in the form of a sentence feature vector string by decomposing the sentence data into Jamo units of Hangul to generate a Jamo unit input, embedding the sentence data, and convolution ( convolution) may be refined using an artificial neural network to form the refined sentence data.

The sentence data analysis unit generates a Jamo unit input by decomposing the sentence data into Jamo units of Hangul, indexes the Jamo unit input and maps it to numeric data, and encodes the numeric data into one-hot encoding (One-hot encoding). ), and multiplying the one-hot encoded vector sequence with the sentence embedding matrix to generate the embedded sentence data composed of a vector sequence having continuous characteristics.

The template, which is an input of the speech feature vector sequence synthesizing unit, may include absolute positional encoding data and relative positional encoding data.

The speech feature vector sequence synthesizing unit generates the template input, and adds the refined sentence data to the template input using an attention mechanism to generate an encoded template. After that, a mel filter bank voice feature vector sequence can be synthesized through decoding the encoded template, and a log power spectrum voice feature vector sequence can be synthesized from the mel filter bank voice feature vector sequence.

The speech reconstruction unit may generate phase information from the speech feature vector sequence having magnitude information using a Griffin-lim algorithm and convert it into the speech data.

A non-autoregressive speech synthesis method based on a deep neural network according to another embodiment includes a sentence data analysis step of analyzing sentence data and outputting refined sentence data; a speech feature vector sequence synthesis step of generating a template input and adding the refined sentence data to the generated template using an attention mechanism to generate a speech feature vector sequence; and a speech reconstruction step of converting the speech feature vector sequence into speech data.

The step of analyzing the sentence data may include: decomposing the sentence data into Jamo units of Hangul to generate a Jamo unit input, then embedding the sentence data to form embedded sentence data in the form of a sentence feature vector string; and refining the embedded sentence data using a convolutional artificial neural network to form the refined sentence data.

The forming of the embedded sentence data in the form of a sentence feature vector string may include: generating a Jamo unit input by decomposing the sentence data into Jamo units of Hangul; indexing the Jamo unit input and mapping it to numeric data; One-hot encoding the numeric data; and multiplying a one-hot encoded vector sequence with a sentence embedding matrix to generate the embedded sentence data including a vector sequence having continuous characteristics.

Here, the template, which is an input of the speech feature vector sequence synthesizing unit, may include absolute positional encoding data and relative positional encoding data.

The step of synthesizing the speech feature vector sequence may include: generating the template input; a voice data encoding step of generating an encoded template by adding the refined sentence data to the template input using an attention mechanism; a speech data decoding step of synthesizing a Mel filter bank speech feature vector sequence through decoding the encoded template; and synthesizing a log power spectrum speech feature vector sequence from the Mel filter bank speech feature vector sequence.

In addition, the generating of the template input may include: generating encoded data of an absolute position; generating encoded data of a relative position; and generating the template by concatenating the generated encoded data of the absolute position and the encoded data of the relative position.

The encoding of the speech data may include: receiving the refined sentence data and the template as inputs, selecting a part required for log power spectrum synthesis by the attention mechanism, and forming a vector of a fixed length; It may include encoding a template containing accurate information by repeating the convolutional artificial neural network and the attention mechanism.

The speech data decoding step may include synthesizing a Mel filter bank speech feature vector sequence from the encoded template through a speech data decoding artificial neural network.

In the speech reconstruction step, phase information may be generated from the speech feature vector sequence having magnitude information using a Griffin-lim algorithm and converted into the speech data.

According to embodiments of the present invention, since speech feature vector sequences are synthesized at once in a non-autoregressive method, a deep neural network-based system capable of synthesizing speech at a faster speed compared to an autoregressive speech synthesis method It is possible to provide a method and system for non-autoregressive speech synthesis of

In addition, according to embodiments of the present invention, there is an advantage in that the size of the output does not gradually decrease in the autoregressive-based model regression model, so that the volume of the voice is kept constant throughout the sentence.

