JPH07134600A - Device for encoding voice and device for decoding voice - Google Patents

Abstract

PURPOSE:To improve the precision by switching the fixed point position even in an adaptive excitation code vector selection part, and to provide voice mean power information for every sub frame by interpolation using simple constitution or processing. CONSTITUTION:A point position selection part 207 decides the fixed point position of an internal variable in the adaptive excitation code vector selection part from fixed point position information stored in a point position buffer part 206 and the voice mean power of the present sub frame related to a synthesis excitation signal of the past sub frame stored in the adaptive excitation code vector selection part 205. Further, a voice power interpolation part 202 is provided with a quantization means 202a of at least the voice mean power, and when the voice mean power information at every sub frame is obtained, the voice mean power at every frame is dealt with a quantized index, and the arithmetical mean processing of the indices is performed for the sub frame obtaining the geometrical mean information of the voice mean power of two pieces of frames.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a speech coder and a speech decoder, for example, VSELP (Vector Sum E).
xcited Linear Prediction: Vector addition excitation linear prediction) This is suitable for application to a speech coding system.

[0002]

2. Description of the Related Art The VSELP speech coding system is a kind of code-excited linear prediction (CELP) speech coding system.
The audio signal is parameter-encoded into vocal tract and vocal cord (excitation source) information. The VSELP speech coding system is different from the CELP coding system particularly in the parameter coding method of the excitation source. This VSELP voice encoding system is, for example, a TIA (Telecommunications Ind.
ustry Association) has been standardized as a voice coding scheme for digital cellular.

Such a VSELP audio encoding system is described in, for example, Documents A and B.

Reference A: “VECTOR SUM EXCITED LINEAR PR
EDICTION (VSELP) SPEECH CODINGAT 8KBPS ", Ira AG
erson and Mark A. Jasiuk, Proc. IEEE Inc. Conf. on A
coustics, Speech and Signal Processing, pp.461-46
4, April 1990 Document B: Japanese Patent Application Laid-Open No. 4-190399 FIG. 2 shows a functional configuration of a speech coder according to the conventional VSELP speech coding system, mainly in accordance with the contents described in Document A.

In FIG. 2, the information about the broken line output line drawn from each functional block shows what is transmitted to the speech decoding apparatus, and the information about the solid line output line drawn from each functional block is shown. The information is information processed when determining (searching) information to be transmitted.

This VSELP speech coding apparatus has an input speech analysis section (which can be further divided into a speech power calculation section, a speech power interpolation section and a vocal tract analysis section) 101, and a target signal generation section 102. , A long-term lag selection unit 103, a first code selection unit 104, a second code selection unit 105, a gain selection unit 106, three amplification circuits 107 to 109, and an addition circuit 110. The device receives the digitized audio signal V sampled at a predetermined frequency and performs encoding.

The input voice analysis unit 101 receives the input voice signal V
, LPC as vocal tract information by performing LPC analysis (a type of vocal tract analysis method) for each 160 samples (1 frame)
The coefficient k and the average voice power R are obtained, and the LPC coefficient k and the average voice power R are quantized and interpolated with respect to a sub-frame consisting of 40 samples.
LPC coefficient kj and speech average power R for each subframe
Outputs 0.

A method of interpolating the voice average power for such a subframe will be described. The voice average power of the first subframe is the voice average power of the previous frame, and the voice average power of the second subframe is the voice average of the previous frame. Power and the average audio power of the current frame are taken as the geometric mean,
The average voice power of the fourth subframe is the average voice power of the current frame.

The target signal creating unit 102 creates a residual signal of the input voice signal V by an analysis filter using the LPC coefficient kj for each subframe, and
By inputting the residual signal to the weighting filter using the PC coefficient kj, it is re-synthesized into the voice signal, and further, by removing the influence of the excitation signal (excitation code vector) ex synthesized in the previous subframe, Create and output the target signal p.

With the target signal p as a target, each functional unit, which will be described later, searches and determines excitation source information related to its own functional unit. Codes representing various kinds of information shown in FIG. 2 are assumed to be in a determined state, and in the following, description will be given by adding a code t to the end of the code (candidate) under search.

The long-term lag selection unit 103 updates and stores the excitation signals ex of a plurality of subframes that have been previously combined each time the processing for each subframe is completed (internal value long-term filter state). ). The long-term lag selection unit 103 extracts an excitation signal c0t of 40 samples at various lags Lt from the stored combined excitation signals (the excitation signal of 40 samples each is a long-term excitation code vector). ),
A long-term excitation code vector c0t corresponding to the lag extracted by the weighting filter using the LPC coefficient kj
Is calculated and the squared error of the difference between the signal f0t obtained by inputting and the target signal p is calculated, and the long-term excitation code vector c0 having the smallest squared error is obtained and output. Note that each long-term excitation code vector c0t corresponds to the lag Lt related to that long-term excitation code vector as an index.

The first code selecting section 104 stores a plurality of fixed data as a basis vector (base vector) v1. The first code selection unit 104 creates a plurality of first excitation code vectors u1t by combining a parameter (Gray code) θ that gives positive and negative polarities of the vector to the basis vector (see equation (1) of document A),
After each excitation code vector u1t is subjected to a weighting filter using the LPC coefficient kj, a signal f1t orthogonal to the long-term excitation code vector c0 selected by the long-term lag selecting unit 103 is created, and this signal f1t is generated. The square error of the difference between the target signal p and the target signal p is calculated, and the first excitation code vector c1 (u1) having the smallest square error is obtained and output. An index It is assigned to each first excitation code vector u1t.

The second code selection unit 105 stores a plurality of basis vectors v2 different from those of the first code selection unit 104. The second code selection unit 105 combines the basis vector with the parameter θ to create a plurality of second excitation code vectors u2t, applies each excitation code vector to a weighting filter using the LPC coefficient kj, and then performs long-term excitation. A signal f2t orthogonalized to the code vector c0 and further orthogonalized to the optimum first excitation code vector c1 selected by the first code selection unit 104 is created, and this signal f2t and the target signal p are generated. The squared error of the difference is calculated, and the second excitation code vector c2 (u2) having the smallest squared error is obtained and output. An index Jt is given to each second excitation code vector u2t.

