JP2898377B2 - Code-excited linear prediction encoder and decoder - Google Patents

Description

Description: BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a code-excited linear prediction encoder for encoding original speech and a decoder corresponding thereto.

[Prior Art] Since voice information has a large amount of information, an original voice is encoded and compressed when transmitting or performing recognition processing. As one of such encoding methods,
A code-excited linear predictive coding scheme has been proposed. For example, it is described in the following literature.

Literature [MRSchroeder, BSAtal
Co-author, "Code-Excited Linear Predictio"
n (CELP): High Quality Speech at Very Low Bit Rate
(Code Excited Linear Prediction: High Quality Speech at Low Bit Rate) ": Proc. ICASSP, pp937-940, March 1985. FIG. 2 shows a conventional conceptual diagram of such a code excited linear predictive encoder. .

In FIG. 2, the code excitation linear predictive encoder 10
The original speech vector S is input to the input terminal 11. The original audio vector S is a vector composed of a plurality of adjacent samples when the original audio is sampled (the same sample may be an element of a plurality of vectors, and one sample is any one of the samples). It may be an element of only one vector). The original speech vector S is provided to the short-time analyzer 12 and the subtractor 18.

The short-time analysis unit 12 calculates a short-term prediction coefficient αj for the original speech vector S. One example is a short LPC
The short-term prediction coefficient αj is calculated by the analysis. The short-term analysis unit 12 calculates the short-term prediction coefficient αj thus obtained.
Is transmitted to the short-time synthesis filter 13 and the quantizer 14. In addition, the short-time analysis unit 12 sends the vector S1 from which the formant component (short-time component) of the voice has been removed to the long-time analysis unit 15.

The long-time analysis unit 15 performs a long-time analysis (LPC analysis as an example of the analysis) on the vector S1, calculates a pitch prediction coefficient β and a lag (pitch cycle) L, To the generator 14.

The quantizer 14 quantizes the short-term prediction coefficient αj, the pitch prediction coefficient β, the lag L, and the index I of the optimal excitation code vector described later, and outputs the result from the output terminal 112 as a total code C.

The index I given to the quantizer 14 is obtained by the processing of the excitation code table 17, the long-time synthesis filter 16, the short-time synthesis filter 13, the subtractor 18, the perceptual weighting filter 19, the square calculation unit 110, and the average operation unit 111. Is determined as follows.

In the excitation code table 17, each index i (i =
Excitation code vectors ei are stored in association with 1 to n), and each excitation code vector ei is applied to the long-time synthesis filter 16 in time sequence or collectively.

The long-time synthesis filter 16 performs inverse processing (synthesis processing) of the long-time analysis unit 15 on the excitation code vector ei using the pitch prediction coefficient β and the lag L, and outputs the processed output vector S
The Ti is sent to the synthesis filter 13 for a short time. The short-time synthesis filter 13 generates a short-time prediction coefficient αj
And the inverse processing (synthesis processing) of the short-time analysis unit 12 is performed, and the processing output vector SWi is sent to the subtractor 18.

Each of the n vectors SWi is represented by each excitation code vector
ei means a synthesized sound (prediction vector) synthesized by ei, and corresponds to the original speech vector S, respectively.

The subtracter 18 calculates each predicted vector SWi and the original speech vector S
And the resulting error vector e
ri is sent to the perceptual weighting filter 19.

The perceptual weighting filter 19 is provided in consideration of human perceptual characteristics, and weights the error vector eri in accordance with the perceptual characteristics to obtain a vector ε.
i is sent to the square calculator 110.

The square calculation unit 110 calculates the square of each component of the perceptually weighted error vector εi, and the average operation unit 111 calculates the average gi of the sum of the squares of the obtained components of the vector εi. Further, the average operation unit 111 detects the minimum average value gI (the one with the smallest prediction error) from the n average values g1 to gn, and uses the index as an index I of the optimal excitation code vector as an excitation code table. 17 to quantizer 14
Output.

