CN1258172C - Apparatus and method for encoding and decoding audio signals - Google Patents

Abstract

The discrete-time audio signal is processed (52) to provide a quantization block with quantized spectral values. Furthermore, an integer spectral representation is generated from the discrete-time audio signal using an integer transform algorithm (56). The quantization block generated using a psychoacoustic model (54) is inversely quantized and rounded (58) to subsequently form a difference between the integer spectral values and the inversely quantized rounded spectral values. After decoding, this quantization block alone provides a lossy psychoacoustic encoded/decoded audio signal; while during decoding, this quantization block, together with the combination module, provides a lossless or near-lossless encoded and re-decoded audio signal. By generating differential signals in the frequency domain, a simple encoder/decoder structure is formed.

Description

Apparatus and method for encoding and decoding audio signals

technical field

The present invention relates to audio encoding/decoding, and in particular to scalable encoding/decoding algorithms comprising a psychoacoustic first extension layer and a second extension including auxiliary audio data for lossless decoding layer.

Background technique

Modern audio coding methods, such as MPEG Layer 3 (MP3) or MPEG ACC, use transforms such as the so-called Modified Discrete Cosine Transform (MDCT) to obtain a block-wise frequency representation of the audio signal. Such audio encoders typically obtain a data stream of time-discrete audio samples. The stream of audio samples is windowed to obtain windowed blocks of eg 1024 or 2048 windowed audio samples. Various window functions are used for windowing, such as sine window and so on.

Subsequently, the windowed time-discrete audio samples are converted to a spectral representation through a filter bank. In principle, the Fourier transform, or a variety of Fourier transforms for special reasons, such as the FFT, or the MDCT explained earlier, can be used for this. The block of audio spectral values at the output of the filter bank can then be further processed as required. In the audio coders cited above, quantization of the audio spectrum is followed, where the quantization level is typically chosen such that the quantization noise introduced by the quantization is below the psychoacoustic masking threshold, that is to say "masked". Quantization is a lossy encoding. In order to obtain a further data volume reduction, the quantized spectral values are entropy coded, for example by Huffman coding. By adding auxiliary information, such as scale factors, a bitstream that can be stored or transmitted is formed from the entropy-encoded quantized spectral values through a bitstream multiplexer.

In the audio decoder, the bitstream is split by a bitstream demultiplexer into encoded quantized spectral values and side information. The entropy encoded quantized spectral values are first entropy decoded to obtain quantized spectral values. The quantized spectral values are then dequantized to obtain decoded spectral values that contain quantization noise, however, such quantization noise is below the physiological acoustic masking threshold and thus is inaudible. These spectral values are then converted to a temporal representation through a synthesis filterbank to obtain time-discrete decoded audio samples. In synthesizing filter banks, a transform algorithm that is the inverse of the transform algorithm must be used. Also, after frequency-to-time conversion or inverse conversion, the window must be canceled.

In order to obtain good frequency selectivity, modern audio coders typically utilize block overlapping. This situation is shown in Figure 4a. First, for example, 2048 time-discrete audio samples are taken out by means 402 and windowed. The means 402 for implementing such a window has a window length of 2N samples and provides at output a data block of 2N windowed samples. To obtain window overlap, by means 404 (which is depicted separately from means 402 in FIG. 4a only for clarity of presentation), a second block of 2N windowed samples is formed. However, instead of the time-discrete audio samples immediately following the first window, the 2048 samples fed into device 404 contain the second half of the samples windowed by device 402, and in addition only contain 1024 "new" sampling. This overlap is schematically illustrated in Fig. 4a by means 406, causing an overlap of 50%. Then, for the 2N windowed sample outputs passing through the device 402 and the 2N windowed sample outputs passing through the device 404, the MDCT algorithm is implemented by means 408 and 410, respectively. The device 408 provides N spectral values for the first window according to the known MDCT algorithm, and the device 410 also provides N spectral values, but for the second window, wherein the first window and the second window There is a 50% overlap between them.

In the decoder, the N spectral values of the first window, as shown in Fig. 4b, are fed into the means 412 to implement the modified inverse discrete cosine transform. The same operation is applied to the N spectral values of the second window. They are fed into means 414, which also implements the Modified Inverse Discrete Cosine Transform. Both means 412 and means 414 provide 2N samples for the first window and the second window, respectively.

In means 416, denoted TDAC (Time Domain Aliasing Cancellation) in Fig. 4b, it is taken into account that the two windows are overlapping. In particular, a sample y ₁ (that is, with coefficient N+k) from the second half of the first window is added to a sample y ₂ (that is, with coefficient k) from the first half of the second window, such that At the output, ie at the decoder, N decoded time-domain samples are generated.

It should be noted that the windowing implemented in the encoder shown in Fig. 4a is taken into account to a certain extent automatically by the function of the means 416, also called the addition function, so it is not necessary in the decoder shown in Fig. 4b There is obvious "reverse windowing" happening.

When the window function implemented by means 402 or 404 is designated as w(k), where the coefficient k represents the time coefficient, the condition that must be satisfied is that the squared window weight w(k) and the squared window weight w(N+k ) equals 1, where k ranges from 0 to N-1. When using a sine window, the weight of the window follows the first half of the sine function, which is always true because the sum of the squared sine and the squared cosine of any angle is 1.

The disadvantage of the windowing method according to the MDCT function described in Fig. 4a is that windowing is performed by multiplying time-discrete samples, which is achieved by a floating point number when considering it as a sinusoidal window, since one at 0 The sine of an angle between 180 and 180 does not produce an integer unless the angle is equal to 90. Even when integer-time discrete samples are windowed, floating-point numbers are produced after windowing.

Therefore, quantization at the output of means 408 or 410 is necessary for proper tractable entropy coding even when psychoacoustic coding is not used, ie when lossless coding is to be obtained.

When known transforms, as described on the basis of Fig. 4a, are applied to lossless audio coding, it is necessary to use very good quantization, so that the resulting errors due to rounding of floating-point numbers can be ignored, or the error signal needs to be e.g. are additionally encoded in the time domain.

Concepts in the prior art, namely in which quantization is adjusted so well that result errors due to rounding of floating-point numbers can be ignored, are disclosed, for example, in German patent DE 19742 201 C1. Here, an audio signal is converted to its spectral representation and quantized to obtain quantized spectral values. The quantized spectral values are then dequantized, transformed into the time domain, and compared with the original audio signal. If the error, that is, the error between the original audio signal and the quantized/inverse quantized audio signal, is above an error threshold, the quantizer is adjusted to be more accurate in the feedback, and the comparison is made again. When it is below the error threshold, stop the iteration. The residual signal that may still be present is encoded by a time-domain encoder and written to a bitstream that includes, in addition to the time-domain encoded residual signal, the encoded spectrum quantized according to the quantizer adjustments that existed at the time of the iterative cancellation value. It should be noted that the quantizer does not necessarily have to be controlled by the psychoacoustic model, so that the coded spectral values are usually quantized more precisely than those obtained due to the use of the psychoacoustic model.

In the publication "A Design of Lossy and Lossless Scalable AudioCoding" (T.Moriya et al., Proc. ICASSP, 2000) a scalable encoder is described which includes e.g. an MPEG encoder as the first Lossy data compression module, this module takes as input a digital signal form in the form of data blocks and generates a compressed bit stream. In another existing native decoder the encoding is canceled again and an encoded/decoded signal is generated. This signal is compared to the original input signal by subtracting the encoded/decoded signal from the original input signal. The error signal is then sent to a second block where a lossless bit converter is used. This conversion has two steps. The first step involves a conversion from two's complement format to signed numeric format. The second step involves the conversion from a vertical sequence of values to a horizontal sequence of bits in one processing block. Lossless data conversion is performed to maximize the number of zeros or to maximize the number of consecutive zeros in a sequence in order to obtain the best possible temporal error signal represented as a digital result. This principle is based on the Bit Slice Algorithm Coding (BSAC) scheme explained in the publication "Multi-Layer Bit Sliced Bit Rate Scalable AudioCoder" (103 ^rd AES Convention, Preprint No. 4520, 1997).

A disadvantage of the above concept is that the data for the lossless extension layer, ie the auxiliary data for decoding to obtain a lossless audio signal, must be obtained in the time domain. This means that obtaining an encoded/decoded signal in the time domain requires a full decoding including frequency/time transforms, so the error signal is computed by forming the sample difference between the original audio input signal and the encoded/decoded audio signal, the encoded/decoded Decoded audio signals are lossy due to psychoacoustic encoding. The disadvantage of this concept is especially that when the encoder generates the audio data stream, two complete time/frequency transformation devices, such as filter banks or algorithms such as MDCT, are required for forward transformation, on the other hand, only for To generate the error signal, a complete inverse filter bank or a complete synthesis algorithm is required. Thus, an encoder must have full decoder functionality in addition to its inherent encoder functionality. If the encoder is implemented by software, there are requirements for both storage performance and processor performance, which results in increased overhead for the implementation of the encoder.

