JP3002299B2 - Audio coding device - Google Patents

Abstract

PURPOSE:To offer the voice encoding device provided with a pitch predictor whose amount of information is small, and also, can predict a signal of pitch length of a high frequency with high accuracy. CONSTITUTION:In a CELP voice encoding processing, transmission of pitch length information is executed by plural sub-frame units, and in order to execute a determination of pitch length, an input sound signal is predicted by a linear predictor 102 at every plural sub-frames, and a linear prediction residual signal is calculated. From a past exciting signal in an adaptive code book 104, and a linear prediction residual signal, pitch length is determined 121 by setting a mutual correlation coefficient as an evaluation reference. In the case of determining a pitch prediction coefficient, a pitch prediction 123 is executed to the whole code book 124 of the pitch prediction coefficient by the pitch length derived previously at every sub-frame, and a reproducing signal is subjected to auditory weighting, and thereafter, a coefficient by which an error to an input signal becomes minimum is selected.

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a device for efficiently compressing, transmitting or storing an audio signal by making use of a pitch structure of the audio signal and performing pitch prediction.

[0002]

2. Description of the Related Art A voice signal has a considerably high correlation between adjacent samples, and means for efficiently compressing information of a voice signal using a linear prediction technique based on a signal sequence of the nearby samples is widely used. The voice signal also has repetitive redundancy depending on the pitch, and means for further improving the efficiency of information compression by pitch prediction is widely used.

As a low bit rate speech encoding method using these two linear prediction means, Code-Excited Linear Prediction (hereinafter referred to as CEL)
Described as P. ) Are being actively researched and developed (for example, "St
ochastic Coding of SpeechSignals at Very Low Bit R
ates: The Importance of Speech Perception "M, R, Sch
roeder and B, S, Atal Speech Communication 4 1985 pa
ge 155-162 North-Holland). In CELP, an excitation signal waveform is treated as a vector having a length of, for example, 40 samples, and the excitation signal waveform is compression-coded at a very low bit rate using a vector quantization technique.

A description of the basic CELP encoder system is as follows.
This is performed using FIG. In this case, a closed-loop one-tap pitch predictor, which is a known technique, is used as the pitch predictor.

First, in the CELP encoder shown in FIG. 5, a speech signal waveform inputted from an input terminal 101 is subjected to linear prediction by a linear predictor 102 at every fixed sample length (for example, 160 samples, which is called a frame). Coefficient (α)
Is calculated. The following processing is performed for each subframe length (for example, 40 samples) obtained by dividing this frame.
Using the linear prediction coefficient (α) as an input, a spectrum predictor 103 and an auditory weighting filter 111 are formed. The linear prediction coefficient is set for each subframe by interpolating from the linear prediction coefficient of an adjacent frame. This processing is known. The adaptive codebook 104 is composed of a memory holding past input signals to the spectrum predictor 103, and assumes an assumed range of the pitch length (tau) (for example, tau = 40 in the case of 8 kHz sampling) according to an instruction of the control unit 114 for each subframe. To 167) are output as pitch prediction signals. This is represented by Equation 1. Here, pit [n] is a pitch prediction signal and exc [nt]
au] represents a past input signal to the spectrum predictor 103, and a represents a pitch prediction coefficient.

[0006]

(Equation 1)

[0007] When the pitch length (tau) is shorter than the subframe length, a method of generating a pitch prediction signal by means called a virtual search is known (for example, "Improve").
d Speech Quality and Efficient Vector Quantization
in SELP "WBKleijn etc.ICASSP 1988 pp.155-158
reference).

The stochastic codebook 106 has a plurality of excitation signal waveforms (referred to as codewords) having a fixed sample length (eg, 40 samples) (eg, 1024).
The type is stored, and all codewords are sequentially output according to an instruction from the control unit 114.

The optimal pitch length (tau) and its prediction coefficient (a), and the index value of the optimal code word (index)
To determine the gain (b) thereof, a sum signal of the pitch prediction signal and the excitation signal amplified by a certain gain is input to the linear predictor 102 to obtain a reproduced signal, and the difference signal between the input signal waveform and the input signal waveform is auditory. After spectrally shaping with the weighting filter 111, the energy calculator 113 analyzes the image by an analysis method (Analys) so that the energy after shaping is minimized.
is-by-Synthesis). The above processing is a known technique.

Next, a pitch predictor used in CELP will be described.

