KR100300963B1 - Linked scalar quantizer - Google Patents

Abstract

The present invention relates to a quantizer and a speech encoding / decoding method for applying a characteristic of a vector quantizer to a scalar quantizer, and a quantizer according to the present invention is characterized in that an arbitrary Nth LSF among M LSFs forms a codebook by a general SQ method (N + K) th LSF (where K is an integer greater than or equal to 0 and equal to or smaller than (MN-1)) forms a codebook based on the distribution of (N + K) th LSFs, 1) th LSF (where K is an integer equal to or greater than 0 and equal to or smaller than (N-2)) forms a codebook based on the distribution of the (NK) th LSFs.

According to the present invention, by applying the present invention to the 4kbps voice coder proposed in ITU-T, SG 16, and QN 21, improved performance can be obtained while saving memory and bits compared to the conventional method.

Description

Linked scalar quantizer

The present invention relates to a quantizer used in a speech encoding / decoding apparatus, and more particularly to a quantizer and a speech encoding / decoding method in which the characteristics of a vector quantizer are applied to a scalar quantizer.

In recent years, standardization of speech coder has been progressing globally. The speech coders participating in the standardization adopt a scheme of expressing the speech by the spectral envelope and the excitation signal, and transmitting the bitstream in which the spectral envelope and the excitation signal are respectively quantized. Generally, linear predictive coding (LPC) coefficients are used to represent the spectral envelope. However, since the LPC coefficients are too variable to be directly quantized, they are converted into line spectrum frequencies (LSFs) and quantized.

There are two methods of LSF quantization: Scalar Quantization and Vector Quantization.

The major difference between designing a M-order LSF with a Scalar Quantizer (SQ) and a Vector Quantizer (VQ) is that the correlation between the orders Correlation is used. For convenience of explanation, the M-order LSF vector is expressed as follows.

{Omega_1, omega_2, omega_3, ... , omega_M}

A code vector obtained by directly vector quantizing the M-th order LSF The from All code vectors up to the car are designed to have a correlation. In other words, silver And therefore, this The effect of limiting the distribution of Therefore, when quantized with the same number of bits, VQ shows superior performance to SQ, but requires more memory requirement and search time complexity. Assuming that M = 16, the number of required bits = 32 bits, the memory required to store one LSF is 1 byte,

16 x 2 ³² x 1 byte = 64 gigabytes (Gbyte)

It is impossible to implement it in an actual system.

The code vector obtained by scalar quantizing the M-th order LSF The from All code vectors to the car are designed independently. In other words, silver So that there are many combinations of code vectors that do not exist. On the other hand, assuming that M = 16, the number of required bits = 32 bits, and the memory required to store one LSF is 1 byte (two bits are assigned to each order)

16 × 2 ² × 1 byte = 64 bytes (byte)

Memory is required, so it is easy to implement in an actual system. However, there is a dramatic performance degradation compared to VQ.

As a result, although the SQ exhibits satisfactory performance in terms of memory and search time, it requires too many transmission bits, which is problematic to be applied to a low transmission speech coder. On the other hand, VQ can reduce the number of transmitted bits drastically compared to SQ, but there are too many search times and memory requirements to implement on a digital signal processor (DSP).

SUMMARY OF THE INVENTION The present invention has been made in order to solve the above problems, and it is an object of the present invention to provide a quantizer which has a satisfactory degree of distortion with as few bits as possible, requires as little memory as possible and requires as little search time complexity as possible, And an encoding / decoding method of LSF.

1 is a configuration diagram of a connected scalar quantizer according to the present invention.

Figure 2 shows the structure of one of the classifiers of Figure 1;

FIG. 3 illustrates a learning process of a connected scalar quantizer according to the present invention.

FIG. 4 illustrates a coding process using a connected scalar quantizer according to the present invention.

FIG. 5 illustrates a decoding process using a connected scalar quantizer according to the present invention.

In order to achieve the above object, the M-th order line spectrum frequencies {omega_1, omega_2, omega_3,. , omega_M} are codebooks to which a large number of codebooks are allocated in each of the M classes; And the M-th order line spectrum frequencies {omega_1, omega_2, omega_3, ... , and omega_M}, quantizing each of the line spectra by selecting one codebook based on the size of the adjacent line spectrum frequency among a plurality of codebooks assigned to the line spectrum frequencies, and referring to the selected codebook .

In order to achieve the above-mentioned other object, a plurality of M-th order line spectrum frequencies {{omega_1}, ... , {Omega_N}, ... , {Omega_M}} comprises the steps of (a) learning a plurality of reference line spectral frequencies {ω _N } to generate one codebook assigned to the N-th order, where 1 ? N? M); (b) a small step of dividing the distribution space of a plurality of line spectrum frequencies {ω _{N + K} } into L by changing K from 0 to (MN-1); (b.2) Each line spectral frequency ω _{N + K + 1} belonging to a plurality of line spectrum frequencies {ω _{N + K + 1} } is divided into L L spectra corresponding to the distribution space to which the line spectrum frequency ω _{N +} Small steps to divide into subsets; And (b.3) a step of generating L codebooks allocated to the (N + K + 1) -th order by learning line spectral frequencies divided into L subsets in the step (b.2) Generating codebooks assigned to higher orders by repetition; (C) a small step of dividing the distribution space of a plurality of line spectrum frequencies {ω _NK } into L by changing K from 0 to (N-2); (c.2) Divide each line spectral frequency ω _NK-1 belonging to a plurality of line spectrum frequencies {ω _NK-1 } into L subsets corresponding to the distribution space to which the line spectrum frequency ω _NK in the same frame belongs step; And (c.3) repeating a small step of generating L codebooks allocated to the (NK-1) th by learning line spectral frequencies divided into L subsets in the (c.2) And generating codebooks assigned to the lower order by the codebooks.

