JPH0481199B2 - - Google Patents

Description

DETAILED DESCRIPTION OF THE INVENTION The present invention relates to a low bit rate waveform encoding system for audio signals, and particularly to an encoding system and apparatus for reducing the amount of transmitted information to 10 kbit/sec or less. An effective method for encoding an audio signal with a transmission information amount of about ok bits/second or less is to minimize the error between the input signal and the signal reproduced using the driving excitation signal sequence of the audio signal. As a condition,
A method of searching at short intervals is well known.
These methods are called tree coding (TREE CODING) and vector quantization (VECTORQUANTIZATION) depending on their search method. In addition to these methods, a plurality of pulse sequences representing the driving sound source signal sequence are transmitted at short intervals.
On the encoder side, A-b-S( A NALYSIS- B Y-
Recently, a method has been proposed that attempts to obtain the information sequentially using the SYNTHESIS method. The present invention relates to this method. For details on this method, please refer to B. S. Attar ((BS
Eye by ATAL) et al. C. A. S.
S. Proceedings of ICASSP, 1982 614
"A. New. Model. of..." published on page 617.
L. P. C. Excitement. Huo. Producing. Natural. Sounding. speech. Att. Row. Bit. Reitz” (“A NEW MODEL OF LPC”
EXCITATION FOR PRODUCING
NATURAL−SOUNDING SPEECH AT
LOW BIT RATES”) (Reference 1)
Since it has been explained previously, a brief explanation will be provided here. FIG. 1 is a block diagram showing the processing on the encoder side in the conventional system described in Document 1. In the figure, 100 indicates an encoder input terminal, into which an A/D converted audio signal sequence x(n) is input. Reference numeral 110 denotes a buffer memory circuit, which stores an audio signal sequence for one frame (for example, in the case of 8KHZ sampling and assuming a frame length of 10 msec, 80 samples). The output value of 110 is the subtracter 12
0 and is output to the K parameter calculation circuit 180. However, according to Reference 1, the reflection.. is used instead of the K parameter. Although it is described as REFLECTION COEFFICIENTS, this is the same parameter as the K parameter. The K parameter calculation circuit 180 includes 11
Using an output value of 0, the K parameter K _i representing the audio signal spectrum for each frame is calculated according to the covariance method.
are obtained for 16 orders (1≦i≦16), and these are output to the synthesis filter 130. 140 is a sound source pulse generation circuit, which generates a predetermined number of pulse sequences in one frame. Here, this pulse sequence is denoted as d(n). An example of a sound source pulse sequence generated by the sound source pulse generation circuit 140 is shown in FIG. In FIG. 2, the horizontal axis indicates discrete time, and the vertical axis indicates amplitude. Here, a case is shown in which eight pulses are generated within one frame. The pulse sequence d(n) generated by the sound source pulse generation circuit 140 is passed through the synthesis filter 1
Drive 30. The synthesis filter 130 inputs d(n), obtains a reproduced signal x~(n) corresponding to the audio signal x(n), and outputs this to the subtracter 120. here,
The synthesis filter 130 inputs the K parameters K _i , converts them into prediction parameters a _i (1≦i≦16), and uses a _i to calculate the reproduced signal x˜(n). x~
(n) can be expressed as shown below using d(n) and a _i . x ~ (n) = d (n) + _p 〓 ^{i = 1} a _i x ~ (ni) - (1) In the above formula, P indicates the order of the synthesis filter, and here P = 16. The subtracter 120 receives the original signal x~
The difference e(n) between (n) and the reproduced signal x(n) is calculated and output to the weighting circuit 190. 190 inputs e(n), uses the weighting function w(n), and calculates the weighting error e _w (n) according to the following equation. e _w (n)=w(n) ^* e(n) −(2) In the above equation, the symbol “*” represents a convolution integral. Furthermore, the weighting function w(n) performs weighting on the frequency axis, and its Z-transformed value is expressed as W(Z).
Then, using the prediction parameter a _i of the synthesis filter, it is expressed by the following equation. W(Z)=(1- _p 〓 ⁱ⁼¹ a _i Z ^-i )/(1- _p 〓 ⁱ⁼¹ a _i・r ⁱ .Z ^-i ) −(3) In the above equation, r is 0≦r It is a constant of ≦1 and determines the frequency characteristics of W(Z). In other words, if r=1,
W(Z)=1, and the frequency characteristics become flat.
On the other hand, when r=0, W(Z) has a frequency characteristic opposite to that of the synthesis filter. Therefore, the characteristics of W(Z) can be changed depending on the value of r. Furthermore, as shown in equation (3), W(Z) is determined depending on the frequency characteristics of the synthesis filter because an auditory masking effect is utilized. In other words, this is due to the perceptual property that even if the error with the spectrum of the reproduced signal is a little large, the error is hard to notice at a location where the input audio signal has a large spectral power (for example, near a formant). Figure 3 shows the input audio signal spectrum in a certain frame and W
An example of the frequency characteristics of (Z) is shown. Here r=
It was set to 0.8. In the figure, the horizontal axis is the frequency (up to 4K
Hz), and the vertical axis shows logarithmic amplitude (maximum 60 dB). Further, the upper curve represents the spectrum of the audio signal, and the lower curve represents the frequency characteristic of the weighting function. Returning to FIG. 1, the weighted error e _w (n) is fed back to the error minimization circuit 150. The error minimization circuit 150 stores the value of e _w (n) for one frame, and uses this to calculate the weighted double error ε according to the following equation. ε _N 〓 ⁿ⁼¹ e _w (n) ² −(4) Here, N indicates the number of samples for calculating the squared error. In the method of Reference 1, this time length is set to 5 msec.
