TW201329960A - Quantization device and quantization method - Google Patents

Abstract

Provided are a quantization device and a quantization method of obtaining enough encoding performance by less computation load for reducing encoding distortion. In a multistage vector quantization unit, in a situation of setting vector quantization of a first order as a pre-assigned alternate number N where 1 is subtracted from the alternate number each time when the vector quantization unit after the second order entering next order, and alternate number is under 3, evaluation is carried out for quantization distortion. Alternate number of the next order is set as a predefined numerical value P when the quantization distortion is greater than a specified threshold value. Alternate number of the next order is set as a numerical value Q less than the predefined P when quantization distortion is less than a specified threshold value.

Description

Quantization device and quantization method Field of invention

The present invention relates to a quantization apparatus and a quantization method for performing quantization using a tree search.

Background of the invention

In mobile communication, in order to effectively utilize the transmission band, it is necessary to compress and encode the digital information of voice and video. Among them, there is a great expectation for a speech codec (encoding/decoding) technique widely used for mobile phones, and a higher quality of sound is strongly required for a conventional high-efficiency coding having a high compression ratio. In addition, in order to make the public use, it must be standardized, and the world is actively researching and developing.

In recent years, in the ITU-T (International Telecommunication Union Telecommunication Standardization Sector) and MPEG (Moving Picture Expert Group), a codec capable of encoding speech or encoding music has been studied. Standardization requires a more efficient and high quality speech codec.

The performance of the speech coding technology is greatly improved by the CELP (Code Excited Linear Prediction) established 20 years ago. The CELP models the speech sounding mechanism and applies the vector quantization skillfully. . In the international standard, CELP is adopted in many standard methods such as ITU-T standard G.729, G.722.2, ETSI standard AMR, AMR-WB, 3GPP2 standard VMR-WB.

The main technique of the above CELP is LPC (Linear Prediction Coding) analysis capable of encoding the outline of the speech spectrum at a low bit rate, and quantization of parameters obtained by LPC analysis. In particular, most of the most recent standard methods are based on line spectrum quantization. Representative LSP (Line Spectral Pair) and ISP (Immittance Spectral Pair) with improved LSP, both of which are interpolated with vector quantization (hereinafter referred to as "VQ (Vector Quantization)" has a higher affinity. By using these for encoding, spectral information can be transmitted at a low bit rate. Thereby, the performance of the CELP-based codec is particularly improved.

Recently, in response to the requirements of high-efficiency and high-quality speech codecs, codecs that encode wide-band signals (16 kbps) and ultra-wideband signals (32 kbps) are gradually standardized in ITU-T, MPEG, 3GPP, and the like. . When an LPC coefficient is used to encode a wide-band or ultra-wideband digital signal, it is necessary to encode an LSP or an ISP having a large number of stages of 16 or more with a large number of bit elements. Therefore, generally, a "divided VQ" in which a coding target (object vector) is divided into a plurality of pieces and quantized separately is used. However, since statistical correlation between elements of a vector cannot be used, coding performance is degraded.

Therefore, as a method capable of obtaining better coding performance, multiple stage quantization is used. This does not divide the target vector, but uses a plurality of smaller vector quantizations to continuously quantize to gradually reduce the error. That is, a method of quantizing the quantized error vector of the previous stage in the next stage. As long as the error is the smallest in the previous stage, the amount of calculation can be greatly reduced. However, if only the quantization result with the smallest error is used as a candidate for multi-stage quantization, the comprehensive coding distortion is not sufficiently small, and the quantization performance is deteriorated.

Therefore, it is considered to use a tree search which retains candidates of quantization results having a small error from the upper level. Thereby, higher encoding performance can be obtained with less calculation amount. In particular, when the number of allocation bits is large, the number of stages is increased in order to suppress the amount of calculation. However, in the multi-level quantization of a large number of stages, if tree search is not used, sufficient Quantitative performance.

Patent Document 1 describes a method of quantizing an excitation vector of a CELP in multiple stages. In addition, it is known that in the case of a large number of stages, an efficient search can be performed by using a tree search. As a multi-level search method that is efficient, it is known to perform a search by setting the number of candidates (quantization results with small errors) remaining in each stage to N, and this method is called "N-optimal search (N Best search)”.

Further, in Patent Document 2, vector quantization is not used, but a search example based on N-optimal search is described.

Prior technical literature Patent literature

[Patent Document 1] Japanese Patent Laid-Open Publication No. 2003-8446

[Patent Document 2] Japanese Patent Laid-Open Publication No. 2000-261321

However, the above-described multi-level vector quantization using the N-optimal search of N>1 enables the final coding distortion to be smaller than the candidate for each level (N=1), but the amount of calculation is increased by N times. On the other hand, if the number of N is suppressed to be small, the coding distortion is large.

