JPH05218980A - Voice coding communication system and its device - Google Patents

Abstract

PURPOSE:To improve the reproduction sound quality by reducing distortion of an especially high frequency region attended with a half rate processing of a transmission line of the analysis synthesis system voice coding communication system. CONSTITUTION:After a DCT coefficient obtained by transforming a long period prediction residual signal into a frequency region by a discrete cosine transformation device 72 is divided into N sets of frequency regions by a divider 73, the result is normalized by normalizing devices 74,75 and quantized with vector quantization devices 76,78 by using a code book 77, pitch information Pa3, maximum values of DCT coefficients Pb3-Pc3, and DCT coefficient vector numbers Pd3-pe3 are coded and sent.

Description

Detailed Description of the Invention

[0001]

[Industrial application] The spread of automobiles and mobile phones is remarkable, and it is expected that the current analog system will not be able to accommodate the increasing number of subscribers. In order to use radio waves more effectively, a plan to shift to a digital system is underway, and standardization specifications for the first generation (full rate) have been released from RCR (Radio System Development Center). The coding speed of the speech coding system in this is 11.2 kbps (bits / second) for speech data and redundant data for error correction. Furthermore, half-rate speech coding is planned for the purpose of doubling the radio wave utilization efficiency. The coding rate of this half-rate voice coding is 5.6 kbps for voice data and redundant data for error correction. The present invention relates to a voice coding communication system for compressing voice data and an apparatus thereof for the purpose of being applied to this half rate system.

[0002]

2. Description of the Related Art FIG. 9 is a block diagram of a conventional adaptive transform coding communication system using pitch prediction, in which (A) is a speech coding apparatus and (B) is a speech decoding apparatus. In this system, audio data is transmitted at 64 kbps (6.4 kHz), for example.
Sampling, 10-bit quantization) to 4.
This is a method of compressing information to 5 kbps. When this method is applied to the half rate, the audio data is 4.5 kbps
Therefore, 1.1 kbps as an error correction redundant bit
Is assigned.

In the following, as an example, voice data is 64 kbps.
To 4.5 kbps will be described.
In FIG. 9A, 10 at 6.4 kHz sampling
Bit-quantized input audio signal (64 kbps) a
Outputs the pitch information Pa extracted by the long-term prediction analyzer 11 for each frame (30 msec: 192 samples) and outputs the long-term prediction residual signal b which is a signal from which the pitch component is removed. The long-term prediction residual signal b is
After being divided into subframes (15 msec: 96 samples), they are transformed into the frequency domain by the discrete cosine transform (DCT) unit 12, and the DCT coefficient c (96 samples / subframe) is output. The DCT conversion formula will be described later. The DCT coefficient c is decimated by the adaptive decimator 13 for each subframe and information is compressed. In the thinning method here, the amplitude of each DCT coefficient changes for each subframe, so a limited number of DCT coefficients with large amplitude are selected to accommodate it, and the remaining DCT coefficients with small amplitude are set to 0. To do. The amplitude information and the position information are output as DCT information Pb. Pitch information Pa, DCT information Pb
Is converted into a digital signal sequence d by the encoder 14, multiplexed and sent to the receiving side.

On the receiving side, the digital signal train e is received and the separation circuit 21 separates the pitch information Pc and the DCT information P.
Separate into d. In the adaptive thinning-out decoder 22, the pitch information P
The transmitted DCT coefficient is reproduced based on the DCT coefficient amplitude information and position information in c, and interpolation is performed by inserting 0 at the position of the DCT coefficient that has not been transmitted. The reconstructed DCT coefficient f is converted to an inverse discrete cosine converter (IDCT device)
It is converted to the time domain by 23 and the long-term prediction residual signal g is reproduced. The long-term prediction synthesizer 24 adds pitch information P _d to the long-term prediction residual signal g to decode and reproduce the voice signal h. Conventional codec frame (30 mse
An example of bit allocation for each c) is shown in Table 1 below. Five synchronization bits are inserted at the beginning of each frame for frame synchronization. Converting the total in Table 1 per second, 135 bits / 30 msec = 4.5 kbps
Becomes

[Table 1] In mobile communication systems such as mobile phones and car phones, the transmission line conditions are harsh, unlike wired or fixed communication systems.
The bit error rate is always 0.1% to 1%, and it is not uncommon to reach about 10%. For this reason, the half-rate speech coding system needs to have a strong error correction function, and about 35% (about 2 kbps) or more of the total coding speed (5.6 kbps) is allocated to redundant bits for error correction. Can be said to be necessary. Therefore, when applied to the half-rate speech coding method, the coding rate of the speech data is about 3.6 kbps or less and high quality (log-P).
It is required to obtain reproduced sound of CM 6 bits or more).