In a predictor such as that described in Equation 1 (referred to as a one-tap pitch predictor), the pitch length (tau) can take only integer values. Therefore, the pitch frequency that can be represented has the disadvantage that it becomes very discrete at high pitch frequencies. To solve this, a multi-tap predictor (three taps are represented by two equations) and a pitch predictor using a non-integer pitch sample length (called Fractional Delay) using an oversampling method are proposed. (For example, "Improve
d Pitch Prediction with Fractional Delays in CELPC
oding "JSMarques etc. ICASSP 1990 page 665-668).

[0012]

(Equation 2)

[0013]

SUMMARY OF THE INVENTION However, CEL
If an attempt is made to reduce the amount of information transmission by P coding, the information amount of a parameter expressing the pitch length occupies a considerable part. For example, in CELP (4.8 kbps) established by the US Government Department of Defense, it is about 20%. In order to obtain a high-quality reproduced sound, it is desired to incorporate a pitch predictor using three taps or a non-integer delay. However, it is difficult to reduce the amount of information transmitted to the pitch predictor.

[0014]

The object of the present invention is achieved.
For this purpose, linear prediction means, pitch length calculation means, adaptive code
Book, prediction coefficient codebook, 3-tap pitch prediction
Means, stochastic code book, multiplication means, adder
Speech note comprising steps, filters, error calculation means, and control means
A linear prediction means for receiving an input speech signal.
Calculate linear prediction coefficient from multiple signals in multiple subframe units
The pitch calculating means outputs a plurality of subs from the input audio signal.
The pitch length is calculated for each frame, and the adaptive codebook is
Stores the past excitation signal and, based on the pitch length,
The adaptive excitation signal is read, and the prediction coefficient codebook is
(1) Store the prediction coefficient corresponding to the index value and
The prediction coefficient is read out by controlling the stage, and the 3-tap pitch
The measuring means calculates the prediction signal based on the adaptive excitation signal and the prediction coefficient.
And the Stochastic code book matches the second index value.
When storing multiple corresponding stochastic excitation signals,
The stochastic excitation signal is read under the control of the control means.
The multiplying means outputs the read stochastic excitation
A multiplication coefficient is calculated and output from the signal, and the adding means calculates the pitch
The excitation signal is output by adding the prediction signal and the output of the multiplication means.
The filter uses a filter based on linear prediction coefficients for the excitation signal.
Filter processing, and the error calculation means outputs
Calculate pitch prediction error energy, which is the error of
The control means controls the pitch prediction error energy in subframe units.
The first and second index values and the multiplication coefficient are determined so that
Set.

[0015]

In the CELP speech coding process, pitch length information is transmitted in units of a plurality of subframes, and pitch prediction coefficients are transmitted in units of subframes. Is determined by the following operation.

In order to determine the pitch length, an input speech signal is predicted by a linear predictor for each of a plurality of subframes, and a linear prediction residual signal is calculated. From the past excitation signal in the adaptive codebook and the linear prediction residual signal, the pitch length is determined using the cross-correlation coefficient as an evaluation criterion. The pitch prediction coefficient is determined based on the pitch length previously obtained for each subframe, and pitch prediction is performed for all codebooks of the pitch prediction coefficient. Select the coefficient that gives Through the above processing, the transmission of the pitch length is performed in units of a plurality of subframes, and the pitch prediction coefficient is transmitted in units of subframes.

In another invention, the linear prediction means, the adaptive
Code book, prediction coefficient code book, 3-tap pitch prediction
Measuring means, stochastic code book, multiplying means, addition
Including means, filter, error calculating means, and control means
An encoding device, wherein the linear prediction means is configured to input speech
Calculate linear prediction coefficients from signals in multiple subframe units
In addition, the adaptive codebook stores past excitation signals
Adaptive excitation signal based on the pitch length specified by the control means.
And the prediction coefficient codebook becomes the first index value.
Stores the corresponding prediction coefficient and controls the control means.
Read out the prediction coefficient from the three tap pitch prediction means,
A prediction signal is calculated based on the adaptive excitation signal and the prediction coefficient, and the
The castec codebook contains multiple index values corresponding to the second index value.
And storing the stochastic excitation signal of
Read the stochastic excitation signal by the control of
The means includes a multiplier for multiplying the read stochastic excitation signal.
The addition means calculates and outputs the pitch prediction signal and the
Output of the excitation means by adding the outputs of
Performs filtering based on linear prediction coefficients on the excitation signal.
Error calculating means calculates the error between the input audio signal and the filter.
Calculating a certain pitch prediction error energy, the control means:
Specify the pitch length in the adaptive codebook, and
Subframe unit so that prediction error energy is minimized
To determine the first and second index values and the multiplication coefficient.
Determine the pitch length in units of frames.

[0018]

1 and 2 show an embodiment of an encoder and a decoder according to the first invention of the present invention, respectively.

