Post-quantization gain correction in audio coding

TECHNICAL FIELD [0001] The subject technology relates to gain correction in audio coding based on quantization schemes in which quantization is split into representation of the gain and shape representation, so-called. gain-shape audio coding, especially enhancement after quantization.

BACKGROUND [0002] Modern telecommunications services are expected to handle many different types of audio signals. Although the main speech content is speech signals, it is desirable to handle more general signals, e.g. music and music and speech connections. Although the capacity of telecommunications networks is constantly increasing, it is still very important to limit the required bandwidth for the communication channel. In mobile networks, lower transmission bandwidth for each connection leads to less energy consumption, both by the mobile device and the base station. This translates into energy and money savings for the mobile operator, and the end user will experience extended battery life and extended talk time. In addition, with less bandwidth per user,

[0003] Currently, the dominant compression technology for mobile voice services is the CELP (Cod Excited Linear Prediction) technique, which achieves good sound quality for speech at low bandwidths. It is widely used in used codecs, eg AMR (Adaptive MultiRate), AMR-WB (Adaptive MultiRate WideBand) and GSM-EFR (Global System for Mobile Communications - Enhanced FulIRate). However, the performance of CELP technology for general audio signals, e.g. for music, is poor. These signals can often be better represented using frequency-based coding, e.g. ITU-T codings G.722.1 [1] and G.719 [2]. However, codecs in the transformation domain generally operate at higher bit rates than speech codecs.

[0004] Codecs in the field of transformation require a compact representation of frequency transformation coefficients. These representations are often based on vector quantization (VQ), in which coefficients are coded in groups. One of the various ways of vector quantizing is GSVQ (gain-shape VQ). This approach applies normalization of vectors before coding of individual coefficients. The normalization factor and normalized coefficients are called the gain and shape of the vector, which can be separately coded. The reinforcement-shape structure has numerous advantages. Separating the gain and shape, the codec can be easily adapted to variable levels of input sources by selecting a gain quantizer. It is also beneficial from a perceptual perspective, because the gain and shape can convey different information in different frequency ranges. Finally, the separation of the reinforcement shape simplifies the construction of the quantizer and reduces its complexity in terms of memory and computing resources, compared to an unlimited vector quantizer. The functional sketch of the gain-shape quantizer is shown in Fig. 1.

-1 ΕΡ 2 681 734 Β1 [0005] When applied to a spectrum in the frequency domain, a gain-shape structure can be used to create a spectral envelope and accurately represent the structure. The gain value sequence creates a spectrum envelope, while the shape vectors indicate the details of the spectrum. From a perceptual perspective, it is advantageous to divide the spectrum using a heterogeneous band structure adapted to the frequency resolution of the human auditory system. In general, this means that narrow frequencies are used for low frequencies, and broader bands are used for high frequencies. The perceptual meaning of the structure of the exact spectrum varies with frequency, but also depends on the characteristics of the signal itself. Transformation coders often use the auditory model to determine the relevant parts of the exact structure and allocate available resources for the most important parts. The spectral envelope is often used as an input for this auditory model. The shape encoder quantizes the shape vectors using the allocated bits. Fig. 2 shows an example of a transformation based coding system with an auditory model.

[0006] Depending on the accuracy of the shape quantizer, the gain value used to reproduce the vector may be more or less appropriate. Especially when a small number of bits is allocated, the gain value deviates from the optimal value. One way to solve this is to encode the correction factor, which takes into account the incompatibility of the gain after quantizing the shape. Another solution is to encode the shape first and then to calculate the optimal amplification factor for the given quantized shape.

[0007] A solution for coding a gain correction factor after shape quantizing may require a substantial bit rate. If this rate is already low, it means that a larger number of bits has to be found elsewhere and can reduce the bit rate available for the exact structure.

[0008] Shape encoding before enhancement encoding is a better solution, but if the bit rate for the shape quantizer is selected based on the value of the quantized gain, then the quantifications of the gain and shape depend on each other. This interdependence would probably be solved by the iterative approach, but it could easily become too complex for real-time use in a mobile device.

[0009] US 2011/0002266 A1 (Yang Gao) describes frequency-domain final processing based on perceptual masking in which an adaptively modified amplification factor for improving the perceived quality of decoded spectral coefficients is applied to each frequency coefficient.

SUMMARY OF THE INVENTION [0010] The aim is to obtain gain control in audio decoding encoded with separate representations of gain and shape.

[0011] This object is achieved according to the appended claims.

[0012] The first aspect includes a gain control method that includes the following steps:

• The measure of the accuracy of the shape representation is estimated.

• Based on the estimated accuracy measure, the gain correction is determined.

• The gain representation is adjusted based on the determined gain correction.

[0013] The second aspect comprises a gain control device that includes:

• Accuracy meter adapted to estimate the accuracy measure of the shape representation and to determine the accuracy of the gain factor on the basis of the estimated measure.

-2 ΕΡ 2 681 734 Β1 • Envelope adjuster adapted to adjust the gain representation based on the determined gain correction.

