BR122019023345B1 - AUDIO ENCODING SYSTEM, AUDIO DECODER, AUDIO ENCODING METHOD AND AUDIO DECODING METHOD - Google Patents

Abstract

The present invention relates to an unprecedented audio encoding system that can encode both general audio and speech signals at low bit rates. A proposed audio coding system comprises the linear prediction unit to filter an input signal based on an adaptive filter; a transformation unit, to transform a frame of the filtered input signal into a transformed domain; and a quantization unit to quantize the signal from the transformed domain. The quantization unit decides, based on the characteristics of the inserted signal, to encode the transformed domain signal with a model-based quantizer or a non-model-based quantizer. Preferably, the decision is based on the duration of the framework applied by the processing plant.

Description

[001] Divided from PI0822236-3 deposited on December 30, 2008.

Technical Field

[002] The present invention relates to the encoding of audio signals and, in particular, to the encoding of any audio signal, not limited to speech, music or a combination thereof.

Background of the Invention

[003] In the prior art, there are speech encoders specifically designed to encode speech signals based on encoding on a source model of the signal, that is, the human vocal system. These encoders cannot handle arbitrary audio signals, such as music, or any signal other than speech. Additionally, in the prior art, there are music encoders, commonly referred to as audio encoders, that base their encoding on understandings about the human auditory system, rather than the source model of the signal. These encoders can handle arbitrary signals very well, albeit at low bit rates for speech signals, and the dedicated speech encoder provides superior audio quality. Therefore, to date, there is no general coding structure for coding arbitrary audio signals that also performs a speech-to-speech encoder and also a music-to-music encoder when operated at low bit rates.

[004] Thus, there is a need for better audio encoder and decoder with higher audio quality and / or lower bit rates.

Summary of the Invention

[005] The present invention relates to the efficient coding of arbitrary audio signals at a level of quality equal to or better than that of a system specifically adapted to a specific signal.

[006] The present invention is directed to audio codec algorithms that contain both a linear prediction encoding (LPC) and a transform encoding part that operates on a processed LPC signal.

[007] The present invention additionally refers to a quantization strategy that depends on the duration of the transformed frame. In addition, an entropy-restricted quantizer based on a model that employs arithmetic coding is proposed. In addition, the insertion of random displacements in a uniform scalar quantizer is provided. The invention further suggests a model-based quantizer, for example, an Entropy Restriction Quantizer (ECQ), which employs arithmetic coding.

[008] The present invention additionally relates to the efficient encoding of scale factors in the transformed encoding part of an audio encoder by exploring the presence of LPC data.

[009] The present invention additionally relates to the efficient use of a bit reservoir in an audio encoder with a variable frame size.

[0010] The present invention additionally relates to an encoder for encoding audio signals and generating a bit stream, and a decoder for decoding the bit stream and generating a reconstructed audio signal which is perceptibly indistinguishable from the input audio signal.

[0011] A first aspect of the present invention relates to quantization in a transform encoder that, for example, applies a Modified Discrete Cosine Transform (MDCT). Preferably, the proposed quantizer quantizes MDCT lines. This aspect is applicable regardless of whether the encoder additionally uses an encoding analysis by linear prediction (LPC) or additional long-term prediction (LTP).

[0012] The present invention provides an audio coding system that comprises a linear prediction unit, to filter an input signal based on an adaptive filter; a transformation unit, to transform a frame of the filtered input signal into a transformed domain; and a quantization unit to quantize the signal from the transformed domain. The quantization unit decides, based on the characteristics of the input signal, to encode the transformed domain signal with a model-based quantizer or a non-model based quantizer. Preferably, the decision is based on the duration of the framework applied by the processing plant. However, other criteria dependent on the input signal to switch the quantization strategy are also designed and are within the scope of this order.

[0013] Another important aspect of the invention is that the quantizer can be adaptive. In particular, the model in the model-based quantizer can be adaptive to fit the input audio signal. The model may vary over time, for example, depending on the characteristics of the input signal. This allows less distortion of the quantization and, thus, higher quality of coding.

[0014] According to one modality, the proposed quantization strategy is conditioned on the duration of the framework. It is suggested that the quantization unit may decide, based on the duration of the frame applied by the transformation unit, to encode the signal of the transformed domain with a model-based quantizer or a non-model-based quantizer. Preferably, the quantization unit is configured to encode a signal from the transformed domain, for a frame with a frame size less than a limit value, by means of model-based entropy-restricted quantization. Model-based quantization can be conditioned to varying parameters. Large frames can be quantized, for example, by a scalar quantizer, for example, with Huffman entropy coding, as used, for example, in the AAC codec.

[0015] The audio coding system may additionally comprise a long-term prediction unit (LTP) to estimate the filtered input signal frame based on a reconstruction of a previous segment of the input signal filtered, and a combination unit of the transformed domain signal, to combine, in the transformed domain, the estimation of the long-term prediction and the transformed input signal, to generate the signal of the transformed domain that is inserted in the quantization unit .

[0016] Switching between different methods of quantizing MDCT lines is another aspect of a preferred embodiment of the invention. By employing different quantization strategies for different transform sizes, the codec can do all quantization and coding in the MDCT domain, without the need to have a speech encoder for the specific time domain running in parallel or in series in the codec of the transformed domain. The present invention provides that, for speech-like signals, when there is an LTP gain, preferably the signal is encoded using a short transform and a model-based quantizer. The model-based quantizer is particularly suitable for the short transform, and provides, as will be outlined later, the advantages of a specific vector quantizer (VQ) of speech in the temporal domain, still being operated in the MDCT domain, and without any requirement that the input signal is a speech signal. In other words, when the model-based quantizer is used for the short transform segments in combination with the LTP, the VQ efficiency of the speech encoder of the dedicated time domain is retained, without loss of generality and without leaving the domain MDCT.

[0017] Furthermore, for more stationary music signals, it is preferred to use a relatively large transform, which is commonly used in audio codecs, and a quantization scheme that can take advantage of the sparse spectral lines discriminated by the large transform. . Therefore, the present invention provides for the use of this type of quantization scheme for large transformed ones.

[0018] Thus, switching the quantization strategy according to the duration of the frame enables the codec to retain both the properties of a dedicated speech codec and the properties of a dedicated audio codec, simply by choosing the size of the transform. This avoids all the problems of prior technology systems, which try to treat speech and audio signals equally well at low rates, since these systems inevitably encounter the problems and difficulties of efficiently combining time domain coding. (the speech encoder) with frequency domain encoding (the audio encoder).

[0019] According to another aspect of the invention, quantization uses adaptive step sizes. Preferably, the size (s) of the quantization step for components of the transformed domain signal is (are) adapted based on parameters of linear prediction and / or long-term prediction. The size (s) of the quantization step can be additionally configured to depend on the frequency. In embodiments of the invention, the size of the quantization step is determined based on at least one of: the adaptive filter polynomial, a coding rate control parameter, a long-term prediction gain value and a variance of the input signal.

[0020] Preferably, the quantization unit comprises uniform scalar quantizers to quantize the components of the transformed domain signal. Each scalar quantizer is applying a uniform quantization, for example, based on a probability model, on an MDCT line. The probability model can be a Laplacian or Gaussian model, or any other probability model that is suitable for signal characteristics. The quantization unit can additionally insert a random displacement into the uniform scalar quantizers. The insertion of random displacement provides advantages of vector quantization to uniform scalar quantizers. According to a modality, random displacements are determined based on an optimization of a quantization distortion, preferably on a perceptive domain and / or on cost in terms of the number of bits required to encode the indices quantization.

[0021] The quantization unit can additionally comprise an arithmetic encoder for coding the quantization indices generated by the uniform scalar quantizers. This achieves a low bit rate, which approaches the minimum possible given by the signal entropy.

[0022] The quantization unit may additionally comprise a residual quantizer to quantize a residual quantization signal resulting from uniform scalar quantizers, in order to further reduce the total distortion. Preferably, the residual quantizer is a fixed rate vector quantizer.

[0023] Multiple quantization reconstruction points can be used in the encoder and / or inverted quantizer quantization unit in the decoder. For example, reconstruction points with minimum mean square error (MMSE) and / or central point (midpoint) can be used to reconstruct a quantized value based on its quantization index. A quantization reconstruction point can additionally be based on a dynamic interpolation between a central point and an MMSE point, possibly controlled by data characteristics. This makes it possible to control the insertion of noise and avoid spectral holes due to the assignment of MDCT lines to a zero quantization position for low bit rates.

[0024] Preferably, a perceptual weighting in the transformed domain is applied when determining the quantization distortion, in order to place different weightings on specific frequency components. Perceptual weightings can be efficiently derived from linear prediction parameters.

[0025] Another independent aspect of the invention refers to the general concept of making use of the coexistence of LPC and SCF (Scale Factor) data. In a transform-based encoder, for example, that applies a Modified Discrete Cosine Transform (MDCT), scale factors can be used in quantization to control the size of the quantization step. In previous technology, these scale factors are estimated from the original signal to determine a masking curve. Now, it is suggested to estimate a second set of scale factors with the help of a perceptual filter or psychoacoustic model that is calculated from the LPC data. This allows for a reduction in the cost of transmitting / storing the scaling factors by transmission / storage only from the difference between the scaling factors actually applied and the scaling factors estimated by LPC, instead of the transmission / storage of the scaling factors. real scale. Thus, in an audio coding system that contains elements of speech coding, such as, for example, an LPC, and elements of transformed coding, such as an MDCT, the present invention reduces the cost of transmitting factor information scale required for the transformed coding part of the codec, by exploiting the data provided by the LPC. It is noticed that this aspect is independent of other aspects of the proposed audio coding system and can also be implemented in other audio coding systems.

[0026] For example, a perceptual masking curve can be estimated based on the parameters of the adaptive filter. The second set of scale factors based on linear prediction can be determined based on the estimated perceptual masking curve. Then, the stored / transmitted information of the scale factor is determined based on the difference between the scale factors actually used in the quantization and the scale factors that are calculated from the perceptual masking curve based on LPC. This removes dynamics and redundancy from the information stored / transmitted, so that fewer bits are needed to store / transmit the scale factors.

