JPH03259199A - Method and device for voice encoding and decoding - Google Patents

Abstract

PURPOSE:To realize the voice encoding and decoding with a small arithmetic quantity by providing plural functions, selecting the best function to optimum section length by using the autocorrelations of the functions and the cross- correlations between the functions and an input signal, and representing spec trum parameters. CONSTITUTION:A time series of spectrum parameters to be quantized is inputted to an input terminal 300. A code book 250 contains plural predetermined functions of predetermined time length as code words previously; and an autocorrelations calculating circuit 252 calculates the autocorrelations of the code words and a cross-correlation calculating circuit 251 calculates the cross- correlations between the code word and input signal so as to calculate the best code word and gain. Then a gain calculating circuit 260 calculates the best gain with which the selected code word is multiplied by using the output of the autocorrelation calculating circuit 252 and the output of the cross- correlation calculating circuit 251. Thus, the spectrum parameters are represented by using the predetermined functions to represent a voice signal excellently with a small calculation quantity and a small transmission informa tion quantity.

Description

DETAILED DESCRIPTION OF THE INVENTION (Industrial field of application) The present invention relates to an audio encoding/decoding method and apparatus for efficiently encoding and decoding audio signals with a low pit-ray.

[Conventional technology]

A multi-pulse encoding method is known for L7 and a method for transmitting an audio signal at a low bit rate, for example, about 16 kb/s or less. These represent the sound source signal as a combination of multiple pulses (multi-pulse) and the characteristics of the vocal tract as a digital filter, and calculate the information on the sound source pulse and the filter coefficients for each fixed time interval (frame). It is transmitting. For details of this method, see, for example, Araseki, Ozawa, and Ono.

'Multi-pulse Exei by Mr. Ochiai
t, ed 5peechCoder Ba5ea on
Maximum Cross-correlation
nSearch A1gorithm', (GLOB
ECOM 83. IEEE GlobalTeleee
omnicat, ion, lecture number 23.3.1.
983) (Reference 1). This method outputs a good audio signal after decoding by separately expressing the vocal tract information and the sound source signal, and by using a combination of multiple pulse trains (multipulse) as a means of expressing the sound source signal. be able to.

In addition, as a method of selecting and expressing the optimal code word from among a plurality of code words prepared in advance for the sound source signal, for example, "C0DE-EXC" by Messrs. Schroeder and Attal
ITED LINEAR PREDICTION (CE
LP) :HIGH-Q[JALITY 5PEECI
(AT VEIi'Y I, OW BIT RATES
'(ICASSP '85 lecture number 25.1.1198
5) It is described in detail in the paper entitled (Reference 2).

[Problem to be solved by the invention] 7 However, with this conventional method, the sound quality was good when the bit rate was high enough and the number of sound source pulses was sufficient, but as the bit rate was lowered, the sound quality deteriorated. There was a problem that the weight was low.

In order to improve this problem, a pitch prediction multi-pulse method has been proposed as a method of reducing the number of sound source pulses to be transmitted, which utilizes the monoperiodicity (pitch correlation) of each pitch of a multi-pulse sound source. Details of this method are disclosed in, for example, Japanese Patent Application No. 58-139022 (Document 3). 1. However, there is a limit to the number of pulses that can be reduced.

On the other hand, there are spectral parameters to be transmitted in the number of analysis orders for 2.1 frames. As a method for quantizing spectral parameters, scalar quantization, which is a well-known method, and vector quantization, which is a well-known method for efficiently quantizing, are often used. Regarding vector quantization, for example, Mr. R.M.
(R, M, Gray.

”Vector quantization for
5peech coding and recognition
on'' (J, Acoust, Soc, Ameri
ca, vol.

80,5upp1.1. Ql, 1986).

Furthermore, there is a method called U2, temporal decomposition, that efficiently expresses spectral parameters using a predetermined number of functions. Regarding temporal decomposition, see, for example, by Bee Nis Attar.
“Efficient Coding of LVC Parameters Pi Temporal Decomposition” (I C Knee Nis Nis B 83.1
4 performance number 2, 61983, reference 5) (B, S
, Atal, “EFFICIENT C0DINGO
F LPCPARAMETER5BY TEMPORA
L DECOMPOSITION” (ICASSP 8
3.2.6 pp. 81-84.1983).

However, when scalar=i quantization is used, there is a problem that more information needs to be transmitted, and if the bit rate continues to be lowered, the sound quality will deteriorate significantly. Also,
- Even when vector vectors are used, the transmission efficiency is improved, but there are problems in that the amount of calculation required to find the optimal vector is extremely large, and the quality of the reproduced sound also deteriorates. Furthermore, even if there is one temporal decomposition as described in the above-mentioned document 5, the amount of calculation required to obtain the optimal function is extremely large, and the calculation method is complicated, making it difficult to put it to practical use.

