KR20090101730A - Method and apparatus for audio signal compression - Google Patents

Abstract

PURPOSE: An audio signal compressing method and a device thereof for improving the efficiency of decoding as an efficient expression are provided to decode FFT(Fast Fourier Transform domain) domain by compressing radius and phase information for reducing entropy. CONSTITUTION: An audio signal compression method is as follows. A quantizing unit is used for the consisting of the index of positive numbers of the plural number production. The index of positive numbers is mapped with the index couple of the radius and phase(S1240). The radius information and the bit stream including the phase information are produced.

Description

Audio signal compression method and apparatus {METHOD AND APPARATUS FOR AUDIO SIGNAL COMPRESSION}

The present invention relates to a method and apparatus for compressing an audio signal, and more particularly, to an audio signal compression method and apparatus for improving efficiency in encoding with an efficient data representation.

The Trellis Coding Quantization (TCQ) method is used for uniform or Gaussian sources (Marcellin and Fisher, "Trellis-coded quantization of memoryless and Gaussian-Markov sources", IEEE.Trans.on Commun., 1990). Vector quantization (VQ) is followed.

A codeword in TCQ refers to a sequence selected in trellis code. Trellis code is selected from uniform or non-uniform scalar quantization that has an extended reproduction level and is divided into subsets. The Trellis Graph can be generated from a finite state machine (FSM) that determines the path of reproduction levels. The path most suitable for the input vectors is selected as a quantization sequence by a Viterbi decoding algorithm.

TCQ may provide an improvement in coding gain, but in order to reconstruct the output of the de-quantizer to prevent errors from increasing, the TCQ output TCQ level and path must be lossless coded. do.

SUMMARY OF THE INVENTION The present invention has been made to improve the prior art as described above. In order to reduce entropy by a lossless coding method combined with a quantizer, a radius and phase information may be configured and compressed, and the existing audio Unlike the coder, the purpose of the present invention is to improve the efficiency of the coder with efficient data representation in encoding in the FFT domain.

In order to achieve the above object and to solve the problems of the prior art, the method of encoding an audio signal according to an aspect of the present invention comprises the steps of framing the signal, a set consisting of a plurality of integer indices (set) Quantizing the frame using a quantizer, mapping the integer index into an index pair of radius and phase, and a bit stream including the radius information and phase information for generation of Generating a step.

According to another aspect of the present invention, the signal is an inverse filtered residual signal interval generated by linear prediction coding (LPC) of a specific section of an input waveform sample or a specific section of an audio waveform.

According to another aspect of the invention, the quantizer is a Uniform Scalar Quantizer (USQ), Lattice Vector Quantizer (LQ), Non-Uniform Scalar Quantizer (Non-USQ), Vector Quantizer (VQ), and Trellis Coded Quantizer (TCQ) ), And a signed or non-negative integer is output by the quantization.

According to another aspect of the present invention, when the quantizer is a Terrellis Coded Quantizer (TCQ), the integer output from the TCQ is, or mapped from the level and path bits of the TCQ, or Signed indices according to the TCQ state, or non-negative indices according to the TCQ state, mapped from the level and path bits of the TCQ.

According to yet another aspect of the present invention, the radius information value is a second index, which is generated by a squared summation of the index pair and is an enumeration code that provides the radius value to all radius sets. Is mapped to.

According to another aspect of the invention, the phase information value is represented by an enumeration value in the coset table of the radius value, the coset table includes a pair of indices having the same radius value.

According to another aspect of the present invention, the method further includes adjusting the gain coefficient by bit rate control (BRC) to tune the interval of the quantization.

A method of decoding an audio signal includes de-multiplexing a bit stream into a radius information value and a phase information value, reconstructing an index pair by the radius information value and the phase information value, and inverse A de-quantizer decodes a set of a plurality of integer indices using the index pair.

According to another aspect of the invention, the radius information value of the index pair is decoded from the bit stream by a lossless decoding process, the phase information value is an enumeration index of the index pair, By means of the enumeration index of the radius information value and the phase information value, the index pair is reconstructed as a leader on a radius coset table.

According to an aspect of the present invention, there is provided a method of encoding a TCX encoder, the method comprising: filtering an input speech signal through a time-varying weighting filter to obtain a weighted signla; and applying windowing to the weighted signal (adaptive windowing), applying a Fourier transform to the windowed signal, pre-shaping the windowed signal, and minimizing encoding noise in a Fourier transform region. The quantized signal is inversed by an inverse pre-shaping function, to provide a quantized time domain signal, and to provide a quantized time domain signal. Fourier transform.

