RU2431892C2 - Parameter decoding device, parameter encoding device and parameter decoding method - Google Patents

Abstract

FIELD: information technology.

SUBSTANCE: parameter decoding device includes a prediction residue decoding unit which finds a quantised prediction residue based on encoded information included in a current frame subject to decoding, and a parameter decoding unit which decodes a parameter based on said quantised prediction residue, wherein said prediction residue decoding unit, when said current frame is erased, finds a current-frame quantised prediction residue from a weighted linear sum of a parameter decoded in the past and a future-frame quantised prediction residue.

EFFECT: high efficiency of frame concealment in order to prevent quality deterioration.

7 cl, 27 dwg

Description

FIELD OF THE INVENTION

The present invention relates to a parameter encoding device that encodes a parameter using a prediction device, and to a parameter decoding device and a method for decoding parameters that decode an encoded parameter.

State of the art

Regarding ITU-T recommendation G.729, 3GPPAMR, or a similar speech codec, some of the parameters obtained by analyzing the speech signal are quantized using a prediction quantization method based on a moving average (MA) prediction model (Patent Document 1, Non-Patent document 1, non-patent document 2). The MA-type prediction quantizer is a model that predicts the current parameter that is being quantized using the linear sum of the previous quantized prediction residuals, and a code-excited linear prediction codec (CELP) is used to predict the spectral line frequency parameter (LSL, LSF ) and power parameter.

As for the MA-type prediction quantizer, since the prediction is based on the weighted linear sum of the quantized prediction residuals in a finite number of previous frames, even if there is an error in the transmission path of the quantized information, its influence is limited by a finite number of frames. On the other hand, using an autoregressive (AR) type prediction quantization tool that uses past decoded parameters recursively, although high prediction and quantization efficiency can generally be obtained, the effect of the remainder continues for a long period. Therefore, the means for quantizing the parameter with the prediction of the MA type can provide higher error resistance than the means for quantizing the parameter with the prediction of the AR type, and it is used, in particular, in a speech codec for mobile communication.

Methods of masking a parameter that will be used when a frame is lost (deleted) on the decoding side has been studied for some time. In general, masking is performed using the parameter of the frame located in front of the remote frame, instead of the parameter of the remote frame. However, in the case of the LSL parameter, the parameters before the deleted frame are gradually changed by gradually approaching the average LSL value, or they are gradually weakened in the case of the power parameter.

This method is also commonly used in a quantizer using an MA-type predictor. In the case of the NLS parameter, processing is performed to update the state of the MA-type prediction device by generating a quantized prediction residual so that the parameter generated in the masked frame is decoded (Non-Patent Document 1), and in the case of the power parameter, processing is performed to update the state of the device MA-type predictions using the result of decreasing the average of the previous quantized prediction residuals in a fixed percentage (Patent Document 2, Not atentny Document 1).

There is also a method by which the parameter of the deleted frame is interpolated after receiving information of the reconstructed frame (normal frame) that follows the deleted frame. For example, Patent Document 3 proposes a method by which pitch gain interpolation is performed and adaptive codebook contents are regenerated.

Patent Document 1: Japanese Patent Application Laid-Open No. HEI 6-175695

Patent Document 2: Japanese Patent Application Laid-Open No. HEI 9-120297

Patent Document 3: Japanese Patent Application Laid-Open No. 2002-328700

Non-Patent Document 1: ITU-T Recommendation G.729

Non-Patent Document 2: 3GPP TS 26.091

SUMMARY OF THE INVENTION

Problems to be Solved by the Invention

The method by which the remote frame parameter is interpolated is used when the prediction quantization is not performed, but when the prediction quantization is performed, even if the encoded information is received correctly in the frame located immediately after the deleted frame, the prediction device is affected by an error in the immediately previous frame, and it cannot get the correct decoded result, and therefore this method is generally not used.

Thus, with the parameter quantization device that uses the conventional MA type prediction device, masking processing of the remote frame parameter is not performed by the interpolation method, and therefore, for example, sound loss due to excessive attenuation of the power parameter may occur, causing subjective quality to deteriorate .

When the prediction quantization is performed, a possible method is to decode the parameter simply by interpolating the quantized decoded prediction residuals, but whereas the decoded parameter varies slightly from one frame to another frame due to the determination of a weighted moving average, even if the quantized decoded prediction remainder changes very much, with this method, the decoded parameter also changes in accordance with the changes in the quantized decoded remainder and predictions, so that when changes are decoded quantized prediction residue great, subjective quality deteriorates.

The present invention is embodied in view of the problems described above, and an object of the present invention is to provide a parameter decoding device, a parameter encoding device and a parameter decoding method that enables parameter masking processing to suppress deterioration of subjective quality when performing prediction quantization.

Problem Solver

The parameter decoding apparatus of the present invention uses a configuration having a prediction residual decoding node that finds a quantized prediction residual based on encoded information that includes the current frame to be decoded, and a parameter decoding node that decodes a parameter based on the quantized prediction residual; moreover, the decoding residual decoding node, when the current frame is deleted, finds the quantized prediction residual of the current frame from the weighted linear sum of the parameter decoded in the past and the quantized prediction residual of the future frame.

The parameter encoding apparatus of the present invention uses a configuration having: an analysis unit that analyzes an input signal and finds an analyzed parameter; a coding unit that predicts the analyzed parameter using the prediction coefficient, and obtains a quantized parameter using the quantized prediction remainder obtained by quantizing the prediction remainder and the prediction coefficient; the masking unit of the previous frame, which stores a plurality of sets of weighting coefficients, finds the weighted sum using the sets of weighting coefficients for the quantized prediction residual of the current frame, the quantized prediction residual two frames back and the quantized parameter two frames back, and finds the set of quantized parameters of the previous frame using the weighted amount; and a determination node that compares the set of quantized parameters one frame back found using the masking unit of the previous frame, and the analyzed parameter found using the analysis node of the previous frame selects one of the quantized parameters of the previous frame and selects and encodes a set of weighting coefficients corresponding to the selected quantized parameter of the previous frame.

The method for decoding parameters of the present invention uses a method having a decoding of a prediction residual consisting of detecting a quantized prediction residual based on encoded information that includes the current frame to be decoded, and a parameter decoding step consisting of decoding a parameter based on the quantized residual predictions; moreover, at the stage of decoding the prediction residue, when the current frame is deleted, the quantized prediction residue of the current frame is found from the weighted linear sum of the parameter decoded in the past and the quantized prediction residue of the future frame.

Advantageous Effect of the Invention

According to the present invention, when the current frame is deleted when the prediction quantization is not performed, parameter masking processing can be performed to eliminate subjective quality degradation by finding the quantized prediction residual of the current frame from the weighted linear sum of the quantized prediction residues of the past frame and the quantized prediction residues of the future frame.

BRIEF DESCRIPTION OF THE DRAWINGS

1 is a block diagram showing a basic configuration of a speech decoding apparatus according to Embodiment 1 of the present invention;

FIG. 2 is a drawing showing an internal configuration of an LPC decoding assembly of a speech decoding apparatus according to Embodiment 1 of the present invention; FIG.

figure 3 is a drawing that shows the internal configuration of the decoding node of the code vector shown in figure 2;

4 is a drawing showing an example of a result of performing conventional processing when there is no deleted frame;

5 is a drawing showing an example of a result of performing masking processing of this embodiment;

6 is a drawing showing an example of a result of performing conventional masking processing;

7 is a drawing showing an example of a result of performing conventional masking processing;

Fig. 8 is a block diagram showing a basic configuration of a speech decoding apparatus according to Embodiment 2 of the present invention;

Fig.9 is a block diagram showing the internal configuration of the LPC decoding node shown in Fig.8;

figure 10 is a block diagram showing the internal configuration of the decoding unit of the code vector shown in figure 9;

11 is a block diagram showing a basic configuration of a speech decoding apparatus according to Embodiment 3 of the present invention;

FIG. 12 is a block diagram showing the internal configuration of the LPC decoding assembly shown in FIG. 11;

Fig.13 is a block diagram showing the internal configuration of the decoding unit of the code vector shown in Fig.12;

FIG. 14 is a block diagram showing an internal configuration of a gain decoding unit shown in FIG. 1;

FIG. 15 is a block diagram showing an internal configuration of a prediction residual decoding unit shown in FIG. 14;

FIG. 16 is a block diagram showing an internal configuration of a quantized prediction residual generation unit of a subframe shown in FIG. 15;

FIG. 17 is a block diagram showing a basic configuration of a speech encoding apparatus according to Embodiment 5 of the present invention; FIG.

Fig. 18 is a block diagram showing a configuration of a voice transmission apparatus and a speech reception apparatus that configure a voice transmission system according to Embodiment 6 of the present invention;

Fig. 19 is a drawing showing an internal configuration of an LPC decoding unit of a speech decoding apparatus according to Embodiment 7 of the present invention;

FIG. 20 is a drawing showing an internal configuration of a code vector decoding assembly shown in FIG. 19;

21 is a block diagram showing a basic configuration of a speech decoding apparatus according to Embodiment 8 of the present invention;

FIG. 22 is a drawing showing an internal configuration of an LPC decoding assembly of a speech decoding apparatus according to Embodiment 8 of the present invention; FIG.

Fig.23 is a drawing that shows the internal configuration of the decoding unit of the code vector in Fig.22;

24 is a drawing showing an internal configuration of an LPC decoding assembly of a speech decoding apparatus according to Embodiment 9 of the present invention;

Fig - drawing, which shows the internal configuration of the decoding node of the code vector shown in Fig; and

FIG. 26 is a block diagram showing a basic configuration of a speech decoding apparatus according to Embodiment 10 of the present invention.

Best Mode for Carrying Out the Invention

Embodiments of the present invention will now be described in detail with reference to the accompanying drawings. In the following embodiments, examples are described of cases in which the parameter decoding apparatus and the parameter encoding apparatus of the present invention are applied to a speech decoding apparatus and a CELP type speech encoding apparatus, respectively.

(Embodiment 1)

1 is a block diagram showing a basic configuration of a speech decoding apparatus according to Embodiment 1 of the present invention. In the speech decoding apparatus 100 shown in FIG. 1, encoded information transmitted from an encoding apparatus (not shown) is divided into a constant codebook code F _{n + 1} ; the adaptive codebook code A _{n + 1} , the gain code G _{n + 1} and the LPC code L _{n + 1} (linear prediction coefficients) using the demultiplexing unit 101. Separately, a frame deletion code B _{n + 1 is} entered into the speech decoding apparatus 100. In this case, the subscript n of each code indicates the number of the frame to be decoded. Those. encoded information is separated in the (n + 1) th frame (hereinafter referred to as the “next frame”) after the nth frame subjected to decoding (hereinafter referred to as the “current frame”).

The constant codebook code F _{n + 1} is entered into the constant codebook vector (VPCK) decoding unit 102, the adaptive codebook code A _{n + 1} is entered into the adaptive codebook vector (VACC) decoding unit 103, the gain code G _{n + 1} is to the gain decoding section 104, and the LPC code L _{n + 1 to the} LPC decoding section 105. The frame deletion code B _{n + 1} is input to the HCC decoding section 102, to the HACC decoding section 103, to the gain decoding section 104 and to the LPC decoding section 105.

The CDPC decoding section 102 generates a constant codebook vector using a constant codebook code F _n if the frame deletion code B _n indicates that "the nth frame is a normal frame" and generates a constant codebook vector by processing the deletion masking if the frame deletion code B _n indicates that “the nth frame is the deleted frame”. The generated constant codebook vector is input to a gain decoding unit 104 and to an amplifier 106.

The HACC decoding section 103 generates an adaptive codebook vector using an adaptive codebook code if the frame deletion code B _n indicates that “the nth frame is a normal frame” and generates an adaptive codebook vector by frame deletion mask processing if the code is B _n frame deletion indicates that "the nth frame is a deleted frame." The generated adaptive codebook vector is input to amplifier 107.

The gain decoding section 104 generates a constant codebook gain and an adaptive codebook gain using the gain code G _n and a constant codebook vector if the frame deletion code B _n indicates that “the nth frame is a normal frame”, and generates a constant codebook gain and an adaptive codebook gain by deletion masking processing if the deletion code B _n indicates that “the nth frame is the deleted frame”. The generated constant codebook gain is input to the amplifier 106, and the generated adaptive codebook gain is input to the amplifier 107.

The KLP decoding unit 105 decodes the KLP parameter using the KLP code L _n if the frame deletion code B _n indicates that “the nth frame is a normal frame” and decodes the KLP parameter by the frame deletion mask processing if the frame deletion code B _n indicates that "the nth frame is the deleted frame." The decoded CLP parameter is introduced into the CLP synthesis unit 109. Details of the LPC decoding unit 105 will be given later in this paper.

The amplifier 106 multiplies the constant codebook gain factor output from the gain decoding section 104 by the constant codebook vector output from the HCC decoding section 102 and outputs the multiplication result to the adder 108. The amplifier 107 multiplies the adaptive codebook gain factor output from the node 104 decoding the gain, by the adaptive codebook vector output from the VACC decoding unit 103, and outputs the multiplication result by the adder 108. The adder 108 adds together the constant vector code book after multiplying by the constant codebook gain factor output from the amplifier 106, and the adaptive codebook vector after multiplying by the adaptive codebook gain factor output from the amplifier 107, and outputs the summation result (hereinafter referred to as the “sum vector”) in node 109 synthesis of CLP.

The LPC synthesis node 109 configures a linear prediction synthesis filter using the decoded LPC parameter output from the LPC decoding section 105, controls the linear prediction synthesis filter using the sum vector output from the adder 108 as an excitation signal, and outputs the synthesized signal obtained in the result of this control, to the output filter (final processing filter) 110. The output filter 110 performs processing for extracting formants and selecting frequencies of the fundamental tone, etc. in the synthesized signal output from the LPC synthesis unit 109, and outputs the signal as a decoded speech signal.

Next, parameter masking processing according to this embodiment will be described in detail, taking as an example the case in which masking of the LPC parameter is performed. FIG. 2 is a drawing showing the internal configuration of the LPC decoding unit 105 shown in FIG. 1.

