JPWO1994007239A1 - Audio encoding method and device - Google Patents

Abstract

(57) [Abstract] This publication contains application data prior to electronic filing, so abstract data is not recorded.

Description

[Detailed Description of the Invention] Audio Coding Method and Apparatus Technical Field The present invention relates to an audio coding method and apparatus for compressing audio signal information, and more particularly to an audio coding method and apparatus using analysis-by-three-dimensional (A-b-S) vector quantization for encoding at transmission rates of 4 to 16 Kbps.

BACKGROUND ART Speech coders using A-b-5 vector quantization, such as Code-Excited Linear Prediction (CELP) coders, have recently been seen as promising speech coders for compressing speech signals while maintaining their quality in corporate communication systems, digital mobile radio systems, and other applications. These vector quantization speech coders (hereafter simply referred to as coders) apply predictive weighting to each code vector in a codebook to create a reproduced signal, evaluate the error power between the reproduced signal and the input speech signal, and determine the index of the code vector with the least error, which is then transmitted to the receiver.

The A-b-3 vector quantization encoder performs a linear predictive synthesis filter process on each of the approximately 1,000 sound source signal vector patterns stored in the codebook, and then searches among those approximately 1,000 patterns for the pattern that minimizes the error between each reproduced sound signal and the input sound signal to be encoded.

However, since the encoder is required to respond to calls immediately, the above search process must be performed in real time. This means that the search process must be performed continuously throughout the call, at short intervals of, for example, 5 ms.

However, as will be explained later, this search process involves complex calculations such as filter and correlation operations, and the amount of computation required for these operations can be enormous, reaching several hundred Mops (Mop. 93 fps). Even with the fastest digital signal processors (DSPs) currently available, several chips are required to handle this, which poses challenges in miniaturization and low power consumption when applied to mobile phones, for example.

As a speech coding method that solves the above problems, the applicant proposed in Japanese Patent Application No. 3-127669 (JP Patent Publication No. 4-352200) a tree-structured delta codebook. This codebook stores delta vectors, which are the differences between signal vectors, instead of storing the code vectors themselves as in the conventional method. The codebook then generates tree-structured code vectors by sequentially adding and subtracting these delta vectors.

This method significantly reduces the memory capacity required for the codebook, and also significantly reduces the amount of computation required, since the filter and correlation operations for each code vector are performed by performing filter and correlation operations on the delta vector and then sequentially adding and subtracting the results.

However, in this method, each code vector is generated as a linear combination of a smaller number of basis vectors, i.e., delta vectors, and therefore contain no components other than the delta vectors. In other words, the code vectors can only be distributed in a subspace with dimensions corresponding to the number of delta vectors (usually 8-10) within the space in which the vectors to be encoded are distributed (usually 40-64 dimensions).

Therefore, even if the basis vectors (delta vectors) are designed based on the statistical distribution of the speech signal to be coded, the tree-structured delta codebook suffers from the problem of degraded quantization characteristics compared to conventional codebooks without structural constraints.

Therefore, in Japanese Patent Application No. 3-515016, the applicant of the present application noted that when the above-mentioned linear predictive synthesis filter operation is performed on a code vector to evaluate distance, all delta vector components are amplified unevenly rather than uniformly, and that the contribution of each delta vector to a code vector in a tree-structured delta codebook can be changed by changing the order of the delta vectors. Therefore, the applicant proposed a method for improving performance by performing a filter operation on each delta vector C each time the coefficients of the linear predictive synthesis filter are determined, comparing the power (vector length), and then sorting the delta vectors in descending order of power before encoding using a tree-structured delta codebook.

However, this method is no different from the conventional method in that it generates code vectors from only a limited number of delta vectors, so there is a limit to the improvement in performance, and further improvement is desired.

Another challenge for speech coders using A-b-5 vector quantization is the realization of variable bit-rate coding. Variable bit-rate coding is a coding method that allows the bit rate to be changed appropriately depending on the transmission channel availability, the importance of the sound source, and other conditions, thereby achieving overall coding efficiency.

When vector quantization is used for variable-rate speech coding, it is necessary to prepare a codebook with a number of patterns for each transmission rate and to switch between them according to the desired transmission rate.

In this case, in the case of a conventional codebook that simply lists code vectors, storing each codebook requires NX × M words of memory, which corresponds to the product of (vector dimension 2N) and (patterns fi:M). Here, the number of patterns M is proportional to a power of 2 of the number of bits in the code vector index, so increasing the variable range of the transmission rate or controlling the transmission rate in fine increments poses the problem of requiring a huge amount of memory.

Furthermore, in variable-rate transmission, the transmission rate of the encoded signal may need to be reduced at the request of the transmission network. In such cases, the decoder must reconstruct the audio signal from the bit-dropped information generated by the encoder.

Scalar quantization, which is less efficient than vector quantization, has traditionally dealt with focus drop by controlling the focus to drop starting with the least significant bit (LSB) or by configuring a high-rate quantizer to encompass the quantization levels of a low-rate quantizer (embedded coding).

However, in the case of vector quantization, which uses a conventional codebook that simply lists code vectors, the codebook itself is not structured in any way, so there is no difference in the importance of the bits in the code vector index (dropping the LSB or MSB results in a completely different vector being called). This means that the same countermeasures as those used in scalar quantization cannot be applied, resulting in significant degradation of sound quality due to bit dropping.

