JPH06138896A - Device and method for encoding speech frame - Google Patents

Abstract

PURPOSE: To relieve the burden of code exciting linear prediction(CELP) voice encoding in terms of a computer by selecting an adaptive code book vector having a min. error function. CONSTITUTION: An adaptive code book searcher 220' uses the inpulse sensible weighting response function H(n) of a short term LPC filter in a convolution generator 510' in order to generate a convolution signal W(n) and uses the frame of target voice X(n) which is sensibly weighted in order to execute a convolution. Then, a self correlative coefficient 552' stored in a table 555' before code book searching is used next for calculating energy as against respective candidate vectors from an adaptive code book 155 later, the result of the normalized correlation of the respective vectors is compared in a peak selector 570' and the vector Ck (m) having a max. mutual correlative value is identified as an optimum pitch cycle vector by the peak selector 570'.

Description

Detailed Description of the Invention

[0001]

FIELD OF THE INVENTION This invention relates to improved means and methods for digital encoding of speech or other analog signals, and more particularly to code excitation or code driven linear predictive coding.
ear predictive coding).

[0002]

Code Excited Linear Prediction (CELP) coding is a well-known stochastic or stochastic coding technique for voice communications. CELP
In encoding, the short time spectrum and long time pitch are modeled by a set of time varying linear filters. In a typical voice coder-based communication system, voice is sampled by an A / D converter at approximately twice the highest frequency desired to be transmitted, eg, for a voice bandwidth of 4 KHz. Typically uses a sampling frequency of 8 KHz. CE
LP coding synthesizes speech using encoded excitation information to excite or drive a linear prediction (LPC) filter. The excitation or excitation used as input to the filter is modeled by a codebook of white Gaussian signals. The optimal excitation is detected by searching the codebook of candidate excitation vectors on a frame-by-frame basis.

LPC analysis is performed on input speech frames to determine LPC parameters. Next LPC
The analysis proceeds by comparing the output of the LPC filter with the digitized input speech as the filter is excited by various candidate vectors from the table, the codebook. The best candidate vector is selected based on how well the speech synthesized using the candidate excitation vector matches the input speech. This is usually done for several subframes of speech.

Information identifying the best codebook entry after the best match is detected,
The LPC filter coefficients and gain coefficients are sent to the synthesizer. The synthesizer has the same copy of the codebook and accesses the appropriate entry in that codebook and uses it to excite the same LPC filter.

A codebook consists of vectors whose components are consecutive excitation samples. Each vector contains as many excitation samples as there are audio samples in a subframe or frame. Excited samples can come from a number of different sources. The encoding of the long term pitch is determined by the proper selection of code vectors from the adaptive codebook. The adaptive codebook is a set of pre-synthesized speech excitation waveforms with different pitch periods.

The optimal choice of code vectors depends on minimizing the perceptually weighted error function, either from the stochastic or adaptive codebook.
This error function is typically obtained by comparing the synthesized speech for each vector in the codebook with the target speech. These exhaustive comparison procedures are computationally intensive and are usually not practical for a single digital signal processor (DSP) to perform in real time. It is important in a digital communication environment to be able to reduce computational complexity without sacrificing voice quality.

[0007]

Error functions, codebook vector searches, and calculations are performed using vector and matrix operations of excitation information and LPC filters. The problem is that many calculations have to be performed, for example about 5 × 10 ⁸ multiply-add operations per second for a 4.8 Kbps vocoder. Prior art arrangements have not been entirely successful in reducing the number of calculations that must be performed. Therefore, there continues to be a need for improved CELP coding means and methods that reduce the computational burden without sacrificing voice quality.

Prior art 4.8 kbit / sec CELP
The encoding system is the US Government's General Ser
Federal Standard FED-issued by Vice Administration
STD-1016. Prior art CEL
P-vocoder systems are described, for example, in US Pat. Nos. 4,899,385 and 4,910 to Ketchum et al.
781, No. 4,220,819 to Atal, Li
n, 4,797,925, and Gerson, 4,817,157, which are hereby incorporated by reference.

A typical prior art CELP vocoder system has a sampling rate of 8 KHz and four 7.
A 30 millisecond frame period divided into 5 millisecond subframes is used. Prior art CELP
The encoding consists of three basic functions. That is,
(1) short delay "spectrum" prediction, (2) long delay "pitch" search, and (3) residual (residu)
al) "Codebook" search.

Although the present invention is described in the case of an analog signal representing human speech, this is merely for convenience of description and as used herein, "speech". The term is intended to include any form of analog signal of bandwidth within the sampling capability of the system.

[0011]

SUMMARY OF THE INVENTION The present invention provides improved means and methods that substantially reduce the computational burden of adaptive and stochastic codebook based CELP speech coding.

In the first embodiment, a recursive computation loop is used to vote the vector of the adaptive codebook to select the optimal excitation vector from it. In the preferred embodiment, the impulse function of the short-term perceptually weighted filter is convolved and autocorrelated with the perceptually weighted target speech and the result correlated with each vector in the adaptive codebook. Combined with the vector and auto-correlated impulse function to generate an error function. The adaptive codebook vector with the smallest error function is selected to represent the particular speech frame (or subframe) being examined.

Another embodiment further provides a recursive loop for the adaptive codebook by reducing the number of autocorrelation operations that must be performed on the K vectors of the adaptive codebook, each having N entries. To simplify. Autocorrelation is initially performed only for a small number P of P << N for N autocorrelation coefficients in each codebook vector, and the detected value is used to operate through all K codebook vectors. Search for S of the K codebook vectors (S << K) that give the best match to the input speech. The autocorrelation function for the S vector is then recomputed for the R autocorrelation coefficient (P <R≤N) and the S codebook vector is any of the S adaptive codebook vectors for the input speech. Re-evaluated to determine which gives the best match.

Yet another embodiment further reduces the number of calculations that must be performed to determine the autocorrelation coefficient when a codebook vector of length M less than the frame length L is being evaluated. . The autocorrelation coefficient U _k (m) of the first vector C _k (n) of length M <L is calculated,
Here, k = 1 and m is an autocorrelation delay index (lag in
dex) and n is the index of consecutive samples in the codebook vector, and L is the analysis frame length according to U ₁ (m) = (L / M) U ′ ₁ (m) (2) Note that m = 0 to T <m. The autocorrelation coefficient U _k (m) of the remaining codebook vectors is incrementally calculated according to the following equation, where k ≧ 2. U ′ _k (m) = [U ′ _k−1 (m) + C _k (M + k−1) · C _k (M + k−1 + m)] (3) U _k (m) = {L / (M + k−1)} _U'k (m) (4)

However, m = 0 to T <M, and the process is repeated until (M + k−1) = L. It is preferred that T = M-1. The obtained values U _l ′ (m) and U _k ′
(M) is scaled by the indicated scaling factor, eg (L / M) for k = 1, k =
L / (M + 1) for 2, and similarly (M + k-1)
= L is performed. The resulting autocorrelation coefficient is used in determining which of the codebook vectors C _k (n) produces the smallest error when compared to the input speech.

In yet another embodiment, the phase of the stochastic codebook vector along with other vectors generated by the CELP encoding process to identify the optimal stochastic codebook vector for replicating the target speech. Means and methods are provided for determining the number of relationships more quickly and easily. More specifically, a first vector V (n) having a value identified by an index n ranging from n = 1 to N, and a set of second vectors S
_k (n), each of the second vectors is identified by an index k, and each of the second vectors is zero (ze).
ro) or non-zero and having values up to N identified by the exponent n from n = 1 to N, where _k is different from S _k (n _i ).
To the index n _{k, i} of S _k (n) for
The values of V (n) corresponding to _{k and i} are added to form a sum Q (k), and the value k = j corresponding to the largest value Q (k = j).
, And combine them by synthesizing the speech using S _{k = j} (n).

In a preferred embodiment, the second of the set
The continuous vector of the vector is the overlap amount Δk,
Determined by the overlap of the preceding second vector according to Δn, and said identifying and adding step is S
₁ (n _i ) is non-zero, the index n of S _k (n)
_{1, i} for k = 1, starting from n _{1, i} and using said overlap amounts Δk, Δn, if S _k (n _{i ′} ) is non-zero, k> 1 , For further exponents n _{k, i ′} , and for summing the values of V (n) for such exponents and further exponents to form sum Q (k).

In yet another embodiment, n = 1 to N
, An N × N multiplexer having n = 1 to N first inputs, a second input, and n = 1 to N selection means are used to combine the codebook vectors. The first logic level provided to the nth selection means couples the nth output to the nth first input and the second logic level provided to the nth selection means is the nth output. To the second input. The second input is expediently a predetermined logic level. The values of n = 1 to N of said first vector are fed to the first inputs of n = 1 to N of the multiplexer and n = of the second vector of index k = 1.
The values 1 to N are supplied to the n = 1 to N selection means of the multiplexer. The second vector provides a first logic level for some values of n and a second logic level for other values of n in the selection means for n = 1 to N. The values of the first vector appearing at the multiplexer output are added in the accumulator to provide the sum.
The providing and summing steps are repeated for further values of k, and speech is synthesized based on which second vector has the sum that gives the closest match to the target speech.

In the preferred embodiment, each second vector is divided into two parts: the value 0, which corresponds to the location of the 0, + 1 value of the second vector.
The first part having 1 and the value 0 of the second vector,
It is divided into a second part having the values 0,1 corresponding to a location of -1. Each part is preferably stored in ROM. Multiplexers are provided for each part, such as those described above and operating in a similar manner, each multiplexer providing an accumulated output based on the 0,1 value of its associated part of the second vector. Offer the sum. The providing and adding steps are repeated for each ROM-multiplexer combination for each value of k as described above,
And the sums resulting from each are further combined in another adder (or subtractor) to provide a combined output for use in selecting which vector gives the closest match to the target speech.

[0020]

DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENT FIG. 1 is a simplified block diagram form of C
1 illustrates a vocoder transmission system utilizing ELP coding.
CELP coder 100 receives input speech 102 and
An ELP encoded output signal 104 is generated. CELP encoded signal 104 is transmitted via transmission path or channel 106 to CELP decoder 300, where a facsimile 302 of original audio signal 102 is reconstructed by synthesis. The transmission channel 106 can be of any form, but is typically a limited bandwidth wired or wireless communication link. CELP coder 100
Is often referred to as an "analyzer", whose function is to determine the CELP code parameters 104 (eg, codebook vectors, gain information, LPC filter parameters, etc.) that best represent the original speech 102. This is because. CELP decoder 300 is often referred to as a synthesizer, whose function is to output synthesized speech 3 based on input CELP encoded signal 104.
02 is to be reproduced. CELP decoder 300 is conventional and not part of the present invention, and thus will not be described in further detail.

2A and 2B show CELP coder 100 in greater detail and in accordance with a preferred embodiment of the present invention. The input analog audio signal 102 is first band-pass filtered by a filter 110 (band-pas).
sed) Prevents aliasing. The bandpass filtered analog audio signal 111 is then sampled by an analog-to-digital (A / D) converter 112. Sampling is usually Nike straight,
For example, 8KH for a 4K CELP vocoder
performed in z. Other sampling rates can also be used. Any suitable A / D converter can be used. The digitized signal 113 from the A / D converter 112 consists of a series of samples, for example a series of narrow pulses whose amplitude corresponds to the envelope of the speech waveform.

The digitized audio signal 113 is then divided into frames or blocks, ie, into consecutive time brackets containing a predetermined number of digitized audio samples, eg 60, 180 or 240 samples per frame. Will be divided. This is customarily called "frame rate" in CELP processing. Other frame rates can also be used. This is done in the framer 114. Means for achieving this are well known in the art. The consecutive audio frames 115 are stored in the frame memory 116. The output 117 of the frame memory 116 is a block 122, 14 whose function is described below for the frame 117 of the digitized audio 115.
2, 162 and 235.

Those skilled in the art will appreciate that a frame of digitized speech can be further divided into subframes, and speech analysis and synthesis can be done using subframes. The term "frame" as used here
Considers to refer to both frames and subframes of digital speech, whether singular or plural.

CELP coder 100 uses two codebooks, an adaptive codebook 155 and a stochastic codebook 180 (see FIG. 2 (b)). For each speech frame 115, the coder 100 calculates an LPC coefficient 123 that represents the formant characteristics of the vocal tract.
The coder 100 also inputs (vectors) from both the stochastic codebook 180 and the adaptive codebook 155.
And the associated scaling (gain) factors, which, when used to excite or drive a filter with LPC coefficients 123, are searched for the one that best approximates the input speech frame 117. L
The PC case, codebook vector and scaling (gain factor) information is processed and channel coder 21
0, where they are combined and encoded CEL
P signal 104 is formed and is transmitted to CELP decoder 300 by path 106. The processing in which the above is performed will be described in more detail.

Referring now to data path 121, which includes blocks 122, 125, 130 and 135, LP
The C analyzer 122 is responsive to the input speech frame 117 to determine the LPC coefficient 123 using well known techniques. LPC coefficient 123 is line spectrum pair (LSP)
Or in the form of Line Spectral Frequency (LSF), these terms are well understood in the art. L
The SP 123 is quantized by the coder 125, and the quantized LPC output signal 126 is the channel coder 21.
0, where the signal is transmitted to decoder 300 via transmit channel 106.
Form a part (i.e., the LPC filter coefficient) of.