The gain selection unit 106 includes amplifier circuits 107-
Together with 109 and the adder circuit 110, the optimum gains β and γ for each excitation code vector c0, c1 and c2 are obtained.
1 and γ2 are determined.

The gain selector 106 includes gains β, γ1, γ2 for the excitation code vectors c0, c1, c2.
A plurality of sets of group information is stored as a set of parameters GS, P0, P1 obtained by equivalently converting these gains β, γ1, γ2. The set of parameters GS, P0, and P1 is vector-quantized, and an index (gain index) G is given to each set. Note that the gains β and γ
The relationship between the set of 1 and γ2 and the set of parameters GS, P0, and P1 is described in the equations (15) to (23) of the above-mentioned document A.

The gain selection unit 106 has a parameter GS.
For each set of t, P0t and P1t, LPC coefficients kj,
Average voice power R0 and each excitation code vector c0, c
Based on 1 and c2, they are converted into gains βt, γ1t and γ2t and given to corresponding amplifier circuits 107, 108 and 109. By the amplifier circuits 107 to 109 and the adder circuit 110, the composite excitation signal ext, which is the product-sum operation result of the excitation code vectors c0, c1, c2 and the corresponding gains βt, γ1t, γ2t, is obtained as a plurality of candidates. The gain selection unit 106 sets the LP for each composite excitation signal ext.
A local reproduction signal is obtained by applying a synthesis filter obtained from the C coefficient kj, a square error of the difference between each local reproduction signal and the target signal p is calculated, and gains β, γ1, γ2 that minimize the square error are calculated. Determine the optimal set of.

Optimal excitation code vectors c0, c1, c
The combined excitation signal ex generated from 2 and the optimum gains β, γ1, and γ2 is used for combining the target signal for the next subframe and updating the long-term filter state.

The VSELP speech coder has the LPC coefficient kj, the average speech power R0, the index (long term lag) L of the long-term excitation code vector c0, the index I of the optimum first excitation code vector c1, and the optimum second. The index J of the excitation code vector c2 and the index G of the set of optimum gain parameters GS, P0 and P1 are transmitted to the VSELP speech decoding apparatus.

Next, although illustration of the VSELP speech decoding apparatus is omitted, its decoding operation will be briefly described.

In the VSELP speech decoding apparatus, the received lag L is applied as an index to the long-term filter state updated based on the past combined excitation signals of several subframes obtained as described later. And decodes the long-term excitation code vector c0. Also, the same stored basis vector as that of the encoding device (provided by the encoding device if the encoding device and the decoding device are installed in the same device) v
The optimum first excitation code vector c1 is decoded from 1 and the received index I for the first excitation code vector, and the stored basis vector is the same as the encoding device (the encoding device and the decoding device are the same). The optimum second excitation code vector c2 is decoded from v2, which is prepared by the encoding device if it is mounted on the device, and the index J for the received second excitation code vector. Furthermore, the set of optimum gain parameters GS, P0, P1 is extracted from the optimum gain index G, and the obtained long-term excitation code vector c0, first excitation code vector c1, second excitation code vector c2, and received LPC Optimal gains β, γ1, and γ2 are decoded from the coefficient kj and the average voice power R0.

The gains β, γ1, and γ2 corresponding to the obtained long-term excitation code vector c0, the first excitation code vector c1, and the second excitation code vector c2 are obtained.
And then add them to obtain a composite excitation signal ex, and use this composite excitation signal ex as the received LPC coefficient kj
By passing through the synthesis filter section configured using
The voice signal V is reproduced, and the noise component of the voice signal V is compressed and output through the post filter unit configured by using the received LPC coefficient kj.

By the way, the VSELP audio encoding system is
US TIA (Telecommunications Industry Associatio)
n) As standardized by the Committee as a voice encoding method for digital cellular, it is mainly used for voice signal compression communication such as digital mobile communication.

Therefore, a processing device (speech encoding device or speech decoding device) that executes the VSELP speech encoding system.
Are required to be as compact and low power consumption as possible. In order to meet such a demand, it is effective to fix the bit length (word length) of the data handled in the processing device as short as possible and to execute the arithmetic processing executed by the processing device in fixed-point representation. Is. Therefore, the conventional VSELP speech coding apparatus shown in FIG.
Even in the SELP speech decoding apparatus, the arithmetic processing is executed by the fixed point representation.

However, when the bit length of the data to be arithmetically processed is selected to be short and the arithmetic processing in the fixed point representation is adopted, if the audio average power of the input audio signal of the subframe at that time changes, the arithmetic operation is performed. The values of various internal variables used for processing also change, and some variables may cause overflow in the selected fixed-point representation. As a result, the calculation accuracy decreases and the quality of the final decoded audio signal significantly decreases. Something was happening.

The above-mentioned document B describes an invention which can solve the inconvenience which has occurred by calculating the data of the short bit length by the fixed point representation. That is, paying attention to the fact that there is a specific relationship between the size of the input audio signal (average power) and the size of the internal variable used for the arithmetic processing, and the arithmetic processing is performed according to the obtained average audio power R0. It is described that the decimal point positions of internal variables and the like used for the above are switched, and a predetermined operation is executed at the fixed decimal point position according to the switched decimal point position.

In FIG. 2, the target signal generator 10
2, long term lag selection section 103, first code selection section 104, second code selection section 105, and gain selection section 10
Reference numeral 6 and the like are functional units that execute various calculations, and are the portions to which the method described in Document B is applied. In FIG. 2, a target signal generation unit 102 and a long term lag selection unit 10
3, first code selecting section 104 and second code selecting section 10
The fact that the voice average power R0 is input in parentheses for the functional block of 5 indicates that the method of Document B is applicable. Also in the gain selection unit 114, the audio average power R0 can be used not only for converting the gain parameter into the gain but also for determining the fixed point position.

