JP2004234021A - Device and method for decoding audio signal or bit stream - Google Patents

Abstract

<P>PROBLEM TO BE SOLVED: To provide concepts for decoding of an audio signal or bit stream which is efficient although errors are permitted. <P>SOLUTION: Disclosed is a device or method for decoding the bit stream, which includes argots which are derived from a code table and have length different from one another; and information regarding the length of the longest argot which is actually generated is included as side information. The decoding device includes a decoding unit which decodes the bit stream by using a code table. THis decoding unit can detect whether an argot extracted from the bit stream is longer than the longest argot and, therefore, whether the argot is wrong and also take a countermeasure when such a wrong argot is detected. <P>COPYRIGHT: (C)2004,JPO&NCIPI

Description

  The present invention relates to an apparatus and a method for decoding an audio signal or a bit stream, which can perform error-tolerant entropy decoding, in particular, error-tolerant Huffman decoding.

  Recent methods of audio encoding or decoding are, for example, based on the MPEG Layer 3 standard, but it is possible to compress the data rate of the audio signal by, for example, a factor of 12, without any noticeable degradation in signal quality. It is possible. To enable such high data rate compression, the audio signal is sampled, resulting in a series of discrete time samples. As is known in the art, to obtain a windowed set of temporal samples, the series of discrete time samples is windowed using a suitable windowing function. A windowed set of temporal samples is transformed into the frequency domain by a filter bank, a modified discrete cosine transform (MDCT) or other suitable method, so that the audio signal in the frequency domain, that is, the Obtain a spectral value that represents the temporal portion of the discrete time sample group. Usually, a group of times that overlap by 50% is generated and transformed into the frequency domain by the MCDT. Due to the nature of MDCT, for example, 1024 discrete time samples always result in 1024 spectral values.

  It is known that the sensitivity of the human ear depends on the temporal spectrum of the audio signal itself. This is reflected in the so-called psychological hearing model. Using this model, it was possible to calculate a masking threshold by temporal spectrum. Masking is that a particular tone or portion of the spectrum is determined to be inaudible, for example, when its adjacent spectral regions have relatively high energy. This masking phenomenon is used to quantize the converted spectral value as coarsely as possible. Therefore, the objective is to encode or quantize as few bits as possible, so as to avoid interference with the decoded audio signal in the audible range. The disturbance caused by the quantization, ie the quantization noise, should be below the masking threshold and should therefore be outside the audible range. According to known methods, the spectral values are therefore pre-divided into so-called magnification bands, which reflect the frequency of the human hearing. A spectrum value of a certain magnification group is multiplied by a certain magnification in order to adjust the spectrum value of the whole magnification band. Then, the magnification band adjusted by the magnification is quantized to generate a quantized spectrum value. Of course, it is not important to group into magnification bands. However, this procedure is used in the MPEG Layer 3 standard and the MPEG-2 AAC (AAC is Advanced Audio Coding) standard.

  A very important element of data compression is the entropy coding of the quantized spectral values. Usually, Huffman coding is used for this. Huffman coding involves coding that varies in length. That is, the length of the code word for the value to be encoded is based on the likelihood that this value will occur. Logically, the most probable symbol is assigned the shortest code, i.e., the code word, so that a very good reduction of extras can be achieved with Huffman coding. One example of a well-known length variation encoding is a Morse code.

  In audio encoding, Huffman codes are used to encode quantized spectral values. For example, modern audio coders that operate on the MPEG-2 AAC standard use a plurality of different Huffman code tables, which are assigned to the spectrum based on a particular standard on a part-by-part basis to encode the quantized spectral values. use. Here, two or four spectral values are always integrally encoded into one codeword.

  One of the differences between the MPEG-2 AAC based method and the MPEG layer 3 method is that different magnification bands, ie, different spectral values, are grouped into any number of spectral parts. In AAC, one spectral portion contains at least four spectral values, preferably more. The entire frequency domain of the spectral values is thus divided into several contiguous parts, each part representing a frequency band, but consequently all these parts span the unconverted spectral values Covers all frequency ranges.