1 is a diagram showing the overall configuration of a non-autoregressive speech synthesis system according to an embodiment.
2 to 4 are flowcharts illustrating a non-autoregressive speech synthesis method according to an exemplary embodiment.
5 is a diagram for explaining a process of generating a Mel filter bank speech feature vector sequence according to an embodiment.

Hereinafter, embodiments will be described with reference to the accompanying drawings. However, the described embodiments may be modified in various other forms, and the scope of the present invention is not limited by the embodiments described below. In addition, various embodiments are provided in order to more completely explain the present invention to those of ordinary skill in the art. The shapes and sizes of elements in the drawings may be exaggerated for clearer description.

The following embodiments of the present invention relate to a deep neural network-based non-autoregressive speech synthesis method and system for non-recursively generating speech vectors by estimating the length of speech feature vectors and generating empty inputs. Embodiments eliminate the autoregressive flow of the existing autoregressive speech synthesis system and configure a new input called a template to generate speech in the test phase in the same way as in the learning phase. let it be

According to embodiments, by synthesizing the speech feature vector sequence in a non-autoregressive method at once, speech may be synthesized at a faster speed than in the autoregressive speech synthesis method.

1 is a diagram showing the overall configuration of a non-autoregressive speech synthesis system according to an embodiment.

Referring to FIG. 1 , a non-autoregressive speech synthesis system 100 based on a deep neural network according to an embodiment includes a sentence data analyzer 110 that analyzes sentence data 101 and a voice feature vector. It may include a speech feature vector sequence synthesizing unit 120 for synthesizing the sequence 104, and a speech reconstruction unit 130 for converting the speech feature vector sequence 104 into speech.

The sentence data analysis unit 110 may analyze the sentence data 101 and output the refined sentence data 102 . The sentence data analysis unit 110 may be divided into a sentence data embedding part that changes the sentence input data received in units of Hangul characters to an input of the deep neural network, and an artificial neural network part that refines the embedded data. More specifically, the sentence data embedding part decomposes the sentence data 101 into Jamo units of Hangul to generate a Jamo unit input, then embeds it to form embedded sentence data in the form of a sentence feature vector string, and the artificial neural network part is The refined sentence data 102 may be formed by refining the embedded sentence data using a convolution artificial neural network.

In particular, the sentence data analysis unit 110 decomposes the sentence data 101 into the units of the Korean alphabet to generate a unit input of the unit of Jamo, indexes the input in units of the unit of Jamo and maps it to numeric data, and converts the numeric data into one-hot encoding ( One-hot encoding and one-hot encoding are multiplied by a sentence embedding matrix to generate embedded sentence data composed of a vector sequence having continuous characteristics.

The speech feature vector sequence synthesizing unit 120 generates a template (Template) 103 input, and adds refined sentence data 102 using an attention mechanism to the generated template (Template 103) to make speech A feature vector sequence 104 may be generated. Here, the template 103, which is an input of the speech feature vector sequence synthesizing unit 120, may include absolute positional encoding data and relative positional encoding data.

Also, the speech feature vector sequence synthesizing unit 120 may include a speech data encoding unit and a speech data decoding unit.

The speech feature vector sequence synthesizing unit 120 generates a template (Template, 103) input, and refines the sentence data (102) using an attention mechanism on the template (103) input through the speech data encoding unit. can be added to create an encoded template. Thereafter, the voice data decoding unit may synthesize the Mel filter bank voice feature vector sequence by decoding the encoded template. In addition, a log power spectrum speech feature vector sequence can be synthesized from the Mel filter bank speech feature vector sequence.

The speech reconstruction unit 130 may convert the speech feature vector sequence 104 into speech data 105 . More specifically, the speech reconstruction unit 130 generates phase information from the speech feature vector sequence 104 having magnitude information using a Griffin-lim algorithm to generate speech data ( 105) can be converted to

2 to 4 are flowcharts illustrating a non-autoregressive speech synthesis method according to an exemplary embodiment.