This device has an input terminal 10 for an input audio signal.
1, linear predictor 102, spectral predictors 103 and 12
0, pitch length calculator 121, 3-tap pitch predictor 12
3, prediction coefficient codebook 124, adaptive codebook 1
04, stochastic codebook 106, multiplier 1
07, adders 108 and 109, subtractors 110 and 122,
It comprises acoustic weighting filters 111 and 112, an energy calculator 113, a control unit 114, and an output terminal 201.

First, the operation of the encoder will be described.

The linear predictor 102 outputs a frame (for example, 1
A linear prediction coefficient (α) is calculated from the audio signal waveform input from the input terminal 101 every 60 samples). Based on the linear prediction coefficients , the prediction coefficients of the spectrum predictors 103 and 120 and the auditory weighting filters 111 and 112 are set. The linear prediction coefficient is set for each subframe by interpolating from the linear prediction coefficient of an adjacent frame.

The following processing is performed for each subframe (for example, 40 samples).

The input signal is passed through an input auditory weighting filter (this output signal is Sw [n], n = 0, 1,.
..., 39). In order to remove the influence of the past sub-frame and perform processing on the signal of the current sub-frame by using the well-known superposition addition method, first, the spectrum predictor 103 and the auditory weighting filter on the reproduction side are used. 11
After evacuating the memory in 2, the 0 of 40 samples is input and its output is subtracted from Sw [n] (this subtracted signal is represented as T [n], n = 0, 1,..., 39) ).
Thereafter, the parameters of the pitch predictor and the stochastic codebook are determined. Here, a method of sequential optimization is used.

First, optimization of the pitch prediction parameter will be described. The pitch length transmits one parameter in a plurality of subframes in order to suppress the amount of transmission information. The prediction coefficient is basically an analysis method based on synthesis, and is determined by fully searching the code book of the prediction coefficient. This will be described below.

First, pitch length determination and information transmission are performed once for a plurality of subframe lengths. Here, for example, two subframes are set. First, from the input signals for two subframes, the spectrum prediction unit 120 uses the linear prediction residual signal (Z
[N] = 0, 1,..., 79). This linear prediction residual signal and past excitation signals (exc [n] =-1, -2,
.., -167) in the assumed range of pitch length (for example, 40 to 167 samples). This method calculates a constant pitch length over a plurality of subframes.

The speech coding apparatus has a 3-tap pitch predictor as a component. The 3-tap pitch estimator has an advantage that the pitch length is interpolated and the prediction accuracy for a high pitch frequency is high, and even if the pitch length proposed here is fixed between a plurality of subframes, If the coefficients can be correctly set in subframe units, pitch prediction can be performed even for a certain change in pitch length ("Efficient Encoding of
the Long-Term Predictor in Vector Excitation Code
rs ", M. Yong and A. Gersho, Advances in Speech Coding,
Kluwer Academic Publishers). On the other hand, even if a non-integer delay is used, basically, a one-tap pitch predictor
Very sensitive to changes in pitch length. If the pitch length cannot be correctly represented by the one-tap pitch predictor, the effect of pitch prediction is almost lost.

The determination of the pitch prediction coefficient is performed for each subframe. A prediction signal is obtained by pitch prediction of three taps for all codewords of the codebook of the prediction coefficient created in advance, and the prediction signal is converted into a spectrum (clearing the contents of the memory) irrespective of the past state. The signal is input to the synthesis filter 103 and the auditory weighting filter 112 to obtain a reproduced signal based on the pitch prediction signal.
[N], n = 0, 1,..., 39). The energy of the error signal between P [n] and the previously obtained T [n] is determined by the energy calculator 113, and the control unit 114 controls the index value of the optimal pitch prediction coefficient (minimum) to minimize the error energy. Index 1) is determined. The above is the procedure for calculating the pitch prediction parameter.

With the above processing, pitch length information is transmitted for each of a plurality of subframes, and a pitch prediction coefficient is transmitted for each of the subframes.

The subsequent processing is the same as that of a known technique, and will be described briefly.

Next, the composite waveform P [n] of the previously determined optimum pitch prediction signal is subtracted from T [n], and the signal waveform of the component that cannot be represented by the pitch component (T2 [n], n =
.., 39).

The optimization of the parameters of the stochastic codebook (excitation signal vector) is performed by an analysis method based on the synthesis of the full search, similarly to the optimization of the pitch predictor. The parameters to be fully searched are the index values of the codebook (Index)
In 2), all the codewords are input to the spectrum estimator 103 having cleared the memory inside the filter and the perceptual weighting filter 112, and the output waveforms (S [n],
n = 0, 1,..., 39) and the energy of the error waveform between T2 [n] are calculated by the energy calculator 113, and the codebook parameter (Index2, b) having the smallest error energy is calculated by the control unit. 114 is determined.