This corresponds to N=40 in the case of 8KHz sampling. Next, the error minimization circuit 15
0 provides pulse position and amplitude information to the sound source pulse generation circuit 140 so as to reduce the squared error ε calculated by the above equation (4). 140 generates a sound source pulse sequence based on this information. The synthesis filter 130 uses this sound source pulse sequence as a driving source to calculate a reproduction signal x~(n). Next, subtractor 12
0, the currently determined reproduced signal x~(n) is subtracted from the previously calculated error e(n) between the original signal and the reproduced signal,
Let this be the new error e(n). Weighting circuit 19
0 inputs e(n), calculates a weighted error e _w (n), and feeds it back to the error minimization circuit 150. The error minimization circuit 150 calculates the squared error again, and adjusts the amplitude and position of the sound source pulse sequence to reduce the squared error. A series of processes from generation of the sound source pulse sequence to adjustment of the sound source pulse sequence by error minimization are repeated until the number of pulses in the sound source pulse sequence reaches a predetermined number, and the sound source pulse sequence is determined. This concludes the explanation of the conventional method. In this method, the information to be transmitted is the K parameter K _i (1≦i≦16) of the synthesis filter, and the pulse position and amplitude of the sound source pulse sequence, depending on the number of pulses generated in one frame. Therefore, any transmission rate can be achieved. Furthermore, the transmission rate
It is considered to be one of the effective methods since good playback quality can be obtained for the range of 10Kbps or less. However, this conventional method has the disadvantage that the amount of calculation is extremely large. When calculating the position and amplitude of a pulse in a sound source pulse sequence, this calculates the error and square error between the reproduced signal and the original signal based on the pulse, and feeds them back to reduce the square error. This is due to the fact that the pulse position and amplitude are adjusted accordingly. Furthermore, this is caused by repeating this process until the number of pulses reaches a predetermined value. Furthermore, according to this conventional method, at a bit rate of about 0 Kbps or less, in the case of an input signal with a high pitch frequency, for example, when a female voice is input,
The drawback was that the playback quality deteriorated. This means that if the pitch frequency is high, many pitch waveforms will be included in the pulse calculation frame, and in order to reproduce this pitch waveform well, it is necessary to
This is because a larger number of sound source pulses are required compared to the case of a speaker with a low pitch frequency. Therefore, for this reason, it has been difficult to significantly lower the transmission bit rate, that is, to significantly reduce the number of pulses within one frame. The purpose of the present invention is to
The purpose of this invention is to provide a high-quality audio encoding system that can be applied to bit rates of 10Kbps or less and its equipment. According to the present invention, on the transmitting side, a discrete audio signal sequence is input, and parameters representing a short-time spectrum including pitch fine structure are extracted and coded, and based on the parameters, a discrete audio signal sequence is input, and a parameter representing a short-time spectrum including a pitch fine structure is extracted and encoded. Calculate the autocorrelations of the impulse response sequence, calculate the crosscorrelations according to the audio signal sequence and the impulse response sequence, and use the autocorrelations and the crosscorrelations to determining and encoding a driving excitation signal sequence for the signal sequence, and outputting a combination of a code representing the driving excitation signal sequence and a code representing the parameter;
On the receiving side, the code sequence is input, and the code sequence representing the driving excitation signal sequence and the parameter code sequence representing the short-time spectrum including the fine structure of the pitch are separated and decoded, and the code sequence representing the driving excitation signal sequence is separated and decoded. A voice encoding method is obtained, characterized in that the voice signal sequence is reproduced using the signal sequence and the decoded parameters. Further, according to the present invention, a parameter calculation circuit inputs a discrete audio signal sequence and extracts and encodes a pitch parameter representing a pitch fine structure and a spectral parameter representing a short-time spectral envelope from the audio signal sequence; an autocorrelation function calculation circuit that inputs the output sequence of the parameter calculation circuit and calculates the autocorrelations of the impulse response sequence according to the short-time spectrum including the pitch structure of the audio signal sequence; a cross-correlation calculation circuit that receives an output sequence of the parameter calculation circuit and calculates a cross-correlation represented by the audio signal sequence and the impulse response sequence corresponding to the short-time spectrum; and the autocorrelation calculation circuit. a driving excitation signal sequence calculation circuit that inputs the output sequence of the output sequence and the output sequence of the cross-correlation coefficient calculation circuit to obtain and encode a driving excitation signal sequence for the audio signal sequence; and an output code sequence of the parameter calculation circuit. There is obtained a speech encoding device characterized in that it has a multiplexer circuit that combines and outputs the output code sequence of the drive excitation signal sequence calculation circuit. Further, according to the present invention, a parameter calculation circuit inputs a discrete audio signal sequence and extracts and encodes a pitch parameter representing a pitch fine structure and a spectral parameter representing a short-time spectral envelope from the audio signal sequence; an autocorrelation calculation circuit that inputs the output sequence of the parameter calculation circuit and calculates the autocorrelation of the impulse response sequence according to the short-time spectrum including the pitch structure of the audio signal sequence; a cross-correlation calculation circuit that receives an output sequence of the parameter calculation circuit and calculates a cross-correlation represented by the audio signal sequence and the impulse response sequence corresponding to the short-time spectrum; and the autocorrelation calculation circuit. a drive excitation signal sequence calculation circuit which inputs the output sequence of the above-mentioned cross-correlation coefficient calculation circuit and the output sequence of the cross-correlation coefficient calculation circuit, and calculates and encodes a drive excitation signal sequence for the audio signal sequence; an output code sequence of the parameter calculation circuit; a multiplexer circuit that combines and outputs the output code sequence of the driving excitation signal sequence calculation circuit; and a multiplexer circuit that receives the code sequence obtained by the combination and outputs a code sequence representing the driving excitation signal