The above-described units 11 to 112 constitute the code excitation linear encoder 10.The short-time analysis unit 12, the short-time synthesis filter 13, the long-time analysis unit 15, and the long-time synthesis filter 16 , Subtractor 18, perceptual weighting filter 1
It can be said that the ninth-square calculation unit 110 and the average operation unit 111 execute arithmetic processing together if their essential processing functions are excluded.

A decoder (not shown) extracts an excitation code vector eI corresponding to the index I from the index I, and inverts the excitation code vector eI of the long time analysis unit 15 using the pitch prediction coefficient β and the lag L ( (Synthesis process) and the inverse process (synthesis process) of the short-time analysis unit 12 using the short-term prediction coefficient αj are sequentially performed to obtain a reproduced speech vector SWI corresponding to the original speech vector S.

[Problem to be Solved by the Invention] The above-described code excitation linear prediction encoder 10 is considered to be mainly applied to compression encoding of voice in digital mobile communication. Mobile communication devices are required to be miniaturized due to their portability and mobility, and low power consumption is required due to the use of a battery power supply. Therefore, the code excitation linear predictive encoder 10 is also required to be as small as possible and consume less power.

By the way, the code excitation linear prediction encoder 10 is provided with each functional unit that performs a desired function by arithmetic processing as described above. In order to reduce the size and power consumption of the encoder 10, it is required to reduce the size and power consumption of each functional unit that performs a desired function through arithmetic processing. When the number of bits provided for the arithmetic processing is small, the size and power consumption of each functional unit that performs a desired function by the arithmetic processing can be reduced.

However, the current code excitation linear predictive encoder 10
Because the internal value that must be represented by the volume of the voice fluctuates greatly, it is very difficult to perform meaningful processing (for example, short-time analysis) with fixed point and without reducing the calculation accuracy. This requires an operation with a large number of bits, which makes it difficult to reduce the size and power consumption.

Note that it is difficult to reduce the size even with an encoder that performs an operation using floating-point representation.

Such a problem occurs not only in the code excitation linear prediction encoder but also in the code excitation linear prediction decoder corresponding to the encoder.

The present invention has been made in view of the above points,
A code-excited linear prediction code capable of performing an operation for realizing a desired function with a small number of bits without lowering the calculation accuracy even if the volume of the input voice fluctuates, and realizing miniaturization and low power consumption. And a decoder and a decoder.

[Means for Solving the Problem] In order to solve the problem, according to a first aspect of the present invention, an analysis parameter obtained by analyzing an original speech vector,
Using the analysis parameters, a plurality of excitation code vectors are synthesized and, among a plurality of prediction vectors, an index of an optimal excitation code vector that minimizes an error from an original speech vector is output as a total code. The code-excited linear prediction encoder was as follows.

That is, at least a part of the plurality of functional units for realizing a predetermined function (analysis or the like) by arithmetic processing can be changed in fixed-point representation, and the magnitude of the original speech vector is correlated with the average speech power. And fixed-point expression control means for changing the fixed-point expression of each functional unit according to the magnitude of the detected original speech vector.

In this case, it is preferable that the control signal from the fixed-point expression control means is also included in the total code and output.

The second invention relates to a code-excited linear prediction decoder corresponding to a code-excited linear prediction encoder that outputs a control signal from a fixed-point representation control unit in a total code. Then, at least a part of the plurality of functional units that realize the predetermined function by the arithmetic processing can have a fixed-point representation changeable, and each functional unit can be fixed according to a control signal included in the total code. It is characterized in that the decimal point expression is changed.

[Operation] The first aspect of the present invention is to reduce the size and power consumption of at least some of the plurality of functional units that realize a predetermined function (analysis or the like) by arithmetic processing.
Therefore, in order to reduce the number of processing bits, each of the functional units can be changed in fixed-point representation. Then, the magnitude of the original speech vector is detected by the detection means, and the fixed-point representation control means changes the fixed-point representation of each functional unit according to the magnitude of the detected original speech vector.

In this case, it is preferable to output a control signal from the fixed-point representation control means in the total code in order to reduce the size and power consumption of the decoder.

