JPH04301900A - Audio encoding device - Google Patents

Abstract

PURPOSE:To provide an audio encoding device which raises the resolution of an adaptive code book to improve the quality and does not increase the calculation volume neither the code volume. CONSTITUTION:The above device is provided with an adaptive code book 110 where driving signals of a synthesizing filter is stored, a minimum distortion search circuit 115 which searches the optimum driving signal in the adaptive code book 110, a synthesizing filter 112 which uses the searched optimum driving signal to synthesize audio signal, an up sampling circuit 131 which up-samples the driving signal read out from the adaptive code book 110, a delay circuit 132 which delays the up-sampled driving signal, a thinning circuit 133 which thins the delayed driving signal and stores the result in the adaptive code book 110, and a start point/end point determining circuit 105 which determines the start point and the end point of the driving signal stored in the adaptive code book 110.

Description

[Detailed description of the invention]

0001

[Industrial Application Field] The present invention relates to a speech encoding device.
In particular, the present invention relates to an audio encoding device suitable for encoding audio signals at a low bit rate of about 8 kbps or less.

0002

[Background Art] Technology to encode audio signals with high efficiency at a low bit rate is an important technology for effective use of radio waves and reduction of communication costs in mobile communications such as car phones and in-house communications. be. CELP (C
ode Excited Linear Predic
tion) method is known.

[0003] This CELP method was developed by AT&T Bell Labs' M
．． R. Schroeder and B. S. “Code-Excited Linear Pr” by Mr. Atal
edition(CELP) “High-Quali
ty Speech at Very LowBit
Rates “Proc. ICASSP; 1985, p.
p. Since it was announced in 937-939 (Reference 1), it has attracted attention as a method that can synthesize high-quality speech, and various studies have been conducted to improve the quality and reduce the amount of calculation. The CELP method is characterized by LPC (Liner Pred
(active coding: linear predictive coding) The drive signal of the synthesis filter is stored in the codebook as a drive signal vector, and the optimal drive signal vector is searched from the codebook while evaluating the error between the synthesized speech signal and the input speech signal. be.

FIG. 5 is a block diagram of a speech encoding device using the latest CELP method. In the same figure, the sampled audio signal sequence that is the input signal is input to the input terminal 60.
It is input in frame units starting from 0. A frame consists of L signal samples, and for a sampling frequency of 8 kHz, L=160 is typically used. Although not shown in FIG. 5, prior to searching for the drive signal vector, LPC analysis is performed on the input audio signal sequence of L samples, and LPC prediction parameters {αi, i=1, 2, . . .
p} is extracted. This LPC prediction parameter αi is supplied to the LPC synthesis filter 630. Note that p is the prediction order, and generally p=10 is used. LPC
The transfer function H(z) of the synthesis filter 630 is [Equation 1]
is given by

[0005]

[Equation 1] Next, a process of searching for an optimal drive signal vector while synthesizing audio signals will be described. First, input terminal 60
0, the subtracter 61
With 0, the influence of the internal state of the synthesis filter 630 in the previous frame on the current frame is subtracted. The signal sequence obtained from the subtracter 610 is divided into four subframes, and becomes a target signal vector for each subframe.

The drive signal vector that is the input signal to the LPC synthesis filter 630 is obtained by multiplying the drive signal vector selected from the adaptive codebook 640 by a predetermined gain in the multiplier 650, and the drive signal vector selected from the white noise codebook 710. The noise vector obtained by multiplying the noise vector by a predetermined gain by the multiplier 720 is added by the adder 660.

[0007] Here, the adaptive codebook 640 is
Analysis by Synth
esis), and the details can be found in W. B.
Kleijin D. J. Krasinski and
R. H. Ketchum, “Improved Sp.
eechQuality and Efficiency
Vector Quantization in C
ELP”, Proc. ICASSP, 1988, pp.
155-158 (Reference 2). According to this document 2, the drive signal of the LPC synthesis filter 630 is processed one sample at a time by the delay circuit 670 over the pitch search range a to b (a, b are the sample numbers of the drive signal, usually a=20, b=147). By delaying a
Create a drive signal vector for a pitch period of ~b samples, which is stored as a codeword in the adaptive codebook.

