JPH1097295A - Coding method and decoding method of acoustic signal - Google Patents

Abstract

PROBLEM TO BE SOLVED: To make decoding possible even when part of information is lost in a network where the usable bit rate is variable. SOLUTION: A first prediction coder 41 codes input voice to obtain the difference signal (43) between its decoding signal and the input voice. It gives gain to the noise waveform vector to supply as a driving vector to a synthesizing filter 14b, selects the noise vector and its gain to mininmize the distortion power of the synthesizing signal to the difference signal. The filtering coefficient of the synthesizing filter in the first coder 41 is set in the filter 14b.

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a cellular phone and a PHS
Method for encoding audio signals such as voice and music to obtain the most desirable quality according to the state of radio waves and the state of the network in a wireless transmission path such as the Internet or an ATM or a LAN such as the Internet, and a decoding method thereof It is about.

[0002]

2. Description of the Related Art In the fields of digital mobile communication and personal computer voice communication, various high-efficiency coding methods are used in order to make effective use of radio waves and networks. Methods for encoding telephone band voice include CELP (code-driven linear prediction), VSELP (vector-addition driven linear prediction),
CS-ACELP and the like are known. For information on each technology, see MRSchroeder and BSAtal: "Code-Excited Line
ar Prediction (CELP): High-quality Speech at Very L
ow Rates ", Proc. ICASSP'85, 25.1.1, pp.937-940, 19
85 (1) and IAGerson and MAJasiuk: "Vector
Sum Excited Linear Prediction (VSELP) Speech Codi
ng at 8kps ", proc. ICASSP'90, S9.3, pp.461-464, 19
90 (Reference 2) and A. Kataoka et al: "ITU-T 8-kbit / s St
andard Speech Coder for Personal Communication Ser
vices, "Int. Conf. on UniversalPersonal Communicat
ions, pp.818-822, 1995 (Reference 3).

A technique for encoding wideband speech by the CELP method in a lump for the entire area is described in A. Fuldseth et al: "Wid
eband speech coding at 16 kbit / s for a videophone
application, "speech communication, No.11, pp.139-
148, 1992 (Reference 4), C. Laflamme et al: "16 kbps w
ideband speech coding technique based on algebraic
CELP, "ICASSP 91, pp.17-20, 1991 (Reference 5), etc. Also, a coding method based on band division using a QMF filter is described in" Some Experiments of 7 kHz "by Rosario DROGO de IACOVO et al. Audio Coding at 16kb
it / s, "Proc. ICASSP 89, S4.19, pp. 192-195, 1989 (Reference 6).

[0004] In these CELP systems, as shown in FIG. 17, a filter coefficient determination unit 12 performs LPC analysis on input speech of a plurality of samples from an input terminal 11 by linear prediction analysis.
The coefficient is calculated, the LPC coefficient is converted into an LSP parameter, and the LSP parameter is
3 and the quantized LSP parameter is
1 / A (Z), which is converted to quantized LPC coefficients and set as a filter coefficient in the synthesis filter 14, is a transfer function of the synthesis filter 14. One pitch period waveform vector is selected from the plurality of pitch period waveform vectors of the adaptive codebook 15, one noise waveform vector is selected from the plurality of noise waveform vectors of the noise codebook 16, and one step is performed. The gain unit 17 multiplies the gain selected from the plurality of gains selected in the eleventh step with the selected pitch period waveform vector and the selected noise waveform vector by the gain unit 17. The sum is supplied to the synthesis filter 14 as a drive vector, and the sound is synthesized. From the noise waveform vector multiplied by the gain, an approximate gain for the noise waveform vector is predicted from a noise waveform vector selected in the past by the gain prediction unit 19, and the prediction gain is multiplied by the prediction gain unit 21. The subtraction unit 22 subtracts a synthesis vector from the synthesis filter 14 from the input audio signal to obtain distortion data.
3 for the weighted distortion data, and for the weighted distortion data, calculate the distortion power of the corresponding data.
Calculate with 4, and minimize this distortion power,
The distortion power calculator 24 selects each vector in both codebooks 15 and 16 and sets each gain of the gain unit 17. In the code output unit 25, a quantization filter coefficient (usually an LSP coefficient), a number and a gain of a selected one of the pitch period waveform vector and the noise waveform vector are represented by a code.
Is output as

[0005] The synthesis filter 14 outputs a so-called zero input response which is affected by past decoded speech even if its input is zero. Therefore, for example, assuming that a decoded waveform as shown in FIG. 18B is obtained from the synthesis filter 14 for an input waveform with auditory weights as shown in FIG. 18A, prior to encoding the current frame of the input waveform, the synthesis filter 14 With a zero drive vector, and as its output:
A zero input response waveform (FIG. 18C) is obtained, and the vector search of the codebooks 15, 16 is performed on the component (target waveform) shown in FIG. 18D, for example, by removing the zero input response waveform from the input waveform of the current frame. . The drive vector of the synthesis filter 14 at the time of encoding is temporarily stored in the adaptive codebook 15.

The decoding of the encoded code is performed as shown in FIG. A pitch period waveform vector is selected from the adaptive codebook 32 based on the pitch code in the input code from the input terminal 31, a noise waveform vector is selected from the noise codebook 33 based on the noise code in the input code, and the selected pitch period waveform is selected. The vector and the noise waveform vector are given respective gains according to the gain code in the input code by the gain unit 34, and are added by the adder 37 to be combined as a driving vector.
8. The gain-given noise waveform vector is input to a gain prediction unit 35, and then an approximate gain for the selected noise waveform vector is predicted, and the prediction gain is calculated by the prediction gain unit 36 as the selected noise waveform vector. Given to. The coefficient code in the input code is decoded by the filter coefficient decoding unit 39, and the decoded filter coefficient is set in the synthesis filter 38. In this case, the data may be decoded as LSP coefficients and converted into LCP coefficients to obtain filter coefficients. Synthetic filter 3
8 is temporarily stored in the adaptive codebook 32, and a decoded speech signal is obtained from the combined vector 38.

