JP3004664B2 - Variable rate coding method - Google Patents

Description

DETAILED DESCRIPTION OF THE INVENTION [Object of the Invention] (Industrial application field) The present invention relates to a variable rate coding method used for a packet communication system or an ATM communication system.

(Prior Art) A packet communication system in which an audio signal is encoded, then packetized and communicated in packet units is being realized. Packet communication systems have the advantage that lines can be efficiently used by transmitting only voiced segments using burst names of other audio signals that can handle various media signals such as voice, images, and data in a unified manner. Having. For this reason, packet communication and ATM communication have attracted attention as a leader of ISDN and BISDN, and research and development are being actively conducted.

However, in packet communication, packets are discarded at the time of congestion in a network or when packet delay is large, thereby deteriorating voice quality. In particular, when ADPCM using adaptive prediction is used as a coding method, quality degradation at the time of packet discarding is large. For this reason, Embede is an encoding method that minimizes quality degradation when discarding packets.
d If the DPCM method is `` Embeded DPCM for variable bit rate t
ransmission ", (IEEE Trans.COM-28,7, pp.1040-104
6 (July 1980)) (referred to as Reference 1).
In addition, CCITT provides CCITT SG XV III “Annex to Question
X / XV (Speech Packetization), (TD131, Geneva 6-17
June 1988) (referred to as Reference 2), the provisional recommendation of Embedded ADPCM as G, EMB as an encoding method for voice packet communication and the provisional recommendation of G, PVNP as a voice packet protocol.

FIG. 20 and FIG. 21 are block diagrams of an encoder unit and a decoder unit of the provisional recommendation G, EMB system. In FIG. 20,
The input of the encoder is a voice signal digitized by a μ-PCM or A-PCM codec. 610 is μ
PCM for converting PCM or A-PCM codes into linear PCM codes
It is a format converter. 630 is an adaptive quantizer, and 670 is an adaptive predictor. The subtraction circuit 620 calculates a difference between the input signal and the prediction signal output from the adaptive predictor 670, and sends the difference to the adaptive quantizer 630. Adaptive quantizer 630
The input prediction difference signal is quantized and output as an ADPCM code. A bit mask circuit 640 masks the lower bits of the output code of the ADPCM by the maximum number of discardable bits and shifts to the right. The output of the bit mask circuit 640 is sent as core bits to the adaptive inverse quantizer 650, and the adaptive inverse quantizer 650 performs inverse quantization of the core bits. The output of the adaptive inverse quantizer is sent to an adaptive predictor 670 and an adding circuit 660.
The addition circuit 660 also creates a local decoded signal by adding the output signal of the adaptive inverse quantizer and the output signal of the adaptive predictor. The adaptive predictor 670 is an adaptive filter having a second-order pole and a sixth-order zero, and receives the local decoded signal and the dequantized prediction difference signal to generate a prediction signal.

The number of bits of adaptive quantizer 630 and the number of core bits fed back depend on the algorithm used. For example 32
In the Kbps (4,2) algorithm, quantization is 4 bits and core bits are 2 bits. In FIG. 20, the adaptive quantizer 630 forms a feedforward path, and the bit mask circuit 640, the adaptive inverse quantizer 650, and the adaptive predictor 670 form a feedback path.

Next, the operation of the decoder will be described. The decoder in FIG. 21 is a feedback path and feedforward adaptive inverse quantizer comprising a bit mask circuit 680, a feedback adaptive inverse quantizer 690, and an adaptive predictor 710, like the coder.
It comprises a feed-forward path including a 720 and a PCM format conversion circuit 740. The feedback path is exactly the same for the coder and the decoder. Bit mask circuit 68
The value 0 indicates that only the core bits are sent to the feedback adaptive inverse quantizer 690 by masking the lower bits and shifting to the right, leaving the upper core bits of the input ADPCM code. The feedback adaptive quantizer 690 performs inverse quantization of the core bit. The adaptive predictor 710 receives the inversely quantized prediction difference signal output from the 690 and the local decoded signal output from the addition circuit 700, and outputs a prediction signal. Bit discarding on the network is performed from the lower bits of the ADPCM code,
Core bit transmission is guaranteed. Therefore, the output of the bit mask circuit 680 on the decoder side is the same as the output of the bit mask circuit 640 on the coder side.

Therefore, the adaptive inverse quantizer 690,650 and the adaptive predictor 710,6
The output of 70 is exactly the same for the coder and the decoder.