Contents of the invention

It is an object of the invention to provide an inexpensive concept with which an audio data stream can be generated which can be decoded in an almost lossless manner.

This object is achieved by a device for encoding a time-discrete audio signal in claim 1, a method for encoding a time-discrete audio signal in claim 21, a device for decoding encoded audio data in claim 22, and a device for encoding a time-discrete audio signal in claim 22. 31, or a computer program in claim 32 or 33.

The invention is based on the discovery that an auxiliary audio signal capable of lossless decoding of the audio signal can be achieved by providing a data block of quantized spectral values as usual and then dequantizing it to obtain dequantized spectral values, The dequantized spectral values are lossy due to quantization using a psychoacoustic model. These inverse quantized spectral values are then rounded to obtain rounded blocks of rounded inverse quantized spectral values. As a reference for forming the difference, according to the invention an integer transformation algorithm is used which generates an integer block of spectral values containing only integer spectral values from a block of discrete samples at integer times. According to the present invention, the combination of the rounding block and the spectral value in the integer block is now realized in the form of spectral value, that is to say in the frequency domain, so no synthesis algorithm is needed in the encoder itself, that is, the reverse Filter bank or inverse MDCT algorithm, etc. Due to the integer transformation algorithm and the rounding of quantization values, a combined block containing different spectral values only contains integer values that can be entropy coded in some known way. It should be noted that any entropy encoder can be used for entropy encoding of combined blocks, such as Huffman encoders and algorithmic encoders.

The encoding of the quantized spectral values of the quantized blocks can also use any encoder, such as known tools commonly used by modern audio encoders.

It is worth noting that the encoding/decoding concept of the present invention is compatible with modern encoding devices such as window switching, TNS, or center/edge encoding of multi-channel audio signals.

In a preferred embodiment of the invention, MDCT is used to provide a quantized block of spectral values quantized using a psychoacoustic model. Also, it is better to use a so-called IntMDCT as an integer transformation algorithm.

In an alternative embodiment of the present invention, the usual MDCT may not be used, and the IntMDCT may be used as an approximation of the MDCT, that is, the integer spectrum obtained by the integer transform algorithm is used in a psychoacoustic quantizer to obtain quantized IntMDCT spectral values, this spectrum The values are then dequantized and rounded again for comparison with the original integer spectral values. In this case, only a single transform is required, namely IntMDCT to produce integer spectral values from integer-time discrete samples.

Typically, processors handle integers, or each floating point number is represented as an integer. If an integer algorithm is used in a processor, it can eliminate the need to round the dequantized spectral values, because due to the processor's algorithm rounding the values, that is, within the precision of the LSB, i.e. the least significant bit, is always existing. In this case, completely lossless processing is achieved, that is, processing within the precision of the processor being used. Optionally, however, rounding to an approximate accuracy is also possible, so that the differential signal in the synthesis block is rounded to an accuracy determined by the rounding function. In order to generate an encoder that is almost lossless in the sense of data compression, rounding is introduced in addition to the rounding of the original processing system, which enhances flexibility and thus affects the degree of lossless encoding.

The decoder according to the invention is itself particularly distinguished both with respect to psychoacoustically encoded audio data and with auxiliary audio data extracted from the audio data, subjected to possible entropy decoding, and then processed as follows. First, the quantized block is dequantized in the decoder and rounded using the same rounding algorithm as in the encoder, which can then be added to the entropy-decoded auxiliary audio data. In the decoder, the spectral representation of the psychoacoustically compressed audio signal then co-exists with a lossless representation of the audio signal, wherein the psychoacoustically compressed spectral representation of the audio signal is transformed into the time domain to obtain a lossless encoded/decoded audio signal, The lossless representation is instead transformed to the time domain by using an integer transformation algorithm inverse to that used to obtain a lossless, or, as mentioned above, substantially lossless encoded/decoded audio signal.

Description of drawings

The above-mentioned and other objects and characteristics of the present invention will be clearer in the following description in conjunction with the accompanying drawings:

1 is a block circuit diagram of a preferred apparatus for processing time-discrete audio samples to obtain integer values from which integer spectral values can be determined;

Figure 2 is a schematic diagram of the decomposition of MDCT and inverse MDCT in Givens rotation and two DCT-IV operations;

Figure 3 is a legend representation of the MDCT decomposition with 50% overlap in rotation and DCT-TV operations;

Figure 4a is a schematic circuit block diagram of a known encoder with MDCT and 50% overlap;

Figure 4b is a block circuit diagram of a known decoder for decoding the values generated in Figure 4a;

Fig. 5 is a preferred schematic circuit block diagram of an encoder according to the present invention;

Fig. 6 is a schematic circuit block diagram of an alternative preferred inventive decoder;

Fig. 7 is a schematic circuit block diagram of an inventive preferred decoder;

Figure 8a is a schematic diagram of a bitstream with a first extension layer and a second extension layer;

Figure 8b is a schematic diagram of a bitstream with a first extension layer and a plurality of other extension layers;

Fig. 9 is a schematic diagram of binary coded differential spectral values for representing possible spreading ratios in relation to the precision (bits) of the differential spectral values and/or in relation to the frequency (sampling rate) of the differential spectral values.

Detailed ways

On the basis of FIGS. 5 to 7, the following will discuss an inventive encoder circuit (FIGS. 5 and 6) or an inventive preferred decoder circuit (FIG. 7). The encoder of the invention shown in Figure 5 comprises an input 50 to which a time-discrete audio signal is fed and an output 52 which outputs encoded audio data. The time-discrete audio signal at the input 50 is fed into a device 52 to provide a quantized block which provides at the output a quantized block of the time-discrete audio signal comprising the time-discrete spectrum using a physiological acoustic model 54 Quantized spectral values of the audio signal 50 . The encoder of the present invention also includes means for generating an integer block using an integer transformation algorithm 56 effective for generating integer spectral values from discrete samples at integer times.

The inventive encoder also includes means 58 for inverse quantizing the output of the quantized block from means 52 and, when a precision different from the processor precision is required, a rounding function. As stated, the rounding function is already inherently included in the inverse quantization of the quantization block if the processor system's precision has been achieved, since a processor with integer arithmetic is by no means capable of providing non-integer values . The means 58 then provide a so-called rounding block comprising inverse quantized spectral values which are inherently or explicitly rounded to integers. Both the rounded block and the integer block are fed to combining means for providing a difference with difference spectral values using difference forming, where the term "difference block" means that the difference spectral value is the difference between the containing integer block and the rounded block value.

Both the quantized block output from means 52 and the differential block output from difference forming means 58 are fed to processing means 60 for performing quantized block processing as usual and for example causing entropy coding of the differential block. The processing means 60 output the encoded audio data at the output terminal 52, these data include the information of the quantization block and also the information of the difference block.

In a first preferred embodiment, as shown in Fig. 6, the time-discrete audio signal is converted into a spectral representation by MDCT method and then quantized. Means 52 for providing quantized blocks have MDCT means 52a and a quantizer 52b.

Also, it is best to use IntMDCT56 as the integer conversion algorithm to generate integer blocks.

In FIG. 6, the processing device 60 shown in FIG. 5 is also described as a bit stream encoding device 60a and an entropy encoder 60b. The bit stream encoding device 60a is used to perform bit stream encoding on the quantized block output by the device 52b. The entropy encoding The unit 60b is used to perform entropy coding on the differential block. The bitstream encoder 60a outputs physiologically-acoustically encoded audio data, while the entropy encoder 60b outputs entropy-encoded difference blocks. The two output data blocks of modules 60a and 60b can be combined in a suitable way into a bitstream with the audio data encoded physiologically as a first extension layer and the auxiliary audio data for lossless decoding as a second extension layer. Two expansion layers. This expanded bitstream then corresponds to the encoded audio data shown in FIG. 5 at the output 52 of the encoder.

In an alternative preferred embodiment, the MDCT block 52a in FIG. 6 may not be used as it has been implied by the dashed arrow 62 in FIG. 5 . In this case, the integer spectrum provided by the integer transformation means 56 is supplied to the difference forming means 58 and the quantizer 52b of FIG. 6 . The spectral values generated by the integer transformation algorithm are used here in one way as an approximation of the usual MDCT spectrum. The benefit of this embodiment is that only the IntMDCT algorithm is present in the encoder, rather than both IntMDCT and MDCT algorithms need to be present.