  To maximize the reduction of extra parts, a so-called Huffman table, one of a plurality of such Huffman tables, is assigned to each part, as in the method with the MPEG layer 3. The bit stream of the AAC method usually contains 1024 spectral values, and the Huffman code words for these spectral values are arranged in order of increasing frequency. The table information used in each frequency part is sent as side information. This situation is shown in FIG.

  In the case shown by way of example in FIG. 2, the bitstream contains ten Huffman code words. If one codeword is always formed from one spectral value, ten spectral values can be encoded here. However, FIG. 2 shows a portion of an encoded bitstream containing 20 or 40 spectral values, since usually two or four spectral values are always integrally coded into one codeword. It will show. If each Huffman codeword contains two spectral values, the codeword designated by number 1 represents the first two spectral values. The length of the code word is relatively short, meaning that the first two spectral values, the two lowest frequency coefficients, occur relatively frequently. On the other hand, the codeword denoted by number 2 is relatively long, and the third and fourth spectral coefficients occur relatively rarely in this encoded audio signal, and therefore these spectral coefficients are relatively large bits. It is encoded by a number. As shown in FIG. 2, the code words 3, 4, and 5 represent the spectral coefficients 5, 6, 7, 8, 9, and 10. Since these code words are relatively short, these code words are used. Occurs relatively frequently. The same idea can be applied to the code words indicated by the numbers 6 to 10.

  As mentioned above, as is evident from FIG. 2, the Huffman code words for the encoded spectral values, in the case of a bit stream generated by a known encoding device, have a higher frequency in the bit stream. They are arranged in a row in the order they go.

  The biggest disadvantage of Huffman coding in the channel where the error occurs is that it conveys the error. For example, it may not be meaningless that the length of the code word 2 may be changed when the code word indicated by number 2 in FIG. 2 is transmitted. Thus, the correct length may be different. In the case of FIG. 2, if the length of the code word 2 has been changed due to interference, the decoder can no longer determine where the code word 3-10 starts, ie almost all parts of the audio signal are affected. Will be. In this way, even all other code words following the disturbed code words will not be correctly decoded. This is because it is not clear where these codewords begin, and this error causes the wrong starting point to be chosen.

  In order to solve the problem of error transmission, EP 0 612 156 proposes that the start of a codeword can be more easily determined even without complete decoding or in case of incorrect transmission. In order to be able to do so, it has been proposed that some of the code words having different lengths are arranged in a raster, and other code words are arranged in the remaining gaps.

  The parameters that determine the efficiency of this known method are how the raster is actually defined, ie how many raster points are needed, the raster distance between the raster points, etc. However, EP 0 612 156 does not go beyond the general suggestion that rasters should be used to suppress the transmission of errors, but not any more to achieve error-tolerant and efficient coding. There is no detailed explanation on how to build a raster efficiently.

EP-A-0 717 503 discloses a digital coding and decoding method for transforming discrete time samples of a music signal into the frequency domain and quantizing and entropy coding the resulting spectral values. This entropy coding produces a certain number of code words of different lengths, some of which are raster arranged and others inserted in the remaining space of the raster.
EP-A-0492537 relates to a video and audio information recording device in which the information is divided into small pixel groups, each pixel group being converted into orthogonal components by orthogonal transformation means. Then, the orthogonal components are encoded using codes having secret words of different lengths. Some of the encoded code words are written to the first memory. If the codeword has more bits than can be written to the first memory, the remaining bits of the codeword are written to another memory.

  It is an object of the present invention to provide a concept for efficient decoding of an audio signal or bit stream despite error tolerance.

  The object is achieved by a bit stream decoding device according to claim 1 and a bit stream decoding method according to claim 6.