The purpose of the present invention is to provide audio encoding and decoding that can reproduce better audio than ever before even when the pit rate is high or when the bit rate is lowered, and that can be realized with a small amount of calculation. An object of the present invention is to provide a method, a speech encoding device, and a speech decoding device.

(Means for Solving the Problems) A first invention inputs a discrete audio signal, extracts a spectral parameter representing a spectral envelope and a multi-pulse sound source signal for each frame of variable time length, and In the audio encoding/decoding method, the parameter and the audio source signal are quantized and combined and transmitted, the audio source signal and the spectrum parameter are separated and restored from the combined signal, and the audio signal is reproduced and output at step 7. , having a plurality of functions in advance and selecting an optimal function over an optimal interval length using the autocorrelation of the functions and the cross-correlation between the functions and the input signal to represent the spectral parameter "ζ" shall be.

The second invention inputs a discrete audio signal, extracts spectral parameters representing the spectral envelope for each frame of variable time length, selects one type from a plurality of codewords prepared in advance, and generates a sound source. representing a signal, quantizing and combining the spectral parameters and an index representing the selected codeword, and separating and restoring the source signal and the spectral parameters from the combined signal; In the audio encoding/decoding method for reproducing and outputting, a plurality of functions are provided in advance, and an optimal one is determined using autocorrelation of the functions prepared in advance and cross-correlation between the functions prepared in advance and the input signal. The method is characterized in that the spectral parameter is expressed by selecting a function over an optimal interval length.

A speech encoding device according to a third aspect of the present invention comprises: 1 a spectral parameter calculation circuit that calculates a spectral parameter representing a short-time spectral characteristic from an input discrete speech signal sequence; 2 Spectral parameter quantum that selects and expresses the optimal one with an optimal time length from among the functions using the autocorrelation of the function prepared in advance and the cross-correlation between the function prepared in advance and the input signal. an inverse quantization circuit that inversely quantizes the quantized spectral parameters; and an inverse quantization circuit that inversely quantizes the quantized spectral parameters; and a source signal of the audio signal that is selected from a plurality of pulse trains or a codebook using the output of the inverse quantization circuit. a sound source signal calculation circuit that represents and encodes a sound source signal using a sound source signal; a code representing a function representing the spectral parameter; a time length during which the function representing the spectral parameter is valid; and a sound source signal selected from the multipulse train or a codebook. It is characterized by having a multiplexer circuit that outputs a combination of the symbols and the symbols it represents.

A voice decoding device according to a fourth aspect of the present invention inputs a code representing a sound source signal, a code representing a function representing a spectral parameter of the audio signal, and a time length in which the function representing the spectral parameter is effective, and decodes the input and separating the codes. a spectral parameter decoding circuit that decodes a spectral parameter representing a short-time spectral characteristic from a function representing the decoded spectral parameter; and a sound source signal restoration circuit that decodes the multipulse and restores the sound source signal. and a synthesis filter circuit that synthesizes the decoded audio signal over the time length using the restored sound source signal and the decoded spectral parameter.

[Operation] The present invention applies the bitge predictive multipulse coding method of the above-mentioned document 3 or the C E 1. of the above-mentioned document 2. ,, [In 3,
In order to express audio signals more efficiently than conventional methods with a small amount of transmitted information, the spectral parameters to be transmitted are selected from a plurality of predetermined functions to optimally express the input signal. Select and represent with the calculation lid.

The operation of the present invention will be explained using FIG. FIG. 3 is a block diagram of a quantizer that expresses spectral parameters using a predetermined number of functions.

In FIG. 3, from the input terminal 300, a time series y of spectral parameters to be made into a single child; (n) is input.

The codebook 250 stores a plurality of predetermined time length functions φ,(n) as code words.

In order to calculate the optimal codeword and gain, the autocorrelation calculating circuit 252 calculates the autocorrelation of the codeword, and the cross-correlation calculating circuit 251 calculates the cross-correlation between the codeword and the input signal.

Output of autocorrelation calculation circuit 252 and cross-correlation calculation - circuit 2
Using the output of 51, gain calculation circuit 260 calculates the optimal gain to be applied to the selected codeword.

The error calculation circuit 265 multiplies the selected code word with a gain, calculates the error with the input signal, and calculates the index and optimal gain of the selected code word, and the time length and time period during which the selected code word is valid. Outputs the error averaged over the length.