Referring again to FIG. 6, it should be noted that the solid boxes and lines represent a general audio coder conforming to an MPEG standard, while the dashed boxes and dashed lines represent extensions of such a general MPEG coder. Therefore, it can be seen that there is no fundamental change to the general MPEG encoder, but the method of adding an integer transformer to capture the lossless encoded auxiliary audio data does not need to change the basic structure of the encoder/decoder.

FIG. 7 shows a schematic block circuit diagram of an inventive decoder for decoding the encoded audio data output at output 52 of FIG. 5 . It is first decomposed into psychoacoustically encoded audio data on the one hand and auxiliary audio data on the other hand. The psychoacoustically encoded audio data is fed to a conventional bitstream decoder 70, while the auxiliary audio data, after being entropy encoded by the encoder, is entropy encoded by an encoder 72. At the output of the bitstream decoder 70 in FIG. 7 there are quantized spectral values which can in principle be fed to an inverse quantizer 74 of the same structure as the inverse quantizer in the arrangement of FIG. 6 . If it is necessary to achieve a precision different from the precision of the processor, a rounding device 76 is also provided in the decoder. The rounding device 76 is the same as the device 58 in FIG. 6, and realizes the same algorithm for mapping a real number into an integer Or the same rounding function. In a decoder combiner 78, the rounded inverse quantized spectral values are preferably combined with the entropy encoded auxiliary audio data by adding the spectral values, so that in the decoder, on the one hand, the inverse quantized spectral values appear At the output of the means 74 , on the other hand integer spectral values appear at the output of a combiner 78 .

The spectral values at the output of the means 74 can then be transformed by means 80 into the time domain in order to perform a modified inverse discrete cosine transform to obtain a lossy psychoacoustically encoded and re-decoded audio signal. In order to perform an inverse integer MDCT (IntMDCT), the output signal of the synthesizer 78 can also be transformed into its time form by means 82 to produce a lossless encoded/decoded audio signal, or in a more coarse rounded Sometimes, an almost lossless encoded and re-decoded audio signal can be produced.

A particularly preferred implementation of the entropy coder 60b in FIG. 6 will now be considered. In typical modern MPEG coders, multiple codetables are selected based on an average statistic of quantized spectral values. Preferably the differential blocks at the output of the combiner 58 are entropy coded using the same code table or codebook. Since the size of the differential block, ie the residual IntMDCT spectrum, depends on the precision of the quantization, the code table selection by the entropy encoder 60b can be performed without auxiliary edge information.

In an MPEG-2 AAC decoder, the spectral coefficients, i.e. quantized spectral values, are grouped into scalefactor bands in quantized blocks, where the spectral values are weighted by gain factors from the corresponding scalefactors associated with the scalefactor bands . Since in this known encoder concept a non-uniform quantizer is used to quantize the weighted spectral values, the size of the residual value, i.e. the spectral value at the output of the combiner 58, depends not only on the scale factor, Also depends on the quantization value itself. However, since the scale factor and the quantized spectral value are included in the bit stream generated by the device 60a of Fig. 6, that is, in the psychoacoustic coded audio data, it is better to realize the codebook selection in the decoder according to the magnitude of the differential spectral value , and on the basis of the scale factor and quantization value transmitted in the bit stream, the code table used in the decoder is determined. Since no side information needs to be transmitted at the output of the combiner 58 to entropy code the differential spectral values, the entropy coding only results in data rate compression without the need to spread any signaling bits in the data stream as side information for the entropy coder 60b .

In an audio codec conforming to the standard MPEG-2 AAC, window switching is used to avoid forward echoes in the transient audio signal domain. This technique is based on the possibility to choose the window shape separately in each half of the MDCT window, enabling the block size to be varied in successive blocks. Likewise, integer transform algorithms of the form IntMDCT (this algorithm is explained with reference to Figures 1 to 3) are also performed using different window shapes for windowing and aliasing in the temporal MDCT decomposition. Thus, preferably the same window discrimination is used for the integer transform algorithm and the transform algorithm that generates the quantized blocks.

In an MPEG-2AAC-compliant encoder, there are many other encoding tools, here only TNS (temporal noise shaping) and mid/edge (CS) stereo coding are introduced. In TNS coding, just as in CS coding, the spectral values are corrected before quantization. Next, the IntMDCT value, ie the difference between the integer blocks, and the quantized MDCT value are increased. According to the invention, an integer transformation algorithm is formed to accommodate TNS coded and intermediate/edge coded integer spectral values. The TNS technique is based on adaptive forward prediction of MDCT values in frequency. The same predictive filter computed by a normal TNS module in a signal-adaptive manner is preferably also used to predict integer spectral values, and if non-integer values are generated therein, rounding down is used to generate integer values again . This rounding preferably happens after each prediction step. In the decoder, the original spectrum can be reconstructed again by using the inverse filter and the same rounding function. Likewise, CS coding can also be based on the lifting method for the IntMDCT spectral values by using a rounded Givens rotation with an angle π/4. Therefore, the original IntMDCT value in the decoder can be reconstructed.

It should be noted that in the preferred embodiment using IntMDCT as the integer transform algorithm, the concept of the present invention can be applied to all MDCT-based auditory adaptive audio coders. Just as an example, these encoders are encoders based on MPEG-4 AAC Scalable, MPEG-4 AAC Low Latency, MPEG-4 BSAC, MPEG-4 Twin VQ, DolbyAC-3, etc.

In particular, note that this inventive concept is backward compatible. The auditory adaptive coder or decoder is not changed, but only extended. Auxiliary information for lossless components may be transmitted in an auditory-adaptive coded bitstream in a backwards-compatible manner, such as MPEG-2 AAC in the "ancillary data" field. An additional part of the preceding auditory-adaptive decoder, shown in dashed lines in Fig. 7, can estimate and reconstruct the auxiliary data together with the quantized MDCT spectrum and the IntMDCT spectrum obtained losslessly from the auditory-adaptive decoder.

The inventive concept of psychoacoustic coding is particularly well suited for generating, transmitting and decoding scalable data streams, complemented by lossless or nearly lossless coding. Scalable data streams are known to contain many different layers of scaling. Of these, at least the lowest extension layer can be transmitted and decoded independently of higher extension layers. In scalable processing of data, other extension or enhancement layers are superimposed on the first extension or base layer. A complete encoder can produce a scalable data stream with a first scalable layer and, in principle, any number of other scalable layers. An advantage of the scalability concept is that the scalable data stream generated by the encoder can be completely transmitted, provided a broadband transmission channel is available. That is to say, including all extensible layers can be transmitted through the broadband transmission channel. However, if there is only one narrowband transport channel, the coded signal can still be sent over the transport channel, but only in the form of the first extension layer or a certain number of other extension layers. Wherein the number of other extension layers is less than the number of all extension layers generated by the encoder. Of course, an encoder coupled to and adapted to the channel may already produce a base or first extension layer and a number of other channel-dependent extension layers.

On the decoder side, the extensibility concept also has the advantage of being backwards compatible. This means that a decoder that can only handle the first extension layer ignores the second and other extension layers in the data stream and can produce a useful output signal. However, if the decoder is a typically more modern decoder capable of handling multiple extension layers in an extension data stream, then this encoder can handle the same data stream as the base decoder.

In the present invention, the basic scalability is that the quantized module, i.e. the output of the bitstream encoder 60a, is written into the first extension layer 81 of Figure 8, which contains the psychological Acoustically encoded data, such as frames. The preferably entropy-coded differential spectral values generated by the combining means 58 are written into the second expansion layer, this simple scalability being indicated by 82 in FIG. 8a. Thus for a frame, auxiliary audio data is included.

If the transport channel from the encoder to the decoder is a broadband transport channel, both extension layers 81 and 82 can be sent to the decoder. But if this transport channel is a narrowband transport channel, only the first extension layer is "compliant", the second extension layer can be directly removed from the data stream before the data is sent, so the decoder only processes the first extension layer.

On the decoder side, a "base decoder" that can only process psychoacoustically encoded data can simply ignore the second extension layer when it is received over a wideband channel. But if this decoder is a full decoder with a psychoacoustic decoding algorithm and an integer decoding algorithm, then it can be decoded with the first and second extension layers to produce a lossless encoded and decoded output signal.