Referring to FIG. 2 , a non-autoregressive speech synthesis method based on a deep neural network according to an embodiment includes a sentence data analysis step (S110) of analyzing sentence data and outputting refined sentence data, a template A speech feature vector sequence synthesis step (S120) of generating a (Template) input and adding refined sentence data to the generated template using an attention mechanism to generate a speech feature vector sequence (S120), and speech A voice reconstruction step (S130) of converting the feature vector sequence into voice data may be included.

Referring to FIG. 3 , in the sentence data analysis step (S110), the sentence data is decomposed into the units of the Korean alphabet to generate the unit input of the characters, and then embedded to form the embedded sentence data in the form of a sentence feature vector column (S111). ) and refining the embedded sentence data using a convolution artificial neural network to form refined sentence data ( S112 ).

Referring to FIG. 4 , in the speech feature vector sequence synthesis step S120, the template input is generated (S121), and the sentence data is refined using an attention mechanism to the template input. A voice data encoding step (S122) of generating an additionally encoded template (Template), a voice data decoding step (S123) of synthesizing a Mel filter bank speech feature vector sequence through decoding the encoded template (S123), and a Mel filter It may include the step of synthesizing the log power spectrum speech feature vector sequence in the bank speech feature vector sequence (S124).

Hereinafter, each step of the non-autoregressive speech synthesis method according to an embodiment will be described in more detail.

The deep neural network-based non-autoregressive speech synthesis method according to an embodiment may be described in more detail by taking the deep neural network-based non-autoregressive speech synthesis system according to the embodiment described with reference to FIG. 1 as an example. Here, the deep neural network-based non-autoregressive speech synthesis system according to an embodiment may include a sentence data analysis unit, a speech feature vector sequence synthesis unit, and a speech reconstruction unit.

In the sentence data analysis step ( S110 ), the sentence data analyzer may analyze the sentence data and output refined sentence data. Here, the sentence data analysis step ( S110 ) may generate refined sentence data by refining it with one artificial neural network. In this case, the artificial neural network can be called a sentence data purification artificial neural network, and it can be trained.

This sentence data analysis step (S110) is a step (S111) of forming embedded sentence data in the form of a sentence feature vector string by decomposing the sentence data into Jamo units of Hangul to generate a Jamo unit input, and then embedding the embedded sentence data. The method may include refining the sentence data using a convolutional artificial neural network to form refined sentence data (S112). Here, embedding is a process for converting discontinuous symbol data such as a sentence into a feature vector having continuous and various characteristics.

The step of forming the embedded sentence data (S111) is a step of decomposing the sentence data into the units of the Korean alphabet to generate a unit input (sentence decomposition step), the step of indexing the input of the Jamo unit and mapping the input to numeric data (index) step), one-hot encoding of numeric data, and one-hot encoded vector sequence multiplied by a sentence embedding matrix to form an embedded sentence consisting of a vector sequence having continuous characteristics It may include a step of generating data (a feature vector transformation step).

The sentence decomposition step is a step of splitting the Hangul sentences into letters, and the index step is a step of numbering the split letters, respectively. Through these two steps, sentence data is converted into numeric data. The generated sentence feature vector sequence (ie, embedded sentence data) can be expressed as the following equation.

[Equation 1]

)를 수학식 2와 같이 임베딩 매트릭스(

)와 각각 곱하여 특징 벡터로 변환할 수 있다. 이에 따라 벡터열로 이루어진 임베딩된 문장 데이터를 생성할 수 있다. In the feature vector conversion step, after one-hot encoding of the generated numeric data, the generated one-hot vector (

) to the embedding matrix (

) and multiplied by each to convert it into a feature vector. Accordingly, it is possible to generate embedded sentence data composed of a vector sequence.

[Equation 2]

In the speech feature vector sequence synthesis step ( S120 ), the speech feature vector sequence synthesis unit generates a template input and adds refined sentence data to the generated template using an attention mechanism to make speech You can create a feature vector sequence.

