FI116597B - In particular, a method for vector quantization of speech signals - Google Patents

Description

116597

In particular, a method for vector quantization of speech signals

The invention provides a method for encoding signal sampling values 5 using vector quantization.

From Speech Communication 8 (1989), pages 363-369, a CELP speech coding method is known in which code variables are jointly optimized. Compared to sequential optimization, the length of the excitation codebook can be significantly reduced by 10.

WO 91/01545 discloses a digital speech decoder which searches for excitation vectors stored in a codebook to select the excitation vector that best represents the original speech sampling value. The speech encoder of Jul-15 WO 91/01545 uses two excitation vectors each time in one of the two codebooks to explain the sampling value of a speech. First, the first excitation vector, regardless of pitch information, is selected and orthogonalized. The second excitation-20 vector is selected similarly. The orthogonalization of the second excitation vector from the second codebook takes into account both the resulting vector and the selected first »· t *. ,: Excitation vector from the first codebook. This selection process is then repeated for the orthogonal excitation signal taken from the second codebook to finally identify the excitation vectors that best match the sample of the original speech. tysarvoa.

It is an object of the invention to improve reliability in the selection of an optimized sampling value without unduly complicating the processing. This task is solved by the measures of claim 1. Alternative direct; other forms are required.

• * t

The invention is based on the following information: When, in contrast to known methods (Speech Communication:., Vol. 35 8 (1989), pages 363-369 or WO 91/01545), the adaptive {t I t | * t 116597 Of the 2 (second) codebooks, more than one minimal error vector is used to combine with all vectors of the first (fixed) codebook, although the workload required for processing increases, but the confidence in optimizing the smallest error sampling value increases. This improvement in reliability translates into better speech quality when processing speech sampling values. Because the amount of work involved in processing, taking into account more than one vector from the adaptive codebook 10, increases less than linearly, a fixed codebook can be obtained by moderately reducing, e.g., codebook thinning "15 (same filing date), the processing workload was kept approximately constant, with the original codebook length without frame reduction as the reference. With the amount of processing remaining approximately the same as in the conventional method, substantially improved speech quality can be achieved by the measures of the invention.

An exemplary embodiment of the invention will now be considered in greater detail on the basis of the drawings.

• 1 1 ·: Figure 1 illustrates the structure of a CELP encoder.

• ·

Figure 2 illustrates the structure of a modified CELP encoder.

1 25 For the sake of clarity in the discussion of the invention, reference is made to "Improving Performance of Code Excitation LPC-Coders by Joint Optimization" (Speech Communication 8: (1989), pages 363-369).

CELP (Code-excited Linear Prediction) encoders are • 1 · 30 licensed to RELP (Residual Excited Linear Prediction) encoders ·., A class in which speech value update queues are generated directly; , ·, By means of a datum representing speech formation. The update queue is obtained by using a codebook from which the best size-codebook vector is selected by the method of "synthesis-based analysis". The best codebook vector here refers to the vector, 116597 3, which has the greatest similarity with the sample speech value of the original speech. This similarity is estimated on the basis of a predefined error criterion, e.g., an average squared error. The codebook is first filled with random values representing normal distribution. The CELP encoder structure shown in Figure 1. In the first step the linear advance tisuodattimen, which in Figure 1 are denoted by the transfer function H (z), a memory portion will reduce the input side of the speech

Ob is sampled and the resulting signal is weighted through a filter comprising a transfer function W (Z). In the second step, the proportion of the weighted memory value of the pitch prediction filter (characterized by the transmission functions H T (Z) and H (Z)) is reduced. Finally, the weighted error signal e (n) is generated by generating the difference w 15 in the filtered codebook vector (filter functions H (Z) and

L

H (Z)) and the predefined signal s (n). The energy of the signal V (w) w e (n) is a function of all code variables, e.g.

20 E = ffa ^ M, b ±, j, Cj), where a. For i = 1,2, ..., P denotes the LP filter-

IM · 1 O

: time odds,: ·, ·! M pitch, ^! 25 b. For i = 1,2, ..., P,

::: i L

coefficients, j = 1.2, ... K codebook entries and •> · b

Cj the corresponding scaling factor.

The best possible speech quality is achieved by • optimizing these signal variables collectively. In the next optimization, the LP variable will not be considered, since taking it into account will hardly do any computational work.

»I

T By minimizing the function E = f (M, j,), 35 suboptimal approximations are obtained.

116597 4

The linear prediction synthesis filter p i -1 H (Z) - (1- E s a. Z) 5 S i = 11 represents the formant structure of the speech spectrum. The weighting filter W (Z) = Hgi Ζ / γ) HS (Z) -110, where 0 <γ <1, provides a spectral noise restriction due to incomplete excitation. HW (Z) produces a chaining of the LP filter and the weighting filter 15 HW (Z) = Hg (Z). W (Z).

A pitch prediction synthesis filter with only

one tap at p = 1 is described by the transfer function L

20 H (Z) = (1 - bZ_M) _1

L

through.

: The memory cells of the filters H (Z), H (Z) and W (Z) are · W Li in Figure 1 at zero. Volume Predictor Variables Daily | *: 25 is always scanned after N sampling values (subframe: contents) and LP filters are updated after every I N sampling values. By default, N> N can be

«M I S

the pitch prediction filter removes the wake branch of Fig. 1 as it does not affect the input of filter H (Z) to w. . 30 streams to reach n <N.

* * · B

For a closer look at the effect of the pitch predictor memory> t, the memory cells of the memory and their connections are described in detail in Fig. 1.