The LPC code L _{n + 1} is entered into the buffer 201 and the code vector decoding section 203, and the frame deletion code B _{n + 1} is entered into the buffer 202, the code vector decoding section 203 and the selection block 209.

The buffer 201 stores the LPC code L _{n + 1 of the} next frame for the duration of one frame, and then outputs this LPC code to the code vector decoding unit 203. As a result of the fact that the code is stored in the buffer 201 for the duration of one frame, the LPC code output from the buffer 201 to the code vector decoding section 203 is the LPC code L _{n of the} current frame.

The buffer 202 stores the next frame deletion code B _{n + 1} for the duration of one frame, and then outputs this deletion code to the code vector decoding unit 203. As a result of the code being stored in the buffer 202 for the duration of one frame, the frame deletion code output from the buffer 202 to the code vector decoding unit 203 is the deletion code B _{n of the} current frame.

The code vector decoding unit 203 has vectors x _n-1 -x _n-m quantized prediction residuals M of the previous frames, decoded vector y _n-1 NUM of the previous frame, LPC code L _{n + 1 of the} next frame, B _{n + 1} code _of deletion of the next frame , the LPC code L _{n of the} current frame and the deletion code B _{n of the} current frame as input, produces a vector x _{n of the} quantized prediction residual of the current frame based on these information elements, and outputs the vector x _{n of the} quantized prediction residual of the current frame to buffer 204-1 and amplifier 205-1. Details of the code vector decoding unit 203 will be given later in this paper.

The buffer 204-1 stores the vector x _{n of the} quantized prediction residual of the current frame for the duration of one frame, and then outputs this vector of the quantized prediction residual to the code vector decoding unit 203, buffer 204-2, and amplifier 205-2. Due to the fact that the vector is stored in the buffer 204-1 for the duration of one frame, the vector of the quantized prediction residual input to the code vector decoding unit 203, the buffer 204-2, and the amplifier 205-2 are the vector x _{n-1 of the} quantized prediction residual previous frame. Similarly, each buffer 204-i (where i is from 2 to M-1) stores the vector x _{n-j + 1 of the} quantized prediction residual for the duration of one frame, and then outputs this vector of the quantized prediction residual to the code vector decoding unit 203 buffer 204- (i + 1), and amplifier 205- (i + 1). The buffer 204-M stores the vector x _{n-M + 1 of the} quantized prediction residual for the duration of one frame, and then outputs this vector of the quantized prediction residual to the code vector decoding unit 203 and amplifier 205- (M + 1).

Amplifier 205-1 multiplies the vector x _{n of the} quantized prediction residual by a predetermined MA prediction coefficient α ₀ and outputs the result to adder 206. Similarly, amplifiers 205-j (where j is 2 to M + 1) multiply the vector x _{n-j + 1 of the} quantized prediction remainder to the predetermined MA prediction coefficient α _j-1 and output the result to the adder 206. The set of MA prediction coefficients can be fixed values of the same type, but in ITU-T G.729 provides two types of sets, which set should be used for run dec dirovaniya is determined on the encoder side, and the active set is encoded and transmitted as part of the information code L _n LPC. In this case, the configuration is used by which the LPC decoding unit 105 is provided with a set of MA prediction coefficients in the form of a table, and the set defined on the encoder side is used as α ₀ -α _M in FIG. 2.

The adder 206 calculates the total sum of vectors of quantized prediction residuals after multiplication by the prediction coefficients MA, derived from amplifiers 205-1-205- (M + 1), and outputs the calculation result, the decoded vector y _n NUM, to the buffer 207 and to the conversion unit 208 KLP.

The buffer 207 stores the decoded vector y _n CHSL for the duration of one frame, and then outputs the decoded vector CHSL a code vector decoding unit 203. As a result, the decoded LSN vector output from the buffer 207 to the decoded code vector node 203 is the decoded LSN vector y _{n − 1 of the} previous frame.

The LPC transform node 208 converts the decoded LSN vector y _n into a set of linear prediction coefficients (decoded LPC parameter), and outputs it to the selection block 209.

The selection unit 209 selects a decoded LPC parameter outputted from LPC conversion unit 208 or a decoded LPC parameter of the previous frame outputted from the buffer 210, based on the code B _n delete the current frame and the code B _{n + 1} of the next frame removal. In particular, the decoded LPC parameter outputted from the LPC conversion unit 208 is selected if the current frame deletion code B _n indicates that “the nth frame is a normal frame”, or the next frame deletion code B _{n + 1} indicates that “( the n + 1) -th frame is a normal frame ", and the decoded LPC parameter of the next frame output from the buffer 210 is selected if the current frame deletion code B _n indicates that" the nth frame is a deleted frame "and the code B _{n + 1} deletion of the next frame indicates that the "(n + 1) th frame is a deleted frame." Then, the selection unit 209 outputs the selection result to the LPC synthesis node 109 and to the buffer 210 as the final decoded LPC parameter. If the selection unit 209 selects the decoded LPC parameter of the next frame output from the buffer 210, then it is not actually necessary to perform all the processing from the code vector decoding section 203 to the LPC conversion section 208, and only processing to update the contents of the 204-1-204-M buffers is necessary perform.

The buffer 210 stores the decoded LPC parameter output from the selection block 209 for a duration of one frame, and then outputs this decoded LPC parameter to the selection block 209. As a result, the decoded LPC parameter output from the buffer 210 to the selection unit 209 is a decoded LPC parameter of the previous frame.

Next, the internal configuration of the code vector decoding unit 203 shown in FIG. 2 will be described in detail using the block diagram of FIG. 3.

In generating the codebook 301 a code vector identified LPC code L _n of the current frame, and outputs it to the switch 309, and also generates a code vector identified code L _{n + 1} LPC next frame, and outputting it to the amplifier 307. As already indicated, in ITU-T Recommendation G.729, information that defines a set of MA prediction coefficients includes the LPC code L _n , in which case the LPC code L _{n is} also used to decode the MA prediction coefficient in addition to decoding the code vector, but a description of this omitted in Anna work. In addition, the codebook may have a multi-stage configuration and may have a split configuration. For example, in ITU-T Recommendation G.729, the codebook has a two-stage configuration with the division of the second stage into two. The vector output from the codebook of a multi-stage configuration or separation configuration is generally not used in the form in which it exists, and if the spacing between its elements is extremely small, or the order of the elements is inverse, then processing is generally performed to ensure so that the minimum interval becomes a specific value, or for ordering.

The vectors x _n-1- x _{n- M} quantized prediction residuals M of the previous frames are input to the respective amplifiers 302-1-302-M and the corresponding amplifiers 305-1-305-M, respectively.

Amplifiers 302-1-302-M multiply the input vectors x _n-1 -x _{n-M of} quantized prediction residues by the coefficients α ₀ -α _{M of the} MA prediction, respectively, and output the results to adder 303. As indicated above, in the case of an ITU recommendation -T G.729, there are two types of sets of MA prediction coefficients, and information about which one is used includes the LPC code L _n . In addition, when the frame for which this multiplication is performed is deleted, the set of MA prediction coefficients used in the previous frame is actually used, since the LPC code L _{n is} deleted. Those. MA prediction coefficient information decoded from the LPC code L _{n-1 of the} previous frame is used. If the previous frame is also a deleted frame, then the information of the frame located in front of it will be used.

Adder 303 calculates the total sum of vectors of quantized prediction residues after multiplication by the prediction coefficients MA output from amplifiers 302-1-302-M, and outputs the vector that is the result of multiplication to adder 304.

The adder 304 subtracts the vector output from the adder 303 from the decoded vector y _n-1 NUM of the previous frame output from the buffer 207, and outputs the vector that is the result of this calculation to the switch 309.

The vector output from adder 303 is an LSN vector predicted by the MA type predictor in the current frame, and adder 304 performs processing to find the quantized prediction residual vector in the current frame necessary to generate the decoded LSL vector of the previous frame. Those. by means of amplifiers 302-1-302-M, adder 303 and adder 304, the vector is calculated so that the decoded vector y _{n-1 of the} number of the previous frame becomes the decoded vector y _{n of the} number of the current frame.

Amplifiers 305-1-305-M multiply the input vectors x _n-1 -x _{n-M of the} quantized prediction residual by the weighting factors β ₁ -β _M , respectively, and outputs the results to adder 308. The amplifier 306 multiplies the decoded vector y _n-1 The LSL of the previous frame, which is output from buffer 207, by the weight coefficient β _-1 , and outputs the result to the adder 308. The amplifier 307 multiplies the code vector x _{n + 1} output from the codebook 301 by the weight coefficient β ₀ , and outputs the result to the adder 308.

Adder 308 calculates the total sum of vectors output from amplifiers 305-1-305-M, amplifier 306, and amplifier 307, and outputs a code vector that is the result of this calculation to switch 309. That is, an adder 308 calculates a vector by performing a weighted summation of the code vector identified by the LPC code L _{n + 1 of the} next frame, the decoded PSL vector of the previous frame, and the quantized prediction residual vectors M of the previous frames.

If the current frame deletion code B _n indicates that “the nth frame is a normal frame”, then the switch 309 selects a code vector output from the codebook 301 and outputs it as the vector x _{n of the} quantized prediction residual of the current frame. On the other hand, if the current frame deletion code B _n indicates that “the nth frame is the deleted frame”, the switch 309 further selects a vector to be output, according to what information the next deletion code B _{n + 1 has} frame.

Those. if the next frame deletion code B _{n + 1} indicates that the “(n + 1) th frame is the deleted frame”, then the switch 309 selects the vector output from the adder 304 and outputs it as the vector x _{n of the} quantized prediction residual of the current frame. In this case, the processing of the vector generation process from the codebook 301 and amplifiers 305-1-305-M to the adder 308 should not be performed.

On the other hand, if the next frame deletion code B _{n + 1} indicates that the “(n + 1) th frame is a normal frame”, then the switch 309 selects the vector output from the adder 308 and outputs it as the quantized remainder vector x _n predictions of the current frame. In this case, the processing of the vector generation process from amplifiers 302-1-302-M to adder 304 should not be performed.

Thus, according to this embodiment, when the current frame is deleted, if the next frame is received in the usual way, the masking processing of the quantized prediction residual decoded for the LSL parameter of the current frame is performed by weighted summation processing (weighted linear summation processing), in particular, masking processing using the parameter decoded in the past, the quantized prediction remainder of the frame received in the past, and the quantized prediction remainder will be his picture, and the decoding parameter CHSL performed by using the masked quantized prediction residue. Thus, it is possible to provide higher masking efficiency than by reusing the past decoded parameter of the LSL.

The results of the masking processing of this embodiment will now be described using FIGS. 4 to 7, which represent actual examples compared to conventional technology. 4 to 7, - denotes a decoded quantized prediction remainder, • denotes a decoded quantized prediction remainder obtained by masking processing, ◊ denotes a decoded parameter, and ♦ denotes a decoded parameter obtained by masking processing.

4 is a drawing showing an example of a result of performing conventional processing when there is no deleted frame in which the decoded parameter y _{n of the} nth frame is found by Equation (1) below from the decoded quantized prediction residual. In Equation (1), c _n is the decoded quantized prediction residual of the nth frame.

5 is a drawing showing an example of a result of performing masking processing of this embodiment, and FIGS. 6 and 7 are drawings showing examples of a result of performing conventional masking processing. 5-7, it is assumed that the nth frame is deleted, and other frames are regular frames.

In the masking processing of the embodiment shown in FIG. 5, the quantized prediction residual C _n decoded for the remote nth frame is found using Equation (3) below to make the sum D (where D is determined using Equation (2) below ) the distance between the decoded parameter y _n-1 (n-1) -th frame and the decoded parameter y _{n the} n-th frame, and the distance between the decoded parameter y _{n the} n-th frame and the decoded parameter y _{n + 1} (n + 1) -th frame minimum, so that changes to the decoded parameter from one frame to another became insignificant.

Then, in the masking processing of this embodiment, the decoded parameter y _{n of the} deleted n-th frame is found by Equation (1) above, using the decoded quantized prediction residual C _{n of the} deleted n-th frame found by Equation (3). As a result, as is clear from the comparison of FIG. 4 and FIG. 5, the decoded parameter y _n obtained by masking processing of this embodiment has almost the same value as the value obtained by conventional processing when there is no deleted frame.

In contrast, in the conventional masking processing shown in FIG. 6, when the nth frame is deleted, the decoded parameter y of the _n-1 (n-1) th frame is used directly as the decoded parameter y _{n of the} nth frame. In addition, in the conventional masking processing shown in FIG. 6, the decoded quantized prediction residue C _{n of the} nth frame is found by the operation inverse to the above Equation (1).

In this case, since the changes in the decoded parameter, accompanied by changes in the decoded quantized prediction remainder, are not taken into account, as is clear from the comparison of FIG. 4 and FIG. 6, the decoded parameter y _n obtained by the usual masking processing in FIG. 6 has a value that very different from the value obtained through normal processing when there is no deleted frame. In addition, since the decoded quantized prediction residual C _{n of the} nth frame is also different, the decoded parameter y of the _{n + 1} (n + 1) th frame obtained by the conventional masking processing in FIG. 6 also has a value different from the value obtained by conventional processing when there is no deleted frame.

In the conventional masking processing shown in FIG. 7, the decoded quantized prediction residual is found by interpolation, and when the nth frame is deleted, the average value of the decoded quantized residual C _n-1 prediction residual of the (n-1) th frame and the decoded quantized remainder is used C _{n + 1} predictions of the (n + 1) th frame as the decoded quantized remainder of C _n predictions of the nth frame.

Then, in the conventional masking processing shown in FIG. 7, the decoded parameter y _{n of the} removed nth frame is found by the above Equation (1) using the decoded quantized prediction residual C _n found by interpolation.