When searching for an optimal drive signal vector, the codewords of the drive signal vector corresponding to each pitch period are read out one by one from the adaptive codebook 640 and multiplied by a predetermined gain in a multiplier 650. . and,
A filter operation is performed by the LPC synthesis filter 630 to generate a synthesized speech signal vector. The generated synthetic speech signal vector is subtracted from the target signal vector by a subtracter 620. The output of this subtracter 620 is input to an error calculation circuit 690 via an auditory weighting filter 680.
The mean squared error is determined. The information on the mean squared error is further input to the minimum distortion search circuit 700, and its minimum value is detected.

The above process is performed for all the codewords of the drive signal vectors in the adaptive codebook 640, and the minimum distortion search circuit 700 finds the number of the codeword that gives the minimum value of the mean squared error. Further, the gain multiplied by the multiplier 650 is also determined so that the mean square error is minimized.

Next, a search for an optimal white noise vector is performed in a similar manner. That is, codewords of noise vectors are read out one by one from the white noise codebook 710, multiplied by a gain in a multiplier 720, and filtered in an LPC synthesis filter 630 to generate a synthesized speech signal vector and obtain the target. A calculation of the mean squared error with the vector is performed for all noise vectors. Then, the number and gain of the noise vector that gives the minimum value of the mean squared error are determined. Note that the perceptual weighting filter 680 is used to shape the spectrum of the error signal output from the subtracter 620 to reduce distortion perceived by humans.

[0011] In this way, the CELP method seeks the optimal drive signal vector that minimizes the error between the synthesized speech signal and the input speech signal, so it is possible to synthesize high-quality speech even at a low bit rate of about 8 kbps. can do. However, at a bit rate of 8 kbps or less, quality deterioration is perceived, and this is still insufficient.

Therefore, P. Mr. kroon and B. S.
“PITCH PREDICTOR” by Mr. Atal
S WITH TEMPORAL RESOLUTIO
N”, proc. ICASSP, pp. 661-66
4, 1990 (Reference 3) and “IMPROVED P
ITCH PREDICTION WITHFRACT
IONAL DELAYS IN SELP CODI
NG”, proc. ICASSP, pp.665-6
68, 1990 (Reference 4). The methods described in these documents 3 and 4 are based on the LP in FIG.
The drive signal of the C synthesis filter 630 is set in the pitch search range a~
b, the sampling frequency of the drive signal is made higher than that of the input audio signal (up-sampling), and after delaying this up-sampled drive signal in units of one sample or less, the adaptive codebook 64
The feature is that it is stored as 0.

Specifically, upsampling can be performed by inserting a predetermined number of 0s between samples and interpolating between sample values using an interpolation filter. The upsampled drive signal is delayed by a delay circuit in units of one sample, and then the samples are thinned out in units of a predetermined number of samples to return to the original sampling frequency and stored in the adaptive codebook. In this case, for example, if the upsampling magnification is doubled, the adaptive codebook will store a drive signal vector with the same sample period as the input audio signal and a drive signal vector with a 1/2 sample period shift between these two. become.

[0014] In this way, by increasing the sampling frequency of the adaptive codebook, pitch search can be performed in units of one sample or less of the input speech signal, which improves pitch prediction accuracy and improves the quality of encoded speech. Improved. According to Document 3, doubling the sampling frequency improves the segmental SNR by 0.9 dB for female voices and by 0.6 dB for male voices.

However, in this method, due to upsampling, the size of the adaptive codebook increases in proportion to the upsampling magnification, and accordingly, the amount of calculation required to search for the drive signal vector from the adaptive codebook increases with the upsampling magnification. As the number increases proportionally, the amount of code required to transmit the code number (index) of the drive signal vector to the receiving side also increases. For example, if the upsampling magnification is doubled,
The amount of calculation required to search for the drive signal vector doubles, and the amount of 4-bit code per frame increases.

[0016]

[Problems to be Solved by the Invention] As mentioned above, in the conventional CELP method having an adaptive codebook, when the resolution of the adaptive codebook is increased by upsampling, although the quality improves, most of the encoding process is The amount of calculation required to search for the drive signal vector from the adaptive codebook increases.