A preferred embodiment of the invention is schematically shown in Figure 8a, the psychoacoustically encoded data for the frame is also placed in the first extension layer. The second extension layer in Fig. 8a is quantized more finely, so that from this second extension layer in Fig. 8 emerges multiple extension layers, e.g. a (smaller) second extension layer, third extension layer, Four expansion layers and more.

The differential spectral values output from adder 58 are particularly suitable for further quantization, as shown based on FIG. 9 . Fig. 9 schematically shows binary coded spectral values. Each row 90 in Fig. 9 represents a binary coded differential spectral value. In FIG. 9 the differential spectral values are classified according to frequency, which is indicated by arrows 91 in the figure. A differential spectral value 92 has a higher frequency than differential spectral value 90 . The first column in the table in FIG. 9 represents the most significant bit in a differential spectrum value; the second number represents the bit whose significance is MSB-1; and the third number represents the bit whose significance is MSB-2. The second-to-last column represents the bit whose significant digit is LSB+2; the penultimate column represents the bit whose significant digit is LSB+1; the last column represents the bit whose significant digit is LSB, that is, the least significant bit of a differential spectrum value.

In a preferred embodiment of the invention, eg the 16 most significant bits of the differential spectral value appear in the second extension layer for precise quantization and thus entropy encoding by entropy encoder 60b, if desired. A decoder employing the second extension layer obtains the differential spectral values at the output with 16-bit precision, so that the second extension layer and the first extension layer together provide a CD-quality lossless decoded audio signal. 16-bit CD-quality audio samples are known to exist.

On the other hand, if the encoder is provided with a studio-quality audio signal, i.e., an audio signal containing 24 bits per sample, the encoder may further generate a third extension layer containing the last 8 bits of the differential spectral value, and Entropy encoding is performed as needed (means 60 of Figure 6).

A complete decoder obtains the data streams of the first extension layer, the second extension layer (16 most significant bits of difference spectrum value) and the third extension layer (8 second most significant bits of difference spectrum value), and this decoder can provide A lossless, studio-quality encoded/decoded audio signal, that is, using all three extension layers to provide a 24-bit sample wordwidth at the output of the decoder.

It should be noted that the audio signal in the studio field has a longer sample word length than the audio signal in the general consumer field. In the consumer field, the signal word width in an audio CD is 16 bits, and in the studio field it is 24 or 20 bits.

Based on the concept of scaling in the IntMDCT domain, all three precisions (16-bit, 20-bit or 24-bit) or any precision quantized with a minimum of 1 bit can be quantized and coded as described above.

Here, an audio signal represented with 24-bit precision is represented in the integer frequency domain by means of an inverse IntMDCT and combined with a hearing-adapted MDCT-based audio coding output signal quantization.

Integer difference values for lossless representation are now not fully encoded in an extension layer, but first encoded at a lower precision. Only the residual values required for an exact representation are sent in a further extension layer. As an alternative, however, a differential spectral value can be represented completely, ie, for example, with 24 bits in the other extension layer, so that no further extension layer is required for decoding this other extension layer. However, this situation leads to higher bitstream sizes, but when the bandwidth of the transmission channel is not an issue, it simplifies at the decoder side, because in the decoder the scalable layers no longer need to be combined, and the decoding is always It is sufficient to adopt an extension layer.

For example, if the lower 8 LSBs, as shown in Figure 9, are not sent at the beginning, scalability between 24 bits and 16 bits can be achieved.

In order to inversely transform the values transmitted with lower precision into the time domain, the transmitted values are preferably extended back to the original area, eg 24 bits, eg by multiplying the transmitted values by ²⁸ . An inverse IntMDCT is applied to the corresponding extended back value.

In precision quantization in the frequency domain according to the invention, it is also advantageous to exploit the redundancy in the LSB. For example if an audio signal has very little energy in the upper frequency domain, this is represented in the IntMDCT spectrum by very small values, e.g. these values are much smaller than what can e.g. be represented by 8 bits (-128, ..... ., 127), which is also reflected in the compressibility of the LSB values of the IntMDCT spectrum. Also, it should be noted that in very small differential spectral values, the bits from MSB to MSB-1 are typically equal to zero; The first 1 in does not exist. In this case, entropy coding is especially suitable for further data compression when the differential spectral values in the second scalable layer contain only zeros.

According to another embodiment of the present invention, sampling rate scalability is preferably used for the second extension layer 82 of FIG. 8a. Sampling rate scalability is achieved by differential spectral values up to the first cutoff frequency contained in the second extension layer, as shown on the right in Figure 9, while in other extension layers, the included frequency is between the first cutoff frequency and the maximum frequency The difference spectrum value between. Of course, further extensions can be implemented to form multiple extension layers across the frequency domain.

In a preferred embodiment of the invention, the second extension layer in FIG. 9 comprises differential spectral values at a frequency of at most 24 kHz, corresponding to a sampling rate of 48 kHz. The third extension layer includes differential spectral values from 24kHz to 48kHz, corresponding to a sampling rate of 96kHz.

It should be further noted that, in the second extension layer and the third extension layer, not all bits in a differential spectrum value need to be coded. In other forms of synthetic extension, the second extension layer may contain bits MSB to MSB-X of a differential spectral value up to a certain cutoff frequency. The third extension layer may then contain the MSB to MSB-X bits of the differential spectral value from the first cutoff frequency to the highest frequency. The fourth extension layer may contain the remaining bits of the differential spectral value up to the cutoff frequency. The last extension layer contains the remaining bits of the difference spectral values at higher frequencies. This concept would result in the table in Figure 9 being divided into four quadrants, each quadrant representing an extension tier.

In the frequency scalability, in a preferred embodiment of the invention, a scalability between 48 kHz and 96 kHz sampling rate is described. The 96 kHz sampled signal is first coded only half in the IntMDCT region of the lossless extension layer and transmitted. If the upper half is not otherwise transmitted, it is assumed to be zero in the decoder. In the inverse IntMDCT (same length as the encoder), a 96kHz signal is generated which contains no energy in the upper frequency domain and thus may be subsampled at 48kHz without loss of quality.

Considering the size of the scalable layer, the differential spectral values in the quadrants with fixed boundaries in Fig. 9 are best quantized on top, because in an extended layer, it is actually only necessary to contain e.g. 16 bits or 8 bits or a maximum of the cutoff frequency or Spectrum value at the cutoff frequency.

An alternative scale "softens" the quadrant boundaries of Figure 9 somewhat. In the case of frequency scalability, this means not applying a so-called "brick-wall low-pass" since the differential spectral value does not change before the cutoff frequency and is zero after the cutoff frequency. Conversely, the differential spectral values can also be filtered by an arbitrary low pass which already somewhat blocks the spectral values below the cutoff frequency, but above the cutoff frequency the differential spectral values still have energy, although the energy is decreasing. In the extension layer thus generated, the spectral values above the cutoff frequency are also included. However, since these spectral values are relatively small, they can be effectively entropy encoded. In this case the highest expansion layer has the difference between the fully differential spectral values and the spectral values contained in the second expansion layer.

Precise quantification can also be softened to some extent. The first expansion layer also contains, for example, spectral values with more than 16 bits, with this difference still being present in the next expansion layer. In general, the second expansion layer has the differential spectral values with less precision, while in the next expansion layer the rest, ie the difference between the full spectral value and the spectral value contained in the second scalable layer, is transmitted. In this way, variable precision reduction is achieved.

The inventive encoding or decoding method is preferably stored on an electronic storage medium, such as a floppy disk, having electronically readable control signals that can cooperate with a programmable computer system to perform the encoding and/or decoding method. In other words, there is a computer program product having computer code stored on a machine-readable carrier for implementing the encoding and/or decoding method when the program product is executed on a computer. The method of the present invention can be realized by a computer program having computer code for executing the method of the present invention when the program is executed in a computer.

Next, as an example of an integer transformation algorithm, the IntMDCT transformation algorithm described in "Audio Coding Based on Integer Transforms" (111 ^th AES convention, New York, 2001) needs to be introduced. IntMDCT is especially favored due to its attractive properties of MDCT algorithms, such as good spectral representation of audio signals, strict sampling and block overlap. A good approximation to MDCT by IntMDCT can use only one transform algorithm in the encoder of FIG. 5, as indicated by arrow 62 of FIG. The important properties of this particular form of integer transformation algorithm are explained on the basis of Figures 1 to 4.