As a result, as is clear from the comparison of FIG. 4 and FIG. 7, the decoded parameter y _n obtained by the conventional masking processing in FIG. 7 has a value that is very different from the value obtained by the conventional processing when there is no deleted frame. This is because when the decoded parameter changes slightly from one frame to another through the determination of a weighted moving average, with this conventional masking process, the decoded parameter also changes along with the changes in the decoded quantized prediction remainder. Furthermore, since the decoded quantized prediction residual C _{n of the} nth frame is also different, the decoded parameter y of the _{n + 1} (n + 1) th frame obtained by the conventional masking processing in FIG. 7 also has a value different from the value obtained through normal processing when there is no deleted frame.

(Embodiment 2)

FIG. 8 is a block diagram showing a basic configuration of a speech decoding apparatus according to Embodiment 2 of the present invention. The speech decoding device 100 shown in FIG. 8 differs from the device in FIG. 1 only by the additional addition of masking type information E _{n + 1} as a parameter input to the LPC decoding unit 105.

FIG. 9 is a block diagram showing the internal configuration of the LPC decoding unit 105 shown in FIG. The LPC decoding unit 105 shown in FIG. 9 differs from the node in FIG. 2 only by the additional addition of masking type information E _{n + 1} as a parameter input to the code vector decoding unit 203.

FIG. 10 is a block diagram showing an internal configuration of a code vector decoding unit 203 shown in FIG. 9. The code vector decoding unit 203 shown in FIG. 10 differs from the node in FIG. 3 only by the addition of a coefficient decoding unit 401.

The coefficient decoding unit 401 stores a plurality of kinds of sets of weights (β _-1 -β _M ) (hereinafter referred to as “sets of coefficients”), selects one set of weights from a plurality of sets of coefficients according to the input masking type code E _{n + 1} , and outputs it amplifiers 305-1-305-M, 306 and 307.

Thus, according to this embodiment, in addition to providing the features described in Embodiment 1, a plurality of weighted summation weighting sets are provided for performing masking processing, information for identifying the optimal set is transmitted to the decoder side after confirmation using which set of weighting factors on the encoder side, high masking efficiency is obtained, and masking processing is performed using the indicated set of weights x coefficients based on information received on the decoder side, providing obtaining even higher masking efficiency than in the embodiment 1.

(Embodiment 3)

11 is a block diagram showing a basic configuration of a speech decoding apparatus according to Embodiment 3 of the present invention. The speech decoding device 100 shown in FIG. 11 differs from the device in FIG. 8 only by the addition of a separation unit 501, which divides the LPC code L _{n + 1} input to the LPC decoding unit 105 into two kinds of codes, V _{n + 1} and K _{n + 1} . The code is a V code for generating a code vector, and code K is the MA prediction coefficient code.

FIG. 12 is a block diagram showing the internal configuration of the LPC decoding unit 105 shown in FIG. 11. The codes V _n and V _{n + 1} that generate the code vector are used in the same way as the LPC codes L _n and L _{n + 1} , and therefore their description is omitted in this paper. The LPC decoding unit 105 shown in FIG. 12 differs from the node in FIG. 9 only by the addition of a buffer 601 and a coefficient decoding unit 602, and the addition of an MA prediction coefficient code K _{n + 1} as a parameter input to the code decoding unit 203 vector.

The buffer 601 stores the MA prediction coefficient code K _{n + 1} for the duration of one frame, and then outputs this MA prediction coefficient code to the coefficient decoding section 602. As a result, the MA prediction coefficient code output from the buffer 601 to the coefficient decoding unit 602 is the MA prediction coefficient code K _{n of} the previous frame.

The coefficient decoding unit 602 stores a plurality of kinds of coefficient sets, identifies a coefficient set by frame deletion codes B _n and B _{n + 1} , a mask type code E _{n + 1,} and a prediction coefficient code K _n and outputs it to amplifiers 205-1-205- (M + 1). In this case, there are three methods by which the identification of a set of coefficients can be performed in the coefficient decoding unit 602 as follows.

If the input frame deletion code B _n indicates that “the nth frame is a normal frame”, then the coefficient decoding section 602 selects a set of coefficients defined by the MA prediction coefficient code K _n .

If the input frame deletion code B _n indicates that “the nth frame is a deleted frame” and the frame deletion code B _{n + 1} indicates that “the (n + 1) th frame is a normal frame”, then the coefficient decoding section 602 determines the set of coefficients to be selected using the masking type code E _{n + 1} , adopted as a parameter of the (n + 1) th frame. For example, if a masking type code E _{n + 1} is determined in advance that indicates the type of the MA prediction coefficient to be used with the nth frame, which is a masked frame, then the masking type code E _{n + 1} can be used directly instead of the coefficient code K _n predictions MA.

In addition, if the input frame deletion code B _n indicates that “the nth frame is a deleted frame”, and the frame deletion code B _{n + 1} indicates that “the (n + 1) th frame is a deleted frame”, then the only the information that can be used is the information of the coefficient set used by the previous frame, and therefore, the coefficient decoding section 602 repeatedly uses the coefficient set used by the previous frame. Alternatively, a set of coefficients of a certain type can be provided in advance so that it is used continuously.

FIG. 13 is a block diagram showing an internal configuration of a code vector decoding unit 203 shown in FIG. 12. As shown in FIG. 13, the code vector decoding unit 203 differs from the node in FIG. 10 only in that the coefficient decoding unit 401 selects a set of coefficients using both the mask type code E _{n + 1} and the prediction coefficient code K _{n + 1} MA.

13, a coefficient decoding unit 401 is provided with a plurality of weighting sets, and a weighting set is prepared according to the MA prediction coefficient used by the next frame. For example, in the case where there are two types of sets of MA prediction coefficients, one is designated type 0 and the other is type 1, the sets of MA prediction coefficients contain a group of sets of weighting coefficients defined for use when the set of MA prediction coefficients of the next frame is type 0, and a group of sets of weights determined for use when the set of MA prediction coefficients of the next frame is type 1.

In this case, the coefficient decoding unit 401 determines one or another group of sets of weighting coefficients from the above groups, selects one set of weighting factors from the set of weighting coefficients according to the input masking type code E _{n + 1} , and outputs it to amplifiers 305-1-305- M, 306 and 307.

An exemplary method for determining weight factors β _-1 -β _{M is} shown below. As already indicated, if the nth frame is deleted and the (n + 1) th frame is received, then the final decoded parameters are unknown in both frames, even if the decoded quantized prediction remainder in the (n + 1) th frame can be decoded correctly. Therefore, the decoded parameters of both frames are not determined unambiguously if an assumption (constraint condition) of some kind is not established. Thus, the quantized prediction residual y _{n is} found by Equation (4) below to minimize D ^(j) , the sum of the distance between the decoded parameter in the nth frame and the decoded parameter in the (n-1) th frame, and the distance between the decoded parameter in the (n + 1) -th frame and the decoded parameter in the nth frame, so that the decoded parameters of the nth frame and (n + 1) -th frame do not differ as much as possible from the already decoded parameter (n- 1st) frame.

When the parameter is an LSN parameter, x _n ^(j) , y _n ^(j) , α _i ^(j) , α ′ _i ^(j) in Equation (4) are as follows:

x _n ^(j) : quantized remainder of the prediction of the jth component of the LSL parameter in the nth frame;

y _n ^(j) : j-th component of the NSR parameter in the nth frame;

α _i ^(j) : jth component of the i-th order component within the set of the MA prediction coefficient in the nth frame;

α ′ _i ^(j) : j-th component of the i-th order component within the set of the MA prediction coefficient in the (n + 1) -th frame.

M: MA prediction order.

In this case, solving the equation obtained by partial differentiation of D ^(j) using x _n ^(j) to obtain 0, x _n ^{(j) is} expressed in the form of Equation (5) below.

In Equation (5), β _i ^(j) is the weight coefficient expressed using α _i ^(j) and α ′ _i ^(j) . That is, if there is only one type of set of prediction coefficient MA, there is also only one type of set of weight coefficients β _i ^(j) , but if there are many types of sets of prediction coefficients MA, then many types of sets of weighting coefficients are obtained using combinations of α _i ^(j) and α ′ _i ^(j) .

For example, in the case of ITU-T Recommendation G.729, there are two types of sets of MA prediction coefficients, and therefore, if they are designated type 0 and type 1, then four types of sets can be obtained - when the nth frame and (n + 1) the nth frame is type 0, when the nth frame is type 0, and the (n + 1) th frame is type 1, when the nth frame is type 1, and the (n + 1 )th frame is type 0 , and when the n-th frame and the (n + 1) -th frame both are of type 1. You can come up with many ways to determine which set of weights should be used from these four types of sets.

The first method consists in generating decoded LSLs of the nth frame and decoded LSLs of the (n + 1) th frame on the encoder side using all four types of sets, calculating the Euclidean distance between the generated decoded LSLs of the nth frame and non-quantized LSLs obtained using analysis of the input signal, calculation of the Euclidean distance between the generated decoded LSLs of the (n + 1) -th frame and non-quantized LSLs obtained by analyzing the input signal, selecting one of the sets of weighting coefficients β, which minimizes mm of these Euclidean distances, encoding the selected set in two bits and transmitting them to the decoder. In this case, two bits in the frame are necessary for encoding the weighting factor β in addition to the ITU-T G.729 recommendation in the coding information. Better sound quality can be achieved by using weighted Euclidean distances, as used when quantizing the BSL recommendation ITU-T G.72 9, instead of Euclidean distances.

The second method is to make the number of extra bits in the frame equal to one when using information such as the prediction coefficient MA of the (n + 1) th frame. Since information such as the prediction coefficient MA of the (n + 1) th frame is on the side of the decoder, combinations of α _i ^(j) and α ′ _i ^{(j) are} limited to two. Those. if the prediction type of the MA of the (n + 1) th frame is type 0, then the combination of the MA prediction type of the nth frame and the (n + 1) th frame is either (0-0) or (1-0), which provides an opportunity to limit the set of weighting factors β to two types. On the encoder side, it is only necessary to perform encoding using the one of these two types of sets of weighting coefficients β that has a smaller remainder with respect to non-quantized LSLs, in the same way as in the first method above, and transmit it to the decoder.

The third method is a method in which no selection information is sent at all using one set of weighting factors, for which there are only two kinds of combinations of the MA- (0-0) or (1-0) prediction type, the first being selected when the coefficient type the predictions of the MA of the (n + 1) th frame is 0, and the latter is selected when the type of the prediction coefficient of the MA of the (n + 1) th frame is 1. Alternatively, a method may be used by which a frame deletion type is set to a specific type, such as (0-0) or (0-1).

Other possible methods are the method by which, for a frame for which it is determined that the input signal is constant, it is ensured that the decoded parameters for the (n-1) th frame and the n-th frame become the same as in the conventional method, and the method which uses the set of weights β found under the condition that the decoded parameters of the (n + 1) th frame and the nth frame become the same.

In this case, pitch period information, information such as MA prediction coefficient or similar information of the (n-1) th frame and the (n + 1) th frame can be used to determine constancy. Those. possible methods are to determine that the signal is constant when the decoded difference between the pitch periods between the (n-1) -th frame and the (n + 1) -th frame is small, or to determine that the signal is constant when the type is selected, suitable for encoding a frame for which the information of the type of the prediction coefficient MA decoded in the (n + 1) th frame is constant (i.e., the type in which the highest prediction coefficient MA also has a certain weight).

Thus, in this embodiment, in addition to what is provided in embodiment 2, there are two types of MA prediction coefficients, enabling the use of different sets of MA prediction coefficients for the constant part and the part that is not constant, and providing the opportunity to improve the efficiency of the means of quantization of the CSL.

In addition, when using Equation (5), a set of weights that minimizes Equation (4), the decoded parameters of the CRL of the remote frame and the normal frame, which is the next frame after the deleted frame, are not guaranteed to become values that are very different from the parameter The number of frames preceding the deleted frame. Therefore, even if the decoded LSL parameter of the next frame is unknown, the received information (quantized prediction remainder) of the next frame can continue to be used efficiently, and the risk of masking performed in the wrong direction - i.e. the risk of a very strong difference from the correct decoded PSL parameter can be kept minimal.

In addition, if the above second method is used as a method of selecting a type of masking, then the information of the type of the prediction coefficient MA can be used as part of the information that identifies a set of weighting factors for using masking processing, which makes it possible to reduce the amount of additionally transmitted information of the set of weighting factors for use of masking processing.

(Embodiment 4)

FIG. 14 is a block diagram showing the internal configuration of the gain decoding section 104 shown in FIG. 1 (the same applies to the gain decoding section 104 in FIG. 8 and FIG. 11). In this embodiment, as in the case of ITU-T G.729, gain decoding is performed once for a subframe, and one frame consists of two subframes, and FIG. 14 shows sequential decoding of codes (G _m and G _{m + 1} ) the gain of two subframes of the nth frame, where n is the frame number and m is the number of the subframe (the numbers of the first subframe and the second subframe of the nth frame are m and m + 1, respectively).

In Fig. 14, the gain code G _{n + 1 of} the (n + 1) th frame is input to the gain decoding unit 104 from the demultiplexing unit 101, the gain code G _{n + 1} is input to the separation unit 700 and divided by the code G _{m + 2} gain of the first subframe and code G _{m + 3} gain of the second subframe of the (n + 1) th frame. The division into codes G _{m + 2} and G _{m + 3} gain can also be performed using the node 101 demultiplexing.

In the gain decoding section 104, a decoded gain of a subframe m and a decoded gain of a subframe m + 1 are obtained in order using G _m , G _{m + 1} , G _{m + 2} and G _{m + 3} generated from the input G _n and G _{n +1}

The operation of each node from the gain decoding section 104 when decoding the gain code G _m will now be described with respect to FIG.

A gain code G _{m + 2} is input to buffer 701 and a prediction residual decoding unit 704, and a frame deletion code B _{n + 1} is input to buffer 703, a prediction residual decoding unit 704 and a selection block 713.

The buffer 701 stores the input gain code for the duration of one frame and then outputs this gain code to the prediction residual decoding section 704, so that the gain code input to the prediction residual decoding section 704 is the gain code of the previous frame. Those. if the gain code entered into the buffer 701 is G _{m + 2} , then the output gain code is G _m . Buffer 702 also performs the same processing as buffer 701. That is, the input gain code is stored for a duration of one frame and then output to the prediction residual decoding unit 704. The only difference is that the input / output of buffer 701 is the gain code of the first subframe, and the input / output of buffer 702 is the gain code of the second subframe.

