JPH04190399A - Vselp coding system - Google Patents

Abstract

PURPOSE:To miniaturize a processing device and to reduce power consumption by changing over a fixed decimal point position of a value in each device into a suitable position in accordance with a degree of an averaged voice power of an inputted voice signal. CONSTITUTION:A Vector Sum Excited Linear Prediction (VSELP) coding system in which a parameter is coded by computation in accordance with the definite word length of a voice signal, incorporates a voice power calculating section 201, an LPC analyzing section 203, a target signal formulating section 206, a long term lag selecting section 208, a code 1 selecting section 210, a code 2 selecting section 212 and a gain selecting section 214. In this arrangement, an averaged voice power indicating an average of the degrees of inputted voice is obtained from an input voice signal, and the decimal point position of a variable used in a computing device is changed over in accordance with thus obtained averaged voice power. The computation is carried out in the expression of a fixed decimal point given by thus changed-over decimal point position. Thereby, it is possible to miniaturize the processing device and to reduce the power consumption.

Description

[Detailed description of the invention] [Industrial application field] The present invention relates to a VSELP encoding method.

``Prior art 1: The VSELP speech coding system is a type of code excitation predictive coding (CELP) that parameter-codes speech into the vocal tract and a vector quantized excitation source,
It has been standardized by the committee as a voice encoding method for digital cellular.

Document A shown below describes the conventional VSELP encoding method.

Document A, “Vector Sum Excited L
inear Prediction (VSELP)
5peech Coding at 8Kbps'', [
raA, Gerson & Mark A, Jas
iuk, Proc, IEEE Int, on,
on Acoustics, 5peech and
Signal Processing, pp, 461-
464. April 1990 Below, the prior art of the VSELP encoding method will be explained based on the above-mentioned document A.

As shown in FIG. 2, the VSELP encoder in the conventional VSELP encoding system includes an LPC analysis section 101, a target signal generation section 105, and a long term lag selection section 107.
, a code 1 selection section 109, a code 2 selection section 111, and a gain selection section 113.

The LPG analysis unit 101 performs LPG analysis on the input audio signal V every 160 samples to obtain LPG coefficients and audio average power R, and further quantizes these LPGP GE3 and audio average power R, and calculates the average audio power R from 40 samples. LPGPGE3 and audio average power Ro are output by interpolating the subframe.

Then, for each subframe, the Tarkerat signal generation unit 105 generates a residual signal of the input audio signal V using the analysis filter using the LPG PGE3, and
Weighting using the PG coefficient kJ applies a filter to the residual signal, and further removes the influence of the excitation signal ex synthesized in the previous subframe, thereby creating and outputting the target signal p.

The long-term lag selection unit 107 has the internal value long-term filter state updated by the excitation signal ex of the previous subframe, and the weighting of the target signal p and the internal value long-term filter state is performed using a filter. An optimal lag is determined by calculating the squared error of the difference, and a long-term code C8 based on the lag is output.

The code 1 selection section 109 has a vane hector inside.
The fixed data stored as ■ is combined using the gray coat θ to create a lid U to the coat, which is then weighted using the LPC coefficient kJ and filtered.
Said long term coat c. The signal f orthogonalized to
o" and evaluate the squared error of the difference between this f.' and the target signal p, the optimal foot 1 code C
, and output it.

The code 2 selection section 111 stores a different basis vector V' from the food 1 selection section.

The code 2 selection unit 111 selects this hesis vector V
' are combined using a gray food θ to create a coat vector U', which is weighted using the LPGPGE3 filter, made orthogonal to the long-term code Co, and then the optimal code obtained from the code I selection unit 109. A signal f1'' that is orthogonalized to the code c1 is created, and the square error of the difference between f1' and the target signal p is evaluated to select and output the optimal code 2 code C2.

The gain selection unit 113 selects the LPGPGE3, audio average power R0, target signal p, long term code C8, code 1 hood c8 . Code 2 code C! parameters (GS, PO, PL) on the vector table in this selection unit 113 based on these input signals.
The optimum set is selected from among them, and the gains β, γ1, . Outputs γ.