The feedforward adaptive quantizer 720 inversely quantizes the core bits of the ADPCM output code and the remaining bits without being discarded. The addition circuit 730 adds the output of the feedforward adaptive quantizer 720 and the output of the adaptive predictor 710 to generate a decoded signal. The resulting decoded signal is output to the PCM format conversion circuit 740, where the μm is converted from the linear PCM code.
-PCM or A-PCM code. 750 is ADPCM-PC
This is a tandem connection correction circuit for preventing an error due to synchronous tandem connection like M-ADPCM.

If bit discarding of an output code occurs in ADPCM that is not normal embedded, the inversely quantized prediction difference signal has different values between the coder and the decoder. As a result, the adaptive processing of the quantizer and the predictor becomes asynchronous operation different between the coder and the decoder, and the error due to discarding is filtered by the synthesis filter, so that the quality deterioration due to bit discarding increases.

On the other hand, in the above-mentioned Embedded ADPCM, only core bits are fed back to the predictor, so even if lower bits excluding core bits are discarded on the network,
Asynchronous operation of the coder and decoder does not occur. In addition, since the prediction signal is the same between the coder and the decoder, the quantization error corresponding to the number of discarded bits is simply added directly to the decoded signal, and the quality degradation due to bit discarding is small.

Reference 2 describes a voice packet configuration method and a protocol that makes use of such characteristics of the Embeded ADPCM.

FIG. 22 shows a packet format described in Reference 2. In the figure, bit 1 represents the LSB and bit 8 represents the MSB.
PD (Protocol Discriminator) is for distinguishing voice packets from other packets. BDI (Blo
ck Dropping Indicator) indicates the number of blocks that can be discarded in the packetized initial state and the number of blocks that can be discarded on each node of the network. Here, a block is information in 128-bit units obtained by collecting encoded audio data for one frame in units of bits, with the encoding frame being 16 ms (128 samples). TS (Time Stamp) indicates the sum of delay amounts generated in each node of the network. CT (Coding
Type) is a field indicating the audio encoding method used when creating the packet. SEQ (Sequence Numbher)
Is a number indicating the continuation order of packets, and is used when a packet is lost. NS (Noise Field) is a field indicating the level of background noise. NON-DROPPABLEOC
TETS is a block of core bits of the embedded ADPCM output, and is a field of information that cannot be discarded on the network. OPTIONAL DROPPABLEBLOCKS is a block of lower bits of the embedded ADPCM, and is an information field that can be discarded when requested by a system on a network. At the beginning and end of the packet, a layer 2 header and trailer are attached. In a packet network protocol using a packet having the format shown in FIG. 22, packet discarding is performed by discarding OPTIONAL DROPPABLE BLOCK in the packet.

The above is the conventional packet discard compensation method using the embedded ADPCM and the packet format. In this method, when information is discarded in a packet, that is, in a bit unit, as described above, quality deterioration is small. However, if discarding occurs on a packet basis, the Embedded AD
Since the PCM core bits are also discarded, quality degradation occurs. Due to the packet discard, the signal for one frame (16 ms) is completely lost, and the original audio signal cannot be reproduced. This state does not end in one frame but continues for one or more frames due to the asynchronous operation of the encoder and the decoder. As a method of compensating for discarding in units of packets, there is a method of interpolating and reproducing the signals of the packets before and after the discarded packet. However, since the prediction difference signal output from the ADPCM is a signal from which correlation has been removed, one frame ( (128 samples) Even if interpolation is performed using separated samples, there is almost no effect of interpolation, and quality deterioration cannot be avoided.

(Problems to be Solved by the Invention) In the conventional encoding method using the Embeded ADPCM, when discarding in packet units occurs, the core bits of the Embedded ADPCM are also discarded, so that the original audio signal cannot be reproduced and the encoder is not reproduced. And the decoder operate asynchronously,
There is a problem that quality deterioration is large.

Further, in the conventional embedded ADPCM, the temporal change of the bit rate is not actively considered, and the control method of the bit rate and the use of fixed-length cells are not sufficiently studied. Since the amount of information of the audio signal changes over time, the fixed bit rate of the embedded ADPCM suffers from the problem that the quality of the encoded audio changes, resulting in annoying sounds and a decrease in encoding efficiency. Have.