[0007] Each of the above-mentioned systems performs high-quality coding with low bits, but can operate only at a constant bit rate. In other words, for example, when transmitting voice at 16 kbit / s, if the line capacity of the network, for example, the number of users using one line is doubled, etc. In the conventional encoding method, the receiving side cannot receive all information, and thus cannot decode. Therefore, there is a problem that the conversation is interrupted or the line is disconnected. Similarly, ATM (asynchronous transfer mode)
In such a system that transmits data in packet units, some of the packets may be lost (discarded) depending on the state of the network (for example, congestion). Even in this case, as described above, there is a problem that decoding cannot be performed in the conventional encoding method because some information is lost. If the time (period) during which all information cannot be received is short, a method of interpolating with past decoded speech as a frame loss is also possible, but if the time during which reception is not possible is long, the quality is significantly degraded.

[0008] In recent years, as seen in the rapid spread of the Internet, network services have become important. In these LANs and ATMs, packet discards and cell losses occur depending on the state of the network. Therefore, the coding method used for them is desirably so-called scalability that can decode even a part of the transmission information.
As such a method, the gods “music and sound coding with a scalable hierarchical structure”, Sound Lecture Book 3-1-5, P
P.277-278, 1995.9 (Reference 6) have been proposed. However, this method requires sampling conversion for the output of the lower (core) encoder, and the upper encoder has information (bits) equal to or greater than that of the lower encoder. Is necessary.

[0009]

SUMMARY OF THE INVENTION An object of the present invention is to provide a network in which a usable bit rate changes or a network in which packet discards occur, even if a part of the information to be transmitted can be decoded even if the information is lost. Accordingly, it is an object of the present invention to provide an encoding method and a decoding method for reducing the deterioration of quality on the receiving side in accordance with the above.

[0010]

According to the encoding method (referred to as a separate encoding method) according to the first aspect of the present invention, as shown in FIG. 1A, an audio signal from an input terminal 11 is converted to a first encoder. 4
1 and outputs the encoded code to a code output unit 42, outputs a decoded signal of the encoded code and auxiliary information used for encoding from a first encoder 41, and outputs the decoded signal The difference signal of the audio signal is obtained by the subtractor 43, the difference signal is encoded by the second encoder 44 using the auxiliary information, and the encoded code is output to the code output unit 42. The code output unit 42 outputs both the encoded codes of the first and second encoders 41 and 44 or only the encoded code of the first encoder 41.

FIG. 1B shows a decoding method for an encoded code according to the separate encoding method. The encoded code of the first encoder 41 in the input code from the input terminal 31 is decoded by the first decoder 45, and the encoded code of the second encoder 44 is supplied to the second decoder 46. , The auxiliary information used for decoding from the first decoder 45
6 and a second decoder 46 decodes the input code using the auxiliary information. First and second decoder 4
The two decoded signals 5 and 46 are combined by an adder 47 and output as a decoded audio signal.

According to the encoding method of the invention of the second aspect (hereinafter referred to as the optimal encoding method), as shown in FIG. 2A, the audio signal from the input terminal 11 is encoded by the first encoder 41, A decoded signal of the encoded code is output, and the input audio signal is encoded using the auxiliary information used for encoding by the first encoder 41, and a decoded signal of the encoded code is output. , The two decoded signals of the second encoders 41 and 44 are combined by the adder 48, and the distortion of the combined decoded signal with respect to the input audio signal is obtained by the subtractor 22, and the distortion is obtained so that the obtained distortion power becomes minimum. The power calculator 24 controls the first and second encoders 41 and 44, and outputs the encoded codes of the first and second encoders 41 and 44 from the code output unit 25 when the minimum distortion power is obtained. Output.

FIG. 2B shows a decoding method for a coded code according to the optimal coding method. The first decoder 41 decodes the encoded code in the input code from the input terminal 31 by the first encoder 41, and decodes the second code in the second decoder 46.
The half-decoded signal is decoded by the encoder 44, the half-decoded signal is supplied to the first decoder 45, and the first decoder 45 outputs a decoded signal including the half-decoded signal. . That is, the second decoder 46 is used for decoding by the first decoder 45 and is also used for decoding by using auxiliary information.

[0014]

FIG. 3 shows an encoder to which the separation type encoding method is applied. In this example, the first encoder 41 shown in FIG. 17 is used, for example, and a second encoder 44 is added to the first encoder 41 to further enhance the performance. As described with reference to FIG. 1A, the difference between the decoded speech obtained by decoding the encoded code of the first encoder 41 and the input speech is obtained by the subtractor 43, and the difference signal is encoded by the second encoder 44. This second encoder 44 is also configured in substantially the same manner as the encoder shown in FIG. 17, and therefore, the portions corresponding to those in FIG. As can be understood from FIG. 3, the adaptive codebook 15, the adder 18, and the filter coefficient quantization unit 13 in FIG. 17 are omitted, that is, the system of the pitch period waveform vector is omitted, and the filter coefficients of the synthesis filter 14b are used. The filter coefficient of the synthesis filter in the first encoder 41 is used, and the filter coefficient of the auditory weighting filter 23b is the same as that of the first encoder 41. That is, here, the filter coefficients of the synthesis filter and the auditory weighting filter of the first encoder 41 are used as the auxiliary information. The noise waveform vector of the noise codebook 16b and the gain of the gain unit 17b are selected so that the distortion power of the decoded signal of the synthesis filter 14a with respect to the difference signal is minimized.