[0014] The third aspect comprises a decoder including a gain control device according to the second aspect.

[0015] A fourth aspect comprises a network node comprising a decoder according to the third aspect.

[0016] The proposed gain correction scheme increases the perceived quality of the gain-shape audio coding system. This scheme has a low computational complexity and may require at most a bunch of additional bits.

BRIEF DESCRIPTION OF THE DRAWINGS [0017] The present technology, together with successive objectives and its advantages, can best be understood by reading the following description together with the figures of the attached drawing, of which:

Fig. 1 illustrates an example of a GSVQ (gain-shape vector quantization) scheme;

Fig. 2 illustrates an example of a coding scheme and decoding in the field of transformation;

Figs. 3A-C illustrate a GSVQ technique in a simplified case;

Fig. 4 illustrates an example of a decoder in the field of transformation using an accuracy measure for determining the boundary correction;

Figs. 5A-B illustrate an example of a synthesis scaling result using gain factors when the shape vector is a rare pulse vector;

Figs. 6A-B illustrate the largest possible pulse height indicative of the accuracy of the shape vector;

Fig. 7 illustrates an example of a velocity damping function for Embodiment 1;

Fig. 8 illustrates an example of a gain control function depending on the speed and maximum pulse height of embodiment 1;

Fig. 9 illustrates another example of a gain control function depending on the speed and maximum pulse height dia of embodiment 1;

Fig. 10 illustrates an embodiment of the technology in the context of the MDCT coding and audio decoding system;

Fig. 11 illustrates an example of the assignment function between a stability measure and a gain control limiting factor;

Fig. 12 illustrates an example of an ADPCM coding and decoding system with an adaptive step size;

Fig. 13 illustrates an example in the context of an ADPCM based audio coding and decoding system for a subband;

Fig. 14 illustrates an example of the technology in the context of the audio coding and decoding system based on the ADPCM for the subband;

Fig. 15 illustrates an example of an encoder in the field of transformation comprising a signal classifier;

Fig. 16 illustrates another example of a decoder in the field of transformation using an accuracy measure for determining envelope correction;

Fig. 17 illustrates an exemplary embodiment of a gain control device according to the present technology; Fig. 18 illustrates in more detail an example of implementing a gain control according to the technology in question;

Fig. 19 is a block diagram illustrating a method according to the technology in question;

-3ΕΡ 2 681 734 Β1

Fig. 20 is a block diagram illustrating an embodiment of the method according to the present technology; and

Fig. 21 illustrates an embodiment of a network according to the technology in question.

DETAILED DESCRIPTION [0018] In the following description, the same reference numerals will be used for elements having the same or a similar function.

[0019] Before describing the technology in detail, the gain-shape encoding (sound encoding using gain and shape) will be illustrated with reference to Figs. 1-3.

[0020] Fig. 1 illustrates an example of a GSVQ scheme (gain-shape vector quantization). The upper part of this figure illustrates the side of the encoder. The input vector x is passed to the calculator of standard 10, which sets the norm of the vector (gain) g, usually

Λ Euclidean norm. This exact norm is quantized in the quantizer of standard 12, and the inverse 1/9

Λ the quantized norm 9 is passed to the multiplication block 14 to scale the input vector x to the shape. The shape is quantized in the shape quantizer 16. The representations of the quantized gain and shape are passed to the multiplexer of bitstreams (mux) 18. These representations are illustrated with dashed lines to indicate that they may e.g. be indexes in tables instead of actual quantized quantities.

[0021] The bottom part of Fig. 1 illustrates the side of the decoder. The bitstream demux (demux) 20 receives representations of gain and shape. The shape representation is passed to the shape dequantizer 22, and the gain representation is passed to the gain 24 dequantizer.

Λ gain 9 is passed to the multiplication block 26, in which it scales the obtained shape, which gives

Λ recreated vector x.

[0022] Fig. 2 illustrates an example of coding scheme and decoding in the field of transformation. The upper part of the figure illustrates the side of the encoder. The input signal is transmitted to a frequency transformer 30, e.g. based on a modified discrete cosine transform (MDCT), to produce an X frequency transform. This frequency transform X is passed to the envelope calculator 32, which determines the energy E ( b) each of the frequency b bands.

AA

These energies are quantized to energy E (b) in the envelope quantizer 34. The quantized energies E (b) are passed to the envelope normalizer 36, which scales the b-band coefficients b

And transforms X according to the inverse of the corresponding quantized energy E (b) of the envelope. Received

Λ the scaled shapes are passed to the precision structure quantizer 38. The quantized energies E (b) are also passed to the bit allocator 40, which allocates bits for quantizing the exact structure for each b-band. As indicated above, the R (b) bit allocation may be based on the layout model

Λ auditory human. Representations of the quantized E-rays (b) and the corresponding quantized shapes are transmitted to the bit stream multiplexer 18.