Thus, in the multi-level vector quantization using the conventional N-optimal search, it is not tried to reduce the coding distortion with a smaller amount of calculation, and thus it is not possible to obtain sufficient coding performance.

It is an object of the present invention to provide a quantization apparatus and a quantization method which reduce coding distortion with a small amount of calculation and obtain sufficient coding performance.

The quantization apparatus of the present invention performs multi-level quantization using a tree search, and the quantization apparatus adopts a structure including: a search unit that matches each target of one or more objects of the encoding target with a code vector stored in the codebook. One or more candidates from the least quantized distortion determined in the previous stage or in a predetermined number of candidates; the calculation unit calculates the quantization error vector by subtracting the code vector from the target for the candidate; and the candidate number The determination unit determines the number of candidates to be used in the next stage based on the number of candidates determined in the preceding stage.

The quantization method of the present invention performs multi-level quantization using tree search, and the method includes the steps of: matching each target of one or more targets of the encoding object with a code vector stored in the codebook, and obtaining a pre-specified first level The candidate number is one or more candidates from the one with the smallest quantization distortion, and the candidate number determined in the previous stage is obtained after the second level, and one or more candidates are selected from the least quantization distortion; for the candidate, The quantization error vector is calculated by subtracting the code vector from the target; and the number of candidates used in the next stage is determined based on the number of candidates determined in the preceding stage.

According to the present invention, it is possible to reduce coding distortion with a small amount of calculation, and to obtain sufficient coding performance.

Simple illustration

Fig. 1 is a block diagram showing the configuration of a CELP encoding apparatus according to a first embodiment of the present invention.

Fig. 2 is a block diagram showing the internal structure of the multilevel vector quantization unit shown in Fig. 1.

Fig. 3 is a block diagram showing the internal structure of the vector quantization unit shown in Fig. 2.

Fig. 4 is a flowchart showing a procedure for determining the number of candidates in the candidate number determining means shown in Fig. 3.

Fig. 5 is a flowchart showing a procedure for determining the number of candidates in the candidate number determining means in the second embodiment of the present invention.

Form for implementing the invention

Hereinafter, embodiments of the present invention will be described in detail with reference to the accompanying drawings.

(First embodiment)

Fig. 1 is a block diagram showing the configuration of a CELP encoding apparatus 100 according to a first embodiment of the present invention. The CELP encoding apparatus 100 encodes the channel information in the voice signal S11 composed of the channel information and the excitation information by obtaining an LPC parameter (linear prediction coefficient). Further, the CELP encoding apparatus 100 determines which of the pattern data of the pre-stored speech pattern is used for determining the excitation information, that is, which excitation vector (code vector) is generated in the adaptation codebook 103 and the fixed codebook 104. Code data, encoding the incentive information.

Specifically, each unit of the CELP encoding apparatus 100 performs the following operations.

The LPC analysis unit 101 performs linear prediction analysis on the speech signal S11, obtains the LPC parameters of the spectral envelope information, and outputs them to the multi-level vector quantization unit 102 and the auditory weighting unit 111.

The multi-level vector quantization unit 102 performs multi-stage vector quantization on the LPC parameters obtained by the LPC analysis unit 101, outputs the obtained quantized LPC parameters to the LPC synthesis filter 109, and outputs the code data of the quantized LPC parameters to the CELP encoding device 100. external.

On the other hand, the adaptation codebook 103 stores the previous drive excitation used by the LPC synthesis filter 109, based on the adaptive codebook delay (lag) corresponding to the code material indicated by the distortion minimization unit 112, based on the stored drive excitation, Generates an excitation vector for 1 sub-frame. The excitation vector is output to the multiplier 106 as an adaptive codebook vector.

The fixed codebook 104 stores a plurality of excitation vectors of a predetermined shape in advance, and outputs an excitation vector corresponding to the code material indicated by the distortion minimizing unit 112 as a fixed codebook vector to the multiplier 107. Here, a case will be described in which the fixed codebook 104 is a digital book, and the weighting is performed by an addition operation in the case of using a codebook based on two types of pulse waves.

Algebraic excitation refers to the excitation used in more standard codecs, and the excitation of a few pulses of size 1 is set only with position and polarity (+-) as information. For example, it is described in Chapter 5.3.1.9 of "CS-ACELP" in Section 5.3 of the ARIB specification "RCR STD-27K", Chapter 5.4.3.7 of "ACELP" in Section 5.4, and the like.

Further, the above-described adaptive codebook 103 is used to indicate a component having a strong periodicity such as voiced speech. On the other hand, the fixed codebook 104 is used to indicate a periodic weak component such as white noise.

The gain codebook 105 generates a gain (adaptive codebook gain) for the adaptive codebook vector output from the adaptive codebook 103, and a fixed codebook vector output from the fixed codebook 104, based on an instruction from the distortion minimizing unit 112. The gain (fixed codebook gain) is output to the multipliers 106 and 107, respectively.