sequence and a code sequence representing the pitch parameter. a demultiplexer circuit that separates the code sequence representing the spectral parameter; a driving excitation decoding circuit that inputs and decodes the code sequence representing the driving excitation signal sequence obtained by the separation; A parameter decoding circuit inputs and decodes a code sequence representing a pitch parameter and a code sequence representing the spectral parameter, and an output sequence of the drive excitation decoding circuit and an output of the parameter decoding circuit are used to reproduce and output an audio signal sequence. There is obtained a speech encoding/decoding device characterized in that it has a synthesis filter circuit. First, the sound source pulse calculation algorithm according to the present invention will be explained in detail. The sound source pulse sequence d(n) at any time n within one frame is expressed by the following equation. d(n)= _K 〓 ⁱ⁼¹ g _i・δ _o,ni (5) Here, δ _o,ni represents Kronetzker's delta, which is 1 when n=mi and 0 when n≠mi. It is. K indicates the number of pulses generated within one frame. g _i indicates the amplitude of the i-th pulse, and m _i indicates the position of the i-th pulse. Next, consider a filter that can represent the spectral structure of the audio signal, including the fine structure of the pitch, as a synthesis filter. This filter can be represented by a cascade of a pitch prediction filter and a spectral envelope prediction filter. A block diagram is shown in Figure 4. In the figure, 191 indicates a pitch prediction filter, and 192 indicates a spectral envelope prediction filter. There are two possible pitch prediction filters: a first-order pitch prediction filter and a high-order pitch prediction filter. Here, for the sake of simplicity, we will consider a case where a first-order pitch prediction filter is used. The reproduction signal x~(n) obtained by driving a synthesis filter consisting of a cascade connection of a pitch prediction filter and a spectral envelope prediction filter using the sound source pulse train d(n) can be expressed as the following equation. I can do it. x~(n)=d(n)+β・x _d (n−M _d )+ _p 〓 ⁱ⁼¹ a _i・x−(ni) −(6) Here, β is the tap coefficient of the pitch prediction filter , and M _d indicates the pitch period of the input signal. l _d (n)
indicates the pitch prediction filter output signal. Also, P
is the envelope prediction order of the spectral predictor and a _i
(1≦i≦P) indicates the prediction coefficient of the spectral envelope predictor. Various methods are known for calculating the tap coefficient β and pitch period M _d of the pitch predictor, but a simple method is, for example, a method of extracting the peak amplitude and its position of the autocorrelation sequence of the input audio signal. well known. For more information on this method, please refer to BSATAL, M.R. Schulder, M.R.
BELL SYSTEM Technical Journal by Mr. SCHROEDER
TECHNICAL JOURNAL) magazine, October 1970 issue,
“ADAPTIVE PREDITIVE CODING” published on pages 1973-1986.
OF SPEECH SIGNALS” (Reference 2), so the explanation is omitted here. Now, let us express the impulse response of the composite filter consisting of the pitch prediction filter and the spectral envelope prediction filter h(i ) (0≦i≦M−1; where M indicates the number of continuous samples of the impulse response), the reproduced signal x~(n) can also be written as the following equation: x~(n )=d(n)*h(n) −(7) Next, the input audio signal x(n) and the reproduced signal x~(n) are 1
The weighted multiplicative error J within a frame can be expressed as follows. J= _N 〓 ⁿ⁼¹ [{x(n)−x〜(n)}*w(n)] ² −(8) Here, w(n) is the impulse response of the weighting circuit, for example, as shown in Fig. 1 It has the same characteristics as the conventional weighting circuit shown in . Further, N indicates, for example, the number of samples in one frame. Next, an algorithm for calculating a sound source pulse train that minimizes the weighted squared error J shown in equation (8) will be derived. First, substitute equation (7) into equation (8) to obtain the following equation. J= _N 〓 ⁿ⁼¹ [{x(n)−d(n)*h(n)}*w(n)] ² −(9) Here, each term on the right side of the above equation is expressed as the following equation. , x _w (n)=x(n)*(n) (10) h _w (n)=h(n)*w(n) −(11) Equation (5), Equation (10), (11) Substituting equation (9) for equation (9), we obtain the following equation. J= _N 〓 ⁿ⁼¹ [x _w (n)− _K 〓 ⁱ⁼¹ g _i・h _w (n−m _i )] ² −(12) The sound source pulse sequence that minimizes equation (12) is ( It is calculated from the following equation obtained by partially differentiating equation 12) with the amplitude g _i of the sound source pulse sequence and setting it to 0. Here, _xh (・) represents the cross-correlation sequence calculated from x _w (n) and h _w (n), and _hh (・) represents the autocorrelation sequence calculated from h _w (n). It can be written like an expression. Note that _hh (·) is often called a covariance sequence in the field of audio signal processing. According to equation (13), the amplitude g _i of the sound source pulse sequence is a function of its position m _i , and it is possible to calculate the optimum amplitude g _i when emitting a pulse at the position m _i . In addition, the position m _i of the sound source pulse can be calculated using equation (13) as (12)
Minimize the squared error J=R _xx (0) − _K 〓 ⁱ⁼¹ g _i・_xh (−m _i ) −(15) obtained by substituting it into the formula, that is, maximize the second term on the right side. All you have to do is choose a location. Furthermore, as an approximate method, a position that maximizes |g _i | may be selected. In equation (15), R _xx (0) indicates the power of the weighting signal x _w (n). Now, if we ignore the influence of the edge of the frame, the covariance function _hh (m _l , m _i ) shown in equation (15) depends on the time difference (|m _l −m _i |) as shown in the following equation. It can be assumed that it is equal to the autocorrelation sequence R _hh (|m _l −m _i |). _hh (m _l , m _i )=R _hh (|m _l −m _i |) −(16) Here, R _hh (・) can be expressed as in the following equation. R _hh (｜m _l −m _i ｜)= _N-( | _nl-ni | ₎ 〓 〓 ⁿ⁼¹ h _w (n-m _l ), h _w (n-m _i ), (1≦m _l , m _i ≦N)
−(17) Therefore, equation (13) is modified as follows using equations (16) and (17). Calculating R _hh (・) requires approximately 1/N the amount of calculation compared to calculating _hh (・,・). Therefore, by using equation (18) to calculate the sound source pulse sequence, the amount of calculation can be reduced to 1/N compared to equation (13). However(18)
When calculating the sound source pulse train according to the formula, if the number of data samples used to calculate the cross-correlation number _xh (・) is not larger than the number of samples in the frame that transmits the pulses, errors will occur in the pulses near the edges of the frame. . Therefore, this problem can be avoided by selecting the number of data samples for calculating the cross-correlation number _xh (.) to be larger than the number of samples of the frame. This concludes the derivation of the sound source pulse calculation algorithm and the explanation regarding its characteristics. Next, a speech encoding method using the sound source pulse calculation algorithm according to the present invention will be explained in detail with reference to FIG. FIG. 5a is a block diagram showing an embodiment of the transmitting side of the speech encoding system according to the present invention, and FIG. 5b is a block diagram showing an embodiment of the receiving side.
In FIG. 5a, a discrete audio signal sequence x(n) is input from an input terminal 195, divided into a predetermined number of samples, and stored in a buffer memory circuit 340. When dividing the input audio signal sequence, the input audio signal sequence is divided with an overlap of a predetermined number of samples. This is because, as described above, the number of data samples used for sound source pulse calculation is made larger than the number of frame samples. Next, the K parameter calculation circuit 28
0 inputs a sequence of a predetermined length among the audio signal sequences stored in the buffer memory circuit 340, and uses this to calculate LPC parameters of a predetermined order P by a well-known method (for example, Calculated according to the linear predictive analysis method).
Various LPC parameters can be considered, but below we will use the K parameter K _i (1≦i≦P)
The explanation will proceed assuming that . The K parameter is the same parameter as the Percoll coefficient. K
The parameter K _i is the K parameter encoding circuit 200
is output to. K parameter encoding circuit 200
encodes K _i based on, for example, a predetermined number of quantization bits, and outputs the code l _ki to gate circuit 460 . Furthermore, the K parameter encoding circuit 200 decodes l _ki and sends k _i ′ to the impulse response calculation circuit 210 and the synthesis filter circuit 400 .
Output to. Next, the pitch analysis circuit 370 inputs the output series of the buffer memory circuit 340, calculates the pitch period M _{d and the pitch gain β according to the method described in the above-mentioned document (2), and calculates the pitch period M d} and the pitch gain β.
Output to 80. The pitch encoding circuit 380 encodes the pit period M _d and the pitch gain β using a predetermined number of bits and outputs _ld and l 〓 to the gate circuit 460 . Further, the pitch encoding circuit 380 outputs M _d ' and β' obtained by decoding l _d and l 〓 to the impulse response calculation circuit 210 and the synthesis filter circuit 400 . Next, the impulse response calculation circuit 210 converts the K parameter decoded value K _i ' to the K parameter encoding circuit 2
00, and also input the decoded values M _d ′ and β′ of the pitch period and pitch gain to the pitch encoding circuit 3.
Enter from 80. Impulse response calculation circuit 21
0 calculates two types of impulse responses. First, using only the K-parameter decoded value K _i ′, the weighted impulse response h _1w in the case of only the spectral envelope prediction filter 192 among the synthesis filters shown in FIG. 4 is calculated using a predetermined number of samples. and outputs the determined h _1w (n) to the autocorrelation number calculation circuit 360 and the cross correlation number calculation circuit 350. Next, the K parameter decoded value
Using K _i ′ and the pitch information (pitch gain β′ and pitch period M _d ′), the weighted impulse response h _2w (n) of the composite filter consisting of the pitch prediction filter and the spectral envelope prediction filter is determined in advance. After calculating the number of samples and completing the correlation calculation process using h _1w (n), it is output to the autocorrelation number calculation circuit 360 and the cross correlation number calculation circuit 350. Next, the autocorrelation calculation circuit 360 inputs the impulse response h _1w (n) and calculates h _1w according to the above-mentioned equation (17).
The autocorrelation sequence R _hh1 (| _ml −m _i |) of (n) is calculated and output to the pulse calculation circuit 390. Next, after completing the above calculation, input the impulse response h _2w (n) and calculate the autocorrelation sequence R _hh2 of h _2w (n).
(|m _l −m _i |), the pulse calculation circuit 390
Output to. Next, the subtracter 285 converts the buffer memory circuit 3
Input the audio signal sequence x(n) accumulated in 40,
The output sequence of the synthesis filter circuit 400 is subtracted by one frame sample from (n), and the subtraction result is output to the weighting circuit 410. The weighting circuit 410 has K
The K-parameter decoded value K _i ' is inputted from the parameter encoding circuit 200, and the weighting function W(n) is calculated so that its z-transformed value is expressed by equation (3). This may be calculated using other frequency weighting methods.
Furthermore, the weighting circuit 410 inputs the subtraction result of the subtracter 285, performs a convolution operation with this and the weighting function w(n), and outputs the obtained x _w (n) to the cross-correlation calculation circuit 350. do. The cross-correlation calculation circuit 350 inputs the impulse response h _1w (n) from the impulse response calculation circuit 210 and calculates the first cross-correlation using h _1w (n) and the above-mentioned x _w (n). Calculate the number _xh1 (−m _i ) (1≦m _i ≦N),
This is output to the pulse calculation circuit 390. Next, input the impulse response h _2w (n) and calculate h _2w (n) and x _w (n)
The second cross-correlation number _xh2 (−m _i )1
≦m _i ≦N) and sends it to the pulse calculation circuit 3.