In order to achieve the above-mentioned other object, the M-th order line spectrum frequencies {omega_1,. , omega_N, ... , omega_M} is encoded by (a) a value obtained by quantizing the reference line spectrum frequency? _N by a codebook allocated to the Nth order And obtaining a coded value i _N (where 1? _N? M); (b) changing K from 0 to (MN-1), (b.1) changing the quantized line spectrum frequency Into a value of one of 1 to L according to its size; (b.2) a small step of finding a codebook corresponding to the value classified in the step (b.1) among the L codebooks allocated to the (N + K + 1) th order; And (b.3) a value obtained by quantizing the line spectrum frequency? _{N + K + 1} by the codebook found in the step (b.2) And a step of obtaining a coded value iN _{+ K + 1} by repeating the steps of: encoding the higher order line spectral frequencies; And (c) changing K from 0 to (N-2), (c. 1) changing the quantized line spectrum frequency Into a value of one of 1 to L according to its size; (c.2) searching a codebook corresponding to a value classified in the (c.1) small step among L codebooks allocated to the (NK-1) th order; And (c.3) a value obtained by quantizing the line spectrum frequency? _NK-1 by the codebook found in the step (c.2) And encoding the lower order line spectral frequencies by repeatedly performing a small step of obtaining the encoded value i _NK-1 .

In order to achieve the above-mentioned another object, the indexes of the M-th order line spectrum frequencies {i_1, ..., , i_N, ... , i_M} are obtained by (a) using the index i _N of the reference line spectral frequency in the codebook allocated to the N-th order to obtain a quantized line spectrum frequency (Where 1 < = N < = M); (b) changing K from 0 to (MN-1), (b.1) changing the quantized line spectrum frequency Into a value of one of 1 to L according to its size; (b.2) a small step of finding a codebook corresponding to the value classified in the step (b.1) among the L codebooks allocated to the (N + K + 1) th order; And (b.3) quantizing the line spectrum frequency using the index iN _{+ K + 1} in the codebook found in the step (b.2) Decoding the indexes of the higher order line spectral frequencies by repeating a small step of finding the higher order line spectral frequencies; And (c) changing K from 0 to (N-2), (c. 1) changing the quantized line spectrum frequency Into a value of one of 1 to L according to its size; (c.2) searching a codebook corresponding to a value classified in the (c.1) small step among L codebooks allocated to the (NK-1) th order; And (c.3) quantizing the line spectrum frequency using the index iNK _-1 in the codebook found in the (c.2) And decoding the indexes of the lower order line spectral frequencies by repeating the steps of searching for the lower order line spectral frequencies.

DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS The present invention will now be described in detail with reference to the accompanying drawings.

Hereinafter, the quantizer according to the present invention will be referred to as a Linked Scalar Quantizer (LSQ).

1, M-th order line spectrum frequencies {omega_1, omega_2, omega_3, ... , omega_M} according to the present invention includes the codebooks 10 for each order, the classifiers 20 for each order, and the quantization control unit 30. [

The LSQ includes one codebook (C _N ) assigned to the N-th order, which is the order of the reference line spectrum frequency, and codebooks allocated to each order other than N, where L is 1? N? M.

Each classifier (20) for each order receives one line spectrum frequency and determines one of the values from 1 to L according to the magnitude of the value.

Then, the quantization controller 30 quantizes the codebook by a corresponding codebook C _N for the N-th order, classifies it by one of the L low-order classifiers of the L codebooks allocated to the line spectrum frequency higher than N Quantized by a codebook corresponding to a value classified by a classifier of one degree higher order among the L codebooks allocated for a line spectrum frequency lower than N,

Fig. 2 shows a classification process of the x-th classifier (1? X? M). Reference numeral 200 denotes LSF {omega_1, omega_2, omega_3, ...} extracted from speech by the xth classifier. , omega_x, ... and it displays the entry of the second data x ω _x of omega_M}. Reference numeral 202 denotes a process of separating the input value ω _x into one of L regions, and reference numeral 204 denotes an index l _x determined by the xth classifier.

3 shows the learning process of the LSQ. Reference numeral 300 denotes M LSF {omega_1, omega_2, omega_3, ... , omega_N, ... a, omega_M N} th data set {ω _N} of illustrates a process of learning by the LBG (Linda Buzo Gray) algorithm.

Reference numerals 302 and 304 denote the learning process when the LSF degree exceeds N. FIG. Reference numeral 302 denotes a data set {omega _{N + K + 1} } to be learned as (N + K + 1) th LSF {omega _{N + K} } of degree (N + K) ) Classifier and divided into L subsets, and reference numeral 304 denotes a range of (N + K + 1) at this time.

