RU2793725C2 - Audio coder and decoder - Google Patents

Abstract

FIELD: audio data processing.

SUBSTANCE: technical result is achieved due to the fact that gain curves of the MDCT area for frames of MDCT coefficients are created based on the response values determined from the sequence of LPC polynomials, and the creation of gain curves in the MDCT area comprises mapping LPC polynomials to the corresponding frames of MDCT coefficients; gain curves in the MDCT area are applied to the MDCT coefficient frames to obtain gain-corrected MDCT coefficient frames; and frames of gain-corrected MDCT coefficients are quantized.

EFFECT: increase in the coding accuracy of both general audio signals and speech signals at low data rates.

16 cl, 34 dwg

Description

The field of technology to which the invention belongs

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

State of the art invention

In the prior art, there are speech encoders specifically designed to encode speech signals based on a signal source model, ie, the human voice system. These encoders cannot process arbitrary audio signals such as music or any other non-speech signal. Also, in the prior art, there are music encoders, commonly referred to as audio encoders, that base their coding on assumptions about the human auditory system rather than on a signal source model. These encoders can handle arbitrary signals very well, although at low speech rates a dedicated speech encoder gives excellent audio quality. Therefore, today there is no general arbitrary audio coding structure that is equally good as a speech encoder for speech and as a music encoder for music when operating at low data rates.

Thus, there is a need for an improved audio encoder and decoder with improved audio quality and/or reduced data rates.

The essence of the invention

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

The present invention addresses algorithms for audio codecs containing both linear predictive coding (LPC) and the signal-based part of the transform encoder based on LPC processing.

The present invention further relates to a quantization strategy depending on the size of the transformed data frame. Additionally, a model-based, entropy-limited quantizer using arithmetic coding is provided. In addition, insertion of random shifts into a uniform scalar quantizer is provided. The invention further provides a model-based quantizer, such as an entropy-constrained quantizer (ECQ), using arithmetic coding.

The present invention further relates to efficient scale factor coding in terms of transform coding of an audio encoder using the presence of LPC data.

The present invention further relates to an efficient implementation of the use of a bit accumulator in a variable frame size audio encoder.

The present invention further relates to an encoder for encoding audio signals and generating a data bitstream and a decoder for decoding a data bitstream and producing a reconstructed audio signal that is perceptually indistinguishable from the input audio signal.

The first aspect of the present invention relates to quantization in a transform encoder, which is applied, for example, in a modified discrete cosine transform (MDCT). The proposed quantizer preferably quantizes the MDCT lines. This aspect applies regardless of whether the encoder uses additional analysis of linear predictive coding (LPC) or additional long-term prediction.

The present invention provides an audio coding system comprising a linear predictor for filtering an input signal based on an adaptive filter; a transform unit for converting a frame of the filtered input signal into a transform region; and a quantizer for quantizing the signal in the transform domain. The quantizer, based on the characteristics of the input signal, decides whether to encode the transform domain signal with a model-based quantizer or a non-model-based quantizer. Preferably, the decision is based on the frame size applied by the transform block. However, other input-signal dependent criteria for switching the quantization strategy are also contemplated and are within the scope of the present application.

Another important aspect of the invention is that the quantizer can be adaptive. In particular, the model in the model-based quantizer may be adaptive in order to adjust the input audio signal. The model may, for example, change over time, for example, depending on the characteristics of the input signal. This makes it possible to reduce quantization distortion and thus improve coding quality.

According to an embodiment, the proposed quantization strategy is based on the frame size. It is proposed that the quantizer, based on the frame size applied by the transform unit, can decide whether to encode the signal in the transform domain with a model-based quantizer or a non-model-based quantizer. Preferably, the quantizer may be configured to encode the transform-domain signal for a frame with a frame size smaller than a threshold value by means of an entropy-limited model-based quantizer. Model-based quantization may be performed based on various parameters. Jumbo frames may be quantized, for example, by a scalar quantizer, for example, using Huffman entropy coding used, for example, in the AAC codec.

The audio coding system may further comprise a long-term prediction (LTP) block for estimating a frame of the filtered input signal based on the reconstruction of the previous segment of the filtered input signal and the signal in the transformation region of the combining block to combine in the transformation region the result of determining the long-term prediction and the transformed input signal to create a signal in the transform domain, which is the input signal for the quantizer.

Switching between different MDCT line quantization methods is another aspect of the preferred embodiment of the invention. By using different quantization strategies for different transform sizes, a codec can perform all of the quantization and coding in the MDCT domain without having to have a dedicated time domain speech encoder operating in parallel or in series with the transform domain codec. The present invention indicates that for speech-like signals where there is LTP gain, the signal is preferably encoded using a fast transform and a model-based canter. The model-based quantizer is particularly suitable for fast transformation and has, as will be described later, the advantages of a dedicated vector quantizer (VQ) of speech in the time domain, while still operating in the MDCT domain, and without any requirement so that the input signal is a speech signal. In other words, when a model-based quantizer is used for fast transform segments in combination with LTP, the performance of a dedicated VQ speech encoder operating in the time domain is maintained without loss of generality and without leaving the MDCT domain.

Furthermore, for more stationary music signals, it is preferable 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 distinguished by the large transform. Therefore, the present invention indicates the use of this kind of quantization scheme for long transforms.

Thus, switching the quantization strategy as a function of the frame size allows the codec to preserve both the properties of the dedicated speech codec and the properties of the dedicated audio codec, simply by selecting the transform size. This completely avoids the problems inherent in prior art systems that tend to process speech and audio signals equally well at low rates, as these systems inevitably face the problems and difficulties of effectively combining time domain coding (speech encoder) with frequency domain coding ( audio encoder).

In accordance with another aspect of the invention, quantization uses adaptive step sizes. Preferably, the quantization step size(s) for the signal components in the transform domain is/are adapted based on the linear prediction and/or long-term prediction parameters. The quantization step size(s) may further be configured to be frequency dependent. In embodiments of the invention, the quantization step size is determined based on at least one of the following: an adaptive filter polynomial, a code rate control parameter, a long term prediction gain value, and an input signal variance.

Preferably, the quantizer comprises homogeneous scalar quantizers for quantizing the signal components in the transform domain. Each scalar quantizer applies uniform quantization to the MDCT line, eg based on a probabilistic model. The probabilistic model may be a Laplace or Gaussian model, or any other probabilistic model suitable for signal characteristics. The quantization unit may additionally insert a random shift into the uniform scalar canting devices. The insertion of a random shift provides uniform scalar quantizers with the benefits of vector quantization. According to an embodiment, the random shifts are determined based on optimizing the quantization distortion, preferably in the perceptual domain, and/or considering the cost in terms of the number of bits required to encode the quantization indices.

The quantizer may further comprise an arithmetic encoder for encoding the quantization indices generated by the homogeneous scalar quantizers. This allows you to achieve a low data rate, approaching the possible minimum, which is set by the entropy of the signal.

The quantizer may further comprise a residual quantizer for quantizing the residual of the quantizer signal resulting from the operation of the uniform scalar quantizers to further reduce overall distortion. The residual quantizer is preferably a fixed frequency vector device.

Multiple quantization reconstruction points may be used in an encoder dequantizer and/or an inverse quantizer in a decoder. For example, a minimum mean square error (MMSE) reconstruction point and/or a reconstruction central point (midpoint) may be used to reconstruct a quantized value based on its quantization index. The quantization reconstruction point may further be based on dynamic interpolation between the center point and the MMSE point, possibly driven by data characteristics. This allows you to control the noise insertion and avoid spectral gaps due to assigning bin 0 to the MDCT lines for low data rates.

Perceptual weighting in the transform domain is preferably applied in determining quantization distortion to give different weights to particular frequency components. The perceptual weights can be efficiently derived from the linear prediction parameters.

Another independent aspect of the invention relates to the general concept of using LPC and SCF (ScaleFactor) data coexistence. In a transform encoder, such as one employing a modified discrete cosine transform (MDCT), scaling factors may be used in quantization to control the quantization step size. In the prior art, these scale factors are determined from the original signal to determine the masking curve. Now it is proposed to determine the second set of scale factors using a perceptual filter or a psychoacoustic model that is calculated from the LPC data. This reduces the cost of transmitting/storing the scale factors by transmitting/storing only the difference between the actual applied scale factors and the LPC-determined scale factors, instead of transmitting/storing the actual scale factors. Thus, in an audio coding system containing speech coding elements such as LPC for example, and transform coding elements such as MDCT, the present invention reduces the cost of transmitting the scale factor information needed for the transform part of the codec coding using the data provided by the LPC. It should be noted that this aspect is independent of other aspects of the proposed audio coding system and can also be implemented in another audio coding system.

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

In the case where LPC and MDCT do not operate at the same frame rate, i.e., have different frame sizes, linear prediction based scaling factors for signal frames in the transform domain can be determined based on the interpolated linear prediction parameters. to match the time window spanned by the MDCT frame.

The present invention therefore provides an audio coding system based on a transform encoder and contains fundamental prediction and generation modules from a speech encoder. The inventive system comprises a linear predictor for filtering an input signal based on an adaptive filter; a transform unit for converting a frame of the filtered input signal into a transform region; a quantizer for quantizing the signal in the transform domain; a scale factor determination unit for generating scale factors based on the masking threshold curve for use in a quantization unit when quantizing a signal in a transform region; a linear prediction scaling factor determination unit for determining a linear prediction based on the scaling factors based on the adaptive filter parameters; and a scale factor encoder for encoding a difference of scale factors based on the masking threshold curve and 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 the available linear prediction information, encoding and storage efficiency can be improved and very few bits need to be stored/transmitted.