The multiplier 106 multiplies the adaptive codebook gain output from the gain codebook 105 by the adaptive codebook vector output from the adaptive codebook 103, and outputs it to the adder 108.

The multiplier 107 multiplies the fixed codebook gain output from the gain codebook 105 by the fixed codebook vector output from the fixed codebook 104, and outputs it to the adder 108.

The adder 108 adds the adaptive codebook vector output from the multiplier 106 to the fixed codebook vector output from the multiplier 107, and outputs the added excitation vector as a drive excitation to the LPC synthesis filter 109.

The LPC synthesis filter 109 uses the quantized LPC parameters output from the multi-stage vector quantization unit 102 as filter coefficients, and uses the filter function generated by the adaptation codebook 103 and the fixed codebook 104 to use the excitation vector as the drive excitation, that is, The LPC synthesis filter generates a composite signal. The composite signal is output to the adder 110.

The adder 110 calculates an error signal by subtracting the synthesized signal generated by the LPC synthesis filter 109 from the voice signal S11, and outputs the error signal to the auditory weighting unit 111. In addition, the error signal is equivalent to coding distortion.

The auditory weighting unit 111 applies an audible weight to the encoded distortion output from the adder 110, and outputs it to the distortion minimizing unit 112.

The distortion minimizing unit 112 obtains, for each subframe, the indices of the adaptive codebook 103, the fixed codebook 104, and the gain codebook 105 which have the smallest coding distortion output from the auditory weighting unit 111, and outputs these indexes as code data. To the outside of the CELP encoding device 100. More specifically, the composite signal is generated based on the adaptive codebook 103 and the fixed codebook 104, and a series of processes for obtaining the coding distortion of the signal is closed-loop control (feedback control), and the distortion minimization unit 112 is in the sub-signal. The code data indicated to each codebook is changed in the frame to search for each codebook, and the code data of each codebook which is finally obtained to minimize the coding distortion is output.

In addition, for each subframe, the drive excitation when the coding distortion is minimized is fed back to the adaptation codebook 103. The adaptation codebook 103 updates the stored drive stimulus by the feedback.

Here, a search method of the fixed codebook 104 will be described. First, the search and code data of the excitation vector are derived by searching for an excitation vector that minimizes the coding distortion of the following equation (1).

(Formula 1)

E = | x - (pHa + qHs) | 2 ‧‧‧ (1)

E: coding distortion, x: coding target, p: gain of adaptive codebook vector, H: auditory weighted synthesis filter, a: adaptive codebook vector, q: gain of fixed codebook vector, s: fixed codebook vector

In general, since the adaptive codebook vector and the fixed codebook vector are searched for in an open loop (respective ring), the fixed codebook 104 is performed by searching for a fixed codebook vector that minimizes coding distortion of the following equation (2). The export of the code.

(Formula 2)

E: coding distortion, x: coding target (audio-weighted speech signal), p: optimal gain for adaptive codebook vector, H: auditory weighted synthesis filter, a: adaptive codebook vector, q: gain of fixed codebook vector , s: fixed codebook vector, y: target vector for fixed codebook search

Here, the gains p and q are determined after searching for the code of the excitation, so the search is performed with the optimum gain. Thus, the above formula (2) can be written as the following formula (3).

(Formula 3)

Further, it is known that the equation for minimizing the distortion is maximized to the same value as the function C of the following equation (4).

(Formula 4)

Therefore, in the case of searching for an excitation composed of a small number of pulse waves excited by the codebook, the above-described function C can be calculated with a small amount of calculation as long as yH and HH are calculated in advance.

Fig. 2 is a block diagram showing the internal structure of the multilevel vector quantization unit 102 shown in Fig. 1. In the present embodiment, as a quantization method of a spectral parameter (LPC parameter), multi-level vector quantization (multi-level VQ) is used. The multi-level VQ refers to a method of continuously advancing a plurality of stages of VQ and quantizing the quantization distortion of the previous stage in the next stage. Here, it is assumed that the number of quantization bits is large and the number of stages is also 6 to 10 or more, and the internal structure of the multi-level vector quantization unit 102 will be described.

The vector quantization unit 201-1 quantizes the LPC parameters obtained by the LPC analysis unit 101, that is, the encoding target (object vector). Specifically, the distance (quantization distortion) between the code vector stored in the codebook is calculated, and vector quantization of the sequence of the minimum distance is performed. In the tree search, the number of several candidates is obtained from the side with the smallest distance (quantization distortion). The vector quantization unit 201-1 obtains the virtual target vector, the code candidate (the column of the serial number (the candidate number column) in the tree search), and the candidate number as the quantization distortion, and obtains the obtained virtual target vector and code candidate. The candidate number is output to the vector quantization unit 201-2, and the code candidates are also output to the code decision unit 202.