When the voice average power R0 is input to each of the functional units, the stage to which the voice average power R0 belongs is detected, and, for example, a table prepared in advance is referred to and the variable used for the calculation is determined. Get fixed-point representation. After such processing, a predetermined calculation (for example, processing of passing a predetermined signal through the filter) is performed.

FIG. 3 shows an example of the structure of a table for switching fixed point positions. For example, when the voice average power R0 is the value PW0 or more and less than PW1, the variable 1
It is defined that the fixed point position of the variable is set to a1, the fixed point position of the variable 2 is set to a2, and the fixed point position of the variable y is set to ay. Incidentally, the table in which the stage to which the average voice power R0 belongs and the fixed point representation of the internal variable are associated with each other is, for example, an output signal from the functional unit corresponding to the stage to which the average voice power R0 belongs (this is also an internal variable. The fixed-point position of the internal variable that can realize such a fixed-point position of the output signal is stored.

[0029]

However, the long term lag selecting section 103 has a character different from other functional sections among the plurality of functional sections whose fixed-point representations of the internal variables can be switched. That is, the long-term lag selection unit 103 stores the combined excitation signals ex of the past several subframes as a long-term filter state, and the fixed-point position of the combined excitation signal may be different depending on the stored subframes. Different from the others. When the fixed-point position of the long-term excitation code vector c0 from the long-term lag selection unit 103 is determined based on the audio average power R0 of the currently targeted subframe, the long-term filter is set to the determined fixed-point position. It is necessary to match the fixed-point positions of the composite excitation signal stored as a state, and to take them out. This means that the data is shifted, and the value may overflow depending on the shift direction.

That is, even if the fixed point position is switched on the basis of the voice average power R0, the long term lag selecting section 103 cannot obtain sufficient accuracy.

As described above, not only the gain selection but also the
The voice average power R0 for each subframe used for switching the fixed point position is, for example, as described above, the voice average power of the first subframe is the voice average power of the previous frame, and the voice average power of the second subframe is The voice average power of the previous frame and the voice average power of the current frame are determined as a geometric mean, and the voice average powers of the third and fourth sub-frames are determined to be the voice average power of the current frame, respectively. The average information of the two audio average powers is calculated by geometric mean for the second subframe because the audio average power is the sum of squares of the sample values, and the geometric mean is more This is because it is closer to the sum of squares of the sample sequence of the intermediate value than the arithmetic mean.

According to the above-mentioned interpolation method, the voice average powers of the first, third and fourth subframes can be easily obtained, but calculation is required for the second subframe. Moreover, since it is a geometric mean calculation, multiplication and square processing are required. Generally, the process of obtaining a square root requires a complicated calculation mechanism or requires many processing steps. Therefore, this is contrary to the possible miniaturization and low power consumption required for a device (speech encoding device or speech decoding device) to which the VSELP speech encoding system is applied.

The problem has been described above with respect to the device conforming to the VSELP speech coding system. However, even in the device conforming to the general CELP speech coding system, the past synthesized excitation signal is stored and a certain kind of excitation code vector is output. Some have an adaptive codebook section, and some interpolate the speech average power for each subframe. That is,
The above problem also occurs in a device that complies with a general CELP speech coding system.

The present invention has been made in consideration of the above points, and also in the adaptive excitation code vector selection unit that stores the past combined excitation signal, the accuracy is improved by switching the fixed point position. It is an object of the present invention to provide a speech coding apparatus and a speech decoding apparatus which can be performed.

Further, the present invention provides a speech coding apparatus capable of performing interpolating calculation of the speech average power of each subframe from the speech average power of each frame with a simple configuration or a simple process. It was something I tried to do.

[0036]

In order to solve such a problem, according to the present invention of claim 1, a voice signal is parameter-encoded by digital processing by a finite word length C.
A speech coder according to an ELP speech coding method, comprising:
A fixed internal variable that saves a composite excitation signal of several past frames (a concept including subframes in this claim) and outputs a part of the signal sequence as an adaptive excitation code vector component The following configuration is provided in a speech coder having an adaptive excitation code vector selector capable of switching the decimal point position.

That is, for each of the composite excitation signals of the past several frames stored in the adaptive excitation code vector selection unit, a decimal point position buffer unit that stores fixed point position information and a decimal point position buffer unit A fixed-point position selection unit that determines the fixed-point position of the internal variable in the adaptive excitation code vector selection unit is provided based on the plurality of fixed-point position information stored and the voice average power for the current frame.

Further, according to the present invention of claim 2, a voice signal is converted into
A speech coding apparatus according to a CELP speech coding method for parameter coding by digital processing with a finite word length, comprising a speech power calculation unit for obtaining a speech average power of each frame of a speech signal, and two consecutive frames. A speech coding apparatus having a speech power interpolating unit that obtains speech average power information for each subframe obtained by equally dividing one frame into several pieces from the speech average power is as follows.

That is, the voice power interpolating unit is provided with at least a means for quantizing the voice average power, and when the voice average power information for each subframe is obtained, the voice average power for each frame is treated with a quantized index. For the subframes for which the geometric mean information of the voice average powers of the individual frames is obtained, the arithmetic mean of the indexes is obtained, and the indexes obtained for each subframe are output, or the obtained indexes are dequantized. It is decided to output the audio average power returned via the.

Furthermore, the present invention of claim 3 is a speech coder according to the CELP speech coding system for parameter-coding a speech signal by digital processing according to a finite word length, wherein speech averaging for each frame of a speech signal is performed. An audio power calculation unit that obtains power, an audio power interpolation unit that obtains audio average power information for each subframe obtained by equally dividing one frame into several parts from the audio average powers of two consecutive frames, and a past number It has an adaptive excitation code vector selector that can switch the fixed-point position of the internal variable, which stores the composite excitation signal of each subframe and outputs a part of the signal sequence as an adaptive excitation code vector component. The speech coder is as follows.