Another independent encoder aspect of the invention relates to bit accumulator processing for variable size frames. In an audio coding system that can encode variable length frames, the bit accumulator is controlled by allocating the available bits between frames. Given a reasonable degree of complexity of the individual frames and a certain size bit bin, a certain deviation from the required constant bit rate allows for better overall quality without violating the buffer requirements imposed by the bit bin size. The present invention extends the concept of using a bit-accumulator to manage a bit-accumulator for a generic audio codec with variable frame sizes. The audio coding system may therefore comprise a bit accumulator control unit for determining the number of bits provided to encode a filtered signal frame based on the frame duration and frame complexity measure. Preferably, the bit accumulator control unit has different control equations for different frame complexity measures and/or different frame sizes. Complexity measures for different frame sizes can be normalized so that they can be compared more easily. In order to control the bit allocation for the variable rate encoder, the bit accumulator manager preferably sets the provided bit control algorithm to a lower allowable limit relative to the average number of bits for the largest allowable frame size.

A further aspect of the invention relates to bit accumulator processing in an encoder using a model-based quantizer, such as an entropy-limited quantizer (ECQ). It is proposed to minimize the change in the ECQ step size. A special control equation is proposed that relates the step size of the quantizer to the speed of ECQ.

The adaptive filter for filtering the input signal is preferably based on a Linear Prediction Coding (LPC) analysis comprising an LPC filter producing a bleached input signal. The LPC parameters for the current frame of input data can be determined using algorithms known in the art. The LPC parameter determiner can calculate any suitable representation of the LPC parameters for the input data frame, such as polynomials, transfer functions, reflection coefficients, discrete spectral frequencies, and so on. The specific type of LPC parameter representation that is used for encoding or other processing depends on the respective requirements. As will be known to those skilled in the art, some representations are more suitable for certain operations than others and are therefore preferred for performing those operations. The linear prediction block may act on the first frame duration, which is set to 20 ms, for example. Linear prediction filtering may further operate on a non-linear frequency axis to selectively emphasize certain frequency ranges, such as low frequencies, over other frequencies.

The transform applied to the frame of the filtered input signal is preferably a modified discrete cosine transform (MDCT) operating on a variable second frame duration. The audio coding system may include a window sequencer that determines, for an input signal block, the frame durations to overlap the MDCT windows, minimizing the coding cost function, preferably the simplified perceptual entropy, for the entire input signal block containing multiple frames. Thus, an optimal segmentation of the input signal block into MDCT windows having corresponding second frame durations is obtained. As a consequence, a transform-domain coding structure containing elements of a speech encoder is proposed, with an adaptive duration MDCT frame as the sole base block for all processing except LPC. Because MDCT frame durations can take on many different values, an optimal sequence can be found and abrupt changes in frame size can be avoided, as is typically the case in the prior art where only a small window size and a large window size are used. In addition, there is no need for the hard-edged transient transducer windows used in some prior art approaches to transition between small and large window sizes.

Preferably, the durations of successive MDCT windows vary by at most a factor of two (2) and/or the durations of the MDCT windows are dyadic values. More specifically, the durations of the MDCT windows may be dyadic portions of an input signal block. The MDCT window sequence is therefore limited to given sequences that are easy to encode with a small number of bits. In addition, the window sequence has smooth frame size transitions, thereby avoiding abrupt frame size changes.

The window sequencer may be further configured to provide long term prediction determination results generated by the long term predictor for window length candidates when searching for a sequence of MDCT window lengths that minimizes the input signal block coding cost function. In this embodiment, the long-term prediction loop is closed in determining the lengths of the MDCT windows, resulting in an improved sequence of MDCT windows used for coding.

The audio coding system may further comprise an LPC encoder for recursively variable rate encoding of discrete spectral frequencies or other appropriate representations of the LPC parameters generated by the linear predictor for storage and/or transmission to the decoder. According to an embodiment, a linear prediction interpolator is provided to interpolate the linear prediction parameters generated at a rate corresponding to the first frame duration so as to match the variable frame durations of the signal in the transform domain.

In accordance with an aspect of the invention, the audio coding system may comprise a perceptual modeler modifying the adaptive filter response by chirping and/or tilting the LPC polynomial generated by the linear predictor for an LPC frame. The perceptual model obtained by modifying the characteristics of the adaptive filter can be used for many purposes in the system. For example, it can be used as a perceptual weighting function in quantization or long-term prediction.

Another aspect of the invention relates to Long Term Prediction (LTP), in particular MDCT domain long term prediction, adapted MDCT frame LTP, and MDCT weighted LTP search. These aspects apply whether or not LPC analysis is present in the upstream of the transform encoder.

According to an embodiment, the audio coding system further comprises an inverse quantization and inverse transform unit for creating a time domain reconstruction of a frame of the filtered input signal. Additionally, a long-term prediction buffer may be provided to store time-domain reconstructions of previous frames of the filtered input signal. These blocks can be organized in a feedback loop from the quantizer to the long-term prediction extractor, which searches the long-term prediction buffer for the reconstructed segment that best matches the current frame of the filtered input signal. In addition, a long-term prediction gain determining unit may be provided that adjusts the gain of a segment selected from the long-term prediction buffer so that it best matches the current frame. Preferably, the long-term prediction determination result is subtracted from the transformed input signal in the transform region. Therefore, a second transform unit may be provided to transform the selected segment into a transform region. The long-term prediction loop may further comprise adding a result of determining the long-term prediction in the transform domain to the feedback signal after inverse quantization and before inverse time domain transformation. Thus, an inverse adaptive long-term prediction scheme can be used that predicts in the transform domain the current frame of the filtered input signal based on previous frames. For greater efficiency, the long-term prediction scheme can be further adapted in various ways, as set out below for some examples.

According to an embodiment, the long-term prediction block comprises a long-term prediction extractor for determining a delay value indicating a reconstructed segment of the filtered signal that best matches the current frame of the filtered signal. The long term prediction gain determiner may determine 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 so as to minimize the distortion criterion related to the difference in the perceptual domain between the long-term prediction estimate and the transformed input signal. The modified linear prediction polynomial can be applied as an equalization gain curve in the MDCT region while minimizing the distortion criterion.

The long-term prediction block may comprise a transform block for transforming the reconstructed segments from the LTP buffer into a transform region. To effectively implement the MDCT transform, such a transform should preferably be a type-IV discrete cosine transform.

Another aspect of the invention relates to an audio decoder for decoding a bitstream created with embodiments of the above encoder. The decoder according to the embodiment comprises a dequantizer for dequantizing a frame of an input bitstream based on scaling factors; an inverse transform block for inverse transforming the signal in the transform region; a linear prediction block for filtering the inversely transformed signal in the transform region; and a scalefactor decoding unit for generating scalefactors used in dequantization based on the obtained scalefactor delta information that encodes a difference between the scalefactors applied in the encoder and the scalefactors generated based on the adaptive filter parameters. The decoder may further comprise a scaling factor determiner for generating scaling factors based on a threshold masking curve obtained from the linear prediction parameters for the current frame. The scalefactor decoder may combine the obtained scalefactor delta information with the generated linear prediction based on the scalefactors to generate scalefactors for input to the dequantizer.

A decoder according to another embodiment comprises a model-based dequantizer for dequantizing a frame of an input bitstream; an inverse transform block for inversely transforming the signal in the transform region; and a linear prediction block for filtering the inversely transformed signal in the transform region. The dequantizer may comprise a model-based dequantizer and a non-model-based dequantizer.

Preferably, the dequantizer contains at least one adaptive probability model. The dequantizer may be configured to adapt the dequantization as a function of the characteristics of the transmitted signal.

The dequantizer may further decide on a dequantization strategy based on the control data for the decoded frame. Preferably, the dequantization control data is received along with the bitstream or obtained from the received data. For example, the dequantizer decides on a dequantization strategy based on the frame transform size.

According to another aspect, the dequantization block contains adaptive reconstruction points. The dequantization unit may comprise homogeneous scalar dequantizers configured to use two dequantization reconstruction points on the quantization interval, in particular a midpoint and an MMSE reconstruction point.

According to an embodiment, the dequantizer uses a model-based quantizer in combination with arithmetic coding.

In addition, the decoder may include many of the aspects disclosed above for the encoder. In general, the decoder will mirror the operations of the encoder, although some operations are performed only in the encoder and will not have any corresponding components in the decoder. Thus, what is described for the encoder should be considered applicable to the decoder as well, unless otherwise stated.

The above aspects of the invention may be implemented as an apparatus, a collection of apparatuses, a method, or a computer program running on a programmable device. Aspects of the invention may further be implemented in signals, data structures, and bit streams.

Thus, the application further discloses an audio coding method and an audio decoding method. An example of an audio coding method comprises the steps of: filtering an input signal based on an adaptive filter; converting a frame of the filtered input signal into a transform region; quantizing the signal in the transform region; creating scaling factors based on the threshold masking curve for use in the quantizer when quantizing the signal in the transform domain; determine based on linear prediction scaling factors, using the parameters of the adaptive filter; and encoding a difference between the scaling factors based on the masking threshold curve and the scaling factors based on the linear prediction.

Another audio coding method comprises the steps of: filtering an input signal based on an adaptive filter; converting a frame of the filtered input signal into a transform region; and quantize the signal in the transform region; in which the quantizer, based on the characteristics of the input signal, decides to encode the signal in the transform domain with a model-based quantizer or a non-model-based quantizer.

An example of an audio decoding method comprises the steps of: dequantizing a frame of an input bitstream based on scaling factors; inversely converting the signal in the conversion area; filtering with linear prediction of the inversely transformed signal in the area of the transformation; determining second scaling factors based on the adaptive filter parameters; and generating scale factors used in dequantization based on the obtained scale factor difference information and the determined second scale factors.

Another audio coding method comprises the steps of: dequantizing a frame of the input bitstream; inversely converting the signal in the conversion area; and linear prediction filtering the inverse transformed signal in the transform region; wherein the dequantization uses a model-based quantizer and a non-model-based quantizer.

The foregoing are merely examples of the preferred audio encoding/decoding methods and computer programs that are provided by the present application and that a person skilled in the art can derive from the following description of exemplary embodiments.

