# How to compress and restore voice messages

(57) Abstract:

The invention relates to telecommunications, and in particular to the field associated with the reduction of the redundancy of the transmitted information. The proposed method can be used to send voice messages via digital communication channels with speeds up to 4 kbit/s and can be classified as encoding form of the speech signal or means of direct encoding restore. The technical result is the development of a method for compressing and restoring voice messages that reduce the required bandwidth of a digital communication channel, which allows the management of telephone calls at low-speed digital communication channels. How to compress and restore voice messages is that discretizing continuous speech signal, quantum discrete samples, form the matrix of the quantized samples of the speech signal, generate a lot of unit and zero elements in the form of rectangular matrices, passed a lot of unit and zero elements of the communication channel, take it from a communication channel, form the matrix of reconstructed samples of the speech signal and convert the matrix uggling matrices provide an incomplete enumeration of all possible values of the elements of columns and rows of matrices, it is possible to impose restrictions on the structure of rectangular matrices and optimization of these matrices in order to reduce the amount of information you will need in the future to transmit via the communication channels. 7 Il. The invention relates to telecommunications, and in particular to the field associated with the reduction of the redundancy of the transmitted information. The proposed method can be used to send voice messages via digital communication channels with speeds up to 4 kbit/s and can be classified as encoding form of the speech signal or means of direct encoding restore.Known methods of encoding the shape of the speech signal, see, for example, the book: J. Cater the Computers, speech synthesizers, M.: Mir, 1985, S. 87-103, including the execution of three operations: the temporal discretization of analog signals, quantization and encoding (representation of the quantized discrete samples of the speech signal binary digits). The considered method is mainly defines a mechanism for encoding and decoding quantized discrete samples of the speech signal.There are also known methods of encoding discrete quantized samples of the speech signal is pulse code modulation, method of block coding from orthogonal transformation, see, for example, the book: M. C. Nazarov, Y. N. Petrov. Methods of digital processing and transmission of digital signals. - M.: Radio and communication, 1985, S. 142-161. The disadvantage of the above methods - analogues is relatively low informational efficiency, which is defined as achieving a certain quality of voice data recovery at a given speed transmission. In some ways analogous to an acceptable quality of voice data recovery is achieved at the transmission rate of more than 16 kbit/sAnalog is also the method described in UK patent 2280827 And IPC^{7}G 10 L 3/02 from 08.02.1995. The known method includes sampling a continuous signal, the quantization of discrete samples, forming a matrix of quantized samples of the speech signal, converting it to a digital mind using the American standard JPEG compression, the transmission of a digital stream over the communication channel, the digital stream from the communication channel, the recovery matrix of quantized samples of the speech signal from the digital stream using the JPEG standard and the inverse transform of the quantized samples in the continuous re which incorporates both the possibility of applying this method to conduct telephone conversations in digital communication channels.The closest in technical essence to the claimed method of the compression and recovery of voice messages is the method described in the patent of Russia 2152646 And IPC

^{7}G 10 L 3/02 2000Prototype method includes the discretization of the continuous speech signal, the quantization of discrete samples, forming a matrix of quantized samples of the speech signal size NN elements, the transformation matrix of the quantized samples of the speech signal size NN elements to digital mind by generating a random square matrix of quantized samples of size mm elements, multiple zero and identity elements in the form of a rectangular matrix of size Nm and mN elements, generate a random matrix of size Nm and mN elements, transformation matrix of size Nm and mN elements by dividing the elements of each row of the matrix of size Nm elements on the amount of units of the corresponding row and dividing the elements of each column of a matrix of size mN elements on the amount of units in the relevant column of matrix computations recovered samples of the speech signal size NN elements, calculate the sum of squared differences between the elements resulting from the matrix multiplication of the speech signal size NN elements, sequential inversion of the i-th element of the matrix N and Nm elements, matrix computations recovered samples of the speech signal size NN elements with inverted element, calculate the sum of squared differences between the elements of the matrix of reconstructed samples of the speech signal size NN elements and matrix elements of the quantized samples of the speech signal size NN elements, comparing the calculated errors and save the inverted values of the i-th element of the matrix mN and Nm elements, if the difference is greater than zero, and re-invert the i-th element of the matrix mN and Nm elements, if the difference is less than zero, transfer multiple zero and unit elements of a rectangular matrix of size m and mN elements in the communication channel, receiving the set of zero and unit elements of matrices of size Nm and mN elements of the communication channel, forming the matrix of reconstructed samples of the speech signal size NN elements of the transformation matrix recovered samples of the speech signal size NN elements in continuous speech signal.Prototype method allows to reduce the value of the time delay of the transmitted information to the amount at which you can conduct telephone conversations on low is more required bandwidth digital communication channel (4-8 Kbps), that limits the applicability of this method for conducting telephone negotiations on a low-speed digital communication channels.The aim of the invention is to develop a method of compression and recovery of voice messages that reduce the required bandwidth of a digital communication channel, which allows the management of telephone calls at low-speed digital communication channels.This goal is achieved by the fact that in the proposed method, compression and recovery of voice messages, pre -, is identical to the transmit and receive sides generate random square matrix of quantized samples of size mm elements, each element of which belongs to the range of the quantized discrete samples of the speech signal. Then discretizing continuous speech signal, quantum discrete samples, form the matrix of the quantized samples of the speech signal size NN elements. For forming the matrix of the quantized samples of the speech signal size NN elements every element a