That is, the voice power interpolating unit is provided with at least a means for quantizing the voice average power, and when the voice average power information for each subframe is obtained, the voice average power for each frame is treated with a quantized index. For the subframes for which the geometric mean information of the voice average powers of the individual frames is obtained, the arithmetic mean of the indexes is obtained, and the indexes obtained for each subframe are output, or the obtained indexes are dequantized. Outputs the voice average power returned via the, and the decimal point position buffer unit that stores fixed point position information for each of the past several subframe composite excitation signals stored in the adaptive excitation code vector selection unit. And a plurality of fixed point position information stored in this decimal point position buffer section and the current subframe And a voice average power of beam, provided with decimal point selection unit which determines a fixed-point position of the internal variables of the adaptive excitation code vector selection unit.

The present invention of claim 4 is the C adopted by the speech encoding apparatus according to any one of the inventions of claims 1 to 3.
The ELP speech coding method is a forward VSELP speech coding method, which is one of the ELP speech coding methods.

The present invention of claim 5 is a speech decoding apparatus according to the CELP speech coding system corresponding to the speech coding apparatus of the present invention of claim 1, wherein several past frames (in this claim) Is a concept that includes subframes) and outputs a part of the signal sequence as an adaptive excitation code vector component. Adaptive excitation code vector selection that can switch the fixed-point position of the internal variable. The following configuration is provided in a speech decoding apparatus having a section.

That is, for each of the composite excitation signals of the past several frames stored in the adaptive excitation code vector selection unit, a decimal point position buffer unit for storing fixed point position information and a decimal point position buffer unit are stored. A fixed-point position selection unit that determines the fixed-point position of the internal variable in the adaptive excitation code vector selection unit is provided based on the plurality of fixed-point position information stored and the voice average power for the current frame.

The present invention of claim 6 is the CEL adopted by the speech decoding apparatus according to any one of the invention of claim 5.
The P voice encoding method is one of them, a forward VS.
It is characterized in that it is an ELP voice encoding system.

[0046]

The present invention according to claims 1 and 5 is a speech coding apparatus and speech decoding apparatus according to the CELP coding system, which parameter-codes a speech signal by digital processing with a finite word length and transmits the speech signal. It has an adaptive excitation code vector selector that can switch the fixed-point position of the internal variable, which stores the composite excitation signal of several frames and outputs a part of the signal sequence as an adaptive excitation code vector component. Regarding things.

There has already been proposed a method of switching the fixed point position of the internal variable used when each functional unit processes by the voice average power. In the adaptive excitation code vector selection unit that saves and uses the composite excitation signals of the past several frames, the fixed-point position is different for each of the composite excitation signals of the past several frames. When the fixed point position is determined from the audio average power of, the fixed point position may not be valid from the viewpoint of the combined excitation signal of the past frames (for example, even if a small number can be expressed accurately, a large number may overflow. May be inaccurate).

Therefore, in the present invention of claim 1 and claim 5, a decimal point position buffer section and a decimal point position selecting section are provided, and the decimal point position selecting section is provided with a plurality of fixed decimal point positions stored in the decimal point position buffer section. The fixed point position of the internal variable in the adaptive excitation code vector selection unit is decided from the information and the speech average power for the current frame. Of course, in the speech coding apparatus and the speech decoding apparatus, the adaptive excitation code vector selection section must output the adaptive excitation code vector in the same state, and the speech coding apparatus must select the decimal point position buffer section and the decimal point position selection section. When a unit is provided, these functional units are also provided in the corresponding voice decoding device.

According to the present invention of claim 2, a voice signal is parameter-encoded by digital processing with a finite word length.
A speech coder according to an ELP speech coding method, comprising:
A voice power calculation unit that obtains a voice average power of each frame of a voice signal, and a voice that obtains voice average power information of each subframe obtained by equally dividing one frame into several voices from the voice average powers of two consecutive frames. It is premised on a speech coding apparatus having a power interpolation unit.

Depending on the sub-frame, the average information of the average voice powers of two adjacent frames is obtained as the interpolation information. Considering that the voice average power is obtained by the sum of squares of sample values, in the past, interpolation was performed using a geometric mean value, but this requires a complicated interpolation configuration or interpolation process.

Therefore, in the present invention of claim 2, the voice power interpolating section is provided with at least a voice average power quantizing means, and when the voice average power information for each subframe is obtained, the voice for each frame is determined. It was decided to handle the average power with a quantized index. Further, depending on the subframe, the geometric mean information of the sound average powers of the two frames may be obtained, but in this case, the arithmetic mean of the indexes is used. Depending on the way of using the audio average power information for each subframe, the index of the audio average power obtained for each subframe may be output, and the obtained index may be output via an inverse quantizer. The returned audio average power may be output.

The speech coder according to the third aspect of the present invention comprises both the features of the first and second aspects of the present invention.

The present invention of claim 4 is the C adopted by the speech encoding apparatus according to any one of the inventions of claims 1 to 3.
The ELP speech coding system is limited to the forward type VSELP speech coding system, which is one of the ELP speech coding systems. The invention of claim 6 is the speech decoding apparatus according to the invention of claim 5. The CELP speech coding method adopted by the present invention is limited to the forward VSELP speech coding method which is one of them. The forward VSELP voice encoding system is adopted in compression communication such as mobile communication, and the present invention of claim 6 and the present invention of claim 6 function effectively. .

[0054]

【Example】

(A) VSELP Speech Encoding Device An embodiment of the speech encoding device according to the present invention will be described in detail below with reference to the drawings. This embodiment is an example of a VSELP speech coder according to the VSELP speech coding method, and FIG. 1 shows the functional block configuration thereof.