1 is a block diagram illustrating a convolutional encoder corresponding to an 8-state trellis using a 4-corset according to an example of the present invention.

2 shows an 8-state trellis using a 4-corset according to an example of the present invention.

3 illustrates a 4-corset TCQ codebook according to an example of the present invention.

4 is a flowchart illustrating a method of configuring a quantizer according to an example of the present invention.

FIG. 5 shows the number of 2-d cubic lattice points of the first quadrant of square radius k in accordance with an example of the present invention.

6 is a diagram for describing an encoding process according to an example of the present invention.

7 is a diagram illustrating a mapping from k to m, a number of enumeration bits, and a number of enumeration bits representing a codevector reader, and a list of k and m values according to an example of the present invention.

8 illustrates a lossless encoding process according to an example of the present invention.

9 illustrates an overload region for lossless coding of index pairs.

10 is a diagram illustrating a lossless decoding method according to an example of the present invention.

11 is a diagram for describing a method for bit rate control according to an example of the present invention.

12 is a flowchart illustrating an audio signal encoding method according to an example of the present invention.

13 illustrates an AMR-WB + encoder to which TCQ is applied according to an example of the present invention.

14 illustrates an AMR-WB + decoder to which TCQ is applied according to an example of the present invention.

15 is a diagram for describing a coding method of a TCX encoder to which TCQ is applied according to an example of the present invention.

Hereinafter, exemplary embodiments of the present invention will be described in detail with reference to the accompanying drawings and the contents described in the accompanying drawings, but the present invention is not limited or limited to the embodiments. Like reference numerals in the drawings denote like elements.

The quantization method according to the present invention can be implemented using a convolutional encoder, a trellis, and a TCLIs coded quantization codebook.

1 is a block diagram illustrating a convolutional encoder corresponding to an 8-state trellis using a 4-corset according to an example of the present invention.

The convolutional encoder receives a path and selects one branch among branches which connect a state provided in a predetermined stage and a state provided in a next stage.

Referring to FIG. 1, '0' or '1', which is a code representing a path, is input through an input terminal x (n), and "D" denotes codes indicating previous paths in which z0, z1, and z2 are sequentially. The code is temporarily stored, and a code indicating a coset, which is an identifier assigned to each branch, is output through output terminals y1 (n) and y2 (n). For example, a corset corresponding to codes output through the output terminals y1 (n) and y2 (n) in FIG. 1 may be implemented as shown in Table 1 below.

Such a convolutional encoder is provided corresponding to the path of a state provided in the trellis. The trellis is composed of a plurality of states, each state provided in a predetermined stage is connected by a branch and at least one of the plurality of states provided in the next stage, each branch being a preset It can be identified by a number of corsets.

FIG. 2 is a conceptual diagram illustrating an embodiment of an 8-state trellis using a 4-corset corresponding to the convolutional encoder shown in FIG. 1.

In FIG. 2, the state is composed of eight states corresponding to 0 to 7, and the corset is composed of four corsets corresponding to D0, D1, D2, and D3 and assigned to correspond to each branch. In FIG. 2, when the coset provided at the front end of the two corsets assigned to each state is detected, the path of x (n) of the converging encoder of FIG. 1 becomes '0' and two branches connected to the nodes of each state. The branch provided at the top center is selected as the path, and when the corset provided at the rear end is detected, the path is '1' and the branch provided at the bottom of the two branches is selected as the path. If the state detects the coset D0 at 4, it corresponds to the corset provided at the rear end and the branch provided at the bottom is selected, and the code 1 indicating the path is selected.

For example, with reference to Figs. 1, 2, and 1, the codes '1', '1', and '1' representing a path to the converging encoder shown in Fig. 1 on the assumption that the initial state is '0'. When ',' 1 'and' 0 'are sequentially input through the input terminal x (n), codes outputted to the output terminals y1 (n) and y2 (n) by the converging encoder shown in FIG. Table 2 shows the states determined by the trellis shown in FIG. 2 according to the corset corresponding to this code in 1 and the path input to x (n).

Assuming that z0, z1, and z2 are all assigned with '0' as an initial value, if 1 is input to x (n) as a code representing a path, '1' in y1 (n) by the converging encoder shown in FIG. Is output and '0' is output from y2 (n), so that 'D2' is selected as the corresponding corset in the table in Table 1, and 'D2' is selected in state 0 provided in the trellis shown in FIG. '1' is selected as the next state.