The second invention relates to a code excitation linear prediction decoder for decoding a total code including a control signal from a fixed-point representation control means, and includes a plurality of functional units for realizing a predetermined function by arithmetic processing. Among them, at least a part of the fixed-point expression can be changed, and such a functional unit is configured to change the fixed-point expression according to a control signal included in the total code.

[Embodiment] Hereinafter, an embodiment of a code excitation linear prediction encoder according to the present invention will be described in detail with reference to the drawings.

FIG. 1 is a functional block diagram showing the configuration of this embodiment, and the same and corresponding parts as in FIG. 2 are denoted by the same reference numerals. FIG. 3 to FIG. 5 are explanatory diagrams of the concept of adopting such a configuration.

Prior to the description of the configuration and operation of the embodiment, the concept of adopting the configuration of the embodiment will be described with reference to FIGS.

As described above, when trying to reduce the size and power consumption of an encoder, the number of bits to be processed by each functional unit that realizes a desired function (short-time analysis or long-time analysis) by arithmetic processing is possible. It must be as small as possible. Therefore, a fixed-point representation method of a signal (vector component) processed by each functional unit will be considered. As an example, consider a method of expressing a vector S1 operated by the long-time analysis unit 15.

This vector S1 has a high correlation with the magnitude of the original voice (vector) S as shown in FIG. That is, the maximum value of the component of the vector S1 and the average value of each component have a high correlation with the average audio power that is the S-mean of each component of the original audio vector S. In FIG. 3, the value of the vector S1 is represented by a logarithm having a base of 2 in consideration of the bit representation.

When considering the number of bits, not the average value but the maximum value is a problem. FIG. 4 shows the maximum value regression line L1 extracted from the maximum value regression line L1 and the average value regression line L2 in FIG.

If at least m-bit precision is required for the vector S1, in order not to lose the precision when the original voice S is small,
The fixed-point representation of the vector S1 must represent from the maximum value (integralized) A of the vector S1 when the original speech S is the smallest to a number smaller by m bits. To avoid overflow, the maximum value of the vector S1 when the original voice S is the largest (integrated)
B must be expressed.

Considering this, in order to express the vector S1 so as to secure the precision with a fixed point, (BA) +
m bits are required. This is the conventional concept of determining the number of bits.

Here, as shown in FIG. 5, the size (average audio power) of the original audio S will be divided into several sections. The same concept as described above can be applied to the case where the required number of bits is independently obtained for each section x (x = 1 to N) irrespective of other sections. That is, for the section x, (Bx−Ax) +
m bits are required.

As is clear from the comparison between FIG. 4 and FIG. 5, if the loudness of the original voice S is divided into several sections and the fixed-point representation method is changed for each section, the long time analysis unit 15 is required. The number of bits to be used can be significantly smaller than in the past. Of course, the maximum number of bits (Bx−Ax) + m in each section is set as the number of bits to be processed so that the original voice S belongs to any section. It costs.

In the above description, the vector S1 handled by the long-time analysis unit 15 has been considered. However, the same concept can be applied to other functional units that execute arithmetic processing.

The embodiment shown in FIG. 1 has been made under such consideration.

Also in the encoder 10A of this embodiment, the basic functional configuration for encoding is the same as that of the related art.

That is, the short-time analysis unit 12 analyzes the short-time prediction coefficient αj obtained by short-time analysis of the original speech vector S given through the input terminal 11 and the quantizer 14 and the short-time synthesis filter
13 and the long-time analysis unit 15
Formant component (short-time component) of the voice given from
The pitch prediction coefficient β and the lag (pitch period) L obtained by analyzing the vector S1 from which the noise has been removed for a long time are supplied to the long-time synthesis filter 16 and the quantizer 14, and the quantizer 14 αj, the pitch prediction coefficient β, the lag L, and the index I of the optimum excitation code vector determined as described later are quantized, and output to the output terminal 112 as the total code C.
The configuration for outputting more is the same as that of the conventional encoder 10.