SUMMARY OF THE INVENTION The present invention has been made in view of the above problems, and has as its object to provide a variable rate coding method which has little quality deterioration even when discarded in packet units, has stable quality, and has high coding efficiency.

[Configuration of the invention]

(Means for Solving the Problems) The present invention divides a signal sequence of an input signal into signals of a plurality of frequency bands, calculates the power of the signal for each of the divided frequency bands, and converts the calculated power value into a value. Change the sum of the number of coding bits for each frequency band on a frame basis based on, and transmit a signal indicating the number of bits to be allocated for coding,
A variable-rate encoding method for encoding a signal for each frequency band, comprising: estimating an SNR of a decoded signal on a reception side;
Determine the allocation of the number of coded bits so that NR is constant, and detect the presence or absence of sound or silence of the input signal, and determine the number of coded bits when coding a signal for each band based on the detection result. This is a variable rate encoding method characterized by changing the total number of encoded bits for each band.

(Operation) In the present invention, an input signal is divided into signals in a plurality of frequency bands, and signals in each band are quantized and encoded. At this time, the number of coded bits in each band is distributed on a frame basis based on the signal power for each band calculated by calculating the signal power for each band. Thereby, the correlation or the redundancy of the input signal is removed, and the input signal can be encoded with high efficiency. At the same time, by changing the sum of the number of coded bits for each band based on the signal power for each band, the SNR of the decoded signal on the receiving side is estimated and the bit rate is controlled so that it is constant. In addition, the quality of the decoded signal can be maintained at a constant level, and the bit rate is controlled by controlling the bit rate so that the bit rate changes in accordance with the time change of the characteristics of the input signal. be able to.

Next, the signal is composed of information units called cells or packets, and the coded signal for each band and the signal representing the number of coded bits of the signal for each band are multiplexed (cellified, packetized) and transmitted. Sent to the road. At this time, priority may be given to each cell or packet. In the present invention, unlike conventional ADPCM, prediction using past signals and synchronous adaptive control of a quantization measure by a coder and a decoder are not performed, signals of a plurality of frequency bands are independently encoded in frame units. Therefore, even if any cell or packet is discarded, the effect of discarding does not affect the next cell or packet. As a result, quality deterioration due to cell discard can be extremely reduced.

Embodiment An embodiment according to the present invention will be described below with reference to the drawings. FIG. 1 is a block diagram of a coder section of a coding apparatus to which a variable rate coding method according to one embodiment of the present invention is applied.

In FIG. 1, reference numeral 1 denotes an input terminal to which a digitized signal sequence is input, and a sequence of a predetermined number of samples is accumulated in an input buffer 101. Reference numeral 102 denotes a filter bank that divides an input signal sequence into a plurality of frequency bands, and a QMF (Quadrature Mirror Filter) bank is known as an excellent filter bank that does not cause aliasing of a spectrum. Is used to divide the signal band up to 4 kHz into eight bands at equal intervals. Second
FIG. 1 is a block diagram showing one configuration example of the QMF bank. In this figure, 201 is a high-pass filter and 202 is a low-pass filter. These filters are composed of 32nd-order FIR filters. The high-pass filter 204 and the low-pass filter 205 are composed of 16-order FIR filters. In this way, the second and
2. Changing the order of the third-stage filter utilizes the fact that the slope of the spectrum of the audio signal is different between the low band and the high band, so that the delay caused by the filtering operation without deteriorating the filter performance Has the effect of reducing Note that the filter coefficients are designed so that aliasing of the spectrum does not occur.
yant, P.Noll “Digieal Coding of Waveformes”, PRENTIC
The description is omitted here because it is described in EHALL, INC (Reference 3).

In FIG. 1, reference numeral 103 denotes a normalization circuit for normalizing a signal for each band, which is an output of the QMF bank 102, as a preprocessing for quantization. A simple specific example of the normalization circuit is a circuit that divides a signal for each band by an RMS (Root Mean Square) for each band. 104 is a quantizer that quantizes the normalized signal of each band by a predetermined number of bits, 105 is a band power calculation circuit that calculates the power of the signal of each band by table lookup, If the signal sequence of the i-th band is x _i (n), i = 1, 2,.
Calculate and output the RMS value σ _i .