DISCLOSURE OF THE INVENTION It is therefore a first object of the present invention to provide a method and apparatus for speech coding using a tree-structured data codebook that is an improvement over the previously described methods.

A second object of the present invention is to provide a speech coding method and apparatus using vector quantization that does not require a large amount of memory for a codebook and that can deal with bit and digital artifacts.

According to the present invention, there is provided a speech coding method for encoding an input speech signal vector using an index assigned to a code vector having a minimum distance from the input speech signal vector among predetermined code vectors, the speech coding method comprising the steps of: (a) storing a plurality of differential code vectors; (b) multiplying each differential code vector by a linear prediction synthesis filter matrix; (c) evaluating the power amplification factors of the differential code vectors multiplied by the matrix; (d) sorting the differential code vectors multiplied by the matrix in order of the magnitude of the evaluated power amplification factors; (e) selecting a predetermined number of vectors from the sorted vectors in order of the magnitude of the evaluated power amplification factors; (f) evaluating the distance between the input speech signal vector and a linear prediction synthesis filter-processed code vector to be generated by sequentially adding and subtracting the selected vectors in a tree structure; and (g) determining the code vector having the minimum evaluated distance.

The present invention also provides a speech coding device for encoding a cough input speech signal vector using an index assigned to a code vector among predefined code vectors that has the smallest distance from the input speech signal vector, the device comprising: means for storing a plurality of differential code vectors; means for multiplying each differential code vector by a linear predictive synthesis filter matrix; means for evaluating the power amplification factors of the differential code vectors multiplied by the cough matrix; means for sorting the differential code vectors multiplied by the matrix in order of the magnitude of the evaluated power amplification factors; means for selecting a predetermined number of vectors from the sorted vectors in order of the magnitude of the evaluated power amplification factors; means for evaluating the distance between the input speech signal vector and a linear predictive synthesis filter-processed code vector to be generated by sequentially adding and subtracting the selected vectors in a tree structure; and means for determining the code vector with the smallest evaluated distance.

The present invention also provides a variable-length speech coding method for variable-length coding an input speech signal vector using a variable-bit-length code assigned to a code vector among predefined code vectors that has the smallest distance from the input speech signal vector, the variable-length speech coding method comprising the steps of: (a) storing ≠X number of differential code vectors; (b) evaluating the distance between the input speech signal vector and a code vector to be generated by sequentially adding and subtracting, from the beginning, a number of differential code vectors corresponding to a desired code bit length in a tree structure; (c) determining the code vector with the smallest evaluated distance; and (d) determining a code of the desired code bit length to be assigned to the determined code vector.

The present invention also provides a variable-length speech coding method for variable-length coding an input cough speech signal vector using a variable-bit-length code assigned to a code vector among predefined code vectors that has the smallest distance from the input speech signal vector. The variable-length speech coding device includes: means for storing multiple differential code vectors; means for evaluating the distance between the input speech signal vector and a code vector to be generated by sequentially adding and subtracting, from the beginning, a number of differential code vectors corresponding to a desired code bit length in a tree structure; means for determining the code vector with the smallest evaluated distance; and means for determining a code of the desired code bit length to be assigned to the determined code vector.

Brief Description of the Drawings Figure 11 is a block diagram showing the concept of a speech generation system; Figure 2 is a block diagram illustrating the principles of general CELP speech coding; Figure 3 is a block diagram showing the configuration of the noise codebook search process for A-b-3 type vector quantization as prior art; Figure 4 is a block diagram modeling the noise codebook search process algorithm.

Figure 5 is a block diagram illustrating the principles of the delta codebook; Figures 6A and 6B are diagrams illustrating the adaptation method for a tree-structured delta codebook; Figures τA, τB, and τC are diagrams illustrating the details of the present invention; Figure 8 is a diagram of the speech coder of the present invention; and Figures 9A and 9B are diagrams illustrating the variable rate coding method of the present invention.

BEST MODE FOR CARRYING OUT THE INVENTION Speech can be divided into voiced and unvoiced sounds. Voiced sounds are generated by a pulse source produced by vocal cord vibration, and are then transformed into voice by the addition of individual vocal tract characteristics of the throat and mouth. Unvoiced sounds are produced without vocal cord vibration; a simple Gaussian noise source is the source, passing through the vocal tract and becoming voice. Therefore, as shown in Figure 1, the speech generation mechanism can be modeled using a pulse sound (5psc) that is the source of voiced sounds, a noise source (NSC) that is the source of unvoiced sounds, and a linear predictive synthesis filter (LPCF) that adds vocal tract characteristics to the signals output from each sound source. Human voices have pitch periodicity, which corresponds to the periodicity of the pulses output from the pulse source and varies depending on the person and the content of speech.

Therefore, if we can identify the period of the pulse source and the noise sequence corresponding to the input speech, we can encode the input speech using a code (index) that identifies these pulse periods and the noise sequence.

As shown in FIG. 2, vector P obtained by delaying past values (bP + gC) by different sample numbers is stored in adaptive codebook 11. Vector P in adaptive codebook 11 is multiplied by gain b to obtain vector bp, which is input to linear predictive synthesis filter 12 for filter processing. The resulting filter processing result bAP is subtracted from input speech signal X, and error power evaluation unit 13 selects vector P from adaptive codebook 11 that minimizes the difference power from the resulting error signal, thereby determining the period.