Reference numerals 306 and 308 indicate a learning process when the LSF degree is less than N. [ Reference numeral 306 denotes a set of L subsets that pass the data set {ω _NK-1 } to be studied through a (NK) classifier with an LSF {ω _NK } of order NK higher than its degree NK- , And reference numeral 308 denotes a range of (NK-1) at this time.

4 shows the encoding process of the LSQ. Reference numeral 400 denotes M LSF {omega_1, omega_2, omega_3, ... , omega_N, ... , encoding the N-th data of the omega_M ω _N} is the process of transmitting the index i _N. At this time, the quantized value of < RTI ID = _0.0 > .

Reference numerals 402 and 404 denote the encoding process when the LSF degree exceeds N. FIG. Reference numeral 402 denotes a quantized value of the N + Kth LSF A second (N + K) and passed through a classifier to select ω _{N + K +} L of the codebook from among l _{N + K-th} codebook to quantize the _first and then, quantizes the ω _{N + K + 1} by using this, the index i _{N + K + 1 < / RTI} > At this time, And using the selected codebook according to the size of < _{RTI ID = 0.0 >} Lt; / RTI > (N + K + 1) is incremented by 1 until the value of M becomes equal to M. [ Reference numeral 404 denotes the (N + K + 1) range of the loop.

Reference numerals 406 and 408 denote the encoding process when the LSF order is less than N. FIG. Reference numeral 406 denotes a quantized value of the (NK) th LSF To a first (NK) and then passed through a classifier to select the L codebook from among l _NK second codebook to quantize the ω _NK-1, with this quantizing ω _NK-1 shows a process of transmitting the index i _NK-1 . At this time, Using the codebook selected according to the size of the symbol _NK-1 Lt; / RTI > (NK-1) is decremented by 1 until the value of (NK-1) becomes 1, Reference numeral 408 denotes the (NK-1) range of the loop.

5 shows a decoding process of the LSQ. Reference numeral 500 decodes the received index i extracted from the audio _N M number of LSF {omega_1, omega_2, omega_3, ... , omega_N, ... , omega_M}, the quantized value of the Nth data? _N .

Reference numerals 502 and 504 denote the decoding process when the LSF degree exceeds N. FIG. Reference numeral 502 denotes a quantized value of the (N + K) -th LSF A second (N + K) and passed through a classifier to select l _{N + K-th} codebook in the L codebook to quantize the ω _{N + K + 1} Then, using an index i _{N + K + 1} received from the codebook ω _{N + K + 1} Is obtained. At this time, I _{+ N + K + 1} in the codebook selected according to the size of Lt; / RTI > (N + K + 1) is incremented by 1 until the value of M becomes equal to M. [ Reference numeral 504 denotes the (N + K + 1) range of the loop.

Reference numerals 506 and 508 denote decoding processes when the LSF degree is less than N. [ Reference numeral 506 denotes a quantized value of the (NK) th LSF A first (NK) to select the second codebook _NK l of the L codebook to quantize and passed through a classifier, with the index i _NK-1 received from the codebook a quantized ω _NK-1 Is obtained. At this time, I _NK-1 in the selected codebook according to the size of Lt; / RTI > (K) is incremented by 1 until the value of (NK-1) becomes M. Reference numeral 508 specifies the (NK-1) range of the loop.

Hereinafter, the operation principle of the present invention will be described in detail.

First, we propose a method of forming a link between each order LSF of SQ to give the characteristics of VQ to SQ. LSF has a correlation between each order, and according to the existing experiment, it is found that LSF has the greatest correlation with the near LSF value. (N + K + 1) -th LSF (where K is an integer equal to or greater than 0 and equal to or smaller than (MN-1)), the Nth LSF of the M LSFs first forms a codebook by a general SQ scheme, (NK-1) th LSF (where K is an integer equal to or greater than 0 and equal to or smaller than (N-2)) forms a codebook based on the distribution of the (N + Thereby forming a codebook based on the distribution. Therefore, although it is designed in the SQ scheme, the connection scalar quantizer (LSQ), which has a memory requirement and a search time that is significantly less than VQ and whose performance is superior to that of SQ, can be designed.

(1) Design and application of Classifier

A second LSF ω _A right side of B using the (B = (A + 1) or (A-1)) classifier for classifying the second LSF ω _B of the L section (Classifier) are designed as follows.

1) If the number of S pieces of the whole learning data of _A ω, ω classifies _A Align the _A ω in descending order the data is distributed to S / L each.

2) At this time, ω _B can be classified using the boundary value of ω _A used in classification, ie, if ω _A entered as the input of the classifier exists in the lower region of the sorted S data

ω _B is limited to a certain period. Here, if L = S, it coincides with the vector quantization of? _A and? _B. If L < S, the necessary memory is reduced and the performance is degraded. Therefore, if an arbitrary value between 1 and S is selected as the value of L, a quantizer satisfactory in memory and performance can be designed.

3) The designed classifier is applied to all learning, coding and decoding processes of LSQ. In particular, LSF value quantized as input of classifier is used for coding and decoding, so that a link loop structure can be used without extra bit transmission.

(2) LSQ learning method

The learning method is as follows.