The reproduced voice quality in the above-mentioned system is a voice coding rate of 4.6 kbps, and only log-PCM 4-bit equivalent is obtained, and the voice coding rate is further increased to 3.6 k.
When it is reduced to bps or less, the number of DCT coefficients that can be transmitted is reduced and distortion in the frequency domain is increased, so that the reproduced voice quality is further deteriorated. That is, in the conventional system, the reproduced voice quality cannot be lowered to 3.6 kbps or less with the log-PCM equivalent to 6 bits, and the performance (quality, error correction capability) required for the half rate voice coding system can be obtained. Can not meet. Therefore, the present inventor
In order to improve this problem, vector quantization is introduced to the above method, and it is applicable to a half rate system (speech coding speed is 3.6 kbps or less, reproduced speech quality is lo).
A voice coding method and its apparatus have been previously proposed (see g-PCM 6 bits or more) (see Japanese Patent Application No. 3-329782). The content is to reduce the coding speed while maintaining the reproduced voice quality higher than that of the conventional method by replacing the adaptive decimator in the above method with a vector quantizer, and FIG. 10 shows an embodiment thereof. It is a block diagram of a voice coding communication system. (A) is a speech encoding device, and (B) is a speech decoding device.

In FIG. 10A, an input voice signal (64 kHz sampling, 10-bit quantized) (64
kbps) a1 is pitch information Pa for each frame (30 msec: 192 samples) by the long-term prediction analyzer 31.
1 is extracted and output, and a long-term prediction residual signal b1 that is a signal obtained by removing the pitch component from the input signal a1 is generated and output. The discrete cosine transformer 32 transforms it into the frequency domain for each subframe (15 msec: 96 samples), and the DCT coefficient c1 as a frequency component.
(96 coefficients) is output. Discrete cosine transform will be described later. The DCT coefficient c1 is converted to DC by the normalizer 33.
The DCT coefficient maximum value Pb is normalized by the maximum value of the T coefficient.
1 and the normalized DCT coefficient d1 are obtained. Next, the normalized DCT coefficient d1 is vector-quantized by the vector quantizer 34 and the codebook 35. The codebook 35 stores, for example, 512 types of vector patterns. The vector quantizer 34 compares the DCT coefficient d1 that is an unknown input vector with the vector in the codebook 35, selects the vector with the smallest inter-vector distance, and outputs the vector number Pc1. The vector number Pc1 is quantized with 9 bits. The vector number Pc1 is encoded by the encoder 36 together with the maximum value Pb1 of the DCT coefficient and the pitch information Pa1 in the form of the digital sequence signal e1 and then multiplexed and transmitted to the transmission line.

The digital column signal f1 received via the transmission line is separated by the separation circuit 41, and DCT vector number Pd1, DCT maximum value information Pe1 and pitch information P are obtained.
f1 is taken out, vector dequantization is performed by the vector dequantizer 42 and the codebook 43 using the DCT vector number Pd1, and the normalized DCT coefficient g1 is reproduced. Here, the codebook 43 has the same contents as the codebook 35 of the encoding apparatus, and by specifying the vector number Pd1, the same vector as the vector selected on the encoding apparatus side can be obtained. In the inverse quantizer 44, the maximum DCT value Pe
The DCT coefficient h1 is reproduced by multiplying the normalized DCT coefficient g1 by 1. The inverse discrete cosine transformer 45 transforms the DCT coefficient h1 into the time domain to reproduce the long-term prediction residual signal i1. The long-term prediction synthesizer 46 adds pitch information Pf1 to the long-term prediction residual signal i1 to decode and reproduce the voice signal j1. The bit allocation in this method is shown in Table 2 below. By using this bit distribution, the coding speed can be reduced to about 1.4 kbps (43 bits / 30 msec), and by appropriately selecting the content and number of each representative vector, log-PCM equivalent to 6 bits can be obtained. It is expected that reproduced voice quality will be obtained.

[Table 2]

[0008]

In the above-mentioned configuration proposed by the present inventor, the voice coding speed can be suppressed low by introducing vector quantization, but the reproduced voice quality is still not sufficiently maintained. There is a problem that is not. The reason for this is that when vector quantization is performed, the vector is selected with emphasis on a low frequency region (about 2 kHz or less) where power is concentrated, and quantization distortion is generated in a high frequency region with low power (about 2 kHz or more). This is because the sound quality becomes large, and the distortion quality deteriorates in the high frequency region of the reproduced sound. Therefore, the lo required in half-rate systems
There is a problem that it is difficult to obtain quality equivalent to 6 bits of g-PMC. An object of the present invention is to further improve the above problems of the prior application and to be applicable to a half rate system (voice coding speed of 3.6 kbps or less, reproduced voice quality of log-PCM equivalent to 6 bits or more) An object of the present invention is to provide a coded communication system and its device.

[0009]