Note that, also in FIG. 1, the information regarding the broken line output lines drawn from the respective functional blocks indicates what is transmitted to the decoding device, and the information regarding the solid line output lines drawn from the respective functional blocks is shown. The information is information processed when determining (searching) information to be transmitted.

Further, the codes representing the various information shown in FIG. 1 are assumed to be in a determined state, and the information (candidates) under search will be described by adding the code t to the end of the code.

The VSELP speech coding apparatus of this embodiment, as shown in FIG. 1, has a speech power calculation section 201, a speech power interpolation section 202, an LPC analysis section 203, a target signal generation section 204, and a long term lag selection section 205. , Decimal point position buffer unit 206, decimal point position selection unit 207,
The first code selecting unit 208, the second code selecting unit 209, the gain selecting unit 210, the three amplifying circuits 211 to 213, and the adding circuit 214 are included.

A digital input voice signal V is given to the voice power calculation unit 201, and the voice power calculation unit 201 calculates the voice average power of the input voice signal V in frames (for example, 16).
It is calculated every 0 samples (for example, the sum of squares of sample values), and the voice average power Rf for each frame is given to the voice power interpolation unit 202 and the LPC analysis unit 203.

The voice power interpolator 202 interpolates the voice average power R0 of each subframe (for example, 40 samples) from the voice average power Rf of the current frame and the previous frame to obtain various functions as described later. Output.

Here, the voice power interpolation unit 202 has a quantization table 202a of the average voice power expressed in logarithmic order, and obtains an index indicating which quantization stage the average voice power belongs to. Is made possible. Also, the audio power interpolation unit 202
Also has an inverse quantization table 202b for obtaining the average voice power from the average voice power index.

In this embodiment, the voice power interpolating unit 202 converts the voice average power Rf of each frame into an index using the quantization table 202a, and performs interpolation processing for subframes at the index stage.
The index of the average voice power of each sub-frame is converted into the average voice power R0 using the inverse quantization table 202b.

Interpolation of the index for each subframe is
For example, do as follows. (1) The index of the average audio power of the previous frame is used as the index of the average audio power of the first subframe. (2) As the index of the average audio power of the second sub-frame, the arithmetic average of the index of average audio power of the previous frame and the index of average audio power of the current frame is used. (3) The index of the average voice power of the current frame is used as the index of the average voice power of the third subframe. (4) The index of the average voice power of the current frame is used as the index of the average voice power of the fourth subframe.

The dequantization table 202b is configured more finely than the quantization table 202a so that an index having a fraction obtained by arithmetic averaging can be converted into a speech average power.

In the interpolation method of this embodiment, the conversion of the voice average power into the index and the conversion of the index into the voice average power are performed by using the tables 202a and 202b, which is a simple process.
The process of obtaining the average information of the individual voice average powers is also a simple process because the arithmetic mean is used instead of the geometric mean.
The geometric mean value is obtained by multiplication and square processing, while the arithmetic mean value is obtained by addition and 1/2 processing. The addition in the arithmetic mean calculation is also a simple process, and is a simple process of ½ times the data of a plurality of bits and shifting by 1 bit.

In the above description, the index is inversely quantized and returned to the voice average power and then given to the various function units. However, the index may be given to the various function units as it is, or the index may be given as it is. You may make it transmit to a decoding apparatus. The audio average power R0 for each subframe is used for the fixed point position of the internal variable in various functional units, as will be described later, and the absolute value is not a problem, and there is a step in relation to the fixed point position. It is sufficient if it is clear, and it is sufficient to output the index. Rather, the form of giving the index can simplify the configuration of various functional units. However, the gain selection unit 2
In No. 10, since the absolute value of the average voice power for each subframe is necessary for the gain selection operation, it is necessary to have an inverse quantization table in the gain selection unit 210 when it is given as an index. Become. However, in the following description, it is assumed that the audio average power R0 itself for each subframe is given to various functional units.

The LPC analysis section 203 is provided as a vocal tract analysis means. The LPC analysis unit 203 switches the fixed point position of the variable in the LPC analysis calculation performed internally based on the given voice average power Rf of the current frame (see FIG. 3). After that, the LPC analysis is performed on the input audio signal V for one frame to obtain the LPC coefficient kj.
And output.

In the prior art, the LPC coefficient is also calculated and output for each subframe, but in this embodiment, the LPC coefficient kj for each frame is output.
Is used in all subframes. However, this difference is irrelevant to the features of the present invention, and the LPC coefficient may be obtained and output for each subframe as in the conventional case.

Target signal generating section 204 performs substantially the same processing as conventional target signal generating section 102 shown in FIG. 2 to form and output target signal p for each subframe. The difference from the conventional target signal generation unit 102 shown in FIG. 2 is that the calculation calculation of the target signal p is performed in accordance with the average audio power R0 of the input subframe before the calculation calculation process of the target signal p is started. The point is to set the fixed point position of the variable used for.

The long-term lag selecting section 205 performs almost the same processing as the conventional long-term lag selecting section 103 shown in FIG. 2 to obtain and output the optimum long-term excitation code vector c0 for each subframe. Of course, at this time, the optimum long-term lug L is also determined.

However, the long term lag selecting section 205 of this embodiment has a function of switching the fixed point position of the internal variable used for the process of determining the optimum long term excitation code vector c0 and the optimum long term lag L. Is different from the conventional one, and the method of switching the fixed point position of the internal variable is also different from the method described in Document B based only on the average sound power R0.

In order to switch the fixed point position, the long term lag selecting section 205 of this embodiment is provided with a decimal point position buffer section 206 and a decimal point position selecting section 207.

FIG. 4 shows the stored contents of the long term filter state buffer section and the decimal point position buffer section 206 in the long term lag selection section 203.

