JPH0451300A - Linear predictive encoder and decoder - Google Patents

Abstract

PURPOSE:To decrease the amount of information which is sent to a reception side by constituting a prediction filter by cascading an instantaneous type exiting wave prediction filter and an instantaneous type exciting wave prediction filter. CONSTITUTION:An error evaluating circuit 311 outputs the best i(t) to an adaptive sound source generation part 312 and an adder 310 adds the best eai and eg(t) to calculate the 2nd difference reproduction vector e2(t) of a vector e2(t) and send it out to an adder 314 and an exciting wave analysis part 317. The adder 314 adds the 2nd difference reproduction vector e2(t) and a 2nd predicted vector S2' to generate and send a 1st difference reproduction vector e1(t)' to an exciting wave analysis part 317, an exciting wave prediction filter part 316, a glottis analysis part 319, and an adder 318. The adder 318 adds the 1st prediction vector S1(t)' and 1st difference reproduction vector e1(t)' to generate a reproduction vector S(t)' to an input vector S(t), which is sent out to the glottis analysis part 319 and a glottis prediction filter part 315. Lastly, the decoder sends decoded betaq<(>G<)> and the best index (i) to a decoder.

Description

DETAILED DESCRIPTION OF THE INVENTION (Field of Industrial Application) The present invention relates to a low rate encoder and decoder using linear prediction.

(Prior Art) FIG. 3 is a block diagram showing the configuration of a conventional linear predictive encoder.

The main parts of the encoder shown in FIG. 3 are explained in the following document.

Literature name N, S, Jayant and J, H, Ch
en,”Speech Coding with Tim
e-Varying Bit Allocations
to Excitation and LPC Para
meters', Proc, ICASSP, pp65
-68, 1989 In FIG. 3, original speech vectors S, which are collected into frame units and input as vectors to a terminal 101, are first inputted to a short-time analysis circuit 102, which performs short-term prediction on the original speech vectors S. coefficient α,
is calculated. The literature does not particularly detail the algorithm for calculating the short-term prediction coefficient α.

This is a normal autocorrelation method that assumes that the input is a stationary signal within a frame. The short-term prediction coefficient α, is the quantizer 1
14, and the quantized value α3. become. Quantizer 11
4 sends αj to the inverse quantizer 115 and the multiplexing circuit 111. The inverse quantizer 115 outputs quantized values α4. are converted into inverse quantized values α, and input α1 to short-time filter 03. Send out.

q Next, both switches 108 and 113 are opened. The adaptive excitation code table 104 has an adaptive excitation code vector e&4.
(t=i~n) is output. adaptive excitation code vector e
, are created by shifting the excitation signals used in the past by one sampling time in the adaptive excitation code tape 104. Adder 110 adds vector e8. In this case, switch 1
Since no signal comes from 13, eal becomes the excitation vector e and is input to the short-time filter 103.

The short-time filter 103 calculates an output vector SWi for the vector e, and the vector SWi sent to the subtracter 109 corresponds to the original speech S.

The subtracter 109 subtracts SW, and S component by component, and sends an error vector er1 to the perceptual filter 105.

The perceptual filter 105 sends the vector e-ik to the perceptual error calculation circuit 106 for the vector eri. The perceptual error circuit 106 calculates the root mean square g of each component of the vector ew□.
i is calculated, and 1, which minimizes gi, is set as the index I of the optimal adaptive excitation code and sent to the adaptive excitation code pull 104 and the multiplexing circuit 111.

Next, the switch 113 is closed while the switch 108 remains open. The adaptive excitation code table 104 outputs the optimal adaptive excitation code vector ea0 with the index 11°. The stochastic excitation code table 107 includes stochastic code vectors e8. (i=1%ct) is output. The adder 110 performs component-by-component addition of the vector e8° and the vector e8, and sends the excitation vector e to the short-term prediction filter 03.

The short-time filter 03 calculates an output vector SW for the vector e and sends it to the subtracter 109. Vector SWi corresponds to original speech S. The subtractor 109 is Swi
and S are subtracted component by component, and an error vector e□ is sent to the perceptual filter 105. Perceptual filter 105 sends an output vector ewi for vector eri to perceptual error calculation circuit 106. The perceptual error circuit 106 calculates the root mean square gf of each component of the vector eW, and uses i, which minimizes gi, as the index Ia0 of the optimal stochastic excitation code, and uses the stochastic excitation code table 107 and the multiplexing circuit 11. Send to.

Next, both the switch 10B and the switch 113f are closed, and an optimal excitation vector eopt is created by adding the optimal adaptive excitation code vector and the optimal probabilistic code vector, and is input to the subframe delay circuit 116. The subframe delay circuit 116 generates the optimum excitation vector e01. for the adaptive excitation code table 104. Send out. The adaptive excitation code table discards the temporally oldest vector previously input and uses the latest vector e sent from the subframe delay circuit 116. , t after the vector sequence.

The multiplexing circuit 111 outputs the quantized values α of the short-time prediction coefficients,
Index I8゜,■,. is multiplexed and transmitted from the output terminal 112 to the transmission path 117 as a tocop-tal code C.
Send to.

