JP2005114814A - Method, device, and program for speech encoding and decoding, and recording medium where same is recorded - Google Patents

Abstract

<P>PROBLEM TO BE SOLVED: To obtain interconnectivity between wide-band encoding of an intermediate band and wider and telephone band encoding of a narrow band. <P>SOLUTION: This device performs a band dividing process of dividing a wide-band speech signal into a telephone-band speech signal of a narrow band, a high-frequency speech signal of an intermediate band, and a high-frequency speech signal of a wide band, a frequency characteristic compensating process of changing frequency characteristics of the band-divided telephone-band speech signal into high-frequency emphasized characteristics, an encoding process of encoding the telephone-band speech signal after the frequency characteristic compensating process, the high-frequency speech signal of the intermediate band, and the high-frequency speech signal of the wide band respectively, and a packet structuring process of sending out respective encoded data as packet signals. <P>COPYRIGHT: (C)2005,JPO&NCIPI

Description

  The present invention relates to an audio encoding / decoding method, an audio encoding / decoding device, an audio encoding / decoding program, which transmits audio in a packet communication network such as the Internet, stores and reproduces an audio signal, The present invention relates to a recording medium on which these are recorded.

Conventional wideband speech coding does not divide a speech signal up to a narrow band telephone band (0 to 3.4 kHz), a medium band (0 to 7 kHz), and a wide band (0 to 15 kHz), but a single band. Encoding methods are often used. Examples of encoding voice signals in a telephone band include nonlinear waveform compression encoding (μ-law / A-law PCM) used in G.711 (Non-Patent Document 1) as waveform encoding, and G.726 (Non-patent). The differential prediction waveform compression coding waveform method (ADPCM) used for literature 2) etc. is mentioned. For the middle band (0 to 7 kHz), a CELP system such as AMR-WB can be used.
ITU-T (Telecommunication Standardization Sector, International Telecommunication Union), Geneva, Switzerland.ITU-T G.711-Pulse code modulation (PCM) of voice frequencies, Nov. 1988. ITU-T (Telecommunication Standardization Sector, International Telecommunication Union), Geneva, Switzerland.ITU-T G.726-40, 32, 24, 16 kbit / s adaptive, differential pulse code modulation (ADPCM), Dec. 1990.

  Wideband coding over the middle band and narrowband telephone band coding, which is widely used, are not interconnected. In addition, since the telephone band coding method cited as the prior art is widely used, it is not easy to change the entire network uniformly to wideband coding. An object of the present invention is to enable playback on a conventional playback device having only a narrow-band playback capability while achieving a high-quality playback function in a wide band.

According to the present invention, a band division process for dividing a wideband audio signal into a narrowband telephone band audio signal, a medium band high band side audio signal, and a wideband high band side audio signal, and a band divisional telephone band Frequency characteristics compensation processing that changes the frequency characteristics of the audio signal to high frequency emphasis characteristics, and each of the phone band audio signal, medium frequency high frequency audio signal, and broadband high frequency audio signal that have been subjected to frequency characteristic compensation processing A speech encoding method is proposed, which includes an encoding process for encoding a packet and a packet construction process for transmitting each encoded data as a packet signal.
The present invention further provides packet decomposition processing for decomposing an incoming packet signal into wideband high-frequency side voice encoded data, medium band high-frequency side voice encoded data, telephone band voice encoded data, and these voice codes. A speech decoding method is proposed that includes a decoding process for decoding each of the encoded data into a voice signal, and a synthesis process for synthesizing the decoded voice signal and reproducing a wideband signal.

  According to the present invention, the voice signal in the telephone band is divided from the wideband signal by the band division, and the frequency characteristic of the divided voice signal is changed to the high frequency emphasis characteristic. It closely matches the characteristics, maintains compatibility with the conventional telephone network, and can reproduce voice at the telephone level. Furthermore, in addition to the voice signal of the telephone band, the high band side of the medium band voice signal and the high band side of the wide band voice signal are also encoded and transmitted as a packet, so even a terminal having a medium band and wide band reproduction function is insufficient. And high-fidelity broadband sound can be reproduced.

In addition, according to the present invention, when connecting to a voice receiving mechanism having only a conventional (G.711 or G.726) voice receiving mechanism, the packet constructing unit uses a medium band and a broadband high band side. If only the telephone voice data is transferred as a payload to the network API without packing the data in a packet, the interconnection can be performed without extra encoding.
Here, since the present invention is a kind of scalable coding, even if an audio signal data packet input and encoded at 32 kHz sampling given in FIG. 1 shown in FIG. In addition, if the pseudo wideband high-frequency code data is ignored and not processed, there is an advantage that sound can be reproduced without any problem.

The band division filter can be composed of a band division filter bank using a wavelet transform technique. A wideband input audio signal is input to the band division filter bank, and a pseudo telephone band audio signal, a pseudo midband high frequency side audio signal, and a pseudo wideband high frequency side audio signal are separated and output on the output side.
The pseudo mid-band high-frequency audio signal and the pseudo wide-band high-frequency audio signal are encoded as they are by an encoding method suitable for each band, and input to the packet construction unit.
The telephone band voice signal is compensated by the high frequency emphasis characteristic in the frequency characteristic compensator so as to match the transmission characteristic in the conventional telephone network, and the telephone band signal compensated for this frequency characteristic is encoded and input to the packet construction unit. .

The packet building unit packetizes the telephone band voice encoded data, the medium band high band side code data, and the wide band high band side encoded data input from each encoding unit, and sends them to the packet communication network.
On the receiving side, the received packet is decomposed into pseudo wideband high band side code data, pseudo medium band high band side code data, and pseudo telephone band band code data by the packet decomposing unit, and these are decoded into voice signals by the decoding unit, The decoded audio signal is synthesized by a resynthesis filter bank to reproduce a wideband audio signal.
When using only the telephone band, the transmitting side only needs to encode the telephone band voice signal and transmit it as a packet. When transmitting the middle band voice signal, the telephone band and the middle band high band voice signal are transmitted. The encoded data may be transmitted as a packet. To transmit a wideband signal, a telephone band voice signal, a middle band high band voice signal, and a wide band high band voice signal are encoded, What is necessary is just to transmit these code | symbol data as a packet. In any state during transmission of a broadband audio signal or transmission of an intermediate band audio signal, a terminal having only a telephone band reproduction function can reproduce audio using only the data of the telephone band.

