JP5174651B2 - Low complexity code-excited linear predictive coding - Google Patents

Abstract

Information (km) about excitation signals of a first signal (sm(n)) encoded by CELP is used to derive a limited set (10') of candidate excitation signals for a second correlated second signal (ss(n)). Preferably, pulse locations of the excitation signals of the first encoded signal (sm(n)) are used for determining the set (10') of candidate excitation signals. More preferably, the pulse locations of the set of candidate excitation signals are positioned in the vicinity of the pulse locations of the excitation signals of the first encoded signal (sm(n)). The first and second signals (sm(n), ss(n)) may be multi-channel signals of a common speech or audio signal. However, the first and second signals (sm(n), ss(n)) may also be identical, whereby the coding of the second signal (ss(n)) can be utilized for re-encoding at a lower bit rate.

Description

  The present invention relates to audio coding, and in particular to code-excited linear predictive coding.

  Existing stereo coding techniques, and more generally multi-channel coding techniques, require very high bit rates. Parametric stereo is often used at very low bit rates. However, these technologies are designed for various types of common audio material: music, voice and mixed content.

  Almost nothing has been done in multi-channel speech coding. Most research has focused on inter-channel prediction (ICP) methods. ICP technology takes advantage of the correlation between the left and right channels. Various ways to reduce this redundancy in stereo signals are described in documents such as [1] [2] [3].

  The ICP method models very well when there is only one speaker, but it cannot model multiple speakers or diffuse sound sources (for example, background noise diffusion). Therefore, the encoding of the ICP residual may be indispensable, in which case a very high bit rate is required.

  Most existing speech codecs are monaural and are based on a code excited linear predictive (CELP) coding model. Examples include AMR-NB (Adaptive Multi-Rate Narrow Band) and AMR-WB (Adaptive Multi-Rate Wide Band). In this model, ie CELP, the excitation signal at the input of the short-term linear prediction synthesis filter is constructed by adding two excitation vectors from the adaptive codebook and the fixed codebook (innovation codebook). The speech is synthesized by supplying two appropriately selected vectors from those codebooks via a short-term synthesis filter. The optimal excitation sequence of the codebook is selected using the AbS search (Analysis-by-Synthesis search) search procedure. In this procedure, the error between the current speech and the synthesized speech is minimized according to the perceptual weighting distortion criterion.

  There are two types of fixed codebooks. The first type of codebook is a so-called stochastic codebook. Probabilistic codebooks often require a large amount of storage capacity. Given an index into the codebook, the excitation vector is obtained by a conventional table lookup. Therefore, the size of the code book is limited by the bit rate and the calculation amount.

  The second type of codebook is an algebraic codbook. In contrast to probability codebooks, algebraic codebooks are not random and do not require substantial storage. An algebraic codebook is a set of indexed code vectors in which the amplitude and position of the pulses making up the kth code vector are derived directly from the corresponding index k. This has virtually no memory requirement. Therefore, the size of the algebraic codebook is not limited by the memory capacity. Algebraic codebooks are also very suitable for efficient search procedures.

  What is important is that a significant portion of the bits available to the speech codec and in most cases the majority are allocated to fixed codebook excitation coding. For example, in the AMR-WB standard, the amount of bits allocated to fixed codebook procedures is between 36% and 76%. In addition, this is a fixed codebook excitation search that represents the majority of the computational effort of the encoder.

  In [7], a multipart fixed codebook including an individual fixed codebook for each channel and a common codebook common to all channels is used. According to this method, an appropriate expression of the correlation between channels can be obtained. However, this increases the storage capacity and the calculation amount. Furthermore, since it is necessary to transmit a common codebook index in addition to each channel codebook index, the bit rate required to encode the fixed codebook excitation is very high. In [8] and [9], a similar method for encoding multi-channel signals is described. In these methods, the coding mode depends on the correlation of different channels. These techniques are well known from Left / Right and Mid / Side coding, and the switching between the two coding modes depends on the residual and thus on the correlation.

  [10] describes a method of encoding a multi-channel signal. The method generalizes the various elements of a single channel linear prediction codec. This method has the disadvantage of requiring a large amount of computation that renders the method unusable in real-time applications such as conversational applications. Another drawback of this technique is the amount of bits required to encode the various inverse correlation filters used for encoding.

[1] H. Fuchs, "Improving joint stereo audio coding by adaptive inter-channel prediction", in Proc. IEEE WASPAA, Mohonk, NY, October 1993 [2] S.A. Ramprashad, "Stereophonic CELP coding using cross channel prediction", in Proc. IEEE Workshop Speech Coding, pp. 136-138, September 2000 [3] T. Liebschen, "Lossless audio coding using adaptive multichannel prediction", in Proc. AES 113th Conv., Los Angeles, CA, October 2002 [4] ITU-R BS.1387 [5] International publication WO96 / 28810 [6] 3GPP TS 26.190, p. 28, table 7 [7] US Patent Application Publication 2004/0044524 [8] US Patent Application Publication 2004/0109471 [9] US Patent Application Publication 2003/0191635 [10] US Patent 6,393,392

  A general problem of conventional speech coding is a high bit rate required and a large amount of calculation of the encoder.