These gains β, γ1 . γ is the aforementioned hood CO+
This is a gain for combining Cl+C1 to synthesize the excitation signal ex.

The code C8 mentioned above is amplified by the gain β in the amplifier circuit 114, the code C1 is amplified by the gain γ in the amplifier circuit 115, and the needle c2 is amplified by the gain β in the amplifier circuit 115.
16 and is amplified by Keyne γ. and,
The amplified codes are added together by an adder circuit 117 to generate a composite excitation signal ex for the current subframe.

The generated synthetic excitation signal ex is used to synthesize a target signal for the next subframe and update the long-term filter state.

In the VSELP encoder as described above, the LPG coefficient kJ, average voice power R, long term lag L1 code 1 intex 11 coat 2 index J1 and gain index obtained from the input voice are transmitted to the decoder.

Also, although it is not shown in the figure, the VSELP decoder in the conventional VSELP encoding system uses the received LPGPGE
Construct a synthesis filter from 3, decode the long-term code C8 from the long-term filter state prepared by the encoder and the received lag L, and decode the long-term code C8 from the basis vector V stored in the encoder and the received code 1 index I. The code 1 code C, is decoded and the code 2 code C1 is decoded from the code 2 index J.
and decodes the Cain β, γ1, γ2 for each coat from the hector quantization table of (GS, PO, PI) and the received Casey/Tex.

Then, the decoded code C8+CI+C2 and gain β
, γ1. γ2 is used to synthesize a synthetic excitation signal ex for the current subframe, and the input audio signal V is decoded by multiplying the received LPG coefficient by the synthetic excitation signal ex through a synthesis filter configured by. The decoded audio signal is then output after further noise compression is performed by a post filter configured using the LPC coefficient kJ.

Note that in the conventional VSELP encoding method, arithmetic processing in each part of the encoder and decoder is performed using fixed-point representation in which the decimal point position is fixed in advance.

[Problems to be Solved by the Invention] Incidentally, the VSELP encoding method is considered to be mainly used for compressed voice communication in digital mobile communication.

Therefore, a processing device (
Encoders, decoders, etc.) are required to be as compact and consume as little power as possible. It is effective to make the length as short as possible and to execute arithmetic processing within the processing device using fixed-point representation.

However, in the case of the conventional VSELP encoding method, for example, if the VSELP encoder performs arithmetic processing using fixed-point representation, or if the bit length of the data to be processed is short,
When the amplitude of the input voice fluctuates greatly, the calculation accuracy decreases and the quality of the decoded voice deteriorates significantly, and as a result, it is not possible to shorten the data bit length and downsize the device. There was a problem.

The present invention has been made in view of the above-mentioned circumstances, and is capable of performing arithmetic processing in fixed-point representation and preventing a decrease in calculation accuracy due to fluctuations in the volume of input audio even if the bit length of data is shortened. It is an object of the present invention to provide a VSELP encoding method that can realize downsizing of a processing device and reduction of power consumption.

[Means for Solving the Problems] The VSELP encoding method according to the present invention does not use fixed-point representation when parameter encoding an audio signal by digital operation with a finite word length, so the bits of variables involved in the operation are This is an attempt to shorten the length.

Specifically, the voice average power indicating the average loudness of the input voice is determined from the input voice signal, and the decimal point position of the variable used within the arithmetic unit is switched according to the magnitude of the determined voice average power. It is characterized by performing arithmetic processing using fixed-point representation using switched decimal point positions.

[Operation] The VSELP encoding method according to the present invention switches the fixed point position of the value in each device to the optimal position according to the magnitude of the average audio power of the input audio signal, so that fluctuations in the magnitude of the input audio are avoided. It is possible to perform high-precision arithmetic processing with a small bit length even for data whose values fluctuate according to This can be achieved without any problems, making it possible to downsize the processing device and reduce power consumption.

[Example] The present invention was obtained as a result of analyzing the causes of the problem of a decrease in calculation accuracy that occurs when attempting to shorten the bit length in the conventional VSELP encoding method, and further intensive research. .