  Furthermore, when a bi-orthogonal filter bank is used as the band dividing means, basically a noise signal (that is, a difference signal between the decoded signal and the original signal) superimposed on the signal obtained through the encoding and decoding processes. Are orthogonalized and added at the time of synthesis, noise is not added and emphasized when passing through the re-synthesis filter, which is convenient for processing such as encoding. That is, the influence of the noise signal can be reduced by using a bi-orthogonal transform or a transform method based on the normal overlap and add filter bank. Here, the band division method includes a conventionally used QMF (Quadrature Mirror Filter). Strictly speaking, such conversion is called “frame”, and it has been proved mathematically that it has such characteristics (Reference 1).

As filter banks satisfying such conditions, discrete Fourier transform (DFT), discrete wavelet transform (DWT) (reference document 2), and the like can be considered, but in this embodiment, fast discrete wavelet transform (FWT) (reference document 3). The implementation when using) is described. Here, the reason why the wavelet transform is used is that, for human hearing, frequency resolution is more important than time resolution in the low frequency band, and conversely time resolution is more important than frequency resolution in the high frequency band. Based on that. In other words, for example, when encoding a sound signal of 32 kHz sampling, 15 kHz is defined as a telephone band (0 to 4 kHz band), a medium band high band (4 to 8 kHz), and a remaining wide band high band (8 to 15 kHz). The division is easier to apply to existing coding schemes, and also matches human auditory characteristics. In addition, the wavelet expansion coefficient and scaling coefficient obtained by wavelet decomposition have the advantage that they can be processed at high speed since the low-pass filter processing and sample thinning processing necessary for sampling conversion are performed simultaneously.
[Reference 1]
S. Mallat. A Wavelet Tour of Signal Processing. Academic Press, San Diego, 2nd edition, 1999.
[Reference 2]
G. Strang and T. Nguyen. Wavelets and Filter Banks. Wellesley‐Ca‐
Mbridge Press, Wallesley, MA, 1996.
[Reference 3]
I. Daubechies. Ten Lectures on Wavelets. SIAM, Philadelphia, PN, 1992.
There are a number of bases (strictly, “frame bases”) such as Mayers, Daubechies, Mexican Hat, etc., but Symmmlet is used, and those having a length of about 14 taps are used. Since this wavelet base has a property that the coefficients have a relatively symmetrical shape in the time domain, the validity of the localization in the time domain is high, and an efficient coding result can be expected. When the number of taps is increased, there is a merit that the overlapping of the divided bands is reduced and noise due to aliasing can be reduced when only the low frequency band is reproduced, but this is not preferable because the amount of calculation increases.

Input that does not conform to the IRS characteristics (high frequency emphasis characteristics) that are expected to be input even when existing high-quality encoding methods such as G.711 and G.726 are used for encoding telephone band audio signals. If a signal (for example, a signal with a flat frequency characteristic) is given, the original performance cannot be exhibited and noise can be perceived remarkably. If frequency characteristic compensation is used, high-quality encoding processing is possible. Thus, the interconnectivity can be maintained.
However, only the compensation of the frequency characteristics can reproduce only a signal without a sense of presence when reproduced as wideband sound because the low frequency is missing. Here, if a mechanism for quantizing the difference signal from the original sound is added to the second stage, the quality of both the reproduction of only the telephone band and the reproduction of the wide band can be kept high.

  Further, in the case of an audio signal, the frequency characteristic is compensated for because the power of the signal is concentrated in a low frequency range, so that much power is lost due to the frequency characteristic correction, and the clarity may be lowered. In order to avoid the deterioration of the sound quality, it is only necessary to adjust the gain. On the other hand, when reproducing in a wide band, the power increases and a mismatch occurs. Here, in the receiving mechanism, it is possible to avoid the state by multiplying the correction gain.

FIG. 1 shows a block diagram on the transmission side when the speech encoding method of the present invention is implemented.
In this embodiment, a wide-band input audio signal is input. This input audio signal may be sampled at 32 kHz, and when a signal having a higher sampling frequency is input, it is necessary to downsample to 32 kHz in advance.
Since the processing system described in this embodiment is intended for real-time processing, processing is usually performed for each short-time processing frame of about 5 ms to 50 ms. Of course, if the operation is performed off-line, the same result can be obtained even if the signal is stored and batch-processed as long as the memory permits.

  First, the wideband input voice signal is divided by using the band division filter bank 10 to obtain a pseudo telephone band voice signal, a pseudo middle band high band voice signal, and a pseudo wide band high band voice signal. For this band division, a filter bank which is bi-orthogonal transformation is used. Examples of the filter bank that satisfies such conditions include a filter bank using a general QMF (Quadrature Mirror Filter) and a fast wavelet transform (FWT). An implementation example of the band division filter bank 10 using FWT is shown in FIG. Here, the case where it comprises with the scaling coefficient analysis filter 10-1, the wavelet expansion coefficient analysis filter 10-2, the scaling coefficient analysis filter 10-3, and the wavelet expansion coefficient analysis filter 10-4 is shown. Since the audio signal of each band obtained here is a scaling coefficient and a wavelet expansion coefficient obtained by wavelet analysis, it is called a pseudo telephone band audio signal, a pseudo medium band high frequency side audio signal, or a pseudo wide band high frequency side audio signal.

The pseudo telephone band audio signal (low frequency audio signal) thus obtained is frequency characteristic compensator 11 which has a flat frequency characteristic and a frequency characteristic of a signal output from a conventional telephone represented by an IRS characteristic or the like. It is changed to the thing according to. The IRS characteristic (reference document 4) here has a gentle high-frequency emphasis characteristic as shown in FIG. This frequency characteristic operation is implemented as a 10-20 tap FIR filter. Further, the frequency characteristic according to IRS can obtain the same audible effect by substituting a high-pass filter as the pseudo IRS characteristic as shown in FIG. In this case, it can be implemented with a filter having a shorter tap length.
[Reference 4]
ITU-T (Telecommunication Standardization Sector, International Telecommunication Union), Geneva, Switzerland.ITU-T P.830 Anner D-modifired IRS send and receive characteristics, Feb. 1996.
Next, the gain of the signal that has passed through the frequency characteristic compensation unit 11 is adjusted by the gain adjustment unit 12. In the case of voice communication, the gain adjustment unit 12 multiplies the sample by a constant g that takes a value in the range of 1.0 to 4.0.