  It is an object of the present invention to improve the method and apparatus for speech coding. A secondary object of the present invention is to reduce the bit rate and the amount of calculation in the CELP encoding method and apparatus.

  The above objective is accomplished by a method and apparatus according to the disclosed claims. In general, the excitation signal of the first signal encoded by CELP is used to derive a limited set of candidate excitation signals for the second signal. The second signal is preferably related to the first signal. In certain embodiments, the limited set of candidate excitation signals is derived by one rule selected from a predetermined set of rules based on the encoded first signal and / or second signal. The pulse position of the excitation signal of the first encoded signal is preferably used to determine a set of candidate excitation signals. More preferably, the pulse positions of the set of candidate excitation signals are positioned close to the pulse positions of the excitation signals of the first encoded signal. The first signal and the second signal can be a multi-channel signal of a common voice or audio signal. However, the first signal and the second signal may be the same so that the encoding of the second signal is utilized to re-encode at a lower bit rate.

  One advantage of the present invention is that the amount of encoding computation can be reduced. Furthermore, in the case of multi-channel signals, the bit rate required to transmit the encoded signal is reduced. The present invention can be efficiently applied to re-encoding the same signal at a lower rate. Another advantage of the present invention is that it is compatible with monaural signals and can be implemented as an extension function of an existing audio codec with little change.

A general CELP speech synthesis model is shown in FIG. 1A. Fixed codebook 10 includes a plurality of candidate excitation signals 30, each characterized by an index k. In the case of an algebraic codebook, only the index k is used to fully characterize the corresponding candidate excitation signal 30. Each candidate excitation signal 30 includes a plurality of pulses 32 having a particular position and amplitude. The index k determines the candidate excitation signal 30 that is amplified in the amplifier 11 that outputs the output excitation signal c _k (n), corresponding to reference numeral 12. Although not the first subject of the present invention, the adaptive codebook 14 outputs an adaptive excitation signal v (n) via an amplifier 15. The adder 17 adds the excitation signal c _k (n) and the adaptive excitation signal v (n), and outputs the addition excitation signal u (n). As indicated by the dashed line 13, the additive excitation signal u (n) affects the adaptive codebook for subsequent signals.

  The additive excitation signal u (n) is used as an input signal for the transformation 1 / A (z) of the linear prediction synthesis unit 20, and a “prediction” signal ^ s (n) indicated by 21 is obtained as a result. This prediction signal ^ s (n) is usually an output signal of the CELP synthesis procedure after post-processing 22.

A CELP speech synthesis model is used in AbS encoding of the speech signal of interest. A target signal s (n), ie a signal to be approximated, is supplied. Long-term prediction using an adaptive codebook is performed, adjusting the past coding for the current target signal and outputting an adaptive excitation signal v (n) = g _p u (n-δ) . It is necessary to find the codebook index k corresponding to the entry c _k that uses the residual error as a target for determining the fixed codebook excitation signal and minimizes the difference according to an objective function such as a mean square criterion. In general, the algebraic codebook is searched by minimizing the mean square error between the weighted input speech and the weighted synthesized speech. The purpose of the fixed codebook search is to find the algebraic codebook entry c _k corresponding to index k so that the following expression is maximized.

The matrix H is a filtering matrix whose elements are derived from the impulse response of the weighting filter. y ₂ is a vector of components depending on the signal to be encoded.

This fixed codebook procedure is shown in FIG. 1B. In FIG. 1B, index k selects entry c _k from fixed codebook 10 as excitation signal 12. In a stochastic fixed codebook, index k usually serves as an input for table lookup. On the other hand, in the algebraic fixed codebook, the excitation signal 12 is directly derived from the index k. In general, multipulse excitation can be written as:

Here, p _{i, k} is the pulse position with respect to the index k, b _{i, k} is the individual pulse amplitude, P is the number of pulses, and δ is the following Dirac pulse function.

  FIG. 1C is a diagram showing an example of candidate excitation signal 30 of fixed codebook 10. The candidate excitation signal 30 is characterized by a plurality of pulses 32, in this example eight pulses. Pulse 32 is characterized by pulse positions P (1) -P (8) and their amplitudes. In a normal algebraic fixed codebook, the amplitude is +1 or -1.

  In a single channel encoder / decoder system, the CELP model is typically implemented as shown in FIG. Each part corresponding to each function of the CELP synthesis model of FIG. 1A is characterized by the functions of those parts, although not substantially at the same level in an actual implementation. It is attached. For example, error weighting filters that normally exist in actual implementations of linear prediction A-b-S are not displayed.

An encoding target signal s (n) indicated by 33 is supplied to the encoder 40. The encoder includes a CELP synthesis block 25 that follows the principles described above (postprocessing is omitted for clarity of illustration). The output from CELP synthesis block 25 is compared with signal s (n) in comparator 31. The The difference 37 weighted by the weighting filter is fed to the codebook optimization block 35, which finds the optimal or at least approximately appropriate excitation signal c _k (n), indicated at 12. Constructed according to conventional principles. The codebook optimization block 35 supplies the corresponding index k to the fixed codebook 10. When the final excitation signal is found, index k and delay δ of adaptive codebook 12 are encoded by index encoder 38 to provide an output signal representing index k and delay δ.

  The representation of the index k and the delay δ is supplied to the decoder 50. The decoder includes a CELP synthesis block 25 that follows the principles described above. (Post-processing is omitted for clarity of illustration.) The representation of index k and delay δ is decoded in index decoder 53, where index k and delay δ are fixed codebook and adaptive as input parameters. Supplied to each codebook. As a result, a composite signal ^ s (n) indicated by 21 that will be similar to the original signal s (n) is obtained.

  The representation of the index k and the delay δ is stored anywhere between the encoder and decoder for a short or long period. Thereby, for example, the storage capacity required for storing audio recordings becomes relatively small.

The present invention relates to speech coding, and more generally to audio coding. In a typical example, the present invention addresses the case where it is desirable to encode the main signal s _M (n) according to the CELP technique and to encode another signal s _S (n). The other signal may be, for example, the same main signal s _S (n) = s _M (n) in re-encoding at a low bit rate, or the encoded version of the main signal s _S (n) = ^ s _M (n) or a signal corresponding to another channel such as stereo or multi-channel 5.1.

  Thus, the present invention is directly applicable to audio stereo coding, and more generally multi-channel coding, in teleconferencing applications. Applications of the invention can include audio coding as part of open loop content dependent coding or closed loop content dependent coding.

In order for the present invention to operate under optimal conditions, it is preferred that there be a correlation between the main signal and other signals. However, the existence of such correlation is not an essential condition for the proper operation of the present invention. In practice, the present invention operates adaptively and is implemented depending on the degree of correlation between the main signal and other signals. In stereo applications, there is a causal relationship between the left and right channels, so in many cases the sum signal is the main signal s _M (n) and the difference signal between the left and right channels is s _S (n) Will be.

The assumption of the present invention is that the main signal s _M (n) is available in the CELP coded representation. One basic concept of the invention is to limit the search in the fixed codebook during encoding of the other signal s _S (n) to a subset of candidate excitation signals. This subset is selected depending on the CELP coding of the main signal. In a preferred embodiment, the subset of candidate excitation signal pulses is limited to a set of pulse positions that depend on the pulse positions of the main signal. This corresponds to defining constrained candidate pulse positions. Usually, the set of available pulse positions is set to the pulse position of the main signal and the adjacent pulse positions.

  By reducing the number of candidate pulses, the calculation amount of the encoder can be greatly reduced.

  The following is a general example of two channel signals. However, this is easily extended to further multi-channels. However, in the case of multi-channel, a different weighting filter is given to each channel and the target is different, and the targets of each channel are delayed from each other.

  The main channel and the sub channel are constituted by the following equations.

However, s _L (n) and s _R (n) are inputs for the left channel and the right channel, respectively. Even though the left and right channels are delayed versions of each other, this is generally not the case with the main and subchannels because the main and subchannels contain information from both channels.

  In the following, it is assumed that the main channel is the first coding channel and the pulse position of the fixed codebook excitation for that coding is available.

  The target of the sub-signal fixed codebook excitation encoding is calculated as the difference between the sub-signal and the adaptive codebook excitation.

Here, g _P v (n) is adaptive codebook excitation, and s _C (n) is a target signal for adaptive codebook search.

  In the present embodiment, the number of pulse positions that can be taken by the candidate excitation signal is defined depending on the pulse position of the main signal. Since those pulse positions are just a few of all possible positions, the amount of bits required to encode a sub-signal that contains excitation signals in a limited set of candidate excitation signals is all Compared with the case where the pulse position of 1 appears, it is greatly reduced.

  Selecting the pulse candidate position with respect to the main pulse position is important in determining the calculation amount and the required bit rate.

  For example, if the frame length is L and the number of pulses in main signal encoding is N, approximately N * log2 (L) bits are required to encode the pulse position. However, if only the main signal pulse position is held as a candidate and the number of pulses of the candidate excitation signal for the sub signal is P, about P * log2 (N) bits are required to encode the sub signal. This means that if the number of N, P, and L is appropriate, the required bit rate is greatly reduced.

  One interesting aspect is when the pulse position for the sub-signal is set to be the same as the pulse position of the main signal. No encoding of the pulse position is necessary, only the encoding of the pulse amplitude is necessary. In the case of an algebraic codebook including a pulse having an amplitude of + 1 / -1, only the code (N bits) needs to be encoded.

Here, the main signal pulse positions are denoted by P _M (i), i = 1,. The pulse position of the candidate excitation signal for the sub-signal is selected based on the main signal pulse position and possible additional parameters. Additional parameters may include time delay between the two channels and / or adaptive codebook index differences.

  In the present embodiment, the set of pulse positions for the sub-signal candidate excitation signal is configured as follows.

Here, J (i, k) represents a delay index. This means that each pulse position generates a set of pulse positions that are used to construct the candidate excitation signal in the sub-signal pulse search procedure. This is shown in FIG. Here, P _M represents the pulse position of the excitation signal of the main signal, P ^* _S shows the Possible pulse position candidate excitation signal in the side signal analysis.

  Of course, this is optimal for highly correlated signals. The reverse strategy is taken for signals with low or no correlation. This means that pulse candidates are all pulses that do not belong to the set of the following equations.

  Since this is a complementary example, those skilled in the art will readily understand that both strategies are similar. Only correlated examples will be described in more detail.

  It will be readily understood that the position and number of pulse candidates depends on the delay index J (i, k). The delay index may depend on the actual delay between the two channels and / or the adaptive codebook index. In FIG. 3 (a), k max = 3 and J (i, k) = J (k) ∈ {-1, 0, +1}.

  In FIG. 3 (b), another slightly different pulse position is selected. Here, k max = 3, but J (i, k) = J (k) ∈ {0, +1, +2}.

  Those skilled in the art will recognize that various methods can be used for the rules of the pulse position selection method. The actual rules used should be adapted to the actual implementation. However, an important feature is that the pulse position candidates are selected depending on the resulting pulse positions as a result of the main signal analysis according to certain rules. This rule may be fixed uniquely, or may be selected from a predetermined set of rules that depend on the correlation between the two channels and / or the delay between the two channels.

  Depending on the rules used, a set of sub-signal pulse candidates is constructed. In general, the set of sub-signal pulse candidates is very small compared to the entire frame length. Thereby, the objective maximization problem based on the thinned frame can be reformulated.

  In a general example, pulses are searched using, for example, a depth-first algorithm as described in [5], or using an exhaustive search if the number of candidate pulses is very small. However, it is recommended to use a fast search procedure even if the number of candidates is small.

  In general, the backward filter signal is pre-calculated using the following equation:

は、h(n)(重み付けフィルタのインパルス応答)の相関の行列であり、その行列の要素は以下の式により計算される。 matrix

Is a matrix of correlation of h (n) (weighting filter impulse response), and the elements of the matrix are calculated by the following equations.

Therefore, the objective function can be written as

  Assuming a set of candidate pulse positions that the sub-signal can take, only a subset of the indices of the backward filtering vector d and the matrix Φ are required. The set of candidate pulses is sorted in ascending order.

P ^* _S (i) is a candidate pulse position, and p is the number thereof. Note that p is always smaller than the frame length L and is usually sufficiently small.

  The thinned signal is expressed by the following equation.

Further, the decimation correlation matrix Φ ₂ is expressed by the following equation.

In this case, Φ ₂ is a symmetric matrix and is positive definite. It can also be written directly as follows:

Here, c ′ _{k ′} is a new algebraic code vector. The index becomes k ′, which is a new entry into the reduced size codebook.

  An outline of these thinning operations is shown in FIG. The upper part of the figure shows that the original size algebraic codebook 10 is reduced to a reduced size codebook 10 '. In the middle stage, it is shown that the original size of the weighting filter covariance matrix 60 is reduced to a reduced weighting filter covariance matrix 60 ′. Also, the lower part shows that the original size of the backward filtered target 62 is reduced to the reduced size of the backward filtered target 62 ′. Those skilled in the art will appreciate that the amount of computation is reduced as a result of such reduction.

Maximizing the objective function of the decimation signal has several advantages. One of them is that the required memory capacity can be reduced. For example, the memory capacity required for the matrix Φ ₂ can be reduced. Another advantage is that the decoder can always use the index of the decimation signal because the pulse position of the main signal is transmitted to the receiver in any case. Thereby, the pulse position of another signal (sub) can be encoded with respect to the pulse position of the main signal, and this uses very few bits. Another advantage is that the amount of computation is reduced because maximization is performed on the decimation signal.

FIG. 5A shows an embodiment of a system of encoders 40A, 40B and decoders 50A, 50B according to the present invention. Many of the details are the same as those in FIG. 2, and therefore, detailed descriptions of the parts whose functions are not essentially changed are omitted. The main signal s _M (n) indicated by 33A is supplied to the first encoder 40A. The first encoder 40A operates according to any conventional CELP coding model, to generate a delayed reference [delta] _m for index k _m and adaptive codebook for a fixed codebook. Details of this encoding are not an important part of the present invention, and will be omitted for easy understanding of FIG. 5A. Parameter k _m and [delta] _m is encoded in the first index encoder 38A, giving a representation k ^* _m and [delta] ^* _m of parameters to be sent to the first decoder 50A. In the first decoder, the expression k ^* _m and [delta] ^* _m are decoded into parameters k _m and [delta] _m in the first index decoder 53A. The original signal is regenerated from those parameters according to any CELP decoding model according to the prior art. Details of this decoding are not important parts of the present invention, and will be omitted to facilitate understanding of FIG. 5A. The regenerated first output signal ^ s _m (n) indicated by 21A is output.

The sub signal s _S (n) indicated by 33B is supplied as an input signal to the second encoder 40B. Most of the second encoder 40B is the same as the encoder of FIG. Here the signal is given an index “s” to distinguish it from the signal used to encode the main signal. The second encoder 40B includes a CELP synthesis block 25. According to the present invention, the expression of the index k _m or index, is supplied from the first encoder 40A to the input 45 of the fixed codebook 10 of the second encoder 40B. Index k _m, in order to extract reduced fixed codebook 10 'according to the principles given above, is used by the candidate deriving means 47. The synthesis of the CELP synthesis block 25 ′ of the second encoder 40B is based on the index k ′ _S representing the excitation signal c ′ _k ′ _S (n) from the reduced fixed codebook 10 ′. The index k ′ _S is found to represent the optimal choice of CELP synthesis. The parameters k ′ _S and δ _S are encoded in the second index encoder 38B and give the parameter representations k ′ ^* _S and δ ^* _S sent to the second decoder 50B.

In the second decoder 50B, the representations k ′ ^* _S and δ ^* _S are decoded into parameters k ′ _S and δ _S by the second index decoder 53B. Furthermore, in order to allow extraction of the second candidate deriving means 57 of those used in the encoder 40B equivalent reduced fixed codebook 10 ', the index parameter k _m, obtained from the first decoder 50A And is fed to the input of the fixed codebook 10 of the second decoder 50B. The original sub-signal is regenerated from the parameters k ′ _S and δ _S and the reduced fixed codebook 10 ′ according to the normal CELP decoding model 25 ″. Details of this decoding are essentially shown in FIG. Performed in the same way as 2, but using the reduced fixed codebook 10 ', in this way, the regenerated output sub-signal ^ s _S (n), denoted 21B, is output.

  The selection of rules for constructing a set of candidate pulses, such as an indexing function J (i, k), is advantageously adapted and depends on additional interchannel characteristics such as delay parameters, degree of correlation and the like. In this case, i.e. in the selection of adapted rules, the encoder preferably sends the selected rules to the decoder to derive a set of candidate pulses for encoding other signals. The selection of the rule is executed by, for example, a closed loop procedure. Here, a plurality of rules are tested, and a rule that finally gives an optimal result is selected.

FIG. 5B illustrates one embodiment using a rule selection method. The mono signal s _M (n) and preferably the sub-signal s _S (n) are also provided to the rule selection unit 39 in this embodiment. Instead of the mono signal may be used a parameter k _m representing the mono signal. In the rule selection unit 39, the signal is analyzed for delay parameters or correlations, for example. Based on the result, for example, the rule represented by the index r is selected from a set of predetermined rules. The index of the selected rule is provided to candidate derivation means 47 that determines how to derive the candidate set. The rule index r is further provided to a second index encoder 38B, which provides an index representation r ^* . The representation r ^* is then sent to the second decoder 50B. The second index decoder 53B decodes the rule index r, which is used to manage the operation of the candidate derivation means 57.

  In this way, a set of rules suitable for different types of signals is provided. Additional flexibility is achieved by simply adding a single rule index during the transfer of data.

  The particular rule used and the resulting number of candidate sub-signal pulses are the main parameters that determine the bit rate and the computational complexity of the algorithm.

As mentioned above, the exact same principle applies equally well to re-encoding of the same channel. FIG. 6 shows an embodiment in which different parts of the transmission path allow different bit rates. This is applicable as part of the rate transcoding scheme. The signal s (n) is provided as the input signal 33A to the first encoder 40A. The first encoder 40A generates parameter representations k ^* and δ ^* to be transmitted according to the first bit rate. In certain locations, the available bit rate is reduced and re-encoding needs to be performed for lower bit rates. The first decoder 50A uses the parameter representations k ^* and δ ^* to generate the regenerated signal ^ s (n) denoted 21A. The regeneration signal ^ s (n) indicated by 21A is provided to the second encoder 40B as the input signal 33B. The index k from the first decoder 50A is provided to the second encoder 40B. The index k is similar to FIG. 6 used to extract the reduced fixed codebook 10 ′. The second encoder 40B encodes the signal ^ s (n) for the low bit rate and provides an index ^ k 'representing the selected excitation signal c' _{^ k '} (n). However, this index ^ k 'is rarely used in the decoder because the remote decoder does not have the information necessary to construct the corresponding reduced fixed codebook. This index ^ k 'needs to be associated with an index ^ k that references the original codebook 10. This is preferably performed in connection with the fixed codebook 10 and is represented in FIG. 6 by an arrow 41 indicating the input of ^ k 'and an arrow 43 indicating the output of ^ k. The encoding of the index ^ k is performed with reference to the entire set of candidate excitation signals.

  In a typical example, the first encoding is performed at a bit rate n, and the second encoding is performed at a bit rate m. However, n> m.

Send live content over different types of networks with different capabilities in real time (eg, in certain applications such as teleconferencing (eg, at several different bit rates to accommodate different types of networks) If real-time coding of the same signal, so-called parallel multi-rate coding, is needed, it would be interesting to provide parallel coding at different bit rates. and shows a system to be supplied to the second encoder 40B similarly to the. previous embodiments, the second encoder, the fixed codebook 10 'that are reduced on the basis of the index k _a representative of the first encoding The second encoding is indicated by the index “b.” The second encoder 40B is independent of the first decoder 50A. To become. Most other parts is the same as FIG. 6, it includes an adaptive indexing.

  For those two applications that re-encode the same signal at a low rate, the present invention inherently reduces the amount of computation and allows the applications to be implemented with low-cost hardware.

  One embodiment of the above algorithm was implemented in connection with the AMR-WB speech codec. To encode the side signal, the same adaptive codebook index is used as was used to encode the mono excitation. LTP gain and improved vector gain were not quantized.

  The algorithm for the algebraic codebook was based on monopulse position. As described in [6], the codebook may be composed of tracks. Except for the minimum mode, the number of tracks is four. For each mode, a certain number of pulse positions is used. For example, in the case of mode 5, that is, in the case of 15.85 kbps, the candidate pulse positions are as follows.

  The implemented algorithm holds all monopulses as sub-signal pulse positions. That is, the pulse position is not encoded. Only the sign of the pulse is encoded.

  Each pulse uses only one bit to encode the code. As a result, the total bit rate becomes equal to the number of monopulses. In the above example, there are 12 pulses per subframe, and the total bit rate for encoding the improved vector is equivalent to 12 bits × 4 × 50 = 2.4 kbps. This is the same as the number of bits required for the minimum AMR-WB mode (2 pulses for the 6.6 kbps mode) but in this case has a higher pulse density.

  Note that no additional algorithm delay is required to encode the stereo signal.

  Figure 8 shows the results obtained with PEAQ [4] for assessing perceptual quality. PEAQ was chosen because it is well known and is the only tool that provides an objective quality standard for stereo signals. The result clearly shows that the stereo 100 actually improves the quality compared to the mono signal 102. There are various sound items used, S1 of Sound 1 is extracted from a movie with background noise, S2 of Sound 2 is a one minute radio recording, S3 of Sound 3 is a car racing sport event And S4, which is Sound 4, is an actual 2-microphone recording.

  FIG. 9 is a flowchart showing an embodiment of the encoding method according to the present invention. The procedure starts at step 200. In step 210, a representation of the CELP excitation signal for the first audio signal is provided. Note that it is not necessary to provide the entire first audio signal, only a representation of the CELP excitation signal is provided. In step 212, a second audio signal associated with the first audio signal is provided. In step 214, a set of candidate excitation signals is derived according to the first CELP excitation signal. The pulse position of the candidate excitation signal is preferably related to the pulse position of the CELP excitation signal of the first audio signal. In step 216, CELP encoding is performed on the second audio signal using the reduced set of candidate excitation signals derived in step 214. Finally, a representation of the CELP excitation signal for the second audio signal, usually an index, is encoded using a reference to the reduced candidate set. The procedure ends at step 299.

  FIG. 10 shows another embodiment of the encoding method according to the present invention. The procedure starts at step 200. In step 211, an audio signal is provided. In step 213, a representation of the first CELP excitation signal for the same audio signal is provided. In step 215, a set of candidate excitation signals is derived according to the first CELP excitation signal. The pulse position of the candidate excitation signal is preferably related to the pulse position of the CELP excitation signal of the first audio signal. In step 217, CELP re-encoding is performed on the audio signal using the reduced set of candidate excitation signals derived in step 215. Finally, a representation of the second CELP excitation signal for the audio signal, i.e. usually an index, is encoded using a reference to the unreduced candidate set, i.e. the set used for the first CELP encoding. . The procedure ends at step 299.

  FIG. 11 shows an embodiment of a decoding method according to the present invention. The procedure starts at step 200. In step 210, a representation of the first CELP excitation signal for the first audio signal is provided. In step 252, a representation of the second CELP excitation signal for the second audio signal is provided. In step 254, the second excitation signal is derived from the representation of the second CELP excitation signal using knowledge of the first excitation signal. Preferably, a reduced set of candidate excitation signals is derived according to the first CELP excitation signal, from which the second excitation signal is selected using an index to the second CELP excitation signal. In step 256, the second audio signal is reconstructed using the second excitation signal. The procedure ends at step 299.

  The above-described embodiments are to be understood as several examples of the invention. Those skilled in the art will appreciate that various modifications, combinations and changes can be made to the embodiments without departing from the scope of the invention. In particular, the various partial solutions in the various embodiments can be combined in other configurations that are technically possible. However, the scope of the invention is defined by the appended claims.

  The present invention uses an algebraic codebook and CELP to greatly reduce the complexity (memory and arithmetic operations) when encoding multiple audio channels and to significantly increase the bit rate. Can be reduced.

It is the schematic which shows a code excitation linear prediction model. It is the schematic which shows the process which derives | leads-out an excitation signal. FIG. 6 is a schematic diagram illustrating one embodiment of an excitation signal for use in a code-excited linear prediction model. FIG. 3 is a block diagram illustrating an embodiment of an encoder and decoder according to a code-excited linear prediction model. (A) is a diagram illustrating an embodiment of the principle of selecting a candidate excitation signal according to the present invention, and (b) is a diagram illustrating another embodiment of the principle of selecting a candidate excitation signal according to the present invention. It is a figure explaining that the data amount required by one Embodiment of this invention can be reduced. FIG. 2 is a block diagram illustrating an embodiment of an encoder and decoder for two signals according to the present invention. FIG. 6 is a block diagram illustrating another embodiment of an encoder and decoder for two signals according to the present invention. FIG. 2 is a block diagram illustrating one embodiment of an encoder and decoder for re-encoding a signal in accordance with the present invention. FIG. 2 is a block diagram illustrating one embodiment of an encoder and decoder for parallel encoding signals for different bit rates in accordance with the present invention. FIG. 6 illustrates perceptual quality achieved by an embodiment of the present invention. 3 is a flowchart showing main steps of an embodiment of the encoding method according to the present invention. 7 is a flowchart showing main steps of another embodiment of the encoding method according to the present invention; 3 is a flowchart illustrating main steps of an embodiment of a decoding method according to the present invention.

Claims (34)

A method for encoding an audio signal, comprising:
And supplying express _{(k, k m, k a} ) of the first excitation signal of a code excited linear prediction of the first audio signal (3 3A),
Deriving a candidate excitation signal (c '(n)) of the set (10') based on the previous SL first excitation signal,
Providing a second audio signal (33B) different from the first audio signal (33A);
Performing code-excited linear predictive coding of the second audio signal (3 3B) using the set (10 ') of the candidate excitation signals (c'(n));
A method characterized by comprising: The method according to claim 1, characterized in that the second audio signal (3 3B) is correlated with the first audio signal (3 3A).   The step of deriving the set (10 ′) of the candidate excitation signals (c ′ (n)) includes: 1 from a set of predetermined rules based on at least one of the first excitation signal and the second audio signal. Method according to claim 1 or 2, characterized in that one rule is selected and the set (10 ') of candidate excitation signals (c' (n)) is derived according to the selected rule. The first excitation signal includes n pulse positions (P _M ) out of a set of N possible pulse positions;
The candidate excitation signal (c ′ (n)) has pulse positions (P ^* _s ) only in a subset of the N possible pulse positions;
The subset of pulse positions (P ^* _s ) is selected based on n pulse positions (P _M ) of the first excitation signal. the method of. If j is an index in the interval {i + L, i + K} (where K and L are integers and K> L) and i is an index of the n pulse positions, the subset of the pulse positions pulse position (P ^* _s) the method of claim 4, characterized in that it is located at the position p _j.   6. The method of claim 5, wherein K = 1 and L = -1. 7. The code-excited linear prediction of the second audio signal ( 33B) is performed using a global search in the set of candidate excitation signals (10 '). The method according to claim 1. Encoding the second excitation signal of the code-excited linear prediction of the second audio signal ( 33B) with reference to the set of candidate excitation signals (10 ');
Wherein the representation of the first excitation signal _{(k, k m, k a} ) and supplying a second excitation signal the encoded with,
The method according to claim 1, further comprising: The method according to claim 3 and 8, characterized by further comprising the step of providing data representative of identification information of said representation of the first excitation signal _{(k, k m, k a} ) said selected rule together. Referring to the set (10) of candidate excitation signals having N possible pulse positions, the method further comprises the step of encoding the second excitation signal of the code excitation linear prediction of the second audio signal ( 33B). A method according to any one of claims 1 to 7, characterized in that Said second excitation signal of m (where, m <n) The method according to any one of claims 1 to 1 0, characterized in that it has a pulse positions. A method for decoding an audio signal (33A, 33B) comprising:
And supplying representations _{(k, k m, k a} ) a first excitation signal of a code excited linear prediction of the first audio signal (33A),
Providing a representation (k ′ _s ) of a second excitation signal for code-excited linear prediction of a second audio signal (33B) different from the first audio signal (33A) ,
The second excitation signal is one of a set of candidate excitation signals (10 ′);
The set (10 ′) of candidate excitation signals is based on the first excitation signal;
The method further comprises:
Deriving the second excitation signal (c ′ _{k ′s} (n)) from the representation (k ′ _s ) of the second excitation signal based on information about the set (10 ′) of candidate excitation signals;
Reconstructing the second audio signal (^ s _S (n)) by predictive filtering the second excitation signal (c '_k's(n));
A method characterized by comprising: It said second audio signal (33B) The method of claim 1 2, characterized in that there is a correlation between the first audio signal (33A). The information regarding the set of candidate excitation signals (10 ′) includes identification information of one rule among a set of predetermined rules, and derivation of the set of candidate excitation signals (10 ′) is determined according to the rules. the method according to claim 1 2 or 1 3, characterized in that it is. The first excitation signal includes n pulse positions (P _M ) out of a set of N possible pulse positions;
The candidate excitation signal has pulse positions (P ^* _s ) only in a subset of the N possible pulse positions;
It said subset of pulse position (P ^* _s) is any one of claims 1 2 to 1 4, characterized in that it is selected based on the n pulse position of the first excitation signal (P _M) The method described in 1. If j is an index in the interval {i + L, i + K} (where K and L are integers and K> L) and i is an index of the n pulse positions, the subset of the pulse positions pulse position (P ^* _s) the method of claim 1 5, characterized in that it is located at the position p _j. The method according to claim 16 , wherein K = 1 and L = −1. An encoder (40B) for encoding an audio signal,
Representation of the first excitation signal of a code excited linear prediction of the first audio signal (3 3A) (k, k m, k a) a means for supplying (45),
Representation before Symbol first excitation signal (k, k _m, k _a) is connected to receive, and means (47) for deriving a set (10 ') of the candidate excitation signal based on said first excitation signal,
Means for providing a second audio signal (33B) different from the first audio signal (33A);
Means (25 ′) for performing code-excited linear prediction, connected to receive a representation of the second audio signal ( 33B) and the set of candidate excitation signals (10 ′), the candidate excitation signal Means for performing code-excited linear predictive coding of the second audio signal (3 3B) using a set (10 ′) of:
An encoder (40B) comprising: The encoder according to claim 18 , wherein the second audio signal (3 3B) is correlated with the first audio signal (3 3A). The means (47) for deriving a set (10 ′) of candidate excitation signals selects one rule from a set of predetermined rules based on at least one of the first excitation signal and the second audio signal. The encoder according to claim 18 or 19 , wherein a set (10 ') of the candidate excitation signals (c' (n)) is derived according to the selected rule. The first excitation signal includes n pulse positions (P _M ) out of a set of N possible pulse positions;
The candidate excitation signal has pulse positions (P ^* _s ) only in a subset of the N possible pulse positions;
It said subset of pulse position (P ^* _s) is any one of claims 1 8 to 2 0, characterized in that it is selected based on the n pulse position of the first excitation signal (P _M) Encoder described in. If j is an index in the interval {i + L, i + K} (where K and L are integers and K> L) and i is an index of the n pulse positions, the subset of the pulse positions pulse position (P ^* _s), the encoder according to claim 2 1, characterized in that it is located at the position p _j. K = 1 and encoder of claim 2 2, characterized in that the L = -1. The means (25 ') for performing code-excited linear prediction of the second audio signal ( 33B) performs a global search in the set of candidate excitation signals (10'). 8 or 2 3 encoder according to any one of. Means (38B) for encoding the second excitation signal of the code-excited linear prediction of the second audio signal (33B) with reference to the set of candidate excitation signals (10 ');
Means for supplying a second excitation signal the encoded said representation _{(k, k m, k a} ) with the first excitation signal,
Furthermore encoder according to any one of claims 1 8 to 2 4, characterized in that it comprises a. The representation of the first excitation signal _{(k, k m, k a} ) together with claim 2 0 and 2 5, characterized in that further comprising a means for supplying data representative of the identity of the selected rule Encoder. Means for encoding the second excitation signal of the code excitation linear prediction of the second audio signal (33, 33B) with reference to a set (10) of candidate excitation signals having N possible pulse positions (38B); the encoder according to any one of claims 1 8 to 2 4), characterized in that it further comprises a. Said second excitation signal of m (where, m <n) encoder according to any one of claims 1 8 to 2 7, characterized in that it comprises a pulse positions. A decoder (50B) for decoding an audio signal,
And means for supplying express (k _m) of the first excitation signal of a code excited linear prediction of the first audio signal (33A) (55),
Means (53B) for supplying a second excitation signal representation (k ′ _s ) of a code-excited linear prediction of a second audio signal (33B) different from the first audio signal (33A) ,
The second excitation signal is one of a set of candidate excitation signals (10 ′);
The set of candidate excitation signals (10 ′) is based on the first excitation signal;
The decoder further comprises:
Connected to receive information associated with the representation of the first excitation signal (k _m ) and the representation of the second excitation signal (k ′ _s ) and based on information about the set of candidate excitation signals (10 ′) Means (57) for deriving the second excitation signal (c ′ _{k ′s} (n)) from the representation (k ′ _s ) of the second excitation signal;
Means (25 ") for reconstructing the second audio signal (^ s _S (n)) by predictive filtering the second excitation signal (c '_k's(n));
A decoder (50B) characterized by comprising: 30. Decoder according to claim 29 , wherein the second audio signal (33B) is correlated with the first audio signal (33A). The information regarding the set of candidate excitation signals (10 ′) includes identification information of one rule among a set of predetermined rules, and derivation of the set of candidate excitation signals (10 ′) is determined according to the rules. the decoder of claim 29 or 3 0 characterized in that it is. The first excitation signal includes n pulse positions (P _M ) out of a set of N possible pulse positions;
The candidate excitation signal has pulse positions (P ^* _s ) only in a subset of the N possible pulse positions;
The subset of pulse positions (P ^* _s ) is selected based on n pulse positions (P _M ) of the first excitation signal, according to any one of claims 29 to 31. The decoder described. If j is an index in the interval {i + L, i + K} (where K and L are integers and K> L) and i is an index of the n pulse positions, the subset of the pulse positions pulse position (P ^* _s), the decoder according to claim 3 2, characterized in that it is located at the position p _j. K = 1 and the decoder of claim 3 3, characterized in that the L = -1.

Images