_{j,i}where j=1,2,..,N; i=1,2,...,N assign quantized discrete value of the reference speech signal, the k-th number of which is determined with regard to ogolnych matrix of size Nm and mN elements, passed a lot of unit and zero elements of the communication channel, take it from a communication channel, form the matrix of reconstructed samples of the speech signal size NN elements and transform the matrix of reconstructed samples of the speech signal in a continuous speech signal.The proposed method differs from the prototype method for forming a variety of unit and zero elements in the form of a rectangular matrix of size Nm and mN elements on the transfer of pre-generated from a variety of unit and zero elements randomly odd columns of the matrix of size mN elements and odd rows of a matrix of size Nm elements, assign the elements of the even columns of a matrix of size mN elements and the even-numbered rows of the matrix of size Nm elements the values of the elements of the odd columns and rows of a matrix of size mN elements and matrix of size Nm elements, respectively. In this case, the elements of each even-numbered column matrix of size mN elements and each even row of the matrix of size Nm elements with numbers from 1 to m/2, assign the value of the element corresponding numbers from the previous column for a matrix of size mN elements and the previous row for a matrix of size Nm leerom Nm elements with numbers from m/2+1 to m, assign the value of the element corresponding number of the next column for a matrix of size mN of the elements and subsequent rows for a matrix of size Nm elements. After this transform matrix of size Nm and mN elements by dividing the elements of each row of the matrix of size Nm elements on the amount of units of the corresponding row and dividing the elements of each column of a matrix of size mN elements on the amount of units of the corresponding column, compute a matrix of reconstructed samples of the speech signal size NN elements by multiplying obtained after the conversion of rectangular matrix of size Nm elements identical with the previously formed on the transmitting and receiving sides of a random square matrix of quantized samples of size mm elements and obtained after the transformation matrix of size mN elements, calculate the sum of the squared differences between the elements obtained by multiplying the matrix of reconstructed samples of the speech signal size NN elements and the corresponding elements of the matrix of quantized samples of the speech signal size NN elements. Then invert each element of the odd-numbered columns of a matrix of size mN elements and Neher mN elements and the even-numbered rows of the matrix of size Nm elements. In this case, the elements of each even-numbered column matrix of size mN elements and each even row of the matrix of size Nm elements with a number from 1 to m/2 assigns the value of the element corresponding numbers from the previous column for a matrix of size mN elements and the previous row for a matrix of size Nm elements, and elements from m/2+l to m assigns the value of the element corresponding numbers in the next column for a matrix of size mN of the elements and subsequent rows for a matrix of size Nm elements. Transform matrix of size Nm and mN elements by dividing the elements of each row of the matrix of size Nm elements on the amount of units of the corresponding row and dividing the elements of each column of a matrix of size mN elements on the amount of units of the corresponding column, compute a matrix of reconstructed samples of the speech signal size NN elements by multiplying obtained after the conversion of rectangular matrix of size Nm elements identical with the previously formed on the transmitting and receiving sides of a random square matrix of quantized samples of size mm elements and obtained after the transformation matrix of size mN elements. Calculate the sum of the squares of the differences between the Ala of N elements and the corresponding elements of the matrix of quantized samples of the speech signal size NN elements and subtract this amount from the amounts received before inverting element of the sum of squared differences between the elements of the matrix of reconstructed samples of the speech signal size NN elements and matrix elements of the quantized samples of the speech signal size NN elements. In case of a positive difference with their values inverted elements, as in the case of a negative difference - re-invert. Then formed many zero and identity elements in the odd columns of the rectangular matrix of size Nm elements and odd rows of the rectangular matrix of size N elements is passed to the communication channel upon reception of the communication channel, restore missing even-numbered columns and rows of matrices of size Nm and mN elements, respectively. If the element number of each even-numbered column matrix of size mN elements and each even row of the matrix of size Nm elements has a value from 1 to m/2, then this element assigns the value of the element corresponding numbers from the previous column for a matrix of size mN elements and the previous row for a matrix of size Nm elements, if the element number of each even-numbered column matrix of size mN elements and each even row of the matrix of size Nm elements has a value from m/2+l to m, it is this element assigns the value of the element corresponding numbers in the next column for a matrix of size mN of the elements and subsequent rows for a matrix of size Nm elements.Featur the rum NN elements to send to each element of which requires 8 bits) to the representations of the speech signal as a product of three matrices, one of which is to transmit over the communication channel is not necessary (it is in advance formed on the transmitting and receiving sides), and the other two is converted into an integer matrix to send each element of which requires 1 bit) and their size is smaller than the original matrix of the quantized samples of the speech signal (Nm and mN elements, m<N). Moreover, the imposition of restrictions on the structure of these matrices allows us to transmit over the communication channel in a truncated form (only the odd-numbered columns or rows), which reduces the value of the required bandwidth to the amount at which you can conduct telephone conversations via low-speed digital communication channels.The analysis of the level of technology has allowed to establish that the analogues, characterized by a set of characteristics is identical for all features of the claimed technical solution is available, which indicates compliance of the claimed method the condition of patentability "novelty". Search results known solutions in this and related areas of technology in order to identify characteristics that match the distinctive features of the prototype of the characteristics of the claimed method, showed that they do not follow explicitly from the level of those who made the invention of the transformations on the achievement of the technical result. Therefore, the claimed invention meets the condition of patentability "inventive step".The claimed method is illustrated by drawings:

- Fig. 1 is a graph curve that describes the shape of the continuous speech signal;

- Fig.2 is an example of quantized discrete samples of the speech signal;

- Fig. 3 is an example of forming the matrix of the quantized samples of the speech signal size NN elements;

- Fig. 4 is a matrix presentation of the restored quantized samples of the speech signal as a product of three matrices;

- Fig.5 - structure of the matrix [Y]

_{Nxm}, [X]

_{mxN}and

- Fig.6 - the transformation matrix of the quantized samples of the speech signal to a digital mind and the inverse transform received from the communication channel digital stream into the matrix of reconstructed samples of the speech signal;

- Fig. 7 - consecutive optimization of the elements of the matrices [X]

_{mxN}and [Y]

^{T}

_{mxN}.Reducing the amount of information that must be sent over the communication channel per unit time, can be achieved by decreasing the sampling frequency continuous speech signal or decreasing the number of quantization levels, but the quality of recovery of speech decreases faster h is s messages while maintaining a good quality of recovery (high intelligibility and speech recognition) is explained by the following. There is the traditional approach, when to reduce the amount of information that must be sent over the communication channel in a unit of time, each of the encoded message (block of consecutive samples in a matrix of quantized samples of the speech signal) is its rating as a product of matrices of support vectors (basis) matrix of coefficients. The recipient, using the received coefficients and basis, performs restoration message source. Examples of this approach are well known: the discrete cosine transform, fast Fourier transform, transformation of karunen-Loev, Wavelet transform, and others. The use of these methods does not allow to achieve the required compression ratio with good quality restore speech. That is, the amount of information that must be sent over the communication channel per unit of time while maintaining good quality restore voice messages is large due to the large dimensions and capacity of the matrix of support vectors (basis).In the proposed method, the support vector (basis) is not transferred and formed on the receiving side of a random matrix and received from a communication channel integer to the of a random square matrix of quantized samples [IN]