However, the section length for calculating the RMS is N. The quantizer 106 calculates the RMS of each band output from the band power calculation circuit.
The value σ _i is quantized by a predetermined number of bits, and the code is
Output to 111 and the inverse quantizer 107. The inverse quantizer 107 is a value obtained by inversely quantizing the sign of the above σ _i Is output. Normalization circuit 103 and bit rate control unit 108,
The bit allocation calculation unit 109 is obtained as the RMS value of each band by the duder Is used. This makes it possible to completely prevent characteristic degradation caused by mismatch between the coder and the decoder in which the number of quantization bits and the normalization parameter are different. Bit rate control unit
Reference numeral 108 controls the bit rate based on the power of each band signal so that the quality of the signal decoded by the dude is constant and the code amount output from the user is constant. However,
Change the bit rate control method between voiced and silent. Details of the bit rate control unit will be described later. The bit allocation calculator 109 calculates the amount of bits to be allocated to the quantum transformer in each band based on the power of each band signal and the bit rate output from the bit rate controller. Details will be described later. 11
Reference numeral 3 denotes a sound / silence detection unit that performs sound / silence detection in units of subframes, and details will be described later. Reference numeral 110 denotes a time stamp calculation circuit which calculates a head subframe number of a frame transmitted by a cell. Specifically, the number of subframes transmitted from one cell output from bit rate control section 108 is integrated. The time stamp of the i-th frame (cell) is T _S (i), the time stamp of the (i−1) -th frame (cell) is T _S (i−1), and the number of subframes is N _S (i− 1) T _S (i) is calculated by the following equation.

T _S (i) = T _S (i−1) + N _S (i−1) (2) In the cell unit 111, the code sequence of each band signal, the code of the RMS value of each band signal, and the The number of frames and the time stamp are converted into cells in the format shown in FIG. In the format of FIG. 3, the entire cell length is 52 bytes, and the information part is 48 bytes. The breakdown of the information part is 1 byte of time stamp, 1 byte of subframe, 4 bytes of band power, and 42 bytes of band signal.

The above is the description of the function of each unit in FIG. Next, the operation will be described.

FIG. 4 is a flowchart showing the entire operation of the coder. First, as initialization, an input buffer, a QMF bank, clearing of a time stamp and a target SNR, a maximum number of subframes, and a subframe length are set. Next, cut out the input signal sequence in subframe units, QMF filtering,
Target SNR for power calculation and bit rate control of each band signal
Is repeated until the above can be achieved, the bit allocation to be allocated to each band is calculated, the signal of each band is quantized based on the bit allocation, and then the celling process is performed. Such a series of processing is repeated for each frame (cell).

The control of the bit rate is performed according to the flowchart of FIG. In the present invention, the bit rate control mode is switched between voiced and silent based on the detection result read from the voiced / silent detecting unit. By switching the bit rate control method between voiced and silent, it is possible to prevent the bitrate from becoming unnecessarily high when there is no sound, and it is possible to transmit the background sound even when there is no sound. Can eliminate the unnaturalness of doing so.

The control of the bit rate in the sound mode is performed according to the flowchart shown in FIG. First, as the initial setting, the target SN
Rd, maximum number of subrooms per cell Ns max, subframe length
Set the L _S sample. Next, I = ∂ is set as an initial value of the number I of subframes. Then set the number of input samples to be input to the QMF into I × L _S samples, direct it to the input buffer. Next, the RMS value of each band calculated by the band power calculator Together with read-free, to calculate a number average bits per sample R needed to transmit a signal sequence of I × L _S samples to be coded in one cell by the following equation.

Here, B is the total number of bits allocated for transmitting the code of the band signal, and in the format of FIG.
= 42 × 8 = 336 bits.

Next, the RMS value of each band Of the signal decoded by the decoder using
The SNR is estimated by the following equation.

Here, Mb is the number of band divisions, and in this embodiment, Mb = 8.

The above SNR estimation formula is based on the result of theoretical analysis of the root-mean-square value of the decoding error when the optimal bit allocation is performed in the sub-band coding scheme. Table 1 compares the value estimated by equation (4) with the SNR value obtained by computer simulation. From this table, it can be seen that the estimated value is in good agreement with the SNR value in the actual encoding.
However, Table 1 shows the case where the bit rate is 16 kbps.