[0017] Therefore, if audio encoding is to be realized through real-time processing, a plurality of high-speed DSPs (digital signal processing circuits) will be required, which poses the problem of increasing the circuit scale, increasing prices, and increasing power consumption. was there.

Furthermore, as the resolution of the adaptive codebook improves, the amount of code necessary to transmit the code number of the drive signal vector increases, resulting in a higher transmission bit rate.

The present invention has been made in view of the above problems, and provides a speech encoding device that can improve the quality by increasing the resolution of an adaptive codebook without increasing the amount of calculation and code. The purpose is to provide.

[0020]

[Means for Solving the Problems] In order to solve the above problems, the present invention provides a codebook (adaptive codebook) in which drive signals are stored as codewords, and an optimal codebook from the adaptive codebook by referring to an input audio signal. a search means for searching for a suitable drive signal; a synthesis filter for synthesizing an audio signal using the searched optimum drive signal; a sampling conversion means for upsampling the drive signal read from the adaptive codebook; The apparatus includes a delay means for delaying a drive signal, a means for thinning out the delayed drive signal and storing it in an adaptive codebook, and a start point/end point determining means for determining a start point and an end point of the drive signal to be stored in the adaptive codebook. It is characterized by

The start and end points of the drive signal stored in the adaptive codebook are determined by, for example, the pitch cycle found by pitch analysis of the input audio signal, or the pitch in the previous frame found in the process of searching for the drive signal. This is done based on the cycle or the sign of the drive signal vector searched from the adaptive codebook in the previous frame.

[0022]

According to the present invention, by upsampling the drive signal and storing it in an adaptive codebook, the pitch period can be determined with a high resolution of one sample or less of the input audio signal, thereby improving the encoding quality.

Furthermore, by determining the start and end points of the drive signal stored in the adaptive codebook in advance, the size of the adaptive codebook or the number of pitch searches can be kept constant despite upsampling.
An increase in the amount of calculation and code amount when searching for a drive signal from an adaptive codebook can be avoided.

[0024] Furthermore, based on the property of the audio signal that the pitch period does not change much in a short period of time in units of frames,
The start and end points of the drive signal are determined based on the pitch period obtained by pitch analysis of the input audio signal or the pitch period obtained in the previous frame, and the scope of the adaptive codebook search is determined. Quality deterioration caused by keeping the number of drive signal searches constant regardless of upsampling is prevented.

[0025]

Embodiments Hereinafter, embodiments of the present invention will be described with reference to the drawings. FIG. 1 is a block diagram of a speech encoding device according to an embodiment of the present invention.

In FIG. 1, the input audio signal is input to input terminal 1.
00 to the frame buffer 101. The frame buffer 101 cuts out the input audio signal sequence in units of L samples and stores them as one frame signal. L is usually 160. One frame of input audio signal sequence from the frame buffer 101 is sent to the LPC analysis circuit 1
02 and weighting filter 106.

[0027] The LPC analysis circuit 102 performs LPC (Linear
Predictive Coding (Linear Predictive Coding) analysis is performed, and P LPC prediction coefficients {αi, i=
1,2,...p,} or reflection coefficient {ki, i=1,
2,...,p} are extracted. The extracted prediction coefficients or reflection coefficients are encoded with a predetermined number of bits in the encoding circuit 103 and then used in the weighting filter 106 and the weighting synthesis filters 107, 112, and 122.

The weighting filter 106 weights the input speech signal sequence when searching for a driving signal vector for the synthesis filter from the adaptive codebook 110 and the noise codebook 120. Synthesis filter 107
, 112, 122 is described by [Equation 1]. At this time, the transfer function W(z) of the weighting filter 106 is expressed by [Equation 2].

[0029]

[Equation 2] However, γ is a parameter that controls the strength of weighting (0≦γ≦1).

The weighted synthesis filters 112 and 122 are
A synthesis filter with a transfer function of H(z) and W(z)
This is a filter in which weighting filters with transfer functions are connected in cascade, and the transfer function HW (z) is described by [Equation 3].

[0031]

[Equation 3] When the weighting filter 106 is used as in this embodiment,
It becomes possible to reduce perceptual coding distortion. Also,
In this embodiment, the weighting filter 106 is provided outside the drive signal vector search loop, and as a result, the amount of calculation required for the search is significantly reduced.

