RU2152646C1 - Method for compression and decompression of voice messages - Google Patents

Abstract

FIELD: telecommunications, in particular, elimination of redundancy in transmitted information. SUBSTANCE: method involves preliminary generation of random square matrix of voice signal samples, which is identical for transmission and reception and contains m*m elements. Then, method involves generation of matrix of voice signal samples, which contains N*N elements and is produced from linear voice signal array, converting said matrix to digital representation by means of its representation as product of three matrices: rectangular N*m matrix, preset random square m*m matrix of samples, and rectangular m*N matrix. Only elements of rectangular matrices are transmitted to digital communication channel. Voice decompression runs in reverse order. Rectangular matrices, which are received from digital communication channel and preset random square matrix, produce N*N matrix of decompressed voice signal samples. Then, the latter matrix is used for generation of linear continuous voice signal. Method is designed for running telephone talks by digital communication channels with rate of 4-16 kbps. EFFECT: decreased information transmission delay caused by compression, increased compression ratio. 7 dwg

Description

 The invention relates to the field of telecommunications, and in particular to the field associated with reducing the redundancy of transmitted information. The proposed method can be used to transmit voice messages over digital communication channels with a speed of up to 4 kbit / s and can be assigned to the class of methods for encoding the shape of the speech signal or methods of direct encoding-recovery.

 Known methods for encoding the shape of a speech signal, see, for example, the book: J. Keyter Computers-speech synthesizers.- M .: Mir, 1985, p.87-103, including the performance of three operations: temporal sampling of analog signals, their quantization and encoding (representation of quantized discrete samples of the speech signal in binary digits). The considered method mainly determines the encoding and decoding mechanism of quantized discrete samples of the speech signal.

 There are also known methods of encoding discrete quantized samples of a speech signal based on delta modulation, adaptive delta modulation, pulse-code modulation, differential pulse-code modulation, orthogonal transform block coding method, see, for example, book: M. V. Nazarov , Yu.N. Petrov Methods of digital processing and transmission of digital signals. - M .: Radio and communications, 1985, p. 142-161. The disadvantage of the above methods - analogues is the relatively low information efficiency, which is understood as the achievement of a good quality of voice information recovery is achieved at a transmission speed of more than 16 kbit / s.

The closest in technical essence to the claimed method of compression and restoration of voice messages is the method described in UK patent N2280827 A, MNC ⁶ G 10 L 3/02 of 08 02 1995. The prototype method includes sampling a continuous signal, quantization of discrete samples, formation matrices of quantized samples of a speech signal, its conversion to digital form using the American JPEG compression standard, transmitting a digital stream through a communication channel, receiving a digital stream from a communication channel, restoring matrices quantized samples of a speech signal from a digital stream using the JPEG standard and the inverse of the quantized samples converted to a continuous speech signal. However, the disadvantage of the prototype method is the large time delay of the speech signal, which excludes the possibility of using this method for conducting telephone conversations on digital communication channels.

 The aim of the invention is to develop a method of compression and restoration of voice messages, which reduces the time delay of transmitted information with a high degree of compression, in which it is possible to conduct telephone conversations on low-speed digital communication channels.