Subsequently, or simultaneously, multiple noise sequences (each represented by an N-dimensional code vector) are prepared in the noise codebook 1, and each code vector C is multiplied by a gain g and processed by the linear predictive synthesis filter 3 to obtain a reproduced signal vector gAC. The error vector evaluator 5 then determines the code vector that minimizes the error between the reproduced signal vector gAC and the input signal vector X (an N-dimensional vector). Once this is determined, speech can be encoded using the period and data (index) that specifies the code vector. Note that the example described above with reference to FIG. 2 is an example in which vector AC and vector AP are orthogonal. If this is not the case, a code vector that minimizes the error between vector X - bAP, which is obtained by subtracting vector bAP from input signal vector X, is determined.

Figure 3 is a block diagram of a speech transmission (encoding) system using vector quantization based on the A-b-3 method, corresponding to the lower half of Figure 2. Specifically, reference numeral 1 denotes a random codebook that stores an N-dimensional code vector C of size M; 2 denotes an amplifier with gain g; 3 denotes a linear predictive synthesis filter that has coefficients determined by linear predictive analysis of the input signal X and performs linear predictive filtering on the output of amplifier 2; 4 denotes an error generator that outputs the error between the reproduced signal vector output from linear predictive synthesis filter 3 and the input signal vector; and 5 denotes an error power evaluator that evaluates the error and selects the code vector that minimizes the error.

Unlike conventional vector quantization, this A-b-3 quantization method multiplies each code vector (C) in the random code I (11) by an optimal gain (g), then filters it through a linear predictive synthesis filter (3). The error signal (E) between the reproduced signal vector (gAC) obtained through this filtering and the input signal vector (X) is generated by an error generator (4). The error power evaluator (5) then searches the random codebook using the power of the error signal as an evaluation function (distance measure) to find the code vector with the smallest error power. The input signal is then encoded using a code (index) that identifies that code vector and transmitted.

The error power in this case is given by the following equation: Equation (1) The optimal code vector and gain g are determined to minimize the error power shown in Equation (1). Note that power varies depending on the voice volume, so the gain g is optimized to match the playback signal power to the input signal power. The optimal gain can be determined by partially differentiating Equation (1) with respect to g and substituting O. That is, dlEl"/dg = Q, so g is given by g = (X" AC)/((AC)"(AC)) (2). Substituting this g into equation (1), we obtain IEI"-IXI"-(X" AC)!/((AC)'(AC)) (3). If the cross-correlation between the input signal X and the output AC of the linear prediction synthesis filter 3 is RXC1 and the auto-correlation of the output AC of the linear prediction synthesis filter 3 is RC, then the cross-correlation and auto-correlation are expressed by the following equations: Equation (4)

The code vector C that minimizes the error power of equation (3) maximizes the second term on the right-hand side of equation (3). Therefore, the code vector C can be expressed as follows: %Equation%(6) The optimal gain is given by g"Rxc/Rcc (7) from equation (2) using the cross-correlation and auto-correlation that satisfy equation (6).

Figure 4 is a block diagram modeling the noise codebook search algorithm that, based on the above equations, selects the code vector that minimizes the error power and encodes the input signal. It includes a calculation unit 6 that calculates the cross-correlation Rxc ("xAC"), a calculation unit 7 that calculates the square of this cross-correlation RXc, a calculation unit 8 that calculates the autocorrelation RCC of AC, a calculation unit 9 that calculates Rxc"/Rcc, and an error power evaluation unit 5 that determines the code vector that maximizes RXc"/Rcc, in other words, minimizes the error power, and outputs a code that identifies that code vector. However, this is equivalent to Figure 3.

The main components of this conventional codebook search process are: (1) filtering of the code vector C, (2) calculation of the cross-correlation RXC, and (3) calculation of the auto-correlation RCC. If the order of the LPC filter 3 is NP and the dimension of the vector quantization (code vector) is N, then the computational complexity required for (1) through (2) for one code vector is NP·N, N, and N, respectively. Therefore, the computational complexity required for a codebook search per code vector is (Np = 2)·N.

A commonly used noise codebook l has approximately 40 dimensions and a codebook size of 1024 (N = 40.M - 1024). Since the analysis order N of the LPG filter 3 is approximately 10th order, one codebook search requires (10 + 2) × 40 × 1024 = 480 × 10' multiplication and addition operations.

To perform this complex codebook search every subframe (5m5ec) of speech coding, a massive processing power of 96 ops (3 mega-operators per second) would be required. Even with the fastest digital processors available today (capacity of 20-40 MOPs), real-time implementation would require several chips.

Furthermore, to store and maintain such a noise codebook 1 as a table, a memory capacity of N × M (= 40 × 1024 = 1024 = 40K would be required.

Furthermore, in the case of mobile phones and cellular phones, which are considered to be potential applications of speech coders using A-b-5 vector quantization, miniaturization and low power consumption are essential requirements. The enormous computational complexity and memory requirements pose significant obstacles to the realization of speech coders.

Based on the above, the applicant of the present application proposed in Japanese Patent Application No. 3-127669 (JP Patent Publication No. 4-352200) the use of a tree-structured delta codebook, as shown in FIG. 5, instead of a conventional stochastic codebook, in order to provide a speech coding method that can reduce the amount of calculation required for stochastic codebook search and also reduce the memory capacity required for storing the stochastic codebook.