[0023] The bottom part of Fig. 2 illustrates the side of the decoder. The bit stream demultiplexer 20 receives the gain and shape representations. These reinforcement representations are transmitted to the dequantizer

A envelope 42. The generated envelope energies E (b) are transmitted to the bit allocator 44, which designates

-4 ΕΡ 2 681 734 Β1 allocation of bits R (b) of the shapes obtained. The shape representations are passed to the precision structure dequantizer 46, which is controlled by the allocation of the bits R (b). The decoded shapes are transmitted to the envelope shaping module 48, which scales them with the appropriate ones

Λ envelope energy H (b) to create a reconstructed frequency transform. This transform is passed to the inverse frequency transformation module 50, e.g. based on the Inverse Modified Discrete Cosine Transform (IMDCT), which generates an output signal representing the synthesized sound.

[0024] Figs. 3A-C illustrate the GSVQ (gain-shape vectorization) technique described above in a simplified case in which the frequency band b is represented by the 2-dimensional vector X (b) of Fig. 3A. simple enough to be illustrated in the figure, but generic enough to illustrate the gain-shape quantization problem (in practice, these vectors usually have 8 or more dimensions.) The right side of Fig. 3A illustrates the exact representation of the X-shape gain ( b) with reinforcement E (b) and shape (unit vector) / \ f (b).

[0025] Thus, as illustrated in Fig. 3B, the exact gain E (b) is encoded to the quantized

Λ Λ gain E (b) on the encoder side. Since the inverse of the quantized gain E (b) is used to scale the vector X (b), the resulting scaled vector N (b) will indicate in the correct direction but will not necessarily have a unit length. When quantizing the shape

Λ The resized vector N (b) is quantized to a quantized N (b) shape. In this case, the quantization is based on the pulse coding scheme [3] that creates the shape (or direction) based on the sum of the total signed pulses. These pulses can be added to each other for each dimension. This means that the possible quantization positions of the shape are represented by large dots w

Λ rectangular nets illustrated in Fig. 3B-C. As a result, the quantized shape / V (b) generally does not coincide with the shape (direction) Λ / (ώ) (and N (b)).

[0026] Fig. 3C illustrates that the shape quantizing accuracy depends on the allocated bits R (b) or equivalently the total number of pulses available for shape quantization. On the left side of Fig. 3C the quantization of the shape is based on 8 pulses, and the quantization of the right shape the page uses only 3 pulses (the example in Fig. 3B uses 4 pulses).

[0027] Thus, it should be noted that depending on the accuracy of the shape quantizer, the gain value

Λ

H (b) used to reproduce the vector X (b) on the decoder side may be more or less suitable. According to the technology, the gain correction can be based on the accuracy of the quantized shape.

[0028] The accuracy measure used to correct the gain may be derived from parameters already available in the decoder, but may also depend on additional parameters intended for this measure of accuracy. Typically, these parameters include the number of bits allocated for a shape vector and the shape vector itself, but may also include a gain value associated with this shape vector and predetermined statistics for signals that are typical for the coding and decoding system. A general view of the system using the accuracy measure and the correction or adjustment of the gain is shown in Fig. 4.

[0029] Fig. 4 illustrates an example of a decoder in the field of transformation 300 using an accuracy measure to determine the correction of the envelope. For the sake of clarity, only the decoder's side is illustrated. Page

-5ΕΡ 2 681 734 Β1 of the encoder can be implemented as in Fig. 2. The new element is a gain control device 60. This gain control device 60 includes an accuracy meter 62 adapted to be estimated

Λ measurement of accuracy A (b) of the A / (b) shape representation, as well as for determining the g gain correction_c(b) based on the estimated accuracy measure A (b). It also contains an envelope regulator 64 adapted for

Λ adjusting the gain representation E (b) on the basis of the determined gain correction.

[0030] As indicated above, in certain example embodiments of the invention, the gain correction can be performed without the cost of additional bits. This is achieved by estimating the gain correction based on parameters already available in the decoder. This process can be described as estimating the accuracy of the coded shape. Typically, this estimation involves determining the measure of accuracy A (b) from the quantization characteristics of the shape indicating the resolution of this quantization of the shape.

Example 1 [0031] In one example, the technology in question is used in a sound coding / decoding system. This system is based on transformation, and the transformation used is a modified discrete cosine transform (MDCT) using sinusoidal windows with 50% overlap. It should be understood, however, that any transformation suitable for coding can be used with appropriate segmentation and windows.

Encoder of Example 1 [0032] The input sound is recorded as frames with 50% overlap and using a symmetrical sinusoidal window. Each windowed frame is then transformed into the MDCT X spectrum. This spectrum is divided into subbands for processing, the widths of these subbands being non-uniform. The spectral coefficients of the m-frame belonging to the b-band are denoted by X (b, tri) and have the bandwidth BW (b). Since most coding and decoding steps can be described for a single frame, we omit the frame index and simply use the X (b) notation. Preferably, bandwidths should increase as the frequency increases to match the frequency resolution of the human auditory system. RMS value

<img file="PL2681734T3_D0001.tif" />

(1) where X (b)^T means transposition X (b).