Figure 1 shows an inventively preferred arrangement for processing time-discrete samples representing an audio signal to obtain integer values for which the IntMDCT integer transform algorithm is efficient. The time-discrete samples are windowed and optionally converted to a spectral representation by the apparatus shown in FIG. 1 . The time-discrete samples fed into the input 10 of the device are windowed by a window w of length 2N time-discrete samples to obtain integer windowed samples at the output 12, which are suitable for passing through the transformation device, especially for Means 14 perform an integer DCT to convert to a spectral representation. Integer DCT is used to generate N output values from N input values, in contrast to MDCT function 408 of Figure 4a, which generates only N spectral values from 2N windowed values according to the MDCT equation.

To window the time-discrete samples, first two time-discrete samples are selected in the device 16, which together represent a vector of time-discrete samples. A time-discrete sample selected by means 16 is located in the first quadrant of the window. Another time-discrete sampling is located in the second quadrant of the window, which is explained in more detail on the basis of Figure 3. To the vectors generated by means 16, a matrix rotation of 2x2 dimension is applied, wherein this operation is not performed immediately, but through a number of so-called "lifting matrices".

A boosting matrix has the property that it contains only one element associated with window w and is not equal to 0 or 1.

The factoring of wavelet transforms into lifting steps is described in "Factoring Wavelet Transforms Into Lifting Steps" (Ingrid Daubechies and Wim Sweldens, preprint, Bell Laboratories, Lucent Technologies, 1996). In general, a lifting scheme is a simple relationship between pairs of perfect reconstruction filters with the same low-pass or high-pass filter. Each pair of complementary filters can be factorized into lifting steps. This is especially true for Givens rotations. Consider the case where the multiphase matrix is a Givens rotation. Then, apply the following formula:

$\left(\begin{array}{c}\mathrm{<span\; class="notranslate">\; cos</span>}\mathrm{<span\; class="notranslate">\; \&alpha;</span>}<span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; sin</span>}\mathrm{<span\; class="notranslate">\; \&alpha;</span>}\\ \mathrm{<span\; class="notranslate">\; sin</span>}\mathrm{<span\; class="notranslate">\; \&alpha;</span>}\mathrm{<span\; class="notranslate">\; cos</span>}\mathrm{<span\; class="notranslate">\; \&alpha;</span>}\end{array}\right)<span\; class="notranslate">\; =</span>\left(\begin{array}{cc}\mathrm{<span\; class="notranslate">\; 1</span>}& \frac{<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; cos</span>}\mathrm{<span\; class="notranslate">\; \&alpha;</span>}<span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; 1</span>}<span\; class="notranslate">\; )</span>}{\mathrm{<span\; class="notranslate">\; sin</span>}\mathrm{<span\; class="notranslate">\; \&alpha;</span>}}\\ \mathrm{<span\; class="notranslate">\; 0</span>}& \mathrm{<span\; class="notranslate">\; 1</span>}\end{array}\right)\left(\begin{array}{cc}\mathrm{<span\; class="notranslate">\; 1</span>}& \mathrm{<span\; class="notranslate">\; 0</span>}\\ \mathrm{<span\; class="notranslate">\; sin</span>}\mathrm{<span\; class="notranslate">\; \&alpha;</span>}& \mathrm{<span\; class="notranslate">\; 1</span>}\end{array}\right)\left(\begin{array}{cc}\mathrm{<span\; class="notranslate">\; 1</span>}& \frac{<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; cos</span>}\mathrm{<span\; class="notranslate">\; \&alpha;</span>}<span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; 1</span>}<span\; class="notranslate">\; )</span>}{\mathrm{<span\; class="notranslate">\; sin</span>}\mathrm{<span\; class="notranslate">\; \&alpha;</span>}}\\ \mathrm{<span\; class="notranslate">\; 0</span>}& \mathrm{<span\; class="notranslate">\; 1</span>}\end{array}\right)<span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; 1</span>}<span\; class="notranslate">\; )</span>$

The three lifting matrices to the right of the equal sign each have 1s as the main diagonal elements. In addition, in each lifting matrix, the elements not on the main diagonal are equal to 0, and the elements not on the main diagonal are related to the rotation angle α.

Now the vector is multiplied by the third lifting matrix, that is, multiplying the rightmost lifting matrix in the above formula, to obtain the first result vector, and this process is described with device 18 in FIG. 1 . As shown in Figure 1 by means 20, the first result vector is rounded by an arbitrary rounding function which maps a set of real numbers to a set of integers. The rounded first result vector is obtained at the output of the device 20 . The first result vector after this rounding is sent to the device 22, multiplied by the middle item, that is, multiplied by the second item on the right, to obtain the second result vector, and then rounded by the device 24 to obtain the rounded After the second result vector. The second result vector after rounding is sent to device 26 and multiplied by the lifting matrix on the leftmost side of the above equation, that is, the first item, to obtain the third result vector, and finally the device 28 is still used to round, and finally output Integer windowed samples are obtained at terminal 12 , and if one wants to obtain its spectrum representation, it needs to be processed by device 14 , so that integer spectrum values can be obtained at spectrum output terminal 30 .

The means 14 are preferably implemented as an integer DCT.

According to type 4 (DCT-IV) of length N, the discrete cosine transform is given by:

${\mathrm{<span\; class="notranslate">\; x</span>}}_{\mathrm{<span\; class="notranslate">\; t</span>}}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; m</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; =</span>\sqrt{\frac{\mathrm{<span\; class="notranslate">\; 2</span>}}{\mathrm{<span\; class="notranslate">\; N</span>}}}\underset{\mathrm{<span\; class="notranslate">\; k</span>}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; 0</span>}}{\overset{\mathrm{<span\; class="notranslate">\; N</span>}<span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; 1</span>}}{\mathrm{<span\; class="notranslate">\; \&Sigma;</span>}}}\mathrm{<span\; class="notranslate">\; x</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; k</span>}<span\; class="notranslate">\; )</span>\mathrm{<span\; class="notranslate">\; cos</span>}<span\; class="notranslate">\; (</span>\frac{\mathrm{<span\; class="notranslate">\; \&pi;</span>}}{\mathrm{<span\; class="notranslate">\; 4</span>}\mathrm{<span\; class="notranslate">\; N</span>}}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; 2</span>}\mathrm{<span\; class="notranslate">\; k</span>}<span\; class="notranslate">\; +</span>\mathrm{<span\; class="notranslate">\; 1</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; 2</span>}\mathrm{<span\; class="notranslate">\; m</span>}<span\; class="notranslate">\; +</span>\mathrm{<span\; class="notranslate">\; 1</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; )</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; 2</span>}<span\; class="notranslate">\; )</span>$

The coefficients of DCT-IV form an orthonormal N×N matrix, as described in the publication “Multirate System And Filter Banks” (P.P. Vaidyanathan, Prentice Hall, Englewood Cliffs, 1993), each orthogonal N×N matrix can Decomposed into N(N-1)/2 Givens rotations. Note that further decompositions are also possible.

For the classification of different DCT algorithms, you can refer to the book "Signal Processing With Lapped Transforms" by H.S. Malvar, published by Artech House Press in 1992. In general, DCT algorithms are differentiated according to their basis function type. Whereas the preferred DCT-IV here contains asymmetric basis functions, that is, a 1/4 cosine wave, a 3/4 cosine wave, a 5/4 cosine wave, a 7/4 cosine wave, etc., Such discrete cosine transforms, such as Type II (DCT-II), have axisymmetric and point-symmetric basis functions. The zero-order basis function is a DC component, the first-order basis function is a half cosine wave, the second-order basis function is a whole cosine wave, and so on. Since the DC component is specifically considered in DCT-II, it is used in video coding but not in audio coding, because the DC component is irrelevant in audio coding, unlike in video coding.

Let's explain how the rotation angle α of the Givens rotation is related to the window function.