An application codebook is also provided in the second encoder 44,
The pitch period waveform vector is selected, and a gain is given to the selected pitch period waveform vector.
The pitch waveform waveform vector and its gain are also selected so that the distortion power is minimized by adding the noise waveform vector to which the gain is applied from b and adding the noise waveform vector to the synthesis filter 14b. In this case, the pitch information is received as auxiliary information from the first encoder 41, the vector of the adaptive codebook of the second encoder 44 is selected based on the auxiliary information, and the search is performed by the distortion power calculator 24b. However, it searches for the weights of the vectors before and after the pitch information that are close to the received pitch information. When the same pitch information is used, the second encoder 44 does not output a pitch code. If you want to search only the previous or next one,
What is necessary is just to output a pitch code of about 4 bits indicating the difference from any of the latter.

Further, the filter coefficient of the synthesis filter 14a is a filter coefficient which can be revised.
It is also possible to prepare an additional filter coefficient in addition to that of 1 so that the synthesis in the synthesis filter 14b is performed with high accuracy. In this case, the first encoder 41 in the filter coefficients that can be taken by the synthesis filter 14b of the second encoder 44
It is necessary to transmit an encoding coefficient code for the filter coefficients that cannot be obtained by the synthesis filter of the above, but the number of bits is small. Note that the difference signal waveform of the subtracter 43 seems to be significantly different from the input speech waveform, and therefore the filter coefficient of the synthesis filter
In this case, it is necessary to analyze and obtain the filter coefficients, and it is necessary to encode the filter coefficients and transmit them as coefficient codes, and the number of encoded bits increases. However, in this embodiment, it has been found that by setting the synthesis filter coefficient of the first encoder 41 to the synthesis filter 14b, encoding with sufficiently high quality is performed. Therefore, the number of coded bits does not need to be increased so much.

FIG. 4 shows an embodiment of the optimal coding method.
The first encoder 41 and the second encoder 44 are of the same method as the encoder of FIG. 17, except that the synthesis filter 14, the auditory weighting filter 23, the distortion power calculator 24, and the code output unit 25 A first encoding excitation source 48 and a second encoding excitation source 49 of each of the encoders 41 and 44 are separately provided, which are also used for the encoders 41 and 44. The parts corresponding to those in FIG. 17 are assigned the same numbers with the letter a for the first encoded excitation source 48,
The second encoded excitation source 49 is shown with a letter b. The first encoding excitation source 48 extends from the adaptive codebook 15 and the noise codebook 16 to the input side of the synthesis filter 14 in FIG.
The coding excitation source 49 is a case where only the system of the noise waveform vector is used. The pitch period waveform vector and the noise waveform vector provided with the gain from the first encoded excitation source 48 and the noise waveform vector provided with the gain from the second encoded excitation source 49 are added by the adder 51. The driving vector is supplied to the synthesis filter 14. The distortion power calculator 24 calculates each codebook 15 so that the distortion power is minimized.
a, 16a, and 15b, and the gains of the gain units 17a and 17b.

In this optimal coding method, the first encoder 41 also searches for a target obtained by subtracting the zero input response of the decoded speech of the entire encoder from the input speech. On the other hand, in the separation type encoding method, a target obtained by subtracting a zero input response of only the decoded voice of the first encoder 41 is targeted. Therefore, the optimal encoding method is the second encoder 44
The efficiency is good by the amount of the zero input response. However, the addition vector of the adder 18a is stored in the adaptive codebook 15a of the first encoding excitation source 48 so that the single operation of the first decoder corresponding to the first encoder can also be performed.

In the configuration shown in FIG. 4, the first encoder 41 and the second encoder 44 commonly use the synthesis filter 14 and the auditory weighting filter 23, that is, the first encoder 4
This means that the filter coefficients of the filters 14 and 23 used for coding 1 are used as auxiliary information in the encoding by the second encoder 44. Regarding the configuration of the second encoder 44, a configuration having an adaptive codebook like a conventional CELP is also possible. In that case, the pitch period T ₀ obtained by the first encoder 41 can be used for selecting an adaptive codebook, and it is also possible to search and use before and after T ₀ using several bits of information. It is. Further, as for the LSP, in addition to the configuration in which the one of the first encoder 41 is used as it is, a more accurate LSP can be used by adding several bits of information. That is, in this case, first, the encoding by the first encoding means 41 is performed using the first encoding excitation source 48, and then the output vector of the adder 18a at that time is supplied to the adder 51, and then the A search is performed for the 2-encoded excitation source 49, that is, encoded by the second encoding means 44,
At this time, the number of filter coefficients that can be used in the synthesis filter 14 is made larger than that in the first encoding unit 41.

In the optimal form shown in FIG. 4, the first and second
Detecting a zero input response at the time of encoding by the encoding means 41, 44, removing the zero input response from the input voice,
Since the search target waveform is used, that is, the zero input response component by the second encoding unit 44 is also taken into consideration,
Since the encoding is performed by the encoding unit 41, the performance of the first decoding unit 45 corresponding to the first encoding unit 41 on the decoding side is the first
Although it is inferior to the decoded signal for the signal encoded by the encoding means 41 alone, the performance of the entire encoder is improved.

