KR20220022337A - A bit-stream-based modified vocoder using autoencoder, computer-readable storage medium and computer program - Google Patents

Abstract

A modified voice encoder of an embodiment may comprise a decoder device of the modified voice encoder comprising: an encoding device of the modified voice encoder comprising a voice compression part that generates a quantized bit string per frame using an original sound, a feature extraction part of a self-encoder encoder part that extracts a bottleneck feature of the self-encoder encoder by inputting the bit string extracted from the voice compression part, and a quantization part that quantizes the bottleneck feature and converts thereof into the bit string; a self-encoder decoder part that restores the bit string extracted from the encoding device; and a voice decoding part that synthesizes a voice by inputting the restored bit string. Therefore, the present invention is capable of having an effect for which can be used in a communication system that requires security such as for the military without loss compared to an existing one.

Description

A bit stream-based modified speech encoder using a magnetic encoder, a computer-readable recording medium, and a computer program

The present invention relates to a technique for transforming a voice encoder for security in digital voice communication.

In general, a low bit rate voice encoder for digital voice communication includes linear predictive coding (LPC)-10e, mixed-exited linear prediction (MELP), code-excited linear prediction (CELP), and the like. In a communication system requiring security, such as for military use, a technology for transforming such a voice encoder is required.

In order to solve the above-mentioned problems, the embodiment aims to provide a bit stream-based transformed speech encoder capable of transforming the speech encoder using an autoencoder utilizing a deep learning technique.

The encoding apparatus of the modified speech encoder of the embodiment includes a speech compression unit that generates a quantized bit stream per frame using the original sound, and a self-encoder extracting bottleneck features of the encoder by inputting the bit string extracted from the speech compression unit as an input The encoder may include a feature extraction unit of the encoder unit, and a quantization unit that quantizes the bottleneck feature and converts it into a bit stream.

The quantization unit may linearly quantize the bit stream.

The bottleneck feature may be a value of a node generated as a bottleneck in the node of the self-encoder.

The speech compression unit may generate a quantized bit stream per frame by passing the original sound through the music sample.

The decoding apparatus of the modified speech encoder of the embodiment may include a self-encoder decoder unit for restoring a bit stream extracted from the encoding device of the modified speech encoder, and a speech decoder for synthesizing speech by inputting the restored bit string as an input.

The speech decoder substitutes 1 for a value larger than a threshold value in the vector output from the self-encoder decoder unit, substitutes 0 for a value smaller than the threshold value, and collects synthesized speech using the substituted value for each frame. can do.

According to an embodiment of the present invention, by synthesizing speech using a modified speech encoder using a self-encoder, there is an effect that it can be used in a communication system requiring security such as military use without loss compared to the existing one.

1 is a block diagram of an encoder device for obtaining a bit stream for transmission in a modified speech encoder according to an embodiment of the present invention.
2 is a block diagram of a decoder device for synthesizing speech when a bit stream is received in the modified speech encoder according to an embodiment of the present invention.
3 is a flowchart illustrating an encoding process for obtaining a bit stream for transmission in a modified speech encoder according to an embodiment of the present invention.
4 is a flowchart illustrating a decoding process in which speech is synthesized when a bit stream is received in the modified speech encoder according to an embodiment of the present invention.

Advantages and features of the present invention and methods of achieving them will become apparent with reference to the embodiments described below in detail in conjunction with the accompanying drawings. However, the present invention is not limited to the embodiments disclosed below, but may be implemented in various forms, and only these embodiments allow the disclosure of the present invention to be complete, and those of ordinary skill in the art to which the present invention pertains. It is provided to fully inform the person of the scope of the invention, and the scope of the invention is only defined by the claims.

In describing the embodiments of the present invention, detailed descriptions of well-known functions or configurations will be omitted except when it is actually necessary to describe the embodiments of the present invention. In addition, the terms to be described later are terms defined in consideration of functions in an embodiment of the present invention, which may vary according to intentions or customs of users and operators. Therefore, the definition should be made based on the content throughout this specification.