The codebook selection circuit 270 holds the output of the error calculation circuit 265 for each codebook7, and selects the index of the optimal codeword that is the minimum among the averaged errors calculated for all the codebooks, and the optimal code. The optimum gain corresponding to the word and the length of time during which the optimum code word is valid are output.

Hereinafter, the operation of the present invention will be explained using equations. Let y=(n) be the first time series of a plurality of spectral parameters calculated from the input speech, and the first one (hereinafter referred to as the i-th parameter time series). The function used to express this i-th parameter time series 3'i (n) is denoted by φ, and the function of the ith order among them is denoted by φ,(n). Using this, the following equation (1) holds true.

9 i (n) = ai kφ, (n)
...(1) Here! , (n) represents the time series obtained by approximating y; (n) using φ8(n), and aik is i
This is the gain at φ,(n) in the next parameter time series. Therefore, for the best approximation, E:
All you have to do is find suitable φk(n) and aik. In the conventional example, an optimal function is calculated using an error scale as shown in the above-mentioned document 5. Furthermore, in document 5,
The calculated function and the calculated gain are alternately corrected again to reduce the error.

On the other hand, in the present invention, the following error E expressed as (2) 2 is considered.

Equation (2) is the function φ for the first-order parameter time series.
, (n), & gain a + kZ The error from the input signal yi (n,) in the case of a saw is expressed by the formula 1.Here, n is the parameter to be expressed using the function φ, (ri). It represents the beginning of a time series section (frame), and Ilz represents the end.

When formula (2) is minimized for aik, 13i
y is determined by equation (3).

a i k ” −' ”””””””” --
−−−−”−”””””””””””−・ ・ ・(3
) At this time, the first-order parameter time series y;, (rl)C
The corresponding error E, has the following form: −(φ, which minimizes E, (n) should be selected so as to maximize the second term of equation (4). The function φ, (ri) may be a divine polynomial function2, or may be obtained by learning from the audio signal.

The length of time during which the optimal function φ, (ri) obtained in this way is valid is obtained by extending 02 in equation (4) by +++] degrees, and also by extending the optimal function φ, (n) together with each of them. Stretch,
The point at which the error E, which is the degree of error E expressed by the following equation (5), exceeds a predetermined threshold TH is also defined as the value of n2.

Here, in equation (5), P is the analysis order. The threshold value 1 and 11 may be fixed, or may be a value L7 that is proportional to the weight of the section in which the error E is being evaluated, ie, the difference between nl and n2. Alternatively, each value of E8 may be weighted and combined, and the value may be evaluated as the error E.

In addition, the optimal method for expanding the function φ,(n) is to uniformly expand the entire function, or to predetermine a part of the function, such as the central part of the function, as the range that can be expanded. Only the predetermined stretchable portion may be uniformly stretched.

By using the method shown below, the total amount of 1i can be significantly reduced compared to the method described in Document 5.
However, it is possible to keep the error reasonably small.

In this example, the measure for determining the error is the sum of the squared distances between each person in the next series of multiple subek [ru parameters, but in addition to kneading, for example, when calculating the sum, it is possible to use Weighted measures or other well-known distance measures can be used.

In addition, the method for determining the end point n2 of the time length at which the optimal function φ, (n) is valid is to use the error E obtained at a predetermined time interval as an initial value and sequentially extend the end point 112. 6. When the ratio of the error and the error obtained at the temporally previous end point 02' is calculated, and the ratio of the gfp error becomes larger than the predetermined threshold value T H'. −11 previous end point A method can be taken in which l is the end point 02 of the time length in which the optimal function φ□(n) is effective.

〔Example〕

FIG. 1 shows a speech encoding device and a speech decoding device that implement the speech encoding/decoding method according to the first invention. 1. In the speech encoding device, a discrete speech signal is inputted from an input terminal 500. do. Spectral parameter if” 3
In one circuit 520, the spectrum parameter representing the spectrum of the input audio signal is determined by the well-known 171'C analysis method. For the obtained spectral parameters, each codeword is calculated in the gain calculation circuit 522 using the method described in the operation section using a predetermined number of codewords prepared in the codebook 521. Calculate the optimal gain for .

The codebook selection circuit 525 selects the one with the smallest error among the outputs of the gain calculation circuit 522 for each codebook, and selects the index of the selected codeword, the optimal gain, and the selected codeword if it is valid. Output a certain length of time.

The backlighting unit 530 inversely quantizes the selected codeword output from the codebook selection circuit 525 using the index and optimal gain and outputs the result.

The weighting circuit 540 applies weighting U to the audio signal using the inverse quantized vector parameter.