In this optimal coding method, in order to increase the search efficiency, as shown in FIG. 5 with the same reference numerals assigned to parts corresponding to FIG. 4, the adaptive codebook 15a in the first coding excitation source 48 is used.
The pitch gain waveform vector obtained from the noise codebook 16a and the noise waveform vector obtained from the noise codebook 16a obtained by adding a prediction gain to the noise waveform vector obtained from the noise codebook 16b in the second encoding excitation source 49 are assigned the prediction gain. What is given is a straightening unit 5
1 and a search is performed. The obtained pitch period waveform vector in the first encoded excitation source 48, the noise waveform vector, and the noise waveform vector in the second encoded excitation source 49 are obtained by the gain unit 17. The gain may be given by three-dimensional vector quantization. By performing the orthogonalization in this way, it is possible to efficiently select, from the noise codebook 16b, noise waveform vectors different from the pitch period waveform vector and the noise waveform vector selected by the first encoding excitation source 48 as much as possible.

Since the second encoder 44 maintains scalability with the first encoder 41, the pitch component is
It is necessary to obtain from one drive vector. Therefore the second
The contribution of the noise waveform vector from the noise codebook 16b of the encoder 44 is not utilized in the pitch period waveform vector. Therefore, FIG. 6 shows an example of a method of obtaining the pitch period waveform vector from the drive vector of the entire encoder while maintaining scalability, with the same reference numerals given to portions corresponding to FIG. The first signal from the adder 18a in the first encoded excitation source 48
The drive vector in the encoder 41 is supplied to the adder 51 through the gain unit 17. An auxiliary adaptive codebook 15b is provided in the second encoding excitation source 49, and the codebook 15b has a drive vector as an entire encoder, that is, the synthesis filter 1b.
4 is temporarily stored. First encoded excitation source 4
8, the respective pitch period waveform vectors of the auxiliary adaptive codebook 15b before and after the pitch period T ₀ obtained before and after, and the determined drive vector of the first encoded excitation source 48,
That is, each correlation value with the output of the adder 18a (see FIG. 4) is calculated, and the pitch period waveform vector of the auxiliary adaptive codebook 15b corresponding to the one having the largest correlation value is selected.
At this time, it is also possible to normalize the waveform before and after the pitch period T ₀ and the respective powers of the drive vector of the first encoder 41.

From the pitch period waveform vector selected in this manner, the drive vector determined by the first encoding excitation source 48 is subtracted by the subtractor 52 to obtain an auxiliary pitch period waveform vector. That is, the sum of the drive vector of the first encoder 41 and the auxiliary pitch cycle waveform vector is regarded as the entire pitch cycle waveform vector. First encoded excitation source 48
, The auxiliary pitch period waveform vector from the subtractor 52, and the noise waveform vector selected from the noise codebook 16b and given the prediction gain,
After the gains are given in step 7, they are added in an adder 51 and supplied to the synthesis filter 14 as a drive vector.

Since the gain search of the gain unit 17 is performed in a closed loop, a predetermined very small gain is given to an auxiliary pitch period vector that need not be used. That is, as the pitch period waveform vector, only the one obtained by the first encoder 41 is sufficient, and a state where the addition of the auxiliary pitch period waveform vector degrades the quality during the search of the auxiliary adaptive codebook 15b. Or when the correlation value with the drive vector of the first encoder 41 is obtained and the correlation value is equal to or less than a predetermined value and the contribution of the auxiliary pitch period waveform vector is small, the auxiliary pitch period It is possible to give a predetermined very small gain to the waveform vector, prevent the quality from deteriorating, and reduce the number of bits of the coding gain code required by the gain unit 17.

Next, another embodiment will be described with reference to FIG. In this example, the decoded speech (auxiliary information) from the first encoder 41 in FIG. 3 and the output combined signal of the synthesis filter 14b of the second encoder 44 are combined by an adder 53, and the backward prediction LPC analysis unit is used. LPC analysis at 54 and synthesis filter 1
4B is different from the case of FIG. In this case, since the decoded signal of the first encoder 41 is obtained by the decoder, there is no need to output the encoded coefficient code of the filter coefficient of the backward prediction LPC analysis unit 54, and the order of analysis is not limited. A higher order (for example, the 50th order) can be used, and not only a speech signal but also a music signal can be encoded with high quality.

As described above, the method of obtaining the filter coefficients of the synthesis filter 14b of the second encoder 44 from the decoded signal can be applied to the optimal coding method. An example is shown in FIG. In this case, the synthesis filter 14 in FIG.
a and 14b, and the synthesis filter 14a is the first filter in FIG.
A vector obtained by adding the pitch period waveform vector to which the gain from the encoding excitation source 48 is given and the noise waveform vector by the adder 18a is given as a driving vector, and the output of the gain unit 17b of the second encoding excitation source 49 is used as the driving vector. To the synthesis filter 14b. These synthesis filters 14a,
14b are added by an adder 53 to be added to a subtractor 22.
And to the LPC analysis unit 54. LP
The input is LPC-analyzed by the C analysis unit 54 to calculate a filter coefficient, which is set in the synthesis filter 14b. in this case,
The zero input response waveforms of the synthesis filters 14a and 14b are synthesized, and the synthesized zero input response is removed from the input speech to be used as a target waveform for search.

When the use of the synthesis filters 14a and 14b for the optimal coding method is applied to the encoder shown in FIG. 4, the number of usable filter coefficients for the synthesis filter 14b is reduced. When the number of filters is larger than that of the synthesis filter 14a, encoding can be performed at the same time without performing encoding twice as described above. FIG.
As shown in (2), a search for the noise codebook 16b may be performed by orthogonalizing the noise waveform vector from the prediction gain unit 21b with respect to both input vectors of the adder 18a. Such an orthogonal search can be applied to the case where the filter coefficient of the synthesis filter 14a is used as the filter coefficient of the synthesis filter 14b in FIG.