An MDCT with a window length of 2N can be reduced to a type IV discrete cosine transform of length N. This can be achieved by performing TDAC operations in the time domain and then applying DCT-IV. Due to the 50% overlap, the left half of the window for block t overlaps with the right half of the previous block, ie block t-1. The overlap of two consecutive blocks t and t−1 is preprocessed in the time domain, i.e. before conversion, that is, between input 10 and output 12 of Fig. 1, as follows:

$\left(\begin{array}{c}{\stackrel{<span\; class="notranslate">\; ~</span>}{\mathrm{<span\; class="notranslate">\; x</span>}}}_{\mathrm{<span\; class="notranslate">\; t</span>}}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; k</span>}<span\; class="notranslate">\; )</span>\\ {\stackrel{<span\; class="notranslate">\; ~</span>}{\mathrm{<span\; class="notranslate">\; x</span>}}}_{\mathrm{<span\; class="notranslate">\; t</span>}<span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; 1</span>}}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; N</span>}<span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; 1</span>}<span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; k</span>}<span\; class="notranslate">\; )</span>\end{array}\right)<span\; class="notranslate">\; =</span>\left(\begin{array}{cc}\mathrm{<span\; class="notranslate">\; w</span>}<span\; class="notranslate">\; (</span>\frac{\mathrm{<span\; class="notranslate">\; <figure-callout\; id="2"\; label="N"\; filenames="C0282897400331.png,C0282897400393.png"\; state="\{\{state\}\}">N</figure-callout></span>}}{\mathrm{<span\; class="notranslate">\; 2</span>}}<span\; class="notranslate">\; +</span>\mathrm{<span\; class="notranslate">\; k</span>}<span\; class="notranslate">\; )</span>& <span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; w</span>}<span\; class="notranslate">\; (</span>\frac{\mathrm{<span\; class="notranslate">\; N</span>}}{\mathrm{<span\; class="notranslate">\; 2</span>}}<span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; 1</span>}<span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; k</span>}<span\; class="notranslate">\; )</span>\\ \mathrm{<span\; class="notranslate">\; w</span>}<span\; class="notranslate">\; (</span>\frac{\mathrm{<span\; class="notranslate">\; N</span>}}{\mathrm{<span\; class="notranslate">\; 2</span>}}<span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; 1</span>}<span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; k</span>}<span\; class="notranslate">\; )</span>& \mathrm{<span\; class="notranslate">\; w</span>}<span\; class="notranslate">\; (</span>\frac{\mathrm{<span\; class="notranslate">\; <figure-callout\; id="2"\; label="N"\; filenames="C0282897400331.png,C0282897400393.png"\; state="\{\{state\}\}">N</figure-callout></span>}}{\mathrm{<span\; class="notranslate">\; 2</span>}}<span\; class="notranslate">\; +</span>\mathrm{<span\; class="notranslate">\; k</span>}<span\; class="notranslate">\; )</span>\end{array}\right)\left(\begin{array}{c}{\mathrm{<span\; class="notranslate">\; x</span>}}_{\mathrm{<span\; class="notranslate">\; t</span>}}<span\; class="notranslate">\; (</span>\frac{\mathrm{<span\; class="notranslate">\; <figure-callout\; id="2"\; label="N"\; filenames="C0282897400331.png,C0282897400393.png"\; state="\{\{state\}\}">N</figure-callout></span>}}{\mathrm{<span\; class="notranslate">\; 2</span>}}<span\; class="notranslate">\; +</span>\mathrm{<span\; class="notranslate">\; k</span>}<span\; class="notranslate">\; )</span>\\ {\mathrm{<span\; class="notranslate">\; x</span>}}_{\mathrm{<span\; class="notranslate">\; t</span>}}<span\; class="notranslate">\; (</span>\frac{\mathrm{<span\; class="notranslate">\; N</span>}}{\mathrm{<span\; class="notranslate">\; 2</span>}}<span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; 1</span>}<span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; k</span>}<span\; class="notranslate">\; )</span>\end{array}\right)<span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; 3</span>}<span\; class="notranslate">\; )</span>$

Values marked with a wavy line above the letters are the values at the output terminal 12 of FIG. The value of the coefficient k ranges from 0 to (N/2)-1, and w represents the window function.

From the TDAC condition of the window function w, we can see the following relationship:

$\mathrm{<span\; class="notranslate">\; w</span>}{<span\; class="notranslate">\; (</span>\frac{\mathrm{<span\; class="notranslate">\; <figure-callout\; id="2"\; label="N"\; filenames="C0282897400331.png,C0282897400393.png"\; state="\{\{state\}\}">N</figure-callout></span>}}{\mathrm{<span\; class="notranslate">\; 2</span>}}<span\; class="notranslate">\; +</span>\mathrm{<span\; class="notranslate">\; k</span>}<span\; class="notranslate">\; )</span>}^{\mathrm{<span\; class="notranslate">\; 2</span>}}<span\; class="notranslate">\; +</span>\mathrm{<span\; class="notranslate">\; w</span>}{<span\; class="notranslate">\; (</span>\frac{\mathrm{<span\; class="notranslate">\; N</span>}}{\mathrm{<span\; class="notranslate">\; 2</span>}}<span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; 1</span>}<span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; k</span>}<span\; class="notranslate">\; )</span>}^{\mathrm{<span\; class="notranslate">\; 2</span>}}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; 1</span>}<span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; 4</span>}<span\; class="notranslate">\; )</span>$

For certain angles α _k , k=0, 1, . . . , (N/2)-1, this preprocessing in the time domain can be written as a Givens rotation, which was explained earlier.

The relationship between the givens rotation angle α and the window function w is as follows:

α＝arctan[w(N/2-1-k)/w(N/2+k)4 (5)

It should be noted that any window function w can be applied as long as it meets the TDAC conditions.

Below, based on Figure 2, a cascaded encoder and decoder is described. Time-discrete samples x(0) to x(2N-1) that are "windowed" together by a window are first selected by means 16 in Fig. 1 such that samples x(0) and x(N-1), i.e. from The samples from the first quarter of the window and the samples from the second quarter of the window are selected to form a vector at the output of the means 16 . Crossed arrows indicate lifting multiplication and successive rounding of means 18, 20 or 22, 24 or 26, 28 to obtain integer windowed samples at the input of the DCT-IV block.

As described above, when the first vector is processed, the second vector is also selected from samples x(N/2-1) and x(N/2), that is, another one from the first vector of the window The samples from the first quarter of the window and the samples from the second quarter of the window are again processed by the algorithm described in Figure 1. All other pairs of samples from the first and second quarters of the window are treated similarly. The third and fourth quarters of the first window are treated similarly. As shown in Figure 2, there are N "windowed" integer samples at output 12, which are sent to the DCT-IV transform. In particular, "windowed" integer samples of the second and third quarters are sent to the DCT. The "windowed" integer samples from the first quarter of the window are fed into the previous DCT-IV along with the "windowed" integer samples from the fourth quarter of the previous window deal with. Similarly, the "windowed" integer samples from the fourth quarter of Figure 2 are sent to DCT-IV along with the "windowed" integer samples from the first quarter of the following window transform. The central integer DCT-IV transform 32 shown in Figure 2 provides N integer spectral values y(0) to y(N-1). Since the windowing process and the transformation process provide integer output values, these integer spectral values can be directly entropy encoded without inverse quantization.

In the right half of Fig. 2 a decoder is depicted. This decoder contains an inverse transform and "inverse windowing", which works in the opposite way to the encoder. It is known that for the inverse transformation of DCT-IV, the inverse DCT-IV as shown in FIG. 2 needs to be used. As shown in FIG. 2, to generate time-discrete audio samples x(0) to x(2N-1) from the integer "windowed" samples at the output of the device 34 again in the previous and next transformations, the previous The values of the first and subsequent transformations are inversely processed to the output values of the decoder DCT-IV34.

Operation at the output is done by a reverse Givens rotation, ie blocks 26, 28 or 22, 24 or 18, 20 are passed in an opposite direction. The second boost matrix based on Equation 1 can be described in more detail. When (in the encoder) the second result vector is formed by multiplying (means 22) the rounded first result vector with the second lifting matrix, the following results:

The values x, y on the right side of Equation 6 are integers. However this does not apply to the value xsinα. Here, we need to introduce the rounding function r, which is expressed by the following equation:

This operation performs the function of means 24 .

The reverse mapping in the decoder can be defined as follows:

Due to the minus sign preceding the rounding operation, it is clear that the integer approximation of the lifting step can be reversed without introducing errors. Application of an approximation to any of these three lifting steps results in an integer approximation of the Givens rotation. The rounding rotation (in the encoder) can be reversed (in the decoder) without introducing errors, i.e. the reverse rounding and lifting steps are passed in reverse order, that is, the algorithm of Fig. It is executed bottom-up.

If the rounding function r is point-symmetric, the rotation of the reverse rounding is the same as the rounding rotation of the angle -α, as follows:

$\left(\begin{array}{cc}\mathrm{<span\; class="notranslate">\; cos</span>}\mathrm{<span\; class="notranslate">\; \&alpha;</span>}& \mathrm{<span\; class="notranslate">\; sin</span>}\mathrm{<span\; class="notranslate">\; \&alpha;</span>}\\ <span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; sin</span>}\mathrm{<span\; class="notranslate">\; \&alpha;</span>}& \mathrm{<span\; class="notranslate">\; cos</span>}\mathrm{<span\; class="notranslate">\; \&alpha;</span>}\end{array}\right)<span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; 9</span>}<span\; class="notranslate">\; )</span>$

The lifting matrix for the decoder, ie for the inverse Givens rotation, is in this case directly derived from equation (1), simply replacing the "sinα" term with "-sinα".