An embodiment of the present invention trains a self-encoder that uses a bit string, which is an output of a speech encoder encoder, as input and output in a speech encoding system that synthesizes speech using the parameters of a speech speech model, and separates the encoder and decoder of the trained self-encoder It is based on quantizing the bottleneck feature of the self-encoder and then transmitting it as a bit stream.

That is, the embodiment of the present invention is based on training the self-encoder with the bit stream that is the output of the speech encoder encoder and transmitting the bit stream using the trained self-encoder encoder and decoder.

Hereinafter, embodiments of the present invention will be described in detail with reference to the accompanying drawings.

1 is a block diagram of an encoder device for obtaining a bit stream for transmission from a modified speech encoder according to an embodiment of the present invention, and FIG. It is a block diagram of a decoder device for synthesizing.

Referring to FIG. 1 , the encoder device of the modified speech encoder according to the embodiment includes a speech compression unit 100 that generates a quantized bit string per frame using an original sound, and the bit string extracted from the speech compression unit 100 . It may include a feature extracting unit 102 of the self-encoder encoder unit for extracting the bottleneck feature of the self-encoder encoder as input, and a quantization unit 104 for quantizing the bottleneck feature and converting it into a bit string.

Here, x(n) input to the speech compression unit 100 represents an original sound input to the speech encoding system, and output y(n) represents a bit stream transmitted from the modified speech encoder.

The speech compression unit 100 may pass a speech sample through an encoder of an existing speech encoder to obtain a quantized bit stream per frame.

The feature extraction unit 102 of the encoder unit of the self-encoder may extract the bottleneck feature by inputting the bit string obtained through the speech compression unit 100 to the encoder of the self-encoder. In this case, the self-encoder can be trained by using the quantized bit stream output from the existing speech encoder encoder as input/output. In such a self-encoder, a model may be constructed using, for example, a dense layer, and an activation function such as softsign or ReLU may be used.

The self-encoder may be a neural network model using deep learning, but is not limited thereto. The magnetic encoder may be composed of an input layer, a hidden layer, and an output layer, and the hidden layer may be composed of two or more layers. A plurality of nodes may be formed between the input layer, the hidden layer, and the output layer, and an encoder may be on the left with respect to the hidden layer and a decoder on the right with respect to the hidden layer, but is not limited thereto.

The quantization unit 104 linearly scalar quantizes the bit stream extracted from the self-encoder. The reason for performing quantization in the embodiment of the present invention is that actual digital communication is considered. For quantization, a conventional quantization technique such as a multi-step quantization technique may be used, but the present invention is not limited thereto.

은 각각 변형 음성 부호화기에서 수신하는 비트 열과 출력되는 합성 음성을 나타낸다.Referring to FIG. 2 , the decoder device of the modified speech encoder according to the embodiment includes a self-encoder decoder unit 200 that restores a bit stream extracted from the encoding device of the modified speech encoder, and receives the restored bit string as an input. It may include a voice decoder 202 for synthesizing. Here, the input y(n) and the output

represents a bit stream received from the modified speech encoder and a synthesized speech output, respectively.

The decoder unit 200 inversely quantizes the bit stream received per frame and inputs it to the decoder of the self-encoder to obtain an estimated value vector of the bit stream of the speech encoder.

The speech decoding unit 202 substitutes 1 for a value larger than a specific threshold value and 0 for a value smaller than a specific threshold in the vector output from the decoder of the self-encoder inside the decoder unit, and inputs the synthesized speech to the decoder of the existing speech encoder. You can get it frame by frame.

Hereinafter, an operation process of the modified speech encoder according to the embodiment will be described with reference to FIGS. 3 and 4 .

3 is a flowchart illustrating an encoding process for obtaining a bit stream for transmission in a modified speech encoder according to an embodiment of the present invention, and FIG. 4 is a flowchart illustrating a bit string received by the modified speech encoder according to an embodiment of the present invention. It is a flowchart illustrating this synthesized decoding process.