In FIG. 5 , an initial vector C0 (= Δ) which is a reference noise sequence and delta vectors Δ1 to ΔL-1 (L = 10) which are (L-1) types (layers) of delta noise sequences are stored in the delta codebook 10 in advance. By adding and subtracting delta vectors Δ1 to ΔL-1 to the initial vector C0 at each layer according to the tree structure, it is possible to sequentially represent code vectors (codewords) C0 to C5°2 of (2-1) types of noise sequences on the tree structure. Alternatively, a zero vector (or a zero vector) can be added to these code vectors to represent code vectors (codewords) C0 to C1°1 of 2-11 noise sequences.

In this way, simply storing the initial vector Δ and (L-1) delta vectors Δ1 to ΔL-1 (L = 10) in the delta codebook 10 allows for the sequential generation of 2L-1 (= 2′-1 = M-1) recode vectors or 2L (= 2L = M) code vectors. This reduces the storage capacity of the delta codebook 10 to L-N (= 10·N), a significant reduction compared to the storage capacity of conventional random codebooks, M-N (= 1024·N).

Using this tree-structured delta codebook IO, the cross-correlation R, c'I+ and autocorrelation R, c'I+ for the code vector Cr (j = 0 to 1022 or 1023) can be expressed by the following recursive formula. That is, when each code vector is expressed as (zk, , w C, -Δ, i -1, 2 - L -1 (8) or Ctw - z Ck -Δ, 2; -1 ≤ k < 2 -1 (9), R, clZk-11 - Rxc + l + l - I - X (AΔi) (10) or R, clZk-H - Rxc N+) X T (AΔ,) (11) and Rec Itk-11 - Rcc tkl + (AΔ,) T (AΔ,) + 2 ( AΔi)' (Ach) (12) or Rcc +! に一! ! l x Rcc + + (AΔ,)' (AΔi) - 2 (AΔi)" (ACk) (13).

Therefore, with regard to the cross-correlation Rxc, once the cross-correlation XT(AΔ,) for each delta vector Δ (i = 0 to L-1; Δ, -C,) is calculated, the cross-correlation R for all code vectors C1 can be immediately calculated by sequentially adding or subtracting these according to the recurrence formulas (10) and (11), i.e., the tree structure of Figure 5. In conventional codebooks, it takes many iterations to calculate the cross-correlation for the code vectors of all noise sequences. Previously, -N (-1024 × N) multiplication and addition operations were required. In contrast, with a tree-structured delta codebook, the cross-correlation R,efj+ is not calculated directly from each code vector C,(j-0,1-2L-1), but is calculated by calculating the cross-correlation with each delta vector Δj(j-0,1,...L-1) and then sequentially adding or subtracting them. This reduces the number of multiplication and addition operations to only L-N (-10 × N), significantly reducing the number of calculations.

Furthermore, the cross term (^Δi)"(AC- in the third term of the autocorrelation equations (12) and (13) can be expressed as Cy = Δ.±Δ1±Δ2...±Δi-1. (AΔi)"(ACk) = (AΔ,)"(AΔ.)±(AΔ,)T(AΔl) =...±(AΔ,)"(AΔi-,) (14). Therefore, the cross-correlation between Δ and Δ.,Δ1...Δi-1 (AΔ,)T(^ Δ. + I + 2", 1-1) and sequentially add or subtract according to the tree structure of Figure 5 to calculate the third term. Furthermore, by calculating the autocorrelation (A Δi)' (AΔ8) of each delta vector Δ3 in the second term and sequentially adding or subtracting it according to equations (12) and (13), i.e., according to the tree structure of Figure 5, the autocorrelations R, c, j+ of all code vectors CJ can be immediately calculated.

That is, with conventional codebooks, calculating the autocorrelation required M-N (= 1024·N) product-sum calculations. In contrast, with tree-structured delta codebooks, the autocorrelation (Rxc+J) is not calculated directly from each code vector CJ (j-0, 1-2L-'-1), but is calculated from the autocorrelation of each delta vector Δj (j = 0, 1...L-1) and the cross-correlation between all combinations of different delta vectors. This reduces the calculation time to just L(LSI)·N/2 (= 55·N) product-sum calculations, significantly reducing the number of calculations.

However, because all codewords (code vectors) in such a tree-structured delta codebook are generated as linear combinations of delta vectors, they contain no components other than delta vectors. In other words, within the space in which the vectors to be coded are distributed (typically 40-64 dimensions), the code vectors can only be distributed in a subspace with dimensions corresponding at most to the number of delta vectors (typically 8-10).

Therefore, even if the basis vectors (delta vectors) are designed based on the statistical distribution of the speech signal to be coded, the tree-structured delta codebook suffers from the problem of degraded quantization characteristics compared to conventional codebooks without structural constraints.

As described above, in the CELP speech coder to which the present invention is applied, vector quantization differs from conventional vector quantization in that it performs distance evaluation in a space of signal vectors obtained by applying a linear predictive synthesis filter having a filter transfer function Az to a code vector, thereby determining the optimal vector.

Therefore, as shown in FIGS. 6A and 6B, the linear predictive synthesis filter transforms the residual signal space (the sphere in FIG. 6A when L=3) into the reproduced signal space. However, during this process, the directional components along each axis are generally not uniform, as shown in FIG. 6B, resulting in distortion-inducing amplification.

That is, the linear prediction synthesis filter characteristic (A) exhibits different amplitude amplification characteristics for each delta vector that constitutes the codebook, and the resulting vectors are not evenly distributed across the entire space.