Brief description of the drawings

The present invention will now be described by means of illustrative examples, without limiting the scope or essence of the invention, with reference to the accompanying drawings, in which:

Fig. 1 shows a preferred embodiment of an encoder and decoder in accordance with the present invention;

Fig. 2 is a more detailed view of an encoder and decoder in accordance with the present invention;

Fig. 3 shows another embodiment of an encoder according to the present invention;

Fig. 4 is a preferred embodiment of an encoder according to the present invention;

Fig. 5 is a preferred embodiment of a decoder according to the present invention;

Fig. 6 shows a preferred embodiment for encoding and decoding MDCT lines in accordance with the present invention;

Fig. 7 shows a preferred embodiment of an encoder and a decoder and examples of corresponding control data transmitted from one to the other in accordance with the present invention;

Fig. 7a is another example of aspects of an encoder in accordance with an embodiment of the invention;

Fig. 8 is an example of a window sequence and relationship between LPC data and MDCT data according to an embodiment of the present invention;

Fig. 9 is a combination of scale factor data and LPC data in accordance with the present invention;

Fig. 9a shows another embodiment of combining scalefactor data and LPC data in accordance with the present invention;

Fig. 9b is another simplified block diagram of an encoder and decoder in accordance with the present invention;

Fig. 10 is a preferred embodiment of translating LPC polynomials into an MDCT gain curve in accordance with the present invention;

Fig. 11 illustrates a preferred embodiment of mapping constant refresh rate LPC parameters to adaptive MDCT window sequence data in accordance with the present invention;

Fig. 12 is a preferred embodiment of perceptual weighting filter adaptation calculation based on transform size and quantizer type in accordance with the present invention;

Fig. 13 is a preferred embodiment of the adaptation of the quantizer depending on the frame size in accordance with the present invention;

Fig. 14 is a preferred embodiment of the adaptation of the quantizer depending on the frame size in accordance with the present invention;

Fig. 15 is a preferred embodiment of adapting the quantization step size as a function of LPC and LTP data in accordance with the present invention;

Fig. 15a - derivation of the delta curve from the LPC and LTP parameters using the delta adaptation block;

Fig. 16 is a preferred embodiment of a model based quantizer using random shifts according to the present invention;

Fig. 17 is a preferred embodiment of a model-based quantizer in accordance with the present invention;

Fig. 17a is another preferred embodiment of a model based quantizer according to the present invention;

Fig. 17b is a schematic representation of a model based decoder 2150 for MDCT lines in accordance with an embodiment of the invention;

Fig. 17c is a schematic representation of pre-processing aspects of a quantizer according to an embodiment of the invention;

Fig. 17d is a schematic representation of aspects of step size calculation in accordance with an embodiment of the invention;

Fig. 17e is a schematic representation of a model based entropy constrained encoder according to an embodiment of the invention;

Fig. 17f is a schematic representation of the operation of a uniform scalar quantizer (USQ) in accordance with an embodiment of the invention;

Fig. 17g is a schematic representation of probability calculations in accordance with an embodiment of the invention;

Fig. 17h is a schematic representation of a dequantization process in accordance with an embodiment of the invention;

Fig. 18 is a preferred embodiment of bit storage management in accordance with the present invention;

Fig. 18a is the basic concept of bit storage management;

Fig. 18b is a bit-accumulator control concept for variable frame sizes in accordance with the present invention;

Fig. 18c is an example of a control curve for driving a bit accumulator according to an embodiment;

Fig. 19 is a preferred embodiment of an inverse quantizer using various reconstruction points in accordance with the present invention.

Description of Preferred Embodiments

The embodiments described below are merely illustrative examples of the principles of the present invention for an audio encoder and decoder. It is intended that modifications and changes to the circuits and details described herein will be apparent to others skilled in the art. The intention, therefore, is to be limited only by the scope of the claims of the accompanying claims, and not by the specific details provided through the description and explanation of the embodiments presented here. Like components of embodiments are denoted by like reference numerals.

In FIG. 1 is a visual representation of an encoder 101 and a decoder 102. The encoder 101 receives a time domain input and produces a bitstream 103 which is subsequently sent to the decoder 102. The decoder 102 produces an output waveform based on the received bitstream 103. The output is psychoacoustically similar to the original input. signal. In FIG. 2 shows a preferred embodiment of an encoder 200 and a decoder 210. The input signal of the encoder 200 passes through an LPC (Linear Prediction Coding) module 201 producing a whitened residual signal for an LPC frame having a first frame duration and corresponding linear prediction parameters. Additionally, gain normalization may be included in the LPC module 201. The residual signal from the LPC is converted to the frequency domain by an MDCT (Modified Discrete Cosine Transform) module 202 operating on a second variable frame duration. In the encoder 200 shown in FIG. 2, an LTP (Long Term Prediction) module 205 is included. LTP will be disclosed in detail in a further embodiment of the present invention. The MDCT lines are subjected to a quantization process 203 as well as a dequantization process 204 to provide the LTP buffer with a copy of the decoded output signal when it is available to the decoder 210. Due to quantization distortion, this copy is referred to as a reconstruction of the corresponding input signal. At the bottom of Fig. 2 shows a decoder 210. Decoder 210 receives the quantized MDCT lines, performs a dequantization process 211, adds a contribution from the LTP module 214, and performs an inverse MDCT transform process 212 followed by synthesis by an LPC filter 213.

An important aspect of the embodiment described above is that the MDCT frame is the only basic unit for encoding, although the LPC has its own (and in one embodiment, constant) frame size and the LPC parameters are also encoded. The embodiment starts with a transform encoder and introduces the underlying prediction and shaping modules from a speech encoder. As will be discussed later, the MDCT frame size is variable and adapts to the block of the input signal, determining the optimal sequence of MDCT windows for the entire block by minimizing the simplified perceptual entropy cost function. This allows scaling to support optimal time/frequency control. Additionally, the proposed unified framework avoids switching or layering combinations of different coding paradigms.

In FIG. The 3 parts of the encoder 300 are described schematically in more detail. The bleached signal as an output signal of the LPC encoder unit 201 shown in FIG. 2 is the input to the MDCT filter block 302. The MDCT analysis may alternatively be a non-linear time scale MDCT analysis, which ensures that the signal step (if the signal is periodic with a well-defined step) is constant across the MDCT transform window.

In FIG. 3, the LTP module 310 is presented in more detail. It contains an LTP buffer 311 storing the reconstructed samples in the time domain of the previous segments of the output signal. LTP extractor 312 finds the segment with the best match in LTP buffer 311 for a given current input segment. Gain block 313 applies a suitable gain value to this segment before it is subtracted from the segment currently being input to quantizer 303 . Obviously, to perform the subtraction before quantization, the LTP extractor 312 also converts the selected signal segment into an MDCT region. The LTP extractor 312 searches for the best gain and delay values that minimize the perceptual domain error function when combining the reconstructed previous output segment with the transformed MDCT domain input frame. For example, the mean square error (MSE) function between the transformed reconstructed segment of LTP module 310 and the transformed input frame (ie, the residual signal after subtraction) is optimized. This optimization can be performed in the perceptual domain, where frequency components (ie, MDCT lines) are weighted according to their perceptual importance. LTP module 310 operates in blocks of MDCT frames, and encoder 300 reads one remainder of an MDCT frame at a time, for example, for quantization in quantizer 303 . Delay and gain searches may be performed in the perceptual domain. Alternatively, LTP may be frequency selective, that is, adapt the gain and/or delay depending on the frequency. Shown block 304 inverse quantization 304 and block 306 inverse MDCT. The MDCT may have a non-linear time scale, as explained later.

In FIG. 4 shows another embodiment of encoder 400. In addition to FIG. 3, LPC analysis 401 is introduced for clarity. Shown is a DCT-IV transform 414 used to transform a selected signal segment into an MDCT region. Additionally, several methods for calculating the minimum error for LTP segment selection are shown. In addition to minimizing the residual signal, as shown in FIG. 4 (identified as LTP2 in FIG. 4) shows the minimization of the difference between the transformed input signal and the dequantized MDCT domain signal before inverse transforming into a reconstructed time domain signal for storage in LTP buffer 411 (labeled LTP3). Minimizing this MSE function will direct the LTP contribution to the optimal (as far as possible) similarity of the transformed input signal and the reconstructed input signal for storage in the LTP buffer 411. Another alternative error function (denoted as LTPl) is based on the difference of these signals in the time domain. In this case, the MSE between the filtered LPC input frame and the corresponding time domain reconstruction in the LTP buffer 411 is minimized. The MSE is preferably calculated based on the MDCT frame size, which may be different from the LPC frame size. Additionally, the quantizer and dequantizers are replaced by a spectral encoder 403 and spectral decoders 404 ("Spec enc" and "Spec dec"), which may contain additional modules in addition to the quantizers as outlined in FIG. 6. Again, MDCT and inverse MDCT may have a non-linear time scale (WMDCT, IWMDCT).

In FIG. 5 shows a proposed decoder 500. The spectrum data from the received bit stream is inversely quantized 511 and added to the LTP contribution provided by the LTP extractor from the LTP buffer 515. Also shown is an LTP extractor 516 and an LTP gain block 517 in the decoder 500. The summed MDCT lines are synthesized in the time domain by the MDCT synthesis block, and the time domain signal is spectrally shaped by the LPC synthesis filter 513.

In FIG. 6, "Spec dec" and "Spec enc" blocks 403, 404 shown in FIG. 4 are described in more detail. The "Spec enc" block 603 shown on the right side of the drawing contains in an embodiment a harmonic prediction analysis module 610, a TNS (temporal noise limiting) analysis module 611 followed by a scale factor scaling module 612 for MDCT lines, and finally quantization and coding lines in module 613 lines Enc. The decoder "Spec Dec" block 604 shown on the left side of the figure performs the inverse process, that is, the received MDCT lines are dequantized in the Dec line unit 620 and the scaling is canceled by the scale factor scaling (SCF) unit 621 . TNS synthesis 622 and harmonic prediction synthesis 623 are applied.