The pseudo telephone band voice signal thus obtained is encoded by the telephone band encoding unit 13. At this time, the telephone band encoding unit 13 may use the conventional waveform encoding method, but may transmit the digital data as it is without performing compression encoding.
At the same time, the phase operation of the pseudo telephone band audio signal is performed by the phase delay compensation unit 14. The phase delay compensation unit 14 has a function of buffering the original voice signal and delaying the telephone band signal. The phase delay compensation unit 14 can also be implemented as an all-pass filter that operates only the phase. This phase lag is the same as the phase lag obtained via the frequency characteristic compensator 11, but when the frequency characteristic compensator 11 is implemented with an FIR filter, it is half the tap length of the filter. Just delay.

Next, the original pseudo telephone band voice signal whose phase delay has been compensated is subtracted by the subtractor 15 from the signal whose frequency characteristic has been changed, passed to the telephone band compensation encoder 16 and encoded.
The middle band high band side (4 to 8 kHz band) and wide band high band side (8 to 16 kHz band) signals are respectively passed to the encoding units 17 and 18 to obtain encoded data. Usually, the conventional telephone band encoding method does not use a wide band, so these encoding methods may be unique, or the same effect can be obtained by using the conventional telephone band encoding.
Thereafter, these encoded data are transferred to the packet construction unit 19, and are transferred to the network API as payload data of the IP packet and transmitted to the IP network. At this time, the priority of the code data of each band can be calculated, and packetized for each priority and transmitted. Note that the packet output method with priority will be described in detail later.

FIG. 5 shows a mechanism block diagram on the receiving side when the speech decoding method of the present invention is implemented. Here, the pseudo telephone band compensation code data, the pseudo telephone band code data, the pseudo medium band high band side code data, and the pseudo wide band high band side code data are divided by the packet decomposition unit 50 from the payload portion of the packet received from the network API. , Respectively, to the corresponding decoding units 51, 52, 53, 54.
Each signal output from the decoding units 53 and 54 other than the telephone band decoding unit is subjected to phase lag compensation by the phase lag compensation unit 55. As described on the transmission side, this phase delay compensation may be implemented as a buffer that delays the signal sample, or may be implemented as an all-pass filter that operates only on the phase. Here, the same phase operation as the phase lag compensation unit 14 of the transmission mechanism shown in FIG.

A gain compensation unit 56 multiplies the signal obtained from the telephone band decoding unit 52 by a correction gain. This correction gain uses the reciprocal (1 / g) of the gain g used in the transmission mechanism (FIG. 1). This signal is summed with the telephone band compensation signal whose phase delay has been compensated by the adder 57 to obtain a pseudo telephone band reproduction signal.
Thus, the pseudo telephone band decoded signal, the pseudo medium band high band side decoded signal, and the pseudo wide band wide side decoded signal are passed to the recombination filter bank 58 and re-synthesized into the wide band voice. An implementation example of the recombination filter bank by inverse FWT at this time is shown in FIG. In the example shown in FIG. 6, the re-synthesis filter bank 58 is configured by two scaling coefficient synthesis filters 58-1 and 58-3 and two wavelet expansion coefficient synthesis filters 58-2 and 58-4.

The wideband audio signal obtained through the resynthesis filter bank 58 is upsampled to a desired sampling frequency greater than 32 kHz. Upsampling is not necessary when using 32 kHz sampling.
When connecting with a voice transmission mechanism having only a conventional (G.711 or G.726) voice reception mechanism, the data in the middle band and the wideband high band is filled with 0 and the resynthesis filter bank 58 is connected. Interconnection is possible by passing the. Further, the same effect can be obtained by moving only the telephone band decoding unit 52 and up-sampling the reproduced audio signal obtained by sampling at 8 kHz to 32 kHz, and can be realized with a small amount of calculation.

  In the above-described embodiment, an example was given in which a high-frequency voice on a wide band at 32 kHz sampling was first coded. However, in a voice conference system using a loudspeaker system, a high-frequency voice on the middle band at 16 kHz sampling is sufficient. There is a case. When applied to such an input audio signal, the band division filter bank 10 can be implemented by dividing the telephone band audio signal and the mid-band high-frequency side audio signal into two so that the telephone band can be output. A mechanism for encoding and decoding the signal is not required. A block diagram of this implementation is shown in FIG. FIG. 8 shows a block diagram of a receiving mechanism that forms a pair. Also, examples of implementation of the band division filter bank 10 and the recombination filter bank 58 in that case are shown in FIGS. 9 and 10, respectively.

FIG. 11 shows an embodiment in which the noise generated in the first-stage telephone band encoding unit 13 is compensated by the second-stage telephone band encoding unit in the mechanism of the first embodiment.
Here, the signal obtained by decoding the code data obtained from the first-stage telephone band encoder 13 by the telephone band decoding unit 111 is multiplied by the correction gain by the gain compensation unit 112, and the phase delay compensation unit 14 is used. A difference signal from the phase-compensated pseudo telephone band signal is obtained by the subtractor 113 and provided to the telephone band compensation encoder 114. As the correction gain of the gain compensator 112, the reciprocal (1 / g) of the gain g is used as in the receiving side of the first embodiment.

The effect obtained here is that if the first-stage telephone band encoding unit 13 cannot obtain a sufficient SNR, it is encoded in a form including noise generated in the first stage in the compensation encoding. The noise can be canceled on the side, and the quality of the reproduced voice can be improved.
The same voice receiving mechanism as that shown in FIG. 5 is used as a pair.
The speech encoding apparatus and speech decoding apparatus of the present invention described above can also be functioned by a computer. In that case, a program for causing the computer to execute the steps of the above-described method of the present invention is installed in a computer functioning as the device from a recording medium such as a CD-ROM or a magnetic disk, or downloaded via a communication line. And execute it.

Hereinafter, a method for outputting a packet signal with priority will be described.
Technology that automatically controls the adjustment of the quality (application quality) of each media so as to obtain the maximum effect using a given network and system resources (voice, video, etc.), so-called Internet As a QoS (Quality of Service) control technique, DiffServ (reference documents 5 and 6) is attracting attention. This method is a mechanism in which packets entering the network are classified in advance according to priority, and packets with low priority are discarded at each network load during congestion. In order to use this mechanism in voice communication, an effective network can be used if the priority is calculated for each voice processing unit (that is, packet).