_{mxm}can be made on the basis of a random numbers generator, for example, on the basis of the noise diode. To fulfill the requirements of the identity matrix [B]

_{mxm}the receiver is similar to the matrix of the transmitter before the beginning of each session the elements of the matrix [B]

_{mxm}can be generated on the transmission and transmitted over a digital communication channel at the receiving side, for example, as part of synchrophasing.The discretization of the continuous speech signal shown in Fig.1, is performed in accordance with the Nyquist theorem. In the proposed method, the selected common frequency sample rate continuous speech signal of 8 kHz.Then carry out the quantization of discrete samples. Quantization is performed on the basis of the methods described, for example, in the book: M. C. Nazarov, Y. N. Petrov. Methods of digital processing and transmission of digital signals. - M.: Radio and communication, 1985, S. 142-161.Next, on the basis of a set of quantized discrete samples of the speech signal are forming the matrix of the quantized samples of the speech signal [A]

_{NxN}. Forming a matrix of quantized samples of the speech signal [A]

_{NxN}it is shown in Fig.3. The matrix of quantized samples, forms,...,N; i=1,2,...,N assigned a quantized value of the reference speech signal, the k-th number of which is determined in accordance with the expression: k=j+N(i-1).To convert a matrix of quantized samples of the speech signal [A]

_{NxN}in order to reduce the amount of information transmitted over the communication channel, using an approach based on the representation of the matrix [A]

_{NxN}as a product of three matrices: a square matrix of size Nm elements (hereinafter denote it as [Y

_{pr}]

_{Nxm}), a random square matrix of quantized samples of size mm elements (hereinafter denote it as [B]

_{mxm}) and a rectangular matrix of size mN elements (hereinafter denote it as [X

_{pr}]

_{mxN}). Then, when the encoding matrix of the quantized samples of the speech signal [A]

_{NxN}the transfer must find the optimal matrix [Y

_{pr}]

_{Nxm}and [X

_{pr}]

_{mxN}that when the multiplication with the matrix [B]

_{mxm}form some matrix restored discrete samples of the speech signal size NN elements (hereinafter denote this matrix as closest to the specified criteria to the matrix of quantized discrete samples of the speech signal [A]

_{NxN}.Osipova mind. This is achieved by the fact that the elements of these matrices has the following limitations:

elements of the matrix [Y

_{pr}]

_{Nxm}and [X

_{pr}]

_{mxN}accept values in the range from zero to one;

- non-zero elements in each row of the matrix [Y

_{pr}]

_{Nxm}equal in amount to form a unit;

- non-zero elements of each column of the matrix [X

_{pr}]

_{mxN}equal in amount to form a unit.Under such restrictions, if the elements of each row of the matrix [Y

_{pr}]

_{Nxm}multiply by the number of nonzero elements in this row, it will obtain the matrix [Y]

_{Nxm}whose elements are defined only on the set of 1's and 0's. Similarly, if the elements of each column of the matrix [X

_{pr}]

_{mxN}multiply by the number of nonzero elements in the column, it will obtain the matrix [X]

_{mxN}whose elements are defined only on the set of 1 and 0.Since the search of the optimal matrix [Y

_{pr}]

_{Nxm}and [X

_{pr}]

_{mxN}carry out the incomplete enumeration of all possible values of the elements of columns and rows of matrices that can impose restrictions on the structure of the matrices and optimization of these matrices in order to reduce the amount of information is reduce optimal matrix [Y]

_{Nxm}and [X]

_{mxN}described in detail in Annex 1.Thus, the representation matrices of the quantized samples of the speech signal [A]

_{NxN}digital mind on the transfer carried out on the basis of multiple zero and identity elements in the form of a rectangular matrix of size Nm (matrix [Y]

_{Nxm}) and mN (matrix [X]

_{mxN}elements.After identifying the optimal matrix [Y

_{pr}]

_{Nxm}and [X

_{pr}]

_{mxN}transmit to the communication channel and not all values of the matrix [Y]

_{Nxm}and only values of the elements of the odd columns of the matrix and the values of elements of odd rows of the matrix [Y]

_{Nxm}. Receive from the communication channel values of the elements of the odd columns of the matrix and the values of elements of odd rows of the matrix [Y]