[0033] The RMS rms value can be seen as the energy value for the coefficient. The sequence of normalization coefficients E (b) for b = ^^ .. ^ Nbands is the envelope of the MDCT spectrum, where N_B3Nds is the number of bands. Then the sequence is quantized for transmission to the decoder. To ensure that normalization can be reversed in the decoder, a quantized envelope E (b) is obtained. In this example, the envelope coefficients are logarithmic scalars using the step size 3dB, and the quantized indices are differentially encoded using Huffman coding. The quantized envelope is used to normalize the spectral bands, i.e .:

<img file="PL2681734T3_D0002.tif" />

(2)

-6ΕΡ 2 681 734 Β1 [0034] It should be noted that if a non-quantized envelope E (b) is used for normalization, the shape would have an RMS value = 1, i.e .:

N '(b)

E (b)

X (b)

BW {b) (3)

[0035] When using the quantized envelope E (b), the shape vector has an RMS value close to 1. This feature is applied in the decoder to produce an approximation of the gain value.

[0036] The normalized N shape vectors (b) form the exact structure of the MDCT spectrum. The quantized envelope is used to generate the R (b) bits for coding the normalized shape vectors N (h). Preferably, the bit allocation algorithm uses an auditory model to arrange the bits in the parts most relevant to the perception. Any quantization scheme can be used to encode a shape vector. It is common for all that they can be created assuming that the input is normalized, which simplifies the quantizer design. In this example, the quantization of the shape is performed using a pulse encoding scheme that creates a synthesized shape from the sum of total pulses with the sign [3]. These pulses can be added to each other to create pulses of different heights.

[0037] The quantizer indices from envelope quantization and shape quantizing are multiplexed in the bit stream for write or send to the decoder.

Decoder of Example 1 [0038] The decoder demultiplexes the indexes from the bitstream and forwards corresponding indexes to each of the decoding modules. First, a quantized envelope E (b) is obtained. Then, from this quantized envelope, the allocation of bits of the exact structure using the allocation is derived

Λ bits identical to those used in the encoder. The shape nectors N (b) of the exact structure are decoded using these indexes and the received bit allocation R (b).

[0039] Now, before corrugating the decoded exact structure using the envelope, additional gain correction factors are determined. First, a gain corresponding to the RMS value is obtained from the formula:

f BW (b)

JV (fc&gt;)^TJV (b) (4) [0040] The coefficient g s (b) is a scale factor that normalizes the RMS value to 1, i.e.;

(&amp; RMS (W))^T (G ^ LWI) _t 'BW (b) {5) [0041] In this example, we intend to minimize the mean squared error (MSE) of the synthesis:

= ^aig ^min - 9 'M (6) which has a solution

-7ΕΡ 2 681 734 Β1 jv (b) W)

.....

(7) [0042] Because gMSE (b) depends on the input form N (b), it is not known to the decoder. In this example, the impact is estimated using the accuracy measure. The ratio of these reinforcements is defined as the gain factor g_c(to

9_c{b) =

9MSE (fr)

9rms [b] (8) [0043] When the shape quantizing precision is good, the correction factor is close to 1, i.e .:

Ń (b} -&gt; JV * (fo) =&gt; g_c(b) -&gt; 1 (9)

[0044] However, when the accuracy of N (b) is low, the gMSE (b) and gRMS (b) coefficients diverge. In this example, in which the shape is encoded using a pulse encoding scheme, the low bit rate causes the shape vectors to take little value, while gRMS (b) gives an overestimation of the MSE-relevant gain. In this case, the value of gc (b) should be less than 1 to compensate for this overshoot. For example, Fig. 5A-B shows an illustration of a case with a low shape pulse rate. Fig. 5A-B illustrates an example of scaling a synthesis with the gain factors gMSE (Fig. 5B) and gRMs (Fig. 5A) when the shape vector is a pulse vector with few possible values. Scaling by g_RMs leads to impulses that are too high in the sense of MSE.

[0045] On the other hand, the target signal with a large number of peaks or sparse can be well represented by the pulse shape. Although the rarity of the input signal may not be known at the synthesis stage, the rarity of the synthesis shape may serve as an indicator of the accuracy of the synthesized shape vector. One of the ways to measure the rarity of the shape of the synthesis is the height of the maximum peak in this shape. This is because a sparse input signal is more likely to cause high peaks in the shape of a synthesis. Fig 7A-B illustrates how the peak height can indicate the accuracy of two pulse vectors with an equal foot. In Fig. 7A there are 5 available pulses (R (b) = 5) to represent the dashed shape. Because this shape is essentially constant,_max = 1. In Fig. 7B there are also 5 pulses available to represent the dashed shape. However, in this case, the shape contains peaks or is sparse, and the largest peak is represented by 3 pulses on top of each other, i.e. pma * = 3. This indicates that the g gain correction_c(ó) depends on the estimated rarity p_max of the quantized shape. [0046] As indicated above, the input shape W (P) is not known to the decoder. Because the value of taste ^ b) depends on the input shape W (b), it means that correction or compensation of gain g_c(b) in practice, can not be based on a perfect equation (8). In this example, the g gain correction_c(b) instead, it is selected based on the bit rate in relation to the number of pulses R (b), the height of the largest pulse in the shape vector pmax (ó) and frequency b band, i.e .:

0cfl&gt;) = p_m, (b), b) (10) [0047] It has been observed that lower rates generally require attenuation of the enhancement to minimize the MSE error. The dependence on the foot can be realized in the form of an array and (R (fc&gt;)), which is learned for the corresponding sound data signal. An example of a table is shown in Fig. 7. Because vectors

-8ΕΡ 2,681,734 Β1 shapes in this example have different widths, preferably the foot can be expressed as the number of pulses per sample. In this way, the same foot dependent attenuation can be used for all bandwidths, another solution that is used in this example is to use the step size T in the array depending on the bandwidth. Here, we use 4 different bandwidths in 4 different groups, so we require 4 step sizes. An example of the step size is shown in Table 1. Using the step size, the value from the table is determined using a rounding operation where LJ represents the rounding to the nearest whole number.

Tab. 1

<td>Group of bands</td><td>Bandwidth</td><td>Step size T</td>

<td>1</td><td>8</td><td>4</td>

<td>2</td><td>16</td><td>4/3</td>

<td>3</td><td>24</td><td>2</td>

<td>4</td><td>34</td><td>1</td>

[0048] Another example of the table is given in Table 2.

Tab. 2

<td>Group of bands</td><td>Bandwidth</td><td>Step size T</td>

<td>1</td><td>8</td><td>4</td>

<td>2</td><td>16</td><td>4/3</td>

<td>3</td><td>24</td><td>2</td>

<td>4</td><td>32</td><td>1</td>

[0049] The estimated rarity can be implemented as another table u {R (b), pmax (b)) depending on both the number of pulses R (b) and the height of the maximum pulse p_mand * (b). An example of a table is shown in Fig. 8. The table u serves as a measure of the accuracy of A (b) for the b-band, i.e .:

^ (b) = u (R (i&gt;), p_m"(B)) (11) [0050] It was noted that from the perspective of perception, the approximation g" se was more appropriate for the lower frequency range. For higher frequencies, the exact structure becomes less important for perception, and the conformity of energy or RMS values ​​becomes crucial. For this reason, gain attenuation can only be used below a certain b-mR band number. In this case, gain correction g_c(b) shows an explicit dependence on the b-frequency band. In this case, the resulting gain correction function can be defined as:

0c (&amp;) b &lt;b_TfJR otherwise (12)

The foregoing description may also be applied to the description of key elements of the embodiment of the invention of Fig. 4. Thus, in the example of Fig. 4, the final synthesis of X (b) is calculated as:

X (b) = Sc (fc)! 7 ™_s(fc) E (n) JV (b) (13)

VW is a different solution, function u (R (b}, p_ma * (b)} can be realized as a linear function of the maximum pulse height p, ™ x and of the allocated bit rate R (b), e.g. as:

U (R (b), p_max (h)) = k (p_mtix (b) - J? (b)) + 1 (14) where the slope k is given by:

* (*&gt;) - 1

Δα = (a,_nax - · a_min) / R [b) (15) [0052] This function depends on the amine tuning parameter that sets the initial attenuation factor for R (b) = 1 and p_{It has}x (b) = 1. This function is illustrated in Fig 9 for the amine tuning parameter = 0.41, Usually u_max 6 [0,7,4,4] and t / min e [0; omax]. In equation (14) u depends linearly on the difference p_max (b) and R (b). Another option is to use other pitch factors for pmax (b) and R (b).

[0053] The bit rate for a given band may vary rapidly for a given band between adjacent frames. This can lead to rapid changes in gain correction. Such changes are particularly important when the envelope is quite stable, i.e. the total changes between the frames are quite small. This is often the case for musical signals that usually have more stable enveloping energies. To avoid introducing instability by suppressing the gain, additional adaptation may be added. A general scheme of such an embodiment of the invention is shown in Fig 10, in which a stability meter 66 has been added to the gain control device 60 in the decoder 300.

[0054] This adaptation may be based e.g. on the envelope stability stability E (b). An example of such a measure is the calculation of the square Euclidean distance between neighboring Iog2 envelope vectors:

ΔΕ (ιη) = -,¹- (16) "liands ί&gt;" θ Here, AE (m) means the square Euclidean distance between the envelope vectors for the frame m and the m-1 frame. The measure of stability can also be filtered low-pass for smoother adaptation:

A £ (m) = - erAE (m) -F (1 - α) ΔΕ (ιη -1) (17) [0056] The value corresponding to the forgetfulness factor a may be 0.1. The smoothed stability measure can then be used to create a damping limitation using, e.g., a sigmoidal function, e.g.

^GMI"i ₊ (¹⁸) in which you can choose Ci = 6, C2 = 2 and C3 = 1.9 as parameters. It should be noted that these parameters should be seen as examples, and current values ​​can be chosen with greater ease. Ex .:

-10EP 2 681 734 B1 c, s [1, 10]

Cr, e [1.4] C₃ e [-5, 10] [0057] Fig. 11 illustrates an example of a function assigning a measure of the stability ΔΕ (η?) to a coefficient for limiting the regulation of communal gain. Preferably, the above expression on communes is implemented as a table or using a simple step function, e.g.