After specifying the SNR, the SNR is compared with the target SNRd. If the SNR is larger than the SNRd, the number of subframes I is equal to the maximum number of subframes Ns.
After checking that the number is equal to or less than max, the number of subframes is incremented, and the process returns to the flowchart. SNR
Is repeated until is smaller than SNRd, and SNR> SNR
The bit rate per sample and the number of subframes (I-1) immediately before d is output. The number of subframes is Ns m
If it exceeds ax, the bit rate per sample and the number of subframes I = Ns max are output. The bit rate control method described here increases the number of input samples to be coded while estimating the SNR, and changes the bit rate, so that the quality can always be kept constant. Encoded data can be accurately placed in fixed-length cells. Since the bit rate is changed according to the temporal change of the property of the input signal, there is an advantage that the coding efficiency is high.

FIG. 7 is a flowchart showing a bit rate control method in the silent mode. In this method, when silence continues, the number of subframes I is set to the maximum number of subframes Ns m
ax and set the bit rate to the lowest rate, i.e. Set to. If a sound is produced in the middle of a frame (I <Ns max), the frame is aborted at that time, the number of subframes is set to I, and the bit rate is set to Set as

Next, the operation of bit allocation calculation section 109 will be described.
FIG. 12 is a flowchart of the bit allocation calculation unit 109. First, the RMS value as the power of each band After reading the bit rate R per sample and the bit rate R per sample from the inverse quantizer 107 and the bit rate control unit 108, the bit allocation amount R _k for each band is calculated according to the following equation.

The above equation is an optimal bit allocation equation that minimizes the mean square value of the decoding error. NSJayant and P. Noll: “Digital Co.
ding of Waveforms ", PRENTICE-HALL, NJ (Reference 4).

The bit allocation amount R _k calculated by equation (5) is a real value, but when a scalar quantizer is used to quantize the signal of each band, it is necessary to set R _k to an integer value. Because then
Correct R _k . 13 is a flow chart showing an example of a correction method of the R _k. First, R _k is converted to an integer by truncation after the decimal point, and then the number of remaining bits R _r generated by the conversion to an integer.
To Calculated by the next remainder bits R _r going to bit by bit reallocation in the order of large band power. Such redistribution of bits in the order of power is effective in reducing decoding errors.

FIG. 8 shows a sound / silence detecting section 11 according to an embodiment of the present invention.
FIG. 3 is a block diagram of FIG. In Figure 8, 1110 LPC cepstrum of signal input from and the input terminal 1100 is LPC cepstrum extraction circuit _{C i (i = 1,2, ...} , P) is computed for each sub-frame by a known method. Here, P is the order of analysis, for example, P = 16. The method of calculating the LPC cepstrum is described, for example, in Sadateru Furui, “Digital Speech Processing” (Tokai University Press, 1985).

LPC cepstrum C _i obtained is input to the feature parameters projective circuit 1140. This circuit 1140 is an inner product operation circuit 11.
20 and a voiced main component vector memory 1130.

As shown in the flowchart of FIG. 9, the voiced main component vector memory 1130 collects voices (learning data) collected in advance in a telephone use environment (step), performs labeling as voiced (step 2), The LPC cepstrum of the sound part is calculated (step 3), and the LPC cepstrum is obtained by performing principal component analysis. Actually LP
The covariance matrix of the C cepstrum is calculated (step 4), eigenvalues are obtained (step 5), and the principal vectors corresponding to eigenvalues having large absolute values are set as principal component vectors in order (step 6). Here, the first to third principal component vectors V ₁ , V ₂ , V ₃ are stored in the memory 130. Inner product calculating circuit 120, a vector _{C = (C 1, C 2} , ..., C P) to the LPC cepstrum C _i as elements an inner product operation between the principal component vectors V _1, V _2, V ₃ do according to: , V ₁ , V ₂ , and V ₃ are used as the coordinate axes to determine the projection point Q of the vector C on the three-dimensional principal component space.

Here, upsilon _ij is the j component Q _i of the principal component vector V _i is a component of the coordinate axes V _i of the projection point Q.

The voiced area defining parameter memory 1160 stores parameters for defining a voiced area on the principal component vector space, and the silent area defining parameter memory 1170 similarly stores parameters for defining a silent area on the principal component vector space. Is stored. The the voiced and silent regions on V _1, V ₂ axes 1
0When the rectangle is defined as a rectangle (shaded area in the figure), the parameters that define the sound area are 有_1l , υ _1h , υ _2l , υ _2h
And the parameters defining the silent region are ζ _1l , ζ _1h , ζ
_2l and ζ _2h . These parameters are determined by performing statistical processing on the LPC cepstrum of the sound section and the LPC cepstrum of the silent section of the voice collected in advance in the telephone use environment. The determination circuit 1150 determines whether the projection point Q belongs to any of a sound area and a silence area or does not belong to any area in the principal component vector space (a) sound, (b)
Silence, (c) Indeterminate determination. That is, it is determined as _{(a) υ 1l Q 1 υ} 1h cutlet upsilon _2l Q _Z silence when upsilon _2h when voiced _{(b) ζ 1l Q 1 ζ} 1h cutlet _{ζ 2l Q 2 ζ 2h (c} ) Other undefined .