The quantized LPC coefficients 126 are decoded by the decoder 130 and the decoded LS
P is a spectral inverse filter, described in connection with data paths 141 and 161, respectively, via output signals 131, 132.
se filters) 145 and 170, and via output signal 133 to bandwidth extension weighting generator 135. Signals 131, 132 and 1
33 contains information about the decoded quantized LPC coefficients. Means for implementing coder 125 and decoder 130 are well known in the art.

The bandwidth extension weight generator 135 provides a scaling factor (typically = 0.8) and performs the function of formant bandwidth extension and contains information about the bandwidth extended LPC filter coefficients. Output signal 1
36 and 137 are generated. Signals 136 and 137 are each sent to cascade weighting filters 150 and 175, whose functions are described below.

Next, blocks 142, 145 and 15
Referring to the data path 141 including 0, the spectrum predictor memory subtractor 142 receives the short-term spectrum predictor filter 195 from the input sample speech 115 arriving from the frame memory 116 via 117 (see (b) of FIG. 2). Subtract the previous state at 196 (ie, the one left by the previous frame). Subtractor 142 provides speech residual signal 143, which is digitized input speech 115 minus what is technically referred to as a filter ringing signal or filtering down. The filter ringing signal is such that the impulse used to drive the filter associated with a given speech frame (eg, the LPC filter 195 of FIG. 2B) has not completely dissipated by the end of that frame, Occurs because it may cause a filter drive or excitation (ie, "ringing") that spans the frame of.

The ringing signal appears as distortion in subsequent frames because it is independent of the speech content of that frame. If the ringing signal is not removed, it affects the selection of code parameters and the decoder 30.
0 deteriorates the quality of speech synthesized.

Filter ringing signal 1 from voice 115
The speech residual or residual signal 143, which contains information about minus 96 (subtracted), is provided from the decoder 130 along with the signal 131 to the spectral inverse filter 145. Filter 145 can typically be implemented as a zero filter (ie, A (z) = A ₀ + A
_1z ^-1 + _... + a A ^{n z -n,} each A LPC here
Is the coefficient of the filter and z is the "Z transform" of the filter
However, other means well known in the art can also be used. Signals 131 and 143 are combined by convolution to produce LPC inverse filtered speech at filter 145. Filter 14
5 output signal 146 is a cascade weighting filter 15
Sent to 0. Filter 150 is typically a pole.
e) implemented as a filter (ie 1 / A
(Z / r), where A (z / r) = A ₀ + A ₁ rz ⁻¹
+ ... + A _n r ⁿ z ⁻ⁿ and each A is an LPC filter coefficient, r is an expansion factor and z is the “Z transform” of the filter), but are well known in the art. Other means can also be used.

The output signal 152 from block 150 is an impulse function (eg 1, 0, 0, ...) Due to the bandwidth expanded LPC coefficient signal 136 arriving from block 135.
0, ..., 0) is the perceptually weighted LPC impulse function H (n) obtained from the convolution of 0 ,. Signal 136 is also combined with signal 146 by convolution in block 150 to output 151.
Generate the perceptually weighted short delay target audio signal X (n) obtained from path 141 at.

The outputs 151 and 152 of weighting filter 150 are provided to adaptive codebook searcher 220. The target speech signal 151 (ie, X (n)) and the perceptually weighted impulse function signal 152 (ie, H (n)) are used by the searcher 220 and the adaptive codebook 155 to produce a pitch period (ie, filter 195). Drive vector) and the gain therefor that most closely correspond to the digitized input speech frame 117. The way in which this is achieved will be explained in more detail in connection with FIGS.

Referring now to datapath 161, which includes blocks 162, 165, 170 and 175, pitch predictor memory subtractor 162 sets previous filter state 192 in long delay pitch predictor filter 190 to 11
7 provides an output signal 163 consisting of the sampled speech minus the long delay pitch predictor filter 190 ringing subtracted from the digitized input sample speech 115 received from memory 116 via 7. Output signal 163
Is supplied to the spectrum predictor memory subtractor 165.

Spectral memory subtractor 165 performs a function similar to that described with respect to block 142, and reduces the short delay spectral predictor ("spectral") filter ringing or ringdown signal 196 to pitch subtractor 162. Subtract from digitized input speech frame 117 transmitted over. This is the remainder consisting of the sampled speech 117 of the current frame minus the ringing of the long delay (“pitch”) filter 190 and the short delay (“spectral”) filter 195 left over from the previous frame. Generate an output signal 166. The remainder signal 166 is provided to a spectral inverse filter 170 similar to block 145.

The inverse filter 170 outputs the remainder signal 16
6 and the output 132 of the decoder 130. Signal 132 contains information about the decoded quantized LPC coefficients. The filter 170 combines the signals 166 and 132 by convolution to produce an output signal 171 with LPC inverse filtered speech. Output signal 1
71 is sent to a cascade weighting filter 175 similar to block 150.

The weighting filter 175 is the filter 170.
171 from and the bandwidth extension weighting generator 13
The signal 137 from 5 is received. Signal 137 contains information about bandwidth-enhanced LPC coefficients. Cascade weighting filter 175 produces output signals 176 and 177.
Filter 175 is typically implemented as a pole filter (ie, only poles in the complex plane), although other means well known in the art can be used.

Signals 137 and 171 are combined by convolution in filter 175 and output 177.
Generate the perceptually weighted LPC impulse function H (n) obtained from path 121 at
At 76, the perceptually weighted long and short delay target speech signal Y obtained from path 161.
(N) is generated. The output signals 176, 177 are sent to a stochastic or stochastic searcher 225. The probabilistic searcher 225 uses the probabilistic codebook 180 to select the optimal white noise vector and the optimal scaling (gain) factor, which when applied to the pitch of the predetermined coefficient and the LPC filter. Gives the best match to audio frame 117. Probabilistic searcher 225 operates in a manner well known in the art and generally operates similar to that achieved by adaptive searcher 220 described in more detail in connection with FIGS.

In summary, in the chain 141,
Spectral inverse filter 145 receives LSP 131 and residue 143 and sends its output 146 to cascade weighting filter 150 to output 152 a perceptually weighted LPC impulse function response H (n).
And a perceptually weighted short-time delayed target speech signal X (n) at output 151. Chain 16
1, the spectral inverse filter 170 receives the LSP 132 and the short and long delay speech residuals 166 and outputs 1
71 to the weighting filter 175 to generate the perceptually weighted LPC impulse function H (n) at the output 177 and the perceptually weighted short-term and long-term delayed target speech signal Y (n) at the output 176. Occur.

Blocks 135, 150, 175 collectively labeled 230 provide a perceptual weighting function. The decoded LSP from chain 121 is used at block 135 to generate a bandwidth expansion weighting factor at outputs 136 and 137. Weighting factors 136, 137 are used to generate perceptually weighted LPC impulse function H (n) in cascade weighting filters 150 and 175. Each element of the perceptual weighting block 230 is an LPC.
In response to the coefficients, compute spectral weighting information in the form of a matrix that emphasizes portions of the speech known to have significant speech content. This spectrum weighting information 1 / A (z / r) is used as the cascade weighting filter 15.
Based on finite impal response H (n) of 0 and 175.
Utilizing the finite impulse response function H (n) significantly reduces the number of calculations that the codebook searchers 220 and 225 must accomplish. The spectral weighting information is utilized by the searcher to determine the best candidate for the driving information from codebooks 155 and 180.

With continued reference to FIGS. 2A and 2B, adaptive codebook searcher 220 produces an optimal adaptive codebook vector index 221 and associated gain 222 to be sent to channel coder 210. Probabilistic codebook searcher 225 produces an optimal stochastic codebook vector index 226 to be transmitted to channel coder 210 and an associated gain 227. These signals are encoded by the channel coder 210.

Channel coder 210 receives five signals. That is, the quantized LSP 12 from the coder 125
6, optimal stochastic codebook vector index 226 and gain setting 227 therefor, and adaptive codebook vector index 221 and gain setting 222 therefor. The output of channel coder 210 is a serial bitstream 104 of encoded parameters. Bitstream 104 is CEL over channel 106
It is sent to a P-decoder 300 (see FIG. 1), where after decoding, the recovered LSP, codebook vector and gain settings are applied to the same filter and codebook to produce synthesized speech 302.

As already explained, the CELP coder 1
00 determines the optimal CELP parameters to be sent to the decoder 300 by the process of analysis, synthesis and comparison. The results of using the trial CELP parameters must be compared on a frame-by-frame basis with the input speech so that the optimal CELP parameters can be selected. Blocks 190, 195, 197, 200, 205 and 23
5 is used to achieve this in combination with the blocks already described in Figures 2a and 2b. The selected CELP parameters (LSP coefficient, codebook vector and gain, etc.) are output 211
To the decoder 182, where those parameters are passed to blocks 190, 195, 197, 200,
205 and 235 and blocks 142, 145, 150, 162, 165, 17 already mentioned above.
Return to 0 and 175.

Block 182 represents signals 126, 221, 2
Identified as a "channel decoder" that has the ability to decode the signal 211 from the coder 210 to recover 22, 226, 227. However, those skilled in the art can omit the code-decode operation represented by blocks 210-182 and use signals 126, 221, 222,
226 and 227 are blocks 182 in uncoded form
Block 182 simply supplies the signal to block 19
It will be appreciated that it can be made to act as a buffer for distribution to 0, 195, 197, 200, 205 and 235. Both configurations are satisfactory and the terms "channel coder 182", "coder 1"
82 "or" block 182 "is intended to indicate an arrangement or any other means for passing such information.

The output signal of the decoder 182 is the block 1
Quantized LSP signal 126 to block 95, block 19
Adaptive codebook exponent signal 221, sent to block 190, adaptive codebook vector gain exponent signal 222, sent to block 180, stochastic codebook exponent signal 226, and stochastic codebook vector gain, sent to block 197. The index signal 227. These signals drive filter 190, thereby producing output 191 that is provided to adaptive codebook 155 and filter 195. Output 191 is coder 1
In combination with output 126 of 82, filter 195 is further driven to produce synthesized speech 196.

The synthesizer 228 is a gain multiplier 19
7, a long delay pitch predictor 190, a short delay spectrum predictor 195, a subtractor 235, a spectrum inverse filter 220 and a cascade weighting filter 205. Decoded parameter 12
6,221,222,226 and 227,
The stochastic code vector 179 is selected and sent to the gain multiplier 197 to be scaled by the gain parameter 226. Output 198 of gain multiplier 197
Is used by the long delay pitch estimator 190 to generate the speech residual 191. Filter status output information 192,
This is also technically the speech residual of the predictor filter 190 (s
Also referred to as "peek residual", it is sent to the pitch memory subtractor 162 for updating the filter memory. A short delay spectrum predictor 195, which is an LPC filter whose parameters are set by the input LPC parameter signal 126, is driven by the speech residual 191 to produce a synthetic digital speech output 196.
The same speech residual signal 191 is used for updating the adaptive codebook 155.

The synthesized voice 196 is subtracted from the digitized input voice 197 by the subtractor 235 to generate a digital voice remainder output signal 236. The speech residue 236 is supplied to the spectrum inverse filter 200 and the residual error signal 2 is supplied.
02 is generated. The output signal 202 is provided to the cascade weighting filter 205, and the output filter state information 206, 207 in combination with the signal paths 141 and 161 is cascade weighting filter 15 as previously described.
Used to update 0 and 175. Output signals 201 and 203, which are filter state information of the spectrum inverse filter 200, are output to blocks 145 and 170.
Used to update the spectral inverse filters 145 and 170 as described above with respect to.

FIGS. 3 and 4 are simplified block diagrams of adaptive codebook searcher 220. FIG. 3 illustrates a suitable configuration of adaptive codebook searcher 220 and FIG.
Shows a further improved configuration. The configuration of Figure 4 is preferred.

Referring now generally to FIGS. 3 and 4, the information in adaptive codebook 155 is the excitation or drive information from the previous frame. For each frame,
The drive information consists of the same number of samples as the original audio sampled. The codebook 155 is conveniently organized as a cyclic list so that a new set of samples is simply shifted into the codebook 155 to replace the previous sample currently in the codebook. New driving sample is long delay pitch predictor 1
90 provided by output 191.

When using the drive information from codebook 155, searcher 220 does not process the vectors in sets, ie, in subframes, and as disjoint samples. Searcher 220 treats the samples in codebook 155 as a linear array. For example, for a 60 sample frame,
Searcher 220 is sample 1 from codebook 155
~ Sample 60 is used to form the first set of candidate information, Samples 2 to 61 are used to form the second set of candidate information, and so on. This type of codebook search is often referred to as an overlapping codebook search. The invention does not relate to the structure and function of codebook 155, but to how codebook 155 is searched to identify the optimal codebook vector.

Adaptive codebook searcher 220 accesses previously synthesized pitch information 156 already stored in adaptive codebook 155 from output 191 in FIG. 2B, and each such set of information 156. Target drive 151 received from block 150 and drive 156 accessed from codebook 155 using
Minimize the error criterion between A scaling factor or gain index 222 is also calculated from each accessed set of information 156, which the information stored in the adaptive codebook 155 allows for changes in the dynamic range of the human voice or other input signal. Because I can't enter.

The preferred error criterion used is Least Squares Prediction Error (MPSE), which is produced at the original speech frame 115 from the frame memory output 117 and at the output of block 195 of FIG. 2B. It is the square of the error between the synthetic speech 196. The synthetic speech 196 is the trial drive information 156 obtained from the codebook 155.
Is calculated with respect to. The error criterion is Codebook 155
Driving information 1 for each candidate vector or set obtained from
The particular set of drive information 156 'that is evaluated for 56 and gives the lowest error value is the set of information used for the current frame (or subframe).