In FIG. 7 shows a very general representation of the coding system according to the invention. An encoder, as an example, takes an input signal and creates a bitstream containing, among other data:

- quantized MDCT lines;

- scale factors;

- polynomial representation of LPC;

- energy of the signal segment (for example, signal dispersion);

- sequence of windows;

- LTP data.

The decoder according to the embodiment reads the provided bitstream and produces an audio output signal that is psychoacoustically similar to the original signal.

In FIG. 7a shows other aspects of an encoder 700 according to an embodiment of the invention. The encoder 700 includes an LPC module 701, an MDCT module 704, an LTP module 705 (shown only simplified), a quantization module 703, and an inverse quantization module 704 for returning the reconstructed signals back to the LTP module 705. Additionally, a step determination module 750 is provided for determining the pitch of the input signal, and a window sequence determination module 751 is provided for determining the optimal MDCT window sequence for a block of the input signal of a larger size (eg, 1 second). In this embodiment, the MDCT window sequence is determined based on an open-loop approach in which a sequence of MDCT window size candidates is determined that minimizes an encoding cost function, such as simplified perceptual entropy. The contribution of the LTP module 705 to the coding cost function that is minimized by the window sequencer 751 may alternatively be taken into account when searching for the optimal MDCT window sequence. Preferably, for each determined window size candidate, the best long-term prediction contribution to the MDCT frame corresponding to the window size candidate is determined and the corresponding coding cost is determined. In general, short MDCT frame sizes are more suitable for speech input, while long transform windows having excellent spectral resolution are preferred for audio signals.

The perceptual weights or perceptual weighting function are determined based on the LPC parameters calculated by the LPC module 701, which will be explained in more detail below. Perceptual weights are provided to LTP module 705 and quantization module 703 703, both operating in the MDCT domain, to weight errors or distortion contributions of frequency components according to their respective perceptual importance. In FIG. 7a further shows which coding parameters are passed to the decoder, preferably by the respective coding scheme, as will be discussed later.

Next, the coexistence of LPC and MDCT data and the emulation of the effect of LPC in MDCT will be discussed, both to counteract and bypass actual filtering.

According to an embodiment, 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 spectrally flat. This is preferred, for example, for LTP operation. However, other parts of the codec that operate on a spectrally flat signal may benefit from knowing which spectral shape of the original signal preceded the LP filtering. Because the post-filtered encoder modules operate on the MDCT transform of the spectrally flat signal, the present invention indicates that the spectral shape of the original signal prior to LP filtering can, if desired, be re-mapped onto the MDCT representation of the spectrally flat signal, reflecting the transfer function of the LP filter being used ( i.e., the spectral envelope of the original signal) on the gain curve or equalization curve that is applied to the frequency bins of the MDCT representation of the spectrally flat signal. In contrast, the LP module can eliminate actual filtering and only determine a transfer function that is subsequently displayed on a gain curve that can be superimposed on the MDCT representation of the signal, thereby eliminating the need for filtering in the time domain of the input signal.

One clear aspect of embodiments of the present invention is that the MDCT-based transform encoder operates using flexible window segmentation on a bleached LPC signal. This is shown in FIG. 8 for an example of a sequence of MDCT windows along with working with LPC windows. Therefore, as is clear from the drawing, LPC operates with a constant frame size (eg 20 ms), while MDCT operates with a variable sequence of windows (eg 4-128 ms). This allows you to independently select the optimal window length for LPC and the optimal window sequence for MDCT.

Fig. 8 further shows the relationship between LPC data, in particular LPC parameters, created at a first frame rate, and MDCT data, in particular MDCT lines, created at a second variable rate. The downward arrows in the drawing symbolize LPC data that is interpolated between LPC frames (circles) to match the corresponding MDCT frames. For example, a perceptual weighting function created using LPC is interpolated for temporal cases as determined by the MDCT window sequence. The upward arrows symbolize refinement data (ie, control data) used to encode the MDCT lines. For AAC frames, the data is typically scaling factors, and for ECQ frames, the data is typically dispersion correction data, and so on. The solid lines relative to the dotted lines represent which data is the "most important" data for coding MDCT lines for a particular quantizer. Double downward arrows symbolize the spectral lines of the codec.

The coexistence of LPC and MDCT data in the encoder can be used, for example, to reduce the need for bits when encoding MDCT scaling factors, taking into account the perceptual masking curve determined from the LPC parameters. Additionally, the perceptual weighting derived from the LPC can be used in determining quantization distortion. As shown in the drawing and as will be discussed below, the quantizer operates in two modes and creates two types of frames (ECQ frames and AAC frames) depending on the frame size of the received data, i.e. corresponding to the size of the MDCT frame or window.

In FIG. 11 shows a preferred embodiment of mapping constant rate LPC parameters to adaptive MDCT window sequence data. The LPC mapping module 1100 receives LPC parameters according to the LPC update rate. In addition, the LPC mapping unit 1100 receives MDCT window sequence information. It then creates an LPC-to-MDCT mapping, such as a mapping of LPC-based psychoacoustic data to corresponding MDCT frames created at a variable frame rate MDCT. For example, the LPC mapper interpolates LPC polynomials or related data for temporal cases corresponding to MDCT frames for use, for example, as perceptual weights in an LTP module or a quantizer.

Now, the specifics of the LPC based perceptual model are discussed with reference to FIG. 9. The LPC module 901 is in an embodiment of the present invention capable of producing a white output signal using linear prediction, for example, order 16 for a signal with a sampling rate of 16 kHz. For example, the output of the LPC module 201 in FIG. 2 is residual after LPC parameters have been determined and filtered. The defined LPC polynomial A(z), as schematically shown at the bottom left of FIG. 9 may be chirped with a bandwidth spreading factor and also tilted, in one embodiment of the invention, by changing the first reflectance of the corresponding LPC polynomial. chirp broadens the bandwidth of the peaks in the LPC transfer function by moving the poles of the polynomial inside the unit circle, thus resulting in smoother peaks. Tilt allows the LPC transfer function to be flatter to balance out the effects of low and high frequencies. These modifications seek to create a perceptual masking curve A'(z) from certain LPC parameters that will be available on both the encoder and decoder sides of the system. Details of LPC polynomial manipulation are shown below in FIG. 12.

The MDCT coding applied to the LPC residual has, in one embodiment of the invention, scaling factors to control the quantizer resolution or quantization step sizes (and thus the noise introduced by the quantization). These scale factors are determined by the scale factor determination module 960 for the original input signal. For example, the scale factors are derived from a perceptual masking threshold curve determined from the original signal. In an embodiment, a separate frequency transform (having possibly a different frequency resolution) may be used to determine the threshold masking curve, but this is not always necessary. Alternatively, the masking threshold curve is determined from the MDCT lines generated by the transform module. In the lower right part in Fig. 9 schematically shows the scaling factors generated by the scaling factor determination module 960 for quantization control so that the introduced quantization noise is limited to inaudible distortion.

If the LPC filter is connected before the MDCT transform module, the bleached signal is converted to the MDCT region. Since this signal has a white spectrum, it is not well suited for deriving a perceptual masking curve from it. Thus, the MDCT gain equalization curve created to compensate for spectrum whitening can be used in determining the masking threshold curve and/or scaling factors. For this reason, scaling factors must be determined for a signal that has the properties of the absolute spectrum of the original signal in order to correctly determine perceptual masking. The calculation of the gain equalization curve for the MDCT region from the LPC polynomial is discussed in more detail below with reference to FIG. 10.

(n) вводится в блок 902 преобразования MDCT, тогда как полином А(z) вводится в блок 970 вычисления кривой усиления MDCT 970 (как показано на фиг. 14). Кривая усиления, определенная из полинома LP, применяется к коэффициентам MDCT или линиям, чтобы сохранить спектральную огибающую первоначального входного сигнала до того, как определять масштабные коэффициенты. Линии MDCT с отрегулированным усилением вводятся в модуль 960 определения масштабных коэффициентов, который определяет масштабные коэффициенты для входного сигнала. An embodiment of the above outlined scaling factor determination scheme is shown in FIG. 9a. In this embodiment, the input signal is input to the LP module 901, which determines the envelope of the spectrum of the input signal described by A(z) and outputs said polynomial as well as a filtered version of the input signal. The input signal is filtered using the A(z) inversion to provide a spectrally white signal that is subsequently used by other parts of the encoder. Filtered signal

(n) is input to the MDCT transform block 902, while the polynomial A(z) is input to the MDCT gain curve calculator 970 970 (as shown in FIG. 14). The gain curve determined from the LP polynomial is applied to the MDCT coefficients or lines to preserve the spectral envelope of the original input signal before scaling factors are determined. The gain-adjusted MDCT lines are input to a scaling factor determination module 960, which determines the scaling factors for the input signal.

Using the approach outlined above, the data transmitted between the encoder and decoder contains both the LP polynomial from which the corresponding perceptual information can be obtained, and the signal model that can be obtained when a model-based quantizer is used, and scaling factors commonly used in a transform codec.

In more detail, returning to FIG. 9, the LPC block 901 shown in the drawing determines from the input signal the spectral envelope A(z) of the signal and obtains the perceptual representation A'(z) from it. In addition, the scaling factors that are commonly used in transform-based perceptual audio codecs are determined from the input signal, or can be determined on the white signal produced by the LP filter if the LP filter's transfer function is taken into account when determining the scaling factors (as described below in the context of FIG. 10). The scalefactors may then be adapted in scalefactor adaptation module 961 for a given LP, as will be described below, to reduce the data rate required to transmit the scalefactors.