A buffer 703 stores the next frame deletion code B _{n + 1} for the duration of one frame, and then outputs this frame deletion code to the prediction residual decoding unit 704, a selection unit 713, and a PAC vector power calculation unit 708. The frame deletion code output from the buffer 703 to the prediction residual decoding unit 704, the selection unit 713 and the PAC vector power calculation unit 708 is the deletion code of the previous frame before the input frame, and is thus the deletion code B _{n of the} current frame.

Prediction residual decoding unit 704 that has logarithmic quantized prediction residuals x _{m − 1} −x _mM (which are the result of calculating the logarithms of quantized prediction residuals MA) M of the previous subframes, decoded power e _{m − 1} (logarithm of the decoded gain) of the previous subframe, coefficient e _B gain error prediction residual, codes G _{m + 2} and G _{m + 3} gain of the next frame, code B _{n + 1} delete the next frame, codes G _m and G _{m + 1} gain of the current frame and code B _n deleted Eating the current frame as input, generates a quantized prediction residue of the current frame based on these pieces of information, and outputs it to the logarithm calculation section 705 and the multiplication section 712. Details of the prediction remainder decoding unit 704 will be provided later in this paper.

The logarithm calculation unit 705 calculates the logarithm x _{m of the} quantized prediction residual output from the prediction residual decoding unit 704 (in ITU-T Recommendation G.729, 20 × log ₁₀ (x), where x is input) and outputs it to the buffer 706 -one.

A buffer 706-1, which has a logarithmic quantized prediction remainder x _m output from the logarithm calculation unit 705, as input information, stores it for the duration of one subframe, and then outputs this logarithmic quantized prediction remainder to the prediction remainder decoding unit 704, a buffer 706-2 and buffer 707-1. Those. the logarithmic quantized prediction residual inputted to the prediction residual decoding unit 704, the buffer 706-2 and the amplifier 707-1, is the logarithmic quantized prediction residual x _{m − 1 of the} previous subframe. Similarly, each of the buffers 706-i (where i is from 2 to M-1) stores the input logarithmic quantized prediction remainder x _mi for the duration of one subframe and then outputs this logarithmic quantized prediction remainder to the prediction remainder decoding unit 704, buffer 706 - (i + 1) and amplifier 707-i. The buffer 706-M stores the input logarithmic quantized prediction residual x _mM-1 for the duration of one subframe and then outputs this logarithmic quantized prediction residual to the prediction residual decoding unit 704 and to the amplifier 707-M.

Amplifier 707-1 multiplies the logarithmic quantized remainder x _{mi of the} prediction by a predetermined MA prediction coefficient α ₁ and outputs the result to adder 710. Similarly, each of the amplifiers 707-j (where j is 2 to M) multiplies the logarithmic quantized remainder x _{mj of the} prediction by a predetermined coefficient α _j MA prediction, and outputs the result to adder 710. Set MA prediction coefficients comprises one kind of fixed values in the recommendation ITU-T G.729, but it is also possible to employ a configuration wherein provide nozhestvo kinds of sets, and choose the one most suitable.

If the current frame deletion code B _n indicates that “the nth frame is a normal frame”, then the GAC vector power calculation unit 708 calculates the power of the GAC vector (constant codebook) decoded separately and outputs the calculation result to the average excess determination unit 709 power. If the current frame deletion code B _n indicates that “the nth frame is a deleted frame”, then the PAC vector power calculation unit 708 outputs the PAC vector power of the previous subframe to the average power excess determination unit 709.

The average power excess determination unit 709 subtracts the GAC vector power output from the GAC vector power calculation unit 708 from the average power and outputs the subtraction result — the coefficient e of _the gain of the prediction residual error — to the prediction residual decoding unit 704 and adder 710. In this case, suggest that the average power is a preset constant. In addition, power addition / subtraction is performed in the logarithmic region.

An adder 710 calculates the total sum of the logarithmic quantized prediction residues after multiplying by the MA prediction coefficients output from the amplifiers 707-1-707-M and the gain e _{in the} error of the prediction residual output from the average power excess determination unit 709, and outputs a logarithmic predicted coefficient amplification, which is the result of this calculation, to the node 711 calculation of the exponential function.

The exponential function computation unit 711 calculates an exponential function (10 ^x , where x is the input information) of the logarithmic predicted gain derived from the adder 710, and outputs the predicted gain that is the result of this calculation to the multiplication unit 712.

A multiplication unit 712 multiplies the predicted gain output from the exponential function calculation section 711 by a quantized prediction residual output from the prediction balance decoding section 704, and outputs the decoded gain, which is the result of this calculation, to the selection block 713.

The selection unit 713 selects one of the decoded gain outputted from the multiplier block 712 or decoded gain factor of the previous frame after attenuation output from the amplifier 715 based on the code B _n delete the current frame and the code B _{n + 1} of the next frame removal. In particular, the decoded gain derived from the multiplier 712 is selected if the current frame delete code B _n indicates that “the nth frame is a normal frame” or the next frame delete code B _{n + 1} indicates that “(n The +1) -th frame is a normal frame ", and the decoded gain of the previous frame after attenuation output from the amplifier 715 is selected if the current frame deletion code B _n indicates that" the nth frame is a deleted frame "and the code B _{n 1} removal of the next frame indicates that the «(n-1) -th frame is the removal of scenes. " Then, selection block 713 outputs the selection result as the final predicted gain to amplifiers 106 and 107, buffer 714, and the logarithm calculation section 716. If the selection block 713 selects the decoded gain of the previous frame after attenuation output from the amplifier 715, then it is not actually necessary to perform all the processing from the prediction remainder decoding unit 704 to the multiplication block 712, and only the processing to update the contents of the buffers 706-1-706-M must be performed.

Buffer 714 stores the decoded gain output from selection block 713 for the duration of one subframe, and then outputs this decoded gain to amplifier 715. As a result, the decoded gain output from buffer 714 to amplifier 715 is the decoded gain of the previous subframe . An amplifier 715 multiplies the decoded gain of the previous subframe output from the buffer 714 by a predetermined attenuation coefficient, and outputs the result to the selection block 713. The value of this predefined attenuation coefficient is 0.98 in ITU-T G.729 recommendation, for example, but the optimal value for the codec can be set accordingly, and this value can also be changed according to the characteristics of the remote frame signal, for example, whether the remote frame is a voiced frame or unvoiced frame.

The logarithm calculation unit 716 calculates the logarithm e _{m of the} decoded gain output from the selection block 713 (in ITU-T Recommendation G.729, 20 × log ₁₀ (x), where x is the input information) and outputs it to buffer 717. Buffer 717, which has the logarithm e _{m of the} decoded gain as information input from the logarithm calculation unit 716, stores it for the duration of one subframe, and then outputs this logarithm of the decoded gain to the prediction residual decoding unit 704. Those. the predicted gain logarithm input to the prediction residual decoding unit 704 is the logarithm of the e _{m − 1} decoded gain of the previous subframe.

FIG. 15 is a block diagram showing an internal configuration of a prediction residual decoding unit 704 shown in FIG. 14. On Fig codes G _m , G _{m + 1} , G _{m + 2} and G _{m + 3} gain are entered into the codebook 801, codes B _n and B _{n + 1} delete the frame is entered into the switch 812, the logarithmic quantized residuals x _{m -1} -x _mM prediction previous subframes M is introduced into the adder 802 and the decoded logarithmic ratio e _m-1 the gain of the previous subframe and the coefficient e _B gain prediction residual error introduced into node 807 generate a quantized residue sub-frame prediction unit 808 and generating a quantized prediction residue subframe.

In codebook 801, the corresponding quantized prediction residues from the input gain codes G _m , G _{m + 1} , G _{m + 2} and G _{m + 3 are} decoded, the quantized prediction residues corresponding to the input gain codes G _m and G _{m + 1} are output, to switch 812 via switch 813, and quantized prediction residues corresponding to the input gain codes G _{m + 2} and G _{m + 3} are output to the logarithm calculation unit 806.

Switch 813 selects any of the quantized prediction residues decoded from gain codes G _m and G _{m + 1} and outputs it to switch 812. In particular, the quantized prediction residual decoded from gain code G _{m is} selected when decoding processing is performed the gain of the first subframe, and the quantized prediction residual decoded from the gain code G _{m + 1} is selected when the gain decoding processing of the second subframe is performed.

Adder 802 calculates the total sum of the logarithmic quantized residuals x _{m − 1} −x _{mM of the} prediction M of the previous subframes and outputs the result of this calculation to amplifier 803. Amplifier 803 calculates the average by multiplying the output of adder 802 by 1 / M and outputs the result of this calculation by Node 804 attenuation by 4 dB (decibels).

The attenuation unit 804 by 4 dB reduces the output signal of the amplifier 803 by 4 dB and outputs the result to the exponential function calculation unit 805. This 4 dB attenuation should prevent the prediction device from outputting an excessively large predicted value in a frame (subframe) recovered from the remote frame, and the attenuation unit is not necessarily an integral part of an exemplary configuration that does not need such a need. With respect to the attenuation of 4 dB, the optimum value can also be freely developed.

The exponential function computation unit 805 computes the exponential function from the output value of the 4 dB attenuation unit 804, and outputs the masked prediction residue that is the result of this calculation to switch 812.

The logarithm calculation unit 806 calculates the logarithms of the two quantized prediction residues output from the codebook 801 (resulting from the decoded gain codes G _{m + 2} and G _{m + 3} ) and outputs the logarithmic quantized prediction residuals x _{m + 2} and x _{m + 3} , which are the results of these calculations, to the sub-frame prediction residual quantization generation unit 807 and the sub-frame prediction quantized residual generation unit 808.

Node 807 of generating a quantized prediction residual subframe, which has logarithmic quantized residuals x _{m + 2} and x _{m + 3} predictions, logarithmic quantized residuals x _m-1 -x _mm predictions M _{of the} previous subframes, decoded power e _{m-1 of the} previous subframe, and coefficient e _In error amplification, the prediction residuals as input information, calculates the logarithmic quantized prediction residual of the first subframe based on these information elements, and outputs it to switch 810. In the same way, the generator node 808 of the quantized prediction residuals of a subframe that has logarithmic quantized residuals x _{m + 2} and x _{m + 3} predictions, the logarithmic quantized residuals x _m-1 -x _{mm of the} prediction M _{of the} previous subframes, the decoded power e _{m-1 of the} previous subframe, and the gain coefficient e _B the error of the prediction residual as input information, calculates the logarithmic quantized prediction residual of the second subframe based on these pieces of information, and outputs it to the buffer 809. Details of the generation nodes 807 and 808 are quantized sub-frame prediction residuals will be given later in this paper.

The buffer 809 stores the logarithmic quantized prediction residual of the second subframe output from the quantized prediction residual subframe generating unit 808 for the duration of one subframe, and outputs this logarithmic quantized prediction residual of the second subframe to switch 810 when the second subframe is processed. During processing of the second subframe, x _{m − 1} −x _mM , e _{m − 1,} and e _B are updated outside the prediction residual decoding section 704, but the processing is not performed by the quantized prediction residual generation section 807 or the quantized subframe prediction residual generation section 808 , and all processing is performed during processing of the first subframe.

During processing of the first subframe, the switch 810 is connected to the quantized prediction remainder sub-frame generation unit 807, and it outputs the generated logarithmic quantized prediction residual of the first subframe to the exponential function computing unit 811, while during the processing of the second subframe, the switch 810 is connected to the buffer 809, and it outputs the logarithmic quantized prediction residual of the second subframe generated by the quantized prediction residual generation unit 808 to the calculation unit 811 reproductive function. The exponential function calculation unit 811 calculates the exponential function of the logarithmic quantized remainder output from the switch 810, and outputs the masked prediction remainder that is the result of this calculation to the switch 812.

If the current frame deletion code B _n indicates that “the nth frame is a normal frame”, then the switch 812 selects the quantized prediction residual output from the codebook 801 via the switch 813. On the other hand, if the current frame deletion code B _n indicates that “the nth frame is a deleted frame”, the switch 812 further selects a quantized prediction remainder that will be output in accordance with the information that the next frame deletion code B _{n + 1 has} .

Those. a switch 812 selects a masked prediction residual output from the exponential function calculation node 805 if the next frame deletion code B _{n + 1} indicates that the “(n + 1) th frame is a deleted frame” and selects a prediction masked residual output from the node 811 calculating an exponential function if the deletion code B _{n + 1 of the} next frame indicates that “the (n + 1) th frame is a normal frame”. Entering data to an input other than the selected input is not necessary, and therefore, in actual processing, it is usually first necessary to determine which input should be selected in the switch 812 and to process it to generate a signal that will be output to a specific input.

FIG. 16 is a block diagram showing the internal configuration of a quantized prediction residual generation unit 807 of a subframe shown in FIG. The internal configuration of the quantized prediction residual sub-frame prediction residual generation unit 808 is also identical to the configuration in FIG. 16, and only the values of the weighting coefficients are different from the values in the quantized prediction residual-generation unit 807 of the sub-frame.

Amplifiers 901-1-901-M multiply the input logarithmic quantized residuals x _m-1 -x _mM predictions by the weighting factors β ₁ -β _M , respectively, and output the results to the adder 906. Amplifier 902 multiplies the logarithmic coefficient e _{m-1 of the} gain of the previous a subframe by a weight coefficient β ₋₁ , and outputs the result to an adder 906. An amplifier 903 multiplies the logarithmic error amplification coefficient e _B by a weight factor β _B , and outputs the result to an adder 906. An amplifier 904 multiplies the logarithmic quantized remainder x _{m + 2} predictions by weighting factor β ₀₀ , and outputs the result to adder 906. Amplifier 905 multiplies the logarithmic quantized prediction remainder x _{m + 3} by weighting factor β ₀₁ and outputs the result to adder 906.