In the following, on the basis of FIG. 3 , the decomposition of a general MDCT with overlapping windows 40 to 60 is mentioned again. Windows 40 to 60 each overlap by 50%. For each window, first a Givens rotation within the first and second quarter of the window, or within the third and fourth quarter of the window is performed, as indicated by arrow 48 . Then, the rotated values, that is, windowed integer samples, are fed into an N by N DCT such that the second and third quarters of one window or the fourth and third quarters of the next window A quarter section is converted into a spectral representation together through the DCT-IV algorithm.

Therefore, the usual Givens rotation is decomposed into lifting matrices, which are executed sequentially, where a rounding step is introduced after each lifting matrix multiplication, so that floating-point numbers are rounded immediately after they are generated, so that in each Before multiplying the result vector with the lifting matrix, the result vector has only integers.

Output values are always integers, and it is preferable to use integer input values as well. This does not represent a limitation on the invention, since each exemplary PCM sample, since they are stored on a CD, is an integer value whose range of values varies according to the width of the bit, that is, according to the time discrete number Whether the input value is sixteen or twenty one to change. However, as explained, the entire process can be reversed by performing the reverse rotation in the reverse order. Therefore, there exists an integer approximation of the MDCT with perfect reconstruction, the lossless transformation.

The conversions shown provide integer output values rather than floating point values. It provides a perfect reconstruction, so no errors are introduced when performing a forward transformation followed by a backward transformation. This transform, according to a preferred embodiment of the present invention, is a replacement for the Modified Discrete Cosine Transform. However, other transformation methods can also be performed in an integer fashion, as long as decomposition into rotations and decomposition of rotations into lifting steps is possible.

Integer MDCT has most of the good properties of MDCT. It has an overlapping structure, whereby better frequency selectivity is obtained than in non-overlapping block switching. Due to the TDAC function, windowing before conversion already takes this function into account, maintaining strict sampling such that all spectral values representing an audio signal are equal to the total number of input samples.

In the preferred integer transform described, the noise is enhanced compared to an ordinary MDCT only in spectral regions with little signal strength, and this noise-enhanced Doesn't qualify as a significant signal strength on its own. For this, integer processing facilitates an efficient hardware implementation because only the multiplication step is used, and multiplication can be easily decomposed into shift and add steps, both of which are easy and fast to implement in hardware. Of course, software implementation is also feasible.

Integer transforms provide a nice spectral representation of the audio signal and still remain in the integer region. When applied to the speech portion of an audio signal, it results in a good energy concentration. With this approach, an efficient lossless coding scheme can be implemented with a simple cascaded windowing/transformation as shown in Figure 1. In particular, stack encoding using escape values, as used in MPEG AAC, is popular. It is best to reduce all values by using a specific power of two until they satisfy a desired code table, and then encode the least significant bits ignored. Compared to the alternative method of using a larger code table, this method is better in terms of the memory consumption required to store the code table. It is also possible to obtain an almost lossless encoder by simply omitting only some of the least significant bits.

Especially for speech signals, entropy coding of integer spectral values enables high coding gains. For the transient part of the signal, the coding gain is low, i.e. due to the flat spectrum of the transient signal, that is to say, due to a small fraction of spectral values equal or almost equal to zero. As described in J. Herre, JD Johnston "Enhancing the Performance of Perceptual AudioCoders by Using Temporal Noise Shaping (TNS)" 101 ^st AESConvention, Los Angeles, 1996, preprint 4384, however this flatness may be obtained by using used for linear prediction. An alternative is to use open-loop forecasting, and an alternative is to use closed-loop forecasting. The first scheme, the open-loop predictor, is called TNS. Post-predictive quantization causes the resulting quantization noise to adapt to the temporal structure of the audio signal, thus preventing forward echoes in psychoacoustic audio coders. For lossless audio coding, the second scheme is more suitable, that is, a closed-loop predictor, because closed-loop prediction allows an accurate reconstruction of the input signal. When this technique is applied to the generated spectrum, a rounding step must be performed after each stage of the prediction filter to keep it in the integer region. By using the inverse filter and the same rounding function, the original spectrum can be generated exactly.

In order to exploit the redundancy between the two channels in data reduction, mid-edge coding can also be used in lossless fashion when using a rounded rotation of α/4 angle. The benefit of this rounding rotation is that it preserves energy compared to methods that calculate the sum and difference between the left and right channels of a stereo signal. Using so-called combined stereo coding it can be switched on or off for each band, as is done in standard MPEG AAC. In order to be able to more flexibly reduce the redundancy between the two channels, other rotation angles are also conceivable.

Claims (31)

1. be used for the device of time-discrete coding audio signal with the voice data after obtaining encoding comprised:

Being used for applied mental acoustic model (54) provides the device (52) of the quantize block of the time-discrete sound signal that is quantized;

Be used for this quantize block of inverse quantization, and the spectrum value of inverse quantization is rounded, with the device that rounds piece (58) of the spectrum value of the inverse quantization that obtains to be rounded;

Be used to utilize the integer transform algorithm to generate the device (56) of the integer piece of integer spectrum value, described integer transform algorithm is used for generating from integer time discrete sampling module the integer piece of spectrum value;

Be used for coupling apparatus (58), to obtain to have the difference block of difference spectrum value according to the difference formation difference block that rounds spectrum value between piece and the integer piece; And

Be used to handle the device (60) of quantize block and difference block, comprise the voice data of coding of the information of the information of quantize block and difference block with generation.

2. device as claimed in claim 1 wherein is used to the device (52) that provides by a MDCT, produces the MDCT module of a MDCT spectrum value from the time block of time audio signal value, and

Quantize this MDCT module with psychoacoustic model, comprise the quantize block of the MDCT spectrum value of quantification with generation.

3. device as claimed in claim 2, the device (56) that wherein is used to produce the integer piece is carried out an IntMDCT on time block, comprise the integer piece of IntMDCT spectrum value with generation.

4. as the described device of the arbitrary claim in front, the device (52) that wherein is used to provide calculates quantize block with the floating-point transfer algorithm.

5. device as claimed in claim 1, the device (52) that wherein is used to provide use the integer piece that produces by the device (56) that is used to generate to calculate quantize block.

6. device as claimed in claim 1,

The device (60) that wherein is used to handle carries out entropy coding (60a) to quantize block, to obtain the quantize block of entropy coding;

Carry out entropy coding (60b) to rounding piece, to obtain the piece that rounds of entropy coding; And

The quantize block of entropy coding is converted to first extension layer of the extended data stream of presentation code voice data, and entropy coding is rounded second extension layer that piece is converted to extended data stream.

7. device as claimed in claim 6,

The device (60) that wherein is used to handle uses in a plurality of code tables also according to the spectrum value that quantizes, and quantize block is carried out entropy coding, and

The device (60) that wherein is used for handling is selected in a plurality of code tables also according to the attribute that quantizes available quantizer, is used for difference block is carried out the quantize block of entropy coding with generation.

8. device as claimed in claim 1,

The device (52) that the provides attribute according to sound signal wherein is provided, selects in a plurality of windows, carry out windowization with time block to audio signal value; And

The device (56) that wherein is used to generate is selected for the integer transfer algorithm carries out identical window.

9. device as claimed in claim 1,

The device that wherein is used to generate has used an integer transfer algorithm, comprising:

Corresponding to the window (w) of 2N time-discrete sampling time-discrete sampling carry out windowization with length, so that the time discrete sampling of windowization to be provided, by producing the conversion of N output valve from N input value, with time-discrete unscented transformation is frequency spectrum designation, and wherein the window process comprises following substep:

Select (16) time-discrete samplings from four of window/part, and select a time-discrete sampling, to obtain the vector of time discrete sampling from the other four/part of this window;

Use a rotation square formation, its dimension and vector are complementary to the dimension of vector, and wherein rotation matrix can be represented with a plurality of lifting matrixes, and one of them promotes matrix and only comprises an element according to window (w), and be not equal to 1 or 0, wherein use substep and comprise following substep:

Multiply each other (18) with promoting matrix and vector, obtain first result vector;

Round the component of first result vector, first result vector that obtains rounding with the bracket function (r) that real number is mapped as integer; And

Execution subsequently promotes matrix multiple (22) with another one and rounds the step of (24), finish up to all lifting matrixes are all processed, obtain a rotating vector, it comprises from the integer window sampling of four of window/part with from the integer window sampling of the other four/part of this window, and

Execution is sampled for all time discretes of the remaining four/part of window and is carried out the step of windowization, obtains 2N filtered round values; And

For second and the 3rd four/a part of filtered integer sample values by window, by integer DCT, be the integer unscented transformation (14) of a N windowization frequency spectrum designation, obtain N integer spectrum value.