Furthermore, in the tree-structured delta codebook shown in Figure 5, the contribution of each delta vector to a code vector varies depending on the delta vector's position in the delta codebook 10. For example, the second delta vector ΔI contributes to all code vectors in the second Vk layer and below, while the third delta vector Δ2 contributes to all code vectors in the third layer and below, and Δ contributes only to code vectors in the tenth layer. In other words, by changing the order of the delta vectors, the contribution of each delta vector to a code vector can be changed.

Based on the above, in Japanese Patent Application No. 3-515016, the applicant calculated the power (amplification factor: if each delta vector is normalized, the power of AΔ itself becomes the amplification factor) I AΔ, :2 = (AΔi)'(AΔ) for each delta vector Δ, vector AΔ with unifilter characteristics (A), and S, and compared them. Then, the delta vectors were rearranged in descending order of power. This resulted in improved performance compared to conventional tree-structured delta codebooks, which have a biased distribution due to the fixed delta vectors.

However, even in this case, the number of delta vectors is the same as the number actually used, and encoding is performed using delta vectors that are rearranged from among them, so the flexibility of the codebook is limited.

For example, to simplify the discussion, consider the case of L = 2, i.e., a tree-structured delta codebook that generates code vectors C0C1 (= Δ0 + Δ1) and C1 (= Δ1 - Δ1) from vector C1 (= Δ0) and delta vector Δ1. As shown in Figure 7A, if the vectors used as Δ1, Δ2 are limited to unit vectors e and ey, then the generated code vectors are limited to the x-y plane indicated by the diagonal lines, even if the order is changed. On the other hand, if two of the three linearly independent unit vectors e z and ey + e are selected as needed to be used as Δ1, Δ1, the degree of freedom in selecting subspaces is expanded, as shown in Figures 7A-7C.

'Announcement-' Filter τ6 To further improve the codebook, the present invention provides a larger number of delta vector candidates (L': L' > L) than the delta vectors actually used to construct the codebook (L = initial vector 〒 L-] delta vectors. These candidates are then reordered using the same procedure as above. After that, the desired number (L) of delta vectors with the highest amplitude amplification factors are selected to construct the codebook. This allows for a codebook with a high degree of flexibility, resulting in improved quantization characteristics.

While the above description applies to the encoder, the decoder can also be configured with the same delta vector candidates and controlled in the same way as the encoder, ensuring compatibility with the encoder by always generating and using the same codebook as the encoder.

Figure 8 is a block diagram showing an embodiment of a speech coding method according to the present invention based on the above concept. In this embodiment, delta vector codebook 10 is configured to store and retain initial vector C0 (=Δ) representing one reference noise sequence and delta vectors Δ1 through ΔL·-1 representing N-dimensional delta noise sequences (L'-1) in number, which is greater than the (L-1) sequences actually used. Initial vector c0 and each delta vector Δ1 through Δ·-1 are each expressed in N dimensions. In other words, the initial vector and delta vector are N-dimensional vectors that each encode the amplitude of N samples of noise occurring in a time series.

In this embodiment, the linear predictive synthesis filter 3 is composed of an IIR filter of order , and an NXN square matrix A generated from the filter's impulse response is multiplied by a delta vector Δ1 to apply filter processing A to the delta vector Δ, outputting vector AΔ. The Np coefficients of the +111 filter vary based on the input speech signal and are determined each time using a well-known method. That is, since there is correlation between adjacent samples of the input speech signal, the correlation coefficient between the samples is obtained, and from this correlation coefficient, a partial autocorrelation coefficient known as the PARCOR coefficient is obtained. The alpha coefficient of the IIR filter is then determined from the PARCOR coefficient, and an NXN square matrix A is created using the filter's impulse response sequence to apply filter processing to vector Δ.

The filtered vectors AΔ, (i = 0.1 ... L'-1) are stored in the memory unit 40, and the power IAΔ112 = (AΔi)"(AΔ,) is evaluated in the power evaluation unit 42. Since each delta vector Δ, is normalized (1Δi" = (Δ,) T(Δ1) = 1), simply evaluating the power directly evaluates the amplification caused by the filter processing A. Based on the results of the power evaluation unit 42, the vectors are sorted in order of power magnitude in the sorting unit 43. For example, in the example shown in Figure 6B, the order is Δ = et, Δ + = ex + Δ2 = ey.

There are a total of L vectors AΔ; (i = 0, 1 ... L'-1) rearranged as described above, but the encoding process is performed using the L vectors AΔ + (i = 0, 1 ... L-1) that are actually used.

Therefore, the selection storage unit 41 selects and stores only L vectors with the largest amplitude amplification factors. For example, in the above example, Δ = e8 and Δ = -6 are selected from the delta vectors. Then, based on the tree-structured delta codebook formed by these vectors, encoding is performed in exactly the same manner as in the conventional tree-structured delta codebook described above.

Details of the Coder 5 Unit The following describes the details of the encoding unit 48, which finds the index of the code vector C that has the smallest distance from the input signal vector X, using a tree-structured delta codebook consisting of vectors AΔ, AΔ1, AΔ2, ... AΔL-1 stored in the selection storage unit 41 and the input signal vector X.