The vector quantization unit 201-2 performs the same quantization as the vector quantization unit 201-1 on the virtual target vector (the plural number sometimes exists in the tree search) output from the vector quantization unit 201-1, and sets the virtual target vector and code. The candidate (candidate number column) and the candidate number are output to the vector quantization unit 201-3, and the code candidates are also output to the code decision unit 202.

The vector quantization units 201-3 to 201-J perform the same quantization as the vector quantization unit 201-1, respectively, and the vector quantization unit 203-J outputs the virtual target vector, the code candidate (candidate number column), and the candidate number to the code decision unit 202. .

The code decision unit 202 integrates the numbers of the candidate number columns with the least quantization distortion among the candidate number columns output from the vector quantization units 201-1 to 201-J into one data string, and transmits them to the CELP encoding apparatus 100 as code data. The outside. Further, if the final distortion is subtracted from the target vector which is the input of the multi-stage vector quantization unit 102, it is a decoded vector obtained as a result of decoding using the code material. Based on the decoded vector, the quantized LPC parameters used by the LPC synthesis filter 109 are obtained and transmitted to the LPC synthesis filter 109.

Fig. 3 is a block diagram showing the internal structure of the vector quantization unit 201-j (1≦j≦J) shown in Fig. 2. Hereinafter, the internal structure of the vector quantization unit 201-j (1≦j≦J) will be described using FIG.

Three signals are input to the vector quantization unit 201-j. One candidate number j is reserved by the quantization unit 201-j as a candidate, and is output to the number of the candidate number column and the virtual target vector of the vector quantization unit 201-(j+1) of the next stage. The next system target vector or virtual target vector (hereinafter, these are collectively referred to as "virtual target vectors"), which is the original encoding target (target vector) or the vector quantization unit 201 of the preceding stage in the middle of the stage. - (j-1) A virtual target vector obtained as a coding distortion. Finally, the candidate sequence number column j is the sequence number of each vector quantization unit with the least distortion to the vector quantization unit 201-j. Further, the target vector is one, but the virtual target vector j and the candidate number column j sometimes have a plurality of numbers.

Here, the candidate number j is taken as K, and the candidate number j-1 is taken as M. Further, in the vector quantization unit 201-1, since the target vector is one, M=1. Further, in the vector quantization unit 201-J of the final stage, only one candidate number column is required, so K=1. Note that the M-system input target vector and the number of the candidate number column j, K means the number of candidates output to the next-stage vector quantization unit 201-(j+1).

The distortion calculation and codebook search unit 301 performs matching of all of the M virtual target vectors with all code vectors stored in the codebook 302 (generally based on the Euclidean distance (as a vector, the difference is obtained for each element) And the distance calculation), searching for K candidates from the side with the smallest distance (quantization distortion), and obtaining the code number. At this time, the original serial number column is also determined. Then, referring to the candidate number column j, the candidate code number is connected to the original number column to calculate the K candidate number columns j+1, and is output to the next-stage vector quantization unit 201-(j+1). Further, the candidate number j, the code vector of the code number of the candidate, and the target vector of the quantization target are output to the virtual target calculation unit 304. Further, one of the candidate number j and the coding distortion is output to the candidate number determining unit 303.

Further, in the case where the vector quantization unit 201-j is the vector quantization unit 201-1 of the first stage, the candidate number j and the candidate number sequence j are set in advance in the vector quantization unit 201-1, and only the target vector is input. Further, in the case where the vector quantization unit 201-j is the vector quantization unit 201-J of the last stage, the number of candidates is 1, and the number whose minimum distance (quantization distortion) is the smallest is connected to the candidate number column corresponding to the target vector. When it is output to the code decision unit 202 as the candidate number column j+1, the candidate number determining unit 303 and the virtual target calculating unit 304 can be made to function.

Hereinafter, a specific processing example of the distortion calculation and codebook search unit 301 will be described. Let j=4, M=4, K=3, and the vector length be L. Since the target (here, the virtual target vector) is x _i ⁰ , x _i ¹ , x _i ² , x _i ³ , the candidate serial number column is j. =4, so assume that there are three levels of vector quantization units using a 64-bit (6-bit) codebook, and (5,12,31)(5,12,48) (31,11,57) ) Four columns of (31, 3, 18). Each of the four columns of candidate columns has a one-to-one relationship with the four virtual target vectors. Set the code vector to C _i ^m . The serial number of the m-code vector. The quantization distortion E _n,m is expressed by the following formula (5).