In general, VAD (Voice Activity Detection) (reference document 7) used for audio signal transmission mainly focuses on the control of the presence / absence of an audio signal. There wasn't. In other words, in the conventional voice signal packet, there are only two stages in which the voice section is set to a high priority and the silent section is set to a low priority.
[Reference 5]
IETF‐RFC2474: Definition of the Differentiated Services Field (DS Field) in the IPv4 and IPv6 Headers, 1998.
[Reference 6]
IETF‐RFC2475: An architecture for Differentiated Services, 1998.
[Reference 7]
3GPP: ETSI TS 146 032, “Digital cellular telecommunications system (Phase 2+); Voice Activity Detection (VAD), 2002
In an integrated network environment of voice, music, video, and data that will become popular in the future, “data communication”, which has a high peak transmission rate, and delay time directly leads to quality degradation even if the transmission rate is low. "Voice, music, video communication" needs to be mixed efficiently. In this case, it is desired to generate packets with smooth staircase priorities so that voice, music, video, etc. can be transmitted efficiently without degrading quality.

In order to meet such demands, the present applicant has created a packet with smooth priority in Japanese Patent Application No. 2003-63445, and can efficiently transmit music and video without degrading quality. A packet output method is proposed.
According to the method of the prior application, a digital signal is divided for each frame, the digital signal for each divided frame is encoded, and the feature amount based on the encoding or / and the feature amount of the digital signal are obtained as explanatory variables. The index value is obtained by linearly combining a plurality of the explanatory variables, the priority is obtained by quantizing the index value, and the priority and the encoding code are output as a packet.

  More preferably, the digital signal is divided into a plurality of bands for each frame, the digital signal for each frame is encoded, the feature amount based on the encoding for each frame and band, and / or the feature amount of each digital signal for each band Are calculated as explanatory variables, and these explanatory variables are combined at one time to obtain one index value, the index value is quantized to obtain a priority, and the priority and coding code for each frame and band are combined. And output as a packet including at least one of the sets. When the frame disappears, the degree of influence of the explanatory variable on the subjective evaluation value is determined in association with each coefficient of the linear combination.

  FIG. 12 shows a first embodiment of the prior application. An audio digital signal (hereinafter referred to as an audio signal) s [n] in which each sample from the input terminal 1211 is a digital value is in units of 5 milliseconds to 20 milliseconds as in this type of general encoder. The frame is divided by the frame dividing unit 1212, and the audio signal s [n] (n is a discrete time) is collectively encoded by the encoding unit 1213 every N samples. For example, in the case of a sound signal of 32 kHz sampling, N = 160 samples to N = 640 samples. Further, the priority determination unit 1214 determines the priority of the packet for each frame. A specific example of the priority determination unit 1214 is shown in FIG. In this example, the feature amount of the audio signal s [n] of the frame is generated as the explanatory variables x1 [i], x2 [i], and x3 [i] by the plurality of explanatory variable generation units 1311, 1312, and 1313, respectively. The explanatory variable of the i-th processing frame is set to xj [i], the audio signal s [n] of the frame is input, and the absolute power is calculated by the following equation (1) by the explanatory variable generation unit 1311.

x1 [i] = (1 / N) Σ _{n = 1} ^N s [Ni + n] ² (1)
Alternatively, as shown in the following equation (2), x1 [i] is obtained as a logarithmic expression of absolute power.
x1 [i] = log ₁₀ ((1 / N) Σ _{n = 1} ^N s [Ni + n] ² ) (2)
The explanatory variable generation unit 1312 receives the explanatory variable x1 [i] from the explanatory variable generation unit 1311 and the explanatory variable x1 [i-1] of the previous frame (i-1), and the power of the previous frame of the power of the current frame. The explanatory variable x2 [i] is output by calculating the ratio to the following equation (3).

x2 [i] = x1 [i] / (x1 [i-1]) (3)
The explanatory variable x1 [i−1] of the previous frame is stored in the previous frame buffer 1312a, the calculation of Expression (3) is performed by the calculation unit 1312b, and the explanatory variable x1 “i” of the current frame is stored in the previous frame buffer 1312a. Update the explanatory variables to be retained.
Further, the explanatory variable generation unit 1313 receives the audio signal s “n”, calculates the maximum value (periodicity) of the autocorrelation function (ρ [n]) by the following equation (4), and calculates the explanatory variable x3 [i ].
x3 [i] = max (ρ _i [k]) (4)
The autocorrelation function ρ [n] normalized here is calculated using the following equation (5).

ρ _i [k] = Σ _{n = 0} ^N (s [Ni + n]) (s [Ni + n + k]) /
Σ _{n = 0} ^N (s [Ni + n]) ² (5)
k is 1, 2,..., and the maximum value of k is approximately equivalent to the pitch period of the audio signal s [n]. At this time, a better result can be obtained by up-sampling the autocorrelation function, that is, by interpolating and calculating a more accurate value.
The index value y [i] is obtained by linearly combining the obtained explanatory variables x1 [i], x2 [i], and x3 [i] by the index value calculation unit 1314. That is, for example, the following equations (6) and (7) are calculated.

y [i] = α0 + Σ _{j = 1} ³ αjxj [i] ^ (6)
xj [i] ^ is obtained by normalizing the average probability distribution of the explanatory variable xj to 0 and the variance to 1, that is, the following equation (7).
xj [i] ^ = (xj [i] -xj ′) / γj (7)
xj ′ and γj are the average value and standard deviation of the explanatory variable xj, respectively.
These linear combination coefficients α0 and α1 use partial regression coefficient values optimized in advance using multiple regression analysis (see, for example, Tadashi Okuno et al .: Multivariate analysis method (revised version), Nikka Giren, 1981). For example, when a MOS value subjectively evaluated by a listener when one packet (frame) is lost is y [i] ′, this index value y [i] calculated by y [i] ′ and Expression (6) is used. The coefficient αj is obtained using the method of least squares so that the error from i] is minimized. α0 is an average value of MOS values 1 to 5. Here, MOS value 1 corresponds to “very bad” and MOS value 5 corresponds to “very good”.

  Since the coefficients α0 to α3 are determined in this way, if the absolute value of αj is large, the explanatory variable (feature value) greatly affects the subjective evaluation quality when the packet (frame) is lost, and the absolute value of αj is small. For example, the explanatory variable (feature amount) has a relatively small influence on the subjective evaluation quality when the packet (frame) is lost. That is, αj is determined so that the coefficient αj increases as the degree of influence on the subjective evaluation quality increases. Since the index value y [i] is obtained by linearly combining a plurality of explanatory variables (feature quantities) x1 [i] to x3 [i] using coefficients α1 to α3, one explanatory variable (feature quantity). As a result, the degree of influence is more correctly shown than the degree of influence of the packet (frame) loss on the subjective evaluation quality. A frame that greatly affects the subjective evaluation quality, in this case, a speech, and those that are audibly important tend to have a small index value y [i], and those that are not important tend to have a large index value.