FIG. 4 shows a transmission line 201 of a conventional linear predictive decoder.
The multiplexed code C sent from the encoder through the demultiplexer circuit 203 is input from the terminal 202 to the demultiplexer circuit 203 .

The demultiplexing circuit 203 separates the multiplexed code C into αj9, I
, I and the corresponding inverse quantization ao
g.

device 204, adaptive excitation code table 205. It is sent to the stochastic excitation code table 206. The inverse quantizer 204 converts the quantized value α into inverse quantized values α1. It is converted into P and sent to the short-term prediction filter 207. Adder 2
0B is the optimal excitation vector e obtained by adding the output vector ea0 of the adaptive excitation code table 205 and the output vector e of the stochastic excitation code table 206. 1. is created and sent to the short-time filter 207 and subframe delay circuit 209. The short time filter 207 reproduces a reproduction signal vector Sw corresponding to the original speech vector. Adaptive excitation code table 205 is constructed by subframe delay circuit 209 to e09. Enter.

(Problem to be solved by the invention) In the configuration of the linear predictive encoder and decoder having the above configuration,
It is necessary to quantize the short-term predictive filter coefficients, excitation waves, and sound source parameters that provide the vocal tract transfer function and send them to the receiving side.

Therefore, since there are many parameters to be transmitted to the receiving side, there is a drawback that the amount of information to be transmitted to the receiving side cannot be reduced without deteriorating the sound quality.

An object of the present invention is to provide a linear predictive encoder/decoder that can reduce the amount of information transmitted to the receiving side without deteriorating the sound quality as described above.

(Means for Solving the Problems) The present invention includes a linear predictive encoder in which a predictive filter is configured by a subordinate connection of an instantaneous vocal tract predictive filter using a reproduced signal and an instantaneous excitation wave predictive filter, The difference signal (sound source signal) between the input signal and the two predicted values is one type selected from amplitude-adjusted white noise and a plurality of quantized sound source signals previously transmitted to the decoder. The receiving side is provided with means for transmitting only the sound source information.

Furthermore, in the linear predictive decoder, the predictive filter is configured by a cascade connection of an instantaneous vocal tract predictive filter and an instantaneous excitation wave predictive filter, and decodes based on the sound source signal sent from the transmitting side.

(Function) First, in the present invention, the vocal tract prediction filter coefficients and the excitation wave prediction filter coefficients are created using signals that can also be obtained on the receiving side, so there is no need to send the filter coefficients to the receiving side.

Therefore, the amount of information transmitted to the receiving side can be reduced.

Conventionally, prediction coefficients have been calculated assuming that the audio signal is a stationary signal. However, audio signals cannot originally be considered stationary signals. Therefore, secondly, since the present invention uses the instantaneous maximum entropy method (instantaneous MEM), which is a prediction filter calculation algorithm that assumes a non-stationary signal, it is possible to calculate accurate prediction filter coefficients, improving prediction ability. .

Improvement in prediction ability reduces the magnitude of the difference signal between the input signal and the prediction signal, which ultimately makes it possible to reduce the amount of information allocated to the difference signal, thereby reducing the amount of information transmitted to the receiving side.

(Embodiment) FIG. 1 is a block diagram showing an embodiment of the encoder of the present invention.

In FIG. 1, a time-series n-dimensional input vector S (t) input from a terminal 301 is subtracted by a subtracter 302 and is converted into a first nΔ-dimensional predicted vector 51 (t), which is the output of a vocal tract prediction filter section 315. is subtracted, resulting in a first n-dimensional difference vector e1ft).

△ The first predicted vector 51 (t) is converted to the second n-dimensional predicted vector △ 52 which is the output of the excitation wave prediction filter unit by the subtracter 303, as shown in equation (1). (t) and the second n-dimensional difference vector e2(
t).

Δ The second predicted vector 52(t) is also expressed as a linear sum of past reproduction vectors, as shown in equation (2), similarly to 51(t).

△ 52 (t) = Σα1 (2) Tobi (t-i) (
2) 1=1 α1(1) in equation (1) is the past playback vector e
.

(t-1), 5(t-i), the vocal tract component of the input vector S is predicted using the calculated vocal tract prediction coefficient using the n-th instantiated maximum entropy method. Similarly, α in equation (2)
I(2) is the past reproduction vector e2(t-1).

From el(t i ), a vector e2(t) that predicts the excitation wave component of the input vector S with the vocal tract prediction filter coefficient calculated by the n-th instantiated maximum entropy method is obtained by the subtracter 304 and the sound source analysis unit 305. will be sent to.

The sound source analysis unit 305 removes the long-period component of the vector e (t), and sends the removed n-dimensional vector e dt) to the Uthian gain calculation unit 306 . The Gaussian gain calculation unit 306 calculates the correlation value between the n-dimensional Gaussian vector prepared in advance and the vector e,(t), calculates the gain coefficient β(0) (1) proportional to the correlation value, The n-pit quantizer 307F, which is input to the n-pit quantizer 307F, sends the n-bit code β(G)(t) to the inverse quantizer 307R.