In order to solve the problems of the above-mentioned proposal, the present invention uses a plurality of DCT coefficients at equal intervals (N
(Where N is an integer of 2 or more) and vector quantization is performed separately for each to reduce the quantization distortion in the high frequency region and improve the quality of reproduced voice. That is, the configuration is such that, on the transmission side, the input speech signal is subjected to long-term prediction analysis to obtain pitch information and a long-term prediction residual signal, and the DCT coefficient obtained by transforming the long-term prediction residual signal into the frequency domain by discrete cosine transform. DCT
The coefficient maximum value is output, and the DCT coefficient normalized by the DCT coefficient maximum value is vector-quantized using a codebook, and the DCT vector number of the approximate pattern read from the codebook, the pitch information, and the DCT coefficient maximum value. Are encoded in the form of a digital signal sequence and transmitted to a transmission line, and the receiving side separates the digital signal sequence from the transmission line and extracts the pitch information, the maximum DCT coefficient value and the DCT vector number. , A vector corresponding to the DCT vector number is read from the codebook, inversely quantized to obtain a normalized DCT coefficient, denormalized by the maximum value of the DCT coefficient to reproduce the DCT coefficient, and then long-term prediction by inverse discrete cosine transform A voice that reproduces a residual signal, adds the pitch information to the long-term prediction residual signal, and decodes and reproduces a voice signal. In Goka communication system, the D together with the DCT coefficients obtained by the discrete cosine transform on the transmission side is divided into equal intervals of a plurality of frequency domain, and outputs the DCT coefficients maximum value of the divided frequency regions, respectively
The DCT coefficients normalized by the CT coefficient maximum value are vector-quantized using a codebook, and the DCT vector number of the approximate pattern read from the codebook, the pitch information, and the plurality of DCT coefficient maximum values are digital signals. The digital signal sequence from the transmission line is separated into the pitch information, the plurality of maximum DCT coefficient values, and the plurality of DCT vector numbers on the receiving side by encoding in the form of a sequence and transmitting to the transmission line. Vectors corresponding to the plurality of DCT vector numbers are read out from the codebook, inversely quantized, normalized DCT coefficients are reproduced and combined, and then a long-term prediction residual signal is reproduced by inverse discrete cosine transform. It is characterized by doing so.

Further, on the transmitting side, the input speech signal is subjected to long-term prediction analysis to obtain pitch information and a long-term prediction residual signal, and the DCT coefficient obtained by transforming the long-term prediction residual signal into the frequency domain by discrete cosine transform. The DCT coefficient maximum value of the DCT coefficient and the DCT coefficient normalized by the DCT coefficient maximum value
Vector-quantizing coefficients using a codebook, coding the DCT vector number of the approximate pattern read from the codebook, the pitch information, and the DCT coefficient maximum value in the form of a digital signal sequence, and sending out to the transmission line; On the receiving side,
The digital signal sequence from the transmission line is separated, the pitch information, the DCT coefficient maximum value and the DCT vector number are taken out, and the vector corresponding to the DCT vector number is read out from the codebook and dequantized and normalized. After obtaining the DCT coefficient, the DCT coefficient is inversely normalized by the maximum value of the DCT coefficient to reproduce the DCT coefficient, the long-term prediction residual signal is reproduced by the inverse discrete cosine transform, and the pitch information is added to the long-term prediction residual signal. In a voice coding communication system for decoding and reproducing a voice signal, a DCT coefficient obtained by the discrete cosine transform on the transmitting side is divided into a plurality of frequency regions at equal intervals, and odd-numbered frequency regions among the divided frequency regions For the DCT coefficients of the above, the maximum value of the DCT coefficient is output respectively, and D normalized by the maximum value of the DCT coefficient is output. The T coefficient vector quantization using a single codebook, DC approximate pattern from said one codebook
Each T vector number is read out, and for the DCT coefficient of the even-numbered frequency domain of the divided frequency domain, the DCT coefficient of each domain is frequency-inverted, and then the DCT coefficient maximum value is output and the DCT coefficient maximum value is output. The normalized DCT coefficients are vector-quantized by sharing the one codebook, DCT vector numbers of approximate patterns are read from the one codebook, and the pitch information, the plurality of DCT coefficient maximum values, and the plurality of DCT coefficients are read out. And the DCT vector number of the DCT vector number are encoded in the form of a digital signal sequence and transmitted to the transmission line, and at the reception side, the digital signal sequence from the transmission line is separated to obtain the pitch information and the maximum values of the plurality of DCT coefficients. And the plurality of DCs
The T vector numbers are respectively taken out, and the vectors respectively corresponding to the plurality of DCT vector numbers are read out from one codebook having the same contents as the one codebook on the transmission side and inversely quantized to obtain the normalized DCT coefficients. DC which is reproduced and frequency-inverted the DCT coefficient in the odd-numbered frequency domain and the DCT coefficient in the even-numbered frequency domain
It is characterized in that the long-term prediction residual signal is reproduced by the inverse discrete cosine transform after synthesizing with the T coefficient.

[0011]

[Example] Here, high frequency range (2 kHz or more)
In order to reduce the distortion at 1, the following description will be made on the assumption that N = 2 and equally divides the high frequency region and the low frequency region and separately performs vector quantization. 1 is a block diagram showing a speech encoding apparatus according to a first embodiment of the present invention, and FIG. 2 is a block diagram showing a speech decoding apparatus according to a first embodiment of the present invention. In FIG. 1, 6.4 kHz sampling, 10-bit quantized input speech signal (64 kbps) a3
Is a frame (10 mse by the long-term prediction analyzer 71).
The pitch information Pa3 is extracted and output for each (c: 64 samples), and the long-term prediction residual signal b3, which is a signal obtained by removing the pitch component from the input signal a3, is generated and output. The discrete cosine transformer 72 transforms it into the frequency domain for each frame (10 msec: 64 samples) and outputs the DCT coefficient c3 (64 coefficients) which is a frequency component. Discrete cosine transform will be described later. DC
The DC coefficient coefficient divider 73 equally divides the T coefficient c3 into N (for example, N = 2) frequency regions, and the T frequency coefficient c3 is in a low frequency region (0
DCT coefficient d3 (32 coefficient) of 1.6 kHz) and DCT coefficient e3 of high frequency region (1.6 to 3.2 kHz)
(32 coefficients). d3 and e3 are #
1 normalizer 74, #N normalizer (N = 2) 75
DCT coefficient maximum values Pb3, Pc3 and normalized DCT coefficients f3, g3 normalized by the maximum value of the CT coefficient
Is output. f3 and g3 are the # 1 vector quantizer 76, the # 1 codebook 77, and the #N vector quantizer (N
= 2) 78 and #N codebook (N = 2) 79, vector quantization is performed, and vector numbers Pd3 and Pe3 are output.
By the encoder 80, the maximum values Pb3, Pc of the DCT coefficients
3 and the pitch information Pa3 are converted into a digital signal train, which is then multiplexed and output to the transmission line as an output signal h3.