The encoding unit 48 includes a calculation unit 50 that calculates the cross-correlation X′′ (AΔ1) between the input signal vector X and each delta vector Δ, a calculation unit 52 that calculates the autocorrelation (AΔ,)′′ (AΔ,) of each delta vector Δ, a calculation unit 54 that calculates the cross-correlation (AΔi)′′ (AΔ, 1□, , , , - ,) between delta vectors, a calculation unit 55 that calculates the cross term (AΔi)′′ (ACk) from the output of calculation unit 54, and a calculation unit 56 that calculates the cross-correlation of each delta vector from calculation unit 50. It consists of a calculation unit 56 that accumulates correlations to calculate the cross-correlation Rxc between the input signal vector X and the code vector C, a calculation unit 58 that accumulates the autocorrelation (AΔi)″ (AΔ,) of each delta vector output by calculation unit 52 and the cross term (AΔi)’ (AC,) output by calculation unit 55 to calculate the autocorrelation of each code vector C, a calculation unit 60 that calculates RC11″/Rcc, a minimum-error noise sequence determination unit 62, and a speech coding unit 64.

Initially, the parameter χ representing the layer being calculated is set to 0. In this state, calculation units 50 and 52 calculate and output X'(AΔ.) and (AΔ.)'(AΔ.), respectively. Calculation units 54 and 55 output 0. The χ'(AΔ.) and (AΔ.)'(AΔ.) output by calculation units 50 and 52 are stored in calculation units 56 and 58 as the cross-correlation Rxc11 and auto-correlation Rcc(11+) at the first layer, respectively, and output. Calculation unit 60 calculates and outputs the value of F(X, C) = Rxc'/Rcc from R, c11, and Rcc(11+).

The minimum-error noise sequence determiner 62 compares the calculated F(X,C) with the maximum value of F(X,C) up to that point, Fwax (initial value 0). If F(X,C) > F+wax, it updates Ftaax as F(x,C) - Fmax, and updates the previous code with a code that specifies the noise sequence (code vector) that gives Fwax.

Next, parameter l is updated from 0 to 1. In this state, calculation units 50 and 52 respectively calculate and output χ"(AΔ+) + (AΔ,)" (AΔ1).

The calculation unit 54 calculates and outputs (AΔ+)' (AΔ). Calculation unit 55 outputs this value as the cross term (AΔl)" (AC,). Calculation unit 56: From the stored Rxe +o+ and the value of X'(AΔ 1) output from calculation unit 50, calculates, outputs, and stores the values of the cross-correlation Rc(1) and Rxc'2) in the second layer according to equations (10) and (11). Calculation unit 56 From the stored R0 0'o and the values of (AΔ 1)T(AΔ+), (AΔ+)" (Ac,) output from calculation units 52 and 55, respectively, calculates, outputs, and stores the values of the autocorrelation RCC(1) and Rcc(2) in the second layer according to equations (12) and (13). The operation of calculation unit 60 and minimum-error noise sequence determination unit 62 is the same as when I = o.

Next, parameter l is updated from 1 to 2. In this state, calculation units 50 and 52 calculate and output X"(AΔz) and (AΔz)'(AΔ2), respectively.

Calculation unit 54 calculates and outputs the cross-correlations (AΔz)' (AΔ) and (AΔz)' (AΔ) between Δ2 and Δ3. Calculation unit 55 calculates and outputs the cross term (AΔz)' (AC+) from these values according to equation (14).

Using the stored RXC(1) + RXC(t') and the value of X"(AΔ2) input from calculation unit 50, calculation unit 56 calculates, outputs, and stores the value of the cross-correlation R, cfl-61 at the third layer according to equations (10) and (11). Using the stored RCC(1) + Rcc" and the values of (AΔz)"(AΔa) and (AΔz)"(AC,) output from calculation units 52 and 55, respectively, calculation unit 58 calculates, outputs, and stores the value of the auto-correlation Rcc0""1 at the third layer according to equations (12) and (13). The operation of calculation unit 60 and minimum-error noise sequence determiner 62 is the same as when i = 0.1.

After repeating the above process up to 1 = L - 1, the speech encoding unit 64 outputs the most recent code stored in the minimum-error noise sequence determination unit 62 as the index of the code vector with the smallest distance from the input signal vector χ.

In addition, in the calculation of (AΔi)" (AΔ,) in the calculation unit 52, the calculation result of the power evaluation unit 42 can be used as is.

By using the aforementioned tree-structured delta codebook and the tree-structured delta codebook improved by the present invention, variable-rate coding can be achieved without the large memory requirements of conventional codebooks and with a built-in countermeasure against bit drop.

That is, a tree-structured delta codebook Δ having the structure shown in FIG.

Δ1, Δ2, etc. are stored, and as shown in Figure 9B, only the first-level vector Δ is used to find C, -O (zero vector) C0-Δ.

By encoding the two code vectors, one bit of information indicating whether or not to select C0 as index data is achieved.

By encoding using the vectors Δ and Δ1 up to the second layer to generate four code vectors: C1 (C3ΔO), C3 (Δ). +Δ1, and C2 (Δ). -Δ1, two-bit encoding is achieved using two bits of information specifying whether C0 is selected and whether ΔC0 or -ΔC1 is selected as index data.

Similarly, using vectors ΔΨ, Δ1, ... Δ up to the i-th stage, i-bit encoding is achieved. Therefore, by simply using a set of tree-structured delta codebooks containing L delta vectors ΔΨ, Δ, ... ΔL, the bit length of the generated index data can be arbitrarily varied within the range of 1 to L.