That is, the inventors of the present invention first found that the deterioration in the quality of decoded speech due to the shortening of the bit length of data handled in the conventional VSELP encoder is caused by LPG analysis section 1 in the figure
01, target/node synthesis section 105, long term lag selection section 107, code 1 selection section 109, two F2 selection section 1
11. It has been found that the values used internally in the gain selection section 113 overflow or underflow, and as a result, no meaningful calculations can be performed.

FIG. 3 shows the relationship between the size of data value and the position of the decimal point when arithmetic processing is performed using fixed-point representation.

If the bit length is sufficient, can both large and small values be expressed?If the bit length is limited, as shown in Figure 3, if the decimal point position is fixed at the upper position, the data value will be larger or larger. An overflow occurs when the
If the decimal point position is fixed to the lower order, an underflow will occur when the data value becomes small, so it is not possible to set a fixed decimal point position that will work in all cases.

Furthermore, the inventors of the present application have found that variables with a large dynamic range handled by the VSELP encoding method often have a strong correlation with the average power of input speech.

FIG. 4 shows, as an example, the correlation between the target signal p handled by the VSELP encoding method and the average voice power Ro.

The fact that it has a strong correlation with the average power R6 of the input voice means that if you understand the fluctuation range of each variable value with respect to the fluctuation range of the average voice power R, in advance by experiment etc. This means that by setting the optimal fixed point position based on the value of , the probability of overflow or underflow occurring due to limiting the bit length can be kept very low.

Based on the above facts, an embodiment of the present invention was proposed.

Hereinafter, based on FIG. 1, VS which is an embodiment of the present invention will be explained.
The ELP encoding method will be explained.

FIG. 1 shows the configuration of a VSELP encoder using a VSELP encoding method according to an embodiment of the present invention.

This VSELP encoder includes a voice power calculation section 201, an LPG analysis section 203, and a target signal creation section 206.
, a long-term lag selection section 208 , a court 1 selection section 210 , a court 2 selection section 212 , and a gain selection section 21
4.

When the audio power calculation section 201 receives the input audio signal (2), it calculates and outputs the average audio power R, which indicates the average loudness of the input audio, for each frame of 160 samples.

The LPC analysis section 203 calculates an LPG coefficient based on the input audio signal V. In the arithmetic processing for calculating this LPC coefficient, the fixed decimal point position of the value inside the analysis section 203 is switched in advance to an optimal position depending on the magnitude of the voice average power R.

In addition, the target signal generation unit 206 generates a residual signal of the input audio signal V by an analysis filter using the LPG coefficient kJ for each subframe, and generates a residual signal of the input audio signal V for each subframe.
Weighting using the PG coefficient kJ applies a filter to the residual signal and further removes the influence of the excitation signal ex synthesized in the previous subframe, thereby creating and outputting the target signal p. When performing the arithmetic processing to calculate p, the fixed point position of the value inside the signal generating section 206 is switched in advance to an optimal position depending on the magnitude of the audio average power R8.

The long-term lag selection unit 208 updates the internal value long-term filter state with the excitation signal ex of the previous subframe, and weights the target signal p and the internal value long-term filter state using a filter. Either the optimum lag is calculated by calculating the squared error of the difference and the long-term code C6 is output based on the lag, or the magnitude of the average voice power R is calculated in advance. Depending on the situation, the fixed-point position of the value inside the selection unit 208 is switched to an optimal position for use.

The coat 1 selection unit 210 creates a code vector u by combining fixed data stored internally as Hoenshector V using a Gray code θ, weights it using the LPG coefficient kJ, and filters it.
Signal f orthogonalized to the long term coat C6
. ', and this f. ” by evaluating the squared error of the difference between the target signal p and the target signal p.
3 or to calculate the code 1 code C1, the fixed decimal point position of the value inside the selection unit 210 is switched to the optimum position in advance according to the magnitude of the audio average power R8. It is used as such.

The code 2 selection unit 212 stores a basis vector V′ that is different from the code 1 selection unit 210.