In the index value calculation unit 1314 in FIG. 13, the explanatory variables x1 to x3 are normalized by the normalization units 1314a1 to 1314a3, respectively, and the normalized explanatory variables x1 to x3 are coefficients α1 to α3 by the normalization units 1314b1 to 1314b3, respectively. Are multiplied by the addition units 1314c1 and 1314c2, and an index value y [i] is output.
The index value y [i] obtained in this way is scalar quantized by the quantizing unit 1315, and a discrete value, for example, a priority p [i] of any one of 0, 1,. The That is, generally, a packet with a small index value is mapped to a high priority packet, and a packet with a large index value is mapped to a low priority packet. The mapping can be expressed by the following function.

p [i] = f (y [i]) (8)
The mapping function f (y) used at this time may use scalar quantization that maps the packet to the total priority step number. The quantization threshold at this time includes a method of dividing the index value y “i” with an equal probability, and a method of equally dividing the range of the index value y [i].
Each value of the linear combination coefficient is, for example, α1 = −0.37, α2 = −0.1, and α3 = −0.2. The larger the absolute value, the greater the influence on the subjective evaluation quality. If a frame with a large absolute power disappears in these three explanatory variables (features), the influence on the subjective evaluation quality is the greatest. This means that if a frame with a large level in the audio signal disappears, it has a significant effect. I mean. When a frame having a large autocorrelation function is lost, the influence on the subjective evaluation quality is the next largest. This means that even if the absolute power of the frame is small, if there is an audio signal, the maximum value of the autocorrelation function becomes large. Thus, the loss of the frame including the audio signal has a relatively large effect even if the absolute power is small. It means that.

Therefore, if at least x1 and x3 are used among the explanatory variables x1 to x3, and further x2 is used, the priority p "i" having a smoother step is obtained, and the influence of the loss of the frame on the subjective evaluation quality is affected. Become more accurate. In the above specific example, the coefficient is a negative value, and the smaller the evaluation value y [i], the higher the priority p [i].
The priority p [i] determined for each frame in this way is output as a packet P "i" of the frame i from the encoding unit 1213 and assembled as a packet by the packet sending unit 1215 (FIG. 12). Is done.

Second Embodiment This second embodiment is applied to a case where a wideband audio signal is divided into a plurality of bands and encoded.
As shown in FIG. 14, the wideband audio signal is divided into frames for each predetermined section by the frame dividing unit 1212, and is divided into F multiple bands by using the bandpass filter by the band dividing unit 1411. For example, if the audio signal s [n] is 16 kHz sampling, the band is divided into upper and lower 4 kHz bands (F = 2), and if it is 32 kHz sampling, F = 3 and 0 to 4 kHz band and 4 kHz to 4 kHz band. It may be divided by wavelets such as 8 kHz band and 8 kHz to 16 kHz band, or may be divided into 4 kHz bands at equal intervals with F = 4. Each band-divided audio signal is encoded for each fixed time length (frame) by an individual encoder. FIG. 15 shows a divided image of the voice block (packet) at this time. In the example of FIG. 15, F = 3, and each band signal is made into a block (packet) for each frame, and three blocks (packets) are generated for each frame.

In the example shown in FIG. 14, the audio signal is divided into two upper and lower bands, and the low-frequency audio signal s1 [n] and the high-frequency audio signal s2 [n] separated are the low-frequency encoding unit 1412L and the high-frequency audio signal s2 [n], respectively. It is encoded by the area encoding unit 1412H. The low frequency audio signal s1 [n] and the high frequency audio signal s2 [n] are respectively input to the low frequency priority determining unit 1413L and the high frequency priority determining unit 1413H, and the packet priority for each frame is determined. .
A specific example of the low frequency priority determination unit 1413L is shown in FIG. In FIG. 16, the same number is attached | subjected to the function structure part corresponding to FIG. 13, and the code | symbol L is attached | subjected to the number. The explanatory variable generator 1311L calculates the absolute power or the logarithm of the low frequency audio signal s1 [n] in the same manner as in the equation (1) or (2) to generate the explanatory variable x1 [1, i]. The explanatory variable generation unit 1312L calculates the previous frame power ratio in the same manner as Expression (3) to generate the explanatory variable x2 [1, i]. Further, the explanatory variable generation unit 1313L calculates the maximum value of the autocorrelation function in the same manner as Expressions (4) and (5) to generate the explanatory variable x3 [1, i].

Furthermore, in this embodiment, the absolute power x1 [f, i] of this band and the absolute power of other bands are input by the explanatory variable generation unit 1316L, and the ratio of the absolute power of this band to the total power is expressed by the following equation (9). And is output as an explanatory variable x4 [f, i].
x4 [f, i] = x1 [f, i] / Σ _{f = 1} ^F x1 [f, i] (9)
In the example of FIG. 16, since F = 2, x4 [1, i] = x1 [1, i] / (x1 [1] is obtained by x1 [1, i] in the low band and x1 [2, i] in the high band. , I] + x1 [2, i])
Is calculated.

The index value calculation unit 1314L linearly combines the explanatory variables x1 [1, i], x2 [1, i], x3 [1, i], x4 [1, i], and the index value y [1, i ] Is calculated.
y [1, i] = α0 + Σ _{j = 1} ⁴ αjxj [1, i] ^
xj [1, i] ^ = (xj [1, i] −xj [1] ′) / γj [1]
The index value y [1, i] is quantized by the quantization unit 1315L, and the priority p [1, i] = f ₁ (y [1, i]) is output.
Similarly, the high band priority determination unit 1413H uses the index value y [2, i] = α0 + Σ _{j = 1} ⁴ αjxj [2, i] ^
xj [2, i] ^ = (xj [2, i] −xj [2] ′) / γj [2]
And the priority p [2, i] = f ₂ (y [2, i]) is output. The packet sending unit 1215 uses the encoded code P [1, i] and the priority p [1, i] from the low frequency encoding unit 1412L as one packet, and the encoded code P [ 2, i] and priority p [2, i] are transmitted as one packet.