Next, an embodiment of the decoding method according to the present invention will be described. First, an embodiment of a decoding method for a separation type encoding method will be described with reference to FIG. In this embodiment, the first
In the case where the same one as that shown in FIG. 19 is used as the decoder 45, the corresponding part is indicated by adding the letter “a” to the same number, and redundant description will be omitted. According to the present invention, a second decoder 46 is provided. This second decoder 46 has substantially the same configuration as that shown in FIG. 19, but omits the adaptive codebook and the filter coefficient decoding unit. For the use part of
Parts corresponding to those in FIG. 19 are indicated by the same numbers with the letter “b” added.

A pitch code, a noise code, a gain code, and a pitch code by the first encoder 41 in the input code from the terminal 31
The audio signal is decoded by the first decoder 45 using the coefficient code as described with reference to FIG. A noise waveform vector is extracted from the noise codebook 33b by the noise code of the second encoder 44 in the input code, a prediction gain is given by the prediction gain unit 36b, and the noise is calculated by the gain unit 34b. , And supplies it as a drive vector to the synthesis filter 38b. The filter coefficient set for the synthesis filter 38a is set for the synthesis filter 38b. This synthesis filter 38b
The decoded difference signal is obtained, and the decoded signal is combined with the decoded signal of the first decoder 45 by the adder 47 to obtain a decoded speech. Gain section 34
The gain for the next noise waveform vector is predicted by the gain prediction unit 35b based on the output of b, and is supplied to the prediction gain unit 36b.

According to this configuration, even if only the encoded code from the first encoder 41 is input, the first decoder 45
Is decrypted. Further, when both the encoded codes of the first and second encoders 41 and 44 are input, the difference signal that cannot be encoded by the first encoder 41 is decoded by the second decoder 46,
A high quality decoded signal is obtained. When a several-bit coefficient code is used in the output of the encoded code of the second encoder 44, the coefficient code of the second encoder 44 is included in the input code of the terminal 31, and this is included in the input code. Although not shown in the figure, the second coefficient code is decoded by the filter coefficient decoding unit and set in the synthesis filter 38b. When there is no second coefficient code, the synthesis filter 38
The filter coefficient of a is used for the synthesis filter 38b.

An adaptive codebook is used for the second encoder 44.
In this case, as shown by the broken line in FIG.
This selection depends on whether the encoding side and the first encoder 41
Outgoing pitch T₀At the same pitch as the adaptive code of the second encoder 44.
When selecting from the number book, use the pitch code in the input code.
Code, that is, the same peak as that taken out from the adaptive codebook 32a.
Pitch code from adaptive codebook 32b
The gain is given by the gain section 34b, and the gain is
Given a noise waveform vector from the given noise codebook 33b,
The sum is supplied to the synthesis filter 38b. First encoder
41 pitch T ₀And only those before and after the second encoder
When searching the 44 adaptive codebooks, only a few bits
The second pitch code is input to the terminal 31 and the second pitch code
Pitch periodic wave from adaptive codebook 32b by pitch code
Select a shape vector.

The optimal encoding method and the corresponding decoding method will be described with reference to FIG. The reference numerals in the figure are shown in FIG.
And the decoding by the first decoder 45 is the same as in the prior art. A second decoding excitation source 61 corresponding to the second encoding excitation source 49 in FIG. 4 is provided, and a noise waveform vector is obtained from the noise codebook 33b by an input code corresponding to the encoding noise code by the second encoder 44. Then, the noise waveform vector is provided with a prediction gain, and a gain corresponding to the second gain code is further provided to decode the drive vector. The drive vector is decoded by the adder 62 and the first decoding excitation source in the first decoder 45. The sum is added to the drive vector from 63 and supplied to the synthesis filter 38. As is apparent from the figure, the first decoding excitation source 63 is a method of obtaining a drive vector from the adaptive codebook 32 and the noise codebook 33 in FIG.

Also in this case, even if only the encoded code from the first encoder 41 is input, it can be correctly decoded, and if the encoded code from the second encoder 44 is also input. , The quality of the decoded signal is further improved. FIG. 11 shows an embodiment of a decoding method for the encoding method shown in FIG.
This will be described with reference to FIG. In FIG. 11, the same reference numerals are given to parts corresponding to the previous figures. The first decoding excitation source 63 in the first decoder 45 adds a pitch period waveform vector and a noise waveform vector by an adder 37a before gain is given by the gain unit 34a unlike the one shown in FIG. I do. The second decoding excitation source 61 of the second decoder 46 has the configuration shown in FIG.
0, the adaptive codebook 32b is added thereto, and the subtractor 64 subtracts the addition vector of the adder 37a from the pitch period waveform vector extracted from the adaptive codebook 32b to obtain an auxiliary pitch period waveform vector. , The gain unit 34 gives each gain according to the gain code to the addition vector from the adder 37a and the noise waveform vector from the noise codebook 33b, adds the gain in the adder 62, and drives the synthesis filter 38. Supply as a vector. In this case, the extraction of the pitch period waveform vector from the adaptive codebook 32b is the same as the extraction of the pitch period waveform vector from the adaptive codebook 32b in FIG.

Also in this case, it can be easily understood that even if only the encoded code by the first encoder 41 is input, it can be correctly decoded. Second encoder 44
In this case, not only the difference of the waveform of the decoded signal with respect to the input voice is decoded by the first decoder 45 with the second decoder 46, but also the pitch component is insufficient for the input speech. As a result, higher quality decoded speech can be obtained.

Also in the configuration of FIGS. 10 and 11, in the description of FIG. 9, when the second code is included in the coded code by the second coder 44, the second The code is decoded into filter coefficients and set in the synthesis filter 38. A decoding method corresponding to the encoding method shown in FIGS. 7 and 8 will be described with reference to FIG. The configuration of the first decoder 45 in this case is the first decoder 45 in FIG.
, But the second decoder 46 uses the synthesis filter 38a as the filter coefficient of the synthesis filter 38b.
The LPC analysis unit 66 performs LPC analysis on the decoded speech signal obtained by adding the respective combined signals of the synthesis filters 38a and 38b by the adder 47 without using that of the synthesis filter 38a, 38b, and sets the filter coefficient in the synthesis filter 38b. This is different from the second decoder 46 in FIG.

In this case as well, even if only the encoded code from the first encoder 41 is input, decoding can be performed correctly. If the encoded code from the second encoder 44 is also input, higher quality is obtained. A decoded signal can be obtained. The order of the synthesis filter 38b can be higher than that of the synthesis filter 38a, and the music signal can be decoded with high quality.

Each of the configurations shown in FIGS. 1 to 12 is a functional configuration, and the first encoder 41 and the second encoder 44 are not necessarily physically separate from each other. For example, one DSP (digital Signal processor). This means that the first decoder 45 and the second decoder 46
Have a similar relationship. Next, an example in which the present invention is applied to actual wideband speech coding will be described. ITU-T for telephone band
This paper describes the standard 8kbit / s G.729 and the scalable 16kbit / s wideband speech coding method. As shown in FIG. 13, the input voice signal is filtered by the QMF filter 71 into the telephone band (0.3
3.43.4 kHz), and are separated into low-frequency components and higher-frequency components, and are coded and decoded by the low-frequency coding / decoding unit 72 and the high-frequency coding / decoding unit 73, respectively. The images are combined by the filter 74. The present invention is applied to the low band in this case. That is, 8 kbit / s G.729 (CS-ACELP), which is an ITU standard, is used as the first encoder 41, and a 4 kbit / s encoder is configured as the second encoder 44. G.729 is used as the first encoder 41, and encoding is performed by using an LSP coefficient, a separation type using an auditory weighting filter, an optimal type, and a method shown in FIG. 6 (referred to as an auxiliary pitch type). The vessel was constructed. That is, one using G.729 as the first encoder 41 in FIG. 3, one using the G.729 excitation source as the first encoding excitation source in FIG. 4, and one in FIG. The first coded excitation source used was a G.729 excitation source. However, although not shown in FIGS. 4 and 6, the drive vector of the second encoder 44 is orthogonalized to the drive vector of the first encoder 41 as shown in FIG. ing. The quality evaluation results of the telephone band and the wideband decoded speech are shown below.

Performance Evaluation of Low-Frequency Portion The performance of each configuration was evaluated by a segmental SNR and a cepstrum distortion (CD) value. FIG. 14 shows the evaluation results of the G.729 part (first encoder 41) and the entire low-frequency part (first encoder 41 + second encoder 44). From Fig. 14, G.72
It can be seen that the CD value of 9 is not good in any format with respect to the CD value of the entire low-frequency part, because G.729 is designed for voice in the 3.4 kHz band. In the entire low-frequency section, each configuration compensates for the degradation of this 3.4 kHz or higher band, and the CD
The value has improved. Also, compared to G.729 (first encoder 41), the separation type configuration is about 0.4 dB, the optimal type configuration is about 2.2 dB,
The auxiliary pitch type configuration has 2.6 dB, each having improved SNR. The performance degradation of the G.729 part due to the difference in the search in consideration of the zero input response of the auxiliary noise codebook (second encoder 44) is only about 0.6 dB.

Quality Evaluation An opinion test was performed to evaluate the quality of the encoder. The evaluation was performed separately for broadband speech and telephone band speech.
Evaluation voice is Japanese (10 men and women each), evaluation is 1-5
At the stage, the subjects are 24 ordinary people. P.34 for wideband audio
(1) 3.4k for telephone band voice
A sound with a flat Hz-band characteristic and a low-frequency sound divided by QMF were used. In addition, G.722, which is an ITU standard for wideband coding, and NMRU (Mod
ulated Noise Reference Unit) signal was used. NMRU
The signal is a signal to which white noise is added so that the amplitude changes in proportion to the amplitude of the original signal. The quality evaluation result is MINRU
It is converted to an equivalent Q value using a signal.

FIG. 15 shows the results of evaluating the quality of wideband speech. A indicates a separated type configuration, B indicates an optimal type configuration, and C indicates an auxiliary pitch type configuration. From this figure, it can be seen that the quality of each configuration is equal to or better than 56 kbit / s G.722. In particular, the one using the auxiliary pitch type configuration has almost the same quality as 64 kbit / s G.722. This is probably because the decoded speech becomes clearer by reinforcing the pitch.

FIG. 16 shows the quality evaluation result of the low-frequency (telephone band) voice. A indicates a flat sound (G.729), B indicates a band-divided sound (G.729), C indicates an optimum type configuration, D indicates an auxiliary pitch type configuration, and E indicates an auxiliary pitch type configuration + low-pass filter.
From this figure, the quality of G.729 is about 1 dB worse for band-divided sound than for flat sound due to the effect of 3.4 kHz or more. The quality of the optimal G.729 part is not different from that of G.729, but the quality of the auxiliary pitch type G.729 part is degraded.
When this sound is processed by a low-pass filter having a cutoff frequency of 3.4 kHz, noise of 3.4 kHz or more is removed, and the quality is improved. It is considered that the reason why the deterioration of the G.729 part of the optimal type and the auxiliary pitch type is small is that the second encoder 44 mainly improves the band of 3.4 kHz or more.

In the above description, one second encoder 44 and one second decoder 46 are used. However, the second encoder 44 and the second decoder 46 are improved by the same method. , A second encoder and a second decoder may be further added. The auditory weighting filter processing may be performed on the input voice of the subtractor 22, and the synthesis filter may also serve as the auditory weighting filter processing.

[0043]

As described above, according to the present invention, decoding can be performed even from a part of transmission bits, so that communication can be performed irrespective of the state of the network.

[Brief description of the drawings]

1A is a block diagram showing a functional configuration example (separated type) of an encoding method of the present invention in which a first encoder 41 and a second encoder 44 operate completely independently, and FIG. 1B is a decoding method thereof; FIG. 2 is a block diagram showing an example of the functional configuration of the embodiment.

FIG. 2A is a block diagram showing an example of a functional configuration (optimal form) of an encoding method according to the present invention in which a first encoder 41 and a second encoder 44 are integrally encoded; FIG. 3 is a block diagram illustrating a functional configuration example.

FIG. 3 is a block diagram showing an example of a functional configuration of a separation type encoding method according to the present invention.

FIG. 4 is a block diagram showing an example of a functional configuration of an optimal coding method according to the present invention.

FIG. 5 is a block diagram showing a functional configuration of another embodiment of the optimal coding method according to the present invention.

FIG. 6 is a block diagram showing a functional configuration of an example in which a noise codebook and an auxiliary adaptive codebook are provided as the second encoder 44 in the optimal coding method of the present invention.

FIG. 7 shows an example of obtaining a combined filter coefficient of the second encoder by backward linear prediction from the sum of the decoded signal of the first encoder and the decoded signal of the second encoder in the separation type encoding method of the present invention. FIG. 2 is a block diagram showing a functional configuration.

FIG. 8 is a block diagram showing a first embodiment of the optimal coding method according to the present invention;
FIG. 10 is a block diagram showing a functional configuration of an example of obtaining a synthesis filter coefficient of the second encoder by backward linear prediction from a sum of both decoded signals of the second encoder.

FIG. 9 is a block diagram showing an example of a functional configuration of a separation type decoding method according to the present invention.

FIG. 10 is a block diagram illustrating a functional configuration example of an optimal decoding method according to the present invention.

FIG. 11 is a block diagram showing a functional configuration example of an auxiliary pitch type decoding method according to the present invention.

FIG. 12 is a block diagram showing a functional configuration of an example of obtaining a synthesis filter coefficient of the second decoder by a sum of both decoded signals of the first and second decoders in the decoding method according to the present invention.

FIG. 13 is a block diagram showing a configuration of subband wideband coding using a QMF filter.

FIG. 14 is a diagram showing a performance evaluation result of a low-frequency part.

FIG. 15 is a diagram showing a quality evaluation result of each configuration for wideband speech.

FIG. 16 is a diagram showing a quality evaluation result of each configuration of a low-frequency part (telephone band voice).

FIG. 17 is a block diagram showing a functional configuration of a conventional predictive encoder.

FIG. 18 is a waveform chart showing a flow of creating a target vector when searching for a predictive encoder.

FIG. 19 is a block diagram showing a functional configuration of a conventional decoder.

 ────────────────────────────────────────────────── ─── Continuing on the front page (72) Inventor Shinji Hayashi 3-19-2 Nishi-Shinjuku, Shinjuku-ku, Tokyo Nippon Telegraph and Telephone Corporation

Claims (21)

