JP3095133B2 - Acoustic signal coding method - Google Patents

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention reduces a signal sequence of an audio signal by performing a predictive coding for synthesizing the audio signal by driving a filter representing a spectral envelope characteristic of an audio signal such as voice or music by a sound source vector. The present invention relates to a high-efficiency speech encoding method for digitally encoding information.

[0002]

2. Description of the Related Art In digital mobile communication, a high-efficiency voice encoding method is used in order to efficiently use radio waves or to efficiently use a communication line or a storage medium for a voice or music storage service. . At present, as a method for encoding speech efficiently, an original speech is divided into sections called frames (or subframes) at a fixed interval of about 5 to 50 ms, and the speech of one frame represents an envelope characteristic of a frequency spectrum. Separation into two information, the characteristics of the linear filter and the driving sound source signal for driving the filter,
Techniques for encoding each have been proposed. In this method, as a method of encoding a drive excitation signal, a method of separating and encoding a periodic component considered to correspond to a pitch period (fundamental frequency) of a voice and other components is known. As an example of the encoding method of the drive excitation information, Code-Excited Linear Pr
ediction: CELP). For details of the above technology, refer to the document MR Schroeder and BS Atal, “Code-Excit
ed Linear Prediction (CELP): High Quality Spe
ech at Very Low Bit Rates ”, IEEE Proc. ICA
SSP-85, pp. 937-940, 1985.

FIG. 8 shows a configuration example of the above-mentioned encoding method. For the speech x input to the input terminal 1-0, the linear prediction analysis unit 1-1 calculates a linear prediction parameter a representing the frequency spectrum envelope characteristic of the input speech. The obtained linear prediction parameter a is encoded by the linear prediction parameter encoding unit 1-2 and sent to the linear prediction parameter decoding unit 1-3. In addition, when distortion calculation is performed using spectral information of an input voice, for example, in consideration of auditory characteristics in distortion calculation, the linear prediction parameter a is also sent to the distortion calculation unit 1-6. Linear prediction parameter decoding unit 1-
In the step 3, the synthesis filter coefficient a か ら is reproduced from the received code and sent to the synthesis filter 1-5. When the auditory characteristics are considered in the distortion calculation, the decoded linear prediction parameter a パ ラ メ ー タ may be used in the distortion calculation instead of using the linear prediction parameter a before quantization in the distortion calculation unit 1-6. The details of the linear prediction analysis and examples of encoding the linear prediction parameters are described in, for example, “Digital Speech Processing” by Sadahiro Furui (Tokai University Press). Here, the linear prediction analysis unit 1-1, the linear prediction parameter encoding unit 1-2, and the linear prediction parameter decoding unit 1-
3 and the synthesis filter 1-5 may be replaced with a non-linear filter.

[0004] A drive excitation vector generation section 1-4 generates a drive excitation vector candidate c having a length of one frame.
Send to synthesis filter 1-5. FIG. 9 shows a configuration example of the driving sound source vector generation unit 1-4. From adaptive codebook 2-1:
The immediately preceding past drive excitation vector stored in the buffer (the drive excitation vector for one to several frames just before quantization) c (t-1) is cut out at a length corresponding to a certain period, and the cut out is performed. was by repeated until the length of the frame vector, candidate v _a time series vector corresponding to the period component of the sound is output. As the “certain period”, a period that reduces the distortion d in the distortion calculator 1-6 is selected. In general, the selected period generally corresponds to the pitch period of voice.
From the fixed codebook 2-2, a time-series code vector candidate v _{r of} one frame length corresponding to the aperiodic component of speech
Is output. The fixed codebook 2-2 stores a predetermined number of candidate vectors according to the number of bits for encoding independently of the input speech. Adaptive codebook 2-
Candidate time-series vector outputted from the first and fixed codebook 2-2, in the multiplication unit 2-4 and 2-5, the weights g _a created in a weight codebook 2-3 respectively, g _r is multiplied These multiplication results are added in an adder 2-6 to become a drive excitation vector candidate c. In the configuration example of FIG. 9, the configuration may be such that only the fixed codebook 2-2 is used without using the adaptive codebook 2-1. When encoding a signal with a small pitch periodicity such as a consonant part or background noise,
In order to save bits, a configuration not using the adaptive codebook 2-1 is often used.