As shown in FIG. 3 , the speech compression unit 100 may extract a bit stream of an existing speech encoder ( S100 ).

Thereafter, the feature extracting unit 102 of the self-encoder encoder unit may extract the bottleneck feature of the self-encoder encoder using the bit string output from the speech compression unit 100 as an input (S102).

The quantization unit 104 may quantize the bottleneck feature extracted by the feature extraction unit 102 of the encoder of the self-encoder (S104). For example, if the voice compression unit 100 uses a 1.2 kbps MELP encoder as the existing voice encoder, the encoder unit 100 uses the 13th order as a bottleneck feature, and the quantizer 104 performs 4-bit scalar quantization, In the conventional speech encoder, the quantized 27 bits extracted per frame are input to the self-encoder to obtain the 13th vector, and finally, the encoder of the transformed speech encoder can transmit the 54th bit stream per frame.

As shown in FIG. 4 , the decoder unit 200 inversely quantizes the bit stream received per frame and then inputs it to the decoder of the self-encoder to obtain an estimated value vector of the bit stream of the speech encoder ( S200 ). In the embodiment of the present invention, inverse quantization is a design of the quantizer used in the quantization unit 104 inversely. For example, if fourth-order linear scalar quantization is used in the quantization unit 104 , the decoder unit 200 inversely quantizes each bit stream by 4 .

The speech decoding unit 202 substitutes 1 for a value larger than a specific threshold value and 0 for a value smaller than a specific threshold in the vector output from the decoder of the self-encoder inside the decoder unit, and inputs the synthesized speech to the decoder of the existing speech encoder. It can be obtained for each frame (S202). For example, if a 1.2 kbps MELP encoder is used as an existing voice encoder in the voice compression unit 100 and the voice decoder 202 and a 54-bit stream per frame is received by the decoder unit 200, the final 27th order per frame A voice may be synthesized through the bit string.

[Equation 1]

In [Equation 1], LSD (Long Spectral Distance) is a value (dB) representing a difference between spectra, and the lower the LSD, the lower the degree of distortion between the original sound and the voiced sound.

In the embodiment of the present invention, the results were simulated by comparing the LSD values, and [Table 1] below shows the sound source synthesized by the MELP vocoder, which is an existing voice encoder, and the sound source synthesized by the modified voice encoder as in the embodiment of the present invention, respectively. This is the result when compared to the original negative sample. LSD values and standard invasive speech intelligibility estimation method (STOI) scores were used to represent the sound quality of synthesized speech.

Vocoder




Measure MELP
1.2 kbps Encoder & Decoder MELP
2.4 kbps Encoder & Decoder Modified MELP Encoder & Decoder LSD 12.39 9.17 12.49 STOI 0.82 0.88 0.80

As in the embodiment of the present invention, when the sound quality of speech synthesized by a modified speech encoder is compared with that of an existing speech encoder, it can be seen that a slight decrease in intelligibility due to a quantization error occurs, but similar performance is shown.

According to the embodiment of the present invention as described above, a modified speech encoder having little difference in sound quality from the speech synthesized by the existing speech encoder is implemented using a self-encoder.

Various embodiments of the present document include software (eg, a machine-readable storage media) (eg, a memory (internal memory or external memory)) including instructions stored in a readable storage medium (eg, a computer). : program) can be implemented. The device is a device capable of calling a stored command from a storage medium and operating according to the called command, and may include the electronic device according to the disclosed embodiments. When the command is executed by the control unit, the control unit may perform a function corresponding to the command directly or using other components under the control of the control unit. Instructions may include code generated or executed by a compiler or interpreter. The device-readable storage medium may be provided in the form of a non-transitory storage medium. Here, non-transitory means that the storage medium does not include a signal and is tangible, and does not distinguish that data is semi-permanently or temporarily stored in the storage medium.