(Formula 5)

Moreover, the first three-digit code number with the smallest quantization distortion E _{n,m is} obtained. As a result of the evaluation, it is assumed that the first three digits are (1) the code number 35 when the virtual target vector is 0, (2) the code number 8 when the virtual target vector is 0, and (3) the code number when the virtual target vector is 3. 52. If the code signal is added last with reference to the candidate serial number column, the three serial numbers of the next transmission are (5, 12, 31, 35), (5, 12, 31, 8), (31, 3, 18). , 52) as the candidate serial number column j+1. Further, three sets of virtual target vectors and code vectors of (x _i ⁰ , C _i ³⁵ ), (x _i ⁰ , C _i ⁸ ), (x _i ³ , C _i ⁵² ) are output to the virtual target calculation unit 304. Further, one of the candidate number 3 and the first three bits (quantization distortion) is output to the candidate number determining unit 303. Further, in the present embodiment, it is also possible to output any one of three distances. This is because no matter which distance is output, there is no big difference in performance.

The candidate number determining unit 303 refers to the candidate number j and the distance (quantization distortion) output from the distortion calculation and codebook search unit 301, and determines the candidate number j+1 used by the vector quantization unit 201-(j+1) of the next stage. And it is output to the vector quantization unit 201-(j+1).

The virtual target calculation unit 304 refers to the group of the target and the code vector output from the distortion calculation and codebook search unit 301, and subtracts the code vector from the target vector to calculate K virtual target vectors j+1. In the above specific example, the three vectors of (x _i ⁰ - C _i ³⁵ ), (x _i ⁰ - C _i ⁸ ), and (x _i ³ - C _i ⁵² ) are virtual target vectors j+1.

Next, the above-described candidate number determining unit 303 will be described in detail including the effect of the arithmetic algorithm. First, in the N-optimal search used by the tree-like search VQ, in the case where the number of stages is large, the amount of calculation increases to N times in proportion to the number of candidates N. Conversely, if N is decreased, the quantization performance deteriorates. . Therefore, the inventors repeated the simulation experiment of the multi-stage VQ using the tree search, performed the performance analysis of the tree search, and extracted the following four trends.

That is, (1) the number of candidates N in the N-optimal search is increased or changed for each stage, and the performance corresponding to the calculation amount cannot be obtained. Retaining a plurality of candidates has an effect on the quantization performance in the initial stage of multi-level quantization.

(2) When one step is advanced, if the number of search candidates is drastically lowered, the quantization performance is greatly degraded.

(3) There is a huge difference between N=2 and N=1. In the case of a large number of stages, substantially sufficient quantization performance can be obtained with N=2.

(4) When the coding distortion does not become small after a plurality of stages are advanced, the probability of deterioration of the final outlier (the ratio of the quantization error to a certain value or more) is increased.

In view of the above tendency, the inventors have created a tree search by combining the following three algorithms. That is, it is performed by the following steps.

(Step 1) In the first stage, only the pre-specified number of candidates N is retained and the next stage is entered.

(Step 2) Starting from the second stage, each time the next stage is entered, the number of candidates is reduced by one as N-1 and N-2.

(Step 3) When the number of candidates is equal to or lower than the predetermined value P, the quantization distortion is evaluated each time. When the number of candidates is greater than a predetermined threshold, the number of candidates for the next stage is P, and when the threshold is equal to or lower than the threshold value, The number of candidates for the next stage is set to a value Q smaller than the predetermined P. In the following description, P=3 and Q=2 will be described as examples of P and Q. In addition, in the case where the amount of calculation is sufficient, the value can also be a larger value. At this time, the coding distortion can be further reduced.

The system candidate number determining unit 303 to which such an operation algorithm is applied, as a result, can be selected in the first part by subtracting 1 each time the candidate is set to the next level (i.e., (step 2)). As a candidate, it is possible to find the minimum number of candidates as early as possible without deteriorating the quantization performance, and it is possible to obtain sufficient quantization performance with a small amount of calculation. In addition, when the number of candidates is 3 (=P) or less, the quantization distortion is evaluated each time, and if the quantization distortion is large, the number of candidates is increased to 3 (= P), and if the quantization distortion is sufficiently small, The number of candidates is reduced to 2 (= Q) (i.e., (step (3)), and thus it is possible to control to achieve sufficiently small coding distortion with a minimum amount of calculation, and to obtain sufficient quantization performance with a small amount of calculation.

Next, the candidate number determining step in the candidate number determining unit 303 will be described using FIG. In the following description, the candidate number j+1 is denoted by KK. The input-to-candidate determination unit 303 is the candidate number j(K) and the distance (quantization distortion) obtained from the distortion calculation and codebook search unit 301. It is assumed that the number of stages J is grasped by the candidate number determining unit 303. In addition, it is assumed that the initial value of K and the reference value of the distance are predetermined before starting the quantization. Further, in Fig. 4, for example, 50000 is used as the reference value of the distance, but there may be cases where other values are appropriate. The appropriate value may be determined according to the dimension of the vector or the value of the element.