Adder 906 calculates the total sum of the logarithmic quantized prediction residues output from amplifiers 901-1-901-M, amplifier 902, amplifier 903, amplifier 904, and amplifier 905, and outputs the result of this calculation to switch 810.

The following shows an exemplary method for determining the weight coefficient β in this embodiment. As already indicated, in the case of ITU-T G.729 recommendation, gain quantization is processing a subframe, and one frame consists of two subframes, and therefore deleting one frame is removing a burst of two consecutive subframes. Therefore, the set of weights β cannot be determined by the method described in Embodiment 3. Thus, in this embodiment, x _m and x _{m + 1} are found that minimize D in Equation (6) below.

In this case, for example, a case is described in which one frame consists of two subframes, as in ITU-T G.729, and the MA prediction coefficient is of only one kind. In Equation

(b) y _m-1 , y _m , y _{m + 1} , y _{m + 2} , y _{m + 3} , x _m , x _{m + 1} , x _{m + 2} , x _{m + 3} , x _B and α _i are following.

y _m-1 : decoded logarithmic gain of the second subframe of the previous frame.

y _m : decoded logarithmic gain of the first subframe of the current frame.

y _{m + 1} : decoded logarithmic gain of the second subframe of the current frame.

y _{m + 2} : decoded logarithmic gain of the first subframe of the next frame.

y _{m + 3} : decoded logarithmic gain of the second subframe of the next frame.

x _m : logarithmic quantized prediction remainder of the first subframe of the current frame.

x _{m + 1} : logarithmic quantized prediction remainder of the second subframe of the current frame.

x _{m + 2} : logarithmic quantized prediction remainder of the first subframe of the next frame.

x _{m + 3} : logarithmic quantized prediction remainder of the second subframe of the next frame.

x _B : logarithmic error gain.

α _i : i-th order prediction coefficient MA.

Determining x _m and x _{m + 1} using the equation obtained by partial differentiation of Equation (6) so that x _m is 0, and the equation obtained by partial differentiation of Equation (6) so that x _{m + 1} is 0 as a system of equations, we obtain Equation (7) and Equation (8). Since β ₀₀ , β ₀₁ , β ₁ -β _M , β _-1 , β _B , β ′ ₀₀ , β ′ ₀₁ , β ′ ₁ -β ′ _M , β ′ _-1 and β ′ _{B are} found from α ₀ -α _M , then they are uniquely determined.

Thus, when the next frame is received in the usual way, the masking processing of the logarithmic quantized prediction residual of the current frame is performed by weighted summing processing, in particular, for masking processing using the logarithmic quantized prediction residual received in the past and the logarithmic quantized prediction residual of the next frame, and gain parameter decoding is performed using masked logarithmic quantized prediscal remainder anija, enabling to provide higher efficiency masking than using the decoded parameter of the previous gain after monotonic decay.

In addition, when using the set of weighting coefficients of Equations (7) and Equations (8), which minimizes Equation (6), the parameters of the decoded logarithmic gain of a remote frame (two subframes) and a conventional frame (two subframes), which is the next frame (two subframe) after the deleted frame is guaranteed to not be very different from the logarithmic gain parameter of the frame preceding the deleted frame. Therefore, even if the parameter of the decoded logarithmic gain of the next frame (two subframes) is not known, the received information (logarithmic quantized prediction residual) of the next frame (two subframes) can continue to be used efficiently, and the risk of masking performed in the wrong direction (risk of strong difference from correct decoded gain parameter), can be kept to a minimum.

(Embodiment 5)

17 is a block diagram showing a basic configuration of a speech encoding apparatus according to Embodiment 5 of the present invention. FIG. 17 shows an example of encoding masking type information E _{n + 1} for determining a set of weights by the second method described in Embodiment 3 — that is, a method by which information of the masking type of the (n-1) th frame is represented with one bit using information of the type of the prediction coefficient MA of the n-th frame.

In this case, the LPC masking unit 1003 of the previous frame finds the masking LSL of the (n-1) -th frame, as described using FIG. 13, by the weighted sum of the decoded quantized prediction residual of the current frame and the decoded quantized prediction residues two frames earlier, by M +1 frames before. Whereas in FIG. 13, the masking LSL of the nth frame is found using the encoding information of the (n + 1) th frame, in this case the masking LSL of the (n-1) th frame is found using the encoding information of the nth frame, and therefore the corresponding dependence is the dependence of the displacement by one frame. Those. the combinations α _i ^{j) and α ′ _i ^{(j) are} limited to two out of four using the code of the prediction coefficient MA of the nth frame (= the current frame) (i.e., when the prediction type MA of the nth frame is type 0, a combination of the prediction types of the MA of the (n-1) th frame and the n-th frame - either (0-0), or (0-1), and therefore the sets of weighting factors β are limited to two types), and the LPC masking unit 1003 of the previous frame produces two types of CLS masking - ω0 _n (j) and ω1 _n (j) - using these two types of sets of weighting factors β.

The masking type determination unit 1004 performs type determination based on the fact that from ω0 _n ^(j) and ω1 _n ^(j) is closer to the input parameter LSL ω _n ^(j) . The relative distances ω0 _n ^(j) and ω1 _n ^(j) from ω _n ^(j) can be determined based on a simple Euclidean distance, or based on a weighted Euclidean distance, for example, as used in quantizing the LSL of ITU-T G.729.

Next, the operation of each node of the speech encoding device shown in FIG. 17 will be described.

The input signal s _n is input to the LPC analysis unit 1001, the target vector calculation unit 1006, and the filter state update unit 1013.

The LPC analysis node 1001 performs the already known linear prediction analysis of the input signal s _n and outputs the linear prediction coefficients a _j (j = 0-M, where M is the order of the linear prediction analysis; and ₀ = 1, 0) to the impulse response calculation node 1005, to the target vector calculation node 1006 and to the LPC encoding node 1002. In addition, the LPC analysis unit 1001 converts the linear prediction coefficients a _j into the LSL parameters ω _n ^(j) and outputs them to the masking type determination unit 1004.

The LPC encoding section 1002 quantizes and encrypts the input LPCs (linear prediction coefficients), and outputs the quantized linear prediction coefficients a ′ _j to the impulse response calculation section 1005, the target vector calculation section 1006, and the synthesis filter section 1011. In this example, LPC quantization and coding are performed in the region of the NUM parameter. In addition, the LPC encoding section 1002 outputs the LPC encoding result L _n to the multiplexing section 1014, and outputs the quantized prediction remainder x _n , the decoded quantized LSN parameter ω ′ _n ^(j) and the MA prediction quantization type K _n to the previous LPC masking section 1003 frame.

Node masking LPC 1003 stores the previous frame decoded quantized parameter ω _'n ^(j) CHSL n-th frame outputted from LPC encoding unit 1002, a buffer for the duration of two frames. The decoded quantized parameter of the PSL is two frames earlier than -ω ′ _n-2 ^(j) . In addition, the LPC masking unit 1003 of the previous frame stores the decoded quantized prediction residual x _{n of the} nth frame for a duration of M + 1 frames. In addition, the LPC masking unit 1003 of the previous frame produces decoded quantized parameters ω0 _n (j) and ω1 _n (j) of the (n-1) -th frame number by weighted summation of the quantized prediction residual x _n decoded the quantized parameter ω ′ _n-2 ^{(j) the} LSL two frames earlier, and the decoded quantized residuals x _n-2 -x _nM-1 predictions two frames earlier — M + 1 frames earlier, and outputs the result to masking type determination unit 1004. In this case, the LPC masking unit 1003 of the previous frame provides four types of sets of weights when a weighted sum is found, but two of these four types are selected according to whether the information K _n of the predicted quantization type MA entered from the LPC encoding unit 1002 is 0 or 1, and used to generate ω0 _n ^(j) and ω1 _n ^(j) .

Unit 1004 determine the type of concealment specifies which of the two parameters ω0 _n ^(j) and ω1 _n ^(j) masking CHSL output from node 1003 concealment LPC previous frame closer to the unquantized parameter ω _n ^(j) CHSL output from node 1001 Analysis KLP, and outputs the code E _n corresponding to the set of weights, with the help of which they create a closer masked parameter LSL, in the node 1014 multiplexing.

The impulse response calculation unit 1005 creates a perceptually weighted impulse response h of the synthesis filter using the non-quantized linear prediction coefficients a _j derived from the LPC analysis node 1001 and the quantized linear prediction coefficients a ′ _j output from the LPC encoding node 1002 and outputs them to the node 1007 encoding the HACC and to node 1008 encoding the HACC.

The target vector calculating unit 1006 calculates the target vector O (a signal in which the characteristic in the absence of an input signal of a perceptually weighted synthesis filter is subtracted from a signal resulting from applying the perceptually weighted filter to an input signal) from an input signal s _n , non- _quantized linear prediction coefficients a _j , output from the LPC analysis unit 1001, and the quantized coefficients a _'j a linear prediction output from LPC encoding unit 1002, and outputs them to the coding unit 1007 Vacca, a coding unit 1009 oeffitsienta gain in unit 1012 updates the filter status.

The VACC coding unit 1007, which has the target vector ○ from the target vector calculation unit 1006, the impulse response h of a perceptually weighted synthesis filter from the impulse response calculation unit 1005 and the excitation signal ex excitation from the excitation generation unit 1010 as input information, searches the adaptive codebook and outputs the resultant adaptive codebook code A _n to the multiplexing unit 1014, the quantized pitch period T to the CKPK coding unit 1008, the vector v AKK to the generation unit 1010 is excited ia, the component p of the filtered ACC vector, to calculate which the convolution of the impulse response h of the perceptually weighted synthesis filter is performed for the vector v ACC - to the filter state update unit 1012 and the gain encoding unit 1009, and the target vector ○ ′ updated to search in Permanent codebook, used for the CCPK coding unit 1008. More specifically, the search method is similar to the method described in ITU-T G.729, etc. Although omitted in FIG. 17, typically the amount of computation required to search in the adaptive codebook does not increase in determining the range in which the pitch search with feedback is performed by searching for the pitch frequency without feedback or the like.

The CKP coding unit 1008, which has the target constant codebook vector ○ ′ and the quantized pitch period T as input from the VAKK coding unit 1007, and the impulse response h of a perceptually weighted synthesis filter as input from the impulse response computing unit 1005, performs search in the constant codebook using the method described, for example, in ITU-T Recommendation G.729, and outputs the code F _{n of the} constant codebook to multiplexing unit 1014, the PAC vector u to node 1010 excitation generation, and component q of the filtered GAC vector, obtained by convolution of the impulse response of a perceptually weighted synthesis filter for the GAC vector u, to filter state updating unit 1012 and gain coefficient encoding unit 1009.

The gain coding unit 1009, which has the target vector ο as input from the target vector computing unit 1006, the p component of the ACC filtered vector as input from the HACC coding unit 1007, and the q component of the filtered PAC vector as input from the node VPKK coding 1008, outputs a pair of ga and gf for which the value | ○ - (ga × p + gf × q | ² becomes minimal, to the excitation generation unit 1010, as a quantized adaptive codebook gain and quantized th constant codebook gain.

The excitation generation unit 1010, which has the adaptive codebook vector v as input information from the HACC coding unit 1007, the constant codebook vector u as input information from the HCC coding unit 1008, the adaptive codebook gain vector ga and the constant codebook gain vector gf as input from the gain coding unit 1009, computes the excitation vector ex ga × v + gf × u, and outputs it to the HACC coding unit 1007 and the synthesis filter unit 1011. The excitation vector e _x output to the VACC coding unit 1007 is used to update the ACB (buffer of the excitation vector generated in the past) in the VACC coding unit.

The synthesis filter unit 1011 controls a linear prediction filter configured by quantized linear prediction coefficients a ′ _j output from the LPC coding unit 1002 by an excitation vector exited from the excitation generation unit 1010, produces a locally decoded speech signal s ′ _n , and outputs it to filter status update node 1013.

The filter state updating unit 1012, which has the synthesized adaptive codebook vector p as input from the HACC coding unit 1007, the synthesized constant codebook vector q as input from the HCC coding unit 1008 and the target vector ○ as input from the node 1006 computing the target vector, generates a state of a perceptually weighted filter in the node 1006 calculating the target vector and outputs it to the node 1006 calculating the target vector.

The filter state update section 1013 computes the remainder between the locally decoded speech signal s ′ _n input from the synthesis filter section 1011 and the input signal s _n and outputs it to the target vector calculation section 1006 as the synthesis filter state in the target vector calculation section 1006.

The multiplexing unit 1014 outputs encoded information in which the codes F _n , G _n , L _n, and B are multiplexed.

In this embodiment, an example is shown in which the remainder related to the non-quantized parameter LSN is calculated only for the decoded quantized parameter LSN of the (n-1) -th frame, but it can also be ensured that the type of masking is determined by considering the remainder between the decoded quantized parameter LSL of the nth frame and the non-quantized parameter of the nth of the nth frame.

Thus, according to the speech encoding apparatus of this embodiment, the optimal set of masking processing weights is identified for masking processing for the speech decoding apparatus of Embodiment 3, and this information is transmitted to the decoder side, making it possible to provide higher masking efficiency and improved decoding quality of the speech signal on the side of the decoder.

(Embodiment 6)

FIG. 18 is a block diagram showing a configuration of a voice transmission apparatus and a speech reception apparatus that form a speech transmission system according to Embodiment 6 of the present invention. The only difference from the conventional system is that the speech encoding device of Embodiment 5 is applied to the speech signal transmission device, and the speech decoding device of any of embodiments 1-3 is applied to the speech signal receiving device.

The speech signal transmission device 1100 has an information input device 1101, an A / D conversion device 1102, a speech encoding device 1103, a signal processing device 1104, an RF modulation device 1105, a transmitting device 1106, and an antenna 1107.

The input of the A / D conversion device 1102 is connected to the information input device 1101. The input of the speech encoding device 1103 is connected to the output of the A / D conversion device 1102. The input of the signal processing device 1104 is connected to the output of the speech encoding device 1103. The input of the RF modulation device 1105 is connected to the output of the signal processing device 1104. The input of the transmitting device 1106 is connected to the output of the RF modulation device 1105. The antenna 1107 is connected to the output of the transmitting device 1106.