[Claims] 1. An input audio signal is input to a first encoding unit,
In a method of encoding using one or more second encoding means, the first encoding means analyzes a spectral envelope of the input audio signal to extract a filter coefficient, and quantizes the filter coefficient. The quantization filter coefficients are set in the synthesis filter, and gains are given to the pitch period waveform vector selected from the first adaptive codebook and the noise waveform vector selected from the first noise codebook, and the synthesis filter is used as a drive vector. To obtain a synthesized audio signal from the synthesis filter, and select the pitch period waveform vector and the noise waveform vector so as to minimize distortion of the synthesized audio signal with respect to the input audio signal, and select the gain. The second encoding means encodes the input audio signal using the auxiliary information used for encoding by the first encoding means. Acoustic signal encoding method and performing. 2. The encoding method according to claim 1, wherein a difference signal between a signal obtained by decoding the encoded code by the first encoding unit and the input audio signal is encoded by the second encoding unit. A sound signal encoding method characterized by the above-mentioned. 3. The encoding method according to claim 2, wherein said second encoding means sets a filter coefficient set in a synthesis filter of said first encoding means in a second synthesis filter as said auxiliary information. Then, a gain is given to the noise waveform vector selected from the second noise codebook, and the gain is given to the second synthesis filter as a drive vector, and the distortion of the synthesized signal from the synthesis filter with respect to the difference signal is minimized. (2) A sound signal encoding method characterized by selecting from a noise codebook and selecting and encoding a gain to be applied to the selected noise waveform vector. 4. The encoding method according to claim 3, wherein a gain is given to the pitch period waveform vector selected from the second adaptive codebook by the encoding by the second encoding means, and the gain is given to the second synthesis filter. It is added to the drive vector and supplied.
The selection of the second adaptive codebook together with the selection of the second noise codebook, the selection of the gain to be applied to the selected pitch period waveform vector, and the extraction of the waveform from the second adaptive codebook as the auxiliary information. A sound signal encoding method is performed based on the pitch period encoded by the first encoding means. 5. The encoding method according to claim 1, wherein said first encoding means and said second
An audio signal encoding method, characterized in that encoding is performed by an encoding unit so as to be optimal. 6. The encoding method according to claim 5, wherein the second encoding means applies a gain to a noise waveform vector selected from a second noise codebook and adds the gain to the driving vector for the synthesis filter. And the selection of the first adaptive codebook and the first noise codebook, the selection of the second noise codebook, and the selection of the gain together with the selection of the second noise codebook.
A sound signal encoding method comprising selecting a gain given to a noise waveform vector selected from a noise codebook. 7. The encoding method according to claim 6, wherein the noise waveform vector selected from the second noise codebook by the pitch period waveform vector selected by the first encoding means and the selected noise waveform vector. Is selected by orthogonalizing, and a gain is given to the selected noise vector to be added to a driving vector. 8. The encoding method according to claim 6, wherein the second encoding means selects an auxiliary pitch period waveform vector selected from a second adaptive codebook and gives a gain to the driving vector. In addition, an audio signal encoding method characterized in that an auxiliary pitch period waveform vector is cut out from the second adaptive codebook based on the pitch obtained by the first encoding means as the auxiliary information. 9. The encoding method according to claim 8, wherein an auxiliary pitch period waveform vector of the second adaptive codebook is created from past drive vectors of the synthesis filter,
Acoustic signal coding characterized by subtracting the pitch period waveform vector and the noise waveform vector selected by the first encoding means from the selected auxiliary pitch period waveform vector to provide a gain and adding the gain to the driving vector of the synthesis filter. Method. 10. The encoding method according to claim 1, wherein said second encoding means applies a gain to a noise waveform vector selected from a second noise codebook to generate a second drive signal as a second drive signal.
The second synthesis filter supplies the input audio signal to the second synthesis filter so as to minimize the distortion of the output synthesized signal of the second synthesis filter with respect to the difference signal from the decoded signal of the first encoding means.
The noise codebook is selected, the gain given to the selected noise waveform vector is selected, and the spectral envelope of the sum of the decoded signal and the output synthesized signal of the second synthesis filter is analyzed to obtain a filter coefficient. A sound signal encoding method, wherein the method is set in a second synthesis filter. 11. The encoding method according to claim 1, wherein an audio weighting filter is used for each of the first encoding means and the second encoding means. An acoustic signal encoding method, wherein a filter coefficient of an auditory weighting filter is set as an auditory weighting filter coefficient of the second encoding means. 12. The encoding method according to claim 1, wherein the number of filter coefficients that can be set in the synthesis filter of the first encoding unit is smaller than the number of filter coefficients that can be set in the second encoding unit. An acoustic signal encoding method, wherein the number of filter coefficients that can be set for a synthesis filter is large. 13. A method for decoding an audio signal by a first decoding means and a second decoding means based on an input code, wherein the first decoding means converts a filter coefficient by a coefficient code in the input code. Decode, set the decoded filter coefficient to the synthesis filter, select the pitch period waveform vector from the adaptive codebook with the pitch code in the input code,
A noise waveform vector is selected from the noise codebook with the noise code in the input code, and each of the pitch period waveform vector and the noise waveform vector is given a gain according to the gain code being input to the synthesis filter as a drive vector. Giving a synthesized sound signal, and decoding by the second decoding means using the one for the second decoding means in the input code and the auxiliary information from the first decoding means. Audio signal decoding method. 14. The decoding method according to claim 13, wherein the sound signal decoded by the first decoding means and the sound signal decoded by the second decoding means are added and decoded. Audio signal decoding method. 15. The decoding method according to claim 14, wherein the second decoding means performs a second decoding based on a second noise code in the input code.
A noise waveform vector is selected from the noise codebook, a gain corresponding to the second gain code in the input code is given to the noise waveform vector, and the noise waveform vector is supplied as a drive vector to the second synthesis filter. An audio signal decoding method, comprising: setting a decoded filter coefficient; adding a synthesized signal of a second synthesis filter and a synthesized audio signal of a first synthesis filter to obtain a decoded audio signal. 16. The decoding method according to claim 15, wherein said second decoding means selects an auxiliary pitch period waveform vector from a second adaptive codebook according to a second pitch code in the input code, and the auxiliary pitch period is selected. A sound signal decoding method, characterized in that a gain according to a second gain code in an input code is given to a waveform vector and added to the drive vector of the second synthesis filter. 17. The decoding method according to claim 13, wherein the second decoding means selects a noise waveform vector from a second noise codebook with a second noise code in the input code,
A sound signal decoding method, characterized in that a gain according to a second gain code in an input code is given to the noise waveform vector and added to a driving vector of the synthesis filter. 18. The decoding method according to claim 17, wherein the second decoding means converts the noise waveform vector selected from the second noise codebook into the selected pitch period waveform vector and the first noise vector. A sound signal decoding method comprising orthogonalizing a noise waveform vector selected from a codebook, giving a gain to the orthogonalized vector, and adding the gain to the drive vector. 19. The decoding method according to claim 13, wherein said selected pitch period waveform vector and said selected noise waveform vector are added, and said second decoding means uses a second pitch code in an input code. An auxiliary pitch period waveform vector is selected from the second adaptive codebook, the above addition vector is subtracted from the auxiliary pitch period waveform vector, and a noise waveform vector is selected from the second noise codebook based on the second noise code in the input code. A sound signal decoding method, wherein each of the gain, the subtraction result, and the addition vector is provided with each gain corresponding to a gain code in an input code and added to obtain the drive vector. 20. The decoding method according to claim 13, wherein the second decoding means selects a noise waveform vector from the second noise codebook in the input code, and substitutes the noise waveform vector with the second noise code in the input code. A gain corresponding to the gain code is given and supplied as a drive vector to the second synthesis filter, and the synthesized signal of the second synthesis filter and the synthesized audio signal of the synthesis filter are synthesized to obtain a decoded audio signal. An acoustic signal decoding method characterized by analyzing and extracting a spectral envelope of an acoustic signal, calculating a filter coefficient, and setting the filter coefficient in the second synthesis filter. 21. The decoding method according to claim 14, wherein a synthesis filter of a second decoding unit can be set based on the number of filter coefficients that can be set in a synthesis filter of the first decoding unit. An acoustic signal decoding method characterized in that more filter coefficients are used.