Returning to the description of FIG. 8, the synthesis filter 1-5
Is a linear filter that uses the output of the linear prediction parameter decoding unit 1-3 as a filter coefficient, and outputs a reproduced sound candidate y using the driving excitation vector candidate c as an input. The order of the synthesis filter 1-5, ie, the order of the linear prediction analysis, is
Generally, about 10 to 16 orders are often used. As described above, the synthesis filter 1-5 may be a non-linear filter.

[0006] In the distortion calculation unit 1-6, the synthesis filter 1-
Then, a distortion d between the reproduced voice candidate y, which is the output of No. 5, and the input voice x is calculated. The calculation of the distortion is often performed in consideration of the coefficient a ＾ of the synthesis filter or the non-quantized linear prediction coefficient a, for example, with auditory weighting. FIG. 11 shows a configuration example in which distortion is calculated in consideration of hearing weighting. Perceptual weighting is configured in the form of a perceptual weight filter using unquantized linear prediction parameters a or quantized synthetic filter coefficients a ＾. The reproduced voice candidate y output from the synthesis filter 4-1 is passed through an auditory weight filter 4-2,
Similarly, a distortion d is calculated between the input voice and the input voice that has passed through the auditory weight filter 4-3. Here, since the hearing weight filters 4-2 and 4-3 usually use the same filter coefficient, the hearing weight filters 4-2 and 4-3 use the distance calculation section 4-4.
Even if one filter is inserted after -4, it is equivalent. However, in terms of the amount of processing, as shown in FIG. 11, it is often divided into two places before the distance calculation unit 4-4.

Furthermore the input time series speech vector x and forth passes through the perceptually weighted filter 4-3 This combining weight calculation unit 1-7, and the targeted voice x _w, is sent to a distance calculation unit 4-4. On the other hand, the driving sound source vector candidate c includes a synthesis filter 4-1 and an auditory weight filter 4-2.
, And becomes a perceptually weighted reproduced voice candidate vector y _w , which is sent to the distance calculator 4-4. Distance calculator 4-4
Then, the distance between the target audio vector x _w and the reproduced audio candidate vector y _w is measured. The distance measure this time for example, d = ‖x _w -y _w ‖ ² (1) such as may be used distance measure. A driving sound source vector that minimizes the distortion measure is selected. In the case of using the configuration of driving excitation vector generation as shown in FIG. 9, a periodic code, a fixed code, and a weight code are determined. Note that the auditory weight filters 4-2 and 4-3 are filters for performing distortion calculation to reduce noise in the reproduced voice using human auditory characteristics, and need not always be used.

At this time, the input time-series speech vector x
May be the input audio signal as it is, but in general,
It is often a time-series signal from which the influence from the previous subframe has been subtracted. In addition, when using the configuration of driving excitation vector generation as shown in FIG. 9, selecting one optimal combination from all possible combinations of the periodic code, the fixed code, and the weight code is an operation process. It is difficult in terms of quantity. For example, it is often determined in the order of, for example, a periodic code, a fixed code, and a weight code, or sequentially searched while appropriately narrowing down candidates, and finally determining a sub-optimal combination.
In the case where the search is performed while sequentially determining or leaving the candidates in this manner, a synthesized component caused by the previously selected code vector (for example, the adaptive code vector) is subtracted from the input speech, and the driving excitation vector candidate c is Only the vector component to be determined (for example, only the fixed code vector)
Is often input to calculate distortion.

In FIG. 8, a codebook search control section 1-8 selects a driving excitation code that minimizes the distortion d between each reproduced speech candidate y and input speech x, and determines a driving excitation vector in the frame. . The adaptive codebook 2-1, the fixed codebook 2-2, and the weighted codebook 2- shown in FIG.
In the case of a configuration composed of three, a periodic code, a fixed code, and a weight code are selected, and these are used as the drive excitation code.

[0010] The excitation code (periodic code, noise code, weight code) determined by codebook search control section 1-8,
The linear prediction parameter code output from the linear prediction parameter coding unit 1-2 is sent to the code transmission unit 1-9 and stored in a storage device or received via a communication channel depending on the form of use. Sent to the side. FIG. 10 shows a configuration example of a decoding method corresponding to the above-described encoding method. Among the codes received at the input terminal 3-0 from the transmission path or the storage medium, the linear prediction parameter code is a linear prediction parameter decoding unit 3-
In step 2, the signal is decoded into a synthesis filter coefficient and sent to the synthesis filter 3-4 and, if necessary, the post-processing unit 3-5. The driving excitation code is sent to driving excitation vector generation section 3-3, and an excitation vector corresponding to the code is generated. The configuration of the driving excitation vector generation unit 3-3 corresponds to the configuration of the driving excitation vector generation unit 1-4 of the encoding method shown in FIG. The synthesis filter 3-4 reproduces a sound by using the driving sound source vector as an input. Post-processing unit 3-5
Performs processing (also referred to as post-filtering) to aurally reduce the sense of noise in the reproduced sound, but the post-processing unit 3-5 is often not used due to a reduction in processing amount or the like.