First, in the step (hereinafter, abbreviated as "ST") 401, it is determined whether or not the level number j = 1, that is, whether it is the vector quantization unit 201-1, and in the case where the level number j = 1 ("Yes"), The process proceeds to ST402, and in the case of the non-level number j=1 ("NO"), the process proceeds to ST405.

In ST402, the candidate number K (in this case, the initial value of K) is taken as an input, and it is determined whether the total number of stages is greater than 7, and if the total number of stages is greater than 7, the process proceeds to ST403, and the total number of stages is not more than 7. In the case, transfer to ST404. Further, in addition to the numerical value of "7", of course, depending on the conditions, there may be cases where other values are appropriate. The appropriate value may be determined in advance based on the total number of stages or the initial value of the number of candidates.

In ST403, KK=K-1 is set, and in ST404, KK=K is set.

In ST405, it is determined in ST401 that the non-level number j=1 (non-vector quantization unit 201-1) is set to KK=K-1, and in ST406, it is determined whether or not the level number j=4 or more and the distance. (Quantization distortion) exceeds the reference value, and if the condition is satisfied (YES), the process proceeds to ST407, and if the condition is not satisfied (NO), the process proceeds to ST409. Here, although the order number j=4 or more is set here, there may be cases where other values are appropriate.

In ST407, it is determined whether KK is less than 3 (=P), and if KK is less than 3 (=P) ("Yes"), the transition to ST408 is set to KK=3, and KK is not less than 3 (=P). ) ("No"), the process moves to ST411.

Further, in ST409, it is determined whether KK is less than 2 (=Q), and when KK is less than 2 (=Q) ("Yes"), the process proceeds to ST410 and is set to KK=2, and KK is not less than 2 ( In the case of =Q) ("No"), the process moves to ST411.

As described above, in ST406 to ST410, when the phase distance (quantization distortion) at which the quantization is performed to a certain extent is sufficiently small, the number of candidates is made small, and when the distance is large, the distance is made larger. The larger the number of candidates, the smaller the integrated quantization distortion. This is an operation algorithm that ensures "2" (=Q) of the lowest candidate number and uses the candidate number "3" (=P) to make the integrated quantization distortion smaller. In the quantification experiment of the present inventors, it was confirmed that the outlier value (the ratio of the quantization distortion to a certain larger value or more) can be reduced by the determination of the distance.

In ST411, it is determined whether or not the level number j=J, that is, whether it is the final level, and if the level number j=J ("Yes"), the process proceeds to ST412, and the non-level number j=J ("No") In the case of this, the candidate number determination step in the stage is ended.

In ST412, KK=1 is set, and the candidate number determination processing in the final stage is ended.

Here, in order to express the effectiveness of the present invention, a quantitative experiment of ISF quantization applicable to CELP is shown. The encoder is based on CELP, the bit rate is about 24 kbps, and the data used is a Japanese 40 sample of the frequency of the wide frequency. The quantified is a 16-dimensional vector of the ISF (Immittance Spectral Frequency). The tree search of the multi-stage VQ system N reference as a reference has six or more stages. The present invention uses the same N as the initial candidate number. Table 1 below shows the results of the quantitative experiments.

As can be seen from the above Table 1, the calculation amount of the maximum frame is reduced by about 1.7 wMOPS (weitghed Mega Oparation Per Second), and the amount of calculation can be drastically reduced. In addition, it can be seen that the S/N ratio (Signal/Noise ratio) is almost constant, and the synthesized sound hardly deteriorates in the objective value. Even if the distortion of the ISF is compared by SD (Spectral Distance), it is only a slight deterioration of 0.01 dB, and in the outlier value of the ratio of 2 dB or more, the deterioration is only 0.2%. This ratio of 500 frames per frame indicates that it hardly deteriorates. Moreover, since the increase in processing caused by the present invention is only a decision of the number of candidates, and the amount of calculation is slight, the overall impact on the operation algorithm is also small.

As described above, according to the first embodiment, in the multi-stage VQ using the tree search, the number of candidates N specified in advance is set in the first stage, and the number of candidates is reduced by 1 each time after entering the next stage after the second stage. When the number of candidates is 3 or less, the quantization distortion is evaluated each time, and when it is larger than a predetermined threshold, the number of candidates in the next stage is set to 3 (=P), and when the threshold is equal to or lower than the threshold value, Set the number of candidates for the next level to 2 (=Q). Thereby, it is possible to select a reliable candidate in the first portion, and it is possible to find the minimum candidate number as early as possible without deteriorating the quantization performance, and it is possible to obtain sufficient quantization performance with a small amount of calculation. In addition, it is possible to control to achieve sufficiently small coding distortion with a minimum amount of calculation.

(Second embodiment)

The configuration of the CELP encoding apparatus according to the second embodiment of the present invention is the same as the configuration shown in the first embodiment of the first embodiment, and differs only in the function of the candidate number determining unit 303 of the vector quantization unit 201-j. It needs to be described with reference to Figs. 1 to 3 .