After the optimal pitch prediction parameters and the parameters of the stochastic code book are determined, an optimal pitch prediction signal waveform and an optimal stochastic codeword (excitation signal waveform) are prepared for data processing of the next subframe. ) Is added to the signal amplified by the optimal prediction coefficient (this is expressed as exc [n], n = 0, 1,...,
39). This is set as a memory in the pitch predictor 104 for the next subframe. And the spectrum predictor 103 and the auditory weighting filter 112
The saved memories are reset and the combined waveforms are calculated using exc [n] as an input, so that the memories of these two filters are also updated.

With the above processing, the speech coding apparatus proposed in the present invention can be realized.

Next, the operation of the decoder will be described.

FIG. 2 shows the decoder. The information transmitted to the decoder includes a linear prediction coefficient (α) , a pitch length (ta )
u), pitch prediction gain index value (index1), stochastic codebook index value (index)
2) The gain (b) of the excitation vector.

Basically, CELP is encoded by using an analysis method based on synthesis, so that the decoding process is included in the encoding process. The procedure of the process will be described below.

First, the pitch predictor 123 outputs a pitch prediction signal based on the transmitted pitch length (tau) and the index value (index1) of the pitch prediction gain. Next, based on the index value (index2) of the codebook transmitted as information of the excitation signal, the codebook 106 outputs an excitation signal, and the gain (b) of the transmitted excitation vector
, The signal is amplified by the multiplier 107 and output to the adder 108.

The above two signals are added by the adder 108, and this is set in the adaptive codebook 104 as data for the next frame. The linear prediction coefficient transmitted
Configure the spectrum predictor 103 based on the number (α) ,
A reproduction signal is calculated by inputting the addition signal, and is output to the output terminal 201.

Next, a second embodiment of the present invention will be described.

In the first aspect of the present invention, the extraction of the pitch length over a plurality of subframes is based on the cross-correlation coefficient. In this method, optimal prediction is performed when signals of a plurality of subframes are predicted with a constant prediction coefficient in the section.
However, as described above, when a constant pitch length is used in a plurality of subframes and a prediction coefficient is set for each subframe, the method of determining the pitch length by the above method is not necessarily good.

The method of the present invention calculates the pitch prediction error energy of the spectrum prediction residual waveform for each subframe, and determines the pitch length so that the error energy in the plurality of subframes is minimized. is there. less than,
Two techniques for determining the pitch length will be described.

First, a first method calculates an optimum prediction coefficient of a 3-tap pitch predictor for each subframe for a certain pitch length, calculates a prediction error energy at the coefficient, and calculates a prediction error energy for each subframe. Are accumulated in a plurality of subframes in units of transmitting the pitch length.
The pitch length that minimizes the prediction error energy of the whole of the plurality of subframes is determined by means of full search. FIG. 3 shows the flow of this processing. In the first method, when determining the pitch length, the prediction coefficient is set by a calculated value. However, in practice, the pitch prediction coefficients can only be of the type that has been quantized and stored in the prediction coefficient codebook 124.
The second method is a method in which this point is strictly considered, and the prediction coefficients are also quantized and determined. This technique will be described.

With respect to a certain pitch length, pitch prediction is performed for all prediction coefficients for each subframe, and a search is made for a pitch length having a minimum prediction error energy. This processing is performed for a plurality of subframes in units of transmitting the pitch length, and the prediction error energy of each subframe is accumulated. The pitch length that minimizes the prediction error energy of the whole of the plurality of subframes is determined by means of full search. This processing flow is shown in FIG.
Shown in

However, the second method requires a very large amount of processing. As described in the first method, the prediction coefficient is calculated not by the full search but by the optimum prediction coefficient of the 3-tap pitch predictor for each subframe, and the minimum prediction coefficient level is calculated from the prediction coefficient codebook. A method of calculating a prediction error energy by setting a distance as a quantization value is also conceivable.

[0045]

As is clear from the above, a linear predictor for calculating a spectrum prediction parameter of an input speech signal;
A spectrum predictor, a pitch predictor, a codebook storing a plurality of excitation signal waveform vectors, an auditory weighting filter for shaping the spectrum of a difference signal between the input audio signal and the reproduced signal, and an auditory weighted difference An energy calculator that calculates the energy of the signal, and a code excitation linear prediction encoder including a control unit that optimally sets parameters of a pitch predictor and a codebook so that an output value of the energy calculator is minimized. Means for calculating a linear prediction residual signal, an adaptive codebook, a pitch length calculator, and a codebook of 3-tap pitch prediction coefficients,
By transmitting the pitch length in units of a plurality of subframes and transmitting the pitch prediction coefficient in units of subframes, it is possible to suppress the information transmission amount of the pitch length low and accurately predict a high-frequency pitch signal.