[0011]

A problem in the CELP system is that a great deal of arithmetic processing is required for calculating a distortion for selecting a driving excitation vector candidate. Algebraic Code-Excited
A method called Linear Prediction (ACELP) has been proposed. This method does not store a fixed codebook as a vector pattern of a frame length, but rather stores several pulses of height 1 in a frame, for example, four pulses in a frame or subframe of 40 samples. By adopting this driving excitation method in a method of setting a fixed code vector by setting it at an appropriate position, and devising the calculation order in distortion calculation, the calculation processing can be significantly reduced as compared with the conventional method. . The details of the ACELP method are described in, for example, Literature, R. Salami, C. Laflamme,
and JP. Adoul, “8 kbit / s ACELP Coding of
Speech with 10 ms Speech-Frame: a Candidate for C
CITT Standardization ”, IEEE Proc. ICASSP-
94, pp. II-97. Also, based on the same processing concept, as a method of higher quality and lower operation amount, “Acoustic signal encoding method and acoustic signal decoding method” already filed by the present inventors (Japanese Patent Application No. 7-150550).
There is. In this method, as a fixed code vector, instead of a pulse having a height of 1, adjacent two to several samples are used as a unit, and a method of arranging a pulse pattern having height information in a frame is used, thereby achieving a lower calculation. It balances quantity and high quality.

However, in these systems,
Synthesis filter or auditory weighting filter for distortion calculation,
Or, a filter combining them is often expressed by an impulse response or FIR type filter. However, when the frame or subframe becomes longer, the number of taps of the FIR filter to obtain a result equivalent to the case of using an IIR type filter Not only increases the amount of computation in comparison to the conventional method, but also has a problem that an extremely large amount of memory is required in order to store an intermediate result of the calculation in the distortion calculation. Therefore, it is difficult to use the above method as it is for low bit rate audio coding that generally lengthens a subframe.

On the other hand, in the configuration of FIG. 11, in order to execute the operation of passing the driving sound source vector candidate c through the synthesis filter 4-1 and the auditory weighting filter 4-2 at high speed, these two filters must be combined. , A single auditory weighted synthesis filter having equivalent filter characteristics.
In order to make an equivalent one filter, for example, it is possible to represent an impulse response from the input of the synthesis filter 4-1 to the output of the auditory weight filter 4-2 as a filter coefficient by using an FIR filter.

FIG. 12 shows a configuration for realizing higher-speed distortion calculation in the configuration expressed by the one equivalent filter. For example, an auditory weighted synthesis filter expressed by an FIR filter is truncated with a finite tap, approximated by an AR filter with a short tap number, or
The number of taps of the IR filter is reduced by a method such as reducing the number of taps required to obtain a result equivalent to that of the IIR filter by a perceptually weighted synthetic approximation filter 5-2 whose filter characteristics do not exactly match each other. . As a result, the amount of arithmetic processing and the amount of memory in the composite distortion calculation can be reduced. However, when the configuration of FIG. 12 is used, if the difference between the filter characteristics of the approximation filter 5-2 and the characteristics of the original synthesis filter 4-1 and the perceptual weighting filter 4-2 increases, an appropriate Since the driving excitation code is no longer selected, which leads to a remarkable deterioration in the quality of the reproduced voice, it was practically impossible to take a longer subframe, that is, lower the bit rate.

SUMMARY OF THE INVENTION It is an object of the present invention to provide a sound such as a sound or a music which can obtain a high quality reproduced sound with a small bit rate, a small memory amount within a range allowed by an inexpensive processor, and a small calculation amount. It is to provide a method for digitally encoding a signal.

[0016]

According to the present invention, an approximate filter which is simplified so that distortion can be calculated at a high speed such as cutting off the tap of the FIR type synthesis filter in the middle is used for the synthesis distortion calculation, and the approximation filter is expressed by this approximation filter. The approximation error generated based on the above is added to the input speech, and this is used as a target vector at the time of codebook search.

