JPH0425560B2 - - 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 apparatus that reduces the amount of transmitted information to 10K bits/second or less.

As an effective method for encoding an audio signal with a transmission information amount of about 10K bits/second or less, the driving sound source signal sequence of the audio signal is conditioned on the minimum error between the signal reproduced using the sequence and the input signal. A method of searching every short period of time is well known. These methods include tree coding (TREECODING) and vector quantization (VECTOR) depending on the search method.
QUANTIZATION). In addition to these methods, a plurality of pulse sequences representing the driving excitation signal sequence are transmitted at short intervals on the encoder side.
A-b-S (ANALYSIS-BY-SYNTHESIS)
Recently, a method has been proposed that attempts to find the value sequentially using the following method. The present invention relates to this method. For details on this method, please refer to the ICAS website by BSATAL et al.
Proceedings of SP), 1982, pp. 614-617, ``A New Model of LBC Excitement for Producing Natural Sounding Speech''.
“At Low Between Rates” (“ANEW
MODEL OF LPC EXCITATION FOR
PRODUCING NATURAL−SOUNDING
SPEECHAT LOW BIT RATES") (Reference 1), so we will briefly explain it here.

FIG. 1 is a block diagram showing processing on the encoder side in the conventional system described in Document 1. In the figure, 100 indicates a code input terminal, into which an A/D converted audio signal sequence Xn is input. 100
is a buffer memory circuit, which stores the audio signal series in one
Accumulates frames (for example, 10 msec, 80 samples in the case of 8KHz sampling). The output value of 110 is sent to a subtracter 120 and a K parameter calculation circuit 18.
It is output as 0. However, according to Document 1, REFLECTION COEFFICIENTS is described instead of the K parameter, but this is the same parameter as the K parameter. K parameter calculation circuit 180
uses an output value of 110 and follows the covariance method:
The K parameter Ki representing the audio signal spectrum for each frame is 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). 140
An example of the sound source pulse sequence generated by
As shown in the figure. In FIG. 2, the horizontal axis indicates discrete time and the vertical axis indicates amplitude. Here, the case where eight pulses are generated within one frame is shown. The pulse sequence generated by 140
dn drives the synthesis filter 130. The synthesis filter 130 inputs d(n), obtains a reproduced signal (n) corresponding to the audio signal x(n), and sends this to the subtracter 12
Output to 0. Here, the synthesis filter 130 is
Input K parameters Ki, convert them to prediction parameters ai (1≦i≦16), and use ai to calculate (n)
Calculate. (n) can be expressed as shown below using d(n) and ai.

X ^~ (n) = d (n) + _P 〓 ^{i = 1} aiX ^~ (ni) - (1) In the above formula, p indicates the order of the synthesis filter, and here p = 16. The subtracter 120 extracts the original signal x(n)
The difference e(n) ^between and the reproduced signal X is calculated and output to the weighting circuit 190. 190 inputs e(n),
Using the weighting function ω(n), the weighting error eω(n) is calculated according to the following equation.

eω(n)=ω(n)*e(n) −(2) In the above equation, the symbol “*” represents a convolution integral. Furthermore, the weighting function ω(n) performs weighting on the frequency axis, and its Z-transformed value is expressed as W(z).
Then, using the prediction parameter ai 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≦1 is a constant that determines the frequency characteristics of W(Z). In other words, if r=1,
W(z)=1, and the frequency characteristic becomes 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). FIG. 3 shows the spectrum of the input audio signal in a certain frame,
An example of the frequency characteristics of W(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 eω(n) is fed back to the error minimization circuit 150. The error minimization circuit 150 stores the values of eω(n) for one frame, and uses them to perform weighting 2 according to the following equation.
Calculate the multiplicative error.

ε= _N 〓 ⁿ⁼¹ eω(n) ² −(4) Here, N indicates the number of samples for calculating the squared error. In the method of Document 1, this time length is set to 5 msec, which corresponds to N=40 in the case of 8KHz sampling. Next, the error minimization circuit 150
gives 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 reproduced signal ^x . Next, the subtractor 120
Then, the error e between the original signal and the reproduced signal calculated earlier is
The currently determined reproduced signal x ^~ is subtracted from (n), and this is set as a new error e(n). Weighting circuit 190
inputs e(n), calculates a weighted error eω(n), and feeds this back to the error minimization circuit 150. 150 again calculates the squared error ε, and adjusts the amplitude and position of the sound source pulse sequence to reduce it. In this way, 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 of the sound source pulse sequence reaches a predetermined number.

This concludes the explanation of the conventional method.