[0027] In the case where the LPC and MDCT do not operate at the same frame rate, that is, they have different frame sizes, the scale factors based on linear prediction for a frame of the transformed domain signal can be estimated with based on interpolated linear prediction parameters, to correspond to the time window covered by the MDCT table.

[0028] Therefore, the present invention provides an audio coding system that is based on a transform encoder, and includes fundamental prediction and modeling modules of a speech encoder. The inventive system comprises a linear prediction unit to filter an input signal based on an adaptive filter; a transformation unit, to transform a frame of the filtered input signal into a transformed domain; a quantization unit, to quantize a signal from the transformed domain; a scale factor determination unit, for generating scale factors based on a masking limit curve, for use in the quantization unit during the quantization of the transformed domain signal; a unit of estimation of the scale factor of linear prediction, to estimate scale factors based on linear prediction based on the parameters of the adaptive filter; and a scale factor encoder to encode the difference between the scale factors based on the masking limit curve and the scale factors based on linear prediction. By encoding the difference between the applied scale factors and the scale factors that can be determined in the decoder based on available linear prediction information, the efficiency of encoding and storage can be increased, and only a few bits need to be stored / transmitted.

[0029] Another specific aspect of the encoder independent of the invention concerns the treatment of the bit reservoir for variable frame sizes. In an audio coding system that can encode frames of varying duration, the bit reservoir is controlled by the distribution of available bits between the frames. Given a measure of reasonable difficulty for the individual frames and a reservoir of bits of a defined size, a certain deviation from the required constant bit rate allows a higher total quality, without a violation of the temporary storage requirements that are imposed by the bit reservoir size. The present invention extends the concept of using a bit reservoir for controlling the bit reservoir for a generalized audio codec with variable frame sizes. Therefore, an audio coding system may comprise a control unit of the bit reservoir, for determining the number of bits allowed to encode a frame of the filtered signal based on the duration of the frame and a measure of difficulty of the frame. painting. Preferably, the bit reservoir control unit has separate control equations for different frame difficulty measures and / or different frame sizes. Difficulty measures for different frame sizes can be standardized, so they can be compared more easily. In order to control the bit allocation for a variable rate encoder, preferably the bit reservoir control unit sets the lowest allowable limit of the bit control algorithm granted in the average number of bits to the highest allowed frame size.

[0030] An additional aspect of the invention relates to the treatment of a reservoir of bits in an encoder that employs a model-based quantizer, for example, an Entropy Restriction Quantizer (ECQ). It is suggested to minimize the variation in the size of the ECQ stage. A particular control equation that relates the size of the quantizer step to the ECQ rate is suggested.

[0031] Preferably, the adaptive filter to filter the input signal is based on an analysis of Linear Prediction Coding (LPC), including an LPC filter that produces a bleached input signal. LPC parameters for the present input data frame can be determined by algorithms known in the art. An estimation unit of the LPC parameter can calculate, for the input data frame, any suitable representation of LPC parameter, such as polynomials, transfer functions, reflection coefficients, spectral line frequencies, etc. The particular type of representation of the LPC parameter that is used for coding or other processing depends on the respective requirements. People skilled in the art realize that some representations are more suitable for certain operations than others and, therefore, are preferred to perform these operations. The linear prediction unit can operate on a first frame size that is fixed, for example, 20 ms. Linear prediction filtering can additionally operate on a deformed frequency geometric axis to selectively emphasize certain frequency ranges, such as low frequencies, over other frequencies.

[0032] Preferably, the transform applied in the frame of the filtered input signal is a Modified Discrete Cosine Transform (MDCT) that operates in a second variable frame size. The audio coding system may comprise a window sequence control unit, which determines, for a block of the input signal, the frame durations for overlapping MDCT windows by minimizing a function of the coding cost, pre¬ ferably, a simplistic perceptual entropy, for the entire block of the input signal that includes several frames. Thus, an optimal segmentation of the input signal block is derived in the MDCT windows with respective second frame durations. As a result, a transformed domain coding structure is proposed, including elements of the speech encoder, with an MDCT frame of adaptive duration only as a basic unit for all processing, except LPC. Since the durations of the MDCT frame can take many different values, an ideal sequence can be found, and abrupt changes in the frame duration can be avoided, as is common in prior art, where only a small window size and a large size windows are applied. Furthermore, transitional transformed windows with sharp edges, as used in some previous technology approaches for the transition between small and large window sizes, are not necessary.

[0033] Preferably, consecutive MDCT window lengths change, at most, by a factor of two (2) and / or the MDCT window lengths are dyadic values. More particularly, the MDCT window lengths can be dyadic partitions of the input signal block. Therefore, the MDCT window sequence is limited to predetermined sequences that are easy to encode with a small number of bits. In addition, the window sequence has smooth frame size transitions, thereby excluding abrupt changes in the frame's duration.

[0034] The window sequence control unit can be additionally configured to take into account the long-term prediction estimates, generated by the long-term prediction unit, for candidate window durations, during the search for the sequence of MDCT window durations that minimize the function of the coding cost for the input signal block. In this modality, the long-term prediction loop is closed during the determination of the MDCT window durations, which results in a better sequence of MDCT windows applied for coding.

[0035] The audio encoding system may additionally comprise an LPC encoder for recursive encoding, at a variable rate, of spectral line frequencies or other appropriate LPC parameter representations, generated by the linear prediction unit, for storage and / or transmission to a decoder. According to one modality, a linear prediction interpolation unit is provided to interpolate linear prediction parameters generated at a rate corresponding to the first frame size to correspond to the variable durations of the transformed domain signal frame.

[0036] According to one aspect of the invention, the audio coding system may comprise a perceptual modeling unit that modifies a characteristic of the adaptive filter, by hissing and / or oscillating an LPC polynomial generated by linear prediction unit for an LPC frame. The perceptual model received by modifying the characteristics of the adaptive filter can be used for many purposes in the system. For example, it can be applied as a perceptual weighting function for quantization or long-term prediction.

[0037] Another aspect of the invention relates to long-term prediction (LTP), in particular, long-term prediction in the MDCT domain, LTP adapted by MDCT framework and MDCT weighted LTP search. These aspects are applicable, regardless of whether an LPC analysis is present upstream of the transform encoder.

[0038] According to a modality, the audio coding system additionally comprises an inverted quantization and inverted transformation unit, for generating a reconstruction of the temporal domain of the filtered input signal frame. Furthermore, a long-term prediction buffer can be provided to store temporal domain reconstructions of previous frames of the filtered input signal. These units can be arranged in a feedback loop from the quantization unit to a long-term prediction extraction unit that seeks, in the long-term prediction temporary storage, the reconstructed segment that best corresponds to the present frame of the signal. filtered input. In addition, a long-term prediction gain estimation unit can be provided that adjusts the gain of the selected segment of the long-term prediction temporary storage, so that it best corresponds to the present table. Preferably, the long-term prediction estimation is subtracted from the transformed input signal in the transformed domain. Therefore, a second transformed unit can be provided to transform the selected segment in the transformed domain. The long-term prediction loop may additionally include adding the long-term prediction estimate in the transformed domain to the feedback signal, after the inverted quantization and before the inverted transformation in the temporal domain. Thus, a regressive adaptive long-term prediction scheme that predicts, in the transformed domain, the present frame of the filtered input signal based on previous frames. In order to be more efficient, the long-term prediction scheme can be additionally adapted in different ways, as set out below for some examples.

[0039] According to one modality, the long-term prediction unit comprises a long-term prediction extractor, for determining a delay value that specifies the reconstructed segment of the filtered signal that best fits the current frame of the filtered signal. A long-term prediction gain estimator can estimate a gain value applied to the signal of the selected segment of the filtered signal. Preferably, the delay value and the gain value are determined to minimize a distortion criterion related to the difference, in a perceptive domain, between the estimation of the long-term prediction and the transformed input signal. A modified linear prediction polynomial can be applied as an MDCT domain equalization gain curve while minimizing the distortion criterion.

[0040] The long-term prediction unit may comprise a transformation unit to transform the reconstructed segment signal from the LTP temporary storage into the transformed domain. For an efficient implementation of an MDCT transform, the transform is preferably a Discrete Cos¬sine Transform type IV.

[0041] Another aspect of the invention relates to an audio decoder for decoding the bit stream generated by the exposed encoder modalities. A decoder according to a modality comprises a unit of quantization, to dequantize a frame of an input bit stream, based on scale factors; an inverted transformation unit, to inversely transform a signal from the transformed domain; a linear prediction unit, to filter the signal from the inversely transformed transformed domain; and a scale factor decoding unit to generate the scale factors used in the quantization based on the delta information of the received scale factor, which encodes the difference between the scale factors applied in the encoder and the scale factors which are generated based on adaptive filter parameters. The decoder can additionally comprise a scale factor de-termination unit, to generate scale factors based on a masking limit curve, which is derived from the linear prediction parameters for the present table. The scale factor deco unit can combine the received scale factor delta information and the scale factors generated based on linear prediction to generate scale factors for entry into the quantization unit.

[0042] A decoder according to another embodiment comprises a model-based quantization unit, to dequantize a frame from an input bit stream; an inverted transformation unit, to inversely transform a signal from the transformed domain; and a linear prediction unit to filter the signal from the inversely transformed transformed domain. The quantization unit may comprise a non-model-based quantizer and a model-based quantizer.

[0043] Preferably, the quantization unit comprises at least one adaptive probability model. The quantization unit can be configured to adapt the quantization according to the transmitted signal characteristics.

[0044] The dequantization unit can additionally decide on a dequantization strategy based on the control data for the decoded frame. Preferably, the quantization control data is received with the bit stream or is derived from the received data. For example, the dequantization unit decides the dequantization strategy based on the size of the frame transform.

[0045] According to another aspect, the quantization unit comprises adaptive reconstruction points. The quantization unit can comprise uniform scalar quantizers that are configured to use two quantization reconstruction points per quantization interval, in particular, a midpoint and an MMSE reconstruction point.

[0046] According to one modality, the quantization unit uses a model-based quantizer in combination with arithmetic codification.

[0047] Furthermore, the decoder can comprise many of the aspects described above for the encoder. In general, the decoder will mirror the operations of the encoder, although some operations are performed only on the encoder and do not have corresponding components on the decoder. Thus, what is disclosed to the encoder is also considerable as applicable to the deco¬difier, if not stated otherwise.