Further, each excitation code vector ei output from the excitation code table 17 is sequentially processed by the long-time synthesis filter 16 and the short-time synthesis filter 13 to form a prediction vector SWi,
After the subtracter 18 obtains the error vector eri of each prediction vector SWi from the original speech vector S, the perceptual weighting filter 19
Performs weighting in consideration of the perceptual characteristic, and calculates a square
Is 2 of each component of the perceptually weighted error vector εi.
The average operation unit 111 calculates the average gi of the sum of the squares of the components of the error vector εi, and calculates the excitation code vector
The configuration for evaluating ei to obtain the index I of the optimal excitation code vector eI is also the same as that of the conventional encoder 10.

However, the encoder 10A of this embodiment includes an average audio power calculation unit 113 and a fixed point control unit 114, and each of the functional units 12, 13, and 15 that realize desired functions by arithmetic processing.
111111 differs from the conventional encoder 10 in that the fixed-point representation can be changed.

The original audio vector S is also provided to the average audio power calculator 113. The average audio power calculation unit 113 is provided to obtain a value representing the magnitude of the original audio vector S. The average audio power calculation unit 113 obtains a root mean square (average audio power) of each component of the original audio vector S, and obtains the power. Value R0 is fixed-point controller 1
Give to 14.

The fixed-point control unit 114 recognizes a section to which the power value R0 belongs (see FIG. 5) and outputs a control signal representing a recognition section x.
Output Cx. The control signal Cx is sent to the short-time analysis unit 12, the short-time synthesis filter 13, the long-time analysis unit 15, the long-time synthesis filter 16, the subtractor 18, the perceptual weighting filter 19, the square calculation unit 110, and the average operation unit 111. Given.

Each of the functional units to which the control signal Cx is applied receives the control signal Cx
Represents and processes the input vectors and coefficients in a fixed-point representation corresponding to the section x indicated by.

In the encoder 10A of this embodiment, the control signal Cx is also included in the total code C in consideration of the processing of the decoder 20 described later.

The above encoding process is repeated every time the original audio vector S changes.

Therefore, according to the above embodiment, the original speech (vector)
Since the fixed-point representation of each functional unit that performs arithmetic processing is changed according to the size of S, the number of processing bits of each functional unit can be small, and even if the size of the original voice fluctuates, the calculation accuracy can be improved. It is possible to perform an operation for realizing a desired function with a small number of bits without lowering the power consumption, thereby realizing miniaturization and low power consumption.

Next, an embodiment of a code excitation linear prediction decoder according to the present invention will be described with reference to FIG.

The decoder 20 shown in FIG. 6 corresponds to the code excitation linear prediction encoder 10A described above.

The decoder 20 includes an input terminal 21, an inverse quantizer 22, an excitation code table 23, a long-time synthesis filter 24, a short-time synthesis filter 25, and an output terminal 26.

The total code C output from the encoder 10A and supplied to the input terminal 21 of the decoder 20 is inversely quantized by the inverse quantizer 22. The index I separated by this processing is provided to the excitation code table 23, the pitch prediction coefficient β and the lag (pitch cycle) L are provided to the long-time synthesis filter 24, and the short-time prediction coefficient αj is provided to the short-time synthesis filter 25. The control signal Cx is provided to the long-time synthesis filter 24 and the short-time synthesis filter 25.

The excitation code table 23 has the same configuration as the excitation code table 17 in the encoder 10A, and outputs the excitation code vector eI corresponding to the given index I to the long-time synthesis filter 24.

The long-time synthesis filter 24 has the same configuration as the long-time synthesis filter 16 in the encoder 10A, including that the fixed-point representation can be changed. The long-time synthesis filter 24 performs inverse processing (synthesis processing) of the long-time analysis unit 15 on the excitation code vector eI using the pitch prediction coefficient β and the lag L, and converts the processing output vector STI into the short-time synthesis filter 25. To send to. At this time, a post-process is executed by expressing a vector or the like to be processed in a fixed-point representation according to the control signal Cx.

The short-time synthesis filter 25 has the same configuration as the short-time synthesis filter 13 in the encoder 10A, including that the fixed-point representation can be changed. After changing the fixed-point representation of the vector STI in accordance with the control signal Cx, the short-time synthesis filter 25 performs the reverse processing (synthesis processing) of the short-time analysis unit 13 using the short-time prediction coefficient αj to process the output vector SWI is obtained, and this is output as a reproduction vector of the original audio vector S via the output terminal 26.