Then, if 1 is entered in x (n), the previously entered x (n) value '1' is shifted to z0, and the value stored in previous z0 '0' is shifted to z1 and the value stored in previous z1 ' 0 'is shifted to z2 and' 0 'is output from y1 (n) and' 0 'is output from y2 (n) by the convolutional encoder shown in FIG. D0 'is selected, and in the state 1 provided in the trellis shown in Fig. 2, the branch is selected as' D0' and '3' is selected as the next state. In the same way, when '1' is input to x (n) next, '0' is output from y1 (n) and '1' is output from y2 (n), so the corset 'D1' is selected and the next state 7 Is selected. Next, when '1' is input to x (n), '1' is output from y1 (n) and '1' is output from y2 (n), so that corset 'D3' is selected and 7 is selected as the next state. do. Finally, when '0' is input to x (n), '0' is output from y1 (n) and '1' is output from y2 (n), thereby selecting corset 'D1'.

The corsets provided in the trellis as shown in FIG. 2 are allocated to each quantization level provided in the TClei coded quantization codebook. The TCQ codebook can determine the quantization level corresponding to each input, and a predetermined corset provided in the trellis is assigned corresponding to each quantization level. In addition, the TCQ codebook classifies consecutive quantization levels in a predetermined unit and is assigned an index.

3 conceptually illustrates an embodiment for a four-coset TCQ codebook. Referring to FIG. 3 as an example, the TCQ index '0' is provided with a quantization level '0' to which corsets D0 and D3 are assigned, and the quantization level '-1.5 with a coset D2 is assigned to TCQ index' -1 '. ', Quantization level' -2.5 'assigned with coset D1, quantization level' -3.5 'assigned with coset D0, quantization level' -4.5 'assigned with coset D3, and coset D1 at TCQ index' +1 ' The assigned quantization level '+1.5', the quantization level '+2.5' assigned to the coset D2, the quantization level '+3.5' assigned to the coset D3, and the quantization level '+4.5' assigned to the corset D0 are provided.

The TCQ codebook presets different cosets to quantization levels included in one index except for an index including quantization level '0'. In other words, the TCQ codebook classifies quantization levels by constructing one index with quantization levels corresponding to the total number of branches of a corset.

For example, if the total number of branches of the corset is four, the number of quantization levels provided in one index is also set to four and different corsets are assigned to each quantization level. As such, by assigning different corsets to quantization levels provided in one index, each quantization level can be identified within one index. The TCQ codebook configured in this manner detects a TCQ index corresponding to the input, detects and quantizes a path corresponding to the quantization level in the trellis, and performs entropy coding.

4 is a flowchart illustrating an embodiment of a method of configuring a quantizer in a new index method according to the present invention.

First, the total number of corsets to be used in the trellis and TCQ codebooks is determined, and the corsets are set correspondingly (S400). In S400, the total number of corsets may be determined as 2 ^ n (where n is an integer greater than 2). For example, if the total number of corsets is determined as 4, four cosets corresponding to D0, D1, D2, and D3 are set in the trellis and TCQ codebooks as shown in FIGS. 2 and 3, and the total number of corsets is set. When 8 is determined, eight cosets corresponding to D0, D1, D2, D3, D4, D5, D6, and D7 are set in the trellis and TCQ codebooks.

A union coset is configured using the coset set in S400 (S410). Here, the union corset constitutes corsets in which corsets allocated to branches connected to nodes of a predetermined state are incompatible. In other words, corsets assigned to branches connected to nodes of a given state are not included in the same union corset. If an embodiment of configuring a union corset in units of two corsets is configured as a table, it can be performed as shown in Table 3 below.

Where N is an integer greater than 2 and K is the total number of branches of the corset. For example, you can configure union corsets with {D0-D2, D1-D3} in the 4-coset TCQ codebook, and {D0-D4, D1-D5} and {D2-D6, D3- in the 8-coset TCQ codebook. The union corset can be configured with D7}.

The quantization levels provided in the TCQ codebook are classified and indexed into predetermined units using the union corset configured in S410 (S420). In indexing at S420, the quantization levels included in one index are indexed such that only corsets included in different union corsets are allocated.

For example, in a 4-coset TCQ codebook in which union corsets are configured with {D0-D2, D1-D3}, no index should contain quantization levels corresponding to D0 and D2 included in the same union corset. The same applies to D3.

An example of the indexing method indexed in S420 includes a first indexing method and a second indexing method described below.