The code 2 selection unit 212 selects this basis vector ■
``is combined with the Gray code θ to create a code vector U'', which is applied to the LPG coefficient, weighted using a filter, and the long term code C0
Create signals f and l that are orthogonalized to the optimal court 1 foot C1 obtained from the code 1 selection unit 210, and evaluate the squared error of the difference between fl'' and the target signal p. Therefore, the optimal code 2 code 1
” c, or to perform the arithmetic processing to calculate the court 2 court C2, the fixed point position of the value inside the selection unit 212 is optimized in advance according to the magnitude of the audio average power R, It is used by changing the position.

The gain selection unit 214 sets the LPG coefficient to 3, the average voice power R6, the target signal p, and the long term code C.
8. Receive code 1 code CI+ code 2 code c2 and select parameters (GS, PO, PL) on the vector table in this selection unit 214 based on these input signals.
The optimum set is selected from among them, and the keys β, γ1, . Outputs γ.

These Keynes β, γ1. γ is the aforementioned code Ca+
This is a gain for combining the CII Ct to synthesize the excitation signal ex.

Note that the gain selection unit 214 selects Cain β.

γ1. When performing the arithmetic processing for calculating γ2, the fixed point position of the value inside the selection unit 214 is switched in advance to an optimal position depending on the magnitude of the audio average power R6.

The code C8 mentioned above is amplified by the gain γ in the amplifier circuit 215, the code C is amplified by the gain γ in the amplifier circuit 216, and the code C2 is amplified by the gain γ in the amplifier circuit 216.
17, the signal is amplified by a gain γ. and,
Each amplified code will be summed by summing circuit 218 to generate the composite excitation signal ex for the current subframe.

As is clear from the above description of the VSELP encoder,
The VSELP encoding method according to an embodiment of the present invention switches the fixed point position of the value in each device to an optimal position according to the magnitude of the average audio power of the input audio signal, so that fluctuations in the magnitude of the input audio It is possible to perform arithmetic processing with high accuracy with a small bit length even on data whose values vary according to the data.

In other words, it is possible to perform arithmetic processing using fixed-point representation and shorten the bit length of data to be handled without reducing the arithmetic accuracy, and it is possible to downsize the processing device and reduce power consumption. It's now possible.

In addition, in FIG. 1, 5ELP, which is an embodiment of the present invention, is shown.
Although a VSELP encoder using a coding method has been shown, it goes without saying that the present invention is applicable not only to a V,5ELP encoder but also to a VSELP decoder.

[Effect of the invention] As is clear from the above explanation, the VSEL according to the present invention
The P encoding method switches the fixed point position of the value in each device to the optimal position according to the magnitude of the audio average power of the input audio signal. In contrast, it is possible to perform arithmetic processing with high precision with a small bit length, and it is possible to achieve arithmetic processing using fixed-point representation and a reduction in the bit length of data to be handled without reducing the arithmetic precision. It is possible to downsize the processing device and reduce power consumption.

[Brief explanation of the drawing]

FIG. 1 is an explanatory diagram of the configuration of an encoder according to an embodiment of the present invention, FIG. 2 is an explanatory diagram of the configuration of an encoder using the conventional VSELP encoding method, and FIG. 3 is an explanatory diagram of problems in fixed-point representation. , Figure 4 shows the average voice power and tarp in the VSELP encoding method. FIG. 201...Audio power calculation section, 203...
・LPC analysis section, 206... Target/to signal creation section, 208... Long term lag selection section, 21
0...--Code 1 selection section, 212...Hood 2 selection section, 214...Gain selection section. What is the input audio R0 in VSELP encoding? - Kate)
Correlation diagram with signal P Figure 4

Claims (1)

[Claims] In the VSELP encoding method in which a speech signal is parameter encoded by digital calculation using a finite word length, the speech average power indicating the average loudness of the input speech is determined from the input speech signal, and the obtained speech average power is A VSELP encoding method characterized in that the decimal point position of a variable used in an arithmetic unit is switched according to the size of the variable, and arithmetic processing is performed in fixed-point representation using the switched decimal point position.