This goal is achieved by the fact that in the known method of compressing and restoring voice messages, including sampling a continuous speech signal, quantizing discrete samples, forming a matrix of quantized samples of a speech signal of size N • N elements, converting it to digital form, transmitting a digital stream through a communication channel, receiving it from a communication channel, restoring a matrix of quantized samples of a speech signal of size N • N elements, and inverting the quantized discrete samples to continuous A speech signal, previously identical in transmission and reception, generates a random square matrix of quantized discrete samples of size m • m elements, each element of which belongs to the range of quantized discrete samples of the speech signal. Then, to transform the matrix of quantized samples of a speech signal of size N • N elements to a digital form in transmission, a set of zero and single elements is formed in the form of rectangular matrices of size N • m and m • N elements and transmit them to the communication channel. These matrices are received from the communication channel. Then they are transformed by dividing the elements of each row of a rectangular matrix of size N • m elements by the sum of units of the corresponding row and dividing the elements of each column of a rectangular matrix of size m • N elements by the sum of units of the corresponding column. After that, the matrix of quantized samples of the speech signal of size N • N elements is restored by sequentially multiplying the rectangular matrix of size N • m elements obtained after conversion, the random square matrix of quantized discrete samples of size m • m elements and obtained after the transformation of a rectangular matrix of size m • N elements. Moreover, to form a matrix of quantized samples of a speech signal of size N • N elements, each of its elements is A _{j, i} , where j = 1,2, ..., N; i = 1,2, ..., N assign the quantized value of the discrete count of the speech signal, the k-th number of which is determined in accordance with the expression: k = j + N • (i-1).

 To form a set of zero and single elements in the form of rectangular matrices of size N • m and m • N elements in the transmission, random rectangular matrices are preliminarily generated from unit and zero elements. Then they are transformed. After that, a matrix of size N • N elements is calculated by sequentially multiplying the obtained rectangular square matrix of size N • m elements, a random square matrix of quantized discrete samples of size m • m, and obtained after the transformation of a rectangular matrix of size m • N elements. Next, the sum of the squared differences between the elements of the matrix obtained by multiplying the size of N • N elements and the elements of the matrix of quantized samples of the speech signal of the size N • N elements is calculated. Then, each element is successively inverted of each element of rectangular matrices of size N • m and m • N elements previously randomly generated in a random manner, while transforming them, and then the rectangular matrix of size N • m elements obtained after the conversion, the random square matrix of quantized discrete samples of size m are multiplied • m elements and a rectangular matrix of m • N elements obtained after the transformation. After that, the sum of the squared differences between the elements of the matrix obtained by multiplying the size of N • N elements and the elements of the matrix of quantized samples of the speech signal of the size N • N elements is calculated. This amount is subtracted from the similar amount obtained in the previous step, and in case of a positive difference, the inverted value of the element is stored, otherwise it is inverted again.

 To convert random rectangular matrices of unit and zero elements, the number of elements of each row of a rectangular matrix of size N • m elements and each column of a rectangular matrix of size m • N elements is divided by the sum of units in this row and this column, respectively.

 The specified new set of essential features allows to reduce the value of the time delay of the transmitted information to a value at which it is possible to conduct telephone conversations via low-speed digital communication channels.

The claimed method is illustrated by drawings:
- Figure 1 Graph of a curve describing the shape of a continuous speech signal;
- Figure 2 Example of quantized discrete samples of a speech signal;
- Figure 3 An example of the formation of a matrix of quantized samples of a speech signal with a size of N • N elements;
- Fig. 4 Presentation of a matrix of reconstructed quantized samples of a speech signal in the form of a product of three matrices;
- Figure 5 The structure of the matrices [Y] _Nxm , [X] _mxN and [Y _pr ] _Nxm , [X _pr ] _mxN ;
- Fig.6 Conversion of the matrix of quantized samples of the speech signal to digital form and the inverse transformation of the digital stream received from the communication channel into the matrix of restored samples of the speech signal.

- Fig.7 Sequential optimization of the elements of the matrices [X] _mxN and [Y] ^T _mxN .