First, when describing the first index scheme, an index corresponding to a positive integer is assigned to quantization levels greater than '0', and an index corresponding to a negative integer to quantization levels smaller than '0'. Are assigned, and the indexes are symmetrically assigned around '0' so that the absolute values of the indices assigned to quantization levels greater than '0' and quantization levels less than '0' are the same, but only different signs are assigned. .

For example, there is a first index shown at the bottom of the 4-coset TCQ codebook shown in FIG. The first index is indexed to +1, +2, ..., +8 with quantization levels larger than '0' around '0', and quantization levels smaller than '0' are -1, -2, .. Indexed by -8, and the indexes are assigned symmetrically with the same absolute value but different signs. In addition, no index includes quantization levels corresponding to D0 and D2 included in the same union corset, and D1 and D3 are not provided at the same index. Therefore, the first index includes quantization levels corresponding to half of the above-described TCQ index as shown in FIG. 3.

By using the above-described TCQ index method, the 4-coset TCQ codebook may be indexed by the first indexing method according to Equation 1 described below.

(Where n is the first index and t is the TCQ index)

Next, referring to the second indexing scheme, the second index does not allocate an integer corresponding to a negative number as an index, but assigns only an integer corresponding to '0' and a positive number as an index so that the index including the smallest quantization levels is '0'. As the quantization level gradually increases, an index corresponding to a large positive integer is sequentially assigned to allocate the index such that the largest index is included in the index including the largest quantization levels. When the second indexing scheme is described in comparison with the first indexing scheme, the first indexing scheme and the scheme of classifying the quantization levels are the same, and the order of indexing sequentially from the smallest quantization level to the largest quantization level is the same. However, in assigning an index, a first index allocates a negative integer to an index smaller than '0' around an index including quantization level 0, a positive integer to an index larger than '0', and a second index. There is a difference in that '0' is allocated to an index including the smallest quantization levels and a positive integer is sequentially assigned to the index including the largest quantization level.

For example, the 4-coset TCQ codebook may be indexed using the TCQ index method described above in the second indexing method by Equation 2 described below.

(Where n is the second index and t is the TCQ index)

When the above-described TCQ index, first index, and second index are performed in the 4-coset TCQ codebook provided with the 8-state trellis shown in FIG. 2 and the two zero levels shown in FIG. It can correspond as well. Here, Table 4 is an index for the union corset C0 composed of the cosets D0 and D2, and Table 5 is an index for the union corset C1 composed of the cosets D1 and D3.

In the indexing by the first index method and the second index method as described above, an index including the quantization level '0' is indexed unlike an index including the quantization levels except '0'. It may be provided as a "two zero level" in which two corsets are assigned to the quantization level '0', a "no zero level" in which no corset is assigned, and a "dead-zone" provided as a dead zone.

An example of a TCQ codebook in which two cosets are allocated at quantization level '0' is shown in FIG. 3. Here, the 4-coset TCQ codebook shown in FIG. 3 uses a convolutional encoder of 8-state trellis code using the 4-coset shown in FIG. 1 and the 4-state trellis shown in FIG. It can be carried out.

An example of a TCQ codebook with no quantization level assigned to 0 'is shown in FIG. 3 with a 4-coset TCQ codebook. Here, the 4-coset TCQ codebook shown in FIG. 3 includes a convolutional encoder of an 8-state trellis code using the 4-coset shown in FIG. 1, and a 4-state trellis shown in FIG. It can be used.

Hereinafter, a lossless coding method will be described.

Assuming that such integer indices correspond to the TCQ of a discrete Fourier transform (DFT) value, the real and imaginary portions of the complex-valued DFT constants are defined as (x (2n), x (2n +). 1)), n = 0,1,2, ..., N-1. The constant points are evenly distributed in phase and combine the real and imaginary parts to the mapping of each "radius" and "phase" processing index pair (real and virtual outputs). Can be performed efficiently.

A lossless code is a non-negative index pair pair (m (n), e (n)) for mapping each real and virtual index pair, (x (2n), x (2n + 1)). It can be composed of). Where m (n) is a parameter representing "radius" and e (n) is a parameter representing "phase" and is used to list possible combinations of index pairs having the same radius.

In this way, lossless coding uses a variable length code Cm to decode m (n) and a fixed length code Ce according to the state of m (n) to decode e (n).

Binary enumeration code is a fixed length code. If there are 2 ^L items to be encoded, each is represented by a unique L-bit binary codeword. The codeword Cm (m) is used to represent m (n), the codeword Ce (e) is used to represent e (n), and (x (2n), x (2n + 1)). The codeword to represent is in the form of (Cm (m), Ce (e)). In order to decode, Cm (m) is decoded as a non-negative integer m, and Ce (e) is decoded by e depending on the state of m. Thus, (m (n), e (n)) is again mapped to (x (2n), x (2n + 1)).