(19)

S min i, a £ (? «) &lt;C₃/ c_! + c₂0, + c₂ [0058] The variable limitation of population suppression 6 [0,1] can be used to produce modifications to the gain adapted to stability g_c(b} as:

g_c(b) = max (g_c(b), 0_luin) (20) [0059] After the gain is estimated, the final synthesis of X (b) is calculated as:

Xtb) = g_c(B) g "_s(b) E (n) Ń (b) (21) t / (n) [0060] In the variations of Example 1 described, the sum of synthesized vectors X (b) forms synthesized

Λ spectrum X, which is further processed using the reverse transformation of MDCT, windowed using a symmetrical sinusoidal window and added to the output synthesis using the overlap-and-add strategy.

Example 2 [0061] In another example, the shape is quantized using a Quadrature Mirror Filter (QMF) and ADPCM (Adaptive Differential Pulse Code Modulation) scheme for shape quantizing. An example of the ADPCM scheme for a subband is ITU-T G.722 [4]. Preferably, the input audio signal is processed in segments. An example of the ADPCM scheme is shown in Fig. 12, with an adapted step size S. Here, the adapted step size of the shape quantizer serves as a measure of the accuracy that is already present in the decoder and does not require additional signaling. However, the size of the quantization step must be derived from the parameters used by the decoding process and not from the synthesized shape itself. The general scheme of this example is shown in Fig. 14.

[0062] Fig. 12 illustrates an example of an ADPCM coding and decoding system with an adapted quantizing step size. The ADPCM 70 quantizer includes a totalizer 72 that receives the input signal and subtracts the estimate of the previous input signal to generate the error signal e. This error signal is quantized in the quantizer 74 whose output is passed to the bit stream multiplexer 18, and to the step size calculator of step 76 and 78. The step size calculator 76 adapts the size of the quantizing step S to obtain a permissible error. The quantization step size S is passed to the bit stream multiplexer 18, and also controls the quantizer 74 and

- 11 ΕΡ 2 681 734 Β1 with dequantizer 78. The dequantizer 78 transmits an estimate of error e to adder 80. The second input of adder 80 receives an estimate of the input signal after delay by delay element 82. This creates a current estimate of the input signal that is transmitted to the delay element 82. The delayed signal is also passed to the step size calculator 76 and (with the sign change) to the adder 72 to generate the error signal e.

[0063] The ADPCM 90 dequantizer 90 includes a step size decoder 92 which decodes the received quantizing step size S and passes it to the dequantizer 94. The dequantizer 94 decodes the error estimation ó which is transmitted to the adder 98 whose second input receives the output from the adder delayed by delay element 96.

[0064] Fig. 13 illustrates an example in the context of a sound encoding and decoding system based on the ADPCM scheme for a subband. The encoder side is similar to the encoder side of the example of Fig. 2. The essential differences are that the frequency transformer 30 has been replaced with the QMF 100 analysis filterbank, and that the exact structure quantizer 38 has been replaced with the ADPCM quantizer, e.g., the quantizer 70 of Fig 12. The decoder side is similar to the decoder side of the example of Fig. 2. The essential differences are that the reverse frequency transformer 50 has been replaced with the QMF 102 synthesis filterbank, and that the dequantizer of the exact structure 46 has been replaced with the ADPCM dequantizer, e.g. the dequantizer 90 of Fig. 12.

[0065] Fig. 14 illustrates an example of the present technology in the context of a sound encoding and decoding system based on the ADPCM scheme for a subband. For the sake of clarity, only the decoder 300 side is illustrated. The encoder side can be implemented as in Fig. 13.

The Encoder of Example 2 [0066] The encoder uses a QMF filter bank to receive subband signals. The RMS values ​​of each of the subband signals are calculated, and these subband signals are normalized. Envelope E (b), bit allocation for subband R (b) and normalized vector vectors W (ó) are obtained as in example 1. Each normalized subband is fed to the ADPCM quantizer. In this example, the ADPCM technique is used in a concurrent adaptive fashion and determines the scaling step S (b) to apply to the subband b. The scaling step is chosen to minimize the MSE error for the subband frame. In this example, this step is selected by trying all possible steps and selecting the one that gives the smallest MSE error:

S (b) = min- ~ (] V (b) Q (JV (b), s))^T(w (b) Q (ff (b), s)) (22) where Q (x, s) is the quantizing function ADPCM of variable x using step s. The selected step size can be used to generate a quantized shape:

Ń (b) = Q (N (b)_&gt;S (b) ') (23) [0067] The quantizer indices from envelope quantization and shape quantizing are multiplexed in the bitstream for recording or transmission to the decoder.