Next, as shown in FIG. 2, the digital signal sequence i3 received via the transmission line is separated by the separation circuit 91 to generate DCT vector numbers Pf3, Pg3, DCT.
The maximum coefficient values Ph3, Pi3 and the pitch information Pj3 are extracted. From the DCT vector numbers Pf3 and Pg3, the # 1 vector dequantizer 92 and the # 1 codebook 93 and #N, respectively.
Vector dequantizer (N = 2) 94 and #N codebook (N =
2) Vector dequantization using 95 and normalized 0
DCT coefficient j3 of up to 1.6 kHz and 1.6 to 3.2 kHz
Reproduce the DCT coefficient k3 of z. Where # 1 codebook 9
The # 3 codebook 95 and the #N codebook 95 have the same contents as the # 1 codebook 77 and the #N codebook 79 of the coding apparatus, respectively. j3
h3 is # 1 inverse normalizer 96 and DCT coefficient maximum value Ph3, #N inverse normalizer 97 and DCT coefficient maximum value Pi3, respectively.
Denormalized by DCT coefficients m3, n of each frequency band
3 is played. The m3 and n3 are combined by the combiner 98 to reproduce the DCT coefficient q3. This DCT coefficient q3
Is transformed into the time domain by the inverse discrete cosine transform (IDCT) device 99 to reproduce the long-term prediction residual signal r3. In the long-term prediction synthesizer 100, pitch information Pj is added to the long-term prediction residual signal r3.
3 is added and the audio signal s3 is decoded and reproduced.

The bit allocation in this method is shown in Table 3 below. If this bit allocation is used, the speech coding rate will be about 3.4 kbps (34 bits / 10 msec),
Distortion in the high frequency region is reduced, and reproduced voice quality equivalent to 6 bits of log-PCM is obtained. However, at this time, make the codebook sufficiently large (for example, (dimension 3 of vector
2 x size 256) x 2 = 16kword, where 16
If word = 2 bytes, then 32 kbytes) are required.

[Table 3]

Next, a second embodiment of the present invention will be described. In the second embodiment, the number N of codebooks in the first embodiment is reduced to 1 / N, and the codebook memory capacity is reduced to 1 / N.
The configuration is reduced to (N = 2). In order to realize this, the principle of upsampling / downsampling continuously shown in FIGS. 3 and 4 was applied. 3 and 4, fs represents the sampling frequency. FIG. 3A shows a signal in the frequency domain, which is divided into N (for example, N = 2) signals, that is, a low frequency domain (c) and a high frequency domain (e). (B) and (d) of FIG.
They are time waveforms corresponding to (c) and (e), respectively. When the time waveforms (b) and (d) are thinned out to 1 / N (N = 2) respectively (when the sampling frequency is downsampled to 1 / N), they become (f) and (h), and their corresponding spectra are ( g) and (i). Here, looking at (i), the spectrum of (e) is frequency-inverted, and (c)
Alternatively, it can be seen that the frequency band is converted to the same frequency band as (g). By performing such a conversion, the frequency bands become equal, so when vector quantization of (i) is performed,
It is possible to use the same codebook as that used in the vector quantization of (g). Therefore, the memory capacity required for the codebook is 1/2 because the number of vector dimensions is 1/2.
Can be reduced to

Next, in order to reproduce the spectrum of (a),
Up-sampling (interpolating the time points thinned by 0) is applied to (f) and (h) to obtain time waveforms (j) and (m) shown in FIG. The spectra corresponding to these are (k) and (n), respectively. If (k) and (n) are passed through a low-pass filter and a high-pass filter as shown in FIG. 4, respectively, they become (q) and (s), respectively.
(E) is reproduced. If (q) and (s) are combined, it becomes (t) and (a) can be reproduced. In the case of N> 2 as well, after dividing into equal intervals, frequency inversion is performed on even-numbered bands counting from the low frequency side, and odd-numbered bands are used as they are, so that one shared codebook is used. It is possible to perform vector quantization on each band. Therefore, the memory capacity required for one codebook can be reduced to 1 / N since the number of vector dimensions is 1 / N.