In the speech feature vector sequence synthesis step (S120), the template input is generated (S121), the template input is encoded by adding refined sentence data using an attention mechanism to the template input (S121). In the speech data encoding step (S122) of generating a template), the speech data decoding step (S123) of synthesizing the Mel filter bank speech feature vector sequence through decoding the encoded template (Template), and the Mel filter bank speech feature vector sequence The step of synthesizing a log power spectrum speech feature vector sequence (S124) may be included.

In addition, the step (S121) of generating a template input includes the steps of generating encoded data of an absolute position, generating encoded data of a relative position, and encoding data of an absolute position and the encoded data of a relative position. It may include the step of creating a template (Template) by concatenating.

Then, the speech data encoding step (S122) includes the steps of receiving refined sentence data and a template as inputs, the attention mechanism selecting a part necessary for log power spectrum synthesis to form a vector of a fixed length, and convolution ( convolution) repeating the artificial neural network and the attention mechanism to encode a template containing accurate information. In this case, a voice data encoding artificial neural network may be trained, and the voice data encoding artificial neural network may be a convolutional artificial neural network.

In addition, the speech data decoding step S123 may include synthesizing a Mel filter bank speech feature vector sequence from a template encoded through a speech data decoding artificial neural network. In this case, the speech data decoding artificial neural network may be trained.

As mentioned earlier, the template (or template data), which is the input of the speech feature vector sequence synthesizing unit, can be composed of relative positional encoding and absolute positional encoding. have. The positional encoding of each position can be configured in the same dimension as the Mel filter bank speech feature vector sequence. That is, it is a matrix having a size of [time, frequency bin]. The composition of the absolute position encoding can be expressed as the following equation.

[Equation 3]

)

)

)

)

세 개의 파라미터를 가지는 sin, cos 파형으로 구성될 수 있다. 이때, pos는 데이터 내의 시간 순서에 따라 1부터 커지는 자연수이며,

는 데이터 내의 주파수 빈의 순서에 따라 1부터 커지는 자연수이다.

은 멜 필터 뱅크의 주파수 빈 개수와 같은 80이다. Absolute position encoding is

It can be composed of sin and cos waveforms with three parameters. At this time, pos is a natural number that increases from 1 according to the time order in the data,

is a natural number that increases from 1 according to the order of frequency bins in the data.

is 80 equal to the number of frequency bins in the Mel filter bank.

Sin waves and cos waves that do not overlap with time add temporal information to data, and unlike recurrent neural networks, convolutional neural networks, which have limited learning of sequence information, help to reflect sequence and temporal information in the results.

The composition of the encoding of the relative position can be expressed as the following equation.

[Equation 4]

)

)

)

)

파라미터 대신에 전체 음성 데이터의 길이인

파라미터가 포함될 수 있다.

파라미터가 위치의 인코딩의 구성에 포함되기 때문에 절대적인 위치의 인코딩과 다르게 전체 길이에 대한 상대적인 시간 정보를 표현할 수 있게 된다. Relative position encoding is similar to absolute position encoding, but

The length of the entire voice data instead of parameters

Parameters may be included.

Since the parameter is included in the composition of the encoding of the position, it is possible to express relative temporal information for the entire length differently from the encoding of the absolute position.

The encoding of each position is concatenated to constitute a Template. Therefore, the entire template is a matrix having a size of [time, frequency bin*2].

5 is a diagram for explaining a process of generating a Mel filter bank speech feature vector sequence according to an embodiment.

Referring to FIG. 5 , the process of generating the Mel filter bank speech feature vector sequence 204 of the speech feature vector sequence synthesizing unit of the deep neural network-based speech synthesis system can be confirmed.

The speech feature vector sequence synthesizing unit may include a speech data encoding unit 220 and a speech data decoding unit 230 . The voice data encoding unit 220 may include a plurality of convolutional networks 221 and an attention mechanism 210 . Through each attention mechanism 210 , the template 202 may be encoded with new information appropriately containing sentence information for generating the Mel filter bank speech feature vector sequence 204 .