Decoder of Example 2 [0068] The decoder demultiplexes the indexes from the bitstream and forwards the respective indexes to

Λ each of the decoding modules, E-book EQ (b) and the allocation of bits R (b) are acquired

Λ as in example 1. The synthesized shape vectors / V (b) are obtained from the decoder or dequantizer

-12 ER 2 681 734 Β1

ADPCM with adaptive step sizes S (b). These step sizes indicate the accuracy of the quantized vector of the shape, with the smaller step size corresponding to greater accuracy and vice versa. One possible implementation is that the accuracy of A (b) is inversely proportional to the size of the step due to the application of the proportionality factor γ:

A (b) = _r (24) where the γ coefficient should be chosen to obtain the desired relationship. One possible choice is γ = Smin, where Smin is the minimum step size, which gives the accuracy of 1 for S (b) = Smin.

[0069] Gain correction factor g_c it can be output using the ordering function:

₀"(B) = b (i? {B), b) A (b) (1) [0070] The mating function h can be implemented as a table based on the foot R {b) and the frequency band b. This table can be defined by grouping the optimal gain correction gMSElgnMs for these parameters and calculating the array element by averaging the optimal gain correction values ​​for each group.

[0071] After the gain correction has been estimated, the synthesis for subband X (b) is calculated as;

* (b) = g_c(B)_gMUb) E (n) W | b) (2) ϊ (η) [0072] The output audio frame is determined by applying a synthesis bank QMF to the subband.

[0073] In the example illustrated in Fig. 14, the accuracy meter 62 in the gain control device 60 has yet to be decoded the size of the quantization step S (b) directly from the received bit stream. As indicated above, another possibility is to decode into the ADPCM 110 dequantizer and transfer it to the accuracy meter 62 in a decoded form.

[0074] It should be noted that the above-described example 2 does not form part of the claimed invention, but provides an example of another embodiment useful for understanding the present invention.

Other possibilities [0075] The measure of accuracy can be supplemented by a parameter of the signal class outputted in the encoder. This can be, for example, the distinction speech / music Sub, estimating the background noise level. A general scheme of a system using a signal classifier is shown in Figs. 15-16. The encoder side of Fig. 15 is similar to the encoder side of Fig. 2, but is provided with a signal classifier 104. The decoder side 300 of Fig. 16 is similar to the decoder side of Fig. 4, but is further equipped with a signal class input for accuracy meter 62.

[0076] The signal class may be considered for gain correction e.g. by class-dependent adaptation. Assuming that the signal classes or the music corresponding to the values ​​C = 1 and C = 0, it is possible to limit the performance of gain control only to speech, i.e .:

't (R [b}) · Α (ή, b &lt;b_mR/ xC = l 1, otherwise (27)

- 13ΕΡ 2 681 734 Β1 [0077] In another embodiment of the invention, the system may act as a predictor along with partially encoded correction or gain compensation. In this example, the precision measure is used to improve the correction prediction or gain compensation so that the remaining gain error can be encoded using a lower number of bits.

[0078] In the production of a correction factor or gain compensation g_c a compromise between the compliance of the RMS value or energy and the minimization of the MSE error may be desirable. In some cases, energy compliance becomes more important than the exact course. This is e.g. for high frequencies. Therefore, in the next example, the final gain correction can be output using a weighted sum of different gain values:

_g' ₌ P9R_Ms + (¹ + (1 - = /? + (1 - fi) g_c (28)

RMS 9 RMS where g_c is a gain correction obtained in accordance with one of the solutions described above. The ß-weighting factor can be adapted e.g. frequency of a frequency, bit rate or signal type.

[0079] The steps, functions, procedures and / or blocks described herein may be implemented in hardware using any conventional technology, e.g. discrete circuit technology or integrated circuits, including both general-purpose electronic circuits and dedicated circuits.

[0080] Alternatively, at least some of the steps, functions, procedures and / or blocks described herein may be implemented using software for execution by a suitable processing device, e.g. a microprocessor, a signal processor (DSP) and / or auburn. any other programmable logic device, e.g. FPGA matrix (Field Programmable Gate Array (FPGA).

[0081] It should also be understood that it may be possible to re-use the overall decoder processing resources. This may, for example, occur as a result of re-programming of existing software or by adding new software components.

[0082] FIG. 17 illustrates an exemplary embodiment of a gain regulating device 60 according to the present technology. This example is based on a processor 110, e.g. a microprocessor, which executes the program component 120 to estimate the measure of accuracy, a software component 130 to determine the gain correction, and a software component 140 for adjusting the gain representation. These software components are stored in the memory 150. The processor 110 communicates with the memory via

Λ Λ system bus. The parameters A / (b), R (b), and (b) are received by an I / O controller 160 (I / O) to which the processor 110 and memory 150 are connected. in this example, the parameters received by the I / O controller 160 are stored in the memory 150 in which they are processed by software components. Program components 120, 130 may implement the functionality of block 62 from the above-described examples. Program component 140 may implement the functionality of block 64 from the examples described above. The adjusted gain representation (b) obtained from the software component 140 is output at the memory output 150 by the i / O controller 160 via the I / O bus.