[Brief description of the drawings]

FIG. 1 is a block diagram of an embodiment of an encoder of a speech encoding device according to the present invention.

FIG. 2 is a block diagram of an embodiment of a decoder of the speech encoding device of the present invention.

FIG. 3 is a diagram for explaining a first pitch length determining method according to the present invention.

FIG. 4 is a diagram illustrating a second pitch length determination technique according to the present invention.

FIG. 5 is a block diagram of a conventional CELP speech encoding device.

[Explanation of symbols]

 101 input terminal 102 linear predictor 103, 120 spectral predictor 104, 123 pitch predictor 105, 107 multiplier 106 stochastic codebook 108, 109 adder 110, 122 subtractor 111, 112 auditory weighting filter 113 energy calculator 114 control unit 121 pitch length calculator 124 prediction coefficient codebook 301 output terminal

Claims (2)

(57) [Claims] A linear prediction means for calculating a pitch length;
Means (120-122), adaptive codebook (10
4) Prediction coefficient codebook (124), 3 tap
H prediction means (123), stochastic code book
(106), multiplication means (107), addition means (10
8), filters (103, 109, 112), error calculation
Means (110, 113) and control means (114)
An audio coding apparatus, wherein a linear prediction means (102) performs decoding from an input audio signal.
The linear prediction coefficient is calculated in units of a number of subframes, and the pitch calculation length output means (120-122) outputs the input speech signal.
, The pitch length is calculated for each of a plurality of subframes, and the adaptive codebook (104) stores the past excitation signal.
Read the adaptive excitation signal based on the pitch length
And the prediction coefficient codebook (124) corresponds to the first index value.
The prediction coefficient to be stored is stored and controlled by the control means.
The prediction coefficient is read, and the 3-tap pitch prediction means (123)
The prediction signal is calculated based on the prediction coefficient, and the stochastic code book (106) calculates the second index value.
When storing multiple stochastic excitation signals corresponding to
Both are controlled by the control means (114).
The excitation signal is read, and the multiplication means (107) reads the read stochastic excitation.
The multiplication coefficient is calculated and output to the vibration signal, and the adding means (108) outputs the pitch prediction signal and the multiplication means (1
07) to output an excitation signal, and the filters (103, 109, 112) apply a line to the excitation signal.
The filter processing is performed based on the shape prediction coefficient, and the error calculation means (110, 113) outputs
Pitch which is the error of the filter (103, 109, 112)
The prediction error energy is calculated, and the control means (114) calculates the pitch prediction energy in subframe units.
First and second index value to measurement error energy becomes minimum, multiply
A speech coding device for determining arithmetic coefficients. 2. A linear prediction means (102) comprising:
(104), prediction coefficient codebook (124), 3
Tap pitch prediction means (123), Stokastec Co
Book (106), multiplication means (107), addition means
(108), filter (103, 109, 112), wrong
The difference calculation means (110, 113) and the control means (114)
A speech coding apparatus comprising, a linear prediction unit (102) is double from the audio signal input
A linear prediction coefficient is calculated in units of a number of subframes, and the adaptive codebook (104) stores a past excitation signal.
Based on the pitch length specified by the control means (114).
The adaptive excitation signal is read out according to the prediction coefficient codebook (124) corresponding to the first index value.
The prediction coefficient to be stored is stored and controlled by the control means.
The prediction coefficient is read, and the 3-tap pitch prediction means (123)
A prediction signal is calculated based on the prediction coefficient, and the stochastic code book (106) calculates the second index value.
When storing multiple stochastic excitation signals corresponding to
Both are controlled by the control means (114).
The excitation signal is read, and the multiplication means (107) reads the read stochastic excitation.
The multiplication coefficient is calculated and output to the vibration signal, and the adding means (108) outputs the pitch prediction signal and the multiplication means (1
07) to output an excitation signal, and the filters (103, 109, 112) apply a line to the excitation signal.
The filter processing is performed based on the shape prediction coefficient, and the error calculation means (110, 113) outputs
Pitch which is the error of the filter (103, 109, 112)
The prediction error energy is calculated, and the control means (114) stores the prediction error energy in the adaptive codebook (104).
Specify pitch length and pitch prediction error energy
, The first and second index values in subframe units so that
Determine the multiplication factor and pitch length in units of multiple subframes
Speech coding device that determines