In the case of this method, the information to be transmitted is the K parameter Ki (1≦i≦16) of the synthesis filter and the pulse position and amplitude of the sound source pulse sequence. Any transmission rate can be achieved. Furthermore, the transmission rate
Good playback quality can be obtained in the region of 10Kbps or less, and it is considered to be an effective method.

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 squared error between the reproduced signal and the original signal based on the pulse, and uses these as feedback to adjust the pulse position and amplitude. This is due to what you are doing. Furthermore, this is caused by repeating this series of processes until the number of pulses reaches a predetermined value.

The purpose of the present invention is to
The object of the present invention is to provide a high-quality audio encoding device that can be applied to transmission rates of 10 Kbps or less.

A speech encoding device according to a second aspect of the present invention includes means for inputting a discrete speech signal sequence, dividing the speech signal sequence into short time periods to obtain a short time speech signal sequence, and calculating a spectral envelope from the short time speech signal sequence. means for extracting and encoding parameters representing the spectral envelope; means for calculating an impulse response sequence based on the parameters representing the spectral envelope; and means for calculating a correlation sequence using the impulse response sequence;
means for calculating a target signal sequence with predetermined corrections based on the short-time audio signal sequence; means for calculating a cross-correlation sequence using the impulse response sequence and the target signal sequence; means for calculating and encoding a driving excitation signal sequence of the short-time audio signal sequence using the correlation sequence and the cross-correlation sequence; and means for calculating and encoding the driving excitation signal sequence of the short-time audio signal sequence, and and a means for outputting the combination with the code representing the information.

A speech encoding device according to a first aspect of the present invention includes means for inputting a discrete speech signal series, dividing the speech signal series into short time periods to obtain a short time speech signal series, and calculating a spectral envelope from the short time speech signal series. means for extracting and encoding parameters representing the spectral envelope; means for calculating an impulse response sequence based on the parameters representing the spectral envelope; and means for calculating a correlation sequence using the impulse response sequence;
means for calculating a correlation sequence using the impulse response sequence and the short-time audio signal sequence; and a driving sound source signal sequence of the short-time audio signal sequence using the autocorrelation sequence and the cross-correlation sequence. and means for outputting a combination of the code of the parameter representing the spectral envelope and the code representing the drive excitation signal sequence.

The speech encoding method according to the present invention is characterized by an algorithm for calculating a sound source pulse sequence. In the following, this algorithm will therefore be explained in particular detail.

First, the sound source pulse sequence dn at an arbitrary time n within one frame is expressed by the following equation.

d(n)= _K 〓 ^k=1 _gk・δ _o,nk −(5) Here, δ _o,nk represents Kronetzker's delta, which is 1 when n=mk and 0 when n≠mk. It is. Furthermore, gk represents the amplitude of the pulse at position mk. The reproduced signal x ^~ obtained by inputting d(n) to the synthesis filter is
Assuming that (1≦i≦N _p ; where Np indicates the order of the synthesis filter), it can be written as the following equation.

x ^~ (n)=dn+ _NP 〓 ⁱ⁼¹ aix ^~ ((n)−i) −(6) Next, the weighted squared error j within one frame between the input audio signal xn and the reproduced signal x ^~ is as follows. It can be written as

J= _N 〓 ⁿ⁼¹ ((x(n)−x ^〜 〓(n))＊ω(n)) ² −(7) Here, ω(n) is the impulse response of the weighting circuit, for example, in the conventional example It is also possible to have the same characteristics as .
Further, N indicates the number of samples in one frame. Equation (7) can be further transformed as follows.

J= _P 〓 ⁱ⁼¹ ((x(n)＊ω(n)−x ^〜 (n)＊ω(n) ² −(8) Here, the term x ^〜 (n)＊ω(n) is It is transformed according to the formula.

Let x ^~ ω(n)=x ^~ (n)＊ω(n) −(9). When both sides of equation (9) are Z-transformed, it can be written as x ^~ ω(z)=x ^~ (z)・W(z) −(10). x ^~ is further multiplied as follows.

x ^~ (z) = H(z)・D(z) − (11) Here, D(z) represents the Z transformation of the sound source pulse sequence equation (5), and H(z) is the impulse response of the synthesis filter. Z
Indicates the conversion value. Substituting equation (11) into equation (10), x ^~ ω(z) = D(z)・H(z)・W(z) −(12), and Hω(z)=H(z)・Letting W(z) be the inverse Z-transform of equation (12) and the inverse Z-transformed value of Hω(z) to be hω(n), the following equation is obtained.