Fig. 5 is a flowchart showing a procedure for determining the number of candidates of the candidate number determining unit 303 in the second embodiment of the present invention. Hereinafter, the candidate number determining step will be described using FIG. In the fifth embodiment, the same reference numerals are given to the portions that are the same as those in the fourth embodiment, and the overlapping description will be omitted.

In addition, in the following description, the same conditions as the fourth figure of the first embodiment are used. That is, the candidate number j+1 is represented by KK. The input-to-candidate determination unit 303 is the candidate number j(K) and the distance (quantization distortion) obtained from the distortion calculation and codebook search unit 301. Further, it is assumed that the number of stages J is determined by the candidate number determining unit 303. In addition, it is assumed that the initial value of K and the reference value of the distance are predetermined before starting the quantization. In addition, in FIG. 5, for example, 50000 is used as the reference value of the distance, but there may be cases where other values are appropriate. The appropriate value may be determined according to the dimension of the vector or the value of the element.

In ST501, it is determined whether or not the level number j=3 or more, or is KK=3 or less, and if the condition is satisfied (YES), the process proceeds to ST502, and the condition is not satisfied (No). In the case, transfer to ST411.

In ST502, it is determined whether the distance (quantization distortion) exceeds the reference value, and if it exceeds (YES), the process proceeds to ST407, and if it does not exceed (NO), the process proceeds to ST409.

As described above, according to the second embodiment, it is confirmed that the candidate number KK is sufficiently small before the evaluation of the quantization distortion, and if the candidate number KK is sufficiently small, the candidate number control using the quantization distortion can be immediately performed, and the number of candidates can be minimized. The amount of calculations is sufficient to quantify performance.

Further, in each of the above embodiments, as shown in FIG. 3, the candidate number determining unit 303 is provided in the subsequent stage of the distortion calculation and codebook search unit 301, but the candidate number determining unit 303 may be set in the distortion calculation and The previous stage of the codebook search unit 301. At this time, it is needless to say that the candidate number determining unit 303 can obtain the same effect by using the distance (quantization distortion) and the number of candidates from the vector quantization unit of the previous stage.

Further, in the above embodiments, the example in the CELP is shown, but the present invention is applicable to the invention of vector quantization, and therefore it is not limited to the CELP system. For example, it can be used for quantization of spectrum using MDCT (Modified Discrete Cosine Transform) or QMF (Quadrature Mirror Filter), and can also be applied to the low frequency domain of the band extension technique. An algorithm for searching for similar spectral shapes in the spectrum. In addition, the present invention can be applied to all encoding methods using LPC analysis.

Further, in each of the above embodiments, an example of encoding the ISF is shown. However, the present invention is not limited thereto, and may be applied to ISP (Immittance Spectrum Pairs) and LSP (Line Spectrum Pairs). The case where the parameters such as spectrum pair) and PARCOR (PARtial autoCORrelation) are quantized. This is because other quantization methods are used instead of the ISF quantization in the embodiment.

Further, in the above embodiments, the present invention has been applied to the tree search VQ of the spectral parameters of the CELP, but the present invention is also self-evident for the quantization of other parameter vectors. This is because the nature of the parameters does not affect the invention.

Further, in the above embodiments, the Euclidean distance is used for the distortion calculation and codebook search unit 301, but other distance scales such as weighted Euclidean distance or city block distance (sum of absolute values) may be used. . Therefore, the present invention relates to an operation algorithm of the candidate number decision unit 303, and the distance scale is not relevant to the present invention.

Further, in the above embodiments, the case where the encoder is applied is shown, but the present invention is also applicable to tree search used in pattern matching such as speech recognition or image recognition. Therefore, the present invention relates to the determination of the number of candidates for tree search, and does not affect the overall purpose of the algorithm.

Further, the encoding device described in each of the above embodiments can be used by being mounted on a communication terminal device or a base station device.

Further, in each of the above embodiments, the reference value compared with the distance (quantization distortion) is used as a predetermined constant. However, it may be a case that the value is different depending on each stage (level number). . Therefore, the present invention does not limit the reference value. A more efficient search can be achieved by changing the reference value at each level (level number).

Further, in each of the above embodiments, the predetermined numerical values of "3 and 2" are used for the control of the candidate number, but numerical values such as "4 and 3", "4 and 2" may be used. In addition, the numerical value may be different in each stage (level number). These values can be set according to various situations such as a case where the amount of calculation is sufficient or a case where higher performance is required.

Further, in the second embodiment, the predetermined numerical values (constants) of "3 and 3" are used for the determination of j and KK, respectively, but may be changed to "2 and 2", "2 and 3", and "4". And 3", "2 and 4", "4 and 4" or "5 and 4", etc. In addition, it may be different in each level (level number). These values can be set in various cases such as a case where the amount of calculation is sufficient and a case where higher performance is required.