An intermediate step exists in the mapping of (x (2n), x (2n + 1)) to (m (n), e (n)). The squared radius is To be defined as, Calculates (x (2n), x (2n + 1)) of square radius k (n), And ) Into the first quadrant of If x (i) is a non-negative integer, (x (2n), x (2n + 1)) will correspond to the integer point of the plane of square radius k (n). Thus, (x (2n), x (2n + 1)) is a two-dimensional integer lattice Belongs to the first quadrant, and the lattice theta function: JH Conway and NJA Sloane, "Voronoi regions of Lattices, second moments of polytopes, and quantization," IEEE Trans. Inform. Th ., Vol. 28, pp 211-226, March 1982.), and also Lattice codevectors (GD Forney, Jr., "Coset codes-Part I: Introduction and geometrical classification," IEEE Trans. Inform. Th . , v. 34, pp. 1123-1151, Sept. 1988).

FIG. 5 shows the number of 2-d cubic lattice points of the first quadrant of square radius k in accordance with an example of the present invention.

Is defined according to the number of 2-d cubic lattice points of the first quadrant of the square radius k. of The value at is as shown in FIG. 7.

In the range of The value of is possible only five values of 0, 1, 2, 3, or 4. If the value of 1 is 1, only the first quadrant lattice point is present and therefore does not require enumeration.

If the value of is 2 or 4, the lattice point can be calculated as 1 or 2 bits, respectively. A value of 3 is relatively rare. It will only occur if the value of k is 50 or 200 in the range of.

When the value is 3, when two bits are used to encode the lattice point, a loss in coding efficiency hardly occurs.

Since there are zero lattice points at a particular value of K, m∈ {0,1,2, .. m _max }, or Only when> 0, it is convenient to apply a one-to-one mapping from k (n) to m (n). In the case of m> m _max , an overload coding rule, such as the Voronoi extension method used in the AMR-WB + standard, may be used. The encoding process as described above is illustrated in FIG. 6.

As the first quadrant two-dimensional (2-d) cubic lattice point of the squared radius k (m) of a number, Therefore, enumeration encoding requires 0, 1, or 2 bits to indicate each e (n).

The enumeration encoding and decoding can be implemented using a reference table. Each value associated with m corresponds to the first quadrant lattice vectors, Will take the form of. For example, a single lattice codevector corresponding to m = 2 (k = 2) to be. The two code vectors corresponding to m = 3 (k = 4) to be. The two code vectors to be described later are permutations, and can store one value corresponding to a larger index value in the upper portion. The permuted code vector is called a coset, and the code vector with the largest value at the top becomes a coset (or code vector) leader.

FIG. 7 is a diagram of a mapping from k to m, a required number of enumeration bits, a number of enumeration bits representing a codevector reader, and a list of k and m values, in accordance with an example of the present invention.

More specifically, the number of enumeration bits, and A list of k and m values is shown, which is m _max = 43 codevector leaders corresponding to k _max = 100 in the interval.

8 is a diagram illustrating a lossless encoding process according to an example of the present invention, and FIG. 9 is a diagram illustrating an overload region for lossless encoding of an index pair.

A lossless encoding process according to an example of the present invention will be described with reference to FIGS. 8 and 9.

(x (2n), x (2n + 1)) and The value is calculated (S900), and k (n) is mapped to m (n) by referring to the reference table using the mapping unit. At this time, m = kToM [k], and kToM [] represents a reference table (S810).

It is determined whether k (n) is less than the preset threshold 'MAXRAD' (S820).

As a result of the determination, when k (n) is less than the preset threshold 'MAXRAD', lossless coding is performed for a radius lower than the threshold (S830).

In this case, a prefix-free lossless encoding module is used to encode m into Cm (m). The prefix-free state allows the codeword to be uniquely parsed from the coded bit-stream.

Thereafter, using the reference table, the number of the bit is determined for enumeration encoding (S840).

As a result of the determination, when the number of enumeration bits is greater than 0, (x (2n), x (2n + 1)) pairs are matched by stored lattice codevector leaders, and the enumeration binary An enumeration binary codevector, that is, ce (e), is selected (S850).

Otherwise, only one index combination k (m) will be present, and coding of e (n) becomes unnecessary (S860). The codeword for (x (2n), x (2n + 1)) is ( c _m ( m ), c _e ( e )).