Typically, the scale factors are sent to the decoder and thus the LP polynomial appears. Now, assuming both are determined from the original input and that both are correlated to some extent with the absolute spectrum properties of the original input, it is suggested to encode a delta representation between them to remove any redundancy, which may arise if both are transmitted separately. According to an embodiment, this correlation is used as follows. Because the LPC polynomial, when properly chirped and tilted, tends to represent the masking threshold curve, the two representations can be combined such that the transmitted transform encoder scale factors represent the difference between the desired scale factors and those that can be obtained from the transmitted LPC polynomial. The scaling factor adaptation module 961 shown in FIG. 9 therefore calculates the difference between the desired scale factors generated from the original input signal and the scale factors obtained from the LPC. This aspect retains the ability to have an MDCT-based quantizer having a scalefactor representation that is commonly used in transform encoders within an LPC structure operating on the LPC residual, and still has the ability to switch to a model-based model-based quantizer that receives quantization step sizes solely from linear prediction data.

In FIG. 9b is a simplified block diagram of an encoder and decoder according to the embodiment. The input signal at the encoder is passed through an LPC module 901 producing a whitened residual signal and corresponding linear prediction parameters. Additionally, the LPC module 901 may contain gain normalization. The residual signal from the LPC is converted to the frequency domain using the transform 902 MDCT. On the right side of Fig. 9b shows a decoder. The decoder receives the quantized MDCT lines, dequantizes 911 them, and applies an inverse MDCT transform 912 followed by LPC synthesis with a filter 913.

The bleached signal as an output signal of the LPC module 901 in the encoder in FIG. 9b is input to the MDCT filter block 902. The MDCT lines as a result of the MDCT analysis are transform encoded with a transform coding algorithm consisting of a perceptual model that controls the desired quantization step size for different portions of the MDCT spectrum. The values that determine the quantization step size are called scale factors, and there is one scale factor value needed for each element of the MDCT spectrum, called the scale factor band. In the prior art of transform coding algorithms, the scale factors are transmitted via the bit stream to the encoder.

In accordance with one aspect of the invention, the perceptual masking curve determined from the LPC parameters, as explained with reference to FIG. 9 is applied in coding the scale factors used in quantization. Another possibility for determining the perceptual masking curve is to use the unmodified LPC filter coefficients to determine the energy distribution across the MDCT lines. With this energy estimate, the psychoacoustic model used in transform coding schemes can be applied at both the encoder and decoder to determine the masking curve.

The two masking curve representations are then combined such that the scale factors to be transmitted by the transform encoder represent the difference between the desired scale factors and the scale factors that can be derived from the transmitted LPC polynomial or be based on the LPC psychoacoustic model. This feature retains the ability to have an MDCT based quantizer having a scalefactor representation commonly used in transform encoders within an LPC structure operating on an LPC residual and still be able to control the quantization noise based on the scalefactor band according to with a psychoacoustic model of a transform encoder. The advantage is that the transfer of the scalefactor difference will cost fewer bits compared to the transmission of the absolute scalefactor values without taking into account already existing LPC data. Depending on the data rate, frame size, or other parameters, the amount of scalefactor residual to be transmitted can be selected. In order to have complete control over the bandwidth of each scale factor, the delta scale factor parameter may be transmitted using an appropriate noiseless coding scheme. In other cases, the cost of transferring the scalefactors can be further reduced by coarser representation of the scalefactor differences. The special case with the lowest overhead is when the scale factor difference is set to 0 for all bands and no additional information is transmitted.

In FIG. 10 shows a preferred embodiment of translating LPC polynomials into an MDCT gain curve. As shown in FIG. 2, the MDCT operates on a bleached signal, where the bleaching is performed by the LPC filter 1001. In order to preserve the spectrum envelope of the original input signal, the MDCT gain curve is computed by the MDCT gain curve module 1070. The gain equalization curve in the MDCT region can be obtained by determining the magnitude of the response of the spectral envelope described by the LPC filter for the frequencies represented by the elements in the MDCT transform. The gain curve may then be applied to the MDCT data, for example, in computing the minimum mean square error signal as shown in FIG. 3 or when determining the perceptual masking curve to determine the scaling factors as shown above with reference to FIG. 9.

In FIG. 12 shows a preferred embodiment of adapting the calculation for a perceptual weighting filter based on the size of the transform and/or the type of quantizer. The LP polynomial A(z) is determined by the LPC module 1201 shown in FIG. 16. The LPC parameter modification module 1271 receives the LPC parameters such as the LPC polynomial A(z) and creates a perceptual weighting filter A'(z) by modifying the LPC parameters. For example, the bandwidth of the LPC polynomial A(z) widens and/or the polynomial slopes. The input parameters for the chirp and slope adaptation module 1272 are the default chirp and slope values, ρ and γ. They change according to predetermined, predetermined rules based on the transform size used and/or the quantization strategy Q used. The modified chirp and slope parameters ρ' and γ' are input to LPC parameter modification module 1271 converting the input signal spectrum envelope represented by A(z) into a perceptual masking curve represented by A'(z).

Next, the strategy of quantization based on frame size and model-based quantization based on various parameters according to an embodiment of the invention will be explained. One aspect of the present invention is that it uses different quantization strategies for different transform sizes or frame sizes. This is shown in FIG. 13, where the frame size is used as a selection parameter for using a model-based quantizer or a non-model-based quantizer. It should be noted that this aspect of quantization is independent of other aspects of the disclosed encoder/decoder and can be applied to other codecs as well. An example of a non-model based quantizer is the Huffman table quantizer used in the AAC audio coding standard. The model-based quantizer may be an entropy-constrained quantizer (ECQ) using arithmetic coding. However, other quantizers may also be used in embodiments of the present invention.

In accordance with an independent aspect of the present invention, it is proposed to switch between different quantization strategies as a function of frame size in order to be able to use the optimal quantization strategy given by a particular frame size. As an example, the sequence of windows may dictate the use of a long transform for a very constant tonal musical signal segment. For this particular type of signal using a long transform, it is very advantageous to use a quantization strategy that takes advantage of the "sparseness" symbol (ie, well-defined discrete tones) in the signal spectrum. The quantization method used in AAC, in combination with Huffman tables and spectral line grouping, also used in AAC, is very advantageous. However, on the other hand, for speech segments, the sequence of windows, given the coding gain given by LTP, may dictate the use of fast transforms. For this signal type and transform size, it is advantageous to apply a quantization strategy that does not attempt to find or introduce sparsity into the spectrum, but instead maintains a wideband energy that, given LTP, will retain the pulse-like symbol of the original input signal.

A more general visual representation of this concept is shown in FIG. 14 where the input signal is converted to the MDCT region and subsequently quantized by a quantizer controlled by the transform size or frame size used for the MDCT transform.

According to another aspect of the invention, the quantizer step size is adapted as a function of the LPC and/or LTP data. This allows you to define the step size depending on the complexity of the frame and control the number of bits allocated to encode the frame. In FIG. 15 shows an example of how model-based quantization can be driven by LPC and LTP data. At the top of Fig. 15 is a schematic visualization of MDCT lines. Shown below is the quantization step size, delta Δ, as a function of frequency. From this particular example, it is clear that the quantization step size increases with frequency, that is, for higher frequencies, more distortion is introduced during quantization. The delta curve is obtained from the LPC and LTP parameters by the delta adaptation module shown in FIG. 15a. The delta curve may further be obtained from the prediction polynomial A(z), by chirp and/or slope, as explained with reference to FIG. 13.

The preferred perceptual weighting function derived from the LPC data is given by the following equation:

,

,

where A(z) is the LPC polynomial, τ is the slope parameter, ρ controls the chirp, and r ₁ is the first reflection coefficient calculated from the A(z) polynomial. It should be noted that the polynomial A(z) can be recalculated to select different representations in order to extract the relevant information from the polynomial. If you are interested in the steepness of the spectrum, to apply "slope" counteracting the steepness of the spectrum, it is preferable to recalculate the polynomial in the reflectances, since the first reflectance represents the steepness of the spectrum.

In addition, the delta values Δ may be adapted as a function of the input signal variance Δ, the LTP gain g, and the first reflection coefficient r ₁ obtained from the prediction polynomial. For example, the adaptation may be based on the following equation:

Δ'= Δ(1+r ₁ (1-g ² ))

The following describes aspects of model-based quantizers according to an embodiment of the present invention. In FIG. 16 shows one aspect of a model-based quantizer. The MDCT lines are input to a quantizer using uniform scalar quantizers. In addition, random shifts are introduced into the quantizer, which are used as shift values for the quantization intervals, shifting the boundaries of the intervals. The proposed quantizer provides the benefits of vector quantization while retaining the search capability of scalar quantizers. The quantizer iterates over a number of different shift values and computes a quantization error for them. An offset value (or offset value vector) minimizing quantization distortion for particular MDCT quantized lines is used for quantization. The offset value is then transmitted to the decoder along with the quantized MDCT lines. The use of random shifts introduces noise stuffing into the dequantized decoded signal and thus avoids spectral gaps in the quantized spectrum. This is especially important at low data rates, where multiple MDCT lines are otherwise quantized to zero, which can lead to audible notches in the spectrum of the reconstructed signal.

In FIG. 17 schematically shows a model-based MDCT line quantizer (MBMLQ) according to an embodiment of the invention. At the top of Fig. 17 shows an MBMLQ encoder 1700. The MBMLQ encoder 1700 receives as input the MDCT lines in the MDCT frame or the MDCT lines of the residual LTP if LTP is present in the system. MBMLQ uses statistical MDCT line models and source codes are adapted to the signal properties on a frame-by-frame MDCT basis, resulting in efficient bitstream compression.

The local gain of the MDCT lines may be defined as the RMS value of the MDCT lines and the MDCT lines normalized in gain normalizer 1720 prior to input to the MBMLQ encoder 1700. Local gain normalizes the MDCT lines and is in addition to LP gain normalization. Since the LP gain adapts to signal level changes over a longer time scale, the local gain adapts to changes over a smaller time scale, resulting in improved quality of transients and speech initiations. The local gain is encoded at a fixed rate or a variable coding rate and transmitted to the decoder.