In the sound / non-sound determining circuit 1180, as shown in the flowchart of FIG. 11, when the output step 1 of the determining circuit 1150 determines that there is a sound or no sound, the effect is output as it is (to the end).

If there is not enough (step 2), the conditional probabilities of speech and silence of the current frame, which are determined based on the past three frame determination results, are obtained from the conditional probability table 1200 by table lookup (step 3). If the conditional probability is larger than the silence conditional probability (step 4), it is determined that there is sound, and vice versa (step 5). Reference numeral 1190 denotes a determination result memory for storing determination results for at least three frames.

Assuming that the determination result of n frames is T _n , the conditional probability P conditioned on the determination results T _n−1 , T _n−2 , T _n−3 of the past three frames
(T _n / T _n−1 , T _n−2 , T _n−3 ) is represented by the following equation.

P (T _n , T _n−1 , T _n−2 , T _n−3 ) and P (T _n−1 , T _n−2 , T _n−3 )
Performs voice / silence labeling of speech (learning data) collected in advance in a telephone use environment by visually recognizing a waveform or spectrum for each frame, and generates voice and silence labels of four consecutive frames and three frames. Is calculated in advance on the basis of The conditional probability obtained by the calculation of the expression (5) is stored in the conditional probability table 1200 in advance.

As described above, performing a sound / non-speech determination based on the conditional probability obtained from the learning data is based on the knowledge about speech that the pattern of sound → no sound → voice → no sound is very small. Since the judgment is made, there is an effect that erroneous judgment of sound / no sound is reduced.

In addition to the LPC cepstrum, a signal power, the number of zero crossings, a linear prediction coefficient, an autocorrelation coefficient, a DFT coefficient, and a combination thereof can be used as the characteristic parameter of the signal. Also, various modifications are possible, such as the number of principal component vectors used in the determination and the number of past frames when calculating the conditional probability can be set to an arbitrary number.

 The above is the description of the coder section.

FIG. 14 is a block diagram of a decoder section of an encoding device to which a variable rate encoding method according to one embodiment of the present invention is applied.

In FIG. 14, reference numeral 301 denotes a cell decomposing unit for decomposing cells in the format shown in FIG. 3 into data of a time stamp, band power, and band signal. Further, reference numeral 302 denotes an inverse quantizer for inversely quantizing the signal of each band, which is realized by table lookup in the same manner as 104 in FIG. Reference numeral 303 denotes an inverse normalization circuit, which outputs the output of the inverse quantizer 302 and the R of each band.
MS value Is multiplied.

Reference numeral 305 denotes a bit allocation calculation unit, which is an RMS value of each band. When calculating the amount of bits allocated to each band like the 109 of FIG. 1 using the sub-frame number N _S in one cell. First,
The average number of bits R per sample is calculated according to equation (3), and then the bit allocation amount R for each band is calculated according to equation (5).
_k (k = 1, 2,..., Mb) is calculated. Cell discard detection circuit 306
Detects whether the cell is discarded using the transmitted time stamp T _S and the number of subframes N _S. FIG. 15 is a flowchart showing the detection method. The detection method will be described with reference to FIG. First, the time stamp T _S and the number N _S of subframes are read, and these are always held for two cells.
Next, the time stamp T _S (n−1) of the cell one time before the cell that has arrived at the current time (time n) and the number of subframes N _S (n−
The expected value T of the current time stamp is calculated by the following equation using 1).

T = T _s (n−1) + N _s (n−1) Next, T is compared with the current time stamp T _s (n). If they match, no discard occurs. It is determined that there was discard immediately before. For example, in the case of Figure 16 _{T S (n-1) =} 1 N S (n-1) = m T = T S (n-1) + N S (n-1) = m + 1 = T s (n) ( 7) Therefore, there is no disposal.