In general, when the band is divided into F, the index value y [f, i] of the f-th band is y [f, i] = α0 + Σ _{j = 1} ⁴ αjxj [f, i] ^
xj [f, i] ^ = (xj [f, i] −xj [f] ′) / γj [f]
The priority p [f, i] is obtained by f _f (y [f, i]).
The coefficient α4 is −0.43, for example, and is larger than α1. In other words, a large ratio of the divided band to the total band power means that there is a voice signal component with a large power in that portion, and it is preferable that the priority is the highest among α1 to α4. ing.

Third Embodiment The third embodiment is an embodiment in which the present invention is applied to a case where speech is encoded using a single-band quality scalable encoder, that is, an encoder capable of performing various quality encodings. It is a form. In this case, the audio block (packet) divided image is shown in parentheses in FIG. 15 to indicate the relationship between the quality q and the frame. FIG. 17 shows a functional configuration applied to a case where an audio signal is encoded in a general fixed processing time unit with a Q = 2 stage configuration.
The audio signal s [n] is divided in frame units by the frame dividing unit 1212, encoded for each frame by the first-stage encoding unit 1412-1, and the priority p in the first-stage priority determining unit 1413-1. [1, i] is determined. The encoded code P [1, i] from the first stage encoding unit 1412-1 is decoded by the first stage decoding unit 1711-1, and this decoded signal is subtracted from the audio signal by the subtraction unit 1712-1. Thus, the first stage residual signal (encoding error signal) e1 [n] is generated. The residual signal is encoded for each frame by the second-stage encoding unit 1412-2, and the priority p2 [2, i] is determined by the second-stage priority determining unit 1413-2. The encoded code P [2, i] from the second-stage encoding unit 1412-2 is decoded by the second-stage decoding unit 1711-2, and the decoded signal is the first-stage residual signal e1 [ n] is subtracted by subtracting unit 1712-2 to generate second stage residual signal e2 [n].

A specific example of the first-stage priority determination unit 1413-1 is shown in FIG. Similar to the priority determination unit 1214 shown in FIG. 13, the explanatory variable x1 [1, i] of the absolute power, the explanatory variable x2 [1, i] of the previous frame power ratio, and the explanatory variable x3 of the autocorrelation function maximum value. [1, i] are generated by the explanatory variable generation units 1311, 1312, and 1313, respectively.
In the third embodiment, the explanatory variable generator 1317 further generates the quality of the code P [1, i], for example, the noise ratio for the signal, as the explanatory variable x5 [1, i]. That is, the signal power calculation unit ^{_{1317a S = Σ n = 1 N}} s [Ni + n] 2 are calculated, also being calculated by the noise calculation unit 1317b is ^{_{E = Σ n = 1 N e1}} [Ni + n] 2, these ratios logarithm Log ₁₀ E / S is calculated by the logarithmic division unit 1317c, and the result is output as the explanatory variable x5 [1, i].

These four explanatory variables are linearly combined by the index calculation unit 1314 to calculate the index value y [1, i]. For example, as in the previous case, the normalization unit 1414aj (j = 1,..., 4) normalizes the explanatory variables xj [1, i], respectively, and the normalized value xj [1, i] ^ is linearly combined y [1, i] = α0 + Σ _{j = 1} ⁴ αjxj [1, i] ^, xj [1, i] ^ = (xj [1, i] −xj [1] ′) γj. The index value y [1, i] is quantized by the quantization unit 1315, and the first-stage priority p [1, i] is output.
The second-stage priority p [2, i] is obtained in the same manner. In this case, as shown in parentheses in FIG. 18, the second-stage residual signal e2 [n] is input instead of the first-stage residual signal e1 [n], and the same applies to these signals. And the second stage priority p [2, i] is output.

In the packet sending unit 1215 (FIG. 17), the first-stage code P [1, i] and the priority p [1, i] are one packet, and the second-stage code P [2, i] and the priority p [2 , I] are output as one packet.
This explanatory variable x5 [q, i] (q = 1, 2,..., Q) can be said to be a feature quantity based on encoding. A general formula for calculating this is as follows.
x5 [q, i] = log ₁₀ (Σ _{n = 1} ^N eq [Ni + n] ² / Σ _{n = 1} ^N s [Ni + n] ² )
In this case, the linear combination coefficient α5 can be about −0.1. A packet having a large q is necessary for reproduction of a high-quality signal. However, when traffic is congested, the meaning content of transmitted information is more demanding than the quality. q, i] becomes small, and α5 is relatively small, so that the priority is not so much involved.

Fourth Embodiment In the case of a general scalable multiband encoder, in addition to the explanatory variables x1 [i], x2 [i], and x3 [i] listed in the first embodiment, they are listed in the second embodiment. The index value y [f, q, i] is calculated using both the explanatory variable x4 [f, i] and the explanatory variable x5 [q, i] mentioned in the third embodiment. FIG. 19 shows a divided image of the voice block (packet) at this time.
That is, in the case of so-called scalable coding in which audio signals of various qualities having combinations of various sampling frequencies and various sample quantization accuracy (number of amplitude bits) are shown, FIG. 19 shows three sampling frequencies and quantization accuracy ( (Quality) is also divided into three stages, the frequency band is divided into three bands of f = 1, f = 2, and f = 3, and the amplitude bit length is divided into three areas of q = 1, q = 2, and q = 3. As a signal block (packet) in the tertiary current space represented by the frequency band axis (band number), quality axis (amplitude bit division number), and time axis (frame number) orthogonal to each other, [f , Q, i].

Each explanatory variable in this case is obtained by the following equation. An audio signal of band f and quality (bit division number q) is represented as sfq.
x1 [f, q, i] = (1 / N) Σ _{n = 1} ^N sfq [Ni + n] ²
Or x1 [f, q, i] = log ₁₀ ((1 / N) Σ _{n = 1} ^N sfq [Ni + n]) ²
x2 [f, q, i] = x1 [f, q, i] / x1 [f, q, i-1]
x3 [f, q, i] = max (ρ _{f, q, i} [k])
ρ _{f, q, i} [k] = Σ _{n = 0} ^N (sfq [Ni + n]) (sfq [Ni + n + k]) / Σ _{n = 0} ^N (sfq [Ni + n]) ²
x4 [f, q, i] = x1 [f, q, i] / Σ _{f = 1} ^F x1 [f, q, i]
x5 [f, q, i] = log ₁₀ (Σ _{n = 1} ^N efq [Ni + n] ² / Σ _{n = 1} ^N sfq [Ni + n] ² )
Index value y [f, q, i] = α0 + Σ _{j = 1} ⁵ αjxj [f, q, i]
Priority p [f, q, i] = f _{f, q} (y [f, q, i])
The priority p [f, q, i] determined in this way and the corresponding encoded code P [f, q, i] are transmitted as one packet.