1) Select an arbitrary Nth LSF ω _N among M LSFs, and LSF ω _N learns by a general SQ method.

2) When learning the (N + 1) th LSF ω _{N + 1} (where (N + 1) is an integer equal to or less than M), the distribution of the _Nth LSF ω _N is used.

2.1) We divide the distribution of ω _N by L,

2.2) ω _N is 1, 2, ..., L _N of the presence of the interval, respectively, the codebook by learning ω _{N + 1} values separately for the _{N + 1} ω L _N different according to the type of _N ω Respectively.

3) (N + K) th LSF ω _{N +} When learning the _K (However, (N + K) is less than or equal to M, an integer exceeding N) (N + K-1 ) th LSF ω _{N + K- 1} is used.

4) (N-K + 1) -th LSF ω _{N + K + 1} when learning the (NK) th LSF ω _{N + K} (NK is less than N and an integer of 1 or more).

(3) LSQ encoding method

The encoding method is as follows.

1) First, the Nth LSF ω _N among M LSFs is encoded. Codebook expressing ω _N Is composed of P _N code vectors, the encoding is performed by transmitting the index when the Euclidean distance measure has the smallest value.

2) (N + K-1) -th LSF (N + K-1) th encoded LSF ω _{N + K} , where N + K is M or less and N is an integer greater than N . That is, the codebook of _{N + K} ω in the learning process, so that L _{N + K-1,} depending on the extent of different types of ω _{N + K-1} exists, the value has already been quantized _{N + K-1-} th codebook among the codebooks of &lt; _{RTI ID = 0.0 &gt;# N + K &} As shown in the following equation.

3) (N-K + 1) -th LSF (N-K + 1) -th encoded LSF ω _NK (NK is an integer smaller than N, . That is, the codebook of _NK ω in the learning process, so that L _{NK + 1} kinds, depending on the extent of the ω _{NK + 1} exists, has already been quantized value Using the distribution of the codebook from the _NK ω L _{'NK + 1} beonjjae As shown in the following equation.

(4) LSQ decoding method

The decoding method is as follows.

1) First, among the M LSFs, the value corresponding to the N-th LSF ω _N Is decoded as follows. When the transmitted index is P ' _N , the decoded LSF is as follows.

2) (N + K-1) -th LSF (N + K-1) th decoded LSF ω _{N + K} , where N + K is an integer smaller than or equal to M, . That is, the codebook of _{N + K} ω in the learning process, so that L _{N + K-1} kinds, depending on the extent of the ω _{N + K-1} exists, the value has already been quantized _{N + K-1-} th codebook among the codebooks of &lt; _{RTI ID = 0.0 &gt;# N + K &} And decodes it as shown in the following equation.

3) (N-K + 1) -th LSF (N-K + 1) th decoded LSF ω _NK (NK is an integer less than N, . That is, in the learning process, the codebook of? _NK has one kind of LN _{-K +} according to? _{N-K + 1} , Using the distribution of the codebook from the _NK ω L _{'NK + 1} beonjjae And decodes it as shown in the following equation.

Hereinafter, experimental results of the present invention will be described.

The speech data sampled at 8000 Hz was subjected to 14th order LPC analysis based on the autocorrelation method every 20ms and converted back to LSF. The 0 to π distribution of the LSF corresponds to 0 to 4000 Hz. The Korean speech of the NATC database used as learning data in this experiment is composed as follows.

1) 4 males and 4 females = (8).

2) Each person pronounces 12 different sentences in 8 seconds.

3) Clean, babble noised, car noised speech environment for each sentence.

(= 8 persons * 12 sentences * 8 seconds * 3 environments = 2304 seconds)

(1) Design of Classifier

The right by using the A-th _{LSF ω A B (B = A} + 1 or A-1) is an actual example of a classifier for classifying the second LSF ω _B with L = 3 different intervals.

A B l = 1 {lower limit, upper limit} l = 2 {lower limit, upper limit} l = 3 {lower limit, upper limit} 2 One {0.f, 0.323399f} {0.323399f, 0.393569f} {0.393569f, 4.f} 3 2 {0.f, 0.492342f} {0.492342f, 0.597209f} {0.597209f, 4.f} 3 {0.f, 4.f} {0.f, 4.f} {0.f, 4.f} 3 4 {0.f, 0.492342f} {0.492342f, 0.597209f} {0.597209f, 4.f} 4 5 {0.f, 0.670035f} {0.670035f, 0.740385f} {0.740385f, 4.f} 5 6 {0.f, 0.901243f} {0.901243f, 0.985694f} {0.985694f, 4.f} 6 7 {0.f, 1.089864f} {1.089864f, 1.182087f} {1.182087f, 4.f} 7 8 {0.f, 1.338076f} {1.338076f, 1.419179f} {1.419179f, 4.f} 8 9 {0.f, 1.564212f} {1.564212f, 1.632792f} {1.632792f, 4.f} 9 10 {0.f, 1.783253f} {1.783253f, 1.835503f} {1.835503f, 4.f} 10 11 {0.f, 1.943743f} {1.943743f, 2.006788f} {2.006788f, 4.f} 11 12 {0.f, 2.155441f} {2.155441f, 2.225683f} {2.225683f, 4.f} 12 13 {0.f, 2.359377f} {2.359377f, 2.422795f} {2.422795f, 4.f} 13 14 {0.f, 2.591981f} {2.591981f, 2.642309f} {2.642309f, 4.f}

The classifier boundary value for L = 6 is as follows.