In FIG. 14, reference numeral 307 denotes a pre-interpolation processing circuit, which bypasses the signal of each band to the QMF bank 308 when there is no discard, and substitutes "0" instead of each band for the QMF bank 308 when there is discard. To enter.

The QMF bank 308 inputs a signal of a divided band and outputs a signal of a full band, and has a configuration in which the input and output are reversed in FIG. The decoded signal obtained by passing through the QMF bank is sent to an interpolation processing section 309, where interpolation of signal dropout due to cell discard is performed. FIG. 17 is a block diagram showing an embodiment of the interpolation processing unit. Also the 18th
The figure is a signal waveform example showing the interpolation processing. In FIG. 17, the decoded signal input from the input terminal 400 is bypassed to the output terminal 310 if there is no cell discard according to the cell discard signal supplied from the terminal 409, and if there is cell discard, the following is performed. The interpolation processing is performed as follows. First, the decoded signal of the cell immediately before discarding is read from the buffer 401 and input to the LPC analysis unit 402. The LPC analysis unit performs an LPC analysis by the autocorrelation method or the covariance method and performs prediction coefficients α ₁ , α ₂ ,.
_P (P is the prediction order, here, 8) and the prediction residual signal e (n) are obtained. LRRab for LPC analysis
Since it is described in detail by S. Suzuki, "digital signal processing of voice", written by iner RWShater, Corona Corporation (Reference 5), the description is omitted. Note that the transfer function H (z) of the prediction filter is It is. Next, pitch analysis is performed on the prediction residual signal e (n) to obtain a pitch period _TP , a gain g, and a prediction residual signal e _P (n). The pitch analysis method is also described in Reference 5, but here, _TP and g are obtained as follows.

In decoded signal of the immediately preceding cell of a cell of discarding (frame), the frame last sample points n = N Distant, define the following error function E (T _P).

Here, L is the section length for evaluating the error, and here L
= 70. Pitch period T _P is determined in the above equation E a (T _P) as T _P minimized. After pitch gain g is obtained the T _P,
It is calculated by the following equation.

The prediction residual signals e (n) and e _P (n) are calculated by the following equations, and e _P (n) is stored in the buffer 404.

Here, x (n) is a signal output from the buffer 401. Then, in the drive signal generation circuit 405, from the end reads T _P samples from the buffer 404 of the prediction residual signal e _P a previous cell (n), pitch gain g in the residual signal
Are repeatedly connected to generate a drive signal as shown in FIG. 18 (c). Next, this drive signal is input to a synthesis filter 406 which is an inverse filter of the prediction filter of Expression (8), and a signal of a discarded cell is synthesized. The combined signal is smoothed at 407 between the cell before the discarded cell and the decoded signal of the cell before and after the discarded cell. Smoothing is performed according to the following equation, where x (n) is the decoded signal of the previous cell, (n) is the combined signal, and y (n) is the smoothing output.

y (n) = {1-W (n)} x (n) + W (n)
(N) (13) where W (n) is a smoothing window function,
The one shown in FIG. 19 is typical.

The above-described interpolation processing in the present embodiment can be said to be interpolation at the drive signal level.However, even if there is a discontinuity between the preceding and following cells and the waveform at the drive signal level, the synthesis filter can be used. By passing, the discontinuous point is smoothed,
There is an effect that it is hardly understood at the audio level. Further, since the continuity between the preceding and succeeding cells is further enhanced by the smoothing circuit, there is an effect that deterioration of the code signal due to cell discard is hardly perceived.

In the above embodiment, the number of bands is 8, and the subframe length is 24.
As a result of computer simulation under the conditions of samples, maximum number of subframes of 12, and target of 8NR22dB, the average bit rate of 21kbps achieves a quality of 32kbps ADPCM or higher, and the cell loss rate is 5%, and quality deterioration is hardly perceived. It was confirmed that it had excellent properties.

〔The invention's effect〕

As described above, according to the present invention, since the encoding bit rate is controlled on a frame basis, the quality of the decoded signal can be kept constant, and the contropy (information amount) of the input signal can be maintained. , The coding rate changes according to, so that high coding efficiency can be obtained. Furthermore, since the number of coded bits in each band is distributed based on the signal power for each frequency band, the SNR of the decoded signal is improved, and high decoding quality is obtained.