_{Nxm}, restore missing even-numbered columns and rows of the matrix [Y]

_{Nxm}and accordingly, if the element number of each even-numbered column of the matrix and each of the even rows of the matrix [Y]

_{Nxm}has a value from 1 to m/2, then this element assigns the value of the element corresponding numbers from the previous column for the matrix and the previous row for the matrix [Y]

_{Nxm}if the element number of each even-numbered column of the matrix and each even row of the matrix is EPA subsequent column for the matrix and subsequent rows for the matrix [Y]

_{Nxm}.After reception of the communication channel digital stream and the recovery matrix [Y]

_{Nxm}and [X]

_{mxN}they transform in matrix [Y

_{pr}]

_{Nxm}and [X

_{pr}]

_{mxN}respectively. The conversion is performed by dividing the elements of each row of the matrix [Y]

_{Nxm}for the amount of units of the corresponding row and dividing the elements of each column of the matrix [X]

_{mxN}for the amount of units of the corresponding column. Then form the matrix of reconstructed samples of the speech signal by multiplying the matrix [Y

_{pr}]

_{Nxm}and [X

_{pr}]

_{mxN}and the previously formed a square matrix of quantized discrete samples [B]

_{mxm}in accordance with the expression: and perform the inverse transform of the matrix of the restored quantized samples of the speech signal in a continuous speech signal. Clearly the matrix presentation of the restored discrete samples of the speech signal as a product of three matrices shown in Fig.4.To evaluate the effectiveness of the proposed method of compression and restore voice messages was conducted simulation on the PC. When encoding speech messages used 8-bit ADC. The size of the coding for the support of the transmitted information below time delay implemented in the method-prototype (in the way the prototype value of the time delay of the speech signal is 0,72 C. ). Delay of 0.28 MS makes it possible to maintain full-duplex phone conversations via low-speed communication channels. The size of the random square matrix of quantized discrete samples was 88 items. In the proposed method, high compression voice data was achieved due to the fact that formation at the reception matrix recovered samples of the speech signal [A]

_{NxN}in a digital communication channel, you need to pass the number of binary units defined by the size of the matrix [Y]

_{Nxm}and [X]

_{mxN}considering the fact that they should pass in a truncated (about 2 times). Thus the achievable compression ratio (reduction ratio to the desired transmission rate of the digital stream) can be found by the formula:

< / BR>

where L is the number of quantization levels of the discrete samples of the speech signal.When choosing N= 15, m=8 (L=256) were provided with a compression ratio of 14 times (transmission rate at the output of the encoder - 4.55 [kbps]). When choosing the size of the random matrix quantized samples of the speech signal 66 elements, the ratio of voice messages status is ALOS ratio signal/noise and amounted to about 9 [d]. When it restored it retains its naturalness and has good clarity. The analysis of computational complexity have shown that the complexity of the encoding/decoding of the proposed procedure (number of operations of multiplication, division, addition, subtraction) is proportional approximately to the value of m