According to an embodiment, the method according to various embodiments disclosed in the present document may be included and provided in a computer program product.

According to an embodiment, as a computer-readable recording medium storing a computer program, generating a quantized bit stream per frame using an original sound, and extracting a bottleneck feature of a self-encoder encoder by inputting the bit string as an input quantizing the bottleneck feature and converting it into a bit stream, restoring the converted bit string, and synthesizing speech using the restored bit string as an input. It may include instructions for causing a processor to perform the method.

According to an embodiment, as a computer program stored in a computer-readable recording medium, generating a quantized bit stream per frame using an original sound, and extracting a bottleneck feature of a self-encoder encoder by inputting the bit string as an input quantizing the bottleneck feature and converting it into a bit stream, restoring the converted bit string, and synthesizing speech using the restored bit string as an input. It may include instructions for causing a processor to perform the method.

Although the above has been described with reference to the drawings and embodiments, it will be understood by those skilled in the art that various modifications and changes can be made to the embodiments without departing from the spirit of the embodiments described in the claims below. will be able

100: voice compression unit
102: self-encoder encoder unit feature extraction unit
104: quantization unit
200: self-encoder decoder unit
202: voice decoder

Claims (10)

a speech compression unit generating a quantized bit stream per frame using the original sound;
a feature extraction unit of the self-encoder encoder unit for extracting bottleneck features of the self-encoder encoder by inputting the bit string extracted from the speech compression unit; and
and a quantization unit that quantizes the bottleneck feature and converts it into a bit stream. According to claim 1,
and the quantization unit linearly quantizes the bit stream. According to claim 1,
The bottleneck feature is a value of a node that is generated in a bottleneck phenomenon in a node of the encoder of the self-encoder unit. According to claim 1,
The speech compression unit passes the original sound through the music sample to generate a quantized bit stream per frame. a self-encoder decoder unit for restoring a bit stream extracted from an encoding device of a transformed speech encoder; and
and a speech decoder for synthesizing speech by inputting the restored bit stream as an input. 6. The method of claim 5,
The speech decoder substitutes 1 for a value larger than a threshold value in the vector output from the self-encoder decoder unit, substitutes 0 for a value smaller than the threshold value, and collects synthesized speech using the substituted value for each frame. A decoder device for a transformed speech encoder. a speech compression unit generating a quantized bit stream per frame using the original sound;
a feature extraction unit of the self-encoder encoder unit for extracting bottleneck features of the self-encoder encoder by inputting the bit string extracted from the speech compression unit; and
an encoding device for a modified speech encoder including a quantizer that quantizes the bottleneck feature and converts it into a bit stream;
a self-encoder decoder unit for restoring the bit stream extracted by the encoding device; and
and a speech decoder for synthesizing speech by inputting the reconstructed bit stream as an input. generating a quantized bit stream per frame using the original sound;
extracting a bottleneck feature of a self-encoder encoder using the bit string as an input;
quantizing the bottleneck feature and converting it into a bit stream;
restoring the converted bit stream; and
synthesizing speech using the restored bit string as an input;
Encoding and decoding method of a transformed speech encoder comprising a. As a computer-readable recording medium storing a computer program,
The computer program is
generating a quantized bit stream per frame using the original sound;
extracting a bottleneck feature of a self-encoder encoder using the bit string as an input;
quantizing the bottleneck feature and converting it into a bit stream;
restoring the converted bit stream; and
A computer-readable recording medium comprising instructions for causing a processor to perform the step of synthesizing speech by inputting the restored bit string as an input. As a computer program stored in a computer-readable recording medium,
The computer program is
generating a quantized bit stream per frame using the original sound;
extracting a bottleneck feature of a self-encoder encoder using the bit string as an input;
quantizing the bottleneck feature and converting it into a bit stream;
restoring the converted bit stream; and
A computer program comprising instructions for causing the processor to perform the step of synthesizing speech by inputting the restored bit string as an input.

Images