Furthermore, weighted synthesis filter 112,1
A weighted synthesis filter 107 having an initial memory is provided so that the filter 22 does not affect the search for the drive signal vector. This weighted synthesis filter 107 is applied to the weighted synthesis filters 112 and 122 at the end of the previous frame.
has as its initial state the internal state held by .

Then, a zero-input response vector for the weighted synthesis filter 107 is created, and the zero-input response vector is subtracted from the output of the weighting filter 106 in the subtracter 108. As a result, the weighted synthesis filter 112
, 122 can be set to zero, and the drive signal vector can be searched without considering the influence of the previous frame.

The LPC analysis circuit 102 performs LPC analysis and calculates a prediction residual signal, and the pitch analysis circuit 104
supply the residual signal to. The pitch analysis circuit 104 determines the pitch period Tp using a known method such as the covariance method, and uses this to determine the pitch period Tp between the start point/end point determination circuit 105 and the multiplexer 142.
give to

All of the above processing is performed frame by frame. Next, a process of dividing a frame into M subframes (usually M=4) and searching for a drive signal vector performed in subframe units will be described.

The search for the drive signal vector is performed in the order of adaptive codebook 110 and noise codebook 120. First, from the adaptive codebook 110, drive signal vector Xj corresponding to pitch period j (vector dimension is L/M
=K) are sequentially read out, Xj is multiplied by a predetermined gain β in a multiplier 111, and then supplied to a weighted synthesis filter 112. The weighted synthesis filter 112 performs a filtering operation to create a synthesized speech vector.

On the other hand, the input audio signal read from the frame buffer 101 is weighted by a weighting filter 106, and then the influence of the previous frame is subtracted by a subtracter 108. Using the audio signal vector Y output from the subtracter 108 as a target vector, a subtracter 113 calculates an error vector Ej with respect to the synthesized audio vector from the weighted synthesis filter 112. Then, the squared error calculation circuit 114 calculates the squared sum of errors ‖Ej ‖,
The minimum value of |Ej || and the index j giving the minimum value are detected by the minimum distortion search circuit 115. This index j is provided to adaptive codebook 110 and multiplexer 142.

Specifically, the error vector Ej is expressed, for example, by [Equation 4]. By partially differentiating this error vector ‖Ej ‖ with respect to β and setting it to zero, the minimum value of ‖Ej ‖ when β is optimized is expressed by [Equation 5]. However, β is the gain given by the multiplier 111.

[0039]

[Math 4]

[0040]

[Equation 5] Here, ‖X‖ represents the square norm, (X, Y) represent the inner product, and H is the impulse response of the weighted synthesis filter (transfer function: HW (z)) given by [Equation 6] It's a queue.

[0041]

[Equation 6] As is clear from [Equation 6], the adaptive codebook 110
The search for the drive signal vector from is performed by calculating the second term on the right side of [Equation 6] for all code words Xj and detecting the index j at which it is maximum.

When the optimal drive signal vector Xopt is searched from the adaptive codebook 110 in this way,
A subtracter 113 subtracts the output of the weighted synthesis filter 112 corresponding to Xopt from the target vector Y, and the output of the subtracter 113 is used as the target vector for noise vector search from the noise codebook 120. The search for the noise vector from the noise codebook 120 can be performed in exactly the same way as the search for the drive signal vector from the adaptive codebook 110 and the like. If the code vector obtained by searching from this noise vector 120 is Nopt, then the drive signal vector X of the synthesis filter is expressed as X=β·Xopt +g·Nopt. However, β and g are multipliers 111 and 1, respectively.
21 is the gain given to the drive signal vector and noise vector searched from the adaptive codebook 110 and the noise codebook 120.

Next, a method of storing the drive signal vector X in the adaptive codebook 110 will be explained. In FIG. 1, the drive signal vector output from the adder 130 is upsampled by inserting zero between samples in an upsampling circuit 131, and then passed through an interpolation filter, so that the sampling frequency is doubled, for example. It is considered to be a signal sequence. Here, the interpolation filter is a low-pass filter having a cutoff frequency at the Nyquist frequency, and is also called a Nyquist filter.