When k (n) is less than the preset threshold MAXRAD, the overload region is determined (S870, S875). When the index pair corresponds to the overload region 1 (1010), the codeword corresponding to the overload region 1 is used. c _m Encode _m ( m ). Then, x (2n + 1) is encoded by binary coding, and binary coding x (2n) is encoded by x (2n) -THR (S880).

As a result of the determination of the overload region, when the index pair corresponds to the overload region 2 (1020), c _m ( m ) is encoded by using a codeword corresponding to the overload region 2. Then, x (2n + 1) is encoded by binary coding, and binary coding x (2n + 1) is encoded by x (2n + 1) -THR (S890).

As a result of the determination of the overload region, when the index pair corresponds to the overload region 3 (1030), c _m ( m ) is encoded using a codeword corresponding to the overload region 3. Then, x (2n) and x (2n + 1) are encoded by binary coding, and x (2n) and x (2n + 1) are encoded by x (2n) -THR (S895).

10 is a diagram illustrating a lossless decoding method according to an example of the present invention.

The bit number numEBits used for the given m and enumeration codewords if codeword c _m is parsed from the bit-stream and the decoded m value (S1200) and m is not greater than MAXRAD. To find the reference, use the reference table shown in FIG.

These bits can be read from the bit-stream, interpreted as a non-signed integer e, and in the case of a normal codeword as a result of the determination (S1210), when numEBits is 0, e is It becomes 0 (S1020, S1030).

If numEBits> 0, e (n) is decoded and parsed from the bit-stream (S1040).

A reference table is used to decode m and e into (x ₁ , x ₂ ) pairs (S1050).

If it is not a normal codeword, and the determination result indicates that the index pair corresponds to the overload region 1, c _m (OV1) (S1060), the x (2n) binary code ov (2n) is parsed from the bit-stream, and x (2n). ) To recover the value, add THR to the value. The binary code x (2n + 1) and ov (2n + 1) are parsed to recover the x (2n + 1) value from the bit-stream (S1070).

As a result of the determination, when the index pair corresponds to the overload region 2, c _m (OV1) (S1065), the x (2n) binary code ov (2n) is parsed from the bit-stream to recover the x (2n) value. do. In order to recover the x (2n + 1) value from the bit-stream, THR is added to the value, and the binary code x (2n + 1) and ov (2n + 1) are parsed (S1080).

As a result of the determination, when the index pair corresponds to the overload region 3, c _m (OV1), the value of x (2n) binary code ov (2n) is parsed from the bit-stream and the value of x (2n) is recovered. Add THR to The binary code x (2n + 1) and ov (2n + 1) are parsed from the bit-stream, and THR is added to the value to recover the x (2n + 1) value (S1090).

The value in that range For this purpose, the preset value k, Can be represented through 0, 1, or 2 bits. Since there is a one-to-one mapping from k to m, later k and m can be referenced to each other alternately, and m can also be referred to as a codevector squared radius, even though k (m) is substantially squared radius. If M (2, m) = 1, there is only one lattice codevector of a given square radius, so that the enumeration (the value of m is decoded and the lattice codevector can be generated using a lookup table) Doesn't need a bit). In this case, the code vector is The enumeration index e is set to 0 and is no longer decoded. If, the lattice codevectors Are formed in the form of permutation pairs. For m in range, there are at most two pairs. For each permutation pair (each codevector corset), Assume that is sorted by. This refers to a lexicographic ordering from the largest value to the smallest value based on the first dimension of the vector. If two pairs exist, the codevectors And when, Sorted by. The enumeration codeword e is 0 or 1 for one lattice codevector pair, or 0, 1, 2 or 3 for two pairs.

Enumeration encoding can be implemented using a look-up table. 0≤m≤m _max represents the index of the code vector square radius. The encoding table indicated by encTable [m] [i] is related to the size and indicates the number of enumeration encoding bits. , And Maintains a codevector of squared radius k (m). When Follows the table shown in Table 6 and FIG.

The enumeration encoding algorithm is described as follows.

to map k to m Wow , And And kToM [ k ], , Equal to x ₁ And set e = i-1.

Enumeration decoding may be implemented using the same reference table as above. The enumeration decoding algorithm is described as follows.

Assuming that the m value is decoded, the number of enumeration decoded codeword bits is set to numBits = encTable [m] [0] to decode.

Then, e = 0 is set, and when numBits> 0, numEBits is read as an unsigned integer, and e = ReadEBits (numBits) is set. If e is even, then (x ₁ , x ₂ ) = (encTable [m] [e + 1], encTable [m] [e + 2]), and if e is odd, (x ₁ , x ₂ ) = ( encTable [m] [e + 1], encTable [m] [e]).

FIG. 11 is a diagram for describing a method for bit rate control (BRC) according to an example of the present invention.

A bit rate control method according to an example of the present invention will be described with reference to FIG. 11 as follows.

One frame of coefficients is rounded to an integer value (S1100), and the total consumption of the current frame BR _INIT is calculated with reference to the table to estimate the current bit rate of each integer (S1120).

Since, BR _INIT And GAIN _INIT the initial gain by the maximum gain magnification calculation module based on the target bit rate BR _TAG . = 2 is calculated as ^{λ △ BR} (S1130).

GAIN _INIT is used to scale input coefficients, and the scaled coefficients are applied to a quantization and lossless coding module (S1140).

If ΔR / BR _TAG is greater than the threshold (S1160) and the current interaction time is less than N, one or more loops are performed (S1170) and the new gain multiplier is obtained in the module GAIN _NEW = GAIN _OLD × 2 ^-λΔBR It is calculated by (S1180).

Otherwise, the interaction ends and the bitrate of the current loop becomes the final coded bitrate.

12 is a flowchart illustrating an audio signal encoding method according to an example of the present invention.

An audio signal encoding method according to an example of the present invention will be described with reference to FIG. 12.

The signal is framed and the framed signal is pre-processed (S1210).

In addition, the preprocessed signal is received and a scaling factor of the dynamic range of the preprocessed signal is adjusted by a gain factor (S1220). The gain factor may be calculated based on the statistics of the input signal and is adjusted to compare the final output bit stream bit rate and the target bit rate.

In this case, the gain factor may be adjusted by a bit rate control unit (BRC). In this case, block forming and critical decision may be performed, and the currently scaled input signal frame may be divided into n-tuple sample lengths, respectively. If the size of all n samples of a particular block is smaller than the preset threshold, then the block is considered to be insignificant, otherwise it is considered to be a significant block. A label for the importance of each block is included in the bit of the last bit stream of each block. A coding bitrate of 1 / n bits per sample overhead for coding significant labels, through block forming and importance checks for scaled coefficients Can be effectively reduced.

Thereafter, in order to generate a set composed of a plurality of integer indices, the frame is quantized using a quantizer (S1230).

Low priority blocks are blocked, and all high priority blocks are connected in a new sequence for quantization. In order to block the low importance block as described above, the quantization unit uses a series of integer indices in a signed or non-signed format. Quantize important sample sequences.

In this case, as the quantizer, a Uniform Scalar Quantizer (USQ), a Lattice Vector Quantizer (LQ), a Non-Uniform Scalar Quantizer (Non-USQ), a Vector Quantizer (VQ), a Trellis Coded Quantizer (TCQ), etc. may be used. Can be.

The integer index is mapped to an index pair of radius and phase (S1240).

Thereafter, a bit stream including the radius and phase information is generated (S1250).

The bit stream forming unit bits the gain factor, significant labels, and lossless coded TCQ indices in the bit stream for transmission or storage. Form into stream.

13 illustrates an AMR-WB + encoder to which TCQ is applied according to an example of the present invention, and FIG. 14 illustrates an AMR-WB + decoder to which TCQ is applied according to an example of the present invention.

As shown in FIG. 13, a TCQ according to an example of the present invention may be configured in an AMR-WB + encoder.

The input signal is divided into two bands. The first band is a low-frequency signal (LF signal), sampled at Fs / 2, and the second band is a high-frequency signal (HF signal), down sampled to obtain a sampled signal.

The low-frequency signal is encoded and decoded using an encoder / decoder based on ACELP and transform coded excitation (TCX). In the ACELP mode, the standard AMR-WB codec is used. In addition, the high-frequency signal is encoded with relatively few bits by a bandwidth extension (BWE) method.

The parameters sent from the encoder to the decoder are mode select bits, which are low-frequency parameters and high-frequency parameters.

In addition, in the AMR-WB + decoder as shown in FIG. 14, the low- and high-frequency bands are each decoded separately and combined in a synthesis filterbank.

15 is a diagram for describing a coding method of a TCX encoder to which TCQ is applied according to an example of the present invention.

To encode a core mono signal that is in the 0-Fs / 4 kHz band, the AMR-WB + codec uses ACELP or TCX coding in each frame. The coding mode is selected based on closed-loop analysis by synthesis.

ACELP is used for 256-sample frames and TCX for 256, 512, or 1024 frame samples.