The information input device 1101 receives a speech signal, converts it into an analog speech signal, which is an electrical signal, and provides this signal to an A / D conversion device 1102. An A / D conversion device 1102 converts an analog speech signal from an information input device 1101 to a digital speech signal, and provides this signal to a speech encoding device 1103. The speech encoding device 1103 encodes the speech signal digitally from the A / D conversion device 1102 and produces a speech encoded binary signal stream, and provides this binary signal stream to the signal processing device 1104. The signal processing device 1104 performs channel coding processing, packetization processing, transmission buffering processing, etc. a speech encoded binary signal stream from a speech encoding device 1103, and then provides this speech encoded binary signal stream to an RF modulation device 1105. The RF modulation device 1105 modulates the speech encoded binary stream of the signal from the signal processing device 1104, in which channel coding processing, etc., is performed and provides a signal to the transmitter 1106. The transmitter 1106 transmits a modulated speech encoded binary signal from the device 1105 RF modulation as a radio wave (RF signal) through an antenna 1107.

In the device 1100 for transmitting a speech signal, the digital signal received through the device 1102 A / D conversion is processed in blocks of a frame with a duration of several tens of milliseconds (ms). If the network that forms the system is a packet-switched network, then one frame or several frames of encoded data are placed in one packet, and this packet is transmitted to the packet network. If the network is a circuit-switched network, packetization processing and transmission buffering processing are not required.

The speech signal receiving device 1150 has an antenna 1151, a receiving device 1152, an RF demodulation device 1153, a signal processing device 1154, a speech decoding device 1155, a D / A conversion device 1156, and an information output device 1157.

The input of the receiving device 1152 is connected to the antenna 1151. The input of the device 1153 RF demodulation is connected to the output of the receiving device 1152. Two inputs of the device 1154 signal processing are connected to the two outputs of the device 1153 RF demodulation. Two inputs of speech decoding device 1155 are connected to two outputs of signal processing device 1154. The input of the D / A conversion device 1156 is connected to the output of the speech decoding device 1155. The input of the information output device 1157 is connected to the output of the D / A conversion device 1156.

A receiving device 1152 receives a radio wave (RF signal), which includes speech encoded information, through an antenna 1151 and produces a coded received speech signal, which is an analog electrical signal, and provides this signal to an RF demodulation device 1153. If there is no signal attenuation or noise over the transmission path, then the radio wave (RF signal) received through the antenna is exactly the same as the radio wave (RF signal) transmitted by the voice signal transmitting device.

An RF demodulation device 1153 demodulates the encoded speech signal received from the receiver 1152 and provides this signal to the signal processing device 1154. The RF demodulation apparatus 1153 also separately provides the signal processing apparatus 1154 with information on whether or not to demodulate a received encoded speech signal in a conventional manner. The signal processing device 1154 performs buffering processing to reduce phase jitter, packet assembly processing, channel decoding processing, etc., received from the encoded speech signal demodulation device 1153, and provides a received speech encoded binary stream to the speech decoding device 1155. In addition, information about whether or not the received encoded speech signal can be demodulated in the usual way is input from the RF demodulation device 1153, and if the information input from the RF demodulation device 1153 indicates that “demodulation cannot be performed in the usual way”, or if packet assembly processing or similar processing in the signal processing device cannot be performed in the usual way and the encoded binary signal stream of the received speech cannot be decoded in the usual way, the appearance of the deleted frame is transmitted to the device 1155 decoding speech as frame deletion information. The speech decoding apparatus 1155 performs decoding processing of a binary coded speech signal encoded speech signal processing apparatus 1154 and produces a decoded speech signal, and provides this signal to a D / A conversion apparatus 1156. The speech decoding apparatus 1155 determines whether decoding processing should be performed in the usual way or decoding processing by the frame deletion masking processing in accordance with the frame deletion information inputted in parallel with the received speech encoded bit string. A D / A conversion device 1156 converts the digital decoded speech signal from the speech decoding device 1155 to an analog decoded speech signal, and provides this signal to the information output device 1157. An information output device 1157 converts an analog decoded speech signal from a D / A conversion device 1156 into air vibrations, and outputs them as a sound wave heard by the human ear.

Thus, by providing a speech encoding device and a speech decoding device shown in Embodiments 1-5, a better quality decoded speech signal can be obtained than before, even if an error occurs in the transmission path (in particular, a frame deletion error that is typical representative packet loss).

(Embodiment 7)

In the above embodiments 1-6, cases are described in which an MA type prediction model is used, but the present invention is not limited to it, and an AR type prediction model can also be used. In embodiment 7, a case is described in which an AR-type prediction model is used. With the exception of the internal configuration of the LPC decoding assembly, the configuration of the speech decoding apparatus according to Embodiment 7 is identical to the configuration shown in FIG.

FIG. 19 is a drawing showing an internal configuration of an LPC decoding unit 105 from a speech decoding apparatus according to this embodiment. Parts of the configuration of FIG. 19, which are common to FIG. 2, have the same reference numbers as in FIG. 2, and a detailed description thereof is omitted in this paper.

The LPC decoding unit 105 shown in FIG. 19 uses a configuration in which, in comparison with FIG. 2, parts related to prediction (buffer 204, amplifiers 205 and adder 206) and parts related to masking frame deletion (node Code vector decoding 203 and buffer 207) were deleted, and parts of the configuration replacing them (code vector decoding unit 1901, amplifier 1902, adder 1903, and buffer 1904) were added.

The LPC code L _{n + 1} is entered into the buffer 201 and the code vector decoding section 1901, and the frame deletion code B _{n + 1} is entered into the buffer 202, the code vector decoding section 1901 and the selection block 209.

The buffer 201 stores the LPC code L _{n + 1 of the} next frame for the duration of one frame and then outputs this LPC code to the code vector decoding section 1901. As a result of the code being stored in the buffer 201 for the duration of one frame, the LPC code output from the buffer 201 to the code vector decoding section 1901 is the LPC code L _{n of the} current frame.

The buffer 202 stores the next frame deletion code B _{n + 1} for the duration of one frame, and then outputs this deletion code to the code vector decoding section 1901. As a result of the code being stored in the buffer 202 for the duration of one frame, the frame deletion code output from the buffer 202 to the code vector decoding unit 1901 is the deletion code B _{n of the} current frame.

A code vector decoding unit 1901, which has a decoded vector y _{n − 1} NUM of the previous frame, LPC code L _{n + 1 of the} next frame, next frame delete code B _{n + 1} , LPC code L _{n of the} current frame and current frame delete code B _n as input, generates a vector x _n of the quantized prediction residual of the current frame, based on these items of information, and outputs the vector x _n of the quantized prediction residual of the current frame to the adder 1903. The node 1901 codevector decoding Details will be given later in this slave those.

An amplifier 1902 multiplies the decoded vector y _{n − 1 of the} PSL of the next frame by a predetermined MA prediction coefficient α ₁ and outputs the result to an adder 1903.

An adder 1903 calculates the sum of the predicted LSL vectors output from the amplifier 1902 (i.e., the result of multiplying the decoded LSL vectors of the previous frame by an AR-type prediction coefficient) and the vector x _{n of the} quantized prediction residual of the current frame, and outputs the result of multiplying the decoded LSL vector y _n to the buffer 1904 and the LPC conversion node 208.

Buffer 1904 stores a decoded vector y _n CHSL for the duration of one frame, and then outputs the decoded vector in CHSL decoding unit 1901 and the code vector to the amplifier 1902. As a result of that vector is stored in the buffer 1904 for the duration of one frame, the decoded vector The LSN input to the code vector decoding section 1901 and amplifier 1902 is the decoded LSL vector y _{n − 1 of the} previous frame.

If the selection unit 209 selects the decoded LPC parameter in the previous frame output from the buffer 210, then it is not actually necessary to perform all the processing from the code vector decoding section 1901 to the LPC conversion section 208.

Next, the internal configuration of the code vector decoding unit 1901 shown in FIG. 19 will be described in detail using the block diagram of FIG. 20.

The codebook 2001 generates a code vector identified LPC code L _n of the current frame, and outputs it to the switch 309, and also generates a code vector identified code L _{n + 1} LPC next frame, and outputting it to the amplifier 2002. In addition, the code a book may have a multi-stage configuration and may have a split configuration.

Amplifier 2002 multiplies the code vector x _{n + 1} output from codebook 2001 by the weight coefficient b ₀ and outputs the result to adder 2005.

Amplifier 2003 performs processing to search for the vector of the quantized prediction residual in the current frame necessary to generate the decoded NUM vector of the previous frame. Those. the amplifier 2003 computes the vector x _{n of the} current frame so that the decoded vector y _{n-1 of the} number of the previous frame becomes the decoded vector y _{n of the} number of the current frame. In particular, amplifier 2003 multiplies the input decoded vector y _{n − 1 of the} PSL of the previous frame by a factor (1-a ₁ ). Then, amplifier 2003 outputs the result of this calculation to switch 309.

Amplifier 2004 multiplies the input decoded vector y _{n-1 of the} PSL of the previous frame by the weight coefficient b _-1 , and outputs the result to adder 2005.

Adder 2005 calculates the sum of vectors output from amplifier 2002 and amplifier 2004, and outputs a code vector that is the result of this calculation to switch 309. That is, the adder 2005 calculates the vector x _{n of the} current frame, performing a weighted summation of the code vector identified by the LPC code L _{n + 1 of the} next frame and the decoded PSL vector of the previous frame.

If the current frame deletion code B _n indicates that “the nth frame is a normal frame”, then the switch 309 selects a code vector output from the codebook 2001 and outputs it as the vector x _{n of the} quantized prediction residual of the current frame. On the other hand, if the current frame deletion code B _n indicates that “the nth frame is the deleted frame”, then the switch 309 further selects a vector to be output according to what information the next frame deletion code B _{n + 1 has} .

Those. if the next frame deletion code B _{n + 1} indicates that the “(n + 1) th frame is the deleted frame”, then the switch 309 selects a vector output from the encoding device 2003 and outputs it as a vector x _{n of a} quantized prediction residual of the current frame . In this case, it is not necessary to process the vector generation process from the codebook 2001 and amplifiers 2002 and 2004 through the adder 2005. In addition, in this case, since y _n-1 can be used as y _n , x _n does not need to be generated with using processing in amplifier 2003.

On the other hand, if the next frame deletion code B _{n + 1} indicates that the “(n + 1) th frame is a normal frame”, then the switch 309 selects a vector output from the adder 2005 and outputs it as a quantized remainder vector x _n predictions of the current frame. In this case, processing with the amplifier 2003 is not required.

In the masking processing of this embodiment, the weighting factors b _-1 and b _{0 are} determined so that the sum D (where D are found, as shown in Equation (9) below) of the distance between the decoded parameter y _n-1 (n-1) -th frame and the decoded parameter y _{n of the} n-th frame, and the distance between the decoded parameter y _{n of the} n-th frame and the decoded parameter y of the _{n + 1} (n + 1) -th frame was small, so that the changes between the decoded frame parameters become insignificant .

An example of a method for determining the weights b _-1 and b _{0 is} shown below. To minimize D in Equation (9), Equation (10) below is solved for the decoded vector x _{n of the} quantized prediction residual of the deleted nth frame. As a result, x _n can be found using Equation (11) below. If the prediction coefficients differ in each group, then Equation (9) is replaced by Equation (12). In this case, a ₁ represents the AR-type prediction coefficient, and a ₁ ^(j) represents the jth element of the set of AR-type prediction coefficients (i.e., the coefficient times y _n-1 ^(j) , the jth element decoded parameter y _n-1 NLS) of the previous frame.

The elements x, y, and a in the above equations are as follows.

x _n ^(j) : quantized remainder of the prediction of the jth component of the LSL parameter in the nth frame.

y _n ^(j) : jth component of the decoded LSL parameter in the nth frame.

a ₁ ^(j) : jth component of the AR-type prediction coefficient set.

Thus, according to this embodiment, in which the AR-type prediction model is used, when the current frame is deleted, if the next frame is received in the usual way, then the masking processing of the decoded quantized residual of the prediction parameter of the NLS parameter of the current frame is performed by weighted summation processing (weighted linear summation processing) ), in particular, for masking processing using a parameter decoded in the past and a quantized prediction remainder of the next frame, and q CHSL parameter encoding is performed using the quantized prediction residual masked. Thus, it is possible to provide a higher masking efficiency than by reusing the decoded LSL parameters in the past.

You can also apply the description of embodiments 2-4 to an embodiment that uses AR type prediction, in which case the same result can also be obtained as described above.

(Embodiment 8)

In the above embodiment 7, a case is described in which there is only one kind of prediction coefficient set, but the present invention is not limited thereto, and it can also be applied to a case in which there are many kinds of sets of prediction coefficient, in the same manner as in the embodiments embodiments 2 and 3. In Embodiment 8, an exemplary case will be described in which AR-type prediction is used for which there are many kinds of sets of prediction coefficients.

21 is a block diagram of a speech decoding apparatus according to this embodiment. With the exception of the difference in the internal configuration of the LPC decoding section and the absence of a masking type E _{n + 1} input line from the demultiplexing section 101 to the LPC decoding section 105, the configuration of the speech decoding apparatus 100 shown in FIG. 21 is identical to the configuration in FIG. 11.

FIG. 22 is a drawing that shows the internal configuration of an LPC decoding unit 105 from a speech decoding apparatus according to this embodiment. Parts of the configuration in FIG. 22, which are common with FIG. 19, have the same reference numbers as in FIG. 19, and a detailed description thereof is omitted in this paper.

The LPC decoding section 105 shown in FIG. 22 uses a configuration in which, compared to FIG. 19, a buffer 2202 and a coefficient decoding section 2203 are added. In addition, the operation and internal configuration of the code vector decoding unit 2201 in FIG. 22 is different from the operation and configuration of the code vector decoding unit 1901 in FIG. 19.

The LPC code V _{n + 1} is entered into the buffer 201 and the code vector decoding section 2201, and the frame deletion code B _{n + 1} is entered into the buffer 202, the code vector decoding section 2201 and the selection block 209.