[0048] The exposed aspects of the invention can be implemented as a device, device, method or computer program that operates on a programmable device. Inventive aspects can additionally be incorporated into signals, data structures and continuous bit streams.

[0049] Thus, the application additionally discloses an audio coding method and an audio decoding method. An exemplary method of audio coding comprises the steps of: filtering an input signal based on an adaptive filter; transform a frame of the filtered input signal into a transformed domain; quantize the signal of the transformed domain; generate scale factors, based on a masking limit curve, for use in the quantization unit during the quantization of the signal of the transformed domain; estimate scale factors based on linear prediction based on adaptive filter parameters; and to encode the difference between the scale factors based on the masking limit curve and the scale factors based on linear prediction.

[0050] Another method of audio encoding comprises the steps of: filtering an input signal based on an adaptive filter; transform a frame of the filtered input signal into a transformed domain; and quantize the signal from the transformed domain; where the quantization unit decides, based on the characteristics of the input signal, to encode the transformed domain signal with a model-based quantizer or a non-model based quantizer.

[0051] An exemplary audio decoding method comprises the steps of: dequantifying a frame from an input bit stream based on scale factors; inversely transform a signal from the transformed domain; filter with linear prediction the signal of the transformed domain inversely transformed; estimate second scale factors based on adaptive filter parameters; and to generate the scale factors used in the quantization based on the information of the difference of the received scale factor and the second estimated scale factors.

[0052] Another method of audio encoding comprises the steps: de-quantizing a frame from an input bit stream; inversely transform a signal from the transformed domain; and filter with linear prediction the signal of the transformed domain inversely transformed; where dequantization uses a non-model-based quantizer and a model-based quantizer.

[0053] These are just examples of preferred methods of encoding / decoding audio and computer programs that are prescribed by the present application and that versed in the technique can derive from the following description of the exemplary modalities.

Brief Description of Drawings

[0054] The present invention will now be described by way of illustrative examples, without limiting the scope or spirit of the invention, in relation to the attached drawings, in which:

[0055] figure 1 illustrates a preferred embodiment of an encoder and a decoder, according to the present invention;

[0056] figure 2 illustrates a more detailed view of the encoder and decoder, according to the present invention;

[0057] figure 3 illustrates another embodiment of the encoder, according to the present invention;

[0058] figure 4 illustrates a preferred embodiment of the encoder, according to the present invention;

[0059] Figure 5 illustrates a preferred embodiment of the decoder, according to the present invention;

[0060] figure 6 illustrates a preferred method of encoding and decoding MDCT lines, in accordance with the present invention;

[0061] Figure 7 illustrates a preferred embodiment of the encoder and decoder, and examples of relevant control data transmitted from one to the other, in accordance with the present invention;

[0062] figure 7a is another illustration of aspects of the encoder, according to an embodiment of the invention;

[0063] figure 8 illustrates an example of a window sequence and the relationship between LPC data and MDCT data, according to an embodiment of the present invention;

[0064] figure 9 illustrates a combination of scale factor data and LPC data, according to the present invention;

[0065] figure 9a illustrates another embodiment of the combination of scale factor data and LPC data, according to the present invention;

[0066] figure 9b illustrates another simplified block diagram of an encoder and a decoder, in accordance with the present invention;

[0067] figure 10 illustrates a preferred embodiment of the translation of LPC polynomials to an MDCT gain curve, according to the present invention;

[0068] Figure 11 illustrates a preferred method of mapping the LPC parameters with a constant update rate for the data of the adaptive MDCT window sequence, according to the present invention;

[0069] figure 12 illustrates a preferred method of adapting the perceptual weighting filter calculation based on the size of the transform and the type of quantizer, according to the present invention;

[0070] figure 13 illustrates a preferred embodiment of the quantizer adaptation depending on the duration of the frame, according to the present invention;

[0071] figure 14 illustrates a preferred embodiment of the quantizer adaptation depending on the duration of the frame, according to the present invention;

[0072] figure 15 illustrates a preferred embodiment of adapting the size of the quantization step according to the LPC and LTP data, according to the present invention;

[0073] figure 15a illustrates how a delta curve is derived from the LPC and LTP parameters by means of a delta adaptation module;

[0074] Figure 16 illustrates a preferred modality of a quantizer based on a model that uses random displacements, according to the present invention;

[0075] figure 17 illustrates a preferred embodiment of a model-based quantizer according to the present invention;

[0076] figure 17a illustrates another preferred embodiment of a model-based quantizer according to the present invention;

[0077] figure 17b schematically illustrates an MDCT line decoder based on model 2150 according to a modality of the invention;

[0078] figure 17c schematically illustrates aspects of the quantizer pre-processing according to an embodiment of the invention;

[0079] figure 17d schematically illustrates aspects of the computation of the step size according to a modality of the invention;

[0080] figure 17e schematically illustrates an entropy-restricted encoder based on a model according to a mode of the invention;

[0081] figure 17f schematically illustrates the operation of a uniform scalar quantizer (USQ) according to an embodiment of the invention;

[0082] figure 17g schematically illustrates probability computations according to an embodiment of the invention;

[0083] figure 17h schematically illustrates a process of quantization according to one embodiment of the invention;

[0084] figure 18 illustrates a preferred embodiment of a bit reservoir control according to the present invention;

[0085] figure 18a illustrates the basic concept of a control of the bit reservoir;

[0086] figure 18b illustrates the concept of a control of the bit reservoir for variable frame sizes, according to the present invention;

[0087] figure 18c shows an exemplary control curve for controlling the bit reservoir according to a modality;

[0088] Figure 19 illustrates a preferred embodiment of the inverted quantizer using different reconstruction points, in accordance with the present invention.

Description of Preferred Modalities

[0089] The modalities described below are merely illustrative for the principles of the present invention for audio encoder and decoder. It is understood that modifications and variations of the arrangements and details described here will be apparent to others skilled in the art. Therefore, it is intended that they are limited only by the scope of the achievements and not by the specific details presented as a description and explanation of the modalities exposed here. Similar components of the modalities are numbered by similar reference numbers.

[0090] In figure 1, an encoder 101 and a decoder 102 are displayed. Encoder 101 takes the input signal from the time domain and produces a bit stream 103 subsequently transmitted to decoder 102. Decoder 102 produces an output waveform based on received bit stream 103. The output signal resembles psychoacoustically to the original input signal.

[0091] In figure 2, a preferred embodiment of encoder 200 and decoders 210 is illustrated. The input signal at encoder 200 passes through an LPC (Linear Prediction Coding) module 201 which generates a bleached residual signal for an LPC frame, with a first frame size, and the corresponding linear prediction parameters. In addition, gain normalization can be included in the LPC module 201. The residual signal from the LPC is transformed into the frequency domain by a module from the MDCT (Modified Discrete Cosine Transform) 202 that operates on a second variable frame size. In the encoder 200 represented in figure 2, a module of the LTP (Long Term Prediction) 205 is included. LTP will be elaborated in an additional embodiment of the present invention. The MDCT lines are quantized 203 and also quantized 204, in order to supply LTP temporary storage with a copy of the decoded output, as will be available to decoder 210. Due to the quantization distortion, this copy is called the reconstruction of the respective input signal. At the bottom of figure 2, the decoder 210 is shown. The deco¬difier 210 takes the quantized lines from the MDCT, dequantizes them 211, adds the contribution of the LTP 214 module, and makes an inverted MDCT transform 212, followed by a synthesis filter from the LPC 213.

[0092] An important aspect of the exposed modality is that the MDCT frame is the only basic unit for encoding, although the LPC has its own (and, in one mode, constants) frame duration and LPC parameters, which are also encoded. Modality starts from a transform encoder, and introduces fundamental prediction and modeling modules of a speech encoder. As will be discussed below, the duration of the MDCT frame is variable and is adapted to a block of the input signal by determining the ideal MDCT window sequence for the entire block, by minimizing a perceptual entropy cost function. simplistic. This allows scaling to maintain optimal time / frequency control. In addition, the proposed unified structure avoids switched or layered combinations of different coding paradigms.

[0093] In figure 3, parts of encoder 300 are schematically described in more detail. The bleached signal transmitted from the LPC 201 module in the encoder in figure 2 is inserted in the filter bank of the MDCT 302. Optionally, the analysis of the MDCT can be an analysis of the MDCT deformed in relation to time, which ensures that the signal step ( if the signal is periodic with a well-defined step) be constant in relation to the MDCT transform window.

[0094] In figure 3, the LTP 310 module is outlined in more detail. It comprises a temporary storage of LTP 311 that contains samples of the reconstructed time domain of the previous segments of the output signal. An LTP 312 extractor finds the best corresponding segment in the temporary storage of LTP 311, given the current input segment. A suitable gain value is applied to this segment by the gain unit 313 before it is subtracted from the segment that is currently inserted in the quantizer 303. Of course, in order to do the subtraction before quantization, the LTP 312 extractor also transforms the segment signal chosen in the MDCT domain. The LTP 312 extractor seeks the best gain and delay values that minimize an error function in the perceptive domain when combining the segment of the reconstructed previous output signal with the input frame of the transformed MDCT domain. For example, an average squared error (MSE) function between the transformed reconstructed segment of the LTP 310 module and the transformed input frame (ie, the residual signal after subtraction) is optimized. This optimization can be performed in a perceptual domain, in which frequency components (that is, MDCT lines) are weighted according to their perceptual importance. The LTP 310 module operates in units of the MDCT frame, and the encoder 300 considers one residual MDCT frame at a time, for example, for quantization in the quantization module 303. The delay and gain search can be performed in a perceptual domain . Optionally, the LTP can be frequency selective, that is, it adapts the gain and / or delay in relation to the frequency. An inverted quantization unit 304 and an inverted MDCT unit 306 are represented. MDCT can be deformed in relation to time, as explained later.