With this configuration, the influence of the approximation is canceled in the distortion calculation, and a high quality low bit rate encoding method can be realized with a very small amount of memory and processing amount even when the subframe is long.

[0018]

DETAILED DESCRIPTION OF THE INVENTION structure underlying the embodiment of the present invention
The result is shown in FIG. Input audio x from input terminal 6-0
Is the quantized (decoded) synthesis filter coefficient a
Inverse filter of synthesis filter by （(synthesis inverse filter)
The signal passes through 6-3 and is converted into an ideal (non-quantized) drive excitation vector r. r is an ideal driving sound source vector in which the distortion from the input sound x becomes zero when the signal passes through the synthesis filter 4-1 having the driving sound vector candidate c as an input in FIG. Ideal driving sound source vector r
Is a modified target speech vector x _'w through auditory weighting synthesis approximation filter 6-4 having the same characteristics as the auditory weighting synthesis approximation filter 5-2. In this case, it is assumed that the approximation error similar to the approximation error caused by the perceptually weighted synthesis filter 5-2 is added to the modified target speech vector x _'w. In the distance calculation unit 4-4, the distance between the perceptually weighted reproduced speech candidate y ′ _w including the approximation error, which is the output of the perceptually weighted synthetic approximation filter 5-2, and the deformed target speech vector x ′ _w Is calculated. Therefore, in this distance calculation, the approximation error generated in the perceptually weighted synthetic approximation filter 5-2 is canceled out in the distance calculation with the approximation error added by the perceptually weighted synthetic approximation filter 6-4, and the distortion d (distance ) Can be calculated with high accuracy.

FIG. 2 specifically shows the synthesis approximation filters 5-2 and 6-4 in the form of the finite tap length FIR filters 7-2 and 7-4 in the method according to the present invention in FIG. . If the number of taps at this time is the same as the number of taps of the subframe length, the encoding result matches that of the conventional method that does not use approximation calculation, but the amount of computation increases. On the other hand, if the number of taps is set to 1 tap that does not use the past sample value (this may be referred to as 0 tap), the distortion between the driving excitation vector candidate c and the ideal driving excitation vector r is reduced to the driving excitation level. , And the amount of calculation processing is extremely small, but sufficient coding quality cannot be obtained. The number of taps is determined in the range of 1 to the subframe length (the number of samples of the subframe) in consideration of the balance between the coding quality and the amount of arithmetic processing. In the method according to the present invention, the number of subframes is, for example, 80. Even when the number of taps is reduced to about 2 to 6 taps at the time of a sample, the approximation error generated in the finite tap length FIR auditory weighted synthesis filter 7-2 causes the finite tap length FIR auditory Since it is added to the weighted synthesis filter 7-4, it has been confirmed that the signal-to-noise ratio (SNR) and the perceptual quality when actual speech is encoded hardly deteriorate.

[0020] Figure 3, in the configuration example of the driving excitation vector generation unit 1-4 is a configuration example using a fixed code vector candidates v _r and a pitch periodic. The ACELP method and “Acoustic signal encoding method and acoustic signal decoding method”
The configuration shown in FIG. 3 is also used in Japanese Patent Application No. 7-150550. The pitch period section 8-7, cycle codes same period code and inputted to the adaptive codebook is input to the period of the output v _r of the fixed codebook 2-2 at a period corresponding to the period code. As a specific periodic operation, a comb filter (comb filter) at a tap position corresponding to the periodic code is often applied to the fixed code vector v _r . The tap position may be an integer sample position, or a comb filter at a non-integer sample position may be realized using an upsampling technique.