After the searcher 220 has determined the best set of driving information 156 'to be used with the corresponding best matching scaling factor or gain 222, the vector exponential output signal 221 corresponding to the best matching exponent 156'. And the best matching scaling factor 22
The scaling factor 222 corresponding to 2'is transmitted to the channel encoder 210.

FIG. 3 shows the adaptive searcher 2 according to the first embodiment.
20 shows a block diagram of 20 and FIG. 4 shows an adaptive searcher 220 'according to a further improved and preferred embodiment.
The adaptive searchers 220 and 220 'are the adaptive codebook 15
5, a vector index C ₁ (n) ... C _k (n) is sequentially searched. During a sequential search operation, the searchers 220, 220 ′ may search for each candidate excitation or drive vector C from the codebook 155.
Access _k (n), where k is an index from 1 to K that identifies a particular vector in the codebook, and n is another index from n = 1 to n = N,
N is the number of samples in a given frame. In a typical CELP application, K = 25
6 or 512 or 1024 and N = 60 or 120 or 240, but other values of K and N can also be used.

Adaptive codebook 155 includes a set of different pitch periods or periods determined from the previously synthesized speech waveform. The first sample vector starts from the Nth sample of the synthesized speech waveform C _k (N), which is N samples back from the synthesized speech waveform of the current last sample. For human voices, the pitch frequency is generally around 40 Hz to 500 Hz. This translates to about 200-16 samples. If fractional (f
rational) If pitch is included in the calculation,
K can be 256 or 512 to represent the pitch range. Therefore, the adaptive codebook contains a set of K vectors C _k (n), which are essentially one for a particular frequency.
Samples of one or more pitch periods.

Referring now to FIG. 3, the convolution generator 510 of the adaptive codebook searcher 220 provides a perceptually weighted LPC impulse response function for each codebook vector C _k (n), or signal 156. H (n), ie, convolve with the signal 152 from the cascade weighting filter 150. The output 512 of the convolution generator 510 is then cross-correlated with the target speech residual signal X (n) (ie, signal 151 of FIGS. 2A and 2B) in a cross-correlator 520. To be done. The convolution and correlation are for each codebook vector C
_{performed on k} (n), where n = 1, ..., N. The operation performed by the convolution generator 510 is mathematically represented by equation (1) below. The operation performed by cross-correlation generator 520 is mathematically represented by equation (2) below.

Output 5 of convolution generator 510
12 is also fed to an energy calculator 535, which has a squarer 552 and an accumulator 5.
53 (accumulator 553 provides the sum of squares determined by squarer 552). Output 554
Is transmitted to a divider 530, which calculates the ratio of signals 551 and 555.
The output 521 of the cross-correlator 520 is fed to a squarer 525, the output of which is also fed to a divider 530. The output 531 of the divider 530 is the peak selection circuit 57.
0, and the function of the peak selection circuit 570 is C
_It is to determine which value C _k (m) of _k (n) produces the best match, ie the maximum cross-correlation.
This is mathematically represented by equations (3a) and (3b). Equation (3a) represents the error E. Minimizing the error E is maximizing the cross-correlation represented by equation (3b) below, where G _k is defined by equation (4).

The identifier (index) of the optimal vector index C _k (m) is transmitted to the output 221. Peak selector 57
The output 571 of 0 is the best matching pitch vector C
_The gain scaling information associated with _k (m) is communicated to gain calculator 580, which provides gain index output 222. The operation performed by the gain calculator 580 is mathematically represented by the following equation (4).

The outputs 221 and 222 are sent to the channel coder 210. A convolution generator 510,
Cross-correlation generator 520, squarers 525 and 552 (these perform similar functions for different inputs), accumulator 553, divider 530, peak selector 570.
And means for providing gain calculator 580 are each well known in the art.

The configuration of FIG. 3 gives satisfactory results,
It requires more computation than necessary to perform the necessary convolution and correlation for each codebook vector. This is because both convolution 510 and correlation 520 must be done for all candidate vectors in codebook 155 for each speech frame 117. This limitation in the configuration of FIG. 3 is overcome by the configuration of FIG.

Adaptive codebook searcher 220 'of FIG.
Is the impulse perceptual weighting response function H (n) of the short-term LPC filter in convolution generator 510 'to generate convolution signal W (n) (ie, output 15 of block 150 of FIG. 2).
2) Use the perceptually weighted frame of the target speech X (n) (ie, signal 151 of FIGS. 2a and 2b) to convolve with 2). This is done only once per frame 117 of the input speech. This immediately reduces the computational load by a large factor that is approximately equal to the number of candidate vectors in the codebook. This is a huge computational savings. The operation performed by the convolution generator 510 is mathematically represented by equation (5) below.

The output 512 'of the convolution generator 510' is then correlated by the cross-correlation generator 520 'with each vector C _k (n) in the adaptive codebook 155. The operation performed by the cross-correlation generator 520 'is mathematically represented by the following equation (6).

The output 551 'is the square of the correlation of each vector _Ck (n) squared by the squarer 525' and normalized by the energy of the candidate vector _Ck (n).
1'is generated. This is by providing each candidate vector C _k (n) (output 156) to the autocorrelation generator 560 ', and the impulse of the filter for the autocorrelation generator 550' whose output is subsequently manipulated and combined. This is accomplished by providing the response H (n) (from output 152). Output 55 of autocorrelation generator 550 '
2'is a lookup table 5 whose function will be described later.
50 '. Output 556 'of table 555'
Is fed to a multiplier 543, where the autocorrelator 560 '
Output 561 of the.

The output 545 'of the multiplier 543' is supplied to the accumulator 540 ', and the accumulator 540'.
Adds the products of n to successive values and sums 54
1'is sent to the divider 530, where the cross-correlation generator 52
0's output 521 'combined. Autocorrelator 56
The operation performed by 0'is mathematically described by equation (7) below, and the operation performed by autocorrelator 550 'is mathematically described by equation (8) below. In this case, C _k (n) is the k-th adaptive codebook vector, each vector is identified by an index k ranging from 1 to K, and H (n) is a perceptually weighted LPC impulse response, N is the number of digitized samples in the analysis frame, and m is a dummy integer exponent, and n is an integer exponent indicating which of the N samples in the speech frame are considered.

The above search operation is performed for each candidate vector C _k.
(N) target speech residual X using MSPE search criteria
Compare with (n). Each candidate vector C _k (n) received from the output 156 of the codebook 155 is sent to an autocorrelation generator 560 'which generates all the autocorrelation coefficients of the candidate vector. To produce an autocorrelation output signal 561 ', which signal 561' comprises blocks 543 'and 540'.
35 '.

The autocorrelation generator 550 'generates all the autocorrelation coefficients of the H (n) function and outputs the autocorrelation output signal 55.
2 ', which signal 552' is provided to energy calculator 535 'via table 555' and output 556 '.

Energy calculator 535 'receives input signal 55
6 ′ and 561 ′ are all _{autocorrelation} coefficients of the candidate vector C _k (n) and cascade weighting filter 15
Combine by summing all product terms of the perceptually weighted impulse function H (n) generated by 0. Ernegi Calculator 535 'is a multiplier 54 for multiplying the autocorrelation coefficient of C _k (n) with the same delay terms (signals 561' and 552 ') of the autocorrelation coefficient of H (n).
3 ', and the output of the multiplier 543' is added and the divider 5
It comprises an accumulator 540 'that produces an output 541' containing information about the energy of the candidate vector sent to 30. Divider 530 'performs the energy normalization used to set the gain. Candidate vector C
_k Erunegi the all products of candidate vector C _{k (n)} of all the autocorrelation coefficients and the perceptually weighted short-term filter 150 perceptually weighted impulse function H (n) of _(n) It is calculated very efficiently by adding the terms. The above-described operation for determining the loop gain G _k is mathematically described by equation (9) below. In this case, C _k (n), X (m), H (n), Φ
_k (n), U _k (n) and N have been previously defined,
And G _k is the loop gain of the kth code vector.

The table 555 'enables the computational load to be further reduced. This is the impulse function H
The autocorrelation coefficient 552 'of (n) is the frame 1 of the input speech.
This is because it only has to be calculated once for every 17. This can be done one before the codebook search and the results can be stored in table 555 '. The autocorrelation coefficient 552 'stored in the table 555' prior to the codebook search is then followed by the adaptive codebook 155 '.
Is used to calculate the energy for each candidate vector. This allows to save even more computation.

The normalized correlation results of each vector in codebook 155 are compared in peak selector 570 'and vector C having the largest cross-correlation value.
_k (m) is identified by peak selector 570 'as the optimal pitch period vector. The maximum cross-correlation can be mathematically represented by equation (10) below. In this case G _k is defined in equation (9) and m is a dummy integer exponent.

The location of the pitch period, that is, the index of the code vector C _k (m) is the channel coder 210.
Is provided at output 221 'for transmission to.

The pitch gain is calculated using the pitch period candidate vector C _k (m) selected by the gain calculator 580 'to generate a gain index 222'. The means and methods described herein reduce the computational complexity with virtually no loss of voice quality. Vocoders using this configuration can be more conveniently implemented by a single digital signal processor (DSP) because of reduced computational complexity. The means and methods of the present invention can also be applied to other areas, such as speech recognition and speech identification, which use least squares prediction error (MPSE) search criteria.

The present invention is sometimes produced by the methods and apparatus described herein, sometimes with a target speech target.
Although described with reference to a perceptually weighted target speech signal X (n), referred to as speech residual), the method of the present invention is now described to obtain a perceptually weighted target speech signal X (n). It is not limited to the particular means and method used and can be used with target speech obtained by other means and methods and with or without perceptual weighting or removal of filter ringing.

The term "residual" as applied to "speech" or "target speech" includes the situation where the ringing signal of the filter is subtracted from the speech or target speech. Thinking. As used herein, the terms "voice residual" or "target voice" or "target voice residual" and the abbreviation "X (n)" are meant to include such variations. The same also applies to the impulse response function H, which can be a finite or infinite impulse response function and with or without perceptual weighting.
The same applies to (n). As used herein, the term “perceptually weighted impulse response function (perceptually weighted i.
mpulse response function ”
Or "filter impulse response (filter im
"pulse response)" and "H (n)"
Is intended to include such variations. Similarly, the word "gain inde
x) ”or“ gain scaling factor (gain)
scaling factor) ”and its notation G _k
Includes many forms that such "gain" or "energy" normalized signals will incorporate in connection with CELP coding of speech.

Even with the advantages offered by the embodiment shown in FIG. 4, a significant computational burden remains. For example, block 560 'of FIG.
Evaluation of autocorrelation coefficient in (see above equation (7))
To compute the energy normalization (gain) coefficients for the K vectors in codebook 155 (K)
• Requires (N!) Multiplication. K is typically 512
Or on the order of 1024 and N is typically 60
Or because it is on the order of 120 or 240, (K)
・ (N!) = (K) ・ (N) ・ (N-1) ・ (N-2)
(2) is usually a very large number. These calculations are best fit with what is required for the operation of blocks 510 ', 520', 550 'and the corresponding values of the particular adaptive codebook vectors C _{k = j} (n) and G _{k = j.} Needed in addition to the stochastic codebook vector and others needed to recursively determine the corresponding gain factor that gives the best fit (minimum error) of the target speech X (n) to the input speech. is there. This requires a significant amount of computational power to make the necessary calculations in a reasonable amount of time.

It has been found that the number of autocorrelation operations that need to be performed on a codebook with K vectors of N entries per vector can be significantly reduced without a significant negative impact on speech quality. This comprises autocorrelating a codebook vector for a first P (P << N) of N entries to determine a first autocorrelation value for it, K codebook vectors and said first Evaluating the codebook vector of K by generating the synthesized speech using the autocorrelation value of S and comparing the result with the input speech, which S (S << K) of the K codebook vectors of K Determines whether to provide the synthesized speech with less error compared to the input speech than the rest of the evaluated K-S vectors, of the K vectors for the R entries in each codebook vector. Autocorrelating a codebook vector (P <R ≦ N) for these S to provide a second autocorrelation value for it, using said second autocorrelation value Re-evaluating the S of the codebook vectors of S to identify which of the S codebook vectors provide the least error compared to the input speech, and of the codebook vector providing the least error. Forming the CELP code for the speech frame using the identification. For K and N of the magnitudes stated here, 5 ≦ P ≦ 10 and 1 ≦ S ≦
P and S in the range of 7 are suitable. R = N or N-
It is desirable that it is 1.

The operations described above can also be described with reference to the equations and figures provided herein. For example, m
= 0 to N−1, n = 1 to N, and k =
Instead of recursively evaluating equation (7) for each value of 1-k = K, the following procedure is used. (1) For m = 0 to m = P, where P << N, autocorrelate the codebook vector C _k (n) in block 550 ′ according to equation (7), and (2) detect it accordingly. Using the value of P of U _k (P) that has been calculated, recursively evaluate all K vectors C _k (n) and S of S of K vectors C _k (n), S << K , Selecting the one that provides the closest match to the input speech, and then (3) more S of the K vectors selected in step (2) above than the value of P originally selected, Recursively reevaluate using preferably all values of m = 0 to m = N−1 to determine U _k (m) in equation (7) and give the j-th best fit to the input speech. Determine the value C _{k = j} (n) and the corresponding gain index or factor G _{k = j} , and ( 4) As before, C _{k = j} (n) and G
Send _{k = j} to the channel coder 210.