[0095] In figure 4, another modality of encoder 400 is illustrated. In addition to figure 3, the LPC 401 analysis is included for clarification. A DCT-IV 414 transform used to transform a selected signal segment in the MDCT domain is shown. In addition, several ways of calculating the minimum error for the selection of the LTP segment are illustrated. In addition to minimizing the residual signal shown in figure 4 (identified as LTP2 in figure 4), the difference between the transformed input signal and the signal from the MDCT domain de-quantized before being inversely illustrated is illustrated. transformed into a reconstructed temporal domain signal for storage in LTP 411 temporary storage (indicated as LTP3). Minimizing this MSE function will direct the LTP contribution towards an ideal similarity (as possible) of the transformed input signal and the reconstructed input signal, for storage in the temporary storage of LTP 411. Another alternative error function (indicated as LTP 1) is based on the difference of these signals in the temporal domain. In this case, the MSE between the filtered LPC input frame and the corresponding reconstruction of the temporal domain in the temporary storage of LTP 411 is mini-minimized. The MSE is advantageously calculated based on the duration of the MDCT framework, which may differ from the duration of the LPC framework. Additionally, the quantizer and dequantizer blocks are replaced by the 403 spectrum coding block and the 404 spectrum decoding blocks ("Spec enc" and "Spec dec") that can contain additional modules, in addition to quantization, as will be outlined in figure 6. Again, the MDCT and the inverted MDCT can be deformed in relation to time (WMDCT, IWMDCT).

[0096] In figure 5, a proposed decoder 500 is illustrated. The received bitstream spectrum data is inversely quantized 511 and added with an LTP contribution provided by an LTP extractor from an LTP 515 buffer. The LTP 516 extractor and the LTP gain unit 517 on the decoder 500 are also illustrated. The combined MDCT lines are synthesized in the time domain by an MDCT synthesis block, and the signal time domain is spectrally modeled by an LPC 513 synthesis filter.

[0097] In figure 6 the blocks "Spec dec" and "Spec enc" 403, 404 of figure 4 are described in more detail. The "Spec enc" block 603 shown on the right of the figure comprises, in one modality, a Harmonic Prediction analysis module 610, a TNS (Temporal Noise Modeling) analysis module 611, followed by a module of analysis ¬ scaling factor 612 of the MDCT lines, and finally, quantizing and coding the lines in an Enc 613 line module. The "Spec Dec" block of decoder 604, shown on the left of the figure, does the reverse process, this that is, the MDCT lines received are de-quantized in a Dec 620 line module, and the scaling is undone by a scale factor scaling module (SCF) 621. TNS 622 synthesis and Harmonic prediction synthesis 623 are applied.

• linhas da MDCT quantizadas;
• fatores de escala;
• representação do polinómio LPC;
• energia do segmento de sinal (por exemplo, variância do sinal);
• sequência de janela;
• dados da LTP.

[0098] In figure 7, a very general illustration of the inventive coding system is outlined. The exemplary encoder takes the input signal and produces a bit stream that contains, among other data:

• quantified MDCT lines;
• scale factors;
• representation of the LPC polynomial;
• energy of the signal segment (for example, signal variance);
• window string;
• LTP data.

[0099] The decoder according to the modality reads the provided bit stream and produces an audio output signal, which resembles psychoacoustically the original signal.

[00100] Figure 7a is another illustration of aspects of a coder 700 according to an embodiment of the invention. The 700 encoder comprises an LPC 701 module, an MDCT 704 module, an LTP 705 module (shown only simplified), a quantization module 703 and an inverted quantization module 704 to feed back reconstructed signals in the LTP 705. A step 750 estimation module is additionally provided to estimate the input signal pitch and a window sequence determination module 751 to determine the ideal MDCT window sequence for a larger signal block input (for example, 1 second). In this modality, the MDCT window sequence is determined based on an open loop approach, in which the candidate MDCT window size sequence is determined, which minimizes a function of the coding cost, for example, a simplistic perceptual entropy. The contribution of the LTP 705 module to the encoding cost function, which is minimized by the window sequence determination module 751, can optionally be considerable when searching for the ideal MDCT window sequence. Preferably, for each candidate window size evaluated, the best contribution of the long-term prediction to the MDCT table corresponding to the candidate window size is determined, and the respective cost of codification is estimated. In general, small MDCT frame sizes are more suitable for speech input, while long windows transformed with fine spectral resolution are preferred for audio signals.

[00101] Perceptual weights or a perceptual weighting function are determined based on the LPC parameters calculated by the LPC 701 module, which will be explained in more detail below. The perceptual weights are supplied in the LTP 705 module and in the quantization module 703, both operating in the MDCT domain, for weighting the contributions of error or distortion of the frequency components according to their respective perceptual importance. Figure 7a further illustrates which encoding parameters are transmitted to the decoder, preferably by an appropriate encoding scheme, as will be discussed below.

[00102] Next, the coexistence of LPC and MDCT data, and the emulation of the effect of LPC on MDCT, both for the opposite action and for the omission of real filtering, will be discussed.

[00103] According to one modality, the LP module filters the input signal, so that the spectral shape of the signal is removed, and the subsequent output of the LP module is a spectrally leveled signal. This is advantageous, for example, for the operation of the LTP. However, other parts of the codec that operate on the spectrally leveled signal can benefit from knowing which spectral shape of the original signal was prior to LP filtering. Since the encoder modules, after filtering, operate on the MDCT transform of the spectral level signal, the present invention provides that, if necessary, the spectral shape of the original signal prior to LP filtering can be reimposed in the MDCT representation of the signal spectrally leveled, by mapping the transfer function of the LP filter used (that is, the spectral envelope of the original signal) into a gain curve, or equalization curve, which is applied at the frequency positions of the MDCT representation of the spectral leveled signal. Conversely, the LP module can omit the actual filtering, and only estimate a transfer function that is subsequently mapped to a gain curve, which can be imposed on the signal's MDCT representation, thus removing the need for temporal domain filtering of the input signal.

[00104] A prominent aspect of modalities of the present invention is that an MDCT-based transform encoder is operated using a flexible window segmentation on a blanched LPC signal. This is outlined in figure 8, in which an exemplary MDCT window sequence is given, along with the LPC window work. Therefore, as is clear from the figure, the LPC operates in a constant frame size (for example, 20 ms), while the MDCT operates in a variable window sequence (for example, 4 to 128 ms). This allows the choice of ideal window durations for the LPC and the ideal window sequence to be MDCT, regardless.

[00105] Figure 8 further illustrates the relationship between the LPC data, in particular, the LPC parameters generated at a first frame rate and MDCT data, in particular, the MDCT lines, generated at a second variable rate . The down arrows in the figure symbolize LPC data that is interpolated between the LPC circles (circles), to correspond to the corresponding MDCT frames. For example, a perceptual weighting function generated by LPC is interpolated for time instances determined by the MDCT window sequence.

[00106] The upward arrows symbolize refinement data (ie control data) used for encoding MDCT lines. For AAC charts, these data are typically scaling factors, and for ECQ charts, typically, data are variance correction data, etc. The solid lines in relation to the dashed lines represent that data is the most "important" data for the codification of MDCT lines, given a certain quantizer. The double down arrows symbolize the spectral lines of the codec.

[00107] The coexistence of LPC and MDCT data in the encoder can be exploited, for example, to reduce the bit requirements of scale factors of the MDCT encoding, by considering an estimated perceptual masking curve of the LPC parameters. In addition, perceptual weighting derived from LPC can be used when determining the quantization distortion. As shown, and as will be discussed below, the quantizer operates in two modes, and generates two types of frames (ECQ frames and AAC frames) depending on the frame duration of the data received, that is, corresponding to the frame. MDCT or window size.

[00108] Figure 11 illustrates a preferred mode of mapping the LPC parameters with constant rate for the data of the adaptive MDCT window sequence. An LPC 1100 mapping module receives LPC parameters according to the LPC update rate. In addition, the LPC 1100 mapping module receives information in the MDCT window sequence. Then, it generates an LPC- to-MDCT mapping, for example, for mapping psychoacoustic data based on LPC to respective MDCT frames generated at the variable MDCT frame rate. For example, the LPC mapping module interpolates LPC polynomials or related data for time instances corresponding to the MDCT frames for use, for example, of the perceptual weights in the LTP module or quantizer.

[00109] Now, specifications of the perceptual model based on LPC are discussed by reference to figure 9. In one embodiment of the present invention, the LPC 901 module is adapted to produce a white output signal, by using linear prediction, for example, order 16, for a 16 kHz sample rate signal. For example, the LPC 201 module output in figure 2 is the residue after estimating and filtering the LPC parameter. The estimated LPC A (z) polynomial, schematically visualized in the lower left part of figure 9, can be hissed by a bandwidth expansion factor, and also oscillated, in an implementation of the invention, by modifying the first coefficient of reflection of the corresponding LPC polynomial. Chiado expands the peak bandwidth in the LPC transfer function, by moving the poles of the polynomial into the unit circle, thus resulting in smoother peaks. Oscillation allows to make the LPC transfer function smoother, in order to balance the influence of lower and higher frequencies. These modifications attempt to generate a perceptual masking curve A '(z) of the estimated LPC parameters that will be available both on the encoder and on the decoder side of the system. Details of the manipulation of the LPC polynomial are presented in figure 12, below.

[00110] The MDCT coding that operates on the LPC residue has, in an implementation of the invention, scale factors to control the quantizer resolution or the quantization step sizes (and thus the noise introduced by the quantization). These scale factors are estimated by a scale factor estimation module 960 in the original input signal. For example, the scaling factors are derived from an estimated perceptual masking limit curve of the original signal. In one embodiment, a separate frequency transform (possibly with a different frequency resolution) can be used to determine the masking limit curve, but this is not always necessary. Alternatively, the masking limit curve is estimated from the MDCT lines generated by the transform module. The right bottom part of figure 9 schematically illustrates scale factors generated by the scale factor estimation module 960, to control quantization, so that the quantization noise introduced is limited to inaudible distortions.

[00111] If an LPC filter is connected upstream of the MDCT transformation module, a blanched signal is transformed in the MDCT domain. Since this signal has a white spectrum, it is not well suited to derive a perceptual masking curve. Thus, an MDCT domain equalization gain curve, generated to compensate for spectrum bleaching, can be used when estimating the masking limit curve and / or scale factors. This is because the scale factors need to be estimated in a signal that has absolute spectrum properties of the original signal, in order to correctly estimate perceptual masking. The calculation of the equalization gain curve of the MDCT domain of the LPC polynomial is discussed in more detail in relation to figure 10 below.