An embodiment in which this principle is realized by using DCT coefficients is shown in FIGS. FIG. 5 is a block diagram of the second embodiment of the speech coding apparatus, FIG. 6 is a block diagram of the second embodiment of the speech decoding apparatus, and FIGS. 7 and 8 show the processing steps of FIGS. The spectrum is shown. In the encoding device of FIG. 5, 6.4 kHz
The input voice signal (64 kbps) a2 which is z-sampled and 10-bit quantized is extracted by the long-term prediction analyzer 51 to output pitch information Pa2 for each frame (10 msec: 64 samples), and the pitch component is input from the input signal a2. Long-term prediction residual signal b which is the signal from which
2 is generated and output. The discrete cosine transform (DC
T) unit 52 transforms each frame (10 msec: 64 samples) into the frequency domain to generate a DCT which is a frequency component.
The coefficient c2 (64 coefficients) (FIG. 7A) is output. DC
The T coefficient c2 is equally divided into N (for example, N = 2) by the DCT coefficient divider 53, and a low frequency band (0 to 1.6 kHz) is obtained.
DCT coefficient d2 (32 coefficients) and high frequency band (1.6
.About.3.2 kHz) and the DCT coefficient e2 (32 coefficients). d2 (FIG. 7B) is directly input to the # 1 normalizer 54, and e2 is frequency-inverted by the frequency inverter 55 to be f2 (FIG. 7C). d2 and f2 are normalized by the maximum value of the DCT coefficient by the # 1 normalizer 54 and the #N normalizer (N = 2) 56, respectively, and the maximum value of the DCT coefficient Pb2 and Pc2 and the normalized DCT coefficient g
2, h2 are output. g2 and h2 are # 1 vector quantizer 57 and #N vector quantizer (N = 2) 5 respectively.
Vector quantization is carried out by one code book 58 shared with No. 9 and vector numbers Pd2 and Pe2 are outputted, and a coder 60 forms a digital signal string together with maximum values Pb2 and Pc2 of DCT coefficients and pitch information Pa2. After that, they are multiplexed and sent as an output signal i2 to the transmission line.

Next, in the decoding device of FIG. 6, the separation circuit 6 separates the digital sequence signal j2 received via the transmission line.
DCT vector numbers Pf2, Pg2 separated by 1
The DCT maximum values Ph2, Pi2 and the pitch information Pj2 are taken out. From the DCT vector numbers Pf2 and Pg2, vector dequantization is performed using one codebook 63 shared with the # 1 vector dequantizer 62 and #N vector dequantizer (N = 2) 64, respectively, and normalized. 0 to 1.6 kH
The DCT coefficient k2 of z and the DCT coefficient m2 of 1.6 to 3.2 kHz are reproduced. Here, the codebook 63 has the same contents as the codebook 58 of the encoding device. k2 and m2 are the # 1 denormalizer 65 and the maximum DCT coefficient Ph respectively.
2, #N denormalizer (N = 2) 66 and the DCT coefficient maximum value Pi2 are denormalized to obtain the DCT coefficient n of each frequency band.
2, p2 is played. n2 (FIG. 8 (d)) is directly input to the combiner 68, and p2 is frequency-inverted by the frequency inverter 67 to become q2 (FIG. 8 (e)) and then input. The combiner 68 combines the high frequency component n2 and the low frequency component q2 into a DCT coefficient r2 (FIG. 8 (f)).
Is played. The inverse discrete cosine transformer 69 transforms the DCT coefficient r2 into the time domain to reproduce the long-term prediction residual signal s2. In the long-term prediction synthesizer 70, the long-term prediction residual signal s2
Is added with pitch information Pj2 to decode and reproduce the audio signal t2. If this method is used, one codebook is provided for each of the transmitting side and the receiving side, and the memory capacity required for the codebook is 1 / N (N = 2), as compared with the first embodiment. Figure 1
1 is a structural example diagram of the second embodiment shown in FIG. 5 when N = 4. Reference numerals are the same as those in FIG.

For reference, DCT and IDCT will be described. These conversion formulas are as follows when the input signal X (n) is used. (1) In the case of DCT, desired DCT coefficient Xc (k)
Is

[Equation 1] However, N is the number of samples per block g (k) = 1 (k = 0) g (k) = √2 (k = 1, 2, ..., N-1) (2) In the case of IDCT, it is restored. The signal X (n) is

[Equation 2]

[0019]

As described in detail above, the following effects can be obtained by implementing the present invention. In the case of the first embodiment, the low frequency component and the high frequency component are separately vector-quantized to reduce the distortion in the high frequency region and improve the quality of reproduced voice. In the case of the second embodiment, the number of codebooks and the memory capacity are 1 / N, respectively, as compared with the first embodiment. Due to these effects, a half rate system (voice coding speed of 3.6 kbps or less,
Since it is possible to realize a voice coding communication system and its apparatus in which reproduced voice quality is log-PCM equivalent to 6 bits or more), there is an extremely great effect in practical use.

[Brief description of drawings]

FIG. 1 is a block diagram showing a speech coding apparatus according to a first embodiment of the present invention.

FIG. 2 is a block diagram showing a speech decoding apparatus according to the first embodiment of the present invention.

FIG. 3 is a diagram illustrating the principle of upsampling / downsampling.

FIG. 4 is a diagram illustrating the principle of upsampling / downsampling.

FIG. 5 is a block diagram showing a speech encoding apparatus according to a second embodiment of the present invention.

FIG. 6 is a block diagram showing a speech decoding apparatus according to a second embodiment of the present invention.

FIG. 7 is a spectrum (DC in the process of the second embodiment.
T coefficient).

FIG. 8 shows a spectrum (DC in the process of the second embodiment.
T coefficient).

FIG. 9 is a block diagram showing a configuration example of a conventional technique.