{0.f, 0.274970f}, {0.274970f, 0.323399f}, {0.323399f, 0.355183f}, {0.355183f, 0.393569f}, {0.393569f, 0.456871f}, {0.456871f, 4.f} }, / * B = 1 * /

0.547863f, 0.597209f}, {0.597209f, 0.638697f}, {0.638697f, 4.f}, {0.f, 0.418122f}, {0.418122f, 0.492342f}, {0.492342f, 0.547863f} }, / * B = 2 * /

0.f, 0.f}, {0.f, 0.f}, {0.f, 0.f}, {0.f, , {0.f, 4.f}}, / * B = 3 * /

0.547863f, 0.597209f}, {0.597209f, 0.638697f}, {0.638697f, 4.f}, {0.f, 0.418122f}, {0.418122f, 0.492342f}, {0.492342f, 0.547863f} }, / * B = 4 * /

{0.f, 0.594666f}, {0.594666f, 0.670035f}, {0.670035f, 0.711242f}, {0.711242f, 0.740385f}, {0.740385f, 0.796516f}, {0.796516f, 4.f} }, / * B = 5 * /

{0.f, 0.853770f}, {0.853770f, 0.901243f}, {0.901243f, 0.939014f}, {0.939014f, 0.985694f}, {0.985694f, 1.054606f}, {1.054606f, 4.f} }, / * B = 6 * /

{1.0f1015f, 1.089864f}, {1.089864f, 1.122802f}, {1.122802f, 1.182087f}, {1.182087f, 1.243816f}, {1.243816f, 4.f} }, / * B = 7 * /

1.377647f, 1.377647f, 1.419179f}, {1.419179f, 1.473773f}, {1.473773f, 4.f}}, {1.2f6068f, 1.338076f}, {1.338076f, }, / * B = 8 * /

{1.50415f, 1.564212f}, {1.564212f, 1.600643f}, {1.600643f, 1.632792f}, {1.632792f, 1.674107f}, {1.674107f, 4.f} }, / * B = 9 * /

1.781556f, 1.783253f}, {1.783253f, 1.809551f}, {1.809551f, 1.835503f}, {1.835503f, 1.876451f}, {1.876451f, 4.f} }, / * B = 10 * /

1.912272f, 1.943743f}, {1.943743f, 1.970872f}, {1.970872f, 2.006788f}, {2.006788f, 2.054019f}, {2.054019f, 4.f} }, / * B = 11 * /

{2.10444f, 2.187412f}, {2.187412f, 2.225683f}, {2.225683f, 2.270792f}, {2.270792f, 4.f} }, / * B = 12 * /

2.38947f, 2.38937f, 2.389477f, 2.389847f, {2.389847f, 2.422795f}, {2.422795f, 2.461599f}, {2.461599f, 4.f} }, / * B = 13 * /

{2.5f961f, 2.591981f}, {2.591981f, 2.617327f}, {2.617327f, 2.641309f}, {2.641309f, 2.672604f}, {2.672604f, 4.f} }, / * B = 14 * /

The value of the classifier boundary for L = 15 is as follows.

0.235044f, 0.262255f, {0.262255f, 0.286506f}, {0.286506f, 0.307375f}, {0.307375f, 0.323399f}, {0.323399f, 0.337399f}, { {0.337399f, 0.349574f}, {0.349574f, 0.361919f}, {0.361919f, 0.376771f}, {0.376771f, 0.393569f}, {0.393569f, 0.417651f}, {0.417651f, 0.446975f}, {0.446975 f, 0.463934f}, {0.463934f, 0.479502f}, {0.479502f, 4.f}}, / * B = 1 * /

{0.f, 0.373994f}, {0.373994f, 0.404027f}, {0.404027f, 0.433275f}, {0.433275f, 0.465560f}, {0.465560f, 0.492342f}, {0.492342f, 0.516309f} {0.516309f, 0.537683f}, {0.537683f, 0.557508f}, {0.557508f, 0.578283f}, {0.578283f, 0.597209f}, {0.597209f, 0.613344f}, {0.613344f, 0.630257f}, {0.630257 f, 0.647859f}, {0.647859f, 0.676275f}, {0.676275f, 4.f}}, / * B = 2 * /

0.f, 0.f}, {0.f, 0.f}, {0.f, 0.f}, {0.f, , {0.f, 0.f}, {0.f, 0.f}, {0.f, 0.f} , {0.f, 0.f}, {0.f, 0.f}, {0.f, 0.f} }, / * B = 3 * /

{0.f, 0.373994f}, {0.373994f, 0.404027f}, {0.404027f, 0.433275f}, {0.433275f, 0.465560f}, {0.465560f, 0.492342f}, {0.492342f, 0.516309f} {0.516309f, 0.537683f}, {0.537683f, 0.557508f}, {0.557508f, 0.578283f}, {0.578283f, 0.597209f}, {0.597209f, 0.613344f}, {0.613344f, 0.630257f}, {0.630257 f, 0.647859f}, {0.647859f, 0.676275f}, {0.676275f, 4.f}}, / * B = 4 * /