  The present invention is based on the discovery that already proposed rasters must be formed or used to enable efficient encoding / decoding as well as error-tolerant encoding / decoding. . Most importantly, the codewords obtained by entropy coding in the form of Huffman coding are of essentially different lengths. The largest coding results appear when the most frequently occurring values are given the shortest code words. On the other hand, values that occur relatively rarely become statistically optimal values for the amount of data even if a long code word is given. The code words obtained by Huffman coding are of essentially different lengths.

  According to a first aspect of the invention, a so-called priority codeword is located at a raster point, so that even if there is an error in the bitstream, the decoder will definitely recognize the beginning of each priority codeword through the raster. Can be. Priority code words are code words that are psychoacoustically important. This means that the spectral values encoded by the so-called priority codewords substantially contribute to the audibility of the decoded audio signal. If the audio signal contains a high percentage of speech, the preferred codeword is a codeword that represents a low spectral value. This is because, in this case, the important spectral information is in the lower region of the spectrum. If the audio signal has a group of tones in the middle region of the spectrum, the preferred codeword is the codeword assigned to the corresponding spectral value in the middle region of the frequency domain. This is because these are spectrally important spectral values. Psychoacoustically important spectral values are also those whose magnitude or signal energy is large compared to other spectral values in the spectrum. On the other hand, psychoacoustically insignificant words, so-called non-prioritized words, fill this raster. Thus, these codewords are not located with the raster point, but are "fitted" into the remaining space after the preferred codeword is placed at the raster point.

  Therefore, according to the first feature of the present invention, the preferred secret words, that is, the secret words given to the psychoacoustically important spectral values are rasterized so that the starting points of these preferred secret words coincide with the raster points. Are arranged.

  According to a second aspect of the invention, the spectral values are grouped into spectral parts, and a different code table is provided for each spectral part. The assignment of a code table to a spectral part is made based on statistical considerations of the signal, for example, which code table is most suitable for coding a spectral part. The assignment of code tables to spectral parts is already well known in the art.

  A raster will be used where several groups of raster points are arranged at equal distances. The distance between raster points within a raster point group is based on a code table used to encode one spectral portion. In other alternative parts of the spectrum, different code tables are used for optimal data compression. The other code table is assigned to another group of equidistantly arranged raster points, and the distance between the two raster points in the raster point group is based on its associated code table. The distance between two raster points in different raster point groups can be determined in at least three ways.

  First, the maximum length of a code word in one code table is determined. Since the distance between the two raster points in the raster point group given this code table is equal to or greater than the maximum length of the code word in that code table, the raster contains the longest code word in this code table. There is space for The distance between two raster points in another raster point group associated with another code table is determined in a similar manner based on the maximum length of the code word in this other code table.

  The second method described below also contributes to increasing the number of raster points. Due to the intrinsic properties of the Huffman code, less frequently occurring code words are longer than more frequently occurring code words. If the distance between raster points is set to be greater than or equal to the maximum length of a word in a table, the word inserted into the raster is usually shorter than the distance between the raster points. Therefore, the distance between raster points can be set shorter than the maximum length of a codeword in a certain table. If a codeword that does not fit into the raster appears during encoding, the remaining part that does not fit in the raster is inserted into the bitstream at some other suitable location outside the raster array. As a result, this fragmented codeword is no longer effectively protected from error transmission. This is very rare and is acceptable for the benefit of increasing the number of raster points.

  A third method of determining the distance between different raster points is to consider the maximum length of the codeword in the bitstream that actually occurs in the encoded spectral portion, rather than the maximum length of the codeword in the code table. .

  According to a third aspect of the invention, instead of an arrangement of the code words in the bitstream in an order that is essentially linearly higher in frequency, an arrangement is used in which the code words are distributed over the frequency domain, This is a method known as "scrambling". This has the advantage that the so-called "sudden errors" do not lead to false compounding of the whole frequency band, but only to the extent that small disturbances are seen in several different frequency ranges.

  According to a fourth aspect of the invention, instead of an arrangement of cryptographic words in an order of essentially linearly increasing frequency, for example, every nth (eg every second, every third, or every fourth) An arrangement method in which only the code words are arranged in the raster can be used. In this way, if the number of possible raster points is smaller than the number of preferential passwords, it is possible to make the spectral region using the preferential passwords as large as possible, ie prevent error transmission.