The final synthetic excitation signal ex for each subframe is given to the long-term lag selecting section 203, and the final synthetic excitation signals of the last few past (four in FIG. 4) subframes are obtained. Signals ex (-1), ex (-2), ex
(-3) and ex (-4) are as shown in FIG.
It is stored in the long-term filter state buffer section. When the composite excitation signal ex is stored, information on the fixed point position fp related to the composite excitation signal is given to the decimal point position buffer unit 206, and as shown in FIG. Position fp (-
Information of 1), fp (-2), fp (-3), and fp (-4) is stored in the decimal point position buffer unit 206.

Here, when searching for the optimum long-term excitation code vector c0 for the current subframe,
It is necessary to extract data for one subframe period while changing the long term lag Lt from the portion stored in the long term filter state buffer unit. In such a case, it is almost the case that the data of the sub-frame period in which the previous two sub-frames are combined into the excitation signal is extracted. The final combined excitation signals ex (−1), ex (−2), ex (−) stored in the long-term filter state buffer unit for the last few past subframes.
3), fixed-point positions fp (-1), f of ex (-4)
p (−2), fp (−3), and fp (−4) are often different, and the fixed point position of the internal variable in the long term lag selection unit 205 is based on only the voice average power R0. It is not reasonable to decide.

The decimal point position selection unit 207 determines the final combined excitation signal ex (-
1), ex (-2), ex (-3), ex (-4) fixed point positions fp (-1), fp (-2), fp (-
3), fp (-4) is considered, and the fixed point position of the internal variable in the long term lag selection unit 205 is determined as follows.

FIG. 5 shows a table structure built in the decimal point position selection unit 207. The decimal point position selection unit 207 accesses the table shown in FIG. 5A by using the voice average power R0 of the current subframe, and estimates the fixed point position fp (0) obtained for the composite excitation signal of the current subframe. To get Next, the decimal point position selection unit 207 stores the fixed decimal point positions fp (−1), fp (−2), fp stored in the decimal point position buffer unit 206.
(-3), fp (-4) and the fixed point position FP representing the largest number among the fixed point positions fp (0) obtained for the current subframe are obtained. After that, the decimal point position selection unit 207 uses the fixed point position FP as shown in FIG.
The table shown in (b) is accessed to determine the fixed point position for each of the internal variables in the long term lag selection unit 205.

By determining the fixed point position as described above, the long-term excitation code vector c0t
It is possible to prevent the accuracy from decreasing when searching for (long term lag Lt).

In the first code selection section 208, the second code selection section 209, and the gain selection section 210 as well, similar to the target signal generation section 204, the fixed point position determination process of the internal variable based on the voice average power R0 is performed. It is executed (see FIG. 3). In the first code selection unit 208, the second code selection unit 209, and the gain selection unit 210, the process itself performed after determining the fixed point position of such an internal variable is the conventional first code selection unit shown in FIG. 10
4, the second code selecting section 105 and the gain selecting section 106 are almost the same, and therefore their explanations are omitted.

According to the VSELP speech coding apparatus of the above embodiment, the fixed decimal point positions of the internal variables used for the calculation in the long term lag selecting section 205 are fixed for the several past composite excitation signals which are stored. Decimal point position,
Also, since the determination is made based on the average voice power of the current subframe, the precision is not affected by the difference in the fixed point position of the past synthesized excitation signal, the encoding precision of the voice signal is improved, and it is decoded. It is possible to improve the quality of the voice signal compared to the conventional one.

The long-term excitation code vector c
For each candidate of 0 (long term lag L), the fixed point position of the internal variable is determined from the fixed point position of the combined excitation signal of the past two subframes related to the candidate and the average voice power R0 of the current subframe. It is also possible to decide.
However, in this case, it is not practical to review the internal variable for each candidate of the long-term excitation code vector c0 (long-term lag L), and it is not practical as in the above-described embodiment. It is preferable to uniformly determine the fixed-point positions of internal variables for the candidates.

Further, according to the VSELP speech coding apparatus of the above-mentioned embodiment, when the speech average power of a certain subframe is obtained as the average information value from the speech average powers of two frames, the quantization of the speech average power, Since the arithmetic mean of the index by quantization is used, the structure for obtaining the average voice power of the subframe can be simplified, or the number of processing steps can be reduced (thus, the power consumption can be reduced. it can).

(B) VSELP Speech Decoding Device Next, a VSELP speech decoding device of an embodiment corresponding to the VSELP speech decoding device according to the above embodiment will be described in detail with reference to the drawings.

FIG. 6 shows a functional block configuration of the VSELP speech decoding apparatus of this embodiment. In FIG. 6, the information about the broken lines of the input lines to each functional block indicates the information transmitted to the decoding device, and the information about the solid lines of the input lines to each functional block is shown. Is information formed in the decoding device.

As shown in FIG. 6, the VSELP speech decoding apparatus of this embodiment has a long term lag selecting section 30.
1, decimal point position buffer unit 302, decimal point position selection unit 3
03, first code selection unit 304, second code selection unit 30
5, gain selection unit 306, three amplification circuits 307 to 30
9, an adder circuit 310, a synthesis filter unit 311, and a post filter unit 312.

In FIG. 6, among the information transmitted from the VSELP speech coding apparatus, the long term lag L is given to the long term lag selecting section 301, and the index I of the first excitation code vector is the first code selecting section. 304
Given to the index J of the second excitation code vector
Is given to the second code selecting unit 305, the gain index G is given to the gain selecting unit 306, and the voice average power R0 is the decimal point position selecting unit 303 and the first code selecting unit 3.
04, a second code selection unit 305, a gain selection unit 306,
The LPC coefficient kj is provided to the synthesis filter unit 311 and the post filter unit 312, and is provided to the gain selection unit 306, the synthesis filter unit 311 and the post filter unit 312.

The long-term lag selecting section 301 is provided with the determined combined excitation signal ex for each subframe,
It is designed to be stored in the internal long-term filter state buffer section, and the composite excitation signal e
Information on the fixed point position fp related to x is given to the decimal point position buffer unit 302 to be stored. That is, the stored contents of the long-term filter state buffer unit and the decimal point position buffer unit 302 before starting the processing of the current subframe are the same as the corresponding buffer units in the encoding device.