[00112] A modality of the estimated sketched scale factor estimation scheme is outlined in figure 9a. In this mode, the input signal is inserted in the LP 901 module which estimates the spectral envelope of the input signal described by A (z), and also transmits said polynomial as a filtered version of the input signal. The input signal is filtered with the inverse of A (z), in order to obtain a spectral white signal subsequently used by other parts of the encoder. The filtered signal - ^ (n) is inserted into an MDCT 902 transformer unit, while polynomial A (z) is inserted into a MDCT 970 gain curve calculation unit (as outlined in figure 14 ). The gain curve estimated from the LP polynomial is applied to the MDCT coefficients or lines, in order to retain the spectral aging of the original input signal prior to the scaling factor estimation. The MDCT lines with adjusted gain are inserted in the 960 scale factor estimation module, which estimates the scale factors for the input signal.

[00113] Using the aforementioned approach, the data transmitted between the encoder and the decoder contains both the LP polynomial, from which the relevant perceptual information as well as a signal model can be derived when a model-based quantizer is used, as well as the scale factors commonly used in a transformed codec.

[00114] With more details, returning to figure 9, the LPC 901 module in the figure estimates, from the input signal, a spectral envelope A (z) of the signal and derives, from this, a perceptual representation A (z). Furthermore, scale factors normally used in transform-based perceptual audio codecs are estimated on the input signal, or they can be estimated on the white signal produced by an LP filter, if the transfer function of the LP filter is taken into account when estimating the scale factor (described in the context of figure 10 below). The scaling factors can then be adapted in the scaling factor 961 adaptation module, given the LP polynomial, as will be outlined below, in order to reduce the bit rate required to transmit scaling factors.

[00115] Normally, the scale factors are transmitted to the decoder, as well as the LP polynomial. Now, given that both are estimated from the original input signal and that both are somewhat correlated to the absolute spectrum properties of the original input signal, it is proposed to encode a delta representation between the two, in order to remove any redundancy that can occur if both were transmitted separately. According to a modality, this correlation is explored as follows. Since the LPC polynomial, when correctly squeaked and oscillated, attempts to represent a masking limit curve, the two representations can be combined, so that the scaling factors of the transmitted encoder represent the difference between the two. desired scale factors and those that can be derived from the transmitted LPC polynomial. Therefore, the scale factor 961 adaptation module shown in figure 9 calculates the difference between the desired scale factors generated from the original input signal and the scale factors derived from the LPC. This aspect retains the ability to have an MDCT based quantizer that has the notion of the scale factors commonly used in transform encoders, in an LPC structure, operating in an LPC residue, and still has the possibility to switch to a model-based quantizer that derives quantization step sizes exclusively from linear prediction data.

[00116] In figure 9b, a simplified block diagram of the encoder and decoder is given according to a modality. The input signal in the encoder passes through the LPC 901 module, which generates a bleached residual signal and the corresponding linear prediction parameters. Additionally, gain normalization can be included in the LPC 901 module. The residual LPC signal is transformed in the frequency domain by an MDCT 902 transform. To the right of figure 9b, the decoder is shown. The decoder takes the quantized MDCT lines, dequantizes them 911 and applies an inverted MDCT transform 912, followed by an LPC 913 synthesis filter.

[00117] The bleached signal transmitted by the LPC 901 module in the encoder of figure 9b is inserted in the filter bank of the MDCT 902. The MDCT lines, as a result of the analysis of the MDCT, are transformed encoded with an encoding algorithm of the transformed, which consists of a perceptual model that guides the size of the desired quantization step for different parts of the MDCT spectrum. The values that determine the size of the quantization step are called scale factors, and there is a required scale factor value for each partition, called the scale factor band, of the MDCT spectrum. In algorithms encoding the transform of the previous technology, the scale factors are transmitted, through the bit stream, to the decoder.

[00118] According to one aspect of the invention, the estimated perceptual masking curve of the LPC parameters, as explained in relation to figure 9, is used when coding the scale factors used in the quantization. Another possibility to estimate a perceptual masking curve is to use the unmodified LPC filter coefficients to estimate the energy distribution in the MDCT lines. With this energy estimation, a psychoacoustic model, used in transformed coding schemes, can be applied both in the encoder and in the decoder, to obtain an estimation of a masking curve.

[00119] Then, the two representations of a masking curve are combined, so that the scale factors to be transmitted by the transform encoder represent the difference between the desired scale factors and those that can be de¬ derived from the transmitted LPC polynomial or the LPC-based psychoacoustic model. This feature retains the ability to have an MDCT-based quantizer that has the notion of scale factors commonly used in transform encoders, in an LPC structure, operating on an LPC residue, and still has the pos ¬ ability to control quantization noise based on a scale factor band according to the psychoacoustic model of the transform encoder. The advantage is that the transmission of the difference of the scale factors will cost less bits, when compared to the transmission of the values of the absolute scale factor, without considering the LPC data already present. Depending on the bit rate, frame duration or other parameters, the amount of residual scale factor to be transmitted can be selected. To have complete control of each scale factor band, a delta scale factor can be transmitted with an appropriate noise-free coding scheme. In other cases, the cost of transmitting scale factors can be further reduced by a more coarse representation of the differences in the scale factor. The special case with lower overprocessing is when the scale factor difference is set to 0 for all bands and no additional information is transmitted.

[00120] Figure 10 illustrates a preferred mode of translating LPC polynomials into an MDCT gain curve. As outlined in figure 2, MDCT operates on a bleached signal, bleached by the LPC filter 1001. In order to retain the spectral envelope of the original input signal, an MDCT gain curve is calculated by the MDCT gain curve module. 1070. The MDCT domain equalization gain curve can be obtained by estimating the magnitude response of the spectral envelope described by the LPC filter, for the frequencies represented by the positions in the MDCT transform. Then, the gain curve can be applied to the MDCT data, for example, during the calculation the minimum mean square error sign outlined in figure 3 or during the estimation of a perceptual masking curve for determining the scale factor as outlined in relation to figure 9 exposed.

[00121] Figure 12 illustrates a preferred method of adapting the calculation of the perceptual weighting filter based on the size of the transform and / or the type of quantizer. The LP A (z) polynomial is estimated by the LPC 1201 module in Figure 16. An LPC 1271 parameter modification module receives LPC parameters, such as the LPC A (z) polynomial, and generates a perceptual weighting A '(z), by modifying the LPC parameters. For example, the bandwidth of the LPC polynomial A (z) is increased and / or the polynomial is oscillated. The parameters entered in the hiss and oscillation module of adaptation 1272 are the standard hiss and oscillation values p and lambda y. These are modified predetermined rules given, based on the size of the transform used and / or the quantization strategy Q used. The modified hiss and oscillation parameters p 'and y' are inserted in the parameter modification module LPC 1271, translating the spectral envelope of the input signal, represented by A (z), into a perceptual masking curve represented by A ' (z).

[00122] Next, the quantization strategy based on the duration of the frame will be explained, and the quantization based on a model conditioned on varied parameters according to a modality of the invention. One aspect of the present invention is that it uses different quantization strategies for different transform sizes or frame sizes. This is illustrated in figure 13, where the frame duration is used as a selection parameter to use a model-based quantizer or a non-model-based quantizer. It is noticed that this aspect of quantization is independent of other aspects of the disclosed encoder / decoder and can also be applied in other codecs. An example of a non-model based quantizer is the Huffman table based quantizer used in the AAC audio coding standard. The model-based quantizer can be an Entropy Restriction Quantizer (ECQ) that employs arithmetic coding. However, other quantizers can also be used in embodiments of the present invention.

[00123] According to an independent aspect of the present invention, it is suggested to switch between different quantization strategies depending on the duration of the frame, in order to be able to use the ideal quantization strategy, given a particular duration of the frame. As an example, the window sequence may indicate the use of a long transform for a very stationary tonal music segment of the signal. For this type of signal in particular, using a long transform, it is highly beneficial to employ a quantization strategy that can take advantage of the "sparse" character (ie, well-defined discrete tones) in the signal spectrum. A quantization method used in AAC, in combination with Huffman tables and grouping of spectral lines, also used in AAC, is very beneficial. However, and on the contrary, for speech segments, the sequence of windows can, given the gain of LTP coding, indicate the use of short transforms. For this type of signal and size of the transform, it is beneficial to employ a quantization strategy that does not try to find or introduce spreading in the spectrum, but instead maintains a broadband energy that, given LTP, will retain the character ¬ pulse type characteristic of the original input signal.

[00124] A more general view of this concept is given in figure 14, in which the input signal is transformed into the MDCT domain and subsequently quantized by a quantizer controlled by the size of the transform or frame duration used for the MDCT transform .

[00125] According to another aspect of the invention, the step size of the quantizer is adapted according to the LPC and / or LTP data. This allows a determination of the size of the step depending on the difficulty of a frame and the controls on the number of bits that are allocated for encoding the frame. In figure 15, an illustration is given of how model-based quantization can be controlled by LPC and LTP data. In the top part of figure 15, a schematic view of the MDCT lines is given. Below the size of the quantization step, delta A is represented as a function of frequency. It is clear from this particular example that the size of the quantization step increases with frequency, that is, more distortion of quantization is incurred for higher frequencies. The delta curve is derived from the LPC and LTP parameters through a delta adaptation module shown in figure 15a. The delta curve can additionally be derived from the prediction polynomial A (z) by hissing and / or oscillation, as explained in relation to figure 13.

em que A(z) é o polinómio LPC, T é um parâmetro de oscilação, p con¬trola o chiado e n é o primeiro coeficiente de reflexão calculado a partir do polinómio A(z). Percebe-se que o polinómio A(z) pode ser recalcu¬lado em uma variedade de diferentes representações, a fim de extrair informação relevante do polinómio. Se houver interesse na inclinação espectral, a fim de aplicar uma "oscilação" para se opor à inclinação do espectro, o recálculo do polinómio em relação aos coeficientes de reflexão é preferido, já que o primeiro coeficiente de reflexão represen¬ta a inclinação do espectro.[00126] A preferred perceptual weighting function derived from LPC data is given in the following equation:

where A (z) is the LPC polynomial, T is an oscillation parameter, p controls the wheezing and n is the first reflection coefficient calculated from the A (z) polynomial. It is realized that polynomial A (z) can be recalculated in a variety of different representations, in order to extract relevant information from the polynomial. If there is an interest in spectral slope, in order to apply an "oscillation" to oppose the spectrum slope, recalculating the polynomial in relation to the reflection coefficients is preferred, since the first reflection coefficient represents the spectrum slope .