The buffer 201 stores the LPC code V _{n + 1 of the} next frame for the duration of one frame, and then outputs this LPC code to the code vector decoding section 2201. As a result of the fact that the code is stored in the buffer 201 for the duration of one frame, the LPC code output from the buffer 201 to the code vector decoding unit 2201 is the LPC code V _{n of the} current frame. In addition, the buffer 202 stores the deletion code B _{n + 1 of the} next frame for the duration of one frame, and then outputs this deletion code to the code vector decoding section 2201.

Node 2201 decoding a code vector that has a decoded vector y _n-1 NUM of the previous frame, VL code _{+ 1} LPC of the next frame, code B _{n + 1} deletion of the next frame, code V _n LPC of the current frame, code K _{n + 1} prediction coefficient of the next frame and the current frame deletion code B _n as input, produces a vector x _{n of the} quantized prediction residual of the current frame based on these pieces of information, and outputs the vector x _{n of the} quantized prediction residual of the current frame to adder 1903. Details of decoding section 2201 the code vector will be given later in this paper.

Buffer 2202 stores the AR type prediction coefficient code K _{n + 1} for the duration of one frame, and then outputs this AR type prediction coefficient code to the coefficient decoding section 2203. As a result, the AR-type prediction coefficient code output from the buffer 2202 to the coefficient decoding unit 2203 is the AR-type prediction coefficient code K _n of the previous frame.

The coefficient decoding unit 2203 stores a plurality of kinds of coefficient sets, and identifies the coefficient set by frame deletion codes B _n and B _{n + 1} and AR type prediction coefficient codes K _n and K _{n + 1} . In this case, there are three ways in which the identification of sets of coefficients can be performed in the coefficient decoding unit 2203 as follows.

If the input frame deletion code B _n indicates that “the nth frame is a normal frame”, then coefficient decoding section 2203 selects a set of coefficients defined by AR-type prediction coefficient code K _n .

If the input frame deletion code B _n indicates that “the nth frame is a deleted frame” and the frame deletion code B _{n + 1} indicates that “the (n + 1) th frame is a normal frame”, then coefficient decoding section 2203 determines the set of coefficients to be selected using the AR-type prediction coefficient code K _{n + 1} , adopted as a parameter of the (n + 1) -th frame. Those. K _{n + 1 is} used directly instead of the AR type prediction coefficient code K _n . Alternatively, it is possible to provide a set of coefficients that will be used in this case, which is determined in advance, and this previously determined set of coefficients will be used without taking into account K _{n + 1} .

If the input frame deletion code B _n indicates that “the nth frame is a deleted frame” and the frame deletion code B _{n + 1} indicates that “the (n + 1) th frame is a deleted frame”, then the only information that can be used is the information of the coefficient set used by the previous frame, and therefore, the coefficient decoding unit 2203 repeatedly uses the coefficient set used by the previous frame. Alternatively, a predetermined set of coefficients can be provided so that it is used continuously.

Then, the coefficient decoding unit 2203 outputs the AR type prediction coefficient a ₁ to the amplifier 1902, and outputs the AR type prediction coefficient a ₁ to the decoding unit of the code vector 2201.

Amplifier 1902 multiplies the decoded vector y _{n − 1 of the} PRF of the previous frame by an AR-type prediction coefficient a ₁ , input from coefficient decoding unit 2203, and outputs the result to adder 1903.

Next, the internal configuration of the code vector decoding unit 2201 shown in FIG. 22 will be described in detail using the block diagram of FIG. 23. Parts of the configuration of FIG. 23, which are common to FIG. 20, have the same reference numbers as in FIG. 20, and a detailed description thereof is omitted in this paper. The code vector decoding section 2201 of FIG. 23 uses a configuration in which a coefficient decoding section 2301 is added to the code vector decoding section 1901 of FIG.

The coefficient decoding section 2301 stores a plurality of kinds of coefficient sets, identifies the coefficient set by the AR type prediction coefficient code K _{n + 1} , and outputs it to the amplifiers 2002 and 2004. It is also possible to calculate the used coefficient set using the AR type prediction coefficient a ₁ , output from coefficient decoding unit 2203 when it is not necessary to store sets of coefficients, and the calculation can be performed after input of AR-type prediction coefficient a ₁ . Details of the calculation method will be given later in this paper.

In codebook 2001, a code vector identified by the LPC code V _{n of the} current frame is generated and output to switch 309, and a code vector identified by the LPC code V _{n + 1 of the} next frame is generated and output to amplifier 2002. In addition, the code a book may have a multi-stage configuration and may have a split configuration.

Amplifier 2002 multiplies the code vector x _{n + 1} output from codebook 2001 by the weight coefficient b ₀ and outputs the result to adder 2005.

The amplifier 2003 multiplies the AR-type prediction coefficient (1-a ₁ ) output from the coefficient decoding unit 2203 by the decoded vector y _{n-1 of the} number of the previous frame and outputs the result to switch 309. In terms of implementation, if this type of route is not created and the switching configuration is provided so that the output of the buffer 1904 is replaced by the output of the adder 1903 and injected into the LPC conversion unit 208 instead of processing in the amplifier 2003, the amplifier 1902 and the adder 1903, then the route through the amplifier 2003 is not needed.

The amplifier 2004 multiplies the input decoded vector y _{n-1 of the} PSL of the previous frame by the weight coefficient b _-1 output from the coefficient decoding unit 2301, and outputs the result to the adder 2005.

In the masking processing of this embodiment, the weighting factors b _-1 and b _{0 are} determined so that the sum D (where D is found, as shown in Equation (13) below) of the distance between the decoded parameter y _n-1 (n-1) -th frame and the decoded parameter y _{n of the} n-th frame, and the distance between the decoded parameter y _{n of the} n-th frame and the decoded parameter y of the _{n + 1} (n + 1) -th frame was small, so that the changes between the decoded frame parameters were small .

An example of a method for determining the weights b _-1 and b _{0 is} shown below. To minimize D in Equation (13), Equation (14) below is solved for the decoded vector x _{n of the} quantized prediction residual of the deleted n-th frame. As a result, x _n can be found using Equation (15) below. If the prediction coefficients differ in each group, then Equation (13) is replaced by Equation (16). In this case, a ′ ₁ represents the AR type prediction coefficient in the (n + 1) th frame, a ₁ represents the AR type prediction coefficient in the nth frame, and a ₁ ^(j) represents the jth element of the set of prediction coefficients AR type (i.e., the coefficient multiplied by y _n-1 ^(j) , the jth element of the decoded parameter y _n-1 NUM of the previous frame)

The elements x, y, and a in the above equations are as follows:

x _n ^(j) : quantized remainder of the prediction of the jth component of the LSL parameter in the nth frame.

y _n ^(j) : jth component of the decoded LSL parameter in the nth frame.

and ₁ ^(j) : the jth component of the set of AR-type prediction coefficients of the nth frame.

a ′ ₁ ^(j) : the jth component of the set of prediction coefficients of the AR type of the (n + 1) th frame.

In this case, if the nth frame is a remote frame, then the set of prediction coefficients of the nth frame is unknown. There are many possible ways to determine a ₁ . Firstly, there is a method by which a _{1 is} sent as additional information in the (n + 1) th frame. However, an extra bit is needed, and modification on the encoder side is also needed. Then there is a method by which the prediction coefficient used by the (n-1) th frame is used, and there is also a method by which the set of prediction coefficients adopted in the (n + 1) th frame is used. In this case, a ₁ = a ′ ₁ . In addition, there is a method by which a specific set of prediction coefficients is always used. However, as described later in this paper, even if another a _{1 is used} , the decoded y _n ′ will be equal when performing AR-type prediction using the same a ₁ . In the case of prediction quantization using the AR type prediction, the quantized prediction residual x _n does not correlate with the prediction, and only the decoded quantized parameter y _n correlates with the prediction, and therefore in this case a ₁ can have an arbitrary value.

If a _{1 is} determined, then b ₀ and b ₁ can be determined from Equation (15) or Equation (16), and a code vector x _{n of the} deleted frame can be generated.

If the frame removal code vector x _n obtained by the above Equation (16) is replaced in an equation representing y _n (y _n = a ₁ y _n-1 + x _n ), then the result is found as shown in the Equation ( 17). Therefore, the decoded parameter in the remote frame generated by the masking processing can be found directly from x _{n + 1} , y _n-1 and a ′ ₁ . In this case, masking processing that does not use the prediction coefficient a ₁ in the remote frame becomes possible.

Thus, according to this embodiment, in addition to providing the features described in Embodiment 7, a plurality of sets of prediction coefficients for performing masking processing are provided, and masking processing is performed, providing even higher masking efficiency than in Embodiment 7.

(Embodiment 9)

In the above embodiments 1-8, cases are described in which the nth frame is decoded after the received (n + 1) th frame, but the present invention is not limited thereto, and it is also possible to generate the nth frame using the decoded parameter the (n-1) -th frame, decode the parameter of the nth frame using the method of the present invention while decoding the (n + 1) -th frame, and decode the (n + 1) -th frame after updating the internal state of the prediction device using this result.

In Embodiment 9, this case will be described. The configuration of the speech decoding apparatus according to Embodiment 9 is identical to that of FIG. 1. In addition, the configuration of the LPC decoding unit 105 may be identical to the configuration in FIG. 19, but it is re-depicted as shown in FIG. 24 to explain that decoding of the (n + 1) th frame is performed for inputted encoded information (n + 1st) frame.

24 is a block diagram showing an internal configuration of an LPC decoding unit 105 from a speech decoding apparatus according to this embodiment. Parts of the configuration in FIG. 24, which are common with FIG. 19, have the same reference numbers as in FIG. 19, and a detailed description thereof is omitted in this paper.

The LPC decoding section 105 shown in FIG. 24 uses a configuration in which, compared to FIG. 19, the buffer 201 is removed, the output of the code vector decoding section is x _{n + 1} , the decoded parameter is the (n + 1) -th parameter frame (y _n ), and a switch 2402 is added. In addition, the operation and internal configuration of the code vector decoding unit 2401 in FIG. 24 differs from the operation and configuration of the code vector decoding unit 1901 in FIG. 19.

The LPC code L _{n + 1} is input to the code vector decoding unit 2401, and the frame deletion code B _{n + 1} is input to the buffer 202, the code vector decoding unit 2401, and the selection unit 209.

The buffer 202 stores the deletion code B _{n + 1 of the} current frame for the duration of one frame, and then outputs this deletion code to the code vector decoding section 2401. As a result of the code being stored in the buffer 202 for the duration of one frame, the frame deletion code output from the buffer 202 to the code vector decoding unit 2401 is the deletion code B _{n of the} previous frame.

The code vector decoding unit 2401, which has the decoded vector y _n-1 NLS two frames earlier, the LPC code L _{n + 1 of the} current frame and the code B _{n + 1 to} delete the current frame as input, produces a quantized remainder vector x _{n + 1} predicting the current frame and a decoded vector y _'n CHSL previous frame, based on these items of information, and outputs them to the adder 1903 and the switch 2402. the node 2401 Details of decoding the code vector will be given later in this paper.

An amplifier 1902 multiplies the decoded vector y _{n − 1} or y ′ _{n of the} LSL of the previous frame by a predetermined AR-type prediction coefficient a ₁ and outputs the result to adder 1903.

An adder 1903 calculates a predicted LSL vector output from an amplifier 1902 (i.e., the result of multiplying a decoded LSL vector of a previous frame by an AR-type prediction coefficient), and outputs the result of this calculation, the decoded LSL vector y _{n + 1} , to a buffer 1904 and a node 208 conversion of KLP.

The buffer 1904 stores the decoded vector LSN _{+ 1} LSL of the current frame for the duration of one frame, and then outputs this decoded LSL vector to the code vector decoding section 2401 and switch 2402. As a result, the vector is stored in the buffer 1904 for the duration of one frame , the decoded LSL vector input to the code vector decoding section 2401 and the LSL switch 2402 is the vector y _{n of the} previous frame.

Switch 2402 selects either the decoded vector y _n CHSL previous frame or the decoded vector y _'n CHSL previous frame generated node 2401 decode the code vector of the current frame using code L _{n + 1} LPC, according to the code B _n delete the previous frame. If B _n indicates the deleted frame, then switch 2402 selects y ′ _n .

If the selection unit 209 selects the decoded LPC parameter of the previous frame output from the buffer 210, then it is not actually necessary to perform all the processing from the code vector decoding unit 2401 to the LPC conversion unit 208.

Next, the internal configuration of the code vector decoding unit 2401 shown in FIG. 24 will be described in detail using the block diagram of FIG. 25. Parts of the configuration of FIG. 25, which are common with FIG. 20, have the same reference numbers as in FIG. 20, and a detailed description thereof is omitted in this paper. The code vector decoding unit 2401 in FIG. 25 uses a configuration in which a buffer 2502, an amplifier 2503, and an adder 2504 are added to the code vector decoding unit 1901 in FIG. In addition, the operation and internal configuration of switch 2501 in FIG. 25 are different from the operation and configuration of switch 309 in FIG.

In codebook 2001, a code vector identified by the LPC code L _{n + 1 of the} LPC of the current frame is generated and output to switch 2501, as well as to amplifier 2002.

Amplifier 2003 performs processing to search for the vector of the quantized prediction residual in the current frame necessary to generate the decoded NUM vector of the previous frame. Those. the amplifier 2003 calculates a vector x _{n + 1 of the} current frame so that the decoded vector y _{n of the} LSF of the previous frame becomes the decoded vector y _{n + 1 of the} LS of the current frame. In particular, amplifier 2003 multiplies the input decoded vector y _{n of the} LSL of the previous frame by a factor (1-a ₁ ). Amplifier 2003 then outputs the result of this calculation to switch 2501.