The drive signal sequence thus upsampled is delayed by one sample in the delay circuit 132 from the start point a given by the start point/end point determining circuit 105 to the end point b. In this example, the start point/end point determination circuit 105 stores samples of the pitch period Tp determined by the pitch analysis circuit 104 in the adaptive codebook 110 so that the size of the adaptive codebook 110 does not change due to upsampling. The starting point a and the ending point b of the drive signal vector are determined by, for example, the following equation.

a=DTp-Np/2b=DTp+
Np /2 However, Np is the size of the adaptive codebook without upsampling, and pitch search is generally 20
Since this is often done for samples up to 147 samples, Np is set to 128. Further, D is the upsampling magnification, and in this embodiment, D=2.

The output of the delay circuit 132 is sent to the thinning circuit 13.
After being thinned out every (D-1) sample in step 3, it is cut out every dimension of the adaptive codebook 110 (40 samples = 1 subframe) and stored in the adaptive codebook 110.

As described above, in the conventional method of increasing the resolution of the adaptive codebook, the size of the adaptive codebook is increased by D times by upsampling.
There is a problem in that the amount of calculation and code required for searching the adaptive codebook increases by a factor of D. Assuming that the subframe length and the dimension of the adaptive codebook are K, and the sampling frequency of the input audio signal is 8kHz, the second right-hand side of [Equation 5]
The number of multiplications per second required to calculate the term is
[{K(K+1)/2}+2K]・Np・8×103
/K. The value of K, Np is usually used K=4
0, Np = 128, the number of multiplications is 22,52
8,000 times. Therefore, if this value is multiplied by D, the amount of calculation becomes enormous, and the circuit for realizing this also becomes large-scale.

On the other hand, according to this embodiment, by determining the starting point a and the ending point b of the drive signal vector to be stored in the adaptive codebook 110 in the starting point/end point determining circuit 105,
Although the drive signal is upsampled by the upsampling circuit 131, the adaptive codebook 11
The size of 0 can be kept constant. Therefore, an increase in the amount of calculation and code can be prevented. Moreover,
Encoding quality can be improved by upsampling.

The encoding parameters obtained through the above processing are multiplexed by the multiplexer 142 and sent as encoded outputs from the output terminal 143 to the transmission path. That is, in the multiplexer 142, the LPC analysis circuit 10
The information on the LPC prediction coefficients obtained in step 2 is sent to the encoding circuit 10.
3, the code of the pitch period Tp for each frame found by the pitch analysis circuit 104, and the adaptive codebook 11 found by the minimum distortion search circuit 115.
0 index code, a code obtained by encoding the gain information multiplied by the multiplier 111 by the gain encoding circuit 140, a code of the index of the noise codebook 120 found by the minimum distortion search circuit 125, and the multiplier. Gain information multiplied by 121 is transmitted to gain encoding circuit 1.
The codes encoded in step 41 are multiplexed.

Next, the configuration of a speech decoding device corresponding to the speech encoding device of FIG. 1 will be explained with reference to FIG. 2. In FIG. 2, input encoding parameters are first decomposed into individual parameters by a demultiplexer 201, and then decoded by decoders 202, 203, and 204, respectively. Then, a drive signal is created based on the index and gain of the decoded adaptive codebook and the index and gain of the noise codebook. This drive signal is filtered by a synthesis filter 215 to create a synthesized audio signal. This synthesized audio signal is subjected to spectrum shaping in a post filter 216 to suppress perceptual distortion, and then outputted from an output terminal 217.

The start point/end point determination circuit 205 in FIG.
The upsampling circuit 220, the delay circuit 221, and the thinning circuit 222 are the circuits 105 and 13 in FIG. 1, respectively.
Since it has the same function as No. 1, 132, and 133, the explanation will be omitted.

FIG. 3 shows a block diagram of a speech encoding device according to another embodiment of the present invention. The difference between this embodiment and the previous embodiment lies in the method of determining the pitch period Tp used in the start point/end point determining circuit 105.