Output to 90. The pulse calculation circuit 390 receives the cross-correlation numbers and the autocorrelation functions in synchronization. That is, at the beginning within a predetermined frame period, the first cross-correlation number _xh1 (-m _i ) and the autocorrelation number R _hh1 (|m _l
−m _i |), and using the aforementioned sound source pulse calculation formula (18), calculate the amplitude of the first sound source pulse train.
A predetermined number of g _i and positions m _i are calculated. Next, after the above processing is completed, the second cross-correlation number _xh2 (−m _i ) and the autocorrelation number
Synchronize and input R _hh2 (｜A _l −m _i ｜) and
The amplitude and position of the second sound source pulse train are calculated according to equation (18). Furthermore, the pulse calculation circuit 390
The power of the error signal between the input signal and the reproduced signal is calculated for each of the first sound source pulse train and the second sound source pulse train according to the following equation. J= _N 〓 ⁿ⁼¹ {(x(n)−x〜(n))＊w(n)) ² =R _xx (0)− _K 〓 ⁱ⁼¹ g _i・_xh (−m _i ) −( 19) The above equation can be obtained by substituting equation (18) into equation (9). Here, R _xx (0) is the weighting circuit 4
The power of the output value x _w (n) of 10 is shown. The error signal power calculated for the first sound source pulse train and the second sound source pulse train is output to the discrimination circuit 430. In addition, instead of formula (U), the residual of the cross-correlation number obtained as a result of calculating the sound source pulse train (subtracting the value found by the pulse and autocorrelation number from the cross-correlation number for each pulse) The power of the error signal can also be approximately calculated using the obtained final value) according to the following equation. J= _N 〓 ^{m=1 2} _xh (-m)/R _hh (0) -(20) Next, the discrimination circuit 430 determines which of the first sound source pulse train and the second sound source pulse train is the input signal. The purpose is to select a sound source pulse train that can be faithfully represented. Therefore, the discrimination circuit 430 compares the power of the error signal for each sound source pulse train input from the pulse calculation circuit 390. If the value for the first sound source pulse is smaller than the value for the second sound source pulse, it is determined that using the first sound source pulse example has better characteristics than using the second sound source pulse train. Since the first sound source pulse train is calculated from the first impulse response without using pitch information (in other words, β'=0, M _d '=0), the discrimination circuit 430 uses the switching circuit 44
0, the first sound source pulse train is encoded by the encoding circuit 4.
70. In addition, the discrimination circuit 4
30 causes the gate circuit 460 to output the K parameter code l _ki to the multiplexer 450 .
Also, the codes (l _d and l〓) representing pitch information are:
A predetermined sign is set in gate circuit 460 and output to multiplexer 450. In the opposite case, it is determined that using the second sound source pulse train has better characteristics than using the first sound source pulse train. Since the second sound source pulse train is calculated from the second impulse response using spectral envelope information and pitch information, the discrimination circuit 430 instructs the switching circuit 440 to transmit the second sound source pulse train to the encoding circuit. 470. Further, the discrimination circuit 430 sends the K parameter code l _ki and the codes l _d , l 〓 representing pitch information to the multiplexer 460 .
output to 50. Next, the encoding circuit 470 switches the switching circuit 44
From 0, input the amplitude and position of the sound source pulse train,
These are encoded using normalization coefficients described later.
In addition, the normalization coefficients are also encoded, and the normalization coefficients,
A code representing the amplitude and position of the sound source pulse train is output to multiplexer 450. It also outputs decoded values g _i ′, m _i ′ of the amplitude and position of the sound source pulse train to the sound source pulse generation circuit 420 . Here, encoding circuit 4
70 can be encoded in various ways. one,
One method is to encode the amplitude and position of a pulse train separately, and the other is to encode the amplitude and position together. An example of the former method will be explained. First, as a method for encoding the amplitude of the sound source pulse, the maximum value of the amplitude of the pulse sequence in the frame is used as a normalization coefficient, and after normalizing the amplitude of each pulse using this value, it is quantized and encoded. There are possible ways. Another possible method is to assume that the amplitude probability distribution is a normal type and use an optimal quantizer for the normal type. Regarding this, please refer to IRE Transactions on Information Theory (IRE) by J. MAX.
TRANSACTIONS ON INFORMATION
“QUANTIZING FOR MINIMUM DISTORTION” published in the March 1960 issue of THEORY, pages 7-12.
MINIMUMDISTOTORTION") (Reference 3), so the explanation is omitted here. Next, various methods can be considered for encoding the pulse position. For example, facsimile signal encoding You may also use a run-length code, etc., which is well known in the field of Conventionally well-known logarithmic compression encoding etc. can be used to encode the coding coefficients. Note that the encoding method of the pulse sequence is not limited to the encoding method described here, but also the best known method. Returning to FIG. 5, the pulse sequence generation circuit 420 uses the input g _i ′ and m _i ′ to generate an amplitude at the position m _i ′.
A sound source pulse sequence having g _i ' is calculated over one frame length N, and is output to the synthesis filter circuit 400 as a drive signal. Synthesis filter circuit 4
00 is K from the K parameter encoding circuit 200.
Input the parameter decoded value K _i ′. Further, pitch information (pitch period decoded value M _d ′ and pitch gain decoded value β′) is input from the pitch encoding circuit 380 . The K parameter decoded value K _i ′ is converted into the predicted parameter a _i
(1≦i≦N _p ) using a well-known method. Further, discrimination information is input from the discrimination circuit 430. When using the first sound source pulse train described above, the pitch information is set to 0. Next, the synthesis filter circuit 400 receives the sound source pulse generation circuit 420 from the sound source pulse generation circuit 420.