10. device as claimed in claim 1,

The device (52) that quantize block wherein is provided is realized prediction for spectrum value on the frequency with a predictive filter, at quantization step (52b) before to obtain being illustrated in the prediction residual spectrum value of the quantize block after quantizing;

A prediction unit wherein also is provided, and it is predicted on frequency the integer spectrum value of integer piece, wherein also provides to round device, rounds with the prediction residual spectrum value that the integer spectrum value that rounds piece owing to expression is obtained.

11. device as claimed in claim 1,

Wherein the time discrete sound signal comprises at least two channels:

The device (52) that provides wherein is provided comes implementation center/edge coding with the spectrum value of time discrete tone signal, after the quantification of center/edge spectrum value, obtaining quantize block, and

The device (56) that wherein is used to generate the integer piece also is provided by the center/edge coding corresponding to the center/edge coding of the device that is used to provide (52).

12. device as claimed in claim 1, the device (60) that wherein is used to handle produces a MPEG-2ACC data stream, has wherein introduced the auxiliary data supplementary that is used for the integer transform algorithm in a zone.

13. device as claimed in claim 1,

The voice data that device (60) the output process that wherein is used to handle is encoded is as the data stream that has a plurality of extension layers.

14. device as claimed in claim 13,

Wherein be used for the device (60) handled and inserted information, and in second extension layer (82), inserted information about difference block about quantize block at first extension layer (81).

15. device as claimed in claim 13,

Wherein be used for the device (60) handled and inserted information, and in the second and the 3rd extension layer, inserted information at least about difference block about quantize block at first extension layer.

16. device as claimed in claim 15,

Wherein in second extension layer, comprise the difference spectrum value that has the precision that is reduced, but in high one-level or more senior extension layer, comprise the residual fraction of difference spectrum value.

17. device as claimed in claim 15,

Wherein the information about difference block comprises binary coding difference spectrum value;

Second extension layer that wherein is used for the difference spectrum value comprises a plurality of bits from the highest significant position (MSB) of difference spectrum value to time high significance bit (MSB-x); And

Wherein comprise a plurality of bits from inferior high significance bit (MSB-x-1) to least significant bit (LSB) (LSB) at the 3rd extension layer.

18. device as claimed in claim 17,

Wherein time discrete sound signal width is that the sampled form of 24 bits is represented, and

The device (60) that wherein is used for handling inserts 16 bits of the higher significance bit of difference spectrum value at second extension layer, in the 3rd extension layer, insert remaining 8 bits of difference spectrum value, demoder has reached CD Quality with second extension layer like this, if wherein adopt the 3rd extension layer, demoder just can reach the tonequality of studio.

19. device as claimed in claim 15,

The device (60) that wherein is used for handling has inserted to small part difference spectrum value at second extension layer, and the expression low-pass filter signal has inserted difference spectrum value and the initial difference between the difference spectrum value in second extension layer in the another one extension layer.

20. device as claimed in claim 15,

The device (60) that wherein is used for handling has inserted the difference spectrum value that is up to certain cutoff frequency to small part at second extension layer, and has inserted in the 3rd extension layer to the difference spectrum value of small part from certain cutoff frequency to higher frequency.

21. time-discrete coding audio signal to obtain the method for coding audio data, being comprised:

Applied mental acoustic model (54) provides the quantize block of spectrum value of the time discrete sound signal of (52) quantifications;

Inverse quantization (58) quantize block, and round the spectrum value of this inverse quantization, to obtain rounding the piece that rounds of inverse quantization spectrum value;

Use an integer transform algorithm to produce the integer piece of (56) integer spectrum values, this integer transform algorithm produces the integer piece of spectrum value from integer time discrete sampling block;

According at the spectral difference score value that rounds between piece and the integer piece, form (58) difference block, to obtain having the difference block of difference spectrum value; And

Handle (60) quantize block and difference block, comprise about the information of quantize block with about the coding audio data of the information of difference block with generation.

22. be used for device that the voice data of having encoded is decoded, this voice data of having encoded produces from a time discrete sound signal, the quantize block of spectrum value of the time discrete sound signal of (52) quantifications is provided by applied mental acoustic model (54), by inverse quantization (58) quantize block and round the spectrum value of inverse quantization, inverse quantization spectrum value after obtaining to round round piece, by using the integer transform algorithm that produces the integer piece of spectrum value from the data block of integer time discrete sampling, produce the integer piece of (56) integer spectrum value, form (58) difference block by basis in the difference that rounds the spectrum value between piece and the integer piece, to obtain the difference block of difference spectrum value, comprising:

Be used to handle the device (70) of coding audio data, obtain a quantize block and difference block;

Be used for inverse quantization and round the device (74) of this quantize block, with the quantize block of the inverse quantization that obtains an integer;

Be used for obtaining a binding modules with the device (78) of spectrum value mode in conjunction with integer quantisation piece and difference block;

Use this binding modules and the integer transform algorithm opposite, produce the device (82) of the time representation of a time discrete sound signal with the integer transform algorithm.

23. the decoding device described in claim 22,

Wherein coding audio data is extendible, and comprises a plurality of extension layers;

The device (70) that wherein is used for handling this coding audio data is determined quantize block from coding audio data, as first extension layer, and determines difference block from coding audio data, as second extension layer.

24. device as claimed in claim 22,

Wherein the information about difference block comprises binary code differential spectrum value,

Wherein coding audio data is extendible, and comprises a plurality of extension layers,

The device (70) that wherein is used for handling this coding audio data determines quantize block from coding audio data, as first extension layer, and extracts representing of difference spectrum value with the precision that has reduced, as second extension layer.

25. device as claimed in claim 24,

The device (70) that wherein is used to handle this coding audio data extracts a plurality of bits from highest significant position to inferior high significance bit as second extension layer, and wherein time high significance bit is higher than the least significant bit (LSB) in the difference spectrum value, and

The device (82) of time representation that is used to generate the discrete tone signal produced the disappearance bit of difference spectrum value with comprehensive method before using the integer transform algorithm.

26. device as claimed in claim 25,

Wherein device (82) is carried out the expansion of second extension layer for comprehensive generation, wherein in expansion, use a scale factor, it equals 2n, and wherein n is the inferior high number of significant bits that is not included in second extension layer, perhaps uses dither algorithm for comprehensive the generation.

27. device as claimed in claim 22,

Wherein coding audio data is extendible, and comprises a plurality of extension layers, and

The device (70) that is used for handling this coding audio data is determined quantize block from coding audio data, as first extension layer, and the difference spectrum value of definite low-pass filtering, as second extension layer.

28. device as claimed in claim 22,

Wherein coding audio data is extendible, and comprises a plurality of extension layers,

The device (70) that wherein is used for handling this coding audio data is determined quantize block from coding audio data, as first extension layer; Determine to be up to the difference spectrum value of first cutoff frequency, as second extension layer, wherein first cutoff frequency is littler than the maximum frequency of the difference spectrum value that can produce in scrambler.

29. device as claimed in claim 28,

Wherein be used for the device (82) that the rise time represents the input value of the integer transform algorithm of total length is made as predetermined value, these are worth on the cutoff frequency of second extension layer; And by by corresponding to the maximum frequency of difference spectrum value and the factor selected by the ratio of frequency, just use after the reverse integer transform algorithm, reduce the time representation of discrete tone signal sample time, wherein difference spectrum value maximum frequency can be produced by scrambler.

30. device as claimed in claim 29,

Wherein the predetermined value of all input values on cutoff frequency is zero.

31. to the method that the voice data of having encoded is decoded, the voice data of wherein having encoded by provide, inverse quantization, generation, formation and processing, from time-discrete sound signal, produce, this method comprises:

Handle (70) coding audio data, to obtain a quantize block and a difference block;

Inverse quantization (74) quantize block also rounds, to obtain the quantize block of an integer inverse quantization;

Mode combination (78) this integer quantisation piece and difference block with spectrum value obtains a binding modules; And

Use this binding modules, and the use integer transform algorithm opposite with the integer transform algorithm, the time representation that produces (82) time discrete sound signal.

Images