[00127] Furthermore, delta A values can be adapted depending on the variance of the input signal p, the gain of LTP g, and the first reflection coefficient n derived from the prediction polynomial. For example, adaptation can be based on the following equation:

[00128] Next, aspects of a quantizer based on a model according to a modality of the present invention are outlined. In figure 16, one of the aspects of the model-based quantizer is visualized. MDCT lines are inserted into a quantizer that employs uniform scalar quantizers. Furthermore, random displacements are inserted into the quantizer and used as displacement values for the quantization intervals that move the limits of the interval. The proposed quantizer provides advantages of vector quantization, while still maintaining the scalability of scalar quantizers. The quantizer iterates over a set of different displacement values, and calculates the quantization error for these. The displacement value (or displacement vector value) that minimizes the quantization distortion for the particular MDCT lines being quantized is used for quantization. Then, the offset value is transmitted to the decoder along with the quantized MDCT lines. The use of random displacements introduces noise filling in the quantized decoded signal and, in doing so, avoids spectral holes in the quantized spectrum. This is particularly important for low bit rates, where many MDCT lines are otherwise quantized to zero, which will lead to audible holes in the reconstructed signal spectrum.

[00129] Figure 17 schematically illustrates a MDCT Line Quantizer based on Model (MBMLQ) according to an embodiment of the invention. The top of figure 17 represents an MBMLQ 1700 encoder. The MBMLQ 1700 encoder takes MDCT lines in an MDCT frame or residual LTP MDCT lines as input if an LTP is present in the system. MBMLQ employs statistical models of MDCT lines, and source codes are adapted to the signal properties based on MDCT frame by frame, producing efficient compression in a bit stream.

[00130] A local gain of the MDCT lines can be estimated as the RMS value of the MDCT lines, and the MDCT lines normalized in the 1720 gain normalization module before insertion into the MBMLQ 1700 encoder. The local gain normalizes the MDCT lines and is a complement to LP gain normalization. While the LP gain adapts to variations in the signal level over a longer time scale, the local gain adapts to variations in a lesser time scale, producing greater quality of transient sounds and speech starts. Local gain is encoded by either fixed rate or variable rate coding and transmitted to the decoder.

[00131] A 1710 rate control module can be employed to control the number of bits used to encode an MDCT frame. A rate control index tracks the number of bits used. The rate control index points to a list of nominal quantizer step sizes. The table can be classified with step sizes in descending order (see figure 17g).

[00132] The MBMLQ encoder is executed with a set of different rate control indexes, and the rate control index which produces a bit count less than the number of bits allocated, given by the control of the bit reservoir, is used for the frame. The rate control index varies slowly, and this can be exploited to reduce complexity of the search and to efficiently code the index. The set of indexes that are tested can be reduced if the test is started around the index in the previous MDCT table. Likewise, efficient entropy coding of the index is obtained if the probabilities reach the maximum around the previous value of the index. For example, for a list of 32 step sizes, the rate control index can be encoded using 2 bits per MDCT frame on average.

[00133] Figure 17 schematically illustrates the MBMLQ 1750 decoder, in which the MDCT frame is renormalized in relation to the gain if a local gain is estimated in the 1700 encoder.

[00134] Figure 17a schematically illustrates in more detail the MDCT line encoder based on model 1700, according to a modality. It comprises a 1730 quantizer preprocessing module (see figure 17c), an entropy restriction encoder based on the 1740 model (see figure 17e), and a 1720 arithmetic encoder that can be an arithmetic encoder of the prior art . The task of the 1730 quantizer pre-processing module is to adapt the MBMLQ encoder to signal statistics, on a MDCT frame-by-frame basis. It takes other codec parameters as input, and derives useful statistics about the signal from them, which can be used to modify the behavior of the entropy-based encoder based on the 1740 model. The entropy-based encoder based on the model 1740 is controlled, for example, by a set of control parameters: a step size of quantizer A (delta, interval duration), a set of variance estimates of MDCT V lines (a vector; an estimated value per line MDCT), a perceptual masking curve PmOd, a matrix or table of displacements (random), and a statistical model of the MDCT lines that describe the shape of the distribution of the MDCT lines and their interdependencies. All of the aforementioned control parameters may vary between MDCT frames.

[00135] Figure 17b schematically illustrates an MDCT line decoder based on model 1750 according to a modality of the invention. It takes bits of complementary information from the bit stream as input and decodes them into parameters that are inserted in the preprocessor module of the 1760 quantizer (see figure 17c). Preferably, the 1760 quantizer preprocessing module has exactly the same functionality in the 1700 encoder as in the 1750 decoder. The parameters that are entered in the 1760 quantizer preprocessing module are exactly the same in the encoder and in the decoder. The preprocessor module of the 1760 quantizer transmits a set of control parameters (even from the 1700 encoder) and these are inserted in the 1770 probability computation module (see figure 17g; even from the encoder, see figure 17e) and in 1780 quantization module (see figure 17h; even for the encoder, see figure 17e). The cdf tables of the 1770 probability computations module, which represent the probability density functions for all MDCT lines, given the delta used for quantization and the signal variance, are inserted in the arithmetic decoder (which can be any arithmetic coder known to those skilled in the art) which then decodes the MDCT line bits into MDCT line indices. Then, the MDCT line indices are dequantized into MDCT lines by the 1780 dequantization module.

[00136] Figure 17c schematically illustrates aspects of the quantizer pre-processing according to a modality of the invention, which consist of: i) computation of the step size, ii) modification of the perceptual masking curve, iii) estimation of the variance of the MDCT lines, iv) construction of the displacement table.

[00137] The computation of the step size is explained in more detail in figure 17d. It comprises: i) a lookup table, in which the rate control index points to a step size table, produces a nominal step size AnOm (delta_nom), ii) low energy adaptation and iii) passive adaptation high.

[00138] Normally, gain normalization results in high energy and low energy sounds being encoded with the same segmental SNR. This can lead to an excessive number of bits being used in low energy sounds. The proposed low-energy adaptation allows for fine adjustment of a hit between low energy and high energy sounds. The step size can increase when the signal energy becomes low, as shown in figure 17d-ii), in which an exemplary curve is shown for the relationship between the signal energy (gain g) and a control factor qi_e. The gain of signal g can be computed as the RMS value of the input signal itself or of the residual LP. The control curve in figure 17d-ii) is just an example, and other control functions to increase the step size for low energy signals can be employed. In the example shown, the control function is determined by gradual line sections that are defined by Ti and T2 limits and the size factor of step L.

[00139] High-pass sounds are perceptibly less important than low-pass sounds. The high-pass adaptation function increases the step size when the MDCT frame is high-pass, that is, when the signal energy in the present MDCT frame is concentrated at the higher frequencies, resulting in fewer spent bits in such frames. If LTP is present and if the gi_TP LTP gain is close to 1, the residual LTP can become high pass; in a case like this, it is advantageous not to increase the step size. This mechanism is represented in figure 17d-iii), where r is the 1st reflection coefficient of the LPC. The proposed high-pass adaptation can use the following equation:

[00140] Figure 17c-ii) schematically illustrates the modification of the perceptual masking curve that employs a low frequency increase (LF) to remove "snoring type" coding artifacts. The LF increase can be fixed or adaptive, so that only a part below the first spectral peak is increased. The LF increase can be adapted by using the LPC envelope data.

[00141] Figure 17c-iii) schematically illustrates the estimation of the variance of the MDCT lines. With an active LPC bleaching filter, all MDCT lines have unit variance (according to the LPC envelope). After perceptual weighting in the entropy-restricted encoder based on the 1740 model (see figure 17e), the MDCT lines have variances that are the inverse of the quadratic perceptual masking curve, or the modified quadratic masking curve PmOd . If an LTP is present, it can reduce the variance of the MDCT lines. In figure 17c-iii), a mechanism is shown that adapts the estimated variances to the LTP. The figure shows a modification function ÇLTP in relation to frequency f. The modified variances can be determined by VLTPmod = V • qi / rp. The LLTP value can be a function of the LTP gain, so that LLTP is closer to 0 if the LTP gain is around 1 (indicating that LTP found a good match), and LLTP is closer to 1 if the LTP gain is around 0. The proposed LTP adaptation of the variances V = {vi, V2, ..., vj, ..., VN} affects only the MDCT lines below a certain frequency (ILTPcutoff). As a result, variances in the MDCT line below the cutoff frequency ÍLTPcutoff are reduced, the reduction being dependent on the LTP gain.

[00142] Figure 17c-iv) schematically illustrates the construction of the displacement table. The nominal displacement table is a matrix filled with pseudo-random numbers distributed between -0.5 and 0.5. The number of columns in the matrix equals the number of rows in the MDCT that are encoded by MBMLQ. The number of columns is adjustable and equals the number of displacement vectors that are tested in RD optimization on the entropy-restricted encoder based on the 1740 model (see figure 17e). The function of the construction of the displacement table scales the nominal displacement table with the size of the quantizer step, so that the displacements are distributed between -A / 2 and + A / 2.

[00143] Figure 17g schematically illustrates a modality for a displacement table. The displacement index is a pointer to the table and selects a chosen displacement vector O = {hi, 02, ..., On, ON}, where N is the number of MDCT lines in the MDCT table.

[00144] As described below, displacements provide a means for filling noise. Best objective and perceptual quality are obtained if the dispersion of displacements is limited to MDCT lines that have low variance Vj, compared to the size of the quantizer step A. An example of a limitation like this is described in figure 17c-iv), where ki and k2 are tuning parameters. The distribution of displacements can be uniform and distributed between + s. The limits can be determined according to

[00145] For low variance MDCT lines (where Vj is small, if compared to A), it may be advantageous to make the displacement distribution non-uniform and dependent on the signal.

[00146] Figure 17e schematically illustrates the entropy restriction encoder based on model 1740 in more detail. The inserted MDCT lines are perceptually weighted by their division with the values of the perceptual masking curve, preferably derived from the LPC polynomial, resulting in the vector of the weighted MDCT lines y = (yi, ..., yN). The purpose of the subsequent coding is to introduce white quantization noise in the MDCT lines in the perceptual domain. In the decoder, the inverse of perceptual weighting is applied, which results in quantization noise that follows the perceptual masking curve.