In the previous embodiment, pitch analysis was performed on a frame-by-frame basis to determine the pitch period, and pitch period information was transmitted on a frame-by-frame basis. On the other hand,
In this embodiment, the pitch period for each subframe obtained by searching the adaptive codebook of the previous frame is stored in the memory 15.
0, and the pitch period estimating circuit 151 estimates the pitch period Tp in frame units based on the stored pitch period. This estimation of Tp may be performed, for example, by averaging the pitch period for each subframe. Furthermore, the pitch period Tp can also be estimated by extrapolating and predicting the pitch period for each subframe.

[0054] This method of estimating the pitch period Tp based on the pitch period of the previous frame and finding the start/end points has the effect of reducing the transmission bit rate because it is not necessary to transmit the information of Tp. be.

The same effect can be expected even if the start point/end point is determined using the sign of the drive signal vector searched from the adaptive codebook in the previous frame instead of the pitch period Tp in the previous frame.

[0056]

As explained above, according to the present invention, by determining the start and end points of the drive signal to be stored in the adaptive codebook, the drive signal is upsampled to improve the encoding quality. Since the size of the adaptive codebook is kept constant, an increase in the amount of calculation and code can be avoided. Therefore, a single or small number of DSPs
It becomes possible to perform real-time processing using , making it possible to lower prices and power consumption, and also to suppress increases in transmission bit rate.

[Brief explanation of the drawing]

[Fig. 1] Block diagram of a speech encoding device according to an embodiment of the present invention

[Figure 2] Block diagram of the audio decoding device according to the same embodiment

[Fig. 3] Block diagram of a speech encoding device according to another embodiment of the present invention

[Figure 4] Block diagram of the audio decoding device according to the same embodiment

FIG. 5 is a block diagram showing a configuration related to a drive signal vector search in a conventional audio encoding device.

[Explanation of symbols]

100...Audio signal input terminal 1
02...LPC analysis circuit 103...encoding circuit
104...Pitch analysis circuit 105...Start point/end point determination circuit 10
6...Weighting filter 107...Weighting synthesis filter 110
...Adaptive codebook 112...Weighted synthesis filter 114
... Square error calculation circuit 115 ... Minimum distortion search circuit
120...Noise codebook 122...Weighting synthesis filter 124
... Square error calculation circuit 125 ... Minimum distortion search circuit
131... Upsampling circuit 132... Delay circuit
133... Thinning circuit 140... Gain encoding circuit 1
41...gain encoding circuit 142...multiplexer
143...Output terminal 150...Memory
151...Pitch period estimation circuit

Claims (3)

[Claims] 1. A codebook storing drive signals as codewords, a search means for searching for an optimal drive signal from the codebook by referring to an input audio signal, and an optimal drive signal searched for by the search means. a synthesis filter for synthesizing audio signals using the above codebook, a sampling conversion means for upsampling the drive signal read from the codebook, a delay means for delaying the drive signal upsampled by the sampling conversion means, and the delay means. A speech code characterized by comprising means for thinning out drive signals delayed by and storing them in the codebook, and start point/end point determining means for determining the start point and end point of the drive signals to be stored in the codebook. conversion device. 2. A codebook storing drive signals as codewords, a search means for searching for an optimal drive signal from the codebook by referring to an input audio signal, and an optimal drive signal searched by the search means. a synthesis filter for synthesizing audio signals using the above codebook, a sampling conversion means for upsampling the drive signal read from the codebook, a delay means for delaying the drive signal upsampled by the sampling conversion means, and the delay means. means for thinning out the drive signal delayed by and storing it in the codebook; a pitch analysis means for pitch-analyzing the input audio signal to obtain a pitch period; and based on the pitch period obtained by the pitch analysis means, A speech encoding device comprising: a start point/end point determining means for determining a start point and an end point of a drive signal stored in the codebook. 3. A codebook in which drive signals are stored as codewords, a search means for searching for an optimal drive signal from the codebook by referring to an input audio signal input frame by frame, and a search means for searching for an optimal drive signal from the codebook. a synthesis filter that synthesizes an audio signal using the optimal drive signal,
a sampling conversion means for upsampling the drive signal read from the codebook; a delay means for delaying the drive signal upsampled by the sampling conversion means; and start point/end point determining means for determining the start point and end point of the drive signal to be stored in the codebook based on the pitch period in the previous frame found in the search process by the search means. A speech encoding device comprising: