JP3490325B2 - Audio signal encoding method and decoding method, and encoder and decoder thereof - Google Patents

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a digital signal with a small amount of information by inputting an audio signal and minimizing distortion between the input audio signal and the synthesized reproduction signal on a predetermined distance scale. The present invention relates to a high-efficiency speech coding method for coding, a decoding method thereof, an encoder and a decoder thereof .

[0002]

2. Description of the Related Art A high-efficiency voice signal coding method is used in order to efficiently use radio waves in digital mobile communications and to efficiently use communication lines and storage media in voice or music storage services. . As a voice encoding method, generally, an encoding method for telephone band voice whose frequency band is limited to 3.4 kHz or less and an encoding method for voice including a frequency band up to 7 kHz band are generally used. It's being used. These encoding methods include ITU-T
G.G. 723.1, G.I. 729, G.I. 7
There are 22 etc.

Of these, the 7 kHz band coding system has a high naturalness, but has a relatively high bit rate, and the telephone band coding system often has a relatively low bit rate, but it is natural. Has a feature that it does not reach the encoding system of the 7 kHz band. In actual applications, these coding methods are often selected and used according to various requirements.

Code-driven linear predictive coding (Cod) is one of the methods for coding speech at a relatively low bit rate.
e-Excited Linear Predicti
The method called on: CELP) is often used. For details of this technique, refer to the document M. R. Schr
oeder and B.I. S. Atal, "Code-
Excited Linear Prediction
(CELP): High Quality Spec
h at Very Low Bit Rates ”,
IEEE Proc. ICASSP-85, pp. 93
7-940,1985.

FIG. 10 shows a functional configuration of this encoding method. A linear prediction parameter representing the frequency spectrum envelope characteristic of the input voice is calculated in the linear prediction analysis unit 1-2 using the acoustic signal (input voice) input to the input terminal. The obtained linear prediction parameter is coded by the linear prediction parameter coding unit 1-3 and sent to the linear prediction parameter decoding unit 1-4. Further, when the distortion calculation is performed using the spectral information of the input voice such as considering the auditory characteristics in the distortion calculation, the linear prediction parameter is also sent to the distortion calculation unit 1-7. The linear prediction parameter decoding unit 1-4 reproduces the synthesis filter coefficient from the received code and sends it to the synthesis filter 1-6. When the auditory characteristics are taken into consideration in the distortion calculation, the decoded linear prediction parameter may be used in the distortion calculation. Details of the linear prediction analysis and an example of coding the linear prediction parameters are described in, for example, "Digital Speech Processing" by Sadahiro Furui (Tokai University Press). Here, the linear prediction analysis unit 1-2, the linear prediction parameter encoding unit 1-3, the linear prediction parameter decoding unit 1-4, and the synthesis filter 1-6 may be replaced with non-linear ones.

The driving sound source vector generation unit 1-5 generates driving sound source vector candidates having a length of one frame and sends them to the synthesis filter 1-6. FIG. 11 shows a functional configuration example of the driving sound source vector generation unit 1-5. From the adaptive codebook 2-1, the immediately preceding past driving excitation vector (driving excitation vector for one to several frames immediately before being already quantized) c (t-1) stored in the buffer is set to a certain cycle. By cutting out with a corresponding length and repeating the cut out vector until the length of the frame is reached, candidates for the time-series vector corresponding to the periodic component of the speech are output. The "certain cycle" is selected as a cycle in which the distortion d in the distortion calculation unit 1-7 is small, and the selected cycle generally corresponds to the pitch cycle of the voice.

From fixed codebook 2-2, a time-series code vector candidate having a length of one frame corresponding to an aperiodic component of speech is output. These candidates are stored as a predetermined number of candidate vectors according to the number of bits for coding, independently of the input speech signal. The fixed code vector candidates output from the fixed codebook 2-2 are, in the periodization unit 2-3, periodicized as necessary with a period (generally equivalent to the pitch period as described above) specified by the periodic code. It Periodization refers to applying a comb filter having a tap at a specified cycle position, or repeating a vector cut out at a length corresponding to a specified cycle from the beginning of the vector as in the adaptive codebook. The periodicization unit 2-3 is often used from the viewpoint of improving coding efficiency, but it may not be used in some cases. Also, such as consonant section,
In the case where the voice itself has no or little pitch component, in some cases, the cycler does not work.

Adaptive codebook 2-1 and periodicizing section 2-3
The candidate of the time series vector output from the multiplication unit 2-
In 4 and 2-5, the weights ga and gf generated in the weight generating unit 2-7 are respectively multiplied, and added in the adding unit 2-6 to become the driving sound source vector candidate c. Figure 11
In the configuration example of 1., the adaptive codebook 2-1 may not be used, and only the fixed codebook 2-2 may be configured. When encoding a signal with a small pitch periodicity such as a consonant part or background noise, bits are In order to save the cost, it is often the case that the adaptive codebook 2-1 is not used.

A synthesis filter 1-6 shown in FIG. 10 is a linear filter which uses the output of the linear prediction parameter decoding unit 1-3 as a filter coefficient, and outputs a candidate y of a reproduced voice with a driving sound source vector candidate c as an input. . The order of the synthesis filter 1-6, that is, the order of the linear prediction analysis is generally 10 to 16.
The following are often used: As described above, the synthesis filter 1-6 may be a non-linear filter.

In the distortion calculation section 1-7, the synthesis filter 1-
The distortion d between the reproduced voice candidate y which is the output of No. 6 and the input voice x is calculated. This distortion calculation is often done taking into account the coefficients of the synthesis filters 1-6 or the unquantized linear prediction coefficients, for example auditory weighting. Figure 1
2 shows an example of a functional configuration in which distortion is calculated in consideration of auditory weighting. Perceptual weighting is configured in the form of perceptual weighting filters 3-2, 3-3 using unquantized linear prediction parameters or quantized linear prediction filter coefficients. The reproduced voice candidate y output from the synthesis filter 3-1 is passed through the perceptual weighting filter 3-2, and the distortion d is calculated with the input voice also passed through the perceptual weighting filter 3-3. Here, the auditory weight filter 3-2
3-3 is equivalent even if it is inserted as one filter after the distance calculation unit 3-4, but in terms of processing amount, as shown in FIG. 12, there are two locations before the distance calculation unit 3-4. Often divided into

In the codebook search controller 1-9 shown in FIG.
A driving sound source code that minimizes the distortion d between each reproduced sound candidate y and the input sound x is selected, and the driving sound source vector in that frame is determined. When the adaptive codebook 2-1, fixed codebook 2-2, and weight codebook 2-3 shown in FIG. 11 are used, the periodic code, fixed code,
And weight codes are selected and used as the driving sound source.

The driving excitation code (periodic code, fixed (noise) code, weight code) determined by the codebook search control unit 1-9 and the linear prediction parameter code output from the linear prediction parameter coding unit 1-2. Is sent to the code sending unit 1-10 and is stored in the storage device or sent to the receiving side via the communication path depending on the form of use. That is, the short-term prediction component for each frame of speech is used as a linear prediction parameter code, and the periodic component longer than the frame in the prediction residual component of this short-term prediction component is used as the periodic code and the remaining components are fixed (noise). ) Code, and the amplitudes of the periodic component and the remaining component are encoded as weight codes.

FIG. 13 shows an example of the functional configuration of a decoding method corresponding to the above encoding method. Among the codes received from the transmission path or the storage medium, the linear prediction parameter code is decoded into the synthesis filter coefficient in the linear prediction parameter decoding unit 4-2, and the synthesis filter 4-4 and, if necessary, the post-processing unit 4-5. Sent to. The received excitation code is
It is sent to the driving sound source vector generation unit 4-3, and the sound source vector corresponding to the code is generated. The driving sound source generator 4
The configuration of -3 corresponds to the drive excitation vector generation unit 1-4 of the encoding method shown in FIG. The synthesis filter 4-4 receives the driving sound source vector as an input and reproduces a voice. The post-processing unit 4-5 performs a process (also referred to as post-filtering) that aurally reduces the noise sensation of the reproduced voice, but the post-processing unit 3-5 is used because of the reduction of the processing amount and the like. There are many things that cannot be done.

As one of the driving sound source vector search methods of the CELP system, Algebraic Code-Exci
ted Linear Prediction (ACE
LP) has been proposed. This method does not store the fixed codebook as a vector pattern of the frame length, but rather several pulses with a height of 1 in a frame, for example, four pulses for a frame or subframe of 80 samples. By setting it to the position, a fixed code vector is adopted, and by adopting this driving sound source method and devising the calculation order in distortion calculation, it is possible to reduce the amount of calculation processing and memory required compared to the conventional method. it can. The details of the ACELP method are described in, for example, the literature, R.M. Salami, C.I. Laflamme, and
J-P. Adoul, “8kbit / s ACELP
Coding of Speech with 10
ms Speech-Frame: a Candida
te for CCITT Standardizatat
Ion ", IEEE Proc. ICASSP-94,
pp. II-97.

[0015]

The frequency components contained in human voice are generally concentrated in a band of 7 kHz or less.
If the band is limited to 3.4 kHz or less, the amount of information is reduced and prediction is facilitated, so that the compression coding efficiency is improved, but part of the naturalness and individuality information is lost. From this, 3.4
When inputting a voice in the kHz band, 8 kbi
It is possible to achieve a relatively high S / N even at a bit rate of about t / s, but the naturalness is deteriorated at the original sound level compared to the sound in the 7 kHz band, so it is necessary to reproduce sound with good naturalness. I can't.

On the other hand, if the voice to be input is acquired in the 7 kHz band, the naturalness of the input voice itself is very high.
Due to the large amount of information, low bit rate, for example 4-6k
It is very difficult to realize high quality voice at about bit / s. 8k to input the sound of 3.4kHz band
A sampling rate of about Hz is required, and if the quantization bit number is 16 bits, the amount of information is 16 × 8000 bits / s. On the other hand, a sampling rate of about 16 kHz is required to input the sound of 7 kHz band,
16 × 16 when the number of quantization bits is 16 bits
The amount of information is 000 bits / s. If you use the feature of voice that power is concentrated in the low frequency range , 0-3.4kHz
Compared with z, 3.4 to 7 kHz can be expressed with a smaller amount of information, but in the case of 7 kHz speech, it is difficult to predict in the high frequency region for both short-term frequency component prediction and long-term pitch component prediction. Therefore, it is very difficult to reduce the bit rate to the same extent as when encoding speech in the 3.4 kHz band.

In addition, in the low bit rate audio encoding method with variable bit rate, when the bit rate is variably controlled in frame units, a problem may occur due to the relationship between encoding efficiency and delay. For example, in encoding systems that support various kinds of bit rate modes, there are many restrictions due to quality restrictions, such that the lower the bit rate, the longer the frame length. In such a variable bit rate realizing method, if the bit rate is changed in units of frames during reproduction, there is a problem that sound is interrupted due to delay constraint. Alternatively, there is a method of adjusting the total delay to the bit rate mode having the longest delay, but in that case, there remains a problem that the performance in the mode having the shortest delay cannot be maximized.

The present invention has been made in view of the above-mentioned drawbacks of the conventional method, and has an overwhelmingly low bit rate while suppressing deterioration of quality as compared with the voice encoding system of the 7 kHz band. An audio signal encoding method and a decoding method capable of reproducing high-quality audio with overwhelmingly high naturalness at a bit rate comparable to that of an audio encoding method in the 4 kHz band.
It is an object of the present invention to provide an encoder and a decoder thereof. Another object of the present invention is to provide an audio signal encoding method and an encoder capable of continuously changing the bit rate in frame units without interruption of sound.

[0019]

According to the present invention, by balancing the frequency band of input speech to about 5 kHz, that is, 4.5 kHz to 5.5 kHz, the amount of information contained in speech and the coding efficiency are balanced. Therefore, it is possible to perform encoding at an overwhelmingly lower bit rate than the conventional encoding method for the input voice of the 7 kHz band, and compared to the encoding method for the input voice of the 3.4 kHz band, It can play overwhelmingly natural sound at the same low bit rate.

For example, when a voice having a sampling rate of 11 kHz is input, the amount of information increases 11/8 times as compared with the case of sampling at 8 kHz. Therefore, the number of pulses in the fixed codebook is generally 11/8. It can be easily inferred that it is necessary to double the number. 11 kHz in this invention
Of the amount of information that increases when the sampling rate is used, only the information within the range that matches the speech model (the short-term prediction component, the long-period component in the prediction residual and the remaining component) By using it, the linear prediction parameter quantization efficiency and the adaptive codebook quantization efficiency are improved, and by using the same characteristics, even in fixed codebook pulse candidates, the pulse arrangement positions are made very sparse and Limit the number of.

As a result, even if the number of pulses is almost the same as in the case of sampling at 8 kHz, it is possible to reproduce a very high quality voice. When lowering the bit rate, in order to improve information efficiency, it is often the case that the frame length is increased to reduce the rate of sending information of the linear prediction parameter. In such a case, since the delay changes as the frame length changes, it is necessary to match the delay to the mode with the longest delay and perform prefetching.

According to the present invention, the frame length and the delay are made constant in all modes, and the information of the prefetched portion can be used without exception in all modes. Therefore, the efficiency reduction for dealing with bit rate switching is minimized. It becomes possible to suppress. Furthermore, by using the adaptive codebook without re-initialization when switching the number of subframes and fixed codebook, the information sent in the past can be used effectively, and the bit rate mode can be switched frame by frame without interruption of sound. It can be carried out. Action While the voice in the 7 kHz band contains a lot of high frequency information that does not match the model of the voice used in the CELP method, the voice in the 3.4 kHz band is within the range that matches the voice model due to the frequency band limitation. Information has been lost.

On the other hand, the voice of the 5 kHz band contains necessary and sufficient information within the range that matches the above-mentioned voice model, and contains almost no other information, and is encoded by the above-mentioned voice model. 7 is suitable for
The bit rate can be overwhelmingly lower than that of voices in the kHz band, and when compared to the encoding method in the 3.4 kHz band, when the bit rate is about the same, the band becomes wider, which is overwhelmingly natural. It can play high quality and high quality audio.

Also, in the case of realizing a variable bit rate, inputting a voice in the 5 kHz band makes it possible to obtain 3.4 k
Higher quality can be realized as compared with the variable bit rate speech coding method for input speech in the Hz band.

[0025]

BEST MODE FOR CARRYING OUT THE INVENTION Embodiments of the present invention will be described below with reference to the drawings. FIG. 1 shows a functional configuration example of a speech coder according to the present invention. The difference from the conventional method is that the input target is a voice (11.02
5 kHz sampling), the subframe length and the fixed codebook are switched while the frame length remains the same during the bit rate switching control, and the adaptive codebook is used continuously without re-initialization for both the encoder and the decoder. It is a point. FIG. 2 shows a functional configuration example of the decoder according to the present invention. <Embodiment 1> As an example of an encoding method according to the present invention, a frame length of 10 ms and a bit rate of 7.8 k are obtained by using the present invention.
An example of designing a bit / s encoding method will be shown.

The input voice signal is filtered by the filter section 5-1.
In, the signal is low-pass filtered and band limited to 5 kHz or less. The analysis frame length is 110 samples. This corresponds to about 10 ms. It is assumed that the number of subframes is 2 and prefetching is 1 subframe 5 ms. FIG. 3 shows an example of bit allocation. In the linear prediction analysis unit 5-2, 14th-order linear prediction analysis is performed, and Le is calculated from the obtained linear prediction coefficients.
1 according to the Vinson-Durbin algorithm
The fourth-order LSP coefficient is calculated. Here, the order is 14th in order to keep the calculation amount low, but the linear prediction order may be about 14th to 20th.

The linear prediction parameter coding unit 5-3 codes the linear prediction parameter by two-stage vector quantization using moving average (MA) prediction. 1 bit for moving average (MA) prediction mode switching, 1
The 7th bit was used for the second stage, and the second stage was divided into two, low order and high order, and 5 bits were used for each. For voice in the 3.4 kHz band, the analysis order is, for example, about 10th order. On the other hand, in the case of speech in the 5 kHz band, the amount of information is increased because the linear prediction analysis order is increased to 14. However, the number of vector quantization bits of LSP is assumed to be about the same even though the amount of information increases. This is because the linear prediction coefficient is correlated up to about 5 kHz, and the characteristic of the 5 kHz band audio signal that it can be quantized with the same number of quantization bits as the audio in the 3.4 kHz band is used. In addition, 7
When a voice in the kHz band is input, it is difficult to set the number of bits to the same level because the high frequency includes uncorrelated components. Since the order of the LSP coefficient is increased, it is usual to faithfully represent it by increasing the number of code vectors of the LSP codebook used for the quantization, and thus increasing the number of coded bits. In the present invention, the number of code vectors is the same as the number of code vectors of the LSP codebook used for encoding in the 3.4 kHz band, and thus the number of encoded bits is the same. Therefore, in the case of the same number of coded bits, the present invention has a larger quantization distortion than that in the 3.4 kHz band. However, 5 kHz
Up to a degree, there is a correlation between linear prediction coefficients, and the increase in the quantization distortion is relatively small. Based on the fact that the band was wider than the increase in the quantization distortion, that is, the order of the LSP coefficient was increased. The improvement in quality has a great influence on the reproduced sound, and the sound becomes natural.

Driving sound source vector generation unit 5-12 to 5-1
4 (in the first embodiment, one driving sound source vector generation unit
Drive source vector to adaptive codebook or fixed codebook
It is generated by adding weights and adding them. Adaptive code
The search of the book is from the beginning of the time series data stored there.
Cut out by changing whether to cut up to the position of
The position is an integer precision search performed in sample position units, and
Triple precision search performed for each position by dividing the sample positions into three equal parts
Searching is usually done. 11.025kHz sample
Adaptive codebook, for example,
Reference position when 8 bits are assigned to the book index
From 27 + 1/3 to 85 + 1/3 samples is 3
Double precision search with integer precision from 86 to 165 samples
You can search. For the second subframe, the previous
Adaptive prediction value T obtained in subframe ₁Int the integer part of
(T₁) [Int (T₁) -5 + 2/3, i
nt (T₁) + 4 + 2/3] with 3 times accuracy
The measurement is searched and represented using 5 bits. Where the adaptation mark
26 if 9 bits are assigned to the issue index
Triple precision from +1/3 to 185 + 2/3 samples
So, search from 186 to 220 samples with integer precision
May be.

When the search range of the adaptive codebook is set in this way, if the same number of bits is assigned, it is 5 kHz.
In the case of using the voice in the z band, the search period with triple precision is shorter than in the case of inputting in the 3.4 kHz band. However, if the characteristic of the 5 kHz band voice that the pitch periodicity is similar to that of the 3.4 kHz band voice is used, equal or higher quality can be obtained with the same number of bits.

8 kHz for voice in the 3.4 kHz band
z is sampled at z, and 11.
It is sampled at 025 kHz. Since the sampling interval of 11.025 kHz sampling is narrower than that of 8 kHz sampling, the time resolution of adaptive codebook search is improved. Substantially, the triple precision in the 5 kHz band can be converted into the 4.13 precision in the 3.4 kHz band. This is also the reason why the quality is improved when the input is in the 5 kHz band.

As indicated above, the 4.5 kHz band
By inputting the voice of the 5.5 kHz band as the input target, the performance of the LSP codebook and the adaptive codebook is remarkably improved even if the bit allocation is similar to that of the encoding method for the voice of the 3.4 kHz band. The bit allocation of the first embodiment is shown in FIG.
As shown in FIG. 729 (8 kb
It is different from the bit allocation of it / s) only in that the weights (gains) are 7 and 7 bits, respectively. As described above, in the first embodiment, the G.I. 2 bits less than 729,
As will be shown later, the quality of reproduced voice is improved in the first embodiment.

From this bit allocation, the first subframe of the adaptive codebook, G. Although both 729 are 8 bits, the number of samples in this subframe is 55. 729 is 40 samples. Therefore, the bit allocation number per sample is 8/55 in the first embodiment, and G. 729 becomes 8/40, and the number of bits is smaller in the first embodiment. As described above, one of the features of the present invention is that the bit allocation per sample is made smaller than that of the 3.4 kHz band.

Also, in the search for the fixed codebook, as compared with the case of the adaptive codebook, 1
Since the sampling interval of 1.025 kHz sampling is narrow, it is possible to obtain the effect of improving the time resolution of the fixed codebook search. By using these things, the number of pulses per 10 ms set in the fixed codebook can be made very sparse, and the bit allocation of the fixed codebook can be reduced to the same level as the 3.4 kHz band. . Even if the bit allocation of the fixed codebook is reduced in this way, the improvement in the performance of the LSP codebook and the adaptive codebook exceeds the quality deterioration due to the reduced number of pulses of the fixed codebook.

FIG. 4 shows an example of a pulse arrangement in which 4 pulses are allocated to a 5 ms subframe to have 17 bits.
The row of i0 in FIG. 4 indicates the position where the 0th pulse can stand, and the positions where the 1st, 2nd and 3rd pulses can stand respectively similarly. Four pulses 0 to 3 are set in each subframe. <Second Embodiment> A second embodiment of the encoding method of the present invention will be described.
In the second embodiment, the frame length is 20 ms, 3.95 kbi.
t / s, 5.75 kbit / s, 7.75 kbit / s
This is a case where the encoding method in which the bit rate is variably controllable in three stages.

The input voice signal is filtered by the filter section 5-1.
, And is bandpass limited to 5 kHz or less. The analysis frame length is 220 samples. This corresponds to about 20 ms. The number of subframes is 3.95 kbit / s, 5.75 kbit / s, 7.7.
It is 2, 3, and 4 in each mode of 5 kbit / s. The prefetch is about 7 ms and about one third of the frame length.

An example of a mode switchable encoder is shown in FIG.
FIG. 5 shows an example of bit allocation in each bit rate mode. Each bit rate mode can be set and switched independently for each frame. In order to inform the decoder of the bit rate mode of the frame, 1 bit to 2 bit of each frame is used separately from the one shown in FIG. For example, 3.95 kbit / s,
5.75 kbit / s and 7.75 kbit / s are represented by 0, 10, and 11, respectively. In this case, each bit rate is 4 kbit / s, 5.85 kbit / s, 7.85.
It becomes kbit / s. When an IP packet or the like is used for communication, the bit rate mode of the frame can be indirectly known from the packet size information, and it is not necessary to send information indicating the bit rate mode. The user switches the bit rate mode, or the upper side of the encoder gives an instruction as to which bit rate mode to use in accordance with the congestion state of communication, for example.

The fixed codebook pulse arrangement in each bit rate mode is shown in FIGS. Linear prediction analysis unit 5
In −2, a 14th-order linear prediction analysis is performed, and a 14th-order LSP coefficient is calculated from the obtained linear prediction coefficient by the Levinson-Durbin algorithm. Although the fourteenth order is used here in consideration of the amount of calculation, a linear prediction order of about 14th to 20th can be used.

The linear prediction parameter coding unit 5-3 codes the linear prediction parameter by two-stage vector quantization using moving average (MA) prediction. 1 bit for moving average (MA) prediction mode switching, 1
8 bits were used for the second stage, and the second stage was divided into two, low order and high order, and 6 bits were used for each. For voice in the 3.4 kHz band, the analysis order is, for example, about 10th order. On the other hand, in the case of speech in the 5 kHz band, the amount of information increases because the linear prediction analysis order is increased to 14. However, the number of vector quantization bits of LSP is assumed to be about the same even though the amount of information increases. This is because the linear prediction coefficient is correlated up to about 5 kHz, and the characteristic of the 5 kHz band audio signal that it can be quantized with the same number of quantization bits as the audio in the 3.4 kHz band is used. Also, 7
When a voice in the kHz band is input, it is difficult to set the number of bits to the same level because the high frequency includes uncorrelated components.

In the driving sound source vector control unit 5-5,
Driving sound source vector generation units 1 to n used in each mode
In all of (5-12 to 5-14), the previous driving excitation vector is commonly stored in each of the adaptive codebooks, and the past adaptive codebook vectors are used as they are without initializing even when the bit rate mode is switched. To do. This makes it possible to effectively use the information of the adaptive codebook sent in the past. Here, the bit rate mode is 3.95 kbi.
t / s, 5.75 kbit / s, 7.75 kbit / s
A driving sound source vector generation unit is prepared for each of the three modes. The number of each subframe is 2, 3, 4
Is.

The adaptive codebook index is 7.75 kb.
9 in the first and third subframes in the it / s mode
Bits are allocated, and in the second and fourth subframes, [int (T ₁ ) -5 + 2/3, int (T ₁ ) + is added to the integer part int (T ₁ ) of the adaptive prediction value obtained in the previous subframe.
4 + 2/3], and search for an adaptive prediction value with triple precision,
It is expressed using 5 bits. In the other bit rate modes, 9 bits are assigned to the first subframe, and 5 bits are used for the remaining subframes.

In this case, since the voice is characterized in that pitch prediction is relatively easy up to about 5 kHz, the adaptive codebook component is expressed by the same number of bits as used for the voice in the 3.4 kHz band. It is possible. Also, 11.0
Since the input signal of 25 kHz sampling is used,
Another reason for the quality improvement is that the time resolution of adaptive prediction is improved as compared with the case where a signal of 8 kHz sampling is used.

Also here, as in the case of the first embodiment, 4.
By inputting speech in the 5 kHz band to 5.5 kHz band, the performance of the LSP codebook and the adaptive codebook is remarkably improved even if the bit allocation is similar to that of the encoding method for speech in the 3.4 kHz band. To do. Also, in the fixed codebook search, as in the case of the adaptive codebook, the sampling interval of 11.025 kHz sampling is narrower than that of 8 kHz sampling, so that the time resolution of the fixed codebook search is improved. You can

By utilizing these things, the number of pulses per subframe set up in the fixed codebook can be made very sparse, and the bit allocation of the fixed codebook is 3.4.
It is possible to reduce to the same level as the kHz band. In the driving sound source vector control unit 5-5, the driving sound source vector switching unit 5-11 changes the bit rate by switching the driving sound source vector to be used. Each driving sound source vector generation unit 1 to n (5-12 to 5-1
In 4), the drive excitation vector buffer is not reinitialized even when the bit rate mode is switched, and the adaptive codebook is searched using the drive excitation vector used in the previous subframe. As a result, even when the bit rate mode is switched from moment to moment, the bit rate can be switched frame by frame without interruption of sound.

When switching the bit rate mode, LS
The P codebook is the same in all bit rate modes, and the weight codebook that represents the gain for the fixed codebook and the adaptive codebook is the MA (moving average) prediction coefficient and weight of the fixed codebook. Only the bias (average value) is switched, and the same weight codebook is used. Here, the weight codebook may be created independently for each bit rate mode and used by switching. The weight codebook,
The drive sound source vector generation units 5-12 to 5-5 regardless of the length of the subframe, that is, the bit rate mode switching.
14 can be made common to the adaptive codebook code vector, the fixed codebook code vector and the driving excitation vector candidate size ratio even if the subframe length changes. It is because there is such a relationship.

In the bit rate switching method according to the present invention, as long as the adaptive codebook can be shared, the bitrate mode can be constructed using fixed codebooks of various systems. For example, Dual-Pulse CS
-CELP, PSI-CELP, MP (Multi
Pulse) CELP and other fixed codebooks based on CELP may be used to configure the bit rate mode. Even in that case, the already transmitted information can be effectively used, and the bit rate mode can be switched without interruption of sound.

In the second embodiment, the read-ahead is about 7 ms, that is, about 1/3 of the frame length regardless of the bit rate mode. Therefore, by switching the bit rate mode, the subframe is about 5 ms, about 7 ms, about 1
Although it is switched to any of 0 ms, the delay for obtaining a code is always constant at 1 frame + 7 ms. That is, since the LSP coefficient is the analysis result of the first subframe of the adjacent frame and interpolates other subframes between them, it is originally 1
Frame + 1 subframe, delayed by L for each frame
Although the SP quantized code is obtained, the LSP analysis result obtained by the 7 ms prefetch is always used as the LSP analysis result for one subframe that is prefetched at the time of mode switching.
Therefore, the delay amount for obtaining the coded result is constant regardless of the mode, it is possible to prevent the communication from being temporarily interrupted, and the delay amount does not have to be large.