Rate control module 1710 may be used to control the number of bits used to encode an MDCT frame. The frequency control index controls the number of bits used. The frequency control index is specified in the list of nominal step sizes of the quantizer. The table can be sorted by step sizes in descending order (see Fig. 17g).

The MBMLQ encoder operates on a set of different rate control indexes, and the frame uses the rate control index resulting in a number of bits less than the number of bits provided by the bit accumulator control. The rate control index changes slowly and this can be used to reduce search complexity and efficiently code the index. The index set that is checked may be reduced if the check starts near the index of the previous MDCT frame. Likewise, an efficient entropy encoding of an index is obtained if the probabilities peak around the previous index value. For example, for a list of 32 step sizes, the frequency steering index may be encoded using an average of 2 bits per MDCT frame.

In FIG. 17 further schematically shows an MBMLQ decoder 1750 where the MDCT frame is gain renormalized if a local gain has been determined in the encoder 1700.

In FIG. 17a schematically shows a model-based MDCT line encoder 1700 according to the embodiment in more detail. It contains a quantizer preprocessor 1730 (see FIG. 17c), an entropy constrained model-based encoder 1740 (see FIG. 17e), and an arithmetic encoder 1720, which may be a prior art arithmetic encoder. The task of the quantizer preprocessor module 1730 is to adapt the MBMLQ encoder to signal statistics on a frame-by-frame basis MDCT. It takes other codec parameters as input and extracts useful statistics about the signal from them that can be used to modify the limited entropy model based encoder 1740. Entropy constrained model-based encoder 1740 is controlled, for example, by a set of control parameters: quantizer step size Δ, set V of MDCT line variance estimates (vector; one estimated value per MDCT line), perceptual masking curve, P _mod , matrix or a table of (random) shifts and a statistical model of MDCT strings that describe the shape of the distribution of MDCT strings and their interdependencies. All of the control parameters mentioned above may change between MDCT frames.

In FIG. 17b schematically shows a model-based MDCT line decoder 1750 according to an embodiment of the invention. It takes the side information bits from the bitstream as input and decodes them into parameters that are input to the quantizer front end module 1760 (see FIG. 17c). The quantizer front end module 1760 preferably has the same functionality in the encoder 1700 as in the decoder 1750. The parameters that are input into the quantizer front end 1760 are exactly the same in the encoder as in the decoder. Quantizer preprocessor module 1760 outputs a set of control parameters (same as in encoder 1700) and these are input to probability calculation module 1770 (see FIG. 17g; same as encoder, see FIG. 17e) and module 1780 dequantization (see Fig. 17h; same in the encoder, see Fig. 17e). The cdf tables from probability calculator 1770 representing the probability distribution densities for all MDCT lines given the delta parameters used for quantization and signal variance are input to an arithmetic decoder (which can be any arithmetic encoder known to those skilled in the art). ), which then decodes the MDCT line bits into MDCT line indices. The MDCT line indices are then dequantized into the MDCT line by dequantization module 1780.

In FIG. 17c schematically shows aspects of the pre-processing of a quantizer according to an embodiment of the invention, which consists of i) calculating the step size, ii) modifying the perceptual masking curve, iii) determining the dispersion of the MDCT lines, iv) constructing a shift table.

The step size calculation is explained in more detail in FIG. 17d. It contains i) a table lookup where the frequency control index indicates in the step size table those dimensions that produce the nominal Δ _nom , ii) low energy adaptation, and iii) high frequency adaptation.

Gain normalization typically results in high energy sounds and low energy sounds being encoded with the same segment signal to noise ratio (SNR). This can result in an excessive number of bits being used for low energy sounds. The proposed low energy adaptation allows fine tuning of the compromise between low energy and high energy sounds. The step size can be increased when the signal energy becomes low, as shown in FIG. 17d-ii), which gives an example of a relationship curve between signal energy (gain g) and control factor q _Le . The signal gain g can be calculated as the RMS value of the input signal itself or the residual LP. The control curve in FIG. 17d-ii) is just one example and other control functions can be used to increase the step size for low energy signals. In the example shown in the drawing, the control function is determined by stepwise linear sections, which are determined by the thresholds T ₁ and T ₂ , and the step size factor L.

High frequency sounds are perceptually less important than low frequency sounds. The high-frequency adaptation function increases the step size when the MDCT frame is high-frequency, that is, when the signal energy in the current MDCT frame is concentrated at higher frequencies, resulting in fewer bits spent on such frames. If LTP is present and if the LTP gain g _LTP is close to 1, the rest of the LTP may become high frequency; in such a case it is advantageous not to increase the step size. This mechanism is shown in Fig. 17d-iii), where r is the first reflectance obtained from LPC. The proposed high frequency adaptation can use the following equation:

In FIG. 17c-ii) schematically shows modification of the perceptual masking curve using low frequency (LF) gain boost to remove "thunder" encoding artifacts. The gain increase at low frequencies can be fixed or made adaptive so that only the part below the first spectral peak receives additional gain. The gain boost at low frequencies can be adapted using the LPC envelope data.

In FIG. 17c-iii) schematically shows the determination of the dispersion of MDCT lines. With an active LPC whitening filter, all MDCT lines have unit dispersion (corresponding to the LPC envelope). After perceptual weighting in a limited entropy model based encoder 1740 (see FIG. 17e), the MDCT lines have dispersions that are the inverse of a quadratic perceptual concealment curve or a quadratic modified concealment curve P _mod . If LTP is present, it can reduce the dispersion of MDCT lines. In FIG. 17c-iii) depicts a mechanism that tailors certain variances to LTP. The drawing shows the modification function q _LTP frequency f. Modified variances can be determined using V _LTPmod = V*q _LTP. The _LLTP value may be a function of the LTP gain, such that _LLTP is closer to 0 if the LTP gain is approximately 1 (indicating that LTP found a good match), and _LLTP is closer to 1 if the LTP gain is approximately 0. Proposed LTP adaptation for dispersions V = {v ₁ , v ₂ ..., v _j ..., v _N } only affects MDCT lines below a certain frequency (f _LTPcutoff ). As a result, the dispersions of the MDCT lines below the cutoff frequency f _LTPcutoff are reduced, the reduction depending on the LTP gain.

In FIG. 17c-iv) schematically shows the construction of the shift table. The nominal shift table is a matrix filled with pseudo-random numbers distributed between -0.5 and 0.5. The number of columns in the matrix is equal to the number of MDCT lines that are encoded with MBMLQ. The number of rows is adjusted to be equal to the number of shift vectors that are checked for RD optimization in the limited entropy model based encoder 1740 (see FIG. 17e). The shift table construction function scales the nominal shift table with the quantizer step size so that the shifts are distributed between -Δ/2 and +Δ/2.

In FIG. 17g schematically shows an embodiment of a shift table. The shift index is a pointer in the table and selects the desired shift vector O = {o ₁ , o ₂ ..., o _n , ..., o _N }, where N is the number of MDCT lines in the MDCT frame.

As described below, offsets provide a means for noise padding. Better objective and perceptual quality is obtained if shift spread is limited for MDCT lines having low dispersion v _j compared to the quantizer step size Δ. An example of such a restriction is described in Fig. 17c-iv), where k ₁ and k ₂ are adjustable parameters. The distribution of shifts can be uniform and be between -s and +s. The boundaries of s can be defined according to the expression:

For low dispersion MDCT lines (where v _j is small compared to Δ), it may be preferable to make the distribution of shifts non-uniform and signal dependent.

In FIG. 17e schematically shows a model based encoder 1740 with an entropy constraint 1740 in more detail. The MDCT input lines are perceptually weighted by dividing them by the values of a perceptual masking curve, preferably derived from the LPC polynomial, resulting in weighted MDCT line vectors y=(y ₁ . .., y _N ). The purpose of the subsequent encoding is to introduce white quantization noise into the MDCT lines in the perceptual domain. The decoder uses perceptual weighting inversion, which results in quantization noise corresponding to the perceptual masking curve.

First, iteration over random shifts is described. In the shift matrix, for each row j in the shift matrix, the following operations are performed: Each row of the MDCT is quantized with uniform scalar shifters (USQ) in which each quantizer is shifted by its own unique shift value taken from the shift row vector.

, где k₃ может выбираться, чтобы быть любым соответствующим числом, например 20. Границей перегрузки является граница, на которой ошибка квантования по величине больше, чем половина размера шага квантования.The minimum distortion interval probability from each USQ is calculated in probability calculation module 1770 (see FIG. 17g). The USQ indices are entropy coded. The cost in terms of the number of bits required to encode the indices is calculated according to FIG. 17e, resulting in the codeword length R _j . The USQ j congestion limit for the MDCT link can be calculated as

, where k ₃ can be chosen to be any appropriate number, such as 20. The overload boundary is the boundary at which the quantization error is greater in magnitude than half the quantization step size.

). d(y,

) может быть среднеквадратичной ошибкой (MSE) или другой перцепционно более подходящей мерой искажения, например, основанной на перцепционной функции взвешивания. В частности, может быть полезна мера искажения, которая взвешивает вместе MSE и рассогласование по энергии между y и y. The scalar reconstruction value for each MDCT line is computed by dequantizer 1780 (see FIG. 17h), resulting in a quantized MDCT vector y. In the RD optimization module 1790, the distortion D _j =d(y,

). d(y,

) may be the mean square error (MSE) or another perceptually more appropriate measure of distortion, such as one based on a perceptual weighting function. In particular, a distortion measure that weights together the MSE and the energy mismatch between y and y can be useful.

In an RD optimization module 1790, a cost C is calculated, preferably based on the distortion D _j and/or the theoretical codeword length R _j for each row j in the shift matrix. An example of a cost function is C=10*log ₁₀ (D _j )+λ*R _j /N. The offset that minimizes C is selected and the corresponding USQ and probability indices are derived from the entropy constrained model based encoder 1780.