To perform variable bit-rate encoding of 1 to L bits using a conventional codebook, where N is the dimension of the vector, the number of memory words required is N x (2° + 2' - ... + 2L) = NX (2L'"' - 1). In contrast, if the tree-structured delta codebook of Figure 9^ is used as shown in Figure 9B, the number of memory words required is xL.

As a tree-structured delta codebook, any of the following can be used: a tree-structured delta codebook that does not perform the reordering described above; a tree-structured delta codebook that reorders delta vectors according to the magnitude of the amplification factor A; or a tree-structured delta codebook that selects and uses one of the L° data vectors.

Note that making the focus rate variable can be easily achieved by terminating the processing in the encoding processing unit 48 in Figure 8 at an intermediate layer depending on the desired number of bits. For example, in the case of 4-bit encoding, i x Q.

Regarding 1, 2 and 3, the processing of the encoding processing unit 48 described above may be performed.

Embedded coding, i.e., a coding system that can reproduce speech at the decoder even if some bits are forcibly dropped during transmission, is achieved by constructing a code system in the tree-structured delta codebook described above so that if some bits are dropped, they are reconstructed as their parent or ancestor code vector in the tree structure. For example, if one bit is dropped in a 4-bit code system CC, CI, ... C, 5, 4, CI3 + CI4 are constructed to be reproduced as a 3-bit system Ch, and C1, CI1 are constructed to be reproduced as a 3-bit system C2. In this way, the code vectors in the parent-child relationship have relatively close values, making it possible to reproduce speech without significant degradation in sound quality.

Tables 1-4 show examples of such code systems.

The above code system is determined as follows, for example, in the case of 4 bits:

C1, = Δ. -Δ1 - Δt - Δ has four delta vector elements, and the signs of each are (-, -, fu, 10) from the most significant, so this is represented as "10 11".

C2 = Δ. -Δ has only two delta vector elements, and the signs are (=, -). In this case, the signs are considered to be (0, 0, 1-) and represented as "0010".

Table 5 shows what happens when a bit drop occurs in the information coded in this way: 4 bits → 3 bits, 1 bit, 1 bit.

As can be seen from Table 5 and Figure 9A, if a single bit is missing, it is regenerated as a vector on the tPi layer and on the upper layer.

Furthermore, if two out-of-focus images occur, they are reproduced as shown in Table 6.

In this case, it is reproduced as the ancestor vector two levels higher.

Tables 7-10 show other examples of embedded code systems of the present invention.

In this coding system, if one bit is missing, the parent vector is regenerated, and if two bits are missing, the ancestor vector two levels above is regenerated.

Claims (1)