Xω(n)=d(n)*h(n)−(13) Here, hω(n) represents the impulse response of the cascade-connected filter of the synthesis filter and the weighting circuit.
Substitute (5) into equation (13) to obtain the following equation.

x ^~ ω(n)= _K 〓 ⁱ⁼¹ g _i hω(n−m _i ) −(14) Here, K indicates the number of pulses generated in one frame. Substituting equations (14) and (9) into equation (8), we get J= _N 〓 ⁿ⁼¹ (xω(n)− _K 〓 ⁱ⁼¹ g _i hω(n−m _i )) ² − (15 ). Therefore, equation (7) can be expressed as equation (15). Formulas for calculating the amplitude g _k and position m _k of the sound source pulse sequence that minimize equation (15) are derived next. By partially differentiating equation (15) with respect to g _k and setting it to 0, the following equation is derived.

Here, _xh・ is the cross-correlation sequence calculated from Xω(n) and hω(n), and ψhh(・) is the impulse response
represents the correlation sequence of hωn. A covariance function sequence or an autocorrelation function sequence is known as a correlation sequence of an impulse response. The cross-correlation sequence and covariance function sequence can be expressed as follows. Furthermore, hh(・)
is often called a covariance function sequence in the field of audio signal processing.

×h(−m _k ) _N 〓 ⁿ⁼¹ x〓(n)h〓(n−m _k )= _hx (m _k ), (1≦m _k ≦N) −(17) hh(m _i,nk )= _N-(ni-nk) 〓 ⁿ⁼¹ hω(n-m _i ) hω(n-m _k ), (1≦m _i , mk≦
N) - (18) According to equation (16), the amplitude gk corresponding to the position mk can be calculated using the pulse position mk as a parameter. In other words, the pulse position mk may be selected from the position mk at which |gk| is maximum for each pulse. This can be proven by solving equation (16) for gi, but the proof will be omitted here.

This concludes the explanation regarding the derivation of this algorithm.

FIG. 4 is a block diagram showing an example of the configuration of the first and second inventions using the sound source pulse calculation algorithm according to equation (16). In the figure, the components given the same numbers as in FIG. 1 have the same functions, so their explanation will be omitted here. Further, in the figure, only the encoder side is shown. The decoder side can be realized with the same configuration as the conventional decoder side, so
The explanation will be omitted here. In FIG. 4, each component performs the following processing for each frame.
The K parameter calculation circuit 280 inputs the audio signal x _o accumulated in the buffer memory circuit 110 and calculates Np predetermined K parameters Ki (1
≦i≦Np). It is output to this K parameter encoding circuit 200. The K parameter encoding circuit 200 encodes Ki based on, for example, a predetermined number of quantization bits, and converts this code into
1Ki is output to multiplexer 260. Also K
The parameter encoding circuit 200 encodes 1KI and outputs the decoded value Ki' (1≦i≦Np) to the excitation pulse calculation circuit 230. Sound source pulse calculation circuit 2
30 inputs the input audio signal _xo accumulated in the buffer memory circuit 110 and the K parameter quantized value Ki′, and calculates the value within one frame based on the above-mentioned equations (17), (18), and (19). The amplitude gk and position mk of the sound source pulse sequence are calculated, and these are sent to the encoding circuit 250.
Output to. Next, the configuration of the sound source pulse circuit 230 will be explained. FIG. 5 is a block diagram showing one embodiment of the sound source pulse calculation circuit 230 in the second invention. In the figure, from terminal 232 to K
The parameter quantized value Ki' is inputted to the impulse response calculation circuit 210 and the weighting circuit 290. The impulse response calculation circuit 210 is
Ki′ is input, hωn (impulse response of a filter consisting of a cascade of a synthesis filter and a weighting circuit) is calculated in equation (13) for a predetermined number of samples, and the obtained hω(n)
is output to the covariance function calculation circuit 220 and the cross-correlation calculation circuit 235. The covariance function calculation circuit 220 calculates a predetermined number of samples.
Input hω(n), calculate the covariance function ψhh(mi, mk) (1≦i, k≦N) of hω(n) according to the above equation (18), and apply this to the pulse sequence calculation circuit. Output to 240. Next, the weighting circuit 290
Ki' is input from 32, and the weighting function ω(n) is calculated, for example, according to the conventional formula (3). This may be calculated using other frequency weighting methods.
Further, the weighting circuit 290 receives xn from the input terminal 231, performs convolution calculation of x(n) and ω(n), and outputs the obtained xω(n) to the cross-correlation coefficient calculation circuit 235. . Cross-correlation calculation circuit 23
5 inputs xω(n) and hω(n) and calculates the cross-correlation number ψxh (−m _k ) (1≦mk≦) according to the above equation (17).
N) and sends it to the pulse sequence calculation circuit 240.
Output to. Next, the pulse sequence calculation circuit 240
is ψxh (−m _k ) from 235, ψhh from 220
(m _i , m _k ) (1≦m _i , m _k ≦N), and using the above-mentioned sound source pulse calculation formula (16),
Calculate the amplitude of the pulse g _k . For example, in equation (16), the first pulse has an amplitude g ₁ when k=1
Find it as a function of the position m ₁ . Next, select m ₁ such that |g ₁ | is maximum, and then set m ₁ and g ₁ to 1
Let be the position and amplitude of the second pulse. Next, the second pulse is found by setting k=2 in equation (16). According to equation (16), it means that the second pulse is found by subtracting the influence of the first pulse. The third and subsequent pulses can be calculated in the same way, and either the predetermined number of pulses is reached or the number of determined pulses is reached.
Pulse calculations are continued until the error value obtained by substituting g _k and m _k into equation (16) falls below a predetermined threshold. g _k and m _k representing the amplitude and position of the pulse sequence are calculated by the pulse sequence calculation circuit 24.
0 through the output terminal 233. The above describes the sound source pulse calculation circuit 23 in the second invention.
Finish explaining 0. Note that the sound source pulse calculation circuit 230 in the first invention is the same as that shown in FIG. 5 without the weighting circuit 290. In this case, the weighting function ωn in equation (7) above is 1, but
When ω(n)=1, equation (7) becomes as shown below.