The RD optimization can be further improved as desired by changing other properties of the quantizer along with the offset. For example, instead of using the same fixed variance estimate V for each shift vector that is tested in RD optimization, the variance estimate vector V can be variable. For the row vector m of the shifts, an estimate of the variance k _m *V can then be used, where k _m can cover, for example, the range from 0.5 to 1.5 as m changes from m=1 to m=(number of rows in shift matrix). This makes entropy encoding and MMSE calculation less sensitive to changes in the input signal statistics that the statistical model cannot capture. This results in a lower cost of C overall.

The dequantized MDCT lines can be further improved by using a residual quantizer as shown in FIG. 17e. The residual quantizer may be, for example, a fixed frequency random vector quantizer.

The operation of a uniform scalar quantizer (USQ) for quantizing an MDCT line n is shown schematically in FIG. 17f, which shows the value of n of the MDCT line, which is in the interval of minimum distortion, having the index i _n . The "x" marks indicate the center (midpoint) of the quantization intervals with step size Δ. The origin of the scalar quantization device is shifted by the amount of shift o _n from the shift vector О = { _о1 , о ₂ ,..., o _n ..., o _N }. Thus, the boundaries of the interval and the midpoints are shifted by the amount of the shift.

The use of shifts introduces encoder-controlled noise filling into the quantized signal, and this makes it possible to avoid gaps in the quantized spectrum. Additionally, shifts improve coding efficiency by providing a range of coding alternatives that fill space more efficiently than a cubic lattice. In addition, the shifts provide a change in the probability tables that are calculated by the probability calculation module 1770, leading to more efficient entropy coding of the MDCT line indices (ie, fewer bits are required).

Using a variable step size Δ (delta) allows for variable quantization precision, so that higher precision can be used for perceptually important sounds and less precision can be used for less important sounds.

In FIG. 17g schematically shows a probability calculation in a probability calculation module 1770. The inputs to this module are the statistical model applied to the MDCT lines, the step size Δ, the variance vector V, the shift index, and the shift table. The output of the probability calculation module 1770 are cdf tables. For each x _j MDCT line, a statistical model (ie, a probability density function, pdf) is determined. The area under the pdf function for the interval I is the probability p _ij of the interval. This probability is used for the arithmetic coding of the MDCT lines.

In FIG. 17h schematically shows how the dequantization process is performed, for example, in a dequantization module 1780. Center of mass (MMSE value) x _MMSE for the minimum distortion interval of each MDCT line is calculated along with the midpoint x _{of the MP} interval. Assuming that an N-dimensional MDCT line vector is quantized, the MMSE scalar value is close to optimal and generally too low. This results in loss of dispersion and spectral imbalance in the decoded output signal. This problem can be mitigated by variance-preserving decoding as described in FIG. 17h, where the reconstruction value is calculated as the weighted sum of the MMSE value and the value at the midpoint. A further optional improvement is to adapt the weight so that the MMSE value dominates for speech and the midpoint dominates for non-speech sounds. This results in cleaner speech while maintaining spectral balance and energy for non-speech sounds.

Variance-preserving decoding in accordance with an embodiment of the invention is achieved by determining a reconstruction point according to the following equation:

x _dequant = (1- Δ)x _MMSE + x _MP

Variance-preserving adaptive decoding can be based on the following rule for determining the interpolation coefficient:

The adaptive weight may further be a function of, for example, _{the LTP} gain g in LTP prediction: Δ = f(g _LTP ) . The adaptive weight changes slowly and can be efficiently encoded with a recursive entropy code.

The statistical model of the MDCT lines used in probability calculations (FIG. 17g) and dequantization (FIG. 17h) should reflect the actual signal statistics. In one version, the statistical model assumes that the MDCT lines are independent and have a Laplacian distribution. Another version models the MDCT lines as independent Gaussian distributions. One version models MDCT lines as Gaussian mixture models containing relationships between MDCT lines within and between MDCT frames. Another version adapts the statistical model to the current signal statistics. Adaptive statistical models may adapt forward and/or backward.

определяется из соответствующего индекса i _n квантования. Как упомянуто выше, значения

реконструкции дополнительно используются, например, в кодере линий MDCT (смотрите фиг. 17), чтобы определить остаток квантования для ввода в устройство квантования остатка. Дополнительно, реконструкция квантования выполняется в инверсном устройстве 304 квантования при реконструкции кодированного кадра MDCT для использования в буфере LTP (смотрите фиг. 3) и, естественно, в декодере.Another aspect of the invention relating to modified quantizer reconstruction points is shown schematically in FIG. 19, which shows the inverse quantizer used in the decoder of the embodiment. The module has, in addition to the usual inverse quantizer input signals, that is, quantized lines and information about the size of the quantization step (quantization type), also information about the reconstruction point of the quantizer. The inverse quantizer of this embodiment can use multiple types of reconstruction points when the reconstructed value

is determined from the corresponding quantization index i _n . As mentioned above, the values

the reconstructions are further used, for example, in the MDCT line encoder (see FIG. 17) to determine the quantization residual to be input to the residual quantizer. Additionally, quantization reconstruction is performed in the inverse quantizer 304 in reconstructing the encoded MDCT frame for use in the LTP buffer (see FIG. 3) and naturally in the decoder.

The inverse quantizer may select, for example, the middle point of the quantization interval as a reconstruction point or an MMSE reconstruction point. In an embodiment of the present invention, the reconstruction point of the quantizer is chosen to be an average between the center point and the reconstruction point of the MMSE. In general, a reconstruction point can be interpolated between a midpoint and an MMSE reconstruction point, for example, depending on signal properties such as signal periodicity. Information about the periodicity of the signal can be obtained, for example, from the LTP module. This feature allows the system to control distortion and energy conservation. The central reconstruction point guarantees energy conservation, while the MMSE reconstruction point guarantees minimal distortion. Given the signal, the system can then adapt the reconstruction point to where the best compromise is achieved.

The present invention further comprises a new window sequence code format. In accordance with an embodiment of the invention, the windows used for the MDCT transform have dyadic sizes and can only change in size from window to window by a factor of two. The dyadic transform sizes are, for example, 64, 128..., 2048 samples corresponding to 4, 8..., 128 ms at a sampling rate of 16 kHz. In general, variable size windows are provided that can take on multiple window sizes between a minimum window size and a maximum window size. In a sequence, successive windows can only be resized by a factor of two, so that smooth sequences of window sizes are formed without abrupt changes. Window sequences as defined by the embodiment, that is, limited to dyadic sizes and being able to change in size from window to window by only a factor of two, have several advantages. First, no special start or stop window is required, that is, windows with hard edges. This contributes to good time/frequency resolution. Second, the window sequence becomes very efficient for encoding, that is, to signal to the decoder which particular window sequence is being used. Finally, the sequence of windows will always fit exactly into the hyperframe structure.

The hyperframe structure is useful when running the encoder in a real system where certain decoder configuration parameters must be passed in order to be able to start the decoder. This data is usually stored in a header field in a bitstream describing the encoded audio signal. To minimize the data rate, no header is transmitted for each encoded data frame, especially in the system proposed in accordance with the present invention where the MDCT frame sizes can vary from very short to very long. Therefore, in accordance with the present invention, it is proposed to group a certain number of MDCT frames together into one hyperframe, in which header data is transmitted at the beginning of the hyperframe. A hyperframe is generally defined as having a certain duration in time. Therefore, care must be taken to ensure that MDCT frame size changes fit within a constant duration, a predetermined hyperframe duration. The inventive window sequence mentioned above ensures that the selected window sequence always fits within the hyperframe structure.

According to an embodiment of the present invention, the LTP delay and the LTP gain are encoded as a variable frequency. This is advantageous because, due to the efficiency of LTP for constant periodic signals, the LTP delay tends to be the same for somewhat long segments. Therefore, this can be used by the arithmetic coding engine, resulting in variable rate LTP delay and LTP gain coding.

Similarly, an embodiment of the present invention takes advantage of bit accumulator and variable rate coding for LP parameter coding. In addition, the present invention provides recursive LP coding.

Another aspect of the present invention is to deal with a bit accumulator for variable frame sizes in an encoder. In FIG. 18 shows a bit drive control unit 1800 according to the present invention. In addition to the measure of complexity provided as input, the bit accumulator control also receives information about the duration of the current frame. An example of a measure of complexity for use in a bit-accumulator control block is the perceptual entropy, or logarithm of the energy spectrum. Bit storage management is important in a system where frame durations may vary according to a set of different frame durations. The proposed bit accumulator manager 1800 considers the frame duration when calculating the number of bits to be provided for a frame to be encoded, as will be described below.

The bit accumulator is defined here as a certain fixed number of bits in the buffer, which must be greater than the average number of bits that a frame is allowed to use for a given data rate. If it has the same size, then no change in the number of bits for a frame can be possible. The bitstore control always checks the level of the bitstore before extracting the bits that will be provided to the coding algorithm as the allowed number of bits for the actual frame. Thus, a full bitaccumulator means that the number of bits available in the bitaccumulator is equal to the size of the bitaccumulator. After a frame has been encoded, the number of bits used will be subtracted from the buffer and the bit accumulator will be updated with some number of bits representing a constant bit rate. Therefore, the bitaccumulator is empty if the number of bits in the bitaccumulator before encoding the frame is equal to the average number of bits per frame.

