JPH04149600A - Voice decoding system - Google Patents

Abstract

PURPOSE:To detect an error in pitch cycle on a reception side without adding redundancy to sent data by delaying the data on a decoding side by a pitch period D received from an encoding side. CONSTITUTION:On the encoding side, the maximum pitch period D among pitch periods in a pitch period search range and the gain (g) at the maximum pitch period are found, and multiplexed with other parameters and sent to a decoder. The pitch period D is received by the decoder and supplied to a buffer 40. Then pitch vectors P1 and P2 of the pitch period D and a pitch period 2D which is double as long as it are extracted from an adaptive code book 10 and the normalization correlation coefficient r12 between the both is calculated from an equation V12=(P1,P2)/(P1,P1). Then an error presence/absence detection part 50 receives the normalization correlation coefficient r12 from the buffer 40 to detect whether there is an error or not. The correlation coefficient r12 is compared with with its threshold value rth and when r12<rth, it is judged that the period D has an error, so that a repair processing part 60 performs repair processing.

Description

[Detailed Description of the Invention] [Summary] Regarding a highly efficient audio decoding method for audio signals transmitted after information compression, the pinch period and pitch prediction coefficient obtained as a result of pitch prediction analysis on the encoding side are In a method for decoding audio received as a coding parameter, the purpose is to detect the presence or absence of a transmission path error with respect to the pitch period without adding redundancy. Calculate the correlation function between the obtained past pitch vector and the past pitch vector obtained from the adaptive codebook by delaying the received pitch period by an integral number of times, and based on the magnitude of this correlation function value. The system is configured to detect whether or not there is an error in the pitch period value due to a transmission path error.

[Industrial application field]

The present invention relates to an audio decoding system, and more particularly to a highly efficient audio decoding system for audio signals transmitted after information compression.

In recent years, in corporate communication systems, digital mobile radio systems, voice storage systems, and the like, there has been a demand for a method for transmitting voice signals in a highly efficient compressed form while maintaining the quality of the voice signals.

C. Prior Art] When transmitting an audio signal, on the side that predicts and encodes the audio, short-term prediction coefficients extracted by short-term (linear) prediction analysis and pitch prediction coefficients extracted by long-term prediction analysis are used for each frame. A common method is to multiplex and transmit the pitch period and a prediction residual signal obtained by applying an inverse characteristic filter (inverse filter) of the short-term prediction filter and the long-term prediction filter to the input signal.

In addition, in order to efficiently transmit this predictive residual signal information, a method (code-driven linear predictive code Method: CELP
), a method (multi-pulse drive coding method: MPC), etc. in which the prediction residual vector is modeled with a finite number of pulse trains and the optimal pulse position and pulse amplitude are transmitted are used.

FIG. 5 shows the pitch period search method used in the above-mentioned long-term prediction analysis on the encoding side, where 1 is an adaptive code that stores the linear prediction residual signal gv from the frame delay device 2 one frame before. A linear prediction residual signal gVi-0 obtained by multiplying a signal delayed by D samples using this adaptive codebook 1 by a pitch prediction coefficient (gain) g is linearly predicted and reproduced. The filter 3 reproduces the auditory weighting.

Then, the subtracter 4 calculates the error between the reproduced signal gy1 and the input audio signal χyo, and the pitch period search unit 5 calculates the power of this error signal. This error power is E.

=Σ　(X, li ...(1) It is expressed as

however, I =v, -, +Σaj)'1-j X; input audio signal V2 past linear prediction residual signal a: linear prediction coefficient p: order of linear prediction It is.

The gain g that minimizes the above equation (1) can be obtained by differentiating both sides with respect to g and setting it to zero, so d Ee /
d g=0 2Σ (xt g)'I I =0 ΣXI')'Direct pa・g8 ......(2) Σ y.

becomes.

Therefore, Error power E9 when 3 is as follows Become.

ED −1(xt gyr )” If the second term on the right side of equation (3) is maximized, the error power E
Since D is the minimum, the pitch period search unit 5 may sequentially shift the illustrated pitch period by, for example, l samples, and select the pitch period when the second term on the right side becomes the maximum as the optimal pitch period. It turns out.

Then, the adaptive codebook 1 is partially updated with the residual vector of the previous frame from the frame delayer 2 at the time when the pitch period search for one frame is completed.

Note that the sound source vector 7 shown in the figure is a code vector selected from a white noise fixed (stochastic) codebook used for short-term predictive analysis in the CELP encoding method described above, multiplied by a gain. CE
In the sequential optimization method in the LP method, it is given as zero, and in the simultaneous optimization method, as shown in the figure, it is added to the pitch period vector multiplied by the gain g in the adder 8 to obtain the residual vector. Become.