{0.f, 0.510562f}, {0.510562f, 0.571390f}, {0.571390f, 0.614788f}, {0.614788f, 0.646839f}, {0.646839f, 0.670035f}, {0.670035f, 0.689459f} {0.689459f, 0.705029f}, {0.705029f, 0.716975f}, {0.716975f, 0.728454f}, {0.728454f, 0.740385f}, {0.740385f, 0.757645f}, {0.757645f, 0.781948f}, {0.781948 0.811476f}, {0.811476f, 0.849480f}, {0.849480f, 4.f}}, / * B = 5 * /

0.838568f, 0.866646f, {0.866646f, 0.885520f}, {0.885520f, 0.901243f}, {0.901243f, 0.916409f}, { {0.916409f, 0.931262f}, {0.931262f, 0.946566f}, {0.946566f, 0.964604f}, {0.964604f, 0.985694f}, {0.985694f, 1.013291f}, {1.013291f, 1.040804f}, {1.040804 f, 1.069675f}, {1.069675f, 1.109874f}, {1.109874f, 4.f}}, / * B = 6 * /

1.038742f, 1.087242f, {1.060728f, 1.076330f}, {1.076330f, 1.089864f}, {1.089864f, 1.101254f}, { {1.101254f, 1.114450f}, {1.114450f, 1.132790f}, {1.132790f, 1.155356f}, {1.155356f, 1.182087f}, {1.182087f, 1.207519f}, {1.207519f, 1.231719f}, {1.231719 f, 1.257682f}, {1.257682f, 1.303627f}, {1.303627f, 4.f}}, / * B = 7 * /

1.224180f, 1.270616f, {1.270616f, 1.299337f}, {1.299337f, 1.320930f}, {1.320930f, 1.338076f}, {1.338076f, 1.354144f}, { {1.354144f, 1.369426f}, {1.369426f, 1.385734f}, {1.385734f, 1.401858f}, {1.401858f, 1.419179f}, {1.419179f, 1.438471f}, {1.438471f, 1.460843f}, {1.460843 f, 1.489174f}, {1.489174f, 1.539309f}, {1.539309f, 4.f}}, / * B = 8 * /

{1.503346f, 1.546492f}, {1.546492f, 1.564212f}, {1.564212f, 1.579060f}, { {1.579060f, 1.593888f}, {1.593888f, 1.607045f}, {1.607045f, 1.619846f}, {1.619846f, 1.632792f}, {1.632792f, 1.646761f}, {1.646761f, 1.663350f}, {1.663350 f, 1.687100f}, {1.687100f, 1.733830f}, {1.733830f, 4.f}}, / * B = 9 * /

1.722448f, 1.752468f, 1.752468f, 1.769501f, {1.769501f, 1.783253f}, {1.783253f, 1.793231f}, { {1.793231f, 1.804241f}, {1.804241f, 1.814543f}, {1.814543f, 1.824683f}, {1.824683f, 1.835503f}, {1.835503f, 1.848488f}, {1.848488f, 1.865541f}, {1.865541 1.889830f}, {1.889830f, 1.932878f}, {1.932878f, 4.f}}, / * B = 10 * /

{1.919255f, 1.932970f}, {1.932970f, 1.943743f}, {1.943743f, 1.954767f}, { {1.954767f, 1.965228f}, {1.965228f, 1.977102f}, {1.977102f, 1.990802f}, {1.990802f, 2.006788f}, {2.006788f, 2.024141f}, {2.024141f, 2.043363f}, {2.043363 2.066415f}, {2.066415f, 2.098800f}, {2.098800f, 4.f}}, / * B = 11 * /

2.10612f, 2.126233f, {2.126233f, 2.141440f}, {2.141440f, 2.155441f}, {2.155441f, 2.169336f}, { {2.169336f, 2.181376f}, {2.181376f, 2.194262f}, {2.194262f, 2.209642f}, {2.209642f, 2.225683f}, {2.225683f, 2.242341f}, {2.242341f, 2.260786f}, {2.260786 f, 2.281577f}, {2.281577f, 2.311329f}, {2.311329f, 4.f}}, / * B = 12 * /

{2.3f140f, 2.359377f}, {2.359377f, 2.371829f}, {2.329068f, 2.345140f}, { {2.371829f, 2.383751f}, {2.383751f, 2.396125f}, {2.396125f, 2.408754f}, {2.408754f, 2.422795f}, {2.422795f, 2.437442f}, {2.437442f, 2.453004f}, {2.453004 f, 2.471189f}, {2.471189f, 2.500069f}, {2.500069f, 4.f}}, / * B = 13 * /

2.565426f, 2.580842f}, {2.580842f, 2.591981f}, {2.591981f, 2.602545f}, {2.505818f, 2.544425f}, { {2.602545f, 2.612613f}, {2.612613f, 2.622403f}, {2.622403f, 2.632054f}, {2.632054f, 2.641309f}, {2.641309f, 2.651491f}, {2.651491f, 2.664897f}, {2.664897 2.681739f}, {2.681739f, 2.709239f}, {2.709239f, 4.f}}, / * B = 14 * /

(2) LSQ learning method

The learning method is as follows.