FIG. 10 is a block diagram showing a configuration example filed earlier.

FIG. 11 is a structural example diagram of the second embodiment of the present invention in FIG. 5 when N = 4.

[Explanation of symbols]

 11 long-term prediction analyzer 12 DCT device 13 adaptive thinning-out device 14 encoder 21 separation circuit 22 adaptive thinning-out decoder 23 IDCT device 24 long-term prediction synthesizer 31 long-term prediction analyzer 32 DCT device 33 normalizer 34 vector quantization 35 Codebook 41 Separation Circuit 42 Vector Dequantizer 43 Codebook 44 Inverse Normalizer 45 Inverse DCT Device 46 Long Term Prediction Synthesizer 51 Long Term Prediction Analyzer 52 DCT Unit 53 DCT Coefficient Splitter 54, 56 Normalizer 55 Frequency inverter 57,59 Vector quantizer 58 Codebook 60 Encoder 61 Separation circuit 62,64 Vector dequantizer 63 Codebook 65,66 Denormalizer 67 Frequency inverter 68 Combiner 69 IDCT device 70 Long term Prediction synthesizer 71 Long-term prediction analyzer 72 DCT device 73 DCT coefficient divider 74, 75 Normalizer 76, 8 vector quantizer 77, 79 codebook 80 encoder 91 separating circuit 94 vector inverse quantizer 93, 95 codebooks 96 and 97 inverse normalizer 98 combiner 99 IDCT unit 100 long-term prediction synthesizer

─────────────────────────────────────────────────── ───

[Procedure amendment]

[Submission date] February 16, 1993

[Procedure Amendment 1]

[Document name to be amended] Statement

[Correction target item name] 0004

[Correction method] Change

[Correction content]

On the receiving side, the digital signal train e is received, and the separation circuit 21 separates the DCT information Pc and the pitch information P.
Separate into d. In the adaptive thinning-out decoder 22, the DCT information P
The transmitted DCT coefficient is reproduced based on the DCT coefficient amplitude information and position information in c, and interpolation is performed by inserting 0 at the position of the DCT coefficient that has not been transmitted. The reconstructed DCT coefficient f is converted to an inverse discrete cosine converter (IDCT device)
It is converted to the time domain by 23 and the long-term prediction residual signal g is reproduced. The long-term prediction synthesizer 24 decodes and reproduces the audio signal h by adding pitch information Pd to the long-term prediction residual signal g. Conventional codec frame (30 mse
An example of bit allocation for each c) is shown in Table 1 below. Five synchronization bits are inserted at the beginning of each frame for frame synchronization. Converting the total in Table 1 per second, 135 bits / 30 msec = 4.5 kbps
Becomes

[Table 1] In mobile communication systems such as mobile phones and car phones, the transmission line conditions are harsh, unlike wired or fixed communication systems.
The bit error rate is always 0.1% to 1%, and it is not uncommon to reach about 10%. For this reason, the half-rate speech coding system needs to have a strong error correction function, and about 35% (about 2 kbps) or more of the total coding speed (5.6 kbps) is allocated to redundant bits for error correction. Can be said to be necessary. Therefore, when applied to the half-rate speech coding method, the coding rate of the speech data is about 3.6 kbps or less and high quality (log-P).
It is required to obtain reproduced sound of CM 6 bits or more).

[Procedure Amendment 2]

[Document name to be amended] Statement

[Correction target item name] 0005

[Correction method] Change

[Correction content]

[0005]

[SUMMARY OF THE INVENTION] reproduced sound quality in the above-described method, the speech encoding rate 4.6Kbps, log-
If only PCM 4 bits are obtained and the voice coding rate is further reduced to 3.6 kbps or less, the number of DCT coefficients that can be transmitted is reduced and distortion in the frequency domain is increased, so that the reproduced voice quality is further improved. to degrade. That is, in the conventional system, the reproduced voice quality cannot be lowered to 3.6 kbps or less with the log-PCM equivalent to 6 bits, and the performance (quality, error correction capability) required for the half rate voice coding system can be obtained. Can not meet.
Therefore, the present inventor introduced vector quantization to the above method in order to improve this problem, and is applicable to a half rate system (speech coding speed is 3.6 kbps or less, reproduced speech quality is log- PCM equivalent to 6 bits or more)
A voice coding method and its apparatus were previously proposed (Japanese Patent Application No.
-329782). The content is to reduce the coding speed while maintaining the reproduced voice quality higher than that of the conventional method by replacing the adaptive decimator in the above method with a vector quantizer, and FIG. 10 shows an embodiment thereof. It is a block diagram of a voice coding communication system. (A) is a speech encoding device, and (B) is a speech decoding device.

[Procedure 3]

[Document name to be amended] Statement

[Correction target item name] 0008

[Correction method] Change

[Correction content]

In the above-mentioned configuration proposed by the inventor of the present invention, the voice coding speed can be suppressed to a low level by introducing vector quantization, but there is a problem that the reproduced voice quality is not sufficiently maintained. is there. The reason for this is that when vector quantization is performed, the vector is selected with emphasis on the low frequency region (about 2 kHz or less) where power is concentrated, and the high frequency region with low power (about 2 kHz or more) is selected.
This is because the quantization distortion becomes large, and the distortion quality deteriorates in the high frequency region of the reproduced voice. Therefore, there is a problem that it is difficult to obtain a quality equivalent to or higher than 6 bits of log-PMC required in the half rate system.
The object of the present invention is to further improve the above problems of the prior application and to be applicable to a half rate system (speech coding rate is 3.6 kbps or less, and reproduced speech quality is log-P.
CM 6 bit or more) to provide a voice coding communication system and its device.