Input one frame worth of driving sound source signal and
One frame worth of zeros is added to the frame worth of signals, and a response signal sequence x~(n) for the two frames of signals is obtained. This is shown in the following equation. x~(n)=d(n)+β・x _d (n−M _d )+ _p 〓 ⁱ⁼¹ a _i・x〜(n−i) −(21) Here, the driving sound source signal d(n) is , 1≦n≦N,
It represents a pulse sequence output from the pulse generation circuit 420, and represents a sequence of all 0s when N+1≦n≦2N. Also, at time N+1≦n≦2N,
For a _i , M _d , and β used in equation (21), values determined at the current frame time may be used, or values determined at the next frame time may be used. x~ calculated according to formula (21)
Among (n), x of the second frame ~ (n) (N+1≦n≦
2N) is output to the subtracter 285. Next, the multiplexer 450
70 and the output code of the gate circuit 460 are input, and these are combined and output from the transmission side output terminal 480 to the communication path. This completes the explanation of the encoder side of the audio encoding system according to the present invention. Next, the receiving side of the speech encoding system according to the present invention will be explained with reference to FIG. 5b. The demultiplexer 500 has a receiving side input terminal 4
Enter the code from 90. Demultiplexer 5
00 separates the code sequence representing the K parameter, the code sequence representing the pitch information, and the code sequence representing the excitation pulse train from the input code, and sends the code sequence representing the K parameter to the K parameter decoding circuit 5.
20, and the code sequence representing pitch information is output to
The code sequence representing the excitation pulse train is output to the pitch decoding circuit 510, and the code sequence representing the excitation pulse train is output to the excitation pulse decoding circuit 530. The K-parameter decoding circuit 520 and the pitch decoding circuit 510 decode the input code sequence and output it to the synthesis filter circuit 550. The sound source pulse decoding circuit 530 inputs a code sequence representing the sound source pulse train, decodes it, and outputs the amplitude and position information of the sound source pulse train to the pulse generating circuit 540.
Output to. The pulse generation circuit 540 inputs the amplitude and position information of the sound source pulse train, generates a sound source pulse train, and outputs it to the synthesis filter circuit 550. The synthesis filter circuit 550 is a cascade of a pitch prediction filter and a spectrum prediction filter, as shown in FIG. The synthesis filter circuit 550 and the pitch decoding circuit 510 input pitch information and K parameter decoded values from the K parameter decoding circuit 520. If the pitch information has a predetermined sign, the synthesis filter circuit 550 that reproduces the signal using only the spectrum prediction filter (that is, the pitch information is set to 0) drives the output pulse train of the pulse generation circuit 540. The signal x~(n) is reproduced as a source and output from the receiving side output terminal 560. This completes the explanation of the decoder side according to the present invention. According to the present invention, since the sound source pulse sequence is calculated according to equation (18), the synthesis filter is driven by pulses, as seen in the conventional method of Reference 1,
There is no path to find the reproduced signal and adjust the pulse by feeding back the error and squared error with the original signal, and there is no need to repeat that process, so the amount of calculation can be significantly reduced, which is good. This has the great effect of providing excellent playback quality. Furthermore, in the calculation of equation (18), _xh (−m _i ) and R _hh (|
By calculating the value of m _l −m _i |) (1≦|m _l −m _i |≦N) in advance for each frame, the calculation of equation (18) It is no longer necessary to perform a correlation calculation every time . Furthermore, compared to other conventional methods for searching for a sound source pulse train, the method according to the present invention has the advantage that better quality can be obtained for the same amount of transmitted information. Further, according to the present invention, the periodicity of the input audio signal, that is, the periodicity of the sound source pulse sequence is used to calculate the sound source pulse sequence using parameters that can reproduce the spectral structure including the pitch structure of the input signal. Since it is possible to predict the sound source pulses that are separated by the pitch period on the sound source pulses, this method has the effect of significantly reducing the number of sound source pulses required to obtain the same characteristics compared to the conventional method.
Therefore, it is extremely effective in reducing the amount of transmitted information.
This also has the effect of improving reproduction quality when the amount of transmitted information is the same as in the conventional method. In particular, for female speakers with high pitch frequencies, which was a problem in the conventional system, good reproduction quality can be obtained even with a transmission information amount of 10 Kbps or less. Further, according to the present invention, the encoder side compares the excitation pulse train calculated using only the spectral parameters and the excitation pulse train calculated using the pitch parameter as well, and transmits the pulse train that can more faithfully reproduce the input signal. Since this is configured to be used for reproduction on the receiving side, it is possible to prevent deterioration caused by non-periodic frames in the transient portion of the input audio signal or errors in pitch parameter extraction. Note that, as a simpler method, it is also possible to adopt a configuration in which the determination is made using the pitch gain β of the pitch parameter. For example, after calculating the pitch gain β, β is compared with a predetermined threshold, and if β is less than the threshold, β is forcibly set to 0. In this case, the source pulse will be calculated using only the spectral parameters. With such a configuration, a discrimination circuit for comparing and discriminating sound source pulse sequences and calculation of the above-mentioned equation (19) or (20) are unnecessary, and the amount of calculation can be reduced. Furthermore, in the sound source pulse calculation method shown in equation (18), suboptimal pulses are calculated one by one. In this method, when calculating the next pulse,
From this, it is also possible to use a method of readjusting the amplitudes of a plurality of pulses found in the past. This method is effective when the pulses are not independent, that is, when the positions of the pulses are determined very close to each other. Furthermore, various other sound source pulse calculation methods can be considered. For example, a method may be used in which the amplitudes of all pulses are readjusted after all pulses within one frame have been determined. Further, according to the present invention, there is a great effect that there is almost no deterioration of the reproduced signal near the frame boundary due to waveform discontinuity at the frame boundary. This effect is achieved by extending the response signal sequence obtained by driving the synthesis filter using the excitation pulse sequence of one frame past to the current frame when calculating the excitation pulse sequence of the current frame on the encoder side. This is due to the configuration in which the sound source pulse sequence of the current frame is calculated from the result of subtracting this from the input audio signal sequence.