[0083] Fig. 18 illustrates in more detail an example of implementing a gain control according to the present technology. The attenuation estimator 200 is adapted to use the received R (b) bit allocation to determine gain attenuation f (R (b)). This dampening estimate 200 can e.g. be implemented in the form of an array or as a program based on a linear equation, e.g. equation (14) above. The allocation of bits R (b) is also passed to a shape accuracy estimator 202, which is also estimated

-14ΕΡ2 681 734 rzad1 rarity p_max (b) of a quantized shape, e.g. represented by the height of the highest impulse w

Λ A / (b) shape representation. The shape accuracy estimator 202 may e.g. be implemented in the form of an array. The estimated attenuation t (R (b)} and the estimated accuracy of the shape A (b) are multiplied in the multiplication block 204. In one example, the product t (R (b)) A {b) gives a straight gain g_c(B). In another example, g gain correction_c(b) is derived according to equation (12) above. This requires a switch 206 controlled by a comparator 208, which determines if the frequency band b is below the frequency limit P_mR. If it is, then g_c(b) equals f (R (b)) - A {b}. Otherwise for g_c(b) the value 1 is selected. G gain correction_c(b) is passed to another multiplication block 210 whose second input receives an RMS value corresponding to gain g_RMA (b). This RMS value corresponding to the g gain_RMA (b) is determined by the corresponding gain calculator

Λ RMS 212 values ​​based on the received / V shape representation (b) and the corresponding BW bandwidth (b), see equation (4) above. The obtained product is passed to another multiplication block 214,

Λ Λ which also receives a N (b) shape representation as well as a E (b) gain representation and produces

Λ synthesis of X (b).

[0084] The stability detection described with reference to Fig. 10 can be used in Example 2 as well as other examples described above. Fig. 19 is a block diagram illustrating a method

Λ according to the technology in question. Step S1 estimates the measure of accuracy A (b) of the A / (b) shape representation. This precision measure may e.g. be derived from the quantization characteristics of the shape, e.g. R (b), S (b), indicating the resolution of the quantization of the shape. Step S2 determines the gain correction, e.g. g_c(b), 9_c(B)

Λ g '_c(b), based on the estimated accuracy measure, Step S3 controls the representation of the gain E (b) based on the determined gain correction.

[0085] Fig. 20 is a block diagram illustrating an embodiment of a method according to the present technology, wherein the shape has been coded using a pulse coding scheme, and the gain correction depends on the estimated rarity of pmax (b) of the quantized shape. It was assumed that the measure of accuracy was already determined in step S1 (Fig. 19). Step S4 estimates the gain attenuation, which depends on the allocated bit rate. Step S5 determines the gain correction based on the estimated measure of accuracy and the estimated gain attenuation. The procedure then proceeds to step S3 (Fig. 19) for adjusting the gain representation, [0086] Fig. 21 illustrates an embodiment of a network according to the technology in question. It comprises a decoder 300 equipped with a device for adjusting the gain according to the technology in question. In this example, a radio terminal is illustrated, but other network nodes are also suitable. For example, if the network uses the voice over łP (Internet Protocol) protocol, these nodes can be computers.

[0087] In the network node of Fig. 21, the antenna 302 receives an encoded audio signal. The radio module 304 converts this signal into sound parameters that are transmitted to the decoder 300 to produce a digital audio signal, as described in various examples above. This digital audio signal is then digitally processed and amplified in the 306 module and then transmitted to the 308 loudspeaker.

- 15ΕΡ 2 681 734 Β1 [0088] Although the above description focuses on sound coding based on transformations, the same principles can also be used for time-domain audio coding with separate reinforcement and shape representations, e.g. CELP encoding.

[0089] Those skilled in the art will recognize that various modifications and changes can be made to the technology without departing from the scope of the present invention as defined by the appended claims.

LIST OF SHORTCUTS [0090]

ADPCM technique Adaptive Differential Pulse-Code Modulation

AMR Adaptive MultiRate

AMR-WB Adaptive MultiRate WideBand

CELP Coda Excited Linear Prediction technique

GSM-EFR Global System for Mobile Communications - Enhanced FulIRate

DSP signal processor (Digital Signal Processor)

FPGA Fieldmap Gate Array

IP Internet protocol

MDCT modified discrete cosine transformation (Modified Discrete Cosine

Transform)

MSE

GMF

RMS

VQ Root Error Error Ouadrature Mirror Filter Root-Mean-Square Vector Quantization (Vector Ouantization)

LITERATURE [1] "ITU-T G.722.1" C: A NEW LOW-COMPLEXITY 14 KHZ AUDIO CODING STANDARD ", ICASSP 2006 [2]" ITU-T G.719: A NEW LOW-COMPLEXITY FULL-BAND (20 KHZ) AUDIO CODING STANDARD FOR HIGH-OUALITY CONVERSAT! ONAL APPLICATIONS ", WASPA 2009 [3] U. Mittal, J. Ashley, E. Cruz-Zeno," Low Complexity Factorial Pulse Coding of MDCT Coefficients using Approximation of Combinatorial Functions ", ICASSP 2007 [4]" 7 kHz Audio Coding Within 64 kbit / s ", [G.722], IEEE JOURNAL ON SELECTED AREAS IN COMMUNICATIONS, 1988