Regarding the weighting method, reference can be made to the weighting circuit 200 in Japanese Patent Application No. 59-272435 (Reference 6).

The pitch parameter calculation circuit 515 calculates pitch parameters representing the fine structure of B...2...C. A method such as the one shown in the above-mentioned document 3 is used as the word one calculation method.

A quantizer 516 quantizes the determined c-pinch parameter.

The dequantizer 51B performs dequantization using the quantized result and outputs the dequantized result.

The impulse response calculation circuit 550 calculates an impulse response using the dequantized pitch parameter and the dequantized scale parameter.

For the specific method, reference can be made to the above-mentioned document 3.

The autocorrelation calculation circuit 560 calculates the autocorrelation of the impulse response and outputs it to the sound source pulse calculation circuit 580-7. The autocorrelation calculation method is described in 1. Autocorrelation function calculation circuit 18 in Reference 3.
0 can be referenced.

Phase W correlator p] circuit 5701. ,, Weighted L:,)
The phase 5-correlation between IrE-”:I- and the impulse response is calculated as 41, and the calculation circuit 580-
Output -4. For the specific method, reference can be made to the above-mentioned document 3.

The combined pulse tolerance circuit 580 obtains a predetermined number of multi-pulses by pitch η measurement. Regarding the method for calculating the multi-pulse train, refer to the sound source pulse 111 calculation circuit 210 in reference 3 above.

A quantizer 590 quantizes the excitation multipulse train and outputs a code.

A code obtained by quantizing the multi-pulse train that is the output of the quantizer 590, a code obtained by quantizing the pitch parameter that is the output of the quantizer 516, and a v-1 book selection circuit 5.
The index of the selected l-toword, which is the output of It becomes input. The multiplex 1/repeater 630 combines and outputs the following lines.

On the other hand, on the receiving side, the demaru (BL/RI4J71. The code representing the valid time length is separated and output 4''.

l-pulse decoder 720 decodes the amplitude position of the multi-pulses.

The pulse generator 730 generates a sound source signal using a multi-pulse train.

Codebook selection circuit 752 decodes the index,
The coden-1 corresponding to the decoded index is selected from the same codebook 751 as the receiving side.

The gain circuit 750 decodes the gain 7 and uses the selected codeword and the decoded gain to decode and output the Svek F. le parameter.

Pitch parameter decoder 740 functions similarly to inverse quantizer 518 on the transmit side.

The pitch recovery filter 755 receives the obtained sound source signal and the decoded pitch parameter as input, and reproduces the pitch and reproduces the synthesized sound source signal.

The subscale envelope filter circuit 760 uses the sound source signal and the decoded subscale I parameter to obtain a Y synthesized speech waveform and outputs it from an output terminal 770.

FIG. 2 shows a speech encoding device and a speech decoding device that implement the speech encoding and decoding method according to the second invention. Second
Among the components in the figure, those with the same reference numbers as in Figure 1 are referred to as the configuration Y with the same reference numbers in Figure 1.
Since the operation is the same as that of ° element, the explanation is omitted here.

In FIG. 2, a sound source two-hood selection circuit 880 uses the method described in the above-mentioned document 2 to select chords from among a plurality of chords stored in a sound source code book 895 prepared in advance. The optimal code word is selected using the 1-code word, and the index representing the 1-word and the optimal gain are output.

The quantizer 890 quantizes the optimum gain and outputs a code of 7.

The index representing the optimal code word, which is the output of the sound source code selection circuit 880, and the optimal gain, which is the output of the 7\engenerator 890, are quantized, and the pinch parameter, which is the output of the quantizer 516, is quantized. sign, and further:
The index of the selected code, the code representing the optimal key, and the length of time that the selected code word is valid, which are the outputs of the l-dog selection circuit 525, are each the output of the multiplexer 9300 Å. The multiplexer 930 combines and outputs the above codes.

On the other hand, on the receiving side, the demultiplexer 910 outputs an index representing the code word, a sign S obtained by quantizing the gain,
It separates and outputs the sign of the pinch parameter, the index of the selected j-doard, the optimum gain, and the sign -5 representing the length of time during which the selected codeword is valid.

The sound source signal decoder 920 stores the sound source codebook 890 and the sound source codebook 9 containing the same code words.
The code sword represented by the index is selected from 25, and the code sword is multiplied by the gain and combined to decode the sound source signal.

The spectral envelope filter circuit 760 uses the restored sound source signal and the decoded spectral parameters to obtain a synthesized speech waveform and outputs it from an output terminal r770.

The configurations described above are merely configurations of the present invention, and various modifications are also possible.