In the configuration shown in FIG.
When searching for a fixed codebook 2-2, the optimum code (or a plurality of periodic code candidates with a small distortion) is searched for assuming that there is no fixed codebook 2-2. an adaptive codebook component y _a obtained by combining the adaptive code vector, as inputs x _r a minus advance from the input speech x, fixed code vectors v _r
Is used to search for a fixed code that minimizes the distortion between the components y _rp and x _r obtained by combining FIG. 4 shows a configuration example of a fixed code vector combined distortion calculation method using this method. Since the pitch periodization unit 8-7 in FIG. 3 can exchange the order with the multiplication unit 2-5, as shown in FIG.
And a pitch filter 8-4 between the synthesizing filter 4-1. The fixed code vector v _r is multiplied by 2
-5. The multiplier unit 2-5 v _r over weight g _r generates a excitation vector candidates c _r, and sends it to the pitch period section 8-4. _cr is pitch-periodicized,
The reproduced speech candidate y _rp passes through the synthesis filter 4-1 and passes through the auditory weight filter 4-2 to pass through the distance calculation unit 4-
4 At this time, if the pitch periodization unit 8-4, the synthesis filter 4-1, and the auditory weight filter 4-2 are represented by one filter having a characteristic obtained by combining three filter characteristics, the amount of arithmetic processing required for the search is reduced. I can do it.
However, when a filter having a synthesis characteristic of the three filters 8-4, 4-1 and 4-2 is expressed by an FIR filter, the synthesis filter 4-1 and the auditory weight filter 4
Unlike the FIR filter having the characteristic of −2 , the coefficient has a large value near the tap position of the cycle considered to correspond to the pitch cycle. Therefore, as shown in the configuration example shown in FIG. I can't stop and search faster.

FIG. 5 shows an embodiment of the present invention which solves this problem and calculates distortion at a high speed even in the case of pitch period. In the configuration example of FIG. 5, similarly to the configuration example shown in FIG.
Synthesis filter 4-1 and auditory weight filter 4 in FIG.
The filter having the characteristic of -2 is replaced with the perceptually weighted synthetic approximation filter 5-2. As in the configuration example of FIG. 1, the input _xr is passed through a synthetic inverse filter 6-3 so that distortion caused by approximation can be canceled between the input side and the input side.
The synthetic approximation filter 6-4 having the same characteristics as the filter 5-2 is passed through the auditory weight, but in this configuration example, an inverse filter (filter for removing the periodicity of the pitch) 10- of the pitch periodicization filter 8-4 in FIG. 4 is input to the input side of the voice x. In this configuration, the synthesis approximation filters 5-2 and 6-4 with auditory weights are replaced with the configuration example shown in FIG.
If a finite tap length FIR-type auditory weighted synthesis filter is used, the codebook can be searched very quickly. The tap length of the FIR filter at this time ranges from one tap (sometimes referred to as zero tap) that does not use a past sample value to the subframe length, as in the configuration example of FIG. In the method according to the present invention, even when the number of taps is reduced to about 2 to 6 taps, the actual audio is encoded in the method according to the present invention. At that time, it was confirmed that the signal-to-noise ratio (SNR) and the auditory quality hardly deteriorated. In the configuration example of FIG. 5,
Synthetic inverse filter 6-3, pitch periodic inverse filter 10-
4. When all of the hearing weighted synthesis approximation filters 6-4 are linear filters, their order may be exchanged.

FIG. 6 shows that in the method according to the invention, F
This shows an example of a configuration in which the distortion calculation is performed efficiently and an extremely high-speed speech encoding is realized, with the advantage that the quality degradation of the encoded sound is very small even if the IR filter is truncated to a finite length. is there. The finite tap length FIR type hearing weighted synthetic filter coefficient calculating unit 11-1 has a perceptual weight having a combined characteristic of a synthetic filter and a perceptual weighted filter from the pertinent synthetic filter coefficient a ＾ and the unquantized linear prediction parameter a. A filter coefficient when the attached synthesis filter is realized by the FIR type is calculated, and a coefficient β obtained by truncating the filter coefficient by a finite tap length is output. The impulse response matrix generation section 11-2 generates a triangular matrix having FIR filter coefficients as elements as shown in the following equation (2). Here, N represents the number of samples of the subframe. In equation (2), since the coefficient β is truncated to a finite length, for example, if the truncation order is k, β _k to β _N−1
Up to 0, which is a matrix as shown in equation (3).