ACELP coding and decoding is similar to the standard AMR-WB speech codec, and ACELP coding consists of LTP analysis, synthesis, and codebook. The ACELP coding mode is performed by the AMR-WB included in the AMR-WB + codec.

In particular, the TCX mode processes signals recognized on the transform domain. Fourier transformed signal (Fourier transformed signal) is quantized by TCQ (Trellis Coded Quantizer). The conversion is within 1024, 512, or 256 samples. The signal is then recovered by inverse filtering of the quantized signal through an inverse filter.

The Trellis Coded Quantizer (TCQ) according to an example of the present invention may be included in the AMR-WB +, and may be configured to be applied to the TCX encoder.

The encoding algorithm of the AMR-WB + codec is based on the ACELP / TCX model. For all inputs of the input signal, the encoder determines which encoding model is best suited, either the ACELP or TCX model, which is more suitable for general music.

In order to obtain a weighted signal, an input speech signal is first filtered through a time-varying weighting filter, which is the same perceptual filter as in AMR-WB (S1510).

Thereafter, windowing is applied to the weighted signal (adaptive windowing) (S1520), and a Fourier transform is applied to the windowed signal (S1530). The signal is pre-shaped to minimize coding noise in a Fourier transform region (S1540) and quantized using TCQ (S1550).

After quantization, an inverse pre-shaping function is applied to the inversely transformed spectrum to provide a quantized time domain signal (S1560). The gain of the frame is rescaled in order to optimize correlation with the original weighted signal (S1570).

After gain scaling, windowing is again applied to the quantized signal in order to maximize the block effect of quantization in the Fourier transform region.

In TCX mode, overlap-and-add is used for the previous frame, and the final signal is generated through inverse filtering with an appropriate filter memory update.

As described above, the present invention has been described by way of limited embodiments and drawings, but the present invention is not limited to the above embodiments, and those skilled in the art to which the present invention pertains various modifications and variations from such descriptions. This is possible.

Therefore, the scope of the present invention should not be limited to the described embodiments, but should be determined not only by the claims below but also by the equivalents of the claims.

Claims (10)

Framing the signal; Quantizing the frame using a quantizer to generate a set consisting of a plurality of integer indices; Mapping the integer index to an index pair of radius and phase; And Generating a bit stream including the radius information and phase information An audio signal encoding method comprising a. The method of claim 1, The signal is, A specific interval of the input waveform sample, or A method of encoding an audio signal, characterized in that it is an inverse filtered residual signal interval generated by linear prediction coding (LPC) of a specific interval of an audio waveform. The method of claim 1, The quantizer is, Any one of the following: Uniform Scalar Quantizer (USQ), Lattice Vector Quantizer (Lattice VQ), Non-Uniform Scalar Quantizer (Non-USQ), Vector Quantizer (VQ), and Trellis Coded Quantizer (TCQ) And a signed or non-negative integer is output by the quantization. The method of claim 3, If the quantizer is a Trellis Coded Quantizer (TCQ), The integer output from the TCQ is The level and path bits of the TCQ, or Signed indices according to the TCQ state, mapped from the level and path bits of the TCQ, or And non-negative indices according to the TCQ state, mapped from the level and path bits of the TCQ. The method of claim 1, The radius information value is, Generated by a squared summation of the index pairs, And mapping to a second index, which is an enumeration code that provides the radius value for every radius set. The method of claim 1, The phase information value is And the coset table includes pairs of indices having the same radius value. The method of claim 1, Adjusting the gain factor by bit rate control (BRC) to tune the interval of the quantization The encoding method of the audio signal further comprising. De-multiplexing the bit stream into radius information values and phase information values; Reconstructing an index pair by the radius information value and the phase information value; And A de-quantizer decodes a set of a plurality of integer indices using the index pairs Audio signal decoding method comprising a. The method of claim 8, The radius information value of the index pair is Decoded from the bit stream by a lossless decoding process, The phase information value is And an enumeration index of the index pair, wherein the index pair is reconstructed as a leader on a radius coset table by the enumeration index of the radius information value and the phase information value. Filtering the input speech signal through a time-varying weighting filter to obtain a weighted signla; Applying windowing to the weighted signal; Applying a Fourier transform to the windowed signal; Pre-shaping the windowed signal to minimize coding noise in a Fourier transform region; Quantizing the noise-minimized signal through a Trellis Coded Quantizer (TCQ); And Inverse Fourier transforming the quantized signal by an inverse pre-shaping function to provide a quantized time domain signal The encoding method of the TCX encoder, characterized in that it comprises a.