; The values of the memory cells are denoted by l (k). Each of the pitch variables M = k produces a delay of the second signal d (n) from the memory cells to the line output. K depends

* * * * * L

116597 5 of the allowed range for the pitch period M. A good choice for M is between 40 and 103. To cover this range, K = 64.

L

These conditions lead directly to the block diagram of Figure 2.

The different signal d (n) of K can be viewed by L K

as if they were put together in a single codebook. In this presentation, there is no difference between the structures of the branch comprising the excitation codebook CB1 and the branch comprising the codebook CB2, which is evident from the pitch memory filter memory. Only the characteristic curves of the codebooks CB1 and CB2 are different: the excitation codebook CB1 is fixed - fixed vectors are recorded therein, whereas the codebook CB2 is time dependent (adjustable) for pitch variables, since the filter memory is modified after each subframe. To optimize these variables, you need to study a large number (K K) of different combinations in order to

L S

we find the minimal error energy E. All of these combinations correspond to the codebook length K K while

L S

The 20-hand optimization for two-stage vector quantization corresponds to two codebooks of length or K i.

According to the block diagram of Figure 2, the error energy E is a function of the codebook entries j and k, and the scaling factors and b:

• »K

. . 25 Ns 2!.:! E (j, k, b, c) = Σ [S (h) - [(b d. (N) + c. T (n) 1 h (hj |) l KJ Π = 1 ^ Λ Λ JJ w 1 1 where h (n) indicates pulse response of the weighted LP filter, 30 background and 1 spread symbol.

In order to obtain the minimum error energy E, the following linear equations must be scaled * for the factors:::, the system must satisfy: • », 35 f <pk (n) 'pk (n)> <pk (n) / qj (n)> \ f bk) \ <Pk (n), q. (n)> <q (n), q (n)> / (d j.

| , J J J J

116597 6 | Sc (n) 'sw (n)> \ \ qj (n)' sw (n)> / where Pk (n) = <* k (n) * ^ (n), 5 q (n) = r (n) * h (n), j J w

OF

and <^ / \ r b> - Σδ a (n) b (n).

(n) (n) n = 1 10 Using these ratios, for a minimal error energy, 15 Emin = <sW (n), sw (n)> - T (j, k, c _., bk).

Since the energy for the subframe is constant, the expression 20 T (j, k, Cj, bk) = bk <pk (n), sw (n)> + <q ^ (n), sw (n)> must be maximized. This maximization is accomplished in two steps • ·:.; · Sa: - Solution of the linear system of equations ί: 25 - Calculation of T (j, k, c., B).

* These steps must be performed K K times. Other Me-

* ·> · L S

: The steps to simplify the method are that, for example, about 90%

* * * I

of the vectors is set to zero, according to DE 3 834 871 II

Inverse filtering according to Cl, allowing only vectors, for example, which always contain at least three values '', 1. zero autocorrelation coefficients.

• *

In accordance with the invention, and unlike hitherto: in known methods, the best vectors are now selected from the second codebook CB2 n> 2, in the example n = 2 (the best vectors mean that these vectors give a certain current I i, e.g. 116597 7 (ie, the best predicted values). These two best vectors are then combined with all vectors of the first codebook CB1 containing solid vectors according to the above system of linear equations 5. For the minimal error energy (same or another error criterion), the values that best match the original sampling value are selected from the combined set, and are prepared, for example, to be transmitted over a lower bit rate transmission channel.

10 If the processing is complicated by handling more than two of the best vectors from the second codebook, the result will be a better speech quality. Without having to reduce this improved speech quality, the processing can be further simplified by reducing the entries in the first 15 codebooks CB1. In addition, the amount of work involved in processing does not increase linearly as the number of vectors selected for processing increases, since many combination results already calculated in the first step can be utilized.

Preferably, codebook thinning is performed without impairing speech quality by defining the basis of the set of thinnings to be the vector i i | : s sumbits, which are then preferably suppressed by * '· y just as many bits that the workmanship is just as large! ! 25 if only one selected best vector from the second codebook CB2 is processed. Codebook Thinning: explained in detail in the above applicant's application · · · 'with the same dates.

Claims (7)

116597 8 A method for vector quantization of signal sampling values, in particular speech, comprising the steps of: - recording fixed vectors in a first codebook (CB1), - vectors in a second codebook (CB2) updated every one time period, from a second codebook (CB2) selecting at least two vectors which give the best predicted values according to a preset first error criterion, - combining at least two vectors selected from the second codebook (CB2) with vectors derived from the first codebook (CB1), selecting a vector to represent the optimized sampling value which, based on the first or some other error criterion, is the shortest distance from the original sampling value. Method according to claim 1, characterized in that excitation vectors for the LPC filter are recorded in the first codebook '(CB1): to predict speech. · Method according to claim 1 or 2, I; 25 characterized in that the second codebook (CB2)) ", * records excitation vectors for the pitch synthesis filter" ;; ·. The method according to any one of claims 1 to 3, characterized in that the entries to the first code letter (CB1) are thinned. A thread according to any one of claims 1-4; method, characterized in that for each excitation vector of the first '!! / codebook (CB1), an error dimension relative to the original sampling value is defined, which takes into account at least two selected from the second codebook 9116597 (CB2). predicted value / pitch predictor. Method according to Claim 4 or 5, characterized in that, in order to attenuate the first codebook (CB1) 5, such vector components are suppressed, which are preferably removed from the sum bits of two frame periods into which the speech signal is divided. Method according to one of Claims 4 to 6, characterized in that the attenuation of the entries in the first codebook (CB1) is carried out to such an extent that the processing job is just as large as processing only one vector selected from the second codebook (CB2). • »* 1 · * I» »116597 10