Further, in the above embodiments, the present invention has been described by way of hardware, but the present invention can also be realized by software in combination with a hardware.

Further, each functional block used in the description of each of the above embodiments is typically implemented as an LSI of an integrated circuit. These may also be individually singulated, or may be singulated in part or in whole. Here, it is an LSI, but it may be called an IC, a system LSI, a super LSI, or an ultra LSI depending on the difference in the degree of integration.

Further, the method of integrating the circuit is not limited to the LSI, and may be implemented by a dedicated circuit or a general-purpose processor. It is also possible to use an FPGA (Field Programmable Gate Array) that can be programmed after the LSI is manufactured, or a reconfigurable processor that can reconfigure the connection or setting of a circuit inside the LSI (Reconfigurable) Processor).

Furthermore, when a technology that replaces the integrated circuit of LSI is developed due to advances in semiconductor technology or other technologies derived therefrom, it is naturally also possible to use a technique to integrate the functional blocks. It is also possible to apply biotechnology and the like.

The disclosure of the specification, drawings, and abstract included in Japanese Patent Application No. 2010-210116, filed on Sep. The contents are all incorporated by reference.

Industrial availability

The quantization apparatus and the quantization method of the present invention can be applied to a speech encoding apparatus or the like.

101. . . LPC analysis unit

102. . . Multilevel vector quantization unit

103. . . Adaptive code book

104. . . Fixed codebook

105. . . Gain codebook

106, 107. . . Multiplier

108, 110. . . Adder

109. . . LPC synthesis filter

111. . . Auditory weighting unit

112. . . Distortion minimization unit

201-1 to 201-J. . . Vector quantization unit

202. . . Code decision unit

301. . . Distortion calculation and codebook search unit

302. . . Code book

303. . . Candidate determination unit

304. . . Virtual target computing unit

ST401~ST412. . . step

ST501~ST502. . . step

Fig. 1 is a block diagram showing the configuration of a CELP encoding apparatus according to a first embodiment of the present invention.

Fig. 2 is a block diagram showing the internal structure of the multilevel vector quantization unit shown in Fig. 1.

Fig. 3 is a block diagram showing the internal structure of the vector quantization unit shown in Fig. 2.

Fig. 4 is a flowchart showing a procedure for determining the number of candidates in the candidate number determining means shown in Fig. 3.

Fig. 5 is a flowchart showing a procedure for determining the number of candidates in the candidate number determining means in the second embodiment of the present invention.

102. . . Multilevel vector quantization unit

201-1 to 201-J. . . Vector quantization unit

202. . . Code decision unit

Claims (6)

A quantization apparatus for performing multi-level quantization using a tree search, the quantization apparatus comprising: a search unit that matches each target of one or more objects of the encoding object with a code vector stored in the codebook, and determines a determination in the previous stage Or one or more candidates from the one having the smallest quantization distortion, or a calculation unit that calculates the quantization error vector by subtracting the code vector from the target for the candidate; and the candidate number determining unit is based on The number of candidates determined in the previous stage determines the number of candidates to be used in the next stage. The quantizing apparatus according to claim 1, wherein the candidate number determining unit determines to use the number of candidates that are decremented by one from the number of candidates determined in the preceding stage in the next stage. The quantizing apparatus of the first aspect of the invention, wherein the candidate number determining means determines that the number of candidates determined by the preceding stage is equal to or less than a predetermined value P, and when the quantization distortion is larger than a predetermined threshold, When the quantization distortion is equal to or less than the predetermined threshold value, it is determined that the value Q smaller than the P specified in advance is used as the candidate number in the next stage. The quantizing apparatus according to claim 1, wherein, in the case of the first stage, the search unit obtains a candidate of a predetermined number of candidates from the least quantized distortion. The quantizing apparatus according to the first aspect of the invention, wherein the candidate number determining means is greater than a predetermined amount when the current number of stages is equal to or greater than a predetermined number of stages, or the number of candidates is equal to or less than a predetermined number of candidates P When the threshold value and the number of candidates are smaller than the predetermined number of candidates R, it is determined that the candidate number R is used in the next stage, and when the quantization distortion is equal to or less than the predetermined threshold value and the candidate number is less than the predetermined candidate number Q, the number of candidates is determined. The first level uses the candidate number Q. A quantization method for performing multi-level quantization using a tree search, the method comprising the steps of: matching each target of one or more objects of the encoding object with a code vector stored in a codebook, and obtaining a pre-specified level at the first level One or more candidates from the candidate having the smallest quantization distortion, and one or more candidates starting from the second stage and having the smallest number of candidates determined in the previous stage, and the candidate is the candidate. In the above object, the quantization error vector is calculated by subtracting the code vector; and the number of candidates used in the next stage is determined based on the number of candidates determined in the preceding stage.