J= _N 〓 ⁿ⁼¹ (x(n)−x ^~ (n)) ^2In other words, in this case, the squared Euclidean distance is used as the evaluation formula for the error between x(n) and x ^~ (n). This is a more general evaluation formula. On the other hand, in equation (7) using ω(n), since ω(n) operates as a weighting function, x
The error between (n) and x ^~ (n) will be evaluated.

Returning to FIG. 4, the encoding circuit 250 inputs the amplitude g _k and position m _k of the pulse sequence from the output terminal 233 of the sound source pulse calculation circuit 230, and encodes these using normalization coefficients to be described later. , g _k , m _k and the codes representing the normalization coefficients are sent to the multiplexer 26
Output to 0. Here, although various encoding methods can be considered, a conventionally well-known method can be used to encode the amplitude gk. For example, a method can be considered in which the amplitude probability distribution is assumed to be a normal type and an optimal quantizer for the normal type is used. Regarding this, J. Maxkus (J.
IRE TR NSACTIONS ON by Mr. MAX)
INFORMATION THEORY) March 1960 issue,
Calculation “Quantizing for
Minimum Distortion” (“
QUANTIZING FOR MINIMUM
This is explained in detail in the paper entitled "DISTORTION" (Reference 2), so the explanation will be omitted here.
Another possible method is to use the maximum value of the amplitude of a pulse sequence within one frame as a normalization coefficient, normalize each pulse amplitude with this value, and then quantize and encode it. In the case of the former method,
rms within one frame (ROOT MEAN
SQUARE) value may be used as the normalization coefficient. Next, various methods can be considered for encoding the pulse position. For example, a run-length code well known in the field of facsimile signal encoding may be used. This represents the length of the code "0" using a predetermined code sequence. Further, for encoding the normalization coefficients, conventionally well-known logarithmic compression encoding or the like can be used.

It should be noted that the coding of the pulse sequence is not limited to the coding method described here, and it goes without saying that the best known method can be used.

Returning to FIG. 4 again, multiplexer 260
inputs the output code of the K-parameter encoding circuit 200 and the output code of the encoding circuit 250, combines them, and outputs the combination from the transmission side output terminal 270 to the communication path. This concludes the description of the speech encoding device according to the present invention.

According to the configuration of the present invention, since the sound source pulse sequence is calculated according to equation (16), the synthesis filter is driven by the pulses found in the conventional method of Document 1,
There is no path to find the reproduced signal, calculate the error and squared error from the original signal, and feed back these to adjust the pulse, and there is no need to repeat this series of processing, so the amount of calculations is significantly reduced. This has the great effect of making it possible to reproduce sound with good playback quality. Furthermore, in the calculation of equation (16), ψxh (−mk) and ψhh (mi, mk) (1
By calculating the values of ≦mi, mk≦N) in advance for each frame, the calculation of equation (16) becomes a very simplified operation of multiplication and subtraction, and further calculations are performed. It has the effect of reducing the amount. Furthermore, compared to other conventional methods for searching for sound source pulse sequences, the method according to the present invention has the advantage that better quality can be obtained for the same amount of transmitted information.