) наиболее близкую по заданному критерию к матрице квантованных дискретных отсчетов речевого сигнала [A]_NxN. После определения оптимальных матриц [Y_pr]_Nxm и [X_pr]_mxN их передают в канал связи, принимают из канала связи и формируют на приеме матрицу восстановленных дискретных отсчетов речевого сигналу

на основе выполнения перемножения: [Y_pr]_Nxm • [B]_mxm • [X_pr]_mxN. Наглядно представление матрицы восстановленных дискретных отсчетов речевого сигнала

в виде произведения трех матриц представлен на фиг. 4. При этом в качестве меры близости матриц [A]_NxN и

выбирают квадратическую ошибку, определяемую по формуле

Особенностью матриц [Y_pr]_Nxm и [X_pr]_mxN является то, что они могут быть легко приведены к цифровому виду. Это достигается тем, что на элементы этих матриц накладываются следующие ограничения:
- элементы матриц [Y_pr]_Nxm и [X_pr]_mxN принимают значения в диапазоне от нуля до единицы;
- ненулевые элементы каждой строки матрицы [Y_pr]_Nxm равны между собой и в сумме образуют единицу;
- ненулевые элементы каждого столбца матрицы [X_pr]_mxN равны между собой и в сумме образуют единицу.The ability to implement the claimed method of compression and restoration of voice messages is explained as follows. The proposed method for compressing and reconstructing a speech signal is based on an approach based on the representation of an encoded block of samples represented by a matrix of quantized samples of a speech signal (hereinafter, denote it as [A] _NxN ) in the form of a product of three matrices: a rectangular matrix of size N • m elements (hereinafter denoted it as [Y _{pr] Nxm),} the random quantized samples of a square matrix size m • m elements (hereinafter denoted it as [B]) _mxm square matrix, and the size m • N elements (henceforth denoted m it as [X _{pr] mxN).} It is believed that the matrix [B] _{mxm is} known in transmission and in reception and does not change during the transmission / reception of information. Then, when encoding the matrix of quantized samples of the speech signal [A] _NxN on the transmission, it is necessary to find such optimal matrices [Y _pr ] _Nxm and [X _pr ] _mxN that, when multiplied with the matrix [B] _mxm, form some matrix of reconstructed discrete samples of the speech signal of size N • N elements (hereinafter, we denote this matrix as

) closest to a given criterion to the matrix of quantized discrete samples of the speech signal [A] _NxN . After determining the optimal matrices [Y _pr ] _Nxm and [X _pr ] _mxN, they are transmitted to the communication channel, received from the communication channel, and a matrix of reconstructed discrete samples of the speech signal is generated at the reception

based on the execution of the multiplication: [Y _pr ] _Nxm • [B] _mxm • [X _pr ] _mxN . Visual representation of the matrix of reconstructed discrete samples of the speech signal

as a product of three matrices is shown in FIG. 4. Moreover, as a measure of proximity of the matrices [A] _NxN and

choose the quadratic error, determined by the formula

A feature of the matrices [Y _pr ] _Nxm and [X _pr ] _mxN is that they can be easily digitized. This is achieved by the following restrictions on the elements of these matrices:
- the elements of the matrices [Y _pr ] _Nxm and [X _pr ] _mxN take values in the range from zero to unity;
- nonzero elements of each row of the matrix [Y _pr ] _Nxm are equal to each other and in total form one;
- nonzero elements of each column of the matrix [X _pr ] _mxN are equal to each other and form one in total.

Under such restrictions, if the elements of each row of the matrix [Y _pr ] _{Nxm are} multiplied by the sum of its nonzero elements, then the matrix [Y] _Nxm will be obtained, the elements of which are defined only on the set "1" and "0". Similarly, if the elements of each column of the matrix [X _pr ] _{mxN are} multiplied by the sum of its nonzero elements, then the matrix [X] _mxN will be obtained, the elements of which are defined only on the set "1" and "0". Thus, the matrix of quantized samples of the speech signal [A] _{NxN is presented} to the digital form in the transmission based on the formation of a set of zero and single elements in the form of rectangular matrices of size N • m (matrix [Y] _Nxm ) and m • N (matrix [X] ] _mxN ) elements. Then, the matrices [Y] _Nxm and [X] _mxN are transmitted to the communication channel and received from the communication channel. Next, the matrix [Y] _{Nxm is transformed} into the matrix [Y _pr ] _Nxm by dividing the elements of each row of the matrix [Y] _Nxm received from the communication channel by the sum of the units of the corresponding row and the matrix [X] _{mxN is transformed} into the matrix [X _pr ] _mxN by dividing the elements each column received from the communication channel of the matrix [X] _mxN for the sum of the units of the corresponding column. Then, the matrix of quantized samples of the speech signal of size N • N elements is restored by successively multiplying the rectangular matrix [Y _pr ] _Nxm obtained after the transformation, the random square matrix of quantized samples of size m • m elements [B] _mxm and the rectangular matrix [X _pr ] obtained after the transformation _mxN . This is illustrated in FIG. 5.