^{2}. Therefore, the proposed method for the compression and recovery of speech can be implemented on modern processors signal processing. How to compress and restore voice messages, namely, that pre-generate identical to the transmit and receive sides of a random square matrix of quantized samples of size mm elements, each element of which belongs to the range of the quantized discrete samples of the speech signal, discretizing continuous speech signal, quantum discrete samples, form the matrix of the quantized samples of the speech signal size NN elements form the set of unit and zero elements in the form of a rectangular matrix of size Nm and mN elements that convey a lot of unit and zero elements of the communication channel, take it from a communication channel, form the matrix of reconstructed speech samples is iwny speech signal, characterized in that for the formation of multiple unit and zero elements in the form of a rectangular matrix of size Nm and mN elements on the transfer of pre-generated from a variety of unit and zero elements randomly odd columns of the matrix of size mN elements and odd rows of a matrix of size Nm elements, assign the elements of the even columns of a matrix of size mN elements and the even-numbered rows of the matrix of size Nm elements the values of the elements of the odd columns and rows of a matrix of size mN elements and matrix of size Nm elements, respectively, in this case, the elements of each even-numbered column matrix of size mN elements and each even row of the matrix of size Nm elements with numbers from 1 to m/2 assigns the value of the element corresponding numbers from the previous column for a matrix of size mN elements and the previous row for a matrix of size Nm elements, and the elements of each even-numbered column matrix of size mN elements and each even row of the matrix of size Nm elements with numbers from m/2+1 to m assigns the value of the element corresponding number of the next column for a matrix of size mN of the elements and subsequent rows for a matrix of size Nm elements, then convert the amount of units of the corresponding row and dividing the elements of each column of a matrix of size mN elements on the amount of units of the corresponding column, compute a matrix of reconstructed samples of the speech signal size NN elements by multiplying obtained after the conversion of rectangular matrix of size Nm elements identical with the previously formed on the transmitting and receiving sides of a random square matrix of quantized samples of size mm elements and obtained after the transformation matrix of size mN of elements, calculate the sum of the squared differences between the elements obtained by multiplying the matrix of reconstructed samples of the speech signal size NN elements and the corresponding elements of the matrix of quantized samples of the speech signal size NN elements, then invert each element of the odd-numbered columns of a matrix of size mN elements and odd rows of a matrix of size Nm elements, and simultaneously invert the elements of the even columns of a matrix of size mN elements and the even-numbered rows of the matrix of size Nm elements, and the elements of each even-numbered column matrix of size mN elements and each even row of the matrix of size Nm elements with a number from 1 to m/2 assigns the value of the element corresponding numbers from the previous column for a matrix of size mN elements and the previous is estoodeeva rooms of the next column for a matrix of size mN of the elements and subsequent rows for a matrix of size Nm elements, transform matrix of size Nm and mN elements by dividing the elements of each row of the matrix of size Nm elements on the amount of units of the corresponding row and dividing the elements of each column of a matrix of size mN elements on the amount of units of the corresponding column, compute a matrix of reconstructed samples of the speech signal size NN elements by multiplying obtained after the conversion of rectangular matrix of size Nm elements identical with the previously formed on the transmitting and receiving sides of a random square matrix of quantized samples of size mm elements and obtained after the transformation matrix of size mN elements, calculate the sum of the squared differences between the elements obtained by multiplying the matrix of reconstructed samples of the speech signal size NN elements and the corresponding elements of the matrix of quantized samples of the speech signal size NN elements and subtract this amount from the previously obtained sum of squared differences between the elements of the matrix of reconstructed samples of the speech signal size NN elements and matrix elements of the quantized samples of the speech signal size NN elements, in the case of a positive difference observed the strain set the zero and unit elements in the odd columns of the rectangular matrix of size Nm elements and odd rows of the rectangular matrix of size N elements are passed to the communication channel, and for forming the matrix of reconstructed samples of the speech signal of length N of elements restore missing even-numbered columns and rows of matrices of size Nm and mN elements, respectively, if the element number of each even-numbered column matrix of size mN elements and each even row of the matrix of size Nm elements has a value from 1 to m/2, then this element assigns the value of the element corresponding numbers from the previous column for a matrix of size mN elements and the previous row for a matrix of size Nm elements, if the item number of each even-numbered column matrix of size mN elements and each even row of the matrix of size Nm elements has a value from m/2+1 to m, then this element assigns the value of the element corresponding numbers in the next column for a matrix of size mN of the elements and subsequent rows for a matrix of size Nm elements.

**Same patents:**

_{s}

FIELD: electric communication, namely systems for data transmitting by means of digital communication lines.

SUBSTANCE: method comprises steps of preliminarily, at reception and transmission forming R matrices of allowed vectors, each matrix has dimension m2 x m1 of unit and zero elements; then from unidimensional analog speech signal forming initial matrix of N x N elements; converting received matrix to digital one; forming rectangular matrices with dimensions N x m and m x N being digital representation of initial matrix from elements of lines of permitted vectors; transmitting elements of those rectangular matrices through digital communication circuit; correcting errors at transmission side on base of testing matching of element groups of received rectangular matrices to line elements of preliminarily formed matrices of permitted vectors; then performing inverse operations for decompacting speech messages. Method is especially suitable for telephone calls by means of digital communication systems at rate 6 - 16 k bit/s.

EFFECT: possibility for correcting errors occurred in transmitted digital trains by action of unstable parameters of communication systems and realizing telephone calls by means of low-speed digital communication lines.

5 cl, 20 dwg