In FIG. 18a shows the basic concept of bit storage management. The encoder provides a means of calculating how difficult it is to encode the actual frame compared to the previous frame. For an average complexity of 1.0, the number of bits provided depends on the number of bits available in the bitstore. According to the predetermined control line, a number greater than the number of bits corresponding to the average data rate will be removed from the bit store if the bit store is completely full. In the case of an empty bit accumulator, fewer bits will be used to encode a frame compared to the average number of bits in the frame. This behavior results in an average bit accumulator level for a longer sequence of frames with medium complexity. For frames of higher complexity, the control line may be shifted upward, resulting in the effect that complex frame coding is allowed to use more bits for the same bit accumulator level. Accordingly, in order to facilitate frame coding, the number of bits allowed for a frame should be reduced by simply shifting down the control line in FIG. 18a in relation to the case of medium complexity to the case of less complexity. Modifications other than simple shifting of the control line are also possible. For example, as shown in FIG. 18a, the steepness of the control curve may vary depending on the complexity of the frames.

When calculating the number of bits to be provided, the limits on the lower bound of the bit accumulator must be obeyed so as not to take more bits from the buffer than allowed. A bit accumulator control circuit comprising calculating the provided bits with the control line as shown in FIG. 18a is just one example of a possible measure of bit storage level and complexity in ratios of provided bits. Other control algorithms will also have, in general, hard limits on the lower bound of the bit-accumulator level that prevent the bit-accumulator from violating the bit-accumulator-empty constraint, as well as limits on the upper bound, where the encoder will be forced to write padding bits if the encoder is running out of too few bits.

For such a control mechanism capable of handling a set of variable frame sizes, this simple control algorithm must be adapted. The measure of complexity used must be normalized so that the complexity values of different frame sizes are comparable. For each frame size there will be a different allowable range of bits to be provided and therefore the average number of bits per frame is different for a variable frame size, hence each frame size has its own control equation with its own limitations. One example is shown in FIG. 18b. An important modification of the fixed frame size case is the reduced margin of the control algorithm. Instead of the average number of bits for the actual frame size, which corresponds to the fixed data rate case, the average number of bits for the largest allowable frame size is now the lowest allowable value for the bit accumulator layer before extracting the bits for the actual frame. This is one of the main differences for bit accumulator management for frames with fixed sizes. This constraint ensures that the next frame with the largest possible frame size can use at least the average number of bits for that frame size.

The measure of complexity may be based, for example, on the calculation of the perceptual entropy (PE), which is obtained from the masking thresholds of the psychoacoustic model, as is done in AAC, or, alternatively, from the number of quantization bits with a fixed step size, as is done in the ECQ part of the encoder. in accordance with an embodiment of the present invention. These values can be normalized with respect to variable frame sizes, which can be achieved by simply dividing by the frame duration and the result is the corresponding PE number of bits per sample. Another normalization step may take place with respect to the average complexity. For this purpose, a moving average over legacy frames can be used, resulting in a complexity value greater than 1.0 for complex frames or less than 1.0 for simple frames. In the case of a two-pass or large look ahead encoder, future frame complexity values may also be taken into account for this normalization of the complexity measure.

Another aspect of the invention relates to the specifics of working with a bit accumulator for ECQ. Bit accumulator control for ECQ operates under the assumption that ECQ produces approximately constant quality when it uses a constant quantizer step size for encoding. A constant quantizer step size creates a variable frequency, and the task of the bit-accumulator is to keep the change in quantizer step-size as small as possible for different frames without violating the bit-accumulator buffer constraints. In addition to the frequency created by ECQ, additional information (eg, LTP gain and delay) is transmitted based on the MDCT frame. The side information is generally also entropy encoded and thus uses a different frequency from frame to frame.

In an embodiment of the invention, the proposed bit accumulator control attempts to minimize ECQ step size variation by introducing three variables (see FIG. 18c):

- R _{ECQ_AVG} : average ECQ frequency per sample used previously;

- Δ _{ECQ_AVG} : average quantizer step size used previously.

Both of these variables are dynamically updated to reflect the latest encoding statistics.

- R _{ECQ_AVG_DES} : ECQ frequency corresponding to the average overall data rate.

This value will be different from R _{ECQ_AVG} if the bit accumulator level changed during the time frame of the averaging window, for example, during this time frame a data rate was used that is higher or lower than the specified average data rate. It is also updated as the frequency of the side information changes so that the overall frequency equals the specified data rate.

The bit accumulator control uses these three values to determine the initial delta guess that should be used for the current frame. This is done by finding _{ΔECG_AVG_DES} on the R _ECQ-Δ curve shown in FIG. 18c which corresponds to R _{ECQ_AVG_DES} . In the second step, this value is possibly changed if the frequency does not meet the limits of the bit accumulator. An example of the R _ECQ-Δ curve shown in FIG. 18C is based on the following equation:

Of course, other mathematical relationships between R _ECQ and Δ can also be used.

In the stationary case, R _{ECQ_AVG} will be close to R _{ECQ_AVG_DES} and the change in Δ will be very small. In the non-stationary case, the averaging operation guarantees a smooth change in Δ.

Although the foregoing has been disclosed in relation to specific embodiments of the present invention, it is understood that the concept of the invention is not limited to the described embodiments. On the other hand, the disclosure provided in this application will enable those skilled in the art to understand and perform the invention. Those skilled in the art will appreciate that various changes may be made without departing from the spirit and scope of the invention, solely as set forth in the accompanying claims.

Claims (39)

1. Audio encoder containing: block (201) linear prediction (LP) for analyzing frames of the first duration of the audio signal to determine the sequence of LPC polynomials; an adaptive duration transform block (202) for converting frames of a variable second duration of the audio signal into frames of modified discrete cosine transform (MDCT) coefficients; a gain curve generating unit (970, 1070) for generating MDCT domain gain curves for frames of MDCT coefficients based on response values determined from a sequence of LPC polynomials, wherein the creation of gain curves in the MDCT domain comprises mapping, by means of a mapping unit (1100), the polynomials LPC into corresponding frames of MDCT coefficients; a gain curve application unit for applying gain curves in the MDCT region to frames of MDCT coefficients to create frames of gain-corrected MDCT coefficients; And a quantizer (203) for quantizing frames of gain-corrected MDCT coefficients. 2. Audio encoder according to claim 1, comprising a window sequencer for determining, for the audio signal block, second frame durations for overlapping MDCT windows. 3. An audio encoder according to any one of the preceding claims, comprising a perceptual modeler modifying the LP filter response by chirping and/or tilting the LPC polynomial generated by the linear predictor for the LPC frame. 4. An audio encoder according to any one of the preceding claims, comprising: a frequency separation unit for separating the audio signal into a low frequency component and a high frequency component; And a high frequency encoder for encoding a high frequency component, wherein the low frequency component is an input to the linear prediction block and the transform block, and wherein the high frequency encoder is a spectral band replication encoder. 5. The audio encoder of claim 4, wherein the frequency splitter comprises a quadrature mirror filter store and a quadrature mirror filter synthesis block capable of downsampling the audio signal. 6. An audio encoder according to claim 4 or 5, wherein the boundary between the low band and the high band can be changed and the splitter determines the split frequency based on the properties of the audio signal and/or the bandwidth requirements of the encoder. 7. An audio encoder according to any one of the preceding claims, comprising a scaling factor determiner (1360) for determining scaling factors to control the quantization noise of the quantization block (203). 8. The audio encoder of claim 7, wherein the scale factors are determined based on the transformed gain curves in the MDCT domain. 9. An audio encoder according to any one of the preceding claims, comprising a parametric stereo block for computing a parametric stereo representation of the left and right input channels. 10. The audio encoder of any one of the preceding claims, wherein the mapper (1500) interpolates LP polynomials generated at a rate corresponding to the duration of the first frame so that they correspond to frames of a signal in the MDCT region generated at a rate corresponding to the duration of the second frame. 11. An audio decoder comprising: a dequantizer (211) for dequantizing the quantized MDCT coefficient frames decoded from the input bitstream, the MDCT coefficient frames representing an audio signal; a gain curve generating unit (970, 1070) for generating MDCT region gain curves for frames of MDCT coefficients based on response values determined from a sequence of LPC polynomials, wherein the LPC polynomials are obtained from an input bitstream, and wherein generating MDCT region gain curves comprises mapping , block (1100) mapping, LPC polynomials in the corresponding frames of the MDCT coefficients; a gain curve application unit for applying gain curves in the MDCT region to frames of MDCT coefficients to create frames of gain-corrected MDCT coefficients; And an adaptive length MDCT inverse transform block (212) for inversely transforming frames of gain-corrected MDCT coefficients into a time domain audio signal, the MDCT inverse transform block operating with variable frame length. 12. An audio coding method, comprising the steps of: performing a linear prediction (LP) analysis on frames of the first duration of the audio signal to determine a sequence of LPC polynomials; converting frames of a variable second duration of the audio signal into frames of modified discrete cosine transform (MDCT) coefficients; generating MDCT domain gain curves for frames of MDCT coefficients based on response values determined from the sequence of LPC polynomials, wherein generating the gain curves in the MDCT domain comprises mapping the LPC polynomials to corresponding frames of MDCT coefficients; applying the gain curves in the MDCT region to the MDCT coefficient frames to obtain gain-corrected MDCT coefficient frames; And quantizing frames of gain-corrected MDCT coefficients. 13. An audio decoding method, comprising the steps of: dequantizing the quantized MDCT coefficient frames decoded from the input bitstream, the MDCT coefficient frames representing the audio signal; generating MDCT region gain curves for MDCT coefficient frames based on response values determined from LPC polynomials, the LPC polynomials being obtained from the input bitstream, and wherein generating the MDCT region gain curves comprises mapping the LPC polynomials to corresponding MDCT coefficient frames; applying gain curves in the MDCT region to frames of MDCT coefficients to create frames of gain-corrected MDCT coefficients; inversely transforming frames of gain-corrected MDCT coefficients into a time domain audio signal using inverse MDCT operating with variable frame length. 14. A programmable device comprising a computer program for causing the programmable device to perform the audio encoding method of claim 12. 15. A programmable device comprising a computer program for causing the programmable device to perform the audio decoding method of claim 13. 16. An audio coding system comprising the audio encoder of claim 1 and the audio decoder of claim 11.

Images