This kind of analysis method (^analysis-by-5yn
According to Lhesis (AB-5 method), on the encoding side, the audio signal is passed once in each pitch period and the one with the smallest error is selected.

[Problem to be solved by the invention]

When using such audio encoding systems in environments with many transmission path errors, such as mobile communications, error correction codes are generally used on the decoding side to suppress quality deterioration when errors occur. Therefore, it is necessary to use both methods and techniques to repair erroneous parameters.

In this case, the parameter restoration process involves interpolating or repeating temporally preceding and subsequent parameters to restore the erroneous parameters, but these processes degrade the playback audio quality when there are no errors. In order to let
When used in conjunction with error detection coding, it is necessary to process only parameters with errors, and parameters that exhibit smooth linear characteristics over time, such as short-term prediction coefficients extracted by short-term prediction analysis. , it is possible to detect the existence of an error when a count value that significantly deviates from the existing range of coefficients estimated from the previous and subsequent coefficients is received, and the parameter repair process can be performed relatively easily. I can do it.

However, as shown in Figure 6, the pitch periods extracted by the long-term prediction analysis described above frequently include double pitches in addition to the basic period, and even triple pitches or more. Therefore, it is not a linear characteristic,
Therefore, conventionally, it is necessary to add some redundancy to detect errors in the pitch period, but the original purpose of audio coding is to remove redundancy from the audio signal and compress information. The application of error detection coding, which will be added, will significantly degrade the reproduced audio quality due to excessive compression of pitch information, which is one of the most important parameters in the voiced part of speech. There was a problem that it was not desirable from the viewpoint.

Therefore, the present invention provides a method for decoding received speech using pitch periods and pitch prediction coefficients obtained as a result of pitch prediction analysis on the encoding side as encoding parameters, without adding redundancy. The purpose is to detect the presence or absence of transmission path errors.

[Means and actions for solving the problem] The optimal pitch period determined by the pitch period search method as shown in Figure 5 is such that the second term on the right side of equation (3) is maximized. This indicates that the correlation value between the input signal vector X and the pitch period vector y in each pinch period (perceptually weighted) by the linear prediction filter 3 is maximum. .

Therefore, in a portion where the input signal X has a periodic waveform such as a voiced section of speech, in addition to its basic pitch period, correlation values at periods that are integral multiples of the basic period, such as double pitch period and triple pitch period, etc. will be similarly high.

That is, as shown in FIG. 2, the correlation value of equation (3) (second term on the right side) for each pitch period when the basic period is 60 samples is, in addition to the basic period of 60 samples, the correlation value for the double pitch period ( 120 samples), the triple pitch period (180 samples), and the intermediate period thereof also have peaks in the correlation function value. f@, pitch period in Figure 6 above (N axis)
corresponds to the double pitch circumference jlI (horizontal axis) shown in FIG.

Therefore, in the present invention, by utilizing the characteristics shown in FIG. 2, as shown in FIG. 1, the decoding side is delayed by the pitch period received from the encoding side. Calculate the correlation function between the past pitch vector obtained from the adaptive codebook l and the past pitch vector obtained from the adaptive codebook 1 by delaying the reception pitch period by an integral multiple (Sl); and , based on the magnitude of this correlation function value, detect whether there is an error in the pitch period value due to a transmission path error (
S2).

This allows pitch period errors to be detected on the receiving side without adding redundancy to the transmitted data.

Furthermore, in the present invention, instead of the pitch vector stored in the adaptive code 4!1, a reproduced signal vector obtained by passing the residual vector through a reproduction filter is used to detect errors in the received pitch period due to similar correlation function values. Detection may also be performed.

Furthermore, if the above correlation calculation is performed only when the pitch prediction coefficient exceeds the threshold, the above correlation calculation for unvoiced parts whose pitch prediction coefficient is lower than the threshold can be omitted.
It is possible to efficiently detect the presence or absence of pitch cycle errors, and it is also possible to avoid erroneous detection when the correlation function value is large in an unvoiced part.

〔Example〕

FIG. 3 shows an embodiment of a decoder for realizing the speech decoding method according to the present invention shown in FIG. 1. In this embodiment, the above-mentioned A
-B-3 method is used, and the pitch period search range is assumed to be 20-147 samples as shown in FIG.

In the figure, 10 is an adaptive codebook corresponding to the adaptive code II (see FIG. 5) on the encoding side, and 20 is a frame delay device that partially updates the adaptive codebook 10 by delaying the residual vector by one frame. , 30 is a linear prediction filter that perceptually weights the residual vector to generate a reproduction vector, 40 is a correlation function calculation buffer that takes in the residual signal stored in the adaptive codebook 10 according to the pitch period, and 50 is the buffer 40. An error detection section 60 detects the presence or absence of an error in the pitch period based on the calculated correlation function and the gain (pitch prediction coefficient) g sent from the receiving side. A repair processing unit for repairing, 70 is an excitation vector multiplied by a gain consisting of a coat' vector selected from a stochastic codebook of white noise as described above, and 80 is an excitation vector 70 and a pitch period vector. (gv) to generate a residual vector.