  Furthermore, priority is given to determining the preferred codeword in such a way as to achieve efficient operation. Preferably, this means dismissing the hypothesis that the psychoacoustically important codeword, ie the preferred codeword, encodes low frequency spectral values. This is often the case, but not always.

  Usually, the preferred codewords encode spectral lines of psychoacoustic importance, which are spectral values that usually have high energy. Spectral lines with high energy are not caused by errors.

  According to the invention, an indicator that has already been implicitly determined is used. This indicator is based on the code table used. In the AAC standard, for example, there are 11 code tables with different absolute value ranges. Code table 1 includes, for example, spectral values having absolute values of −1 to +1, and code table 11 encodes spectral values of −8191 to +8191. The larger the code table number, the larger the range in which it can be encoded. This means that a low numbered code table represents only relatively small values and therefore only causes a relatively small error, while a high numbered code table represents a relatively large range and causes a relatively large error. Means

  If an error occurs in a small numbered code table, it will not be heard, and the result of the wrong spectral line will not be very different from the original correct spectral line. However, if an error occurs in the highest numbered code table, this error can in principle be some of the absolute values of this code table. The spectral line encoded with the code table with the highest number has a small value and is combined in the decoder as the spectral line with the highest absolute value by this code table, for example due to an error that has occurred during transmission. If so, this incorrect spectral line will be audible.

With regard to error tolerance, the most important code table is therefore the highest numbered code table (code table 11 in the AAC standard). This is because this code table misses the value between −2 ¹³ +1 (−8191) and +2 ¹³ −1 (+8191).

  According to a further aspect of the invention, a short window is used for the AAC standard transmission. With a short window, the frequency resolution is reduced for higher temporal resolution. The preferred codeword is determined such that the psychoacoustically significant spectral values, i.e. low frequency spectral values or spectral values from the higher numbered code table, are reliably located on the raster points. Multiplier band interleaving, which is a feature of the AAC standard, is defeated for this purpose.

  Preferred embodiments of the present invention will be described with reference to the accompanying drawings.

  For the purpose of explaining the present invention, priority code words are shown by hatching in FIG. FIG. 2 shows an arrangement of known lengths of codewords, in order of linearly increasing frequency. In FIG. 2, the preferred secret words are secret words 1 to 5. As mentioned above, if the audio signal contains, for example, a high proportion of speech, or contains many low-frequency sounds, the secret word given to the low-frequency spectral value is the priority secret word. In FIG. 2, the code words 6 to 10 relate to high-frequency spectral values, which contribute to the overall impression of the decoded signal, but do not significantly affect the auditory perception, Not very psychoacoustical.

  FIG. 1 shows a bit stream having raster points 10-18. In FIG. 1, the distance between the raster points 10 and 12 is D1, and the distance between the raster points 14 and 16 is D2.

  For the description of the first aspect of the present invention, only the bit stream between raster points 10 to 14 will be considered. Priority code words 1 and 2 are important spectral portions located in the low frequency range in the example shown in FIG. 2, but are arranged in a raster so that they are not subject to error transmission during decoding. ing. The non-priority code words, which are not hatched in FIGS. 1 and 2, are arranged to fill the raster after the priority code word. Since the length of the Huffman codeword is known from the codeword itself, there is no need to put non-priority codewords together in the raster. The decoder can determine if what is read is just a part of the code word. In this case, the decoder automatically adds a certain number of bits to the first part of the codeword, following the preferred codeword after the next raster point. Therefore, the first part of the non-priority codeword is rasterized so that each of the non-priority codewords 7, 8, 9 is split into two in the bitstream, ie, 7a, 7b, 8a, 8b, 9a, 9b. Can be inserted into the first empty space, and the rest can be inserted elsewhere.