Sampling the continuous speech signal of FIG. 1, are performed in accordance with the Kotelnikov theorem. In the proposed method, a conventional sampling rate of a continuous speech signal of 8 kHz is selected. Then quantize discrete samples. Quantization is carried out on the basis of the methods described, for example, in the book: M.V. Nazarov, Yu.N. Petrov Methods of digital processing and transmission of digital signals. - M .: Radio and communications, 1985, p. 142-161. Then, based on the set of quantized discrete samples of the speech signal, a matrix of quantized samples of the speech signal [A] _{NxN is formed} . The formation of a matrix of quantized samples of the speech signal [A] _{NxN is} shown in _Fig.3 . The matrix of quantized samples is formed from N ² quantized discrete samples of the speech signal, each element of which A _ji , where i = 1,2, ..., N; r = 1,2, ..., N assign a quantized value of the count of the speech signal, the k-th number of which is determined in accordance with the expression: k = j + N • (i-1).

The formation of a random square matrix of quantized discrete samples [B] _mxm can be performed based on a random number sensor, for example, based on a noise diode. To fulfill the identity requirement of the receiver matrix [B] _{mxm to} a similar transmitter matrix, before the start of each communication session, the elements of the [B] _mxm matrix can be generated in transmission and transmitted over the digital communication channel to the receiving side, for example, as part of a sync packet.

To convert the matrix of quantized samples of the speech signal [A] _NxN to a digital form in transmission, a set of zero and single elements is formed in the form of matrices [Y] _Nxm and [X] _mxN and they are transmitted to a digital communication channel. The procedure that implements a search on the transmission of optimal matrices [Y] _Nxm and [X] _mxN is described in detail in Appendix 1.

в соответствии с выражением:

и выполняют обратное преобразование из матрицы восстановленных квантованных отсчетов речевого сигнала

в непрерывный речевой сигнал.After receiving the matrices [Y] _Nxm and [X] _mxN from the digital communication channel, they are transformed into the matrices [Y _pr ] _Nxm and [X _pr ] _mxN, respectively. The conversion is performed by dividing the elements of each row of the matrix [Y] _Nxm by the sum of the units of the corresponding row and dividing the elements of each column of the matrix [X] _mxN by the sum of the units of the corresponding column. After that, the matrix of quantized samples of the speech signal is restored

in accordance with the expression:

and perform the inverse transformation from the matrix of restored quantized samples of the speech signal

into a continuous speech signal.

где L - число уровней квантования дискретных отсчетов речевого сигнала.To assess the effectiveness of the proposed method of compression and restoration of voice messages, simulation was carried out on a PC. When encoding voice messages, an 8-bit ADC was used. At the same time, the size of the encoded block was 24 • 24 elements, which ensured a delay in speech transmission of 0.72 ms. This value of the time delay of the transmitted information is significantly lower than the time delay implemented in the prototype method (in the prototype method, the value of the time delay of the speech signal is 6 s). At the same time, a delay of 0.72 ms allows for duplex telephone conversations over low-speed communication channels. The size of the random square matrix of quantized discrete samples was 12 • 12 elements. In the proposed method, a high degree of compression of speech information was achieved due to the fact that in order to restore the matrix of quantized samples of the speech signal [A] _NxN at the _reception , it is necessary to transfer the number of binary units determined by the sizes of the matrices [Y] _Nxm and [X] _mxN . In this case, the achieved compression ratio (reduction coefficient of the required transmission rate of the digital stream) can be determined by the formula:

where L is the number of quantization levels of discrete samples of the speech signal.