In the above-described embodiment of the present invention, the sound source pulse sequence within one frame is encoded by the component 250 in FIG. 4 after all pulse sequences have been determined. The encoding may include the calculation of the pulse sequence, and each time one pulse is calculated, the encoding may be performed, and the next pulse may be calculated. By adopting such a configuration,
Since a pulse sequence that minimizes errors including encoding distortion can be found, the quality can be further improved.

Furthermore, in the above embodiment, the pulse sequence was calculated on a frame-by-frame basis, but the frame may be divided into several subframes, and the pulse sequence may be calculated for each subframe. . According to this configuration, if the frame length is N, the amount of calculation can be approximately 1/d times as large as that of the configuration shown in FIG. Here, d indicates the number of frame divisions. For example, if d=2, the amount of calculation can be reduced to about 1/2. Of course, equivalent characteristics can be obtained.

Furthermore, in the configuration example described above, the frame length is constant, but it may be variable.
The characteristics will improve if it is made variable. Further, although the K parameter is used as the parameter representing the spectral envelope of the short-time audio signal sequence, other well-known parameters (for example, LSP parameters, etc.) may also be used. Note that, as is well known in the field of digital signal processing, the autocorrelation function may be calculated from the power spectrum. The cross-correlation function may also be calculated from the cross-power spectrum.

In addition, in the sound source pulse calculation formula (16) according to the present invention, the covariance function was calculated according to formula (18) as ψhh. It may be configured.

Ψhh(｜m _i −m _k ｜)= _N- | _ni-nk | 〓 〓 ⁿ⁼¹ h _w (n)h _w (n-｜m _i −m _k ｜), (1≦｜m _i −m _k ｜≦N
)−(19) By adopting this configuration, ψhh(・)
It is possible to significantly reduce the amount of calculation required for the calculation of , and there is an effect that the amount of calculation as a whole can also be reduced.

[Brief explanation of drawings]

Fig. 1 is a block diagram showing the configuration of the conventional system, Fig. 2 shows an example of a sound source pulse sequence, and Fig. 3 shows the frequency characteristics of the input audio signal sequence and the frequency characteristics of the weighting circuit shown in Fig. 1. Diagram showing an example, No. 4
The figure shows a block diagram showing an embodiment of the speech encoding device according to the present invention, and FIG. 5 shows a block diagram showing an example of the structure of the sound source pulse calculation circuit 230 shown in FIG. 4. In the figure, 110...buffer memory circuit,
120...subtraction circuit, 130...composition filter,
140... Sound source pulse generation circuit, 150... Error minimization circuit, 180, 280... K parameter calculation circuit, 190, 290... Weighting circuit, 20
0... K parameter encoding circuit, 230... Sound source pulse calculation circuit, 210... Impulse response calculation circuit, 220... Covariance function calculation circuit, 235...
. . . cross-correlation calculation circuit, 240 . . . pulse calculation circuit, 250 . . . encoding circuit, 260 . . . multiplexer, respectively.

Claims (1)

[Scope of Claims] 1. Means for inputting a discrete audio signal sequence, dividing the audio signal sequence into short-time intervals to obtain a short-time audio signal sequence, and determining a parameter representing a spectral envelope from the short-term audio signal sequence. means for extracting and encoding; means for calculating an impulse response sequence based on parameters representing the spectral envelope; means for calculating a correlation sequence using the impulse response sequence; means for calculating a cross-correlation sequence using a short-time audio signal sequence; and calculating and encoding a drive excitation signal sequence of the short-time audio signal sequence using the correlation sequence and the cross-correlation sequence. and means for outputting a combination of the code of the parameter representing the spectral envelope and the code representing the drive excitation signal sequence. 2. means for inputting a discrete audio signal sequence and dividing the audio signal sequence into short-term audio signal sequences to obtain a short-term audio signal sequence; and means for calculating parameters representing a spectral envelope from the short-term audio signal sequence;
means for extracting and encoding the impulse response sequence and the target signal sequence; means for calculating an impulse response sequence based on a parameter representing the spectral envelope; and calculating a correlation sequence using the impulse response sequence. means for calculating a target signal sequence with a predetermined correction based on the short-time audio signal sequence; and calculating a cross-correlation sequence using the impulse response sequence and the target signal sequence. means for calculating and encoding a driving excitation signal sequence of the short-time audio signal sequence using the correlation sequence and the cross-correlation sequence; and a code of a parameter representing the spectral envelope and the driving excitation source. 1. A speech encoding device comprising means for outputting a combination of a code representing a signal sequence and a code representing a signal sequence.