The decimal point position selection unit 303 obtains an estimated value of the fixed decimal point position fp (0) required for the composite excitation signal of the current subframe based on the voice average power R0 of the current subframe, and then the decimal point position buffer. The composite excitation signals ex (−1), ex (−2), ex (−3), ex (−) of the previous several past subframes stored in the unit 302.
4) fixed point positions fp (-1), fp (-2), f
The fixed point position FP that represents the largest number among p (−3) and fp (−4) and the fixed point position fp (0) obtained for the current subframe is calculated, and the fixed point position FP is obtained.
The fixed point position is determined for the internal variable in the long term lag selection unit 301.

Long term lag selector 30 of the decoding device
In 1, it is not necessary to search for the optimum long-term excitation code vector c0, so the variable whose fixed point is switched is the output long-term excitation code vector c0.
Itself.

When the fixed-point position for the current subframe is determined, the long-term lag selecting section 301 extracts the long-term excitation code vector c0 from the long-term filter state buffer section based on the given long-term lag L. And outputs it to the gain selection unit 306 and the amplification circuit 307.

First code selecting section 304, second code selecting section 305, gain selecting section 306, synthesis filter section 31.
1. The post-filter unit 312 determines the fixed point position of the internal variable to be used for processing, based on the average voice power R0 given from the encoding device side, before executing the original processing assigned to the self block. (See FIG. 3).
The first code selection unit 30 in the speech decoding device
4, second code selection unit 305 and gain selection unit 306
Since there is no need to search for the optimum value, the number of types of internal variables whose fixed-point positions are determined is smaller than that of the corresponding functional block in the encoding device.

The first code selection unit 304 stores the same basis vector v1 as that of the first code selection unit 208 in the encoding device, and when the fixed point position is determined, the first vector received with this basis vector is determined. The optimum first excitation code vector c1 is decoded from the index I of the excitation code vector and output to the amplification circuit 308 and the gain selection unit 306.

The second code selecting unit 305 stores the same basis vector v2 as the second code selecting unit 209 in the encoding device, and when the fixed point position is determined, this second basis vector v2 is received. The optimum second excitation code vector c0 is decoded from the index J for the excitation code vector and output to the amplifier circuit 309 and the gain selection unit 306.

The gain selection unit 306 has the same gain parameter GS as the gain selection unit 210 in the encoding device,
The set information of P0 and P1 is stored, and when the fixed point position is determined, the optimum set of gain parameters GS, P0 and P1 is extracted from the received gain index G, and the obtained long term excitation code vector c0 is obtained. , A first excitation code vector c1, a second excitation code vector c2,
Optimal gains β, γ1 and γ2 are decoded from the received LPC coefficient kj and average voice power R0 and output to the corresponding amplifier circuits 307, 308 and 309, respectively.

The long-term excitation code vector c0, the first excitation code vector c1, obtained as described above,
The second excitation code vector c2 is supplied to each amplifier circuit 307, 3
08, 309, corresponding gains β, γ1, γ2
Gain control based on
Are added to obtain a composite excitation signal ex for the current subframe. This combined excitation signal ex is given to the combined filter section 311 and the long term lag selecting section 301.

The synthesis filter unit 311 sets the LPC coefficient kj received in the determined fixed-point position representation to define the filter characteristic, and filters the input synthetic excitation signal ex with the filter characteristic to produce a voice signal. Reproduce. The reproduced voice signal thus obtained may have a large noise component.

The post-filter unit 312 sets the LPC coefficient kj received in the determined fixed-point position representation to define the filter characteristic, and filters the reproduced voice signal by the synthesis filter unit 311 with the filter characteristic to generate a noise component. A final reproduced audio signal V obtained by compressing is obtained and output.

According to the VSELP speech decoding apparatus of the above embodiment, the fixed decimal point positions of the internal variables used for the calculation in the long term lag selecting section 301 are fixed for the several past composite excitation signals which are stored. Decimal point position,
Also, since the determination is made based on the average voice power of the current subframe, the precision is not affected by the difference in the fixed point position of the past synthesized excitation signal, the encoding precision of the voice signal is improved, and it is decoded. It is possible to improve the quality of the voice signal compared to the conventional one. Of course, the same effect can be obtained by adopting the same fixed-point position determination method as that of the encoding device.

(C) Other Embodiments In the description of the above embodiments, other embodiments have been mentioned, but the following other embodiments can be mentioned in addition to the above description.

The above-mentioned embodiment has two main features. The first feature relates to a method of interpolating a voice average power,
That is, the arithmetic mean of the indexes is used to obtain the average information from the two voice average powers. The second feature relates to a fixed-point representation method when the past excitation signal of the sub-frame is used as the excitation code vector component of the current sub-frame. That is, the fixed point position is determined. However, the device may be configured to have only one of the features.

In the above embodiment, the present invention is applied to the device conforming to the forward type VSELP speech encoding system, but it may be applied to the device conforming to the backward type VSELP speech encoding system. CELP
It may be applied to a device that complies with the audio encoding system.

For example, even a device according to a general CELP speech coding system has an adaptive codebook unit (corresponding to the long term lag selection unit of the embodiment) for storing and using the past synthesized excitation signal. The second feature can be applied to such a device. In this case, the device may not decompose the frame into subframes.

Further, for example, even if the device does not use the average information of the two voice average powers for switching the fixed point representation, the device that needs to obtain the average information of the two voice average powers. If so, the first characteristic configuration can be applied.

[0103]

As described above, according to the speech coding apparatus and the speech decoding apparatus of the present invention, for each of the composite excitation signals of the past several frames stored in the adaptive excitation code vector selection unit. , The fixed point position buffer section for storing fixed point position information, a plurality of fixed point position information stored in this decimal point position buffer section, and the voice average power for the current frame, Since the decimal point position selection unit that determines the fixed point position of the variable is provided, it is possible to improve accuracy by switching the fixed point position even in the adaptive excitation code vector selection unit that stores the past composite excitation signal. become able to.