In the above fifth embodiment , each divided audio block is output as one packet with its encoding code and priority set as a set. In the fifth embodiment, the codes of signal blocks having the same priority are combined into one packet. Send out as a packet.
For example, as shown in FIG. 20, a frame-divided audio signal is divided into F bands by a band dividing unit 1411, and the 1st to Fth band signals are encoded by encoding units 2011-1 to 2011-F, respectively. And priorities are determined by the priority determination units 2012-1 to 2012 -F. In the fifth embodiment, the encoded codes P [1, i] to P [F, i] and the priorities p [1, i] to p [F, i] are supplied to the packet aggregating unit 2013, and a predetermined frame is received. For each number, codes of the same priority are collected and sent from the sending unit 1215 as one packet.
Input audio signal s [n] is divided into F = 3 bands of 0-4 kHz, 4 kHz-8 kHz, and 8-16 kHz using, for example, wavelet analysis, divided in the time direction in 5 ms, and transmitted in packets every 20 ms And The frame number i = 1,..., 4 in each packet transmission number t, the encoding code of the signal block of the band number f of the frame number i is P [f, i], and the priority is p [f, i]. Represent each. When the code P [f, i] and the priority p [f, i] of each block in each t-th transmission section are as shown in FIG. 21A, the packet aggregating unit 2013 is the same as shown in FIG. 21B. Each block having priority is aggregated to form one packet. In this example, the codes P [1,2] and P [1,3] of the blocks (1,2) and (1,3) with the priority p = 4 are put together, and the codes P [1,2], The position information (1, 2), (1, 3) on the band-time coordinate of P [1, 3] is incorporated into the packet with the priority p = 4. Codes P [2, 2], P [1, 4] and their position information (2, 2), (1, 4) are incorporated into a packet of priority p = 3. In the same manner, codes having the same priority are collected and incorporated as one packet together with the position information.

In this example, packets in which codes having the same priority are aggregated are sent to the network every 20 ms. At this time, a packet having a low priority has little influence on the quality according to the state of the network, and therefore does not need to be transmitted. Further, even if a low-priority packet is discarded in each node of the network according to traffic congestion, the influence on the call quality is kept to a minimum.
In this way, the packets sent to the network are all packets in the t-th sending section at the receiving side as shown in FIG. 22 on the receiving side, and in the case of FIG. Two packets P [1, t] to P [4, t] are reconstructed on band-time coordinates through the reverse procedure of the assembly shown in FIG. 21, and each band code P [1, i] to P [ F, i] are decoded by the decoding units 2212-1 to 2212 -F into band audio decoding. At this time, if there is a low priority code that has not reached the receiving side, basically the operation of the decoding unit for that code is stopped. When the high priority code does not arrive, the frame (block) loss countermeasure is taken in the corresponding part of the block loss compensation units 2213-1 to 2213 -F to avoid quality degradation. Each band audio signal decoded in this way and subjected to erasure compensation as necessary is synthesized by the band synthesizing unit 2214 and output as a reproduced audio signal s [n]. Note that block loss information is supplied from the packet decomposition unit 2211 to the block loss compensation units 2213-1 to 2213 -F. This block disappearance compensation may be performed by a known technique.

As shown in FIG. 19, in the case of the fourth embodiment in which the audio signal is blocked in three-dimensional coordinates (space), a block code having the same priority is grouped together with its position information for each predetermined number of frames. May be sent out.
In the above description, the prioritized packet output method is applied to the audio signal, but it can also be applied to the music signal and the video signal. In addition, the following may be considered as explanatory variables of feature amounts based on encoding. For example, depending on the speech coder using predictive coding, if a packet such as a word head is discarded, the speech quality (S / N ratio) thereafter may be significantly degraded. Degradation of the S / N ratio that propagates when such a packet is discarded may be used as the explanatory variable xj (m, j). The explanatory variable of the feature amount of the audio signal and the explanatory variable of the feature amount based on the encoding are not limited to the above-described examples, and various types can be used.

  The prioritized packet transmission method described above can be applied to the speech encoding method and the encoding method of the present invention described with reference to FIGS. As a result, it is possible to obtain an effect that the influence on the call quality is kept to a minimum even in a situation where traffic is congested.

  The speech encoding method and decoding method of the present invention can be used in a packet communication network such as the Internet, and can be interconnected from a low function terminal to a high function terminal.

BRIEF DESCRIPTION OF THE DRAWINGS The block diagram for demonstrating 1st Example of the audio | voice encoding apparatus which performs the audio | voice encoding method of this invention. The block diagram for demonstrating the structure of the band division | segmentation filter bank used for the Example shown in FIG. The graph for demonstrating an example of the frequency characteristic of the frequency characteristic compensation part used for the Example shown in FIG. The graph for demonstrating the other example of the frequency characteristic of the frequency characteristic compensation part used for the Example shown in FIG. The block diagram for demonstrating the Example of the audio | voice decoding apparatus which performs the audio | voice decoding method of this invention. The block diagram for demonstrating the structure of the resynthesis filter bank used for the audio | voice decoding apparatus shown in FIG. The block diagram which shows the Example of the audio | voice encoding apparatus at the time of applying the audio | voice encoding apparatus shown in FIG. 1 to 16kHz sampling. The block diagram which shows the structure of the audio | voice decoding apparatus which arrives and decodes the packet transmitted with the audio | voice encoding apparatus shown in FIG. The block diagram for demonstrating an example of a structure of the band division | segmentation filter bank used for the audio | voice coding apparatus shown in FIG. The block diagram for demonstrating an example of a structure of the resynthesis filter bank used for the audio | voice decoding apparatus shown in FIG. The block diagram for demonstrating the Example of the audio | voice coding apparatus demonstrated in Example 2 of this invention. The block diagram which shows the function structural example of 1st Embodiment of the transmission method of the packet signal with a eugenicity applicable to this invention. FIG. 14 is a block diagram illustrating a specific functional configuration example of a priority determination unit 1214 in FIG. 13. The block diagram which shows the function structural example of 2nd Embodiment of the transmission method of this packet signal with a priority. The figure which shows the example which divided | segmented the signal into the block of a band-time coordinate. FIG. 15 is a block diagram showing a specific functional configuration example of a low frequency priority determination unit 1413L in FIG. 14. The block diagram which shows the function structural example of 3rd Embodiment of the packet transmission method with a priority. FIG. 18 is a block diagram illustrating a specific functional configuration example of a first-stage priority determination unit 1413-1 in FIG. The figure which shows the example which divides | segments a signal into the three-dimensional coordinate of the quality-band-time used for 4th Embodiment of the packet transmission method with a priority. The block diagram which shows the function structural example of 5th Embodiment of the packet transmission method with a priority. The figure for demonstrating the process of the packet aggregation part 2013 in FIG. The block diagram which shows the function structural example of the packet receiver corresponding to the packet transmitter shown in FIG.