When choosing N = 24, m = 12 (L = 256), a compression ratio of 8 times was provided (the transmission rate at the encoder output was 8 [kbit / s]). When choosing the size of a random square matrix of quantized discrete samples of a speech signal of 6 • 6 elements, the compression ratio of voice messages was 16 times (the transmission speed at the encoder output was 4 (kbit / s)). The quality of the restored speech at the reception was estimated by the signal-to-noise ratio and amounted to about 14 [db]. At the same time, the restored speech retains its naturalness, naturalness and has good intelligibility. Analysis of computational complexity showed that the complexity of encoding / decoding the proposed procedure (quantitative of multiplication, division, addition, subtraction) is proportional to the value of about ² m. Therefore, the proposed method for compression and decompression of speech may be realized on modern signal processors.

Claims (3)

 1. The method of compression and restoration of voice messages, which consists in sampling a continuous signal, quantizing discrete samples, forming a matrix of quantized samples of a speech signal of size N • N elements, converting it to digital form, transmitting a digital stream through a communication channel, receiving a digital stream from a communication channel restoring the matrix of quantized samples of a speech signal of size N • N elements and the inverse transformation of quantized discrete samples into a continuous speech signal, characterized in that a random square matrix of quantized samples of size m • m elements, each element of which belongs to the range of quantized discrete samples of a speech signal, is generated previously identically in transmission and reception, and for the transformation of a matrix of quantized samples of a speech signal of size N • N elements to a digital form in transmission, a set zero and single elements in the form of rectangular matrices of size N • m and m • N elements, moreover, to form a set of zero and single elements in the form of a rectangular of oly matrices of size N • m and m • N elements in the transmission, random rectangular matrices are preliminarily generated from unit and zero elements, they are converted, a matrix of size N • N elements is calculated by multiplying the rectangular matrices obtained after the transformation with a square matrix previously identically formed in transmission and reception quantized samples of size m • m elements, calculate the sum of the squared differences between the elements obtained by multiplying a matrix of size N • N and the elements of the quantum matrix samples of a speech signal of size N • N elements, then each element is sequentially inverted of randomly generated rectangular matrices of size N • m and m • N elements, they are converted, and then rectangular matrices of size N • m and m • N elements are multiplied from previously formed by a matrix of quantized samples of size m • m elements, calculate the sum of the squares of the differences between the elements obtained by multiplying a matrix of size N • N and elements of the matrix of quantized samples of a speech signal a, subtract this amount from the same amount obtained in the previous step and, in the case of a positive difference, save the inverted value of the element, and otherwise, perform its repeated inversion, then the formed sets of zero and single elements in the form of rectangular matrices of size N • m and m • N are transmitted to the communication channel, received from the communication channel, converted by dividing the elements of each row of the matrix N • m by the sum of the units of the corresponding row and dividing the elements of each column of the matrix of size m • N by the sum of one corresponding column, then reduced matrix of quantized speech signal samples size N • N elements by multiplying obtained after conversion rectangular matrix size N • m and m • N elements and previously formed square matrix of quantized discrete sample size m • m elements.  2. The method according to claim 1, characterized in that for the formation of a matrix of quantized samples of a speech signal of size N • N elements to each of its elements Aj, i, where j = 1, 2, ..., N; j = 1, 2, ..., N is assigned the quantized value of the discrete count of the speech signal, the kth number of which is determined in accordance with the expression: k = j + N • (i -1).  3. The method according to claim 1, characterized in that for converting random rectangular matrices of unit and zero elements, the number of elements of each row of the matrix N • m and each column of the matrix of size m • N is divided by the sum of units in this row and this column, respectively.

Images