Claims (6)

[Claims] 1. A DCT coefficient obtained by, on the transmitting side, long-term prediction analysis of an input speech signal to obtain pitch information and a long-term prediction residual signal, and transforming the long-term prediction residual signal into a frequency domain by discrete cosine transform. Of the DCT coefficient maximum value and the DCT coefficient normalized by the maximum DCT coefficient value are vector-quantized using a codebook, and the DCT vector number of the approximate pattern read from the codebook, the pitch information, and the DCT coefficient The maximum value and the maximum value are encoded in the form of a digital signal sequence and transmitted to the transmission line, and on the receiving side, the digital signal sequence from the transmission line is separated to obtain the pitch information, the maximum DCT coefficient value and the DCT vector number. Of the DCT vector number, the vector corresponding to the DCT vector number is read from the codebook and inversely quantized to obtain the normalized DCT coefficient. After the DCT coefficient is reproduced by inverse normalization with the maximum value of the CT coefficient, the long-term prediction residual signal is reproduced by the inverse discrete cosine transform, and the pitch information is added to the long-term prediction residual signal to decode and reproduce the voice signal. In the voice coding communication system, the DCT coefficient obtained by the discrete cosine transform on the transmitting side is divided into a plurality of frequency regions at equal intervals, and the DCT coefficient maximum value of each of the divided frequency regions is output and the DCT coefficient is output. The DCT coefficients normalized by the maximum value are vector-quantized using a codebook, and the DCT vector number of the approximate pattern read from the codebook, the pitch information, and the maximum values of the plurality of DCT coefficients are stored in a digital signal sequence. The digital signal sequence from the transmission line is separated at the receiving side by being encoded into a form and transmitted to the transmission line. The pitch information, the plurality of DCT coefficient maximum values, and the plurality of DCT vector numbers, respectively, and the vectors corresponding to the plurality of DCT vector numbers are read from the codebook and inversely quantized to obtain the normalized DCT coefficients. A speech coding communication method characterized in that after reproducing and synthesizing, a long-term prediction residual signal is reproduced by inverse discrete cosine transform. 2. A maximum DCT coefficient of DCT coefficients obtained by performing long-term prediction analysis of an input speech signal to obtain pitch information and a long-term prediction residual signal, and transforming the long-term prediction residual signal into a frequency domain by discrete cosine transform. The value is output and D
The DCT coefficient normalized by the CT coefficient maximum value is vector-quantized using a codebook, and the DCT vector number of the approximate pattern read from the codebook, the pitch information, and the DCT coefficient maximum value are converted into a digital signal string form. A speech coding apparatus for coding and transmitting to a transmission line, a DCT coefficient divider for dividing a DCT coefficient obtained by the discrete cosine transform into a plurality of frequency regions at equal intervals, and outputting a plurality of DCT coefficients; The DCT coefficient maximum value is output from each of the plurality of DCT coefficients from the coefficient divider, and D
A plurality of normalizers that output DCT coefficients normalized with the maximum CT coefficient value, and the normalized DCT coefficients from the plurality of normalizers are vector-quantized using a codebook and read from the codebook. A plurality of vector quantizers that output DCT vector numbers of the approximate pattern, and code the pitch information, the plurality of DCT coefficient maximum values, and the plurality of DCT vector numbers in the form of a digital signal sequence and send it to the transmission line. A speech coding apparatus, comprising: 3. A signal encoded in the form of a digital signal sequence is received that includes pitch information of a voice signal, a plurality of maximum DCT coefficient values corresponding to divided frequency domains, and a plurality of DCT vector numbers. , A separation circuit for separating the digital signal sequence and extracting the pitch information, the plurality of DCT coefficient maximum values and the plurality of DCT vector numbers, and vectors respectively corresponding to the plurality of DCT vector numbers from respective codebooks. A plurality of vector dequantizers that output read dequantized and normalized DCT coefficients, respectively, and denormalize respective outputs of the plurality of vector dequantizers by the corresponding plurality of DCT coefficient maximum values. A plurality of denormalizers that output the normalized DCT coefficients for each frequency domain, and the outputs from the plurality of denormalizers are combined. A DCT coefficient to output a DCT coefficient, an inverse discrete cosine converter that reproduces a long-term prediction residual signal by performing an inverse discrete cosine transform on the output of the synthesizer, and the pitch information is added to the long-term prediction residual signal. And a long-term predictive synthesizer for decoding and reproducing a voice signal. 