According to the decoder 20 of this embodiment, the fixed-point representation of signals processed by the long-time synthesis filter 24 and the short-time synthesis filter 25 is changed according to the transmitted control signal x. The number of processing bits of the filter can be small, and even if the volume of the original voice fluctuates, it is possible to perform an operation for realizing a desired function with a small number of bits without lowering the calculation accuracy. Electricity can be realized.

Although the encoder 10A of the above-described embodiment also includes the control signal Cx in the total code C, the control signal Cx may not be included, and the encoder 10A of the above-described embodiment may be omitted. It has the same effect as. In this case, the same decoder as the conventional one is applied.

In the encoder 10A of the above-described embodiment, the control signal Cx is provided to all the functional units that perform the arithmetic processing. However, the control signal Cx may be provided only to a part of the functional units that perform the arithmetic processing. Further, the method of classifying the loudness of the original voice may be different for each functional unit.

Further, in the above-described embodiment, the case where the present invention is applied to the encoder having both the short-time analysis unit and the long-time analysis unit has been described. However, the present invention can be applied to at least the encoder having the short-time analysis unit. . The present invention can also be applied to an encoder in which the excitation code table is of an adaptive type.

[Effects of the Invention] As described above, according to the present invention, the fixed-point representation of the functional unit that performs arithmetic processing is changed according to the magnitude of the original signal. An operation for realizing a desired function can be performed with a small number of bits without deteriorating the calculation accuracy, so that downsizing and low power consumption can be realized.

[Brief description of the drawings]

FIG. 1 is a block diagram showing one embodiment of a code-excited linear prediction encoder according to the first invention, FIG. 2 is a block diagram showing a conventional encoder, and FIGS. FIG. 6 is a functional block diagram of an embodiment of a code-excited linear prediction decoder according to the second invention, which is an illustration of the concept of adopting the configuration shown in FIG. 12 ... short-time analysis unit, 13, 25 ... short-time synthesis filter,
14 Quantizer, 15 Long time analysis unit, 16, 24 Long time synthesis filter, 17, 23 Excitation code table, 18
... Subtractor, 19 ... Perceptual weighting filter, 110 ... Square calculator, 111 ... Average operation unit, 113 ... Average voice power calculator, 114 ... Fixed point controller, 22 ... Dequantizer .

──────────────────────────────────────────────────続 き Continuation of the front page (72) Inventor Kenichiro Hosoda 1-7-12 Toranomon, Minato-ku, Tokyo Oki Electric Industry Co., Ltd. (56) References JP-A-1-54497 (JP, A) 63-261925 (JP, A) (58) Fields investigated (Int. Cl. ⁶ , DB name) G10L 3/00-9/18 H03M 7/30 JOIS file (JICST)

Claims (4)

(57) [Claims] An analysis parameter obtained by analyzing an original speech vector, and a plurality of prediction vectors obtained by synthesizing a plurality of excitation code vectors using the analysis parameter, the error from the original speech vector is minimized. In a code excitation linear prediction encoder that collectively outputs an index of an optimal excitation code vector and a total code, a fixed-point representation is used for at least a part of a plurality of functional units that realize a predetermined function through arithmetic processing. Detecting means for detecting the magnitude of the original speech vector based on the correlation with the average speech power, and the above-mentioned functions capable of changing the expression according to the magnitude of the detected original speech vector Fixed-point representation control means for changing a fixed-point representation of a section. 2. The code-excited linear prediction according to claim 1, wherein said detecting means has a table in which a section of the average audio power of the original audio vector is associated with a maximum value of the original audio vector. Encoder. 3. The code-excited linear predictive encoder according to claim 1, wherein a control signal from said fixed-point representation control means is included in a total code and output. 4. A code-excited linear predictive decoder for receiving a total code output from the code-excited linear predictive encoder according to claim 3, wherein said plurality of functional units realize predetermined functions by arithmetic processing. Wherein the fixed-point representation is at least partially changeable and the fixed-point representation of each of the functional units is changed in accordance with the control signal contained in the total code. Decoder.