In general, there is a difference in sampling rate between the speech feature vector sequence and the sentence data. Compared to the sentence data, the sampling rate of the speech feature vector sequence is generally much larger. When information on sentence data is added to the template 202, a problem arises as to which sentence data to add for each time of the template 202 due to the difference in the sampling rate. The attention mechanism 210 has shown a great effect in solving this problem in a general autoregressive speech synthesis system. The present invention also appropriately adds the refined sentence data 201 to the template 202 using the attention mechanism 210 .

However, sentence information and speech information of the previous time are contained in the Mel filter bank speech feature vector string 204, which is an input of the autoregressive model, whereas only time information is contained in the template 202. Accordingly, in order to solve the lack of template (Template) 202 information, the present invention can provide a method of increasing the accuracy of information so that the attention mechanism 210 is recursive as shown in the structure of FIG. 5 . The speech data encoding unit 220 may appropriately add the refined sentence data 201 information to the template 202 by using the attention mechanism 210 and the convolutional network 221 a total of three times.

After passing through the speech data encoding unit 220 , the template 202 is mapped to an encoded template 203 including sentence information for synthesizing the Mel filter bank speech feature vector sequence 204 . The process of each attention mechanism 210 is as follows.

First, the refined sentence data 201 is divided into a V matrix and a K matrix. At this time, the dimensions of the V and K matrices are the same as [sentence time, channel]. A template that has passed through the convolutional network 221 is called a Q matrix, and the size is composed of [voice time, channel]. In this case, the channel values of the sentence data and the voice data are the same.

[Equation 5]

는

와

의 스칼라 곱으로 구해진다. 이때, 행렬

의 크기는 [음성 시간, 문장 시간]이다. 행렬

의 원소를

로 나타낼 수 있다.[Equation 5] is an expression related to a method of obtaining a matching degree between two inputs in the attention mechanism 210 . A matrix indicating the degree of matching between two matrices.

Is

Wow

is obtained as the scalar product of In this case, the matrix

The size of is [voice time, sentence time]. procession

the elements of

can be expressed as

[Equation 6]

행렬의 데이터를 소프트맥스(soft-max) 함수를 이용해 확률의 의미를 가지는

로 변환할 수 있다.

로 이루어진 행렬을 A로 나타내고 얼라인먼트(alignment) 행렬이라 부른다. 어텐션 메커니즘(210)의 최종 결과물인

행렬을 구하는 과정을 다음 식과 같이 나타낼 수 있다.[Equation 6] is

Using the soft-max function for matrix data,

can be converted to

A matrix consisting of is denoted by A and is called an alignment matrix. The final result of the attention mechanism 210 is

The process of finding a matrix can be expressed as the following equation.

[Equation 7]

행렬은

행렬과 동일한 차원 [음성 시간, 채널]을 가지기 때문에

와

를 병합(concatenate)하여 네트워크의 다음 입력으로 사용할 수 있다.The C matrix contains information about sentence data and is called context data. this

the matrix is

Because it has the same dimension [voice time, channel] as the matrix

Wow

can be concatenated and used as the next input to the network.

The attention mechanism 210 is applied a total of three times, and may sequentially form a more accurate alignment matrix. The encoded template 203 generated through the same process as described above may be mapped to the Mel filter bank voice feature vector sequence 204 through the voice data decoding unit 230 .

The Mel filter bank speech feature vector sequence 204 synthesized through the above process may be mapped to a log power spectrum speech feature vector sequence through a post-processing artificial neural network. The artificial neural network used in the sentence data analyzer and the speech feature vector sequence synthesizer can be learned through an error between the two synthesized feature vector sequences and the actual speech feature vector data. That is, it is possible to train a speech feature vector sequence synthesis artificial neural network.

In the speech reconstruction step ( S130 ), the speech reconstruction unit may convert the speech feature vector sequence into speech data.

The voice reconstruction unit may recover voice data using the log power spectrum voice feature vector sequence finally synthesized in the previous step. Since the log power spectrum speech feature vector sequence synthesized by the deep neural network has only magnitude information without phase information, it must be reconstructed by generating new phase information using the Griffin-lim algorithm. do.