[00147] First, the iteration over random displacements is outlined. The following operations are performed for each column j in the displacement matrix: Each MDCT line is quantized by a uniform displacement scalar quantizer (USQ), where each quantizer is displaced by its own unique displacement value taken from the column vector displacement.

[00148] The probability of the minimum distortion interval for each USQ is computed in the 1770 probability computations module (see figure 17g). USQ indices are encoded by entropy. The cost, in terms of the number of bits required to encode the indexes, is computed as shown in figure 17e, producing a length of theoretical code word Rj. The overload limit of the USQ k of the MDCT j line can be computed as 3 v J, where k3 can be chosen to be any appropriate number, for example, 20. The overload limit is the threshold for which the error of quantization is greater than half the size of the quantization step in magnitude.

[00149] A scalar reconstruction value for each MDCT line is computed by the 1780 quantization module (see figure 17h), producing the quantized MDCT vector y. In the RD 1790 optimization module, a distortion Dj = d (y, y) is computed. d (y, y) can be the mean squared error (MSE), or another perceptually more relevant measure of distortion, for example, based on a perceptual weighting function. In particular, a distortion measure that weighs MSE and the divergence in energy between y and y can be used.

[00150] In the RD 1790 optimization module, a cost C is computed, preferably, based on the distortion Dj and / or the length of the theoretical code word Rj for each column j in the displacement matrix. An example of a cost function is C = 10 * logw (Dj) + À * Rj / N. The offset that minimizes C is chosen, and the corresponding USQ indices and probabilities are transmitted by the entropy-restricted encoder based on the 1780 model.

[00151] Optionally, the RD optimization can be further improved, by varying other properties of the quantizer along with the displacement. For example, instead of using the same fixed variance estimate V for each displacement vector that is tested in RD optimization, the variance estimate vector V can vary. Then, for the vector of the displacement column m, an estimate of variance km * V will be used, in which km can cover, for example, the range of 0.5 to 1.5, since m varies from m = 1 am = (number of columns in the displacement matrix). This makes entropy coding and MMSE computing less sensitive to variations in the input signal statistics that the statistical model cannot capture. This results in an overall lower C cost.

[00152] The quantified MDCT lines can be further refined by using a residual quantizer represented in figure 17e. The residual quantizer can be, for example, a fixed rate random vector quantizer.

[00153] The operation of the Uniform Scalar Quantizer (USQ) for quantization of line n of the MDCT is schematically illustrated in figure 17f, which shows the value of the line of the MDCT n in the minimum distortion interval with index in. The 'x' markings indicate the center (midpoint) of the quantization intervals with step size A. The origin of the scalar quantizer is displaced by the displacement on of displacement vector O = {01, o2, on, ON }. Thus, the limits and midpoints of the interval are displaced by the displacement.

[00154] The use of displacements introduces encoder-controlled noise filling in the quantized signal, and in doing so, avoids spectral holes in the quantized spectrum. In addition, displacements increase coding efficiency by providing a set of coding alternatives that fill the space more efficiently than a cubic structure. Also, offsets provide variation in the probability tables that are computed by the 1770 probability computations module, which leads to more efficient entropy coding of the MDCT row indices (ie, fewer required bits).

[00155] The use of a variable step size A (delta) allows variable precision in quantization, so that more precision can be used for perceptually important sounds, and less precision can be used for less important sounds.

[00156] Figure 17g schematically illustrates the probability computations in the 1770 probability computing module. The inputs for this module are the statistical model applied to the MDCT lines, the step size of the quantizer A, the vector of vari¬ tance V, the displacement index and the displacement table. The output of the 1770 probability computation module is cdf tables. For each MDCT line Xj, the statistical model (ie, it is a function of probability density, pdf) is evaluated. The area under the pdf function for an interval i is the pij probability of the interval. This probability is used for the arithmetic coding of MDCT lines.

[00157] Figure 17h illustrates schematically the quantization process carried out, for example, in the 1780 quantization module. The center of mass (MMSE value) XMMSE for the minimum distortion interval of each MDCT line is computed together with the XMP midpoint of the range. Considering that an N-dimensional vector of the MDCT lines is quantized, the scalar MMSE value is sub-ideal and, in general, very low. This results in a loss of variance and spectral imbalance in the decoded output. This problem can be mitigated by decoding the conservation of variance described in figure 17h, in which the reconstruction value is computed as a weighted sum of the MMSE value and the midpoint value. An additional optional improvement is to adapt the weighting so that the MMSE value dominates for speech and the midpoint dominates for non-speaking sounds. This produces speech, while spectral balance and energy are conserved for non-speaking sounds.

[00158] Decoding with variance conservation according to one embodiment of the invention is achieved by determining the reconstruction point according to the following equation:

[00159] Decoding with adaptive variance conservation can be based on the following rule for determining the interpolation factor:

[00160] For example, the adaptive weighting can be additionally depending on the prediction gain LTP gi_Tp: * = The adaptive weighting varies slowly and can be efficiently coded by a recursive entropy code.

[00161] The statistical model of the MDCT lines that is used in the probability computations (figure 17g) and in the quantization (figure 17h) must reflect the statistics of the real signal. In one version, the statistical model considers that the MDCT lines are independent and distributed by the Laplacian model. Another version models the MDCT lines as independent Gaussian models. One version models the MDCT lines as Gaussian mix models, including interdependencies between the MDCT lines in the MDCT tables and between them. Another version adapts the statistical model to online signal statistics. Adaptive statistical models can be adapted forwards and / or backwards.

[00162] Another aspect of the invention that relates to the modified reconstruction points of the quantizer is schematically illustrated in figure 19, in which an inverted quantizer used in the decoder of a modality is represented. The module also has, in addition to the normal inputs of an inverted quantizer, that is, the quantized lines and information about the size of the quantization step (type of quantization), information about the reconstruction point of the quantizer. The inverted quantizer of this modality can use multiple types of reconstruction points when determining a reconstructed value of the corresponding quantization index in. As exposed, reconstruction values y are additionally used, for example, in the MDCT line encoder (see figure 17) to determine the residual quantization for entry into the residual quantizer. In addition, quantization reconstruction is performed on the inverted quantizer 304, to reconstruct an encoded MDCT frame for use in LTP temporary storage (see figure 3) and, of course, in the decoder.

[00163] For example, the inverted quantizer can choose the midpoint of a quantization interval as the reconstruction point, or the MMSE reconstruction point. In one embodiment of the present invention, the quantizer reconstruction point is chosen to be the average value between the center and the MMSE reconstruction points. In general, the reconstruction point can be interpolated between the midpoint and the MMSE reconstruction point, for example, depending on signal properties, such as signal periodicity. Signal periodicity information can be derived from the LTP module, for example. This feature allows the system to control distortion and energy conservation. The central reconstruction point will guarantee energy conservation, while the MMSE reconstruction point will guarantee minimal distortion. Then, given the signal, the system can adapt the reconstruction point to where the best fit is provided.

[00164] The present invention additionally incorporates a new window sequence encoding format. According to an embodiment of the invention, the windows used for the MDCT transform are dyadic sizes, and can only vary by a factor of two, from window to window. For example, dyadic transform sizes are 64, 128, ..., 2048 samples corresponding to 4, 8, ..., 128 ms at a sampling rate of 16 kHz. In general, windows of varying size are proposed that can take a plurality of window sizes between a minimum window size and a maximum size. In a row, consecutive window sizes can vary by only a factor of two, so that smooth sequences of window sizes develop without sudden changes. The window sequences defined by a modality, that is, limited to dyadic sizes and allowed to vary only in factor two in size, from window to window, has several advantages. First, no specific start or end windows are needed, that is, windows with sharp edges. This maintains a good time / frequency resolution. Second, the window sequence becomes very efficient for encoding, that is, for signaling to a decoder which particular window sequence is used. Finally, the window sequence will always skilfully fit into a hyperframe structure.

[00165] The hyperframe structure is used during the operation of the encoder in a real-world system, in which certain configuration parameters of the decoder need to be transmitted in order to be able to start the decoder. This data is commonly stored in a header field in the bit stream that describes the encoded audio signal. In order to minimize bit rate, the header is not transmitted for each frame of the encoded data, particularly in a system proposed by the present invention, in which the MDCT frame sizes can vary from very small to very large. Therefore, the present invention proposes to group a certain number of MDCT frames into a hyperframe, in which the header data is transmitted at the beginning of the hyperframe. Typically, the hyperwater is defined with a specific length of time. Therefore, it is necessary to take care, so that the variations in the sizes of the MDCT frame are suitable for a predefined constant hyperframe duration. The exposed inventive window sequence ensures that the selected window sequence always fits into a hyperframe structure.

[00166] In accordance with an embodiment of the present invention, the LTP delay and the LTP gain are encoded in a variable rate manner. This is advantageous since, depending on the effectiveness of LTP for periodic stationary signals, the LTP delay tends to be the same during rather long segments. Therefore, this can be exploited through arithmetic encoding, resulting in a variable rate LTP delay and LTP gain encoding.

[00167] Similarly, one embodiment of the present invention also takes advantage of a bit rate and variable rate encoding reservoir for encoding LP parameters. Furthermore, encoding the recursive LP is prescribed by the present invention.

[00168] Another aspect of the present invention is the treatment of a bit reservoir for variable frame sizes in the encoder. In figure 18, a control unit of bit reservoir 1800 according to the present invention is outlined. In addition to a measure of difficulty provided as input, the control unit of the bit reservoir also receives information about the duration of the current frame. An example of a measure of difficulty for use in the control unit of the bit reservoir is perceptual entropy, or the logarithm of the power spectrum. Control of the bit reservoir is important in a system where the frame durations can vary over a set of different frame durations. The suggested bit reservoir control unit 1800 considers the frame size when calculating the number of bits allowed for the frame to be encoded, as will be outlined below.