[Claims] 1. A speech coding method for encoding an input speech signal vector using an index assigned to a code vector among predefined code vectors that has the smallest distance from the input speech signal vector, the speech coding method comprising the steps of: a) storing a plurality of differential code vectors; b) multiplying each differential code vector by a linear predictive synthesis filter matrix; c) evaluating the power amplification factor of the differential code vector multiplied by the matrix; d) sorting the differential code vectors multiplied by the matrix in order of the magnitude of the evaluated power amplification factor; e) selecting a predetermined number of vectors from the sorted vectors in order of the magnitude of the evaluated power amplification factor; f) evaluating the distance between the input speech signal vector and a linear predictive synthesis filtered code vector to be generated by sequentially adding and subtracting the selected vectors in a tree structure; and g) determining the code vector with the smallest evaluated distance. 2. The method of claim 1, wherein each of the differential code vectors is normalized. 3. 3. The method of claim 1, wherein step f) includes calculating the cross-correlation Rxc between the input speech signal vector and the linearly predictive synthesis filtered code vector by calculating the cross-correlation between each of the selected vectors and the input speech signal vector and sequentially adding and subtracting them in a tree structure; calculating the auto-correlation Rcc of the linearly predictive synthesis filtered code vector by calculating the auto-correlation of each of the selected vectors and the cross-correlation of all combinations of different vectors and sequentially adding and subtracting them in a tree structure; and calculating Rxc2/Rcc, which is the square of the cross-correlation Rxc divided by the auto-correlation Rcc, for each code vector; and wherein step g) includes determining the code vector providing the largest Rxc2/Rcc value as the code vector having the smallest distance from the input speech signal vector. A coding device that encodes an input speech signal vector using an index assigned to a code vector among predefined code vectors that has the smallest distance from the input speech signal vector, the device comprising: means for storing a plurality of differential code vectors; means for multiplying each of the differential code vectors by a linear predictive synthesis filter matrix; means for evaluating the power amplification factor of the differential code vector multiplied by the matrix; means for sorting the differential code vectors multiplied by the matrix in order of the magnitude of the evaluated power amplification factor; means for selecting a predetermined number of vectors from the sorted vectors in order of the magnitude of the evaluated power amplification factor; means for evaluating the distance between the input speech signal vector and a linear predictive synthesis filtered code vector to be generated by sequentially adding and subtracting the selected vectors in a tree structure; and means for determining the code vector with the smallest evaluated distance. 5. The device according to claim 4, wherein the names of the differential code vectors are normalized. 6. The apparatus 7 according to claim 4, wherein the distance evaluation means includes means for calculating the cross-correlation Rxc between the input speech signal vector and the linear predictive synthesis filtered code vector by calculating the cross-correlation between each of the selected vectors and the input speech signal vector and sequentially adding and subtracting the cross-correlations in a tree structure; means for calculating the auto-correlation Rcc of the linear predictive synthesis filtered code vector by calculating the auto-correlation of each of the selected vectors and the cross-correlation of all combinations of different vectors and sequentially adding and subtracting the cross-correlations in a tree structure; and means for calculating Rxc2/Rcc, which is the square of the cross-correlation Rxc divided by the auto-correlation Rcc, for each code vector; and the code vector determination means includes means for determining the code vector providing the largest Rxc2/Rcc value as the code vector having the smallest distance from the input speech signal vector. A variable-length speech coding method for variable-length coding an input speech signal vector using a code of a variable bit length assigned to a code vector among predefined code vectors that has the smallest distance from the input speech signal vector, the method comprising the steps of: a) storing a plurality of differential code vectors; b) evaluating the distance between the input speech signal vector and a code vector to be generated by sequentially adding and subtracting, in a tree structure, a number of differential code vectors corresponding to a desired code bit length, starting from the first; c) determining the code vector with the smallest evaluated distance; and d) determining a code of the desired code bit length to be assigned to the determined code vector. 8. The method of claim 7, further comprising the step of multiplying each differential code vector by a linear predictive synthesis filter matrix, wherein in step b), the distance between the input speech signal vector and a linear predictive synthesis filtered code vector to be generated by sequentially adding and subtracting, in a tree structure, the differential code vectors multiplied by the matrix. 9. 10. The method of claim 8, wherein step b) includes: calculating the cross-correlation Rxc between the input speech signal vector and the linear predictive synthesis filtered code vector by calculating the cross-correlation between each of the differential code vectors multiplied by the matrix and the input speech signal vector, and sequentially adding and subtracting these values in a tree structure; calculating the auto-correlation Rcc of the linear predictive synthesis filtered code vector by calculating the auto-correlation of each of the differential code vectors multiplied by the matrix and the cross-correlation of all combinations of different vectors, and sequentially adding and subtracting these values in a tree structure; and evaluating Rxc2/Rcc, which is the square of the cross-correlation Rxc divided by the auto-correlation Rcc, for each code vector; and wherein step c) includes determining the code vector yielding the largest Rxc2/Rcc value as the code vector having the smallest distance from the input speech signal vector. The method of claim 9 further includes a step of evaluating the power amplification factor of the differential code vector multiplied by the matrix and sorting the differential code vectors multiplied by the matrix in order of the magnitude of the evaluated power amplification factor, wherein in step b), addition and subtraction on the tree structure are performed in accordance with the sorted order. 11. The method of claim 10 further includes a step of selecting a predetermined number of vectors from the sorted vectors in order of the magnitude of the evaluated power amplification factor, wherein in step b), addition and subtraction on the tree structure are performed using the selected vectors. A variable-length speech coding device that performs variable-length coding on an input speech signal vector using a code of a variable bit length assigned to a code vector among pre-specified code vectors that has the smallest distance from the input speech signal vector, the variable-length speech coding device comprising: means for storing a plurality of differential code vectors; means for evaluating the distance between the input speech signal vector and a code vector to be generated by sequentially adding and subtracting, in a tree structure, a number of differential code vectors corresponding to a desired code bit length, starting from the first; means for determining the code vector with the smallest evaluated distance; and means for determining a code of a desired code bit length to be assigned to the determined code vector. 13. The device of claim 12 further comprises means for multiplying each differential code vector by a linear predictive synthesis filter matrix, wherein the distance evaluation means evaluates the distance between the input speech signal vector and a linear predictive synthesis filtered code vector to be generated by sequentially adding and subtracting, in a tree structure, the differential code vectors multiplied by the matrix. The apparatus of claim 13, wherein the distance evaluation means includes: means for calculating the cross-correlation Rxc between each of the differential code vectors multiplied by the matrix and the input speech signal vector, and sequentially adding and subtracting the results in a tree structure to calculate the cross-correlation Rxc between the input speech signal vector and the linear predictive synthesis filtered code vector; means for calculating the auto-correlation Rcc of the linear predictive synthesis filtered code vector by calculating the auto-correlation Rxc of each of the differential code vectors multiplied by the matrix and the cross-correlation of all combinations of different vectors, and sequentially adding and subtracting the results in a tree structure; and means for evaluating Rxc2/Rcc, which is the square of the cross-correlation Rxc divided by the auto-correlation Rcc, for each code vector; and the code vector determination means includes means for determining the code vector providing the largest Rxc2/Rcc value as the code vector having the smallest distance from the input speech signal vector. The apparatus of claim 14 further comprises means for evaluating a power amplification factor of a differential code vector multiplied by the matrix, and means for sorting the differential code vectors multiplied by the matrix in order of the magnitude of the evaluated power amplification factor, wherein the distance evaluation means performs tree-structured addition and subtraction in the sorted order. 16. The apparatus of claim 15 further comprises means for selecting a predetermined number of vectors from the sorted vectors in order of the magnitude of the evaluated power amplification factor, wherein the distance evaluation means performs tree-structured addition and subtraction on the selected vectors. 16. The method of claim 7, wherein each code vector is assigned a code such that, if one bit is missing, it corresponds to the code vector corresponding to its parent in the tree structure. 17. The apparatus of claim 12, wherein each code vector is assigned a code such that, if one bit is missing, it corresponds to the code vector corresponding to its parent in the tree structure.