According to embodiments, since the speech feature vector sequence is synthesized at once in a non-autoregressive method, speech can be synthesized at a higher speed than in the autoregressive speech synthesis method. In addition, a phenomenon in which the size of the output gradually decreases does not appear in the autoregressive-based model regression model, so that the size of the speech may be maintained constant throughout the sentence.

The device described above may be implemented as a hardware component, a software component, and/or a combination of the hardware component and the software component. For example, devices and components described in the embodiments may include, for example, a processor, a controller, an arithmetic logic unit (ALU), a digital signal processor, a microcomputer, a field programmable array (FPA), It may be implemented using one or more general purpose or special purpose computers, such as a programmable logic unit (PLU), microprocessor, or any other device capable of executing and responding to instructions. The processing device may execute an operating system (OS) and one or more software applications running on the operating system. The processing device may also access, store, manipulate, process, and generate data in response to execution of the software. For convenience of understanding, although one processing device is sometimes described as being used, one of ordinary skill in the art will recognize that the processing device includes a plurality of processing elements and/or a plurality of types of processing elements. It can be seen that can include For example, the processing device may include a plurality of processors or one processor and one controller. Other processing configurations are also possible, such as parallel processors.

The software may comprise a computer program, code, instructions, or a combination of one or more thereof, which configures a processing device to operate as desired or is independently or collectively processed You can command the device. The software and/or data may be any kind of machine, component, physical device, virtual equipment, computer storage medium or device, to be interpreted by or to provide instructions or data to the processing device. may be embodied in The software may be distributed over networked computer systems, and stored or executed in a distributed manner. Software and data may be stored in one or more computer-readable recording media.

The method according to the embodiment may be implemented in the form of program instructions that can be executed through various computer means and recorded in a computer-readable medium. The computer-readable medium may include program instructions, data files, data structures, etc. alone or in combination. The program instructions recorded on the medium may be specially designed and configured for the embodiment, or may be known and available to those skilled in the art of computer software. Examples of the computer-readable recording medium include magnetic media such as hard disks, floppy disks and magnetic tapes, optical media such as CD-ROMs and DVDs, and magnetic such as floppy disks. - includes magneto-optical media, and hardware devices specially configured to store and execute program instructions, such as ROM, RAM, flash memory, and the like. Examples of program instructions include not only machine language codes such as those generated by a compiler, but also high-level language codes that can be executed by a computer using an interpreter or the like.

As described above, although the embodiments have been described with reference to the limited embodiments and drawings, various modifications and variations are possible for those skilled in the art from the above description. For example, the described techniques are performed in a different order than the described method, and/or the described components of the system, structure, apparatus, circuit, etc. are combined or combined in a different form than the described method, or other components Or substituted or substituted by equivalents may achieve an appropriate result.

Therefore, other implementations, other embodiments, and equivalents to the claims are also within the scope of the following claims.

Claims (15)