4. A DCT coefficient obtained by, on the transmitting side, long-term prediction analysis of an input speech signal to obtain pitch information and a long-term prediction residual signal, and transforming the long-term prediction residual signal into a frequency domain by discrete cosine transform. Of the DCT coefficient maximum value and the DCT coefficient normalized by the maximum DCT coefficient value are vector-quantized using a codebook, and the DCT vector number of the approximate pattern read from the codebook, the pitch information, and the DCT coefficient The maximum value and the maximum value are encoded in the form of a digital signal sequence and transmitted to the transmission line, and on the receiving side, the digital signal sequence from the transmission line is separated to obtain the pitch information, the maximum DCT coefficient value and the DCT vector number. Of the DCT vector number, the vector corresponding to the DCT vector number is read from the codebook and inversely quantized to obtain the normalized DCT coefficient. After the DCT coefficient is reproduced by inverse normalization with the maximum value of the CT coefficient, the long-term prediction residual signal is reproduced by the inverse discrete cosine transform, and the pitch information is added to the long-term prediction residual signal to decode and reproduce the voice signal. In the voice coding communication system, the DCT coefficient obtained by the discrete cosine transform on the transmitting side is divided into a plurality of frequency regions at equal intervals, and the DCT coefficient of the odd-numbered frequency region of the divided frequency regions is DCT. The maximum value of each coefficient is output, and the DCT coefficients normalized by the maximum value of the DCT coefficient are vector-quantized using one codebook, and the DCT vector numbers of the approximate patterns are read out from the one codebook, and the division is performed. Regarding the DCT coefficient of the even-numbered frequency domain of the frequency domain thus obtained, the frequency of the DCT coefficient of each domain is inverted. DC normalized respectively by said DCT coefficient maximum value and outputs the Chi DCT coefficients maximum respectively
The T coefficient is vector-quantized by sharing the one codebook, the DCT vector numbers of the approximate patterns are read out from the one codebook, and the pitch information, the plurality of DCT coefficient maximum values and the plurality of DCT vector numbers are read. Are encoded in the form of a digital signal sequence and transmitted to a transmission line, and at the reception side, the digital signal sequence from the transmission line is separated to obtain the pitch information, the plurality of DCT coefficient maximum values and the plurality of DCT coefficient maximum values. The DCT vector numbers are respectively taken out, and the vectors respectively corresponding to the plurality of DCT vector numbers are read out from one codebook having the same contents as the one codebook on the transmitting side and inversely quantized to obtain the normalized DCT coefficients. Reproduce and frequency-invert the DCT coefficient in the odd-numbered frequency domain and the DCT coefficient in the even-numbered frequency domain A voice coding communication system characterized in that the long-term prediction residual signal is reproduced by inverse discrete cosine transform after synthesizing with the DCT coefficient. 5. A maximum DCT coefficient of DCT coefficients obtained by performing long-term prediction analysis of an input speech signal to obtain pitch information and a long-term prediction residual signal, and transforming the long-term prediction residual signal into a frequency domain by discrete cosine transform. The value is output and D
The DCT coefficient normalized by the CT coefficient maximum value is vector-quantized using a codebook, and the DCT vector number of the approximate pattern read from the codebook, the pitch information, and the DCT coefficient maximum value are converted into a digital signal string form. In a speech coding apparatus for coding and transmitting to a transmission line, a DCT coefficient divider for dividing the DCT coefficient obtained by the discrete cosine transform into a plurality of equally spaced frequency regions and outputting the divided DCT coefficient, and a DCT coefficient divider A frequency inverter for inverting the frequency of each of the even-numbered frequency domain DCT coefficients of the divided DCT coefficients, and each of the odd-numbered frequency domain DCT coefficients of the divided DCT coefficients from the DCT coefficient divider. , The DCT coefficient maximum from each of the frequency-inverted even-numbered frequency domain DCT coefficients from the frequency inverter. A plurality of normalizers that respectively output a value and a DCT coefficient that is normalized by the maximum value of the DCT coefficient, and one codebook that commonly uses the normalized DCT coefficients from the plurality of normalizers A plurality of vector quantizers that perform vector quantization by using and read out and output DCT vector numbers of approximate patterns from the one codebook; and the pitch information, the plurality of DCT coefficient maximum values, and the plurality of DCT vector numbers. And a coder that encodes and are encoded in the form of a digital signal sequence and is sent to a transmission line. 6. A plurality of DCs corresponding to pitch information of a voice signal and a plurality of frequency regions divided at equal intervals.
A signal including a T coefficient maximum value and a plurality of DCT vector numbers and encoded in the form of a digital signal sequence is received, the digital signal sequence is separated, and the pitch information, the plurality of DCT coefficient maximum values, and A separation circuit for extracting the plurality of DCT vector numbers, and a plurality of vector dequantizations for reading out vectors corresponding to the plurality of DCT vector numbers from one codebook and dequantizing them to output normalized DCT coefficients, respectively. And a plurality of denormalizations for denormalizing the outputs of the plurality of vector dequantizers by the corresponding maximum values of the plurality of DCT coefficients and outputting normalized DCT coefficients for each frequency domain. A frequency inverter that inverts the DCT coefficients in the even-numbered frequency regions of the outputs from the plurality of denormalizers, respectively, A synthesizer that synthesizes the DCT coefficients in the odd-numbered frequency domain of the outputs from the plurality of denormalizers and the frequency-inverted DCT coefficients in the even-numbered frequency domain and outputs the DCT coefficients And an inverse discrete cosine transformer that reproduces a long-term prediction residual signal by performing an inverse discrete cosine transform on the output of the synthesizer, and a long-term decoder that decodes and reproduces a speech signal by adding the pitch information to the long-term prediction residual signal. A speech decoding apparatus having a predictive synthesizer.