Next, the operation of this embodiment will be explained.

First, on the encoding side, the second term on the right side of the above equation (3) is calculated for each pinch period without a pitch period search range,
The pinch period at which this becomes the maximum and the gain g at that time are determined, multiplexed with other parameters, and transmitted to the decoder shown in FIG.

Then, this decoder receives the pitch period and provides it to the buffer 40. In the buffer 40, this pitch circumference 3
Each pitch with a pitch period of 2D, which is twice that of IIID,
The vectors P and P2 are extracted from the adaptive code [10] and the normalized correlation coefficient r11 between them is calculated as shown in the following equation [4].
Calculate.

Next, the error detection unit 50 receives the normalized correlation coefficient r+z from the buffer 40 and performs error detection as shown in FIG.

That is, first, the gain g transmitted from the receiving side is compared with a predetermined threshold g (step TI).

As a result of this comparison, if g<g, the received audio signal is considered to be an unvoiced part, so no error detection is performed.
The pitch circumference MD is used as is (step T2).

On the other hand, in the case of gag, it is considered to be a voiced part, so the normalized correlation coefficient of the pitch vector between the pitch period calculated in the buffer 4o by the above equation (4) and the double pitch P1 period 2D rI! is compared with its threshold value r (step T3), and when r+I<r, it is a voiced part, but 2
Since the correlation with the double pitch cycle is small, it is assumed that there is an error in the received pitch cycle (step T4), and the repair processing unit 6o performs repair processing. As this repair process, for example, a method such as replacing with a recent pitch period D' without an error is performed, and a residual vector gVo. and a reproduced vector gFe. are obtained.

Also, when rlz>r, it is assumed that there is no transmission path error, and a detection output of "no error" is generated, and no post-processing is performed.In this case, the contribution of the pitch vector to the reproduced audio is small. Therefore, the influence of errors is small and can be ignored.

In the above embodiment, a correlation calculation is performed using the pitch vector to detect whether there is an error. A gain g is applied to the reproduced signal vector obtained through
Similar detection may be performed by calculating the correlation function value between the residual vector multiplied by the linear prediction filter 30 and the reproduced signal vector obtained.

〔Effect of the invention〕

As explained above, according to the present invention, the past pitch vector obtained from the adaptive codebook by delaying the reception pitch period and the past pitch vector obtained from the adaptive codebook by delaying the reception pitch period by an integral multiple Since the configuration is configured to calculate the correlation function between the past pitch vector and detect the presence or absence of an error in the pitch circumference M value due to transmission error based on the magnitude of this correlation function value, an error detection code is used. Since errors in pitch period information can be detected on the receiving side without adding redundancy such as can be converted into

[Brief explanation of drawings]

FIG. 1 is a diagram conceptually illustrating the audio decoding method according to the present invention; FIG. 2 is a diagram illustrating an example of correlation function values at each pitch period in audio with a predetermined pitch period (60 samples); FIG. 3 is a block diagram showing an embodiment of the audio decoding method according to the present invention, FIG. 4 is a flowchart showing the error detection process used in the audio decoding method according to the invention, and FIG. FIG. 6 is a block diagram conceptually showing a pitch period search method for an input audio signal on the encoding side, which has been conventionally common. FIG. 6 is a diagram showing an example of a change in pitch period value over time.
It is. In the figure, 10 is an adaptive codebook, 20 is a frame delay device, 30 is a linear prediction filter, 40 is a correlation function calculation buffer, and 50 is an error detection unit. In the figures, the same reference numerals indicate the same or corresponding parts.

Claims (3)

[Claims] (1) In a method for decoding received speech using the pitch period and pitch prediction coefficient obtained as a result of pitch prediction analysis on the encoding side as encoding parameters, adaptation is achieved by delaying the received pitch period (D). The correlation function between the past pitch vector obtained from the codebook (1) and the past pitch vector obtained from the adaptive codebook (1) by delaying the reception pitch period (D) by an integral multiple is calculated. 1. An audio decoding method characterized in that the presence or absence of an error in a pitch period value due to a transmission path error is detected based on the magnitude of the correlation function value. (2) The audio decoding method according to claim 1, wherein a reproduced signal vector obtained by passing a residual vector through a reproduction filter is used instead of the pitch vector. (3) The audio decoding method according to claim 1 or 2, wherein the correlation calculation is performed only when the pitch prediction coefficient exceeds a threshold value.