[0024]At this time, the part where the element of the matrix is 0 is recorded in the memory etc.
No need to remember. In the correlation matrix generation unit 11-3,
From the impulse response matrix H, H^tCalculate H
You. At this time, the coefficient β_kFrom β_N-1Up to 0
It is not necessary to perform an N × N matrix calculation, and a k × k matrix
H by calculation^tH can be determined. For example, k
Means that the quality of the coded sound is almost
80 × 80 rows when N = 80
Compared to column calculation, for example, 5 × 5 matrix calculation is a remarkable operation
The processing amount is reduced. Input speech without adaptive codebook components
x _rPasses through the synthesis inverse filter 6-3, and is pitch-cycled
It passes through the inverse filter 10-4 and enters the convolution unit 11-6.
Is forced. In the convolution unit 11-6, the pitch period reverse
Output r of filter 10-4_pIs the FIR filter of coefficient β
Target sound including tap truncation distortion
Voice x '_rpX '_rpAnd H matrix, time axis inversion
X ′ by convolution or matrix operation_rp ^t
H (A^tIndicates the transpose of matrix A). This and
Very high speed if the cutoff order k is small
Can be calculated. Folding section 11-6 is separate
Can be used, and the correlation matrix calculation unit 11-3
Output H^tH and the pitch inverse periodic filter 11-5
Output r_pFrom the matrix operation, r_p ^t(H^t
H) can also be calculated. At this time, the above x '
_rp ^tH and r_p ^t(H^tH) has the same value. Most
In the final distance scale calculation unit 11-7, the driving sound source vector candidate
Fixed codebook component c_rAnd H^tH, x '_rp ^tH
(Or r_p ^tH ^tH), the distance scale d '= (x'_rp ^tHc_r)^Two/ (C_r ^tH^tHc_r) (4) is calculated. d 'is sent to the codebook search control unit,
The degree d 'is maximized (e.g., the distortion measure d is minimized)
Value) code is selected.

In the above description, the synthetic approximation filter does not necessarily need to be one having the auditory weighting characteristic. In the claims, the “frame” may be either a frame or a sub-frame obtained by dividing the frame.

[0026]

The following experiments were conducted to confirm the effects of the present invention. 4.6kbit / s Dual-Pulse
CS-CELP was configured. The frame length is 20 ms, the sub-frame length is 10 ms (80 points), and LPC quantization is performed for each frame, and the rest is performed for each sub-frame. Bit allocation is LSP 22 bits per frame, adaptive code 8
× 2 bits, Dual-Pulse code 20 × 2, gain code 7 × 2 (92 (4.6 kbit / s) in total)
Pulses are arranged in three sets per subframe, and 11 bits for the position, 6 bits for the pattern, and 3 bits for the sign are assigned.

Inputting the actual audio data to the encoder,
The performance of the inventive method was investigated. Audio data is 8kHz
In sampling, ITU-TG. 712 band filters were used. FIG. 7 shows the cutoff order and W when the tap of the FIR filter is cut off at a finite length.
The relationship of SNR was shown. Since the WSNR is measured between the final synthesized speech and the input speech, the WSNR has the same scale irrespective of the number of taps for censoring. Method (1) in the figure
Is a case where a target speech for minimizing distortion is obtained by a conventional method, and only taps of a filter for searching a codebook are cut off. In this case, the quality rapidly deteriorates when the number of taps becomes equal to or less than 20 taps. The method (2) is a case where the method of the present invention shown in FIG. 2 without using the pitch periodic inverse filter is applied. With this method, the WSNR hardly changes until the number of taps is about two. Method (3) is a case in which the method of the present invention shown in FIG. 6 using a pitch period inverse filter is applied. 4.6 kbit / s Dual-
Pulse CS-CELP is Dual Pulse
Method (3) in which the pitch period is used for the driving sound source
, Very high-speed encoding can be realized. Comparing the quality in this case with that of method (2),
Although it is reduced by about 0.3 dB as a whole, as in the case of the method (2), even if the number of taps is reduced, the WSNR is reduced.
Did not drop much.

In terms of audibility, if about 6 taps are used, deterioration is hardly felt as compared with the case where all taps are used. Further, the method (3) is slightly deteriorated as compared with the method (2). As described above, according to the present invention, it has been confirmed that quality degradation is extremely small even when the cutoff is performed with a very small number of taps and a high-speed codebook search, that is, a high-speed voice coding is realized.

[Brief description of the drawings]

FIG. 1 is a diagram showing a functional configuration of a method for calculating a distance between a perceptually weighted reproduced speech candidate including an approximation error and a deformed target speech also including an approximation error, which is a premise of the present invention.

FIG. 2 is a functional configuration diagram showing an example of expressing a synthetic approximation filter with auditory weights in the form of a finite tap length FIR filter in the method shown in FIG. 1;

FIG. 3 is a diagram illustrating an example of a functional configuration in which a fixed code vector candidate is pitch-periodically used in a configuration of a driving excitation vector generation unit.