As shown in FIG. 2, the functional structure of the decoding method of the present invention is almost the same as that of the conventional method shown in FIG. However, when switching the bit rate mode, a plurality of driving sound source vector generation units 5 are provided corresponding to each mode.
-12 to 5-14 are configured as the same as those of the encoder. The codebook used for decoding the code (index) of the short-term prediction component, in the embodiment, the LSP codebook is 3.4 kH.
The quantization distortion is larger than that of a codebook used for decoding a code (index) having the same number of bits in the z band. That is, when the index of the short-term prediction component has the same number of bits, the LS of the LSP codebook of this 5 kHz band is higher than the order of the LSP coefficient of the LSP codebook of 3.4 kHz band.
The order of the P coefficient is high.

Further, in the present invention, the adaptive codebook used for decoding the code (index) of the periodic component is 3.
It is periodic with the same time accuracy as the adaptive codebook used for decoding the periodic component code (index) of the same number of bits in the 4 kHz band. In other words, as described in the embodiment of the coding method, the value obtained by dividing the number of bits of the periodic component code (index) by the number of samples of one code vector of the adaptive codebook has the same number of index bits. In this case, the 5 kHz band of the present invention is smaller than that of the 3.4 kHz band.

In the encoding method of the present invention, the voice band is limited to approximately 5 kHz, but if it is smaller than 4.5 kHz band, the naturalness of the original sound deteriorates to the 3.4 kHz band to the same extent and the naturalness of the original sound deteriorates. I can't get the excellent playback sound. On the other hand, if the band is greater than 5.5 kHz,
A component other than the periodicity, which is a feature of speech, is included, and unless the number of bits in each of the LSP codebook and the adaptive codebook is considerably increased, the performance is significantly deteriorated. In short, by setting the frequency range from 4.5 kHz to 5.5 kHz, the naturalness of the original sound is remarkably enhanced compared to the 3.4 kHz band, and the periodicity which is a characteristic of voice is used to obtain 3.4 kHz.
With bit allocation similar to Hz band, short-term prediction component,
The periodic components can be encoded while maintaining high quality.

[0050]

In order to clarify the effect of the present invention,
A MOS evaluation test using the encoding method shown in the embodiment was conducted. The following was used as the encoding system set to be evaluated. Frequency band 7kHz, 5kHz, 3.4
Original sound and MNRU for each kHz audio signal
(Voice with amplitude correlation noise) 40 dB, 30 dB, 20 d
B, 10 dB.

As a voice coding system in the 3.4 kHz band, G. 723.1 (6.3 kbit / s, 5.3 kbit
/ S), G.I. 729 (8 kbit / s). G. 7 kHz band. 722 (64 kbit / s, 56 kbit / s,
48 kbit / s). As a reference value when a sound of 11 kHz sampling is input to the existing 3.4 kHz band encoding method, G. G.72 which inputs the voice of 11.25 kHz sampling into the 5.3kbit / s mode and the 6.3kbit / s mode of 723.1. 723.1Base1
1 kHz 7.29 bit / s, 8.68 kbit / s.

As the coding system of the 5 kHz band according to the present invention, the 7.8 kbit / s coding system of the first embodiment and the 5.75 kbit / s coding system of the second embodiment. For each of the above, a total of 14 voices for male and female voices were used for evaluation, and 16 subjects performed a MOS evaluation test with 5 grade absolute evaluation. The evaluation test results are shown in FIG.

From the result of the experiment, 8k in the 3.4kHz band
G. Hz sampling voice as input 723.1
The MOS evaluation values of 5.3 kbit / s and 6.3 kbit / s are 2.7589 and 2.8884, respectively.
The evaluation value of 729 was 3.2054. 11 kHz for a coding scheme designed for 8 kHz sampling
G723.1Base1 that input sampling voice
The values were 2.9464 and 3.0402 at 1 kHz 7.29 bit / s and 8.68 kbit / s, respectively.

On the other hand, in the 7.8 kbit / s encoding system of the first embodiment using the present invention, the MOS evaluation value is 3.
442 and 3.2902 in the 5.75 kbit / s mode of Example 2. From these results, the existing 3.
The MOS evaluation value of the present invention, which is specialized in the band of about 5 kHz, is more suitable than the coding system of the 4 kHz band or the existing coding system of the 3.4 kHz band in which a signal of 11.25 kHz sampling is simply input. Was significantly higher.

In the present invention, the frequency band of the input voice is set to 5
By specializing in about kHz, it is possible to perform encoding at an overwhelmingly lower bit rate than the conventional encoding method for 7 kHz band input speech, and encoding for 3.4 kHz band input speech. Compared to the system, it is possible to play overwhelmingly natural sound at the same low bit rate. In addition, since the input voice is set wider than the band of about 5 kHz and the band of 3.4 kHz, and the performance of the linear predictive quantization codebook and the adaptive codebook is improved as described above, the bit rate is maintained while maintaining a relatively high quality. Can be variable.

[Brief description of drawings]

FIG. 1 is a diagram showing a functional configuration example of an encoding method according to the present invention.

FIG. 2 is a diagram showing a functional configuration example of a decoding method according to the present invention.

FIG. 3 is a diagram showing an example of bit allocation in the 7.8 kbit / s encoder of the first embodiment.