[00169] The bit reservoir is defined here with a certain fixed amount of bits in a temporary storage that needs to be greater than the average number of bits that a frame is allowed to use for a given bit rate. If it is the same size, no variation in the number of bits for a frame is possible. The control of the bit reservoir always considers the level of the bit reservoir before eliminating bits that will be assigned to the encoding algorithm as the allowed number of bits for the real frame. Thus, a full bit reservoir means that the number of bits available in the bit reservoir equals the size of the bit reservoir. After encoding the frame, the number of bits used will be subtracted from the temporary storage and the bit pool is updated by adding the number of bits that represents the constant bit rate. Therefore, the bit pool is empty if the number of bits in the bit pool before encoding a frame is equal to the average number of bits per frame.

[00170] In figure 18a, the basic concept of the control of the bit reservoir is represented. The encoder provides a means of calculating how difficult it is to encode the actual frame, compared to the previous frame. For an average difficulty of 1.0, the number of bits allowed depends on the number of bits available in the bit pool. According to a given control line, more bits that correspond to an average bit rate will be eliminated from the bit reservoir, if the bit reservoir is almost full. In the case of an empty bit reservoir, fewer bits, compared to the average of bits, will be used for encoding the frame. This behavior produces an average level of the bit reservoir for a larger sequence of frames with medium difficulty. For frames with greater difficulty, the control line can be moved upwards, with the effect that frames that are difficult to encode use more bits at the same level as the bit reservoir. In this way, for easy-to-encode frames, the number of bits allowed for a frame will be smaller just by moving down the control line in figure 18a from the medium difficulty case to the easy difficulty case. Modifications other than simple displacement of the control line are also possible. For example, as shown in figure 18a, the slope of the control curve may change, depending on the difficulty of the frame.

[00171] When calculating the number of bits granted, the limits at the lower end of the bit reservoir need to be obeyed in order not to eliminate more bits from the temporary storage than allowed. A control scheme of the bit reservoir that includes the calculation of the bits granted by a control line, shown in figure 18a, is just an example of the possible relations of the level of the bit reservoir and the measure of difficulty by the bits granted. Also, other control algorithms will have in common the severe limits at the lower end of the bit reservoir level that prevents a bit reservoir from violating the restriction of the empty bit reservoir, also the limits at the upper end, where the encoder will be forced to write filled bits, if a very low number of bits is consumed by the encoder.

[00172] In order for a control mechanism like this to handle a set of variable frame sizes, this simple control rhythm needs to be adapted. The difficulty measure to be used needs to be normalized, so that the difficulty values of frames of different durations are comparable. For each duplication of the frame, there will be a different range allowed for the given bits and, because the average number of bits per frame is different for a variable frame size, consequently, each frame duration has its own control equation with its own limitations. An example is shown in figure 18b. An important modification of the case of the fixed duration frame is the lowest allowed limit of the control algorithm. Instead of the average number of bits for the actual frame size, which corresponds to the case of the fixed bit rate, the average number of bits for the largest allowed frame size is now the lowest allowed value for the bit reservoir level before clearing the bits for the real frame. This is one of the main differences in relation to the control of the bit reservoir for fixed frame sizes. This restriction ensures that a next frame with the largest possible frame size can use at least the average number of bits for this frame size.

[00173] The measure of difficulty can be based, for example, on a perceptual entropy (PE) calculation that is derived from the masking limits of a psychoacoustic model, as is done in AAC, or as an alternative to the bit counting of a quantization with size of the fixed step, as is done in the ECQ part of an encoder according to an embodiment of the present invention. These values can be normalized in relation to the variable frame sizes, which can be accomplished by a simple division by the frame size, and the result will be a PE respectively of a bit count per sample. Another stage of normalization can occur in relation to the average difficulty. For this purpose, a moving average in relation to past frames can be used, resulting in a difficulty value greater than 1.0 for difficult frames or less than 1.0 for easy frames. In the case of a two-step encoder or greater anticipation, difficulty values for future frames can also be taken into account for this normalization of the difficulty measure.

[00174] Another aspect of the invention relates to specifications of the treatment of the bit reservoir for ECQ. The management of the bit reservoir for ECQ works under the assumption that ECQ produces an approximately constant quality when using a constant quantizer step size for encoding. Constant quantizer step size produces a variable rate, and the purpose of the bit reservoir is to keep the variation in the size of the quantizer step between different frames as small as possible, while still not violating the reservoir's temporary storage restrictions. bits. In addition to the rate produced by the ECQ, additional information (for example, LTP gain and delay) is transmitted based on the MDCT framework. In general, the additional information is also encoded by entropy and thus consumes a different rate from frame to frame.

• RECQ_AVG: Taxa média do ECQ por amostra previamenteusada;
• AECQ_AVG: Tamanho médio da etapa do quantizador previ¬amente usada.

[00175] In one embodiment of the invention, a control of the proposed bit reserve tries to minimize the variation in the size of the ECQ step by introducing three variables (see figure 18c):

• RECQ_AVG: Average rate of ECQ per sample previously used;
• AECQ_AVG: Average size of the quantizer step previously used.

[00176] Both of these variables are dynamically updated to reflect the most recent encoding statistics.RECQ_AVG_DES: The ECQ rate corresponding to the average total bit rate.

[00177] This value will differ from RECQJWG in the event that the bit reservoir level changes during the time frame of the averaging window, for example, a bit rate higher or lower than the specified average bit rate was used during this time frame. It is also updated as the rate of supplementary information changes, so that the total rate equals the specified bit rate.

[00178] The bit reservoir control uses these three values to determine an initial assumption in the delta to be used for the current frame. It does this by finding AECG_AVG_DES on the RECQ-A curve shown in figure 18c, which corresponds to RECQ_AVG_DES. In a second stage, this value is possibly modified if the rate does not comply with the restrictions of the bit reservoir. The exemplary RECQ-A curve in figure 18c is based on the following equation:

[00179] Certainly, other mathematical relationships between RECQ and A can also be used.

[00180] In the stationary case, RECQ_AVG will be close to RECQ_AVG_DES, and the variation in A will be very small. In the non-stationary case, the averaging operation will guarantee a smooth variation of A.

[00181] Although the above has been disclosed in relation to the particular modalities of the present invention, it is understood that the inventive concept is not limited to the described modalities. On the other hand, the disclosure presented in this application will enable those skilled in the art to understand and carry out the invention. Experienced in the technique, they understand that several modifications can be made without departing from the spirit and scope of the invention, in the form exposed exclusively by the embodiments.

Claims (14)

Audio coding system, characterized by the fact that it comprises: a linear prediction (LP) unit (201) for filtering an audio signal based on an LP filter, the LP unit operating on a first frame size of the audio signal; an adaptive-length transformation unit (202) for transforming an audio signal frame into a transformed domain, the transformed being a Modified Discrete Cosine Transform (MDCT) operating on a second variable frame size; a quantization unit (203) for quantizing a signal from the MDCT domain; a gain curve generation unit (1470) for generating gain curves in the MDCT domain based on magnitude responses from the LP filter; and a mapping unit (1500) for mapping gain curves in the MDCT domain to corresponding signal frames of the MDCT domain. Audio coding system, according to king¬ indication 1, characterized by the fact that it comprises: a window sequence control unit to determine, for a block of the audio signal, the second frame durations for overlapping MDCT windows. Audio coding system according to claim 1 or 2, characterized by the fact that it comprises a perceptual modeling unit that modifies a characteristic of the LP filter by the hiss and / or oscillation of an LPC polynomial generated by the linear prediction unit for an LPC frame. Audio coding system according to any one of claims 1 to 3, characterized by the fact that it comprises: a frequency division unit for dividing the audio signal into a low band component and a high band component; and a high band encoder to encode the high band component, where the low band component is inserted into the linear prediction unit and the transformation unit. Audio coding system, according to King 4, characterized by the fact that the frequency division unit comprises a quadrature specular filter bank and a quadrature specular filter synthesis unit, configured to make descending sampling of the audio signal. Audio coding system, according to King 4 or 5, characterized by the fact that the limit between the low band and the high band is variable and the frequency division unit determines the crossing frequency based on on audio signal properties and / or encoder bandwidth requirements. Audio encoding system according to any one of claims 4 to 6, characterized by the fact that the high band encoder is a spectral band replication encoder. Audio encoding system according to any one of claims 1 to 7, characterized by the fact that it comprises: a scale factor estimation unit (1360) to estimate scale factors to control the quantization noise of the quantization unit (203). Audio coding system, according to Rei¬vindication 8, characterized by the fact that the scale factors are determined based on the gain curves in the mapped MDCT domain. Audio encoding system according to any one of claims 1 to 9, characterized by the fact that it comprises a parametric stereo unit for calculating a parametric stereo representation of the inserted left and right channels. Audio coding system according to any one of claims 1 to 10, characterized by the fact that the mapping unit (1500) interpolates gain curves in the MDCT domain generated at a rate corresponding to the first frame length to correspond to the frames of the signal of the MDCT domain generated at a rate corresponding to the second frame length. Audio decoder, characterized by the fact that it comprises: a quantizing unit (211) to dequantize a frame from an input bit stream and generate a signal from the transformed domain; an inverted MDCT transformation unit of adaptive length (212) for inversely transforming the transformed domain signal, the inverted MDCT transformation unit operating in a variable frame length; a gain curve generation unit (1470) for generating gain curves in the MDCT domain based on magnitude responses from the linear prediction filters, in which parameters for the linear prediction filters are received in the bit stream; and a mapping unit (1500) for mapping gain curves in the MDCT domain to corresponding signal frames of the MDCT domain. Audio coding method, characterized by the fact that it comprises the steps of: perform a linear prediction analysis (LP) on an audio signal, the LP analysis operating on a first frame length and generating LP filter parameters; transform a frame of the audio signal into a domain of the Modified Discrete Cosine Transform (MDCT), the MDCT operating in a second variable frame length; quantize a signal from the MDCT domain; generate gain curves in the MDCT domain based on magnitude responses from the generated LP filters; and map gain curves in the MDCT domain to corresponding signal frames of the MDCT domain. Audio decoding method, characterized by the fact that it comprises the steps of: dequantize a frame from an input bit stream and generate a signal from the transformed domain; make the inverted MDCT transform of the transformed domain signal, the inverted MDCT transform operating in a variable frame length; generate gain curves in the MDCT domain based on magnitude responses from linear prediction filters, in which parameters for the linear prediction filters are received in the bit stream; and map gain curves in the MDCT domain to corresponding signal frames of the MDCT domain.

Images