The inverse quantizer 5oyRId sends the inverse quantizer β, (0)(t) of the gain coefficient β (t to 0) to the coefficient circuit 308.

Quantization normameter i (t
) can be obtained using the procedure shown below. Gaussian signal generation section 309 sends vector g prepared in Gaussian gain calculation section 306 to coefficient circuit 30B. Coefficient circuit 30& multiplies vector g by β(G)(t) and sends n-dimensional vector e (t) to adder 310.

The adder 310 generates an n-dimensional vector eat (t) (i-
1 to t) and the vector e(t), and the subtractor 304
and the vector e2(t) and (eai (t) + eg(
The error vector eri(t) with respect to t) ) is obtained and input to the error evaluation circuit 311. The error evaluation circuit 31 is
It has a built-in perceptual weighting filter, and the power of the error vector that has passed through the perceptual weighting filter is minimized by 1 (
Select t).

e-(t) is the past spectrum e2(t1) *...e2(t-
j) and the gain coefficient β9. Made by (P).

The error evaluation circuit 311 outputs the optimal i (t) to the adaptive excitation generator 312, and the adder 310 adds the optimal e&1(t) and eg(t) to reproduce the second difference of the vector e2(t). Calculate vector e2(t), and delay circuit 313;
It is sent to an adder 314 and an excitation wave analysis section 317.

The adaptive sound source generator is currently the oldest e2(t −j)
After discarding , enter e, (t) and get e, (t+1)
make.

l Vector i (t) is created, excitation wave analysis section 317, excitation wave prediction filter section 316, vocal tract analysis section 319, adder 3
Send on 18th.

Δ Adder 318 adds the first prediction vector S,(t) and the first
Add the differential reproduction vector e, (t) of , and create a reproduction vector S (t) for the input vector S (t),
It is sent to the vocal tract analysis section 319 and the vocal tract prediction filter section 315.

Finally, the encoder sends the code β(0) and the optimal index i to the decoder.

The process described above is such that the new input vector S is
Occurs every time the .

FIG. 2 shows an embodiment of the decoder of the present invention, in which 401 is a coefficient circuit, 402 to 404 are adders, and 40
5 is a terminal. The decoder inputs β(0) sent from the encoder to the coefficient circuit 401 and the optimal index i to the adaptive excitation generator 312.

In the subsequent processing, if β(0) and 1 are transmitted normally, the operation is exactly the same as that in the encoder, and the final vector S is equal to the reproduced vector S obtained in the encoder. is obtained at terminal 412.

(Effects of the Invention) As described in detail above, according to the present invention, it is not necessary to send the coefficients of the vocal tract prediction filter and the coefficients of the excitation wave prediction filter to the decoder side, and the prediction difference signal is Since sufficient whitening can be achieved, it can be expected that the amount of information allocated to sound source information may be small.

Therefore, it is possible to achieve a lower bit rate than the conventional linear predictive encoder/decoder.

[Brief explanation of the drawing]

FIG. 1 is a block diagram showing one embodiment of the encoder of the present invention, FIG. 2 is a block diagram showing one implementation of the decoder of the present invention, and FIGS. 3 and 4 are explanations of the prior art. It is a diagram. 301...Terminal, 302-304...Subtractor, 30
5... Sound source analysis unit, 306... Gaussian gain calculation unit, 307F... n-bit quantizer, 307R... inverse quantizer, 308... coefficient circuit, 309... is Gaussian signal generation unit , 310... Adder, 31... Error evaluation circuit, 312... Adaptive sound source generation unit, 313...
Delay circuit, 314... Adder, 315... Vocal tract prediction filter section, 316... Excitation wave prediction filter section, 31
7... Excitation wave analysis section, 318... Adder, 3/9.
... Vocal tract analysis section, 401 ... Coefficient circuit, 402-40
4...Adder. 5...Terminal.

Claims (2)

[Claims] (1) An instantaneous maximum entropy method using a reproduction signal in which the prediction filters are connected in a dependent manner and include a vocal tract prediction filter and an excitation wave prediction filter, and the tap coefficients of the two prediction filters correspond to the input signal. The difference signal between the prediction signal of the two prediction filters and the input signal is updated by a candidate difference created by white noise whose amplitude has been adjusted and two or more types of past reproduced difference signals whose amplitudes have been adjusted at least. Linear prediction characterized in that it is created by adding a candidate signal that has the smallest weighting error with the actual difference signal among the signals, and sends an amplitude parameter and an index indicating the optimal candidate signal to a decoder. encoder. (2) An instantaneous maximum entropy method using a reproduced signal in which the prediction filters are connected in a dependent manner and include a vocal tract prediction filter and an excitation wave prediction filter, and the tap coefficients of the two prediction filters correspond to the input signal. A linear predictive decoder updated by the above, and decodes a reproduced signal corresponding to an input signal of the encoder based on transmission information from the linear predictive encoder described in Patent No. 1.