As the multi-pulse calculation method in the embodiment of the first invention, in addition to the method shown in Document 1, various known methods can be used.

Further, the number of multi-pulses to be obtained may be proportional to the length of time during which the spectral parameter is valid.

Further, as the spectral parameter, other well-known parameters (line spectrum pair, kebstrum melkebstrum, logarithmic cross-sectional area ratio, etc.) can also be used.

〔Effect of the invention〕

According to the present invention, by representing spectral parameters using a plurality of predetermined functions, the amount of calculation is extremely small compared to the conventional method, and the transmission information 1 is reduced compared to the conventional method.
This has the great effect of being able to represent audio signals well.

[Brief explanation of drawings]

FIG. 1 is a block diagram showing the configuration of an embodiment of a speech encoding/decoding method and apparatus according to the first aspect of the present invention, and FIG. 2 is a speech encoding method according to the second aspect of the present invention. FIG. 3 is a block diagram showing the structure of an embodiment of the decoding method and its device. FIG. 3 is a block diagram showing the operation of the present invention. 500 ... Input terminal 515 ... Pinch parameter 520 ... Svek[-le parameter calculation circuit 51
6, 590.890...Quantizer 518, 530
...Inverse quantizer 521, 751...
・ ・Codebook 522, 260...Gain calculation circuit 525, 752...Codebook selection circuit 540...Weighting circuit 550...Impulse response correction circuit 560, 252
...Autocorrelation calculation circuits 570, 251...
...Cross correlation calculation circuit 580 ...Sound source pulse calculation circuit 630, 930...Multiplexer? +, 0.
910...Demultiplexer 720...
- Sound source pulse decoder 730 ... Pulse generator 740 ... Pitch parameter decoder 750 -
... Gain circuit 755 ... Binge regeneration filter 760 ... Spectral envelope filter 770
... Output terminal 880 ... Sound source code selection circuit 895, 925... ... Sound source code book 9
20...Sound source signal decoder

Claims (4)

[Claims] (1) Input a discrete audio signal, extract a spectral parameter representing the spectral envelope and a multi-pulse sound source signal for each frame of variable time length, quantize the spectral parameter and the sound source signal, combine them, and transmit. In the audio encoding/decoding method, the method includes separating and restoring the sound source signal and the spectrum parameter from the combined signal, and reproducing and outputting the audio signal, which includes a plurality of functions in advance; A speech encoding/decoding method characterized in that the spectral parameter is represented by selecting an optimal function over an optimal interval length using correlation and cross-correlation between the function and the input signal. (2) Input a discrete audio signal, extract spectral parameters representing the spectral envelope for each frame of variable time length, and select one type from among multiple codewords prepared in advance to represent the sound source signal. , quantizing and combining the spectral parameter and an index representing the selected codeword, transmitting the combination, separating and restoring the sound source signal and the spectral parameter from the combined signal, and reproducing the audio signal. In the speech encoding/decoding method for outputting, a plurality of functions are provided in advance, and an optimal function is optimized using autocorrelation of the functions prepared in advance and cross-correlation between the functions prepared in advance and the input signal. A speech encoding/decoding method characterized in that the spectral parameter is represented by selecting it over a certain interval length. (3) a spectral parameter calculation circuit that calculates a spectral parameter representing a short-time spectral characteristic from an input discrete audio signal sequence; a spectral parameter quantization circuit that selects and expresses an optimal one using an autocorrelation of the function prepared in advance and a cross-correlation between the function prepared in advance and the input signal; and the quantized spectral parameter. an inverse quantization circuit that inversely quantizes the inverse quantization circuit; and an excitation signal that uses the output of the inverse quantization circuit to represent and encode the excitation signal of the audio signal using a plurality of pulse trains or an excitation signal selected from a codebook. a calculation circuit; a multiplexer circuit that outputs a combination of a code representing a function representing the spectral parameter, a length of time during which the function representing the spectral parameter is effective, and a code representing the sound source signal selected from the multipulse train or the codebook; A speech encoding device comprising: (4) a demultiplexer circuit that inputs, separates, and decodes a code representing a sound source signal, a code representing a function representing a spectral parameter of the audio signal, and a time length in which the function representing the spectral parameter is effective; a spectral parameter decoding circuit that decodes a spectral parameter representing a short-time spectral characteristic from a function representing the converted spectral parameter; a sound source signal restoration circuit that decodes a multipulse train or a codebook index to restore a sound source signal; An audio decoding device comprising: a synthesis filter circuit that synthesizes the decoded audio signal over the time length using the audio source signal and the decoded spectral parameter.