As used herein, "recursive" is an analysis-by-synthesis codebook search and FIGS. 2 (a) and 2 (b) and FIG. It refers to the error minimization procedure described above.

It has been found that the output voice quality improves as P increases up to about P = 10, with a slight improvement over P> 10. Good voice quality is obtained for 5 ≦ P ≦ 10. The voice quality deteriorates rapidly for P <5. Since N is typically on the order of 60 or more, considerable computational savings are obtained.

It has been found that useful voice quality reduces the value of S to S = 1, and voice quality increases as S increases. Above about S = 7, further improvement in voice quality becomes difficult to detect. Therefore, 1
≦ S ≦ 7 is a useful operating range, which significantly reduces the number of calculations that must be accomplished during the recursive search for the optimal codebook vector and corresponding gain index or factor. This makes it easier to achieve the desired vocoder function using a single digital signal processor.

Yet another problem exists as to how codebook entries are constructed and autocorrelated. This occurs as a result of a procedure called "copy-up" that is often used in the prior art to provide for the identification of short pitch periods (see, for example, the aforementioned Ketchum et al. Patent). . This is explained as follows.

The energy term of the error function in the adaptive codebook search for the optimum pitch period can be reduced to a linear combination of the autocorrelation coefficients of the two functions. These two functions are the impulse response function H of the perceptually weighted short-term linear filter.
(N) and the codebook vector C _k (n) of the adaptive codebook. The computational complexity of the adaptive codebook is greater than that of the probabilistic codebook, because the autocorrelation coefficient for the adaptive codebook vector cannot be pre-computed and stored.

Each adaptive codebook vector is a linear array of N entries, also called samples or values. Each entry is identified by an index n ranging from 1 to N or N to 1. Adjacent vectors in the codebook differ from each other by one entry, that is, each subsequent vector has one new entry added at the end of the vector and one old entry dropped from the other end of the vector. The intervening entries have remain the same. Therefore, except for the ends of the vector, adjacent vectors have the same entry offset by one index. If adjacent vectors are placed side by side, they would be balanced if they were offset by one entry or sample. This is outlined below for a virtual adjacency vector k, k 'with an arbitrary entry value between 0 and 9 and an index n = 1-60. This displacement is referred to as the codebook "overlap".

Example I-Illustration of Vector Overlap k (n): 1,2,3,4,5,6,7, ..., 55,56,57,58,59,60 (Index) 4,6 , 9,3,5,1,8, ……, 0, 4, 6, 8, 2, 3 (value) k ′ (n): 1,2,3,4,5,6,7, …… , 55,56,57,58,59,60 (index) 6,9,3,5,1,8,5, ……, 4, 6, 8, 2, 3, 7 (value) Vector k ′ is It has the same entries as the neighbor vector k displaced by one exponent, and old entries are dropped from one end of the vector (eg, the value 4 is dropped on the left) and new entries are added to the pattern (eg, the value 7). Is added to the right edge).

The autocorrelation function U _k (m) is given by equation (7), where m = 0 to N−1 is the product C _k (n) * C
“lag” value in _k (n + m), and n = 1 to N is the index of the vector entry
Is. So far, the vector length N (ie the number of entries per codebook vector) and frame length L
(Ie the number of audio samples per analysis frame) has been assumed to be the same. But this is not always the case. Different strategies are used to determine the autocorrelation coefficient depending on whether N and L are the same or different.

If the vector length N is equal to or larger than the frame length L, the autocorrelation coefficient is added-deleted (ad
d-delete) It can be calculated by a process called end collection. For example, a zero-order or zero-delay (delay m = 0) autocorrelation coefficient, such as a subsequent adaptive codebook vector C _k , C _k ′ C _k ″, calculates the sum of (C _k (n)) ² for the first vector. And can be determined by detecting another vector with end collection, which requires adding the square of the newly added vector value and subtracting the square of the just deleted vector value. This same procedure is followed (with some changes) for m = 1, 2, 3, etc., so that the computational burden is calculated separately for each vector by equation (7) for each autocorrelation coefficient. This is reduced compared to the case of computing .. This add-delete end-correction process for determining the autocorrelation coefficient is well known in the art.

If the number of samples in the vector is less than the frame length L, it is common to "copy-up" the vector to fill the frame (eg, Ketchum et al., Supra, US). See patent). For example, if the frame length is 60 and only 20 entries are used in the analysis, these 20 entries are repeated 3 times to get 60 vector lengths. This is explained below in terms of vector-valued exponents.

Illustrative Example II- Copyup Vectors 1,2, .................. 59,60 Copyup Vectors 1,2, ..., 19,20,1,2, ..., 19,20,1,2, ..., 19,20.

This duplication or "copy-up" results in an error if one tries to use the previously mentioned add-delete end correction method for calculating the autocorrelation coefficient. These errors degrade the quality of the synthesized speech.

The error of the end correction increases for larger values of m, ie higher orders (larger "lag" terms) in the autocorrelation function. The simple add-delete end correction procedure described above no longer works satisfactorily for copied-up vectors. Thus accepting poorer voice quality due to having a smaller computational burden (eg easy end-collection), or higher voice quality and greater computational burden (eg computing each vector separately). There will be some unwanted choices left. It has been found that the computational burden of obtaining the autocorrelation coefficient for situations where the number of samples in the vector is less than the frame length can be reduced without loss of synthesized speech quality by the improved computational procedures and apparatus described below. It was

A codebook in which the analysis frame has lengths L (eg, 60) and N samples or values (eg, 60) is used to determine the adaptive codebook vector that produces the best match for the target speech. 2 to 4 are assumed to be used for the apparatus and procedure. In addition, a smaller subset of vector values M <N (eg M
~ 20) are initially used for the analysis. In the past, M samples or values were copied up to fill a frame of length L and the analysis was based on the copied up frames. According to the method of the present invention, it is not necessary to copy up a subframe of M values.

The description given in connection with this embodiment is specifically directed to the efficient determination of the autocorrelation coefficient of the adaptive codebook vector, and has a smaller error and the best match to the target speech. Reference should be had to FIGS. 2-4 for a description of the other parts of the analysis process used to select the codebook vector.

Also, reference should be made to the equation (7), and in this case, n = 1 to N of the product [C _k (n) * C _k (n + m)].
And the sum U _k (n) over m = 0 to N−1 is the autocorrelation coefficient of the kth vector. The index m is usually 0 to N-
Identify the "lag" that is to be 1 and that is used to calculate the autocorrelation coefficient. An index k in the range 1 to K identifies a codebook vector and an index n indicates an individual sample or value within the vector. The number of samples used in the analysis depends on the pitch period detected. For example, about 20 samples are required for the shortest pitch period associated with the human voice and about 147 samples are required for the longest pitch period.

The zero-order autocorrelation coefficient corresponds to m = 0, and 1
The next coefficient corresponds to m = 1 and so on. The "pitch lag" M <N is defined as the number of values in the vector to be used for analysis. Therefore, m varies from zero to M in determining the autocorrelation coefficient for the voice component of the short pitch period. "Frame size"
L is defined as the number of audio samples in the frame. Usually L = N. A typical value for L is 60
And a typical value for M is 20, but L <
If M, other values can be used for both of them. For convenience of description, values of L = 60 and M = 20 are assumed in the following description. However, those skilled in the art will understand based on the description provided herein that these values are not limiting and that other values for M and L can be used.

The present invention provides means and methods for reducing the computational burden of determining autocorrelation coefficients and avoiding copy-up errors. It requires a limited number of codebook samples (eg, 20) for which copy-up was previously used, i.e. to quickly identify the shortest pitch period, but that limited number of samples is It applies to the part of the analysis procedure with recursive synthesis, which has to be extended to the analysis frame length (eg 60) to avoid energy normalization problems.
Once the first M + k-1 vectors have been analyzed and the vector expansion is complete, resulting in N = L, the autocorrelation coefficient is calculated by the add-delete end correction process described above.

In the preferred embodiment, the method of the present invention self-phases for the first vector k by evaluating (7) for (1) m = 0 to T <M and n = 1 to M. Determine the relation number U _k and multiply the result by L / M. In this case L, M, P, n and m have the meanings given above. For L = 60 and M = 20, L
/ M = 3. The parameter T determines how many values of the autocorrelation delay m are used, ie how many autocorrelation coefficients are calculated. Typically T = M-1, but other smaller values of T can be used. Using a smaller value of T is advantageous when the dominant values of the codebook vector are calculated and the dominant autocorrelation coefficient is for small values of m. (2) Find the sum of the products in equation (7) for each value of m previously obtained in step (1) and add (C _k ′ (n = M + 1)) ² to the term m = 0. Then C
_k '(n = M + 1) * C _k ' (n = M + 2) is added to the term of m = 1, and C _k '(n = M + 1) * C _k ' (n = M +
3) is added to the m = 2 term, and so on until the Tth term is similarly determined, the autocorrelation coefficient U _k for the second vector k ′ is determined, and the result is L / (M +
1) and (3) sum the products for each value of m previously obtained in step (2) and add (C _k ′ (n = M + 2)) ² to the term m = 0. To C
_k ″ (n = M + 2) * C _k ″ (n = M + 3) is added to the term of m = 1, and C _k ″ (n = M + 2) * C _k ″ (n = M +
4) is added to the term of m = 2, and the same procedure is performed up to the T-th term to determine the autocorrelation coefficient U _k for the third vector k ″, and the result is determined by L / (M + 2). Multiply, and (4) steps (1)-
L / (M + k−1) = (3) for the remaining vectors
1 for each additional vector until 1
Increment by only to determine the remaining autocorrelation coefficient. afterwards,
The autocorrelation coefficient is calculated by the traditional prior art add-drop procedure described above.

Alternatively, the autocorrelation coefficient of the codebook vector can be calculated using the following equations (11a)-(11b) for the coefficient U of the first vector k = 1 for m = 0-T <M.
It is determined by calculating _k (m). U ₁ (m) = (L / M) U ′ ₁ (m) (11b)

The determination of the autocorrelation coefficient of the above codebook vector then remains for m = 0-T <M and (M + k-1) ≤L using the following equations (12a)-(12b). It is determined by calculating the autocorrelation coefficient U _k (m) of the codebook vector of _U'k (m) = [ _U'k-1 (m) + _Ck (M + k-1) _Ck (M + k-1 + m)] (12a) _Uk (m) = {L / (M + k-1)} U ′ _K (m) (12b)

Analysis by synthesis uses vectors of increasing length (and their corresponding autocorrelation coefficients), starting with a vector of length M and until the vector length equals the frame length, ie ( This is done by increasing the length of each subsequent vector by one sample until M + k-1) = L. The extension of the short pitch samples to match the frame length is then complete. Subsequent vectors have the same length as the frame length and each successive vector in the overlapping codebook corresponds to removing old samples from one end of the vector and adding new samples at the other end of the vector. ..
Prior art add-delete end correction methods are then used to determine the autocorrelation coefficients of the remaining vectors to be analyzed.

The sum of products in equation (11a) need only be evaluated once for the first vector and then (M
Other vectors up to + k-1) are C _k for additional values or samples included from the first vector term.
It can be calculated by adding the product contributions of * C _k .
No copy-up procedure is required and the auto-correlation coefficient error produced by copy-up does not occur. This substantially reduces the computational burden in the synthetic analysis procedure described with respect to FIGS.

The differences between the prior art copy-up procedure and the procedure of the present invention are outlined below with respect to the vector index. The calculation of the autocorrelation coefficient involves adding the product of the vector with itself to the delay m, i.e. various amounts of relative displacement of the vector. The following example shows which values are multiplied together for various amounts of lag m for the copy-up method and the method of the present invention. The numbers in this example are vector values or exponents of entries, not the values themselves, but can be thought of as a measure of the position of each entry along the vector.

Illustrative Example III-Copy-Up Autocorrelation For copy-up, we multiply by terms and add for each n and m. For example, for (k = 1, m = 0), 1,2,3, ..., 19,20,1,2,
3, ..., 19,20,1,2,3, ..., 19,20 to 1,2,3, ..., 19,20,1,2,
Multiply by 3, ..., 19,20,1,2,3, ..., 19,20 ;, for (k = 2, m = 0), 1,2, ..., 19,20,21,1,
2, ..., 19,20,21,1,2, ..., 17,18 to 1,2, ..., 19,20,21,1,
Multiply by 2, ..., 19,20,21,1,2, ..., 17,18; and for (k = 3, m = 0) 1,2, ..., 19,20,21,22,
1,2, ..., 20,21,22,1,2, ..., 15,16 to 1,2, ..., 19,20,21,2
2,1,2, ..., 20,21,22,1,2, ..., 15,16 ;, and so on for all k, m and n.

Illustrative Example IV-Improved Autocorrelation (m = 0) For the configuration of the present invention, m = 0 to M-1 is multiplied and added by the first (eg, 20) entry, Then, products of entries such as n = M + 1 and n = M + 2 are added next. For example, for (k = 1, m = 0), 1,2,3, ..., 19,20 × 1,
Calculate 2,3, ..., 19,20 and multiply by L / M, and for (k = 2, m = 0), 1,2,3, ..., 19,20 ,twenty one
× 1,2,3, ..., 19,20,21 is obtained by adding 21 · 21 to the previously calculated value for k = 1, and L / M +
Multiply by 1, and for (k = 3, m = 0), 1,2,3, ..., 19,20,21,2
2 × 1,2,3, ..., 19,20,21,22 is obtained by adding 22 · 22 to the previously calculated for k = 2, and multiplied by L / M + 2, Then continue for all m until the vector length equals the frame length and the last term 60.60 has been added, then proceed as in the prior art.