Further, in this embodiment, the case where the frame length is constant has been described, but of course the same effect can be obtained by using a variable length frame in which the frame length is changed over time. Furthermore, according to the configuration of the embodiment of the present invention, as a method of obtaining a response signal sequence derived from a sound source pulse sequence of one frame past, pitch information input one frame past is used as a filter parameter of the response signal calculation circuit. Although the K parameter values are used as they are, pitch information and K parameter values input at the current frame time may be used when calculating the response signal sequence derived from the sound source pulse of the past frame. Further, according to the present invention, in the synthesis filter circuit 400 on the transmitting side, pitch information is used or not in accordance with the determination result of the determination circuit 430 when determining a response signal sequence derived from a sound source pulse of one frame past. However, the pitch information may be always set to 0 when calculating the response signal sequence. Also, on the encoder side, a synthesis filter circuit 4
When calculating the response signal sequence derived from the sound source pulse of the past frame in 00, the response signal sequence calculated using only the K parameter and the response signal sequence calculated using the K parameter and pitch information are different. Calculate two types of response signal sequences, and decide which response signal sequence should be used in the next frame. For example, calculate the power of the weighted error between the input signal and each response signal sequence. , the characteristics are further improved by selecting the one with smaller error power. However, with such a configuration, the selection information necessary for selecting one of the two types of response signal sequences on the decoder side must be transmitted in extra bits for each frame. In this case, the increase in the amount of transmitted information is 50 bits/second when the frame length is 20 msec.
A very small amount is required. Further, according to the configuration of the embodiment of the present invention, the weighting circuit 410 on the transmitting side performs weighting according to equation (3) used in the conventional system. This weighting is related to the spectral envelope and does not include weighting using the pitch structure. Therefore, by using the weighting function W _p (n) that utilizes both the spectral envelope and pitch structure shown in the following equation, more effective weighting can be performed. Here, W _p (z) is the Z-transformed representation of the weighting function w p (n), r and r' are weighting coefficients, and 0 <
A value of r, r′<1 is chosen. Further, according to the present invention, the determination circuit 43 on the encoder side
0, when determining which of the two sound source pulse trains should be used for better characteristics, the power of the weighted error signal obtained by equations (19) and (20) is used as the criterion for judgment. did. Other best methods can be used as criteria. For example, a configuration may be adopted in which prediction gains are calculated when pitch prediction is performed, and these values are compared with a predetermined threshold value to be used as a judgment criterion. Further, in the present invention, when calculating the autocorrelation sequence of the impulse response sequence representing the short-time spectral structure, the impulse response calculation circuit 210 uses the K parameter decoded value and the pitch information to calculate the impulse response sequence. After calculating,
The autocorrelation sequence was calculated using this impulse response sequence. As is well known in the field of digital signal processing, the autocorrelation sequence of the impulse response sequence corresponds to the power spectrum of the short-time spectrum. Therefore, a configuration may be adopted in which the power spectrum of the short-time spectrum is determined using the K-parameter decoded value and the pitch information, and the autocorrelation sequence is calculated. On the other hand, when calculating the cross-correlation sequence between the audio signal sequence and the impulse response sequence representing the short-time spectral envelope, in the configuration of this embodiment, the signal sequence x _w (n), which is the output value of the weighting circuit 410, is , impulse response calculation circuit 2
A cross-correlation calculation circuit 350 calculates the cross-correlation using the impulse response sequence obtained in step 10. As is well known, the cross-correlation number corresponds to the cross-power spectrum. Using this relationship, the cross-power spectrum may be obtained using the audio signal sequence, the K-parameter reconstruction value, and the pitch information, and the cross-correlation sequence may be calculated. Regarding the correspondence between the power spectrum and the autocorrelation sequence, and the correspondence between the cross power spectrum and the cross-correlation sequence, please refer to A.V. Otzpenheim (AV
“DIGITAL SIGNAL PROCESSING” by Mr. OPPENHEIM et al.
Since it is explained in detail in Chapter 8 of the monograph titled (Reference 4), the explanation will be omitted here. In addition, in the embodiment of the present invention described above, the encoding of the sound source pulse sequence within one frame is performed by the encoding circuit 47 in FIG. 5 after all pulse sequences have been determined.
Although encoding was performed using 0, the encoding is included in the calculation of the pulse sequence, and each time one pulse is calculated,
A configuration may also be used in which encoding is performed and the next pulse is calculated. By adopting such a configuration, a pulse sequence that minimizes errors including encoding distortion can be found, so that quality can be further improved. Regarding code assignment in the encoding circuit 470, in the configuration of the present invention, the encoding efficiency is considerably improved by variable length code assignment rather than equal length code assignment. This is because, by determining the sound source pulse train using pitch information, the amplitude distribution of the sound source pulse train becomes even more skewed. In addition, in the embodiments described above, the K parameter was used as the parameter representing the spectral envelope of the short-time audio signal sequence, but other well-known parameters (for example, LSP parameters) may also be used. good. Furthermore, the weighting function w(n) may not be included in the above equation (8). In addition, in this embodiment, in order to prevent quality deterioration due to discontinuity of the reproduced waveform at frame boundaries, a response signal sequence derived from the sound source pulse one frame past the current frame is calculated, and the input audio of the current frame is After subtracting this response signal from , the driving sound source pulse was calculated, as shown in Figure 6.