FIG. 4 is a diagram showing an example of a functional configuration of a method for calculating a fixed code vector combined distortion when the configuration shown in FIG. 3 is used.

FIG. 5 is a diagram showing a functional configuration of a distortion calculation method to which the present invention is applied and a pitch period inverse filter is provided on the input side in a case where there is pitch period shown in FIG. 3;

FIG. 6 is a diagram showing an example of a functional configuration of a method for realizing a very high-speed speech encoding by truncating an FIR filter to a finite length and efficiently performing distortion calculation according to the method of the present invention;

FIG. 7 is a graph showing the relationship between the cutoff order of FIR filter taps and WNSR when the present invention is applied to actual speech coding.

FIG. 8: Code-driven linear prediction coding (Code-Excited)
FIG. 3 is a diagram illustrating a functional configuration example of Linear Prediction (CELP).

FIG. 9 is a diagram showing a functional configuration example of a driving sound source vector generation unit in FIG. 8;

FIG. 10: Code-driven linear predictive coding of speech (Code-Excit)
FIG. 3 is a diagram showing an example of a functional configuration of a decoding method corresponding to ed Linear Prediction (CELP).

FIG. 11 is a diagram showing an example of a functional configuration for calculating distortion in consideration of auditory weighting.

FIG. 12 is a diagram showing an example of a conventional high-speed distortion calculation method, showing a functional configuration example in which an approximation filter of a perceptually weighted synthesis filter is used for the synthesis distortion calculation.

────────────────────────────────────────────────── ─── Continuation of the front page (56) References JP-A-8-248996 (JP, A) JP-A-7-506202 (JP, A) Miki et al. “Basic algorithm of PSI-CELP speech coding” NTT R & D Vol . 43 No. 4, pp363-372 (1994) (58) Fields investigated (Int. Cl. ⁷ , DB name) G10L 11/00-21/06 H03M 7/30 H03M 7/42 JICST file (JOIS)

Claims (5)

(57) [Claims] An adaptive codebook in which an adaptive codebook vector is recorded and a fixed codebook in which a fixed codebook vector is recorded, wherein a driving excitation vector based on a fixed codebook vector candidate extracted from the fixed codebook and an adaptive excitation vector are adapted. An audio signal encoding method for selecting the fixed codebook vector that maximizes a distance measure from an input audio signal from which a codebook component has been removed, wherein a step of calculating a linear prediction parameter from an input audio signal; Calculating the synthesized filter coefficient by quantizing the equation, a step of approximating the synthesized filter coefficient to a finite-length impulse response, and a step of generating an impulse response matrix expressed by a triangular matrix having the impulse response as an element. Calculating a correlation matrix consisting of the product of the impulse response matrix and the transpose of the impulse response matrix; Converting the input acoustic signal from which the adaptive codebook component has been removed through a synthetic inverse filter having an inverse filter characteristic of the synthetic filter coefficient to convert the input acoustic signal into an ideal drive excitation vector; And a convolution process of obtaining a target speech vector by multiplying the impulse response matrix by a convolution of the target speech vector multiplied by the impulse response matrix and a fixed codebook vector candidate. And a step of calculating a distance measure by dividing by a product of the correlation matrix and the transposed vector of the fixed codebook vector candidate. 2. The audio signal encoding method according to claim 1, wherein a tap length of the synthesis filter is set to be 2 taps or more and 6 taps or less. A step of obtaining a driving excitation vector by periodicizing a fixed codebook vector candidate extracted from the fixed codebook at a period corresponding to a periodic code input to the adaptive codebook by a periodic filter; 2. The method according to claim 1, further comprising the step of passing the input speech or the ideal drive excitation vector or the target speech vector from which the adaptive codebook component has been removed, through a periodic inverse filter having an inverse characteristic of the periodic filter. Alternatively, the audio signal encoding method according to claim 2. 4. A step of calculating an auditory weighted synthetic filter coefficient truncated at the finite length from the synthetic filter coefficient and the linear prediction parameter, wherein the auditory weighted synthetic filter coefficient is used as the synthetic filter coefficient. The acoustic signal encoding method according to any one of claims 1 to 3, wherein the method is used. 5. The method according to claim 1, wherein said correlation matrix is calculated, developed and stored in a memory, and said distance scale calculation is performed by referring to a value of the correlation matrix stored in said memory. The audio signal encoding method according to any one of the above.