In addition, since the sum of the coded data is controlled so as to fit in the fixed-length cell, there is no decrease in the effect due to excess or deficiency of the coded data. Further, according to the present invention, since signals of a plurality of frequency bands are independently encoded in frame units, a cell such as an ADPCM or an Embedded ADPCM which performs prediction and quantization control using past signals is used. The quality can be kept only in the discarded cells without the quality degradation due to discarding, and the effect is small even if any cells are discarded. Therefore, it is not necessary to perform the priority control of the cells, thus simplifying the system. effective.

In addition, there is an effect that the discarded cells can be reproduced by the interpolation processing with almost no perceived deterioration in quality.

[Brief description of the drawings]

FIG. 1 is a block diagram of a coder section of an encoding apparatus according to one embodiment of the present invention, FIG. 2 is a block diagram showing an example of a configuration of a QMF bank in FIG. 1, and FIG. FIG. 4 is a flowchart for explaining the operation of the coder, FIG. 5 is a flowchart for explaining the operation of the bit rate control unit in FIG. 1, and FIGS. 6 and 7 are diagrams for explaining the present invention. FIG. 8 is a block diagram of a sound / non-sound detector according to one embodiment of the present invention, FIG. 9 is a flowchart showing a procedure for obtaining a principal component vector according to one embodiment of the present invention, and FIG. FIG. 11 is a diagram showing a sound area on a principal component vector space according to an embodiment of the present invention, FIGS. 11 to 13 are flowcharts for explaining an embodiment of the present invention,
14 is a block diagram of a decoder unit, FIG. 15 is a flowchart showing the operation of the cell discard detection circuit of FIG. 14, FIG. 16 is a diagram for explaining a cell discard detection method, and FIG. 17 is FIG. 18 is a block diagram showing an example of the configuration of an interpolation processing unit, FIG. 18 is a waveform example for explaining the interpolation processing, FIG. 19 is a diagram showing a window function of smoothing, and FIG. 20 is a block diagram of a coder unit of the conventional embedded ADPCM. FIG. 21 is a block diagram of a decoder section of a conventional embedded ADPCM, and FIG. 22 is a view showing a conventional packet format. 100, 200, 300, 400, 600 ... input terminal, 101 ... input buffer, 102, 308 ... QMF bank, 103 ... normalization circuit, 104, 1
06: Quantizer, 105: Band power calculation circuit, 107, 304
…… Inverse quantizer, 108 …… Bit rate control unit, 109,305
... Bit allocation calculation unit 110 110 Time stamp calculation circuit 111 Cell conversion unit 112 Output terminal 1100 Input terminal 1110 LPC cepstrum extraction circuit 1120 Inner product calculation circuit 1130 ...... Voiced principal component vector memory, 1140
…… Feature parameter projection circuit, 1150 …… Judgment circuit, 1160
…… Sounded area definition parameter memory, 1170 …… Silence area specified parameter memory, 1180 …… Sound / silence determination circuit,
1190: determination result memory, 1200: conditional probability table, 201, 204: high-pass filter, 202, 205 ... low-pass filter, 203: down-sampler, 301: cell decomposition unit, 303: denormalization circuit, 306 ... … Cell discard detection circuit,
307 ... interpolation pre-processing circuit, 309 ... interpolation processing unit, 310 ...
Output terminal, 401, 404: buffer, 402: LPC analyzer, 4
03 ... Pitch analysis unit, 405 ... Drive signal generation circuit, 406 ...
... Synthesis filter, 407 ... Smoothing circuit, 408 ... Switch, 610,740 ... PCM format conversion circuit, 620 ...
… Subtraction circuit, 630… Adaptive quantizer, 640,680… Bit mask circuit, 650… Adaptive inverse quantizer, 660,700,730…
Adder circuit, 690 ... feedback adaptive inverse quantizer, 720
…… feedforward adaptive inverse quantizer, 750 …… tandem connection correction circuit.

Claims (1)

(57) [Claims] A signal sequence of an input signal is divided into signals of a plurality of frequency bands, the power of the signal is calculated for each of the divided frequency bands, and coding for each frequency band is performed based on the calculated power value. The sum of the number of bits is changed for each frame, and a signal indicating the number of bits to be allocated for encoding is transmitted.
A variable rate encoding method for encoding a signal for each frequency band, comprising: estimating an SNR of a decoded signal on a receiving side; determining an allocation of the number of encoded bits so that the SNR is constant; Detecting the presence or absence of the input signal, silence, and changing the sum of the number of coded bits and the number of coded bits for each band when coding a signal for each band based on the detection result. Encoding method.