In the above example of the autocorrelation process for the prior art technique and the technique of the present invention, only the 0th order term was shown, but one skilled in the art will know based on the description herein that m = 1, m =
2, we will see how to shift the vector to represent the product term for others. As an aid to that process, the following examples are presented for the present invention for k = 1, k = 2 and m = 1. Example V-For improved autocorrelation (m = 1) (k = 1, m = 1), 1,2,3, ...,. 19,20 x
Calculate 1,2,3, ..., 18,19 and multiply by L / M + 1, for (k = 2, m = 1), 1,2,3, ..., 19 , 20,21
X 1,2,3, ..., 19,20 is obtained by adding 20 · 21 to the one previously calculated for k = 1, and L / M
Multiply by +2, and for (k = 3, m = 1), 1,2,3, ..., 19,20,21,2
2 × 1,2,3, ..., 19,20,21 is obtained by adding 21 · 22 to the previously calculated for k = 2 and then multiplied by L / M + 3 and evaluated All k done
And m until L / (M + k-1) = 1.

An apparatus suitable for determining the autocorrelation coefficient as described above according to a preferred embodiment of the present invention is shown in FIG. Autocorrelation device 6 corresponding to the present invention
00 comprises a signal input 602 whose vector samples C _k (n) are received from the adaptive codebook 155 of FIG. The vector sample or value C _k (n) follows two paths 604,606. Path 606 passes through switch 608 the initial vector (ie, k = 1) autocorrelator 6
Up to 10. The initial vector autocorrelator 610 achieves the function illustrated by equation (11a) above, ie, it has an autocorrelation coefficient corresponding to k = 1, m = 1, 2, 3, ..., T-1, T. Calculate U ₁ (m). These autocorrelation coefficients are transmitted to the end correction coefficient calculator 622 via the switch 620.

First vector autocorrelation coefficient calculator 61
0 comprises registers 612 and 614, which contain the first M (eg 2
0) sample is loaded. Registers 612, 61
4 is a well known serial input / parallel output register for convenience, but other configurations well known in the art can be used.

The sampled values are transferred to an autocorrelator 616, which in turn multiplies the product U for m = 0.
₁ (m) = SUM [C ₁ (n) C ₁ (n + m)] (ie, U ₁ (0)) is determined and this coefficient is clocked into block 622 via switch 620.
The autocorrelator 616 then shifts the sample in register 614 by one sample and computes U1 (1), corresponding to m = 1, through block 618, which is then clocked out to block 622. . This procedure continues until all autocorrelation coefficients for the initial vector C ₁ (n) have been determined and loaded into block 622. Switches 608 and 620 then disconnect autocorrelation generator 610 from clock 622.

Block 622 performs the functions described by equations (11b) and (12a) through (12b) above. For convenience, this is a register 624, a multiplier 626, an adder 628, a register-accumulator 63.
0, generated by the combination of multiplier 632 and output buffer 634. Registers 624, 630 and buffer 634 are expediently of the same length as registers 612, 614 (shown in FIG. 5, for example), but how many autocorrelation coefficients are evaluated and updated for subsequent vectors. It can be longer or shorter depending on what is desired. For example, registers 624, 6
30 and buffer 634 can be as large as the frame length.

Register element 630 contains the autocorrelation coefficient calculated before which the end correction should be added to determine the autocorrelation coefficient for the subsequent vector. The end correction is provided by register 624 associated with multiplier 626. Multiplier 6
The end correction from 26 is added with the previously calculated coefficient from register 630 in adder 628,
And is fed back via loop 629 to update register 630. From the register 630, the autocorrelation coefficients are transferred to a multiplier 632 where they are scaled by an appropriate factor of L / (M + k−1) and sent to an output buffer 634 where they are, for example, The output 561 'of FIG. 4 is formed, in this case the autocorrelation generator 600 is the element 56
0'is described in more detail for (M + k-1) ≤L.

To describe the operation of block 622 in more detail, register 624 is loaded simultaneously with registers 612 and 614 by vector value. Register 630
Is the output U of the first vector autocorrelation coefficient generator 610 before the autocorrelator 610 is disconnected from the block 622.
Loaded by ₁ (m). These initial autocorrelation coefficients are copied to multiplier 632 where they are L / M.
Are sent to a buffer 634, from which they are extracted during the analysis procedure by synthesis described with respect to FIGS.

After register 630 is loaded with the first T autocorrelation coefficient value, additional vector values are clocked into register 624 and the vector value of each stage of register 624 is clocked as indicated by arrow 625. Is output. Assuming the initial vector has a value of M, the most recent value now in register 624 is n.
= M + 1. This corresponds to the vector k = 2,
It means that each vector is the same as the previous vector, n = (M + k-1)
This is because one entry is different up to = L.

The new value n = M + 1 is multiplied by itself in the multiplier 6261 and the result is the adder 6281.
Where it is transmitted to it is already U _{1 (m} = 0) of the 0th order stored in the register element 6301 combined with coefficient. Register element 6301 is then updated as indicated by arrow 6291, which causes U ₁ (0) + C _k (M +
1) The sum of C _k (M + 1) is now the register element 6301
Is present at L / (M +
1) and loaded in buffer 634 with other updated coefficient values from other elements of register 630 multiplied by 632 by the same factor. A counter 640 is provided to keep track of the number of entries in the codebook vector loaded into register 624 and adjust the multiplication factor in multiplier 632 so that it is L / (M) for k = 1. Correspondingly,
For k = 2, it corresponds to L / (M + 1), and similarly up to (M + k-1) = L.

Sample C _k (M) from register 624
Is multiplied by C _k (M + 1) in multiplier 6262 and register element 6302 in adder 6228.
From U ₁ (1) from which the sum updates register element 6302 via connection 6292. The updated value is sent to multiplier 632, where it is multiplied by L / (M + 1) and sent to buffer 634. Register 62
The remaining samples in 4 are processed in a similar manner and then another sample, for example n = M + 2, is registered in register 62.
4 is clocked in and the process is repeated. In this way, the autocorrelation coefficient is
Obtained in buffer 634 for each new vector formed by adding the two samples, this being done in a manner similar to that shown in simplified form in Examples IV-V.

Temporary storage elements 612, 614, 624,
Although 630 and 634 are described as registers or buffers, one of ordinary skill in the art, based on the disclosure herein, is for convenience of illustration only and is not limited to, but is not limited to, random. It will be appreciated that other types of data storage such as accessible memory, content addressable memory, etc. may be used. Further, such memory can be in a wide variety of physical configurations, such as flip-flops, registers, cores and semiconductor memory devices. The terms "register" and "buffer" as used herein are intended to include any modifiable information store, whether singular or plural, of any kind or configuration. .. Similarly, for example, an autocorrelator 616, an indexer 618, a switch 608,
620, adder 628, multiplier 626 and / or other blocks shown as counter 640, whether discrete elements or combinations of elements, standard or application specific integrated circuits, are programs capable of achieving the desired functionality. General purpose processor, or any combination thereof, either separately or in combination, is intended to be included.

The present invention provides a shorter set of codebooks needed to detect short pitch periods without introducing the traditional copy-up errors that occur in extending a small number of codebook samples to standard frame lengths. It provides a quick and easy way to determine the autocorrelation coefficient for a standard analysis frame length (eg 60) based on vector samples (eg 20). The computational burden is reduced without sacrificing voice quality because the end autocorrelation add-delete errors associated with prior art copy-up configurations are avoided. Copy-up is completely avoided.

Although the method of the present invention for generating an autocorrelation coefficient has been described above in connection with hardware registers, autocorrelators, multipliers, adders, switches, etc., those skilled in the art will understand that these are software. And configuring a computer to carry out the method of the invention based on the detailed description of the embodiments described herein, which can be realized by the following, thereby achieving the same functions as described herein for the device. It will be understood that, and such variations are within the scope of the invention.

Although the improvements described above significantly reduce the computational burden associated with determining the optimal codebook vector for replicating the target speech, further improvements are desired. In particular, improvements are desired in the manner in which the optimal vector of the probabilistic codebook is identified.

US Pat. No. 4,797,925, Lin, describes a procedure that reduces the computational burden of considering all vectors in a stochastic codebook by using overlapping stochastic vectors. . According to Lin's construction, each subsequent vector in the codebook differs from the preceding vector by the old value being dropped from one end of the vector and the new value being added to the other end of the vector. With this configuration, the number of unique values in the codebook of 1024 vectors, each having 60 values, is 1024 × 60.
= 61,440 to 60 + 1023 = 1,083. Even then, a large number of computations are still required to perform the analysis and the steps are time consuming as they involve subsequent multiplications and additions.

Probabilistic codebook 180 ((b) of FIG. 2)
Contains a vector S _k (n) of K of length N, where k = 1 to K and n = 1 to N, and K is conveniently 512, 1024, 2048, and so on, Typically 1024, and N is conveniently 20,40,6
0, 120, etc., and typically 60. The indices k and n for the stochastic codebook vector S _k (n) have the same meaning as for the adaptive codebook vector C _k (n), ie k identifies which vector is considered and n is Identify the values considered in the vector k. The exponential bounds K and N for the vectors of the stochastic codebook 180 are adaptive codebook 1
It is convenient to have the same magnitude as the exponential limits K and N for the 55 vectors, but this is not essential. For convenience of explanation only and not by way of limitation, K and N may have the same value for both codebooks, eg K = 1024 and N.
= 60.

Vectors in the probabilistic codebook 180 are conveniently pseudo-random 0 and 1 or 0,1.
And a linear array of -1. That is, each vector S _k (n) is a concatenation of N values, each value identified by an index n. FIG. 6 shows a codebook 180.
An exemplary ternary (e.g., 0,1, -1) stochastic codebook 180 'with K = 8 and N = 20, similar to
Indicates. Those of ordinary skill in the art will understand, based on the description herein, how the characteristic features of the codebook of FIG. 6 apply to larger values of K and N. further,
Although FIG. 6 shows a ternary (eg, 0,1, −1) codebook, a binary (eg, 0,1 or 0, −)
1) or other types of codebooks can also be used. Three
A hex codebook is preferred.

The vector S _k (n) of FIG. 6 for each successive value of _k overlaps by N-2. For example,
Vector _{S k =} 2 (n) adds two new values at the right end of the vector _{S k =} 1 from (n) to drop two old value from the left end of the vector _S 1 (n) and the vector _S 1 (n) It depends on things. Therefore, the value of vector S ₂ (n) is shifted by two places to the left compared to vector S ₁ (n) and there are two new values at the right end. Each successive vector is as different as the previous vector. Overlap amount, for example N in FIG.
The selection of -2 is convenient, but not essential. Any value of overlap can be used, for example 1-N-1. Also, although the vector has been described as being shifted to the left by adding a new value to the right, the reverse method can also be used, i.e. shifting to the right and adding the new value to the left. You can also

The analysis procedure for identifying the optimal stochastic codebook vector is essentially the same as for the adaptive codebook vector, except that S _k (n) is used instead of C _k (n), That is, the codebook 155
Codebook 180 is used instead of and the perceptually weighted short delay target speech signal X (n) (FIG. 2).
Instead of (a) to (b) 151 of FIG. 2), the perceptually weighted short and long delay target speech signal Y (n) (176 of (a) and (b) of FIG. (See) is used. The following equations (1 '), (2'),
(5 ') and (6') are similar to equations (1), (2), (5), and (6) given above, respectively, but instead of those previously described for the adaptive codebook. Appropriate variables for probabilistic codebooks are used in.

The important difference between the probabilistic codebook and the adaptive codebook is that the vectors that make up the probabilistic codebook 180 do not change as a result of the synthetic analysis process as occurs in the codebook 155 and are fixed. ing. Therefore, many of the calculations represented by equations (1 ') through (6') can be done once per frame and the results stored and reused. For example, autocorrelation of a stochastic codebook vector only needs to be done once, because the result is invariant. The autocorrelation coefficient is conveniently stored in a look-up table and does not need to be calculated again. This greatly simplifies the computational burden.

The process required to determine which of the probabilistic codebook vectors best represents the target speech is the equations (1 '), (2') of the values of the probabilistic codebook vector S _k (n). , (5 '), (6'), it has been found that it can be substantially simplified and faster by eliminating the multiplication by other signals nominally required. While the means and methods of the present invention are most usefully applicable to cross-correlation operations involving stochastic codebook vectors, they are also applicable to convolution operations involving stochastic codebook vectors. For convenience of explanation, the configuration of the present invention describes a correlation operation or operation, but those skilled in the art will understand how based on the description herein how it can be applied to a convolution operation.

Cross-correlation is achieved in the first embodiment by a multiplexer-accumulator combination, the select lines of the multiplexer being driven by the codebook or one or more replicas of the codebook. This will be explained in more detail with reference to FIGS.

FIG. 7 is a simplified block diagram of a stochastic codebook cross-correlator 700 according to the present invention. Correlator 700 is shown for the ternary (eg, 0,1, -1) codebook case. Those skilled in the art will understand based on the description herein that the present invention is equally applicable to binary and other forms of codebooks. The procedure described below can also be used to convolve the codebook vector with other signals.

Correlator 700 has an input 701, where signal (s) 702 are to be cross-correlated with the codebook vector, 702, but not in a limiting sense, eg, signal W '(from equation (5')). n), or another signal to be correlated with the codebook vector S _k (n). The signal 702 received at input 701 is generally a vector having N values identified by some index, for example n or m ranging from 1 to N. For example, if equation (6 ') is evaluated or estimated, W' (n) appears at input 701. If equation (1 ') is estimated, H (n-m + 1) appears at input 701. The arrangement of the invention is particularly useful for voice vocoders, but it can also be used for any signal or similar type of signal string.

For convenience of explanation, the means and method of the present invention will be described in terms of evaluation of equation (6 '), but one of ordinary skill in the art will understand that one vector or vector array , For example, but not limiting, 1,0 or -1,0 or -1,
It will be appreciated that it is applicable to any other sum of the products of two vectors or vector arrays that have fixed values, such as 0, 1 and the other is variable. The evaluation of equation (6 ') produces a single cross-correlation value Q (k) for each value of index k, ie To occur.

Vector signal 70 applied to input 701
2 (eg W '(n)) is the multiplexer 704, 7
05. Multiplexer 704 is shown in more detail in FIG. 8 and multiplexer 705 is substantially the same. The multiplexer 704 has a memory 706, for example, “1” in the codebook 180.
ROM or EPR with non-zero entries corresponding to
OM is connected. FIG. 9 shows the contents of memory 706 ', which is similar to memory 706, but with K = 9 and N =
20 and corresponds to the contents of the codebook 180 'of FIG. The indices k and n have the same function with respect to memory 706 (and memory 707) as in codebook 180, that is, k identifies a vector or another data string corresponding to the vector, and n
Identifies the values in the vector or string. Memories 706 and 706 'have zeros except where ones appear in codebooks 180 and 180' (compare FIGS. 6 and 9). The output of memory 706 is connected to select line 708 of multiplexer 704, thereby controlling the particular select line n with which each value k, n acts on the value of the vector provided at input 701.

Connected to the multiplexer 705 is a memory 707 similar to the memory 706 but having a non-zero entry corresponding to -1 in the codebook 180. FIG. 10 shows the contents of memory 707 'which is similar to memory 707, but with K = 8 and N = 20, and which corresponds to codebook 180' of FIG. Memory 707,
707 'has a 1 everywhere a -1 appears in the codebook 180, 180' and a zero elsewhere (compare FIGS. 6 and 10). Memory 70
The output of 7 is connected to select line 709 of multiplexer 705, thereby controlling the particular select line n with which each value k, n acts on the value of the vector provided at input 701.

The memories 706 and 707 are controlled by the address sequencer 714. The first (ie, k =
1) When the signal vector 702 is applied to the input 701 to be correlated with the codebook vector, the sequencer 7
14 accesses the k = 1 data set of the memories 706, 707 and outputs the values n = 1 to n = N for k = 1 to the corresponding multiplexers 7 on the select lines 708, 709.
04,705. The values appearing on select lines 708 and 709 pass through multiplexers 704 and 705 when the appropriate values of input vector 702 are accumulators 712 and 713.
And the values are added in the accumulators 712 and 713 to generate outputs 716 and 717. Outputs 716,717 are combined in combiner 720 to provide a first cross-correlation, Q (1), at output 721.

The sequencer 714 then proceeds to the memory 706, 7
The value of k = 2 at 07 and the values of n = 1 to N therein for k = 2 are multiplexers 704, 70.
5 select lines 708 and 709, and so on, to produce a second cross-correlation, Q (2), at output 721. This process is repeated until the input vector signal 702 for a speech frame is correlated with the codebook vector represented by the entries in memories 706, 707 to obtain the cross-correlation values Q (1), ..., Q (K). Probabilistic vectors of index k = j with large values of Q (k = j) usually give better phonetic representation than other vectors k = i with smaller values of Q (k = i).

Although it is convenient to use two memories 706, 707 for a ternary codebook, it may be possible to use more or less memory depending on the type of coding used in codebook 180. it can. For example, for a binary codebook it is necessary to use only one memory, and the codebook itself has a value of 0,1 for each index k, which corresponds to n = 1 to N. If it can be transmitted to the select line of the multiplexer, it is sufficient as a memory. Therefore, in the case of a binary codebook or its equivalent, no separate memory is needed and the codebook itself can be used to feed the select lines of the multiplexer.

Next, referring to FIG. 8, the multiplexer 70
The operation of No. 4 will be described. The configuration and operation of the multiplexer 705 are the same. The multiplexer 704 is generally
, GN, N × N multiplexer with N gates 715. Each gate 715
Input 701 to receive a particular value of input signal vector 702 (identified by index n).
And the other input 703 is connected to the system logic zero reference level, eg, ground. Gate 715 outputs 710 as input 701 (ie, signal 702) or input 703 (ie, as determined by the logic signal present on select line 708).
"0"). For the configuration shown, for example, a value of 1 on line n = i of select line 708 produces a value of n = i of input vector 702 (appears on line n = i of input 701), n = of output 710. to be transferred to line i, otherwise a value of 0 is transferred. Any equivalent logic configuration with similar results can be used.

Multiplexer 704 receives N input signal values 702 at input 701 and N on select line 708.
71 of the input signals 702 to N depending on whether the select line 708 driven by the memory 706 is set to 0 or 1
It is possible to transfer to 0. Multiplexer 705
Is similar for input 702, select line 709 driven by memory 707 and output 711, except that multiplexer 705 changes the value of input vector signal 702 at input 701 to −1 if the codebook vector value is −1. Output 711 for indices k and n
While the multiplexer 704 passes the input vector value 702 to the output 710 for the indices k, n if the codebook vector value is +1.

Outputs 710 and 711 are connected to accumulators 712 and 713, respectively, and input vector signal values 702 transferred through multiplexers 704 and 705 are added together to produce Q ⁺ (k) and Q ⁺ , respectively. ^- (K) Generate outputs 716,717 corresponding to the correlation values. The outputs 716 and 717 are combined in combiner 720 to produce a correlated output value Q (k) at 721. If the codebook 180 is a ternary codebook, as in this example, then the combiner 720 will be the accumulator 71.
The difference between outputs 716 and 717 from 2, 713 is output 7
21, ^{i.e., Q (k) = Q +} (k) -Q - generating a (k). This is multiplexer 705, memory 707.
Taking into account that the operations performed by and accumulator 713 correspond to a value of -1 in codebook 180, eg, see FIG. Combiner 72
0 performs subtraction in this particular embodiment, but those skilled in the art will understand that the same result can be obtained by many other means based on the description herein. For example, this is not meant to be limiting, but the same output 72
1 is a multiplexer 705 or an accumulator 713
Can be obtained by inverting the output of and combining combiner 720 as an adder.

Correlation generator 700 of FIG. 7 corresponds, for example, to correlation generator 520 or 520 'of FIGS. 3 and 4, and output 721 of correlation generator 700 is output 5 of FIG.
21 or output 551 'of FIG. 4, but depending on what particular input signal vector is processed, the stochastic codebook vector S _k (rather than the adaptive codebook vector C _k (n). n) and for the target audio signal Y (n) rather than X (n).

Another embodiment of the present invention will be described with reference to the above equation (6 ″) and FIGS. 6, 9 and 10. The equation (6 ″) is applied to the codebook 180 ′ of FIG. Thus, as shown in Table I below, n = 1-2
0 and the correlation values Q (1) to Q (w) for the value of W '(n)
(8) is obtained. Table I Q (1) = + W '(04) -W' (05) -W '(09) + W' (14) -W '(18) + W' (19) Q (2) = + W ′ (02) -W ′ (03) -W ′ (07) + W ′ (12) -W ′ (16) + W ′ (17) Q (3) =-W ′ (01) -W ′ (05 ) + W ′ (10) -W ′ (14) + W ′ (15) + W ′ (20) Q (4) =-W ′ (03) + W ′ (08) -W ′ (12) + W ′ (13) + W ′ (18) -W ′ (19) Q (5) =-W ′ (01) + W ′ (06) -W ′ (10) + W ′ (11) + W ′ (16 ) -W ′ (17) Q (6) = + W ′ (04) -W ′ (08) + W ′ (09) + W ′ (14) -W ′ (15) -W ′ (19) Q ( 7) = + W ′ (02) -W ′ (06) + W ′ (07) + W ′ (12) -W ′ (13) -W ′ (17) Q (8) =-W ′ (04) + W '(05) + W' (10) -W '(11) -W' (15) + W '(20) The array of Table I was rearranged to produce the Q array as shown in Table II below. ^{(k) = [Q + (} k)] - is - ^{[(k) Q]} grouped term corresponding to the term and codebook values of -1, corresponding to the codebook value of +1 to represent a correlation value as be able to.

Table II Q (1) = [W ′ (04) + W ′ (14) + W ′ (19)]-[W ′ (05) + W ′ (09) + W ′ (18)] Q (2) = [W ′ (02) + W ′ (12) + W ′ (17)]-[W ′ (03) + W ′ (07) + W ′ (16)] Q (3) = [W ′ (10) + W ′ (15) + W ′ (20)]-[W ′ (01) + W ′ (05) + W ′ (14)] Q (4) = [W ′ (08) + W ′ (13) + W ′ (18)]-[W ′ (03) + W ′ (12) + W ′ (19)] Q (5) = [W ′ (06) + W ′ (11) + W ′ (16)]-[W ′ (01) + W ′ (10) + W ′ (17)] Q (6) = [W ′ (04) + W ′ (09) + W ′ (14)]- [W ′ (08) + W ′ (15) + W ′ (19)] Q (7) = [W ′ (02) + W ′ (07) + W ′ (12)]-[W ′ (06) + W ′ (13) + W ′ (17)] Q (8) = [W ′ (05) + W ′ (10) + W ′ (20)]-[W ′ (04) + W ′ (11) + W '(15)] The value shown in the leftmost bracket of Table II corresponds to the input to the accumulator 712, and the value in the rightmost bracket of Table II corresponds to the input to the accumulator 713. .

With reference to courtbook 180 'of FIG. 6, it is clear that the codebook is sparsely filled, that is, most entries are zero.
Further, referring to Tables I and II, it is clear that the overlap characteristics of successive vectors are reflected in the exponent of the value of W '(n) added to obtain the correlation value Q (k). . Therefore, the structure of this codebook is suitable for a more economical way of generating the sums shown in Tables I and II. These will be described below.

Than storing all codebook values
Rather, a non-zero entry exponent for each value of k
Only (ie, the value of n) can be stored. This
This is Q ⁺(K) and Q⁻Perform separately for the value of (k)
But this is not mandatory. k
Correlation value Q for each value of⁺(K) and Q⁻(K)
Is simply W ′ corresponding to the stored value of n for each value of k.
By adding (n), i.e. Table I or I
It is obtained by performing the sum shown in I.

Computational and / or address storage requirements can be further reduced and faster operation can be obtained by using a recursive calculation method that takes into account the overlapping properties of codebook entries. According to the preferred method, the exponent value n of the codebook entry for k = 1 is stored and the exponent n for the vectors k = 2, k = 3, etc. is k based on the codebook overlap.
Calculate from the index value of = 1. The exponents of any new codebook entries added to the end of each vector are also considered.

For example, for the +1 entry of FIG. 9 and the Q ⁺ (k) portion of Table II (ie, the leftmost parenthesized quantity), n = 4, 14, 19 and codebook overlap. And compute the contribution to Q (k) resulting from the value of W ′ (n), which in this case is N−2 (ie Δk = + 1, Δn = −2) and is: . That is, the indices of k = 1, n = 4 are evaluated first and contribute to the values of Q ⁺ 1 (1) and Q ⁺ (2) in the following terms. Table III Q ⁺ (1) = W '(04) Q ⁺ (2) = W' (02) The Q (2) terms W '(02) for the indices k = 2, n = 2 are codebook overlaps ( Δk = + 1, Δn =-
2) is applied to the first index k = 1, n = 4.

The indices k = 1, n = 14 are then evaluated and the additional terms W '(14), W' (12), W '(1
0), W '(08), W' (06), W '(04),
Give W '(04) and W' (02). W '(1
All terms except 4) are determined by applying the codebook overlap to the starting indices k = 1, n = 14. This results in the following: Table IV Q ⁺ (1) = W '(04) + W' (14) Q ⁺ (2) = W '(02) + W' (12) Q ⁺ (3) = W '(10) Q ⁺ (4) = W '(08) Q ⁺ (5) = W' (06) Q ⁺ (6) = W '(04) Q ⁺ (7) = W' (02)

The indices of k = 1, n = 19 are then evaluated and the additional terms W '(19), W' (17), W '(1
5), W '(13), W' (11), W '(09),
W '(07), W' (05), W '(03) and W'
Give (01). All terms except W '(19) start codebook overlap index k = 1, n = 19
Detected by applying to. The result is then where the sequence is expanded for vectors with k> 8 to show how the contribution continues for higher vector numbers. Table V Q ⁺ (1) = W '(04) + W' (14) + W '(19) Q ⁺ (2) = W' (02) + W '(12) + W' (17) Q ⁺ (3) = W '(10) + W' (15) Q ⁺ (4) = W '(08) + W' (13) Q ⁺ (5) = W '(06) + W' (11) Q ⁺ (6) = W ' (04) + W '(09) Q ⁺ (7) = W' (02) + W '(07) Q ⁺ (8) = W' (05) Q ⁺ (9) = W '(03) Q ⁺ (10 ) = W '(01) This exhausts all the values that can be determined from it based on the index and codebook overlap for k = 1. No additional non-zero values appear at the end of the vectors k = 1, k = 2, so the correlation values Q ⁺ (1), Q
⁺ (2) is now perfect.

The next indices to be included are k = 3, n = 20 and the additional terms W '(20), W' (18), W '.
(16), W '(14), W' (12), W '(1
0), W '(08), W' (06), W '(04) and W' (02). Again, all terms except W '(20) start codebook overlap index k = 3, n =
Identified by applying 20. The result is as follows, where the sequence is expanded for vectors with k> 10 to show how the contribution continues for higher vector numbers. Table VI Q ⁺ (1) = W '(04) + W' (14) + W '(19) Q ⁺ (2) = W' (02) + W '(12) + W' (17) Q ⁺ (3) = W '(10) + W' (15) + W '(20) Q ⁺ (4) = W' (08) + W '(13) + W' (18) Q ⁺ (5) = W '(06) + W' ( 11) + W '(16) Q ⁺ (6) = W' (04) + W '(09) + W' (14) Q ⁺ (7) = W '(02) + W' (07) + W '(12) Q ⁺ (8) = W '(05) + W' (10) Q ⁺ (9) = W '(03) + W' (08) Q ⁺ (10) = W '(01) + W' (06) Q ⁺ ( 12) = W '(04) Q ⁺ (13) = W' (02) This can be determined from the index and codebook overlap for k = 0 to k = 7. Use all values. No additional non-zero values appear at the ends of the vectors k = 1 to k = 7, so the correlation values Q ⁺ (1) to Q ⁺ (7) are now perfect.

The process described above exhausts all non-zero entries in the codebook and all Q's.
⁺ (K) Continue until the correlation value is determined. Q ^- (k)
The process used for the value of is substantially the same. Splitting the ternary codebook into separate parts to compute Q ⁺ (k) and Q ⁻ (k) is described above for computing the correlation value of Q (k) by taking advantage of the codebook overlap. While avoiding the need to take into account the sign of individual entries during the process of, it is not disturbed. Q (k) is the difference ^{Q (k) = Q + (} k) -Q -
Detected by (k). The vector of index k = j with the largest correlation value Q (j) is Q for k = 1 to k = K (or at least for some of its subsets) using means well known in the art. It is identified by comparing the values of (k). The calculated value determined as above is the optimal stochastic codebook vector, i.e., the stochastic codebook vector which, when used to synthesize the speech, gives the least error compared to the input target speech, Used with other information in the synthetic analysis process described above to identify the. This optimal stochastic codebook vector from codebook 180 is then finally transmitted at the receiver using the transmitted vocode (VOCOD) which is used to replay the input speech.
E) is used to build.

More generally described is a set of a first vector V (n) and a second vector S _k (n) having values identified by an index n ranging from n = 1 to N. Each of the second vectors is identified by an index k and each of the second vectors is zero or non-zero and identified by an index n ranging from n = 1 to N.
The process of encoding speech described above using a combination with said set of second vectors S _k (n) having values up to and including _k is different for S _k (n _i ) is non-zero.
For identifying the index n _{k, i} of S _k (n) with respect to, the value of V (n) corresponding to the index n _{k, i} is added and the sum Q
Forming (k), identifying the value k = j corresponding to the maximum value Q (k = j), and synthesizing the speech using S _{k = j} (n). .

Furthermore, said set of second vectors is preferably determined by overlapping the preceding second vectors according to the amount of overlap Δk, Δn, in which case said identifying and summing step is n _{1 , I} , and using the amount of overlap Δk, Δn, let S _i (n _i ) be non-zero, and for k = 1, S
identifying the indices n _{1, i} of _k (n), S
determining further exponents n _{k, i ′} for k> 1 given that _k (n _{i ′} ) is non-zero, and adding the value of V (n) to such exponents and further exponents And forming a sum Q (k). Furthermore, for k ≧ 2, we identify S _k (n _{i ″} ) as non-zero and identify a previously unidentified first index n _{k, i ″} , and then from index n _{k, i ″} Starting and assuming that S _k (n _{i ″ ′} ) is non-zero, the amount of overlap is used to determine yet another index n _{k, i ″ ′} for k ≧ 3, and such an additional index. The value of V (n) is added to
It is convenient to form (k).

The method described above can be implemented by a general purpose computer, in which case the computer
S _k (n _i ) is non-zero and S for different k
Means for identifying index n _{k, i} of _k (n), index n
Means for adding the values of V (n) corresponding to _{k, i} to form the sum Q (k), and means for identifying the value k = j corresponding to the maximum value Q (k = j) , And S _{k = j} (n)
Should be programmed to provide means for synthesizing speech using. Those of ordinary skill in the art will understand how to do this.

Further, the computer is S ₁ (n _i )
Is non-zero, the index n _{1, i} of S _k (n) is k = 1.
Means for identifying, S _k (n _{i ′} ) is non-zero, starting from n _{1, i} and using the overlap amounts Δk, Δn, yet another index n for k> 1.
_A program for providing means for determining _{k, i '} , and means for adding such values of V (n) to such and other indices to form a sum Q (k). It is desirable to be done. Further, the means for identifying, determining and adding has a first index n _{k, i ″} not previously identified as S _k (n _{i ″} ) being non-zero.
Means for identifying to _{k ≧ 2, S k (n} i "')
Means for determining yet another exponent n _{k, i ″ ′} for k ≧ 3, starting from the exponent n _{k, i ″,} where _k is non-zero, and using the amount of overlap, and such further It is desirable to include means for adding the value of V (n) to the index of to further form the sum Q (k).

[0150]

EFFECT OF THE INVENTION Equivalence of the sum of the above-mentioned products of the first vector having the values of n = 1 to N and the second set of k = 1 to K having the values of n = 1 to N. Means and methods for providing an object by exploiting the sparse non-zero values of a codebook and the overlapping properties of the codebook vector are substantially less computationally intensive and faster than prior art techniques. And can be achieved with significantly less computational resources than required by the prior art. The processes described above are accomplished by a general purpose or special purpose computer programmed to perform the procedures described herein and illustrated in Tables I-VI. Those skilled in the art will understand how to program a computer based on the description herein and using means well known in the art to accomplish the above steps.

Those skilled in the art will appreciate that the means and methods described above based on the description herein are related to determining which of the stochastic codebook vectors provides the best match with the target speech. It is clear that it has the same effect as the multiplication step normally required for the correlation process. Eliminating the multiply operations makes the correlation operations faster and reduces the number of vector-valued operations required. These advantages are extremely advantageous.

Finally, the above-described embodiments of the present invention are considered to be exemplary only. Many alternative embodiments can be devised without departing from the spirit and scope of this invention.

[Brief description of drawings]

FIG. 1 is a simplified block diagram of a CELP vocoder system in general form.

FIG. 2 is a simplified block diagram showing a CELP coder according to a preferred embodiment of the present invention.

FIG. 3 is a block diagram showing in detail a part of the coder of FIG. 2 (b) according to the first embodiment.

FIG. 4 (b) of FIG. 2 according to a preferred embodiment of the present invention.
FIG. 3 is a block diagram showing in greater detail a portion of the coder of FIG.

FIG. 5 is a block diagram illustrating an apparatus for providing an autocorrelation coefficient of an adaptive codebook vector according to a preferred embodiment of the present invention.

FIG. 6 is an illustration showing the contents of a small probabilistic codebook of the type used for CELP coding.

FIG. 7 is a simplified block diagram illustrating the cross-correlation function of the present invention.

8 is a block circuit diagram showing further details of the multiplexer used in FIG. 7. FIG.

9 is an illustration showing the contents of a first memory means whose entry corresponds to a non-zero entry in the codebook of FIG.

10 is an illustration showing the contents of a second memory means whose entry corresponds to a non-zero entry in the codebook of FIG.

[Explanation of symbols]

100 CELP coder 106 Transmission path or transmission path 300 CELP decoder 110 Bandpass filter 112 A / D converter 114 Framer 116 Frame memory 122 LPC analyzer 125 Coder 130 Bandwidth extension weight generator 142,165 Spectrum memory subtractor 145,170 Spectrum Inverse filter A
(Z) 150,175 Cascade weighting filter 1 / A
(Z / r) 162 Pitch memory subtractor 155 Adaptive codebook 180 Stochastic codebook 182 Channel decoder 190 Long delay pitch predictor 195 Short delay vector predictor 1 / A (z) 197 Gain multiplier 200 Spectral inverse filter A (z) 205 Cache weighting filter 1 / A (z / r) 210 Channel coder 220 Adaptive codebook searcher 225 Probabilistic codebook searcher

[Procedure amendment]

[Submission date] October 22, 1993

[Procedure Amendment 1]

[Document name to be corrected] Drawing

[Correction target item name] All drawings

[Correction method] Change

[Correction content]

[Figure 1]

[Figure 3]

[Fig. 2]

[Figure 4]

[Figure 6]

[Figure 9]

[Figure 10]

[Figure 5]

[Figure 7]

[Figure 8]

 ─────────────────────────────────────────────────── ─── Continuation of front page (31) Priority claim number 722, 572 (32) Priority date June 27, 1991 (33) Priority claiming country United States (US) (72) Inventor David El Baron United States Arizona State 85251, Scottsdale, East Thomas 8449

Claims (19)

[Claims] 1. An optimal codebook vector C _{k = j} (n) for coding a speech frame with N consecutive samples n = 1 to n = N of the input analog speech to best synthesize said speech frame. Device for determining
0,220 ') and K possible perceptually weighted drive vectors C
_An adaptive codebook (155) containing _k (n),
Where k is an integer exponent ranging from 1 to K for identifying the vector, and n = 1 to n = N is an integer exponent identifying consecutive audio samples in the audio frame. , One or more LPC filters (190, 19) for synthesizing a trial copy of the speech frame when driven by the codebook vector C _k (n)
5,200,205,170,145,175,15
0), and the LPC filters (190, 195, 2
00, 205, 170, 145, 175, 150) having an impulse response H (n), a codebook for determining an optimal codebook vector C _{k = j} (n) that best synthesizes the speech frame. One or more LPs according to the vector C _k (n)
C filter (190, 195, 200, 205, 17
0, 145, 175, 150), and the residual X of the target speech perceptually weighted for comparison with the result.
Means (150) for generating (n), for convolution of X (n) with H (n) for each value of n once per frame for transmission to the cross-correlator (520 '). Means (510 ') for producing a convolved output (512'), cross-correlate the convolved output (512 ') with C _k (n) for each value of n and k, and output (52
1 ') is connected to a dividing means (530') and is a squarer (5)
25 ') and the cross-correlated output (55
1 ') for producing (1') an autocorrelation of C _k (n) for each value of n and k for transmission to a multiplier means (543 '). Means (56) for providing a correlation output (561 ')
0 '), H (n) are autocorrelated to a second autocorrelated output (552', 552 ', for transmission to said multiplier means (543').
Means (550 ') for generating an output product for transmission to the adder (540') by multiplying the first (561 ') and second (552') autocorrelation outputs. Means (54 ') for generating (545')
3 '), the product is added to each value of k and a divider (53
0 ') means (540') for producing an added output (541 '), a squared cross-correlation output (521') and an added output of the adder (541). Dividing means (53) for detecting the ratio of ′ ′ and transmitting it to the selector means (570 ′).
0 '), and for transmission to the channel coder (210), selects the value C _{k = j} (n) of C _k (n) that produces the output of maximum amplitude from said dividing means (530'). An apparatus for encoding a speech frame, characterized in that it comprises means (570 ') for 2. Encoding a speech frame comprising N consecutive samples of the input analog speech, k being an integer index from 1 to K, and n being consecutive speech samples n = 1 in said speech frame. , ..., n = N is another integer index that identifies the speech frame using an adaptive codebook (155) containing perceptually weighted drive vectors C _k (n) of K targets. A method for determining an optimal codebook vector C _{k = j} (n) to synthesize best, the method comprising synthesizing a trial copy of the speech frame when driven by the codebook C _k (n) vector. One or more LPC filters (190, 19)
5,200,205,170,145,175,15
0), said one or more LPC filters (190, 195, 200, 205,
170, 145, 175, 150) are impulse responses H
With (n), one or more LPC filters (190, 195, 200, 2) depending on the codebook vector C _k (n)
05,170,145,175,150) to provide a perceptually weighted target speech residual X (n) for comparison with the result, H for each value of n once for each frame. Convolving X (n) with (n) to produce a convolved output W (n) for transmission to the cross-correlator (520 '), for each value of n and k C _k (n ), Cross-correlating the output W (n) convolved with the output to produce a cross-correlated output for transmission to a squarer (525 ') coupled to a divider (530'). autocorrelating C _k (n) for each value of n and k to provide a first autocorrelation output U _k (m) for transmission to a multiplier (543 ′), In this case m is m
A dummy exponent ranging from = 0 to m = N−1, a second autocorrelated output Φ (m) for transmission to the multiplier (543 ′) by autocorrelating H (n). In which m is a dummy exponent from m = 0 to m = N−1, wherein the multiplier (543 ′) multiplies the first and second autocorrelation outputs. Generating an output product (545 ') for transmission to an adder (540'), adding the product to each value of k and adding a divider (53).
0 ') to produce a summed output (541') for transmission to the peak selector (570 ') and the squared cross-correlation output (521') for addition to the adder The step of performing a division to obtain the ratio of the outputs (541 '), and C _k (n) producing the maximum magnitude output from the divider (530') for transmission to the channel coder (210). ) Value C _{k = j} (n) is selected in the peak selector. 3. A method for providing CELP coding for digitized input speech frames based on the use of a codebook (155) containing K vectors each having N entries, the method comprising: Autocorrelating the codebook vector for the first P (P << N) to determine a first autocorrelation value (561 ') for it, using K codebook vectors and said first autocorrelation value Evaluating the codebook vector of K by generating a synthesized speech and comparing the result with the input speech, which S (S << K) of the K codebook vectors is compared with the input speech. Determining whether to provide synthetic speech with a smaller error than the remaining vector of K-S evaluated by Autocorrelating a codebook vector for S (P <R ≦ N) of K vectors for a bird to provide a second autocorrelation value for it, using the second autocorrelation value Re-evaluating the S of the K vectors and comparing with the input speech to identify which of the S codebook vectors provides the minimum error, and the codebook giving the minimum error Forming a CELP code for the speech frame using a vector identifier. 4. The P, S and R are 5 ≦ P ≦ 10,
The method according to claim 3, wherein 1 ≦ S ≦ 7 and R = N or N−1. 5. The frame of digitized input speech designated X (n) comprises n = 1 to n = N consecutive samples of the input analog speech, and the codebook is K.
Target perceptually weighted drive vector C
_k (n), where k is an integer exponent ranging from 1 to K, and n is another integer exponent identifying consecutive speech samples n = 1 to n = N in the speech frame. , And C _{k = j} (n) represents the optimal codebook vector that best synthesizes the target speech frame X (n), and the first autocorrelation step is Autocorrelating with respect to m = 0 to m = P, where P << N is the codebook vector C _k (n), and said evaluating step comprises: U detected from said equation in a codebook searcher. comprising recursively evaluating all K vectors C _k (n) using the value of P of _k (P) to determine a mean square error probability, and said determining step comprising S << K, and selecting S of the K vectors C _k (n) that provide the closest match to the target speech X (n), and the second autocorrelation and reevaluation step comprises: In the codebook searcher all m = 0 to determine U _k (m) in the above equation
Recursively re-evaluate S of the K vectors chosen above using ~ m = N-1, thereby yielding the j-th value C
_{k = j} (n) and selecting the corresponding gain index _{Gk = j} that gives the best fit to the target speech X (n), and said forming step for transmitting to the CELP combiner. The method of claim 3, comprising transmitting C _{k = j} (n) and G _{k = j} to a channel coder. 6. An apparatus (100, 220, 220 ', 600) for CELP coding of speech using vector autocorrelation coefficients of an adaptive codebook (155), the analysis initially comprising a length. Utilizing a subset of samples M in association with L> M speech analysis frames, the apparatus is k = 1 and m is the autocorrelation delay index and n is the index of consecutive samples in the codebook vector. And the autocorrelation coefficient U _{k ′} (m) of the first vector C _k (n) of length M is given by the first equation, for m = 0 to T <M, Means (610) for determining according to k ≧ 2 and incrementally the autocorrelation coefficient U _k (m) of the remaining codebook vectors (M +) for m = 0 to T <M.
Up to k-1) = L, the second equation, _U'k (m) = [ _U'k-1 (m) + _Ck (M + k-
1) Means (624, 626, 62) for determining according to C _k (M + k-1 + m)]
8, 630), scaling the result of the first equation according to the following equation: U ₁ (m) = (L / M) U ′ ₁ (m) and the result of the second equation: , U _k (m) = {L / (M + k−1)} U ′ _k (m), and these scalings are m =
Means (632) for producing a result for each m and each k, for which 0 to T <M, and which codebook vector using said result is the smallest compared to the input speech. An apparatus for CELP coding of speech, characterized in that it comprises means (220, 220 ') for assessing whether to provide an error. 7. An adaptive codebook of vector length N (15)
CEL of speech using vector autocorrelation coefficient of 5)
A method for P coding, wherein the analysis initially uses a subset of samples of M <N having speech analysis frames of length L, said method being: k = 1 and m is an autocorrelation delay exponent. And n is the index of consecutive samples in the codebook vector, and for m = 0 to T <M, Calculating the autocorrelation coefficient U _k (m) of the first vector C _k (n) of length M according to U ₁ (m) = (L / M) U ′ ₁ (m), k ≧ 2 Then, for m = 0 to T <M, the following equation: U ′ _k (m) = [U ′ _k−1 (m) + C _k (M + k−
1) C _k (M + k−1 + m)] U _k (m) = {L / (M + k−1)} U ′ _k (m) according to the autocorrelation coefficient U of the remaining codebook vector
calculating _k (m), repeating the second calculation until (M + k−1) = L, and using the autocorrelation coefficient determined as described above, the codebook vector C _k Determining which of (n) produces the smallest error when compared to the input speech. A method for CELP coding of speech. 8. The method according to claim 6 or 7, wherein T and M have a relationship of T = M-1. 9. A method for CELP coding of speech using the autocorrelation coefficients of a vector of an adaptive codebook (155) identified by an index k, the analysis by synthesis being initially M <L. A codebook value is used, where L is the length of the speech analysis frame and m is an index from 0 to M-1 that describes the autocorrelation delay, the method is as follows: n is the codebook vector value N is the exponent of
Calculating the autocorrelation coefficient of m = 0 to M−1 of the first codebook vector k having values of = 1 to M therein, m = 0 to M−1 in the temporary memory (630). Storing the calculated autocorrelation coefficient of the multiplicative factor L /
Scale by M and output the result (634)
To n = M + j, where the codebook value is n = M +
multiplying by a codebook value falling from j to n = 1 and adding the product to the autocorrelation coefficients from said temporary store (630) for m = 0 to M-1, respectively, to produce a result. Replacing the autocorrelation coefficient in the temporary storage device (630) with the result, scaling the coefficient in the temporary storage device (630) by a multiplication factor L / (M + j) and outputting the result to the output (634). Transferring, repeating k, k = (L + 1−M) and j = 2 to j = k−1 for the multiplication, replacement, scaling and transfer steps, and using the autocorrelation coefficient transferred to the output And which codebook vector has better voice C
A method for CELP encoding of speech, comprising: determining whether to provide ELP encoding. 10. Apparatus (100,7) for performing CELP coding of speech by combining a first vector with a second set of vectors identified by an index k.
00) and the first and second vectors are n =
Having a value identified by an index n ranging from 1 to N, the device comprising: n = 1 to N outputs, n = 1 to N first inputs, second inputs, and n = 1 to N N × N having selection means of
Multiplexer (704) for providing a first logic level to the nth selecting means for coupling the nth output to the nth first input and to the nth selecting means. A second logic level provided is for coupling the nth output to the second input, n = 1 to N values of the first vector being n = of the first multiplexer (704). 1-N first input (70
1) means for supplying n = 1 to N values of the second vector of index k = 1 to n = 1 to N selection means of the first multiplexer (704) (708) 706), the second vector providing the first logic level for some values of n in the selection means of n = 1 to N and An output (710) of the first multiplexer (704) coupled to an output (710) of the first multiplexer for providing a second logic level for other values.
First accumulator means (712) for adding together the values of said first vector to provide a first sum (716), indexing k from k = 1 to k = K Means (714), and means (228) for synthesizing the speech based on which sum identifies the second vector that gives the closest match to the target speech. For CELP encoding of. 11. The second vector has values 0, + 1,-
Said means for providing a value of n = 1 to N of said second vector having two parts, namely an entry corresponding to the location of the value of 0, + 1 of said second vector. A first portion (706) having a value of 0,1 and a second portion (707) having a value of 0,1 corresponding to a location of a value 0, -1 of the second vector,
And said providing means provides its first part to said selecting means (708) of said first multiplexer (704), and said device further comprises: n = 1 to N outputs, n = 1 to N A second N × N having a first input, a second input, and n = 1-N selection means.
Multiplexer (705) for providing a first logic level to the nth selecting means for coupling the nth output to the nth first input and to the nth selecting means. A second logic level provided for coupling the nth output to the second input and a second of the providing means.
Is coupled to the selection means of the second multiplexer, the value n = 1 to N of the first vector to the first input n = 1 to N of the second multiplexer. Second means for supplying (701), together adding together the values of the first vector coupled to the output of the second multiplexer and transferred to the output of the second multiplexer to obtain a second sum ( Second accumulator means (713) for providing 717), and said first
Means for combining the (716) and second (717) sums to produce a result (721) that is used to determine which input vector (702) gives the closest match to the target speech (721). 720), The apparatus according to claim 10, further comprising: 12. A first vector V (n) having values identified by an index n ranging from n = 1 to N, and a second vector.
For the CELP coding of speech using a combination of a set of vectors S _k (n) of each of the second vectors, each of the second vectors being identified by an index k and each of the second vectors being Zero or non-zero and n = 1 to
Having values up to N identified by an index n leading to N,
The apparatus is means (704, 706) for identifying an index n _{k, i} of S _k (n) for different k, said S _k (n
_i ) is non-zero, means (712, 721) for adding the values of said V (n) corresponding to the exponent n _{k, i} to form sum Q (k), maximum value Q ( means (714) for identifying the value k = j corresponding to k = j) and means (228) for synthesizing the speech using S _{k = j} (n). A device for CELP encoding of speech. 13. Continuing vectors of said set of said second vectors are determined by the overlap of the preceding second vectors according to the amount of overlap Δk, Δn, means for said identification and means for addition. Is a means (705, 707) for identifying the index n _{1, i} of S _k (n) for k = 1, said S ₁ (n _i ) being non-zero, S _k (n _{i ′} ) Is non-zero, and for k> 1, n
_{1, i} and the overlap amount Δk, Δ
Means (705, 707) for determining yet another index n _{k, i '} using n, and adding the values of V (n) for such index and still other indices to sum Q. 13. Device according to claim 12, comprising means (713, 721) for forming (k). 14. The means for identifying, the means for determining, and the means for adding assume that S _k (n _{i ″} ) is non-zero and were not previously identified for k ≧ 2. First
Means for identifying the index n _{k, i ″} of S _k (n
_{i ″ ′} ) is non-zero, and for k ≧ 3, another index n
Means for determining _{ki "'} starting from the index n _{k, i"} and using said amount of overlap, and adding the values of V (n) for such yet another index and further summing Q 14. The apparatus of claim 13, comprising means for forming (k). 15. A method for CELP encoding speech by combining a first vector (701) with a second set of vectors identified by an index k, said first and second vectors. Has a value identified by the index n leading to n = 1 to N, the method comprising: n = 1 to N outputs, n = 1 to N first inputs, second inputs, and N × N first with n = 1 to N selection means
A multiplexer (704) of
The first logic level provided to the nth selection means couples the nth output to the nth first input, and the second logic level provided to the nth selection means is the nth output. A second output coupled to the second input, providing the n = 1 to N values of the first vector (701) to the n = 1 to N first inputs of the first multiplexer. Step, n = 1 to 2 of the second vector (702) with index k = 1
The value of N is set to n = 1 of the first multiplexer (704).
To N selection means, said second vector providing said first logic level for some values of n in said selection means of n = 1 to N and n. Providing the second logic level for other values, adding together the values of the first vector coupled to the output of the first multiplexer to provide a sum, and yet another of k Repeating the providing and summing steps for the values, and synthesizing the speech based on which sum identifies the second vector that gives the closest match to the target speech. For CELP encoding of speech to be played. 16. The second vector has values 0, + 1, −.
1 and providing a value of n = 1 to N of said second vector comprises a first entry having an entry 0,1 corresponding to a location of a 0, + 1 value of said second vector. Providing a portion to a first multiplexer and a second portion having a value 0,1 corresponding to a location of a value 0, -1 of the second vector, similar to the first multiplexer and Providing a second multiplexer responsive to the first input vector at its input and the second portion of the second vector at its selection means, similar to the first malplexer, said first multiplexer The values of the first vector coupled to the output of the multiplexer are added together to provide a first sum, and the values of the first vector coupled to the output of the second multiplexer are added together. To provide a second sum, the first and second sums are combined to provide an output useful for determining which input vector provides the closest match to the target speech. 16. The method of claim 15, comprising: 17. CELP for speech using a combination of a set of a first vector V (n) and a second vector S _k (n), where n = 1 to N and having a value identified by an index n. A method for encoding, wherein each of the second vectors is identified by an index k and each of the second vectors is zero or non-zero and n = 1 to N.
With values up to N identified by an exponent n, where S _k (n _i ) is nonzero and S is different for different k.
_k index _{n k} of the _(n), identifying a _i, the index _{n k,} sums by adding the value of V (n) corresponding to the _i Q
Forming (k), identifying the value k = j corresponding to the maximum value Q (k = j), and synthesizing the speech using S _{k = j} (n). A method for CELP encoding of speech, comprising: 18. The successive vectors of said set of said second vectors are determined by the overlap of the preceding second vectors according to the amount of overlap Δk, Δn, said discriminating and summing steps comprising S ₁ (n _i ) is non-zero, the index n of S _k (n)
_{Discriminating 1, i} with respect to k = 1, starting from n _{1, i} and overlapping amounts Δk, Δn
Using S _k (n _{i ′} ) is non-zero and k>
The step of determining yet another index n _{k, i ′} for _1, and adding the values of V (n) to such index and still other indices to form the sum Q (k). 18. The method of claim 17, comprising: 19. S _k (n _″ ) is non-zero, and
For k ≧ 2, the previously unidentified first index n
_{k, i ″} , and then using the amount of overlap, starting from the index n _{k, i ″} , S _k (n _{i ″ ″} )
Is non-zero, and for k ≧ 3, another index n
the step of determining _{k, i ″ ′} , and adding the values of V (n) for such further exponents and further summing Q
19. The method of claim 18, further comprising forming (k).