Explanation of symbols

DESCRIPTION OF SYMBOLS 10 Band-division filter bank 50 Packet decomposition part 11 Frequency characteristic compensation part 51 Telephone band compensation decoding part 12 Gain adjustment part 52 Telephone band decoding part 13 Telephone band encoding part 53 Middle band high band side decoding part 14 Phase delay compensation part 54 Wide band High band side decoding unit 15 Subtraction unit 55 Phase delay compensation unit 16 Telephone band compensation coding unit 56 Gain adjustment unit 17 Middle band high band side coding unit 57 Addition unit 18 Wide band high band side coding unit 58 Resynthesis filter bank 19 Packet building department

Claims (18)

Band division processing for dividing a wideband audio signal into a narrowband telephone band audio signal, a medium band highband audio signal, and a wideband highband audio signal;
Frequency characteristic compensation processing for changing the frequency characteristic of the telephone band voice signal that has been divided into high-frequency emphasis characteristics;
An encoding process for encoding each of the telephone band audio signal subjected to the frequency characteristic compensation process, and the medium-band high-frequency audio signal and the broadband high-frequency audio signal;
A packet construction process for sending each encoded data as a packet signal;
A speech encoding method comprising:   The speech encoding method according to claim 1, wherein gain compensation is performed on the telephone band speech signal before the encoding process.   3. The speech encoding method according to claim 1, wherein the priority of each band speech signal is obtained and the digital signal packet output method with priority is used together.   4. The voice encoding method according to claim 1, wherein a difference between the telephone band voice signal subjected to the frequency characteristic compensation processing and the telephone band voice signal not subjected to the frequency characteristic compensation processing is obtained, and the difference signal is obtained. An audio encoding method comprising: encoding processing; packetizing the encoded data as compensation code data;   4. The speech encoding method according to claim 1, wherein the decoding processing for decoding the encoded data obtained by encoding the telephone band speech signal, the telephone band signal subjected to the decoding processing, and the frequency characteristic compensation. A subtraction process for obtaining a difference from the unprocessed telephone band voice signal, an encoding process for encoding the difference signal obtained by the subtraction process, and the compensation code data obtained by the encoding process are transmitted as a packet signal. A speech coding method characterized by adding a packet construction process. Packet decomposition processing for decomposing an incoming packet signal into wideband high-frequency voice encoded data, medium-band high-frequency voice encoded data, and telephone band voice encoded data;
A decoding process for decoding each of these audio encoded data into an audio signal;
A synthesis process for synthesizing the decoded audio signal and reproducing a wideband signal;
A speech decoding method comprising:   7. The voice decoding method according to claim 6, wherein a process of extracting compensation code data by the packet decomposition process, decoding the compensation code data, and adding the decoded compensation component to the telephone band voice signal is added. A speech decoding method characterized by the above.   8. The speech decoding method according to claim 6, wherein a gain correction process is performed on a speech signal obtained by decoding the telephone band encoded data. A band dividing unit that divides a wideband audio signal into a narrowband telephone band, a medium-band wide-area voice signal, and a wide-band wideband voice signal;
A frequency characteristic compensator for changing the frequency characteristic of the telephone voice signal subjected to the band division to a high frequency emphasis characteristic;
An encoding unit that encodes each of the telephone band audio signal frequency-compensated by the frequency characteristic compensator, the medium-band wide-band audio signal, and the wide-band wide-band audio signal;
A packet construction unit that sends out the encoded data encoded by each encoding unit as a packet signal;
A speech encoding apparatus comprising:   10. The speech coding apparatus according to claim 9, further comprising a gain compensation unit that performs gain compensation on a telephone band speech signal input to the coding unit.   The subtractor for obtaining a difference between the telephone band voice signal frequency-compensated by the frequency characteristic compensator and the telephone band signal not subjected to frequency characteristic compensation, according to any one of the voice encoding devices according to claim 9 or 10, An audio comprising: an encoding unit that encodes a difference signal obtained by subtraction by the subtraction unit; and a packet construction unit that transmits the differentially encoded data encoded by the encoding unit as a packet. Encoding device.   11. The voice encoding device according to claim 9, wherein a decoding unit that decodes encoded data of the telephone band voice signal, a telephone band signal decoded by the decoding unit, and a frequency characteristic by the frequency characteristic compensation unit. A subtractor for obtaining a difference from a non-compensated telephone band signal, an encoding process for encoding the difference signal obtained by the subtractor, and a packet for transmitting the compensation code data obtained by the encoding process as a packet signal A speech coding apparatus comprising: a construction unit. A packet decomposing unit for decomposing an incoming packet signal into wideband high-frequency side voice encoded data, medium-band high-frequency side audio encoded data, and telephone band voice-encoded data;
A decoding unit for decoding each of these audio encoded data into an audio signal;
A synthesis unit that synthesizes the decoded audio signal and reproduces the wideband signal;
A speech decoding apparatus comprising:   14. The speech decoding apparatus according to claim 13, wherein an addition unit is provided for extracting compensation code data by the packet decomposition unit, decoding the compensation code data, and adding the decoded compensation component to the telephone band speech signal. A speech decoding apparatus characterized by that.   15. The speech decoding apparatus according to claim 13, further comprising: a gain adjusting unit that performs gain correction on the speech signal obtained by decoding the telephone band encoded data. apparatus.   A speech encoding program for causing a computer to execute at least one of the speech encoding methods according to claim 1 according to each processing step.   A speech decoding program for causing a computer to execute at least one of the speech decoding methods according to claim 6 according to each processing step. 18. A recording medium comprising a computer-readable recording medium on which at least one of the voice encoding program and the voice decoding program according to claim 16 and 17 is recorded.

Images