1) The third LSF ω ₃ of the 14th LSF is learned by the LBG method.

2) ω is passed through the _third value of the classifier L (L = 3 or 6 or 15) gaejung and include one of the fourth LSF ω ₄ in the training set, each of the training set ω ₄ learns the LBG method.

3) (N + K) When learning a second LSF ω _{N + K} (However, (N + K) is less than or equal to 14, an integer exceeding N) (N + K-1 ) th LSF ω _{N + K- 1} is used.

4) (N-K + 1) th LSF ω _{N-K + 1} when learning the (NK) th LSF ω _NK (NK is an integer less than N and equal to or greater than 1).

(4) Encoding and decoding method

The third LSF? ₃ of the 14th LSF is encoded / decoded by the SQ scheme, and the remaining LSFs select one of the L (L = 3 or 6 or 15) codebooks to be encoded / decoded using a classifier. At this time, since the input of the classifier uses quantized values differently from the learning process, additional additional bit transmission is not required.

(5) Experimental results

For fair evaluation, test voice used English voice of NATC database as follows. (Learning data is NATC's Korean voice)

1) 4 males and 4 females = (8)

2) Each person pronounces 12 different sentences in 8 seconds.

3) For each sentence, clean, babble noise 30dB, car noise 30dB, car noise 15dB

(= 8 persons * 12 sentences * 8 seconds * 4 environments = 3072 seconds)

Spectral distortion (SD) measurement was used for the performance evaluation. The SD of the i-th frame is as follows,

Where P _j represents the power spectrum of the original LSF (power spectrum of the original LSF), Represents the power spectrum of the quantized LSF. In accordance with the characteristics of the human ear, a was chosen to be 125 Hz and b was chosen to be 3400 Hz.

Performance evaluation table (SD) of quantizer Various environment-specific SD quantizer types clean Bubble 30dB 30dB for cars 15dB for car Sum Average Distortion (dB) 2dB ~ 4dB (%) 4dB to (%) Average Distortion (dB) 2dB ~ 4dB (%) 4dB to (%) Average Distortion (dB) 2dB ~ 4dB (%) 4dB to (%) Average Distortion (dB) 2dB ~ 4dB (%) 4dB to (%) Average Distortion (dB) SQ ^* 1.56 13.9 2.33 1.34 8.95 1.49 1.43 9.52 1.51 1.20 4.77 0.42 1.41 SVQ ^** 1.20 6.30 0.47 1.06 3.97 0.11 1.09 4.12 0.11 0.98 1.52 0.00 1.08 LSQ ^*** : 3 links 1.13 6.52 0.79 1.01 4.65 0.36 1.04 4.61 0.38 0.94 2.04 0.01 1.03 LSQ ^*** : 6 links 1.04 5.34 0.04 0.94 3.93 0.15 0.96 3.94 0.16 0.87 1.52 0.00 0.95 LSQ ^*** : 15 links 0.99 5.29 0.16 0.91 3.87 0.07 0.94 3.78 0.07 0.86 1.57 0.00 0.93

* SQ: Scalar quantizer

** SVQ: Split vector quantizer (14th divided into 2-2-2-2-2-2-2)

*** LSQ: Link Scalar Quantizer: 3 or 6 or 15 Link

Quantizer Performance Table (Memory and Search Time Complexity) SQ * Memory 2 ⁴ + 7 × 2 ³ + 5 × 2 ² + 2 ¹ = 94 bytes time 0.038 MIPS SVQ ** Memory 6 x 2 x 2 ⁵ + 2 x 2 ⁶ = 512 bytes time 0.256 MIPS LSQ ***: 3 links Memory 2 ⁴ + 7 × 3 × 2 ³ + 5 × 3 × 2 ² + 3 × 2 ¹ = 250 bytes time 0.038 MIPS LSQ ***: 6 links Memory 2 ⁴ + 7 × 6 × 2 ³ + 5 × 6 × 2 ² + 6 × 2 ¹ = 484 bytes time 0.038 MIPS LSQ ***: 15 links Memory 2 ⁴ + 7 × 15 × 2 ³ + 5 × 15 × 2 ² + 15 × 2 ¹ = 1186 bytes time 0.038 MIPS

Table 1 shows the performance of various quantizers by the SD method. As a result, it is seen that the performance is increased when the number of links of LSQ is increased to 3, 6, or 15. Also, LSQ is superior to SQ and SVQ in terms of Average Spectral Distortion. In addition to the SD performance, considering the memory and search time complexity in Table 2, we can see that SQ is superior in terms of memory, but SD performance is far behind, and SVQ can be used for both SD performance, memory, and search time complexity LSQ 6 link structure.

According to the present invention, by applying the present invention to the 4kbps voice coder proposed in ITU-T, Study Group 16, and Question Number 21, improved performance can be obtained while saving memory and bits compared to the conventional method.