a sentence data analysis unit that analyzes the sentence data and outputs refined sentence data;
a speech feature vector sequence synthesizing unit for generating a template input and adding the refined sentence data to the generated template using an attention mechanism to generate a speech feature vector sequence; and
A speech reconstruction unit that converts the speech feature vector sequence into speech data
including,
The sentence data analysis unit,
After decomposing the sentence data into Jamo units of Hangul to generate a Jamo unit input, embedding to form embedded sentence data in the form of a sentence feature vector string, and using a convolution artificial neural network for the embedded sentence data to form the refined sentence data by refining it,
The template, which is an input of the speech feature vector sequence synthesizing unit,
After generating absolute positional encoding data including order and time information and relative positional encoding data including relative temporal information for the entire length, the generated absolute positional encoding data By concatenating data and the encoded data of the relative position, the template having time information is generated, and the encoded data of the absolute position is a natural number increasing from 1 according to the temporal order in the data, the frequency bin in the data. It is composed of sin and cos waveforms having three parameters: a natural number that increases from 1 according to the order, and the frequency bin number of the Mel filter bank to help the convolutional artificial neural network reflect the order and time information in the result, and the relative position of the The encoded data expresses relative temporal information about the total length, including parameters of the length of the entire voice data,
The speech feature vector sequence synthesizing unit,
After generating the template input and adding the refined sentence data to the template input using an attention mechanism to generate an encoded template, the encoded template ( Template) to synthesize a Mel filter bank speech feature vector sequence through decoding, and synthesizing a log power spectrum speech feature vector sequence from the Mel filter bank speech feature vector sequence
A non-autoregressive speech synthesis system based on a deep neural network. delete According to claim 1,
The sentence data analysis unit,
By decomposing the sentence data into Jamo units of Hangul to generate a Jamo unit input, indexing the Jamo unit input and mapping it to numeric data, one-hot encoding the numeric data, and one-hot Multiplying the encoded (One-hot encoded) vector sequence with the sentence embedding matrix to generate the embedded sentence data consisting of a vector sequence having continuous characteristics
A non-autoregressive speech synthesis system based on a deep neural network. delete delete According to claim 1,
The voice reconstruction unit,
Generating phase information from the speech feature vector sequence having magnitude information using a Griffin-lim algorithm and converting it into the speech data
A non-autoregressive speech synthesis system based on a deep neural network. A sentence data analysis step of analyzing the sentence data and outputting refined sentence data;
a speech feature vector sequence synthesis step of generating a template input and adding the refined sentence data to the generated template using an attention mechanism to generate a speech feature vector sequence; and
Speech reconstruction step of converting the speech feature vector sequence into speech data
including,
The sentence data analysis step is,
forming embedded sentence data in the form of a sentence feature vector string by decomposing the sentence data into Jamo units of Hangul to generate a Jamo unit input; and
Refining the embedded sentence data using a convolutional artificial neural network to form the refined sentence data
including,
The step of synthesizing the speech feature vector sequence,
generating the template input;
a voice data encoding step of generating an encoded template by adding the refined sentence data to the template input using an attention mechanism;
a speech data decoding step of synthesizing a Mel filter bank speech feature vector sequence through decoding the encoded template; and
Synthesizing a log power spectrum speech feature vector sequence from the Mel filter bank speech feature vector sequence
includes,
The step of generating the template input includes:
generating encoded data of an absolute position including order and time information;
generating encoded data of a relative position including relative time information for an entire length; and
generating the template having time information by concatenating the generated encoded data of the absolute position and the encoded data of the relative position
including,
The encoded data of the absolute position is a sin, cos waveform having three parameters: a natural number increasing from 1 according to the time order in the data, a natural number increasing from 1 according to the order of frequency bins in the data, and the frequency bin number of the Mel filter bank. It is configured to help the convolutional neural network reflect order and time information in the result, and the encoding data of the relative position includes the parameter of the length of the entire voice data to express relative temporal information for the entire length.
Characterized in, a deep neural network-based non-autoregressive speech synthesis method. delete 8. The method of claim 7,
The step of forming the embedded sentence data in the form of the sentence feature vector column comprises:
generating a Jamo unit input by decomposing the sentence data into Jamo units of Hangul;
indexing the Jamo unit input and mapping it to numeric data;
One-hot encoding the numeric data; and
generating the embedded sentence data consisting of a vector sequence having continuous characteristics by multiplying a one-hot encoded vector sequence with a sentence embedding matrix
Including, a deep neural network-based non-autoregressive speech synthesis method. delete delete delete 8. The method of claim 7,
The voice data encoding step includes:
receiving the refined sentence data and the template as inputs and forming a vector of a fixed length by the attention mechanism selecting a part necessary for log power spectrum synthesis; and
Encoding a template containing accurate information by repeating the convolutional artificial neural network and the attention mechanism
Including, a deep neural network-based non-autoregressive speech synthesis method. delete 8. The method of claim 7,
The voice reconstruction step is
Generating phase information from the speech feature vector sequence having magnitude information using a Griffin-lim algorithm and converting it into the speech data
Characterized in, a deep neural network-based non-autoregressive speech synthesis method.

Images