If the current frame deletion code B _{n + 1} indicates that “the (n + 1) th frame is a normal frame”, then the switch 2501 selects a vector output from the codebook 2001 and outputs it as the vector x _{n + 1 of the} quantized remainder predictions of the current frame. On the other hand, if the deletion code B _{n + 1 of the} current frame indicates that the “(n + 1) th frame is a deleted frame”, then the switch 2501 selects the vector output from the amplifier 2003 and outputs it as the vector x _{n + 1} quantized remainder of the prediction of the current frame. In this case, the processing of the vector generation process from the codebook 2001 and from the amplifiers 2002 and 2004 through the adder 2005 should not be performed.

Buffer 2502 stores a decoded vector y _n CHSL previous frame for the duration of one frame and then outputs the decoded vector CHSL amplifier 2004 and amplifier 2503 as a decoded vector y _n-1 CHSL two frames before.

Amplifier 2004 multiplies the input decoded vector y _n-1 NLS two frames earlier by the weight coefficient b _-1 and outputs the result to adder 2005.

Adder 2005 calculates the sum of the vectors output from amplifier 2002 and amplifier 2004, and outputs a code vector that is the result of this calculation to adder 2504. That is, the adder 2005 calculates the vector x _{n of the} previous frame by performing a weighted summation of the code vector identified by the LPC code L _{n + 1 of the} current frame and the decoded NUM vector two frames earlier and outputs it to the adder 2504.

Amplifier 2503 multiplies the decoded vector y _n-1 NLS two frames earlier by a prediction coefficient a ₁ and outputs the result to adder 2504.

The adder 2504 adds together the output of the 2005 adder (decoded vector x _{n of the} previous frame, recalculated using the LPC code L _{n + 1 of the} current frame) and the output of the amplifier 2503 (vector resulting from multiplying the decoded vector y _n-1 NLS two frames earlier by prediction coefficient a _1), and recalculates the decoded vector y _'n CHSL previous frame.

The method of recalculating the decoded LSN vector Y ′ _{n of} this embodiment is the same as the masking processing in Embodiment 7.

Thus, according to this embodiment, using the configuration whereby the decoded vector x _n obtained by masking processing of Embodiment 7 is used only for the internal state of the prediction device when decoding the (n + 1) th frame, makes it possible to decrease by one frame processing delay required in embodiment 7.

(Embodiment 10)

In the above embodiments, 1-9 provide only features related to the internal configuration and processing of the LPC decoding node, but the configuration of the speech decoding device according to this embodiment has a feature related to the configuration outside the LPC decoding node. Although the present invention can be applied to any of FIG. 1, FIG. 8, FIG. 11 or FIG. 21, in this embodiment, for example, a case has been described in which the present invention is applied to FIG. 1.

Fig. 26 is a block diagram showing a speech decoding apparatus according to this embodiment. Parts of the configuration in FIG. 26, which are common with FIG. 21, have the same reference numbers as in FIG. 21, and a detailed description thereof is omitted in this paper. The speech decoding apparatus 100 shown in FIG. 26 uses a configuration in which, compared to FIG. 21, a filter gain calculating section 2601, a drive power control section 2602, and an amplifier 2603 are added.

The LPC decoding section 105 outputs the decoded LPC to the LPC synthesis section 109 and the filter gain calculation section 2601. In addition, the LPC decoding section 105 outputs the frame deletion code B _n corresponding to the decoded nth frame to the driving power control section 2602.

The filter gain factor calculating section 2601 calculates a filter gain for the synthesis filter configured by inputting the LPC from the LPC decoding section 105. As an example of a method for calculating a filter gain, there is a method by which the square root of the power of the impulse response is found and used as a filter gain. This is based on the fact that if the input signal is considered as a pulse with a power of 1, then the power of the impulse response of the synthesis filter configured by the input LPC is actually filter gain information. Another example of a method for calculating a filter gain is a method in which since the RMS residual of the linear prediction can be found from the LPC using the Levinson-Darbin algorithm, its inversion is used as filter gain information, and the square root of the inversion of the rms of the linear prediction is used in as a filter gain. The found filter gain is output to the drive power control unit 2602. The rms power of the impulse response or the remainder of the linear prediction can also be output to the excitation power control unit 2602 without finding the square root.

The drive power control unit 2602, which has a filter gain from the filter gain calculation unit 2601 as input, calculates a scaling factor for controlling the amplitude of the drive signal. The drive power control unit 2602 is provided with an internal memory and stores a filter gain of the previous frame in this memory. After the scaling factor is calculated, the contents of the memory are overwritten with the input filter gain of the current frame. The scaling factor SG _{n is} calculated using the equation SG _n = DG _max × FG _n-1 / FG _n , where FG _n is the filter gain of the current frame, FG _n-1 is the filter gain of the previous frame, and DG _max is the upper limit value gain increase rate. In this case, the gain increase rate is described as FG _n / FG _n-1 , and it indicates how many times the filter gain of the current frame is greater than the filter gain of the previous frame. The upper limit value of the rate of increase of the gain is determined in advance as DG _max . If the filter gain increases sharply relative to the filter gain of the previous frame in the synthesis filter created by the masking deletion processing, the output of the synthesis filter also increases sharply, and the decoded signal (synthesized signal) has a large amplitude in a certain place, producing an explosive sound . In order to avoid this, if the gain of the synthesis filter configured by the decoded LPC generated by the frame deletion mask processing exceeds a predetermined gain increase rate relative to the filter gain of the previous frame, then the power of the decoded excitation signal, which is the synthesis filter control signal, is reduced . For this purpose, a scaling factor is used, and the predetermined rate of increase of the gain is the upper limit value DG _{max of the} rate of increase of the gain. Normally, an explosive sound can be prevented by setting the DG _max to 1 or to a value less than 1, for example, 0.98. If FG _n / FG _{n-1 is} less than or equal to DG _max , then SG _{n is} taken equal to 1, 0, and scaling is not required to be performed in amplifier 2603.

Another way to calculate the scaling factor SG _n is to use, for example, the equation SG _n = Max (SG _max , FG _n-1 / FG _n ). In this case, SG _max represents the maximum value of the scaling factor and has a value slightly greater than 1, for example, 1.5, and Max (A, B) is the function that outputs A or B, depending on which has the highest value. If SG _n = FG _n-1 / FG _n , then the excitation signal power decreases in proportion to the increase in the filter gain, and the power of the decoded synthesized signal of the current frame becomes the same as the power of the decoded synthesized signal of the previous frame. In this way, a sharp increase in the power of the synthesized signal described above can be avoided, and a sharp decrease in the power of the synthesized signal can also be avoided. In this case, if SG _n = FG _n-1 / FG _n , then SG _n has a value of 1 or higher, and plays a role in preventing local attenuation of the power of the synthesized signal. However, since the drive signal generated by the frame deletion mask processing is not necessarily suitable as the drive signal, using too large a scaling factor can lead to noticeable distortion and poor quality. Therefore, if an upper limit value of the scaling factor is provided, and FG _n-1 / FG _n exceeds this upper limit value, then FG _n-1 / FG _{n is} cut to the upper limit value.

The filter gain of the previous frame, or a parameter representing the gain of the filter (such as the power of the impulse response of the synthesis filter), can be entered from the external drive power control unit 2602, instead of being stored in the drive power control unit 2602 in memory. In particular, if the information related to the filter gain of the previous frame is used in a different way than the speech decoder, it is ensured that the parameter described above is input from the outside rather than being overwritten in the drive power control unit 2602.

Then, the drive power control unit 2602, which has the frame deletion code B _n as input from the LPC decoding unit 105, outputs the calculated scaling factor to the amplifier 2603 if B _n indicates that the current frame is a deleted frame. On the other hand, if B _n indicates that the current frame is not a remote frame, then the drive power control unit 2602 outputs “1” to the amplifier 2603 as a scaling factor.

An amplifier 2603 multiplies the scaling factor input from the drive power control unit 2602 by a decoded drive signal input from the adder 108, and outputs the result to the LPC synthesis unit 109.

Thus, according to this embodiment, if the gain of the synthesis filter configured by the decoded LPC generated by the frame deletion mask processing is changed relative to the filter gain of the previous frame, the occurrence of explosive sound or sound loss can be prevented by correcting the power of the decoded speech signal, which is the control signal of the synthesis filter.

Even if B _n indicates that the current frame is a remote frame, the drive power control unit 2602 can output the calculated scaling factor to amplifier 2603 if the immediately previous frame is a remote frame (i.e., if B _n-1 indicates that the previous frame is a deleted frame). This is because when predictive coding is used, there may be a residual effect of the remainder on the frame recovered from the deleted frame. In this case, the same result can also be obtained as described above.

This completes the description of the embodiments of the present invention.

In the above embodiments, it is assumed that the encoding parameter is an LSN parameter, but the present invention is not limited thereto, and it can be applied to a parameter of any kind as long as it is a parameter with a slight change between frames. For example, immitance spectrum frequencies (ISFs) may be used.

In the above embodiments, it is assumed that the encoding parameter is the LSN parameter itself, but the LSL parameter “after eliminating the average value”, which is the result of extracting the difference from the average LSL value, can also be used.

In addition to being used in a speech decoding apparatus and a speech encoding apparatus, it is also possible to install a parameter decoding apparatus and a parameter encoding apparatus according to the present invention in a communication terminal apparatus and a base station apparatus in a mobile communication system, whereby it is possible to provide a communication terminal apparatus, a basic apparatus stations and a mobile communication system that have the same processing efficiency as described above.

For example, a case has been described in which the present invention is configured as hardware, but it is also possible to implement the present invention with software. For example, the same kind of functions as the functions of the parameter encoding device according to the present invention can be realized by recording the algorithm of the method of decoding the parameter according to the present invention in a programming language, storing this program in memory and executing it using the information processing means.

Function blocks used in the descriptions of the above embodiments are typically embodied as LSIs (large integrated circuits), which are integrated circuits. They can be implemented separately, as single chips, or one chip can implement some or all of them.

The term LSI is used in this document, but the terms integrated circuit (IC), system LSI, ultra-large integrated circuit (VLSI), ultra-large integrated circuit (UBIS), etc. may also be used according to differences in degree of integration.

The method for implementing integrated circuits is not limited to LSIs, and implementation through specialized circuits or a universal processor may also be used. PPVM can also be used (user-programmable gate array), for which programming is possible after manufacturing the LSI, or a reconfigurable processor that allows reconfiguration of the connections between circuit elements and settings within the LSI.

In the case of the introduction of the implementation technology of an integrated circuit by which LSIs will be replaced by another technology that is a development or derivative of semiconductor technology, the integration of functional blocks can, of course, be performed using this technology. The use of biotechnology or similar technology is also possible.

Disclosures of Japanese Patent Application No. 2006-305861, filed November 10, 2006, Japanese Patent Application No. 2007-132195, filed May 17, 2007, and Japanese Patent Application No. 2007-240198, filed September 14, 2007, which include a description, drawings and abstract are presented in their entirety for reference.

Industrial applicability

A parameter decoding apparatus, a parameter encoding apparatus, and a parameter decoding method according to the present invention are suitable for use in a speech decoding apparatus and a speech encoding apparatus, and furthermore, in a communication terminal apparatus, a base station apparatus, and the like. in a mobile communication system.

Claims (7)

1. A device for decoding parameters, comprising:
a prediction residual decoding unit that finds a quantized prediction residual based on encoded information included in the current frame to be decoded; and
a parameter decoding unit that decodes the parameter based on the specified quantized prediction residual,
wherein said decoding residual decoding node, when said current frame is deleted, finds the quantized prediction residual of the current frame from the weighted linear sum of the parameter decoded in the past and the quantized prediction residual of the future frame. 2. The parameter decoding apparatus according to claim 1, wherein said decoding residual decoding node, when said current frame is deleted, finds a quantized prediction residual of the current frame so that the total sum of the distance between the decoded parameter of the past frame and the decoded parameter of the current frame and the distance between the decoded parameter of the current frame and the decoded parameter of the future frame was minimal. 3. The parameter decoding apparatus according to claim 1, wherein said decoding residual decoding unit stores a plurality of sets of weighting factors, and when said current frame is deleted, selects a set of weighting factors based on a command from the interacting party and multiplies said set of weighting factors by said a parameter decoded in the past and the specified quantized remainder of the prediction of the future frame. 4. The parameter decoding apparatus according to claim 1, wherein said decoding residual decoding node, when said current frame is deleted, finds a quantized prediction residual of a current frame from a weighted linear sum of a parameter decoded in the past, a quantized prediction residual of a past frame, and a quantized future prediction remainder frame. 5. A parameter encoding device, comprising:
an analysis node that analyzes the input signal and finds the analyzed parameter;
a coding unit that predicts the specified analyzed parameter using the prediction coefficient, and obtains a quantized parameter using the quantized prediction remainder obtained using the quantized prediction remainder and the specified prediction coefficient;
the masking unit of the previous frame, which stores a plurality of sets of weighting coefficients, finds the weighted sum using the specified sets of weighting coefficients for the specified quantized prediction residual of the current frame, the specified quantized prediction residual two frames back and the specified quantized parameter two frames back, and finds the set of specified quantized parameters previous frame using the specified weighted sum; and
a determination node that compares the plurality of said quantized parameters of the previous frame found using the specified masking unit of the previous frame, and the specified parsed parameter found using the specified analysis node of the previous frame, selects one of the specified quantized parameters of the previous frame, and selects and encodes a set of weight coefficients corresponding to the selected specified quantized parameter of the previous frame. 6. A method for decoding parameters, comprising:
a prediction residual decoding step of detecting a quantized prediction residual based on encoded information included in the current frame to be decoded; and
a parameter decoding step consisting of decoding a parameter based on the specified quantized prediction remainder,
moreover, at the indicated stage of decoding the prediction residual, when the specified current frame is deleted, the quantized prediction residual of the current frame is found from the weighted linear sum of the parameter decoded in the past and the quantized prediction residual of the future frame. 7. The method of decoding the parameters according to claim 6, in which at the indicated step of decoding the prediction residual, when the specified current frame is deleted, the quantized prediction residual of the current frame is found so that the total sum of the distance between the decoded parameter of the last frame and the decoded parameter of the current frame and the distance between the decoded parameter of the current frame and the decoded parameter of the future frame was minimal.

Images