Claims (8)

M-order line spectral frequencies {omega_1, omega_2, omega_3, ... , omega_M} in a quantizer, Codebooks to which a plurality of codebooks are allocated according to M levels; And The M-th order line spectrum frequencies {omega_1, omega_2, omega_3, ... , omega_M} are quantized, one codebook is selected based on the magnitude of the line spectrum frequency adjacent to the line spectrum frequency to be quantized among a plurality of codebooks allocated to each line spectrum frequency, and the quantization And a quantization unit for quantizing the line spectrum frequency to be quantized. M-order line spectral frequencies {omega_1, ... , omega_N, ... , omega_M} in a quantizer, Codebooks allocated to N-th order and codebooks assigned to L other than the N-th order (where 1? N? M); A first classifier to an M classifier for classifying the line spectrum frequency inputted for each order into L kinds according to the magnitude; And The reference spectrum frequency ωN is quantized by a codebook assigned to the N-th order, and for a line spectrum frequency higher than ωN, one of the L codebooks assigned to the order is assigned to a value classified by a low- Quantized by a codebook corresponding to a value classified by a class of one order higher than that of L codebooks allocated to the degree of linear spectrum frequency lower than the above-mentioned frequency N by a corresponding codebook, And a quantizer for quantizing the quantized signal. 3. The method of claim 2, wherein the line spectral frequencies input to the first to M &lt; th & Wherein each quantizer is a quantized linear spectral frequency of the same order as the classifier. Multiple M-order line spectrum frequencies {{omega_1}, ... , {Omega_N}, ... , {Omega_M}} for learning a connected scalar quantizer, (a) learning a plurality of reference line spectral frequencies {? N} to generate one codebook assigned to the N-th order (where 1? N? M); (b) While changing K from 0 to (M-N-1) (b.1) a small step of dividing the distribution space of a plurality of line spectrum frequencies {? N + K} into L; (b.2) Each line spectral frequency? N + K + 1 belonging to a plurality of line spectrum frequencies {? N + K + 1} is divided into L subsets corresponding to the distribution space in which the line spectrum frequency? ; And (b.3) repeating a small step of generating L codebooks allocated to (N + K + 1) -th order by learning line spectral frequencies divided into L subsets in the (b.2) Generating codebooks assigned to higher orders by performing; And (c) While changing K from 0 to (N-2) (c.1) a small step of dividing the distribution space of a plurality of line spectrum frequencies {? N-K} into L; (c.2) Each line spectral frequency? N-K-1 belonging to a plurality of line spectral frequencies {? N-K-1} is divided into L subsets corresponding to the distribution space to which the line spectral frequency? ; And (c.3) Repeating the small step of generating L codebooks allocated to the (NK-1) difference by learning each of the line spectrum frequencies divided into L subsets in the step (c.2) And generating codebooks assigned to the lower-order codebooks. M-order line spectral frequencies {omega_1, ... , omega_N, ... , omega_M}, comprising: (a) The value obtained by quantizing the reference line spectral frequency? N by the codebook allocated to the Nth order And obtaining a coded value iN (where 1? N? M); (b) While changing K from 0 to (M-N-1) (b.1) quantized line spectrum frequency Into a value of one of 1 to L according to its size; (b.2) a small step of finding a codebook corresponding to the value classified in the step (b.1) among the L codebooks allocated to the (N + K + 1) th order; And (b.3) a value obtained by quantizing the line spectrum frequency? N + K + 1 by the codebook found in the step (b.2) And a small step of obtaining a coded value iN + K + 1; And (c) While changing K from 0 to (N-2) (c.1) Quantized line spectrum frequency Into a value of one of 1 to L according to its size; (c.2) searching a codebook corresponding to a value classified in the (c.1) small step among L codebooks allocated to the (N-K-1) th order; And (c.3) a value obtained by quantizing the line spectrum frequency? N-K-1 by the codebook found in the (c.2) And encoding the lower-order line spectral frequencies by repeatedly performing a small step of obtaining a coded value iN-K-1. 6. The method of claim 5, wherein the quantized values in steps (a), (b.3), and (c.3) are quantized from values included in a corresponding codebook, And a Claryian distance scale is set to a smallest value. The method of claim 5, wherein the values encoded in steps (a), (b.3), and (c.3) are each an index of a corresponding codebook for a corresponding quantized value, A method of coding a line spectral frequency. The indexes of M-order line spectral frequencies {i_1, ... , i_N, ... , i_M}, comprising: (a) the quantized line spectrum frequency using the index iN of the reference line spectral frequency in the codebook assigned to the Nth order (Where 1 &lt; = N &lt; = M); (b) While changing K from 0 to (M-N-1) (b.1) quantized line spectrum frequency Into a value of one of 1 to L according to its size; (b.2) a small step of finding a codebook corresponding to the value classified in the step (b.1) among the L codebooks allocated to the (N + K + 1) th order; And (b.3) quantizing the line spectrum frequency using the index iN + K + 1 in the codebook found in the step (b.2) Decoding the indexes of the higher order line spectral frequencies by repeating a small step of finding the higher order line spectral frequencies; And (c) While changing K from 0 to (N-2) (c.1) Quantized line spectrum frequency Into a value of one of 1 to L according to its size; (c.2) searching a codebook corresponding to a value classified in the (c.1) small step among L codebooks allocated to the (N-K-1) th order; And (c.3) The quantized line spectrum frequency (iN-K-1) using the index iN-K-1 in the codebook found in the (c.2) And decoding the indexes of the lower order line spectral frequencies by repeating the step of finding the lower frequency spectral frequencies.