According to another speech coding apparatus of the present invention, the speech power interpolating section includes at least a quantizing means for the speech average power, and when the speech average power information for each subframe is obtained, The audio average power of is treated as a quantized index, and the geometric mean information of the audio average powers of two frames is calculated by the arithmetic mean of the index numbers, and the index calculated for each subframe is output or calculated. Since the voice average power obtained by returning the index through the inverse quantization means is output, the voice average power information for each subframe can be obtained by a simple configuration or a simple process.

[Brief description of drawings]

FIG. 1 is a functional block diagram of a VSELP audio encoding device according to an embodiment.

FIG. 2 is a functional block diagram of a conventional VSELP speech coder.

FIG. 3 is an explanatory diagram of a general fixed point representation switching method.

FIG. 4 is an explanatory diagram (Part 1) of a fixed-point representation switching method for the long-term lag selecting unit according to the embodiment.

FIG. 5 is an explanatory diagram (part 2) of the fixed point representation switching method for the long term lag selection unit of the embodiment.

FIG. 6 is a functional block diagram of a VSELP audio decoding device according to the embodiment.

[Explanation of symbols]

 201 ... Voice power calculation unit, 202 ... Voice power interpolation unit, 202a ... Quantization table of voice average power, 203 ... LPC analysis unit, 204 ... Target signal creation unit, 205, 301 ... Long term lag selection unit, 206, 302 ... Decimal point position buffer section, 207, 303 ... Decimal point position selection section, 208, 304 ... First code selection section, 209, 305 ... Second code selection section, 210, 306 ... Gain selection section, 211-213, 307-309 ... amplification circuit 214, 310 ... addition circuit 214, 311 ... synthesis filter section, 312 ... post filter section.

Front Page Continuation (72) Inventor Katsutoshi Ito 1-7-12 Toranomon, Minato-ku, Tokyo Oki Electric Industry Co., Ltd.

Claims (6)

[Claims] 1. A speech coder according to a CELP speech coding method for parameter-coding a speech signal by digital processing according to a finite word length, wherein synthetic excitation signals of several past frames are stored, In a speech coder having an adaptive excitation code vector selection unit that can switch the fixed-point position of an internal variable that outputs a part of the signal sequence as an adaptive excitation code vector component, save it in the adaptive excitation code vector selection unit For each of the composite excitation signals of the past several frames, the decimal point position buffer section that stores the fixed point position information, the multiple fixed point position information stored in this decimal point position buffer section, and the current frame , And the fixed-point position of the internal variable in the adaptive excitation code vector selector Determining speech coding apparatus is characterized in that a decimal point selection unit which. 2. A speech coding apparatus according to the CELP speech coding system for parameter-coding a speech signal by digital processing according to a finite word length, comprising: a speech power calculation unit for obtaining a speech average power of each frame of a speech signal. And a voice power interpolating unit that obtains voice average power information for each subframe obtained by equally dividing one frame into several frames, from the voice average power of two successive frames. The unit includes at least a speech average power quantizing means, and when obtaining the speech average power information for each sub-frame, the speech average power for each frame is treated by a quantized index, and the speech average power of two frames is calculated. When calculating the geometric mean information, the arithmetic mean of the indexes is calculated, and the index calculated for each subframe is calculated. Force and, or, the speech coding apparatus and outputs a voice average power back through the inverse quantization means indexes obtained. 3. A speech coding apparatus according to the CELP speech coding system for parameter-coding a speech signal by digital processing with a finite word length, comprising: a speech power calculation unit for obtaining a speech average power for each frame of a speech signal. , A voice power interpolator for obtaining voice average power information for each sub-frame obtained by equally dividing one frame into several voices from the voice average powers of two consecutive frames, and a composite excitation signal of several past sub-frames In the speech encoding apparatus having an adaptive excitation code vector selection unit capable of switching the fixed point position of the internal variable, which outputs a part of the signal sequence as an adaptive excitation code vector component. When the power interpolator includes at least a speech average power quantizing unit and obtains the speech average power information for each subframe, The speech average power of each frame is treated as a quantized index, and when the geometric mean information of the speech average powers of two frames is calculated, the arithmetic average of the indexes is calculated and the index calculated for each subframe is output. , Or output the speech average power obtained by returning the obtained index through the dequantization means, and for each of the past composite excitation signals of several subframes stored in the adaptive excitation code vector selection unit. , The fixed-point position buffer unit that stores fixed-point position information, the plurality of fixed-point position information that is stored in the fixed-point position buffer unit, and the average voice power for the current subframe, in the adaptive excitation code vector selection unit. A sound characterized by having a decimal point position selection unit that determines the fixed point position of an internal variable. Encoding device. 4. The speech coding apparatus according to claim 1, wherein the CELP speech coding method is a forward VSELP speech coding method which is one of the CELP speech coding methods. 5. C corresponding to the speech coding apparatus according to claim 1.
A voice decoding device according to an ELP voice encoding method, comprising:
Save the composite excitation signals of the past few frames,
In a speech decoding device having an adaptive excitation code vector selection unit that can switch the fixed-point position of an internal variable that outputs a part of the signal sequence as an adaptive excitation code vector component, save it in the adaptive excitation code vector selection unit For each of the composite excitation signals of the past several frames, the decimal point position buffer section that stores the fixed point position information, the multiple fixed point position information stored in this decimal point position buffer section, and the current frame And a decimal point position selection unit that determines a fixed decimal point position of an internal variable in the adaptive excitation code vector selection unit. 6. The speech decoding apparatus according to claim 5, wherein the CELP speech coding method is a forward VSELP speech coding method which is one of the CELP speech coding methods.