The data used for the sound source pulse calculation may include data of a frame that transmits the pulse and data from the past. In Figure 6,
N _T indicates a frame for transmitting pulses, and N indicates a frame for calculating source pulses. Such a configuration has the effect that it is not necessary to calculate a response signal sequence derived from a sound source pulse one frame past.

[Brief explanation of the drawing]

Figure 1 is a block diagram showing the configuration of the conventional system, Figure 2 is a diagram showing an example of a sound source pulse sequence, and Figure 3 shows the frequency characteristics of the input audio signal sequence and the frequency characteristics of the weighting circuit shown in Figure 1. Diagram showing an example, No. 4
Figure 5 shows an example of a synthesis filter used to explain the sound source pulse calculation algorithm according to the present invention.
The figure is a block diagram showing an embodiment of the speech encoding system according to the present invention, and FIG. 6 is a diagram for explaining the positional relationship between the pulse transmission frame and the sound source pulse calculation frame. In the figure, 110, 340... buffer memory circuit, 120, 285... subtraction circuit, 130,
400,550...Synthesis filter circuit, 140,
420, 540...Sound source pulse generation circuit, 150
... Error minimization circuit, 180, 280 ... K parameter calculation circuit, 190, 410 ... Weighting circuit, 200 ... K parameter encoding circuit, 191
... Pitch prediction filter, 192 ... Spectrum envelope prediction filter, 210 ... Impulse response calculation circuit, 350 ... Cross correlation calculation circuit, 360 ...
... Autocorrelation calculation circuit, 370 ... Pitch analysis circuit, 380 ... Pitch encoding circuit, 390 ... Pulse calculation circuit, 430 ... Discrimination circuit, 440 ...
Switching circuit, 470... Encoding circuit, 450...
...Multiplexer, 460 ... Gate circuit, 50
0... Demultiplexer, 510... Pitch decoding circuit, 520... K parameter decoding circuit, 530
... Each shows a sound source pulse decoding circuit.

Claims (1)

[Claims] 1. On the transmitting side, a discrete audio signal sequence is input, parameters representing a short-time spectrum including pitch fine structure are extracted and encoded, and based on the parameters, a signal is encoded according to the short-time spectrum. Calculate the autocorrelations of the impulse response sequence, calculate the crosscorrelations according to the audio signal sequence and the impulse response sequence, and use the autocorrelations and the crosscorrelations to A driving excitation signal sequence for the signal sequence is determined and encoded, a code representing the driving excitation signal sequence and a code representing the parameter are combined and output, and the receiving side inputs the code sequence to represent the driving excitation signal sequence. Separate and decode the code sequence and the code sequence of the parameter representing the short-time spectrum including the pitch fine structure, and use the decoded drive excitation signal sequence and the decoded parameter to generate the audio signal sequence. 1. A voice encoding method characterized in that the voice encoding method reproduces 2. A parameter calculation circuit that inputs a discrete audio signal sequence and extracts and encodes a pitch parameter representing a pitch fine structure and a spectral parameter representing a short-time spectral envelope from the audio signal sequence, and an output sequence of the parameter calculation circuit. an autocorrelation calculation circuit that calculates an autocorrelation coefficient of an impulse response sequence according to a short-time spectrum including a pitch structure of the audio signal sequence; and an output sequence of the audio signal sequence and the parameter calculation circuit. a cross-correlation coefficient calculating circuit that receives the input signal and calculates a cross-correlation coefficient represented by the audio signal sequence and the impulse response sequence according to the short-time spectrum;
a drive excitation signal sequence calculation circuit that inputs the output series of the autocorrelation coefficient calculation circuit and the output series of the cross-correlation coefficient calculation circuit to obtain and encode a drive excitation signal sequence for the audio signal sequence; and the parameter A speech encoding device comprising: a multiplexer circuit that combines and outputs the output code sequence of the calculation circuit and the output code sequence of the drive excitation signal sequence calculation circuit. 3. A parameter calculation circuit that inputs a discrete audio signal sequence and extracts and encodes a pitch parameter representing a pitch fine structure and a spectral parameter representing a short-time spectral envelope from the audio signal sequence, and an output sequence of the parameter calculation circuit. an autocorrelation calculation circuit that calculates an autocorrelation coefficient of an impulse response sequence according to a short-time spectrum including a pitch structure of the audio signal sequence; and an output sequence of the audio signal sequence and the parameter calculation circuit. a cross-correlation calculation circuit that calculates the cross-correlation numbers represented by the audio signal sequence and the impulse response sequence corresponding to the short-time spectrum; a driving excitation signal sequence calculation circuit that inputs the output sequence of the correlation coefficient calculation circuit, obtains and encodes a driving excitation signal sequence for the audio signal sequence, and calculates the output code sequence of the parameter calculation circuit and the driving excitation signal sequence. a multiplexer circuit that combines and outputs the output code sequences of the circuit; and a multiplexer circuit that receives the code sequence obtained by the combination and represents the code sequence representing the drive excitation signal sequence, the code sequence representing the pitch parameter, and the spectral parameter. a demultiplexer circuit that separates the code sequence, a driving excitation decoding circuit that inputs and decodes the code sequence representing the driving excitation signal sequence obtained by the separation, and a code representing the pitch parameter obtained by the separation. a parameter decoding circuit that inputs and decodes a sequence and a code sequence representing the spectral parameter, and a synthesis filter that reproduces and outputs an audio signal sequence using the output sequence of the driving excitation decoding circuit and the output of the parameter decoding circuit. A speech encoding/decoding device comprising a circuit.