FIG. 4 is a diagram showing an example of pulse arrangement in the 7.8 kbit / s encoder of the first embodiment.

FIG. 5 is a diagram showing an example of bit allocation in each bit rate mode of the encoder according to the second embodiment.

FIG. 6 is 7.75 kbit / sM of the encoder according to the second embodiment.
The figure which shows the example of the pulse arrangement in ODE.

FIG. 7 is 5.75 kbit / sM of the encoder according to the second embodiment.
The figure which shows the example of the pulse arrangement in ODE.

FIG. 8 is 3.95 kbit / sM of the encoder according to the second embodiment.
The figure which shows the example of the pulse arrangement in ODE.

FIG. 9 is a diagram showing a MOS evaluation test result.

FIG. 10 is a diagram showing a functional configuration of a conventional encoder.

FIG. 11 is a diagram showing a conventional driving sound source vector generation unit.

FIG. 12 is a diagram showing a configuration of a conventional synthetic distortion calculation method.

FIG. 13 is a diagram showing a functional configuration of a conventional decoder.

─────────────────────────────────────────────────── ─── Continuation of the front page (56) Reference JP-A-7-36495 (JP, A) JP-A-6-153119 (JP, A) JP-A-5-199071 (JP, A) JP-A-6- 124100 (JP, A) (58) Fields surveyed (Int.Cl. ⁷ , DB name) G10L 19/00 G10L 19/04

Claims (10)

(57) [Claims] 1. A short-term prediction of an input speech signal for each frame
Component and its prediction residual component, and the prediction residual component
To a method of encoding using the adaptive codebook and the fixed codebook.
Be careful Limit the input audio signal to the band of about 5kHzThen When the above short-term prediction components are coded with the same bit allocation
Has a larger quantization distortion than the encoding in the 3.4 kHz band
To be Encoding using the adaptive codebook is performed in the 3.4 kHz band.
Fewer bits allocated to one sample than encoding
It is characterized by doingSoundVoice signal coding method. 2. A coding method according to claim 1 Symbol placement, by switching the fixed code book, the speech signal encoding method characterized by the bit rate varying. 3. The encoding method according to claim 2 , wherein when the bit rate is switched, the adaptive codebook immediately before switching is used without re-initializing the adaptive codebook. . 4. The encoding method according to claim 2 or 3 , wherein the frame length is constant at all bit rates,
A voice encoding method characterized in that the number of subframes for each frame is changed according to a bit rate. 5. A speech signal is divided into a short-term prediction component for each frame and its prediction residual component, and the prediction residual component is
A method of decoding a code encoded by a method of separately encoding a periodic component longer than a frame and the remaining component, wherein the code of the short-term prediction component is the same number of bits in the 3.4 kHz band. Decoding is performed using a codebook that has a larger quantization distortion than the codebook used to decode the code, and the code of the periodic component is of the same time as the code of the same number of bits in the 3.4 kHz band. A method for decoding a voice signal, characterized in that the adaptive codebook is cyclicized with high accuracy and then decoded. 6. An encoding method in which an input speech signal is divided into a short-term prediction component for each frame and a prediction residual component thereof, and the prediction residual component is encoded using an adaptive codebook and a fixed codebook.
In vessels, the filter for limiting the input audio signal in the band of about 5kHz
Comprising a part, if the number of same bits are allocated, the short-term prediction
The linear prediction analysis order of the components of the encoding in the 3.4 kHz band
Is it larger than, the encoding using the adaptive codebook, than the encoding of 3.4kHz band, is it small number of bits allocated to one sample
Rot characteristics and be Ruoto voice signal encoder that is. 7. The speech signal encoder according to claim 6, wherein a plurality of fixed code vectors in which the number of stored fixed code vectors is different.
A constant codebook, the speech signal encoder characterized in that it comprises a vector switching unit using these fixed codebook, the <br/> re coded cut by the bit rate control signal. 8. The encoder according to claim 7 , wherein the fixed codebook is switched by a bit rate control signal.
Was sometimes not performed reinitialize the adaptive codebook, the speech signal encoder according to claim Rukoto adaptive codebook immediately before the switching is utilized. 9. Te encoder smell <br/> according to claim 7, against the all bit rates, the frame length is constant,
The higher the bit rate, the number of subframes per frame
There speech coder, characterized in that there is a large. 10. A speech signal is divided into a short-term prediction component for each frame and its prediction residual component, and the prediction residual component is divided into a periodic component longer than a frame and its remaining component and coded. a decoder of an encoded code by the method of reduction, the sign of the short-term prediction component, linear prediction order <br/> than the codebook used for decoding the codes of the same number of bits of 3.4kHz band It is decoded using the code book entry size, the sign of the periodic components, the decoding of the codes of the same number of bits of 3.4kHz band, the adaptive codebook is circumferential <br/> initialized at time system comparable Te, the audio signal decoding, wherein Rukoto decoded
Bowl .