  As already explained, the second part of the bit stream of FIG. 1 illustrates the second aspect of the invention. Unless the raster distance D1 is changed to a small raster distance D2, a raster having a distance D1 that can contain any of the preferred code words 1 to 5 has, so to speak, not enough non-preferred code words to fill the rest of the raster. Such a long bit stream can result. Therefore, only a number of priority code words to be inserted into the bit stream are extracted from the audio signal so as not to leave a fundamentally empty space, and the bit stream is not unnecessarily extended.

  The second feature of the present invention will be described in detail with reference to FIG. In the case of an encoding method based on the MPEG-2 AAC standard, 11 different Huffman code tables are used for encoding. For most of these tables, the maximum possible codeword length is 10-20 bits. However, a special table, the so-called "miss table", contains a maximum length of 49 bits. If the length of the longest code word in all the tables is used as the raster distance D, a raster distance of 49 bits is set. This results in a raster of very large width, and if all preferred codewords are arranged at raster points, the bitstream would be too long and inefficient for almost all tables. Thus, according to the invention, the width of the raster is adjusted based on the code table used. As mentioned above, the spectral values are grouped into several spectral parts, each of which is given an optimal code table taking into account the statistical elements of the signal. The maximum code word length in one code table is usually different from the maximum code word length in other code tables.

  It is assumed that the spectral values represented by code words 1 and 2 belong to the first spectral part and the spectral values represented by code words 3 to 10 belong to the second spectral part. In this case, the bitstream is rastered by two raster point groups. The first raster point group consists of raster points 10, 12, and 14, and the second raster point group consists of raster points 14, 16, 18. Further, spectral part 0 is given a Huffman code table n, spectral part 1 is given a Huffman code table m, and code word 2 is the longest code word of table n given to spectral part 0. The raster distance of the first group of raster points is greater than or preferably equal to the maximum code word length in Table n, that is, the code word 2 length in this example.

  On the other hand, as can be seen from the portion between the raster point 14 and the end of the bitstream at the code word 10, the longest code word in the code table m does not appear in this example. Therefore, there is no codeword of length D2 in the raster of the bitstream represented by group 2.

  According to a second aspect of the invention, the width of the raster is selected based on the code table used. However, in this case, the tables used must be recognized as they are decoded at the decoder. However, if the code table number is always transmitted as side information of each spectral part, the decoder can recognize which code table of a certain different, in this example eleven, Huffman tables. it can.

  As described above, even if the raster distance is determined by the code table used, optimal data compression is not achieved, as can be seen by considering the missed table containing the 49-bit codeword. This is because in the case of this missing table, the raster distance would be adjusted to 49 bits so that the maximum size spectral value could be encoded. The miss table is used to have a short code table so that relatively large values can be encoded using the short code table along with the miss table. For values beyond the range of one code table, the code word for this spectral value will be a particular value, which will indicate to the decoder that the miss table has also been used in the coder. . If a code table contains the values 0-2, for example, a value of 3 in the code table will indicate to the decoder that the miss table has been used. A codeword having a value of 3 in the "basic" code table, together with the maximum value in that basic code table, gives the value of the missed table forming the corresponding spectral value.

  According to a further embodiment of the invention, the distance between raster points in one group (eg group 1 or group 2) is no longer the same as the length of the longest codeword in a code table, but rather the length of the bit stream belonging to the code table. It is the same as the length of the longest code word that actually occurs. In the first embodiment of the second feature point of the present invention, the encoding efficiency in the missing table is not yet optimal, so this embodiment is further improved. The maximum length (within the spectrum) of the code words in this table is usually quite short for technical coding reasons. The longest codeword in the missed table is, for example, 49 bits.

  The longest escape table codeword that actually occurs in a normal audio signal is typically about 20 bits long. Therefore, by transmitting the number of raster points and the length of the longest codeword in one block, the number of priority codewords that can be arranged at the raster points can be further increased. The length of the raster is then equal to the length of the longest code word that actually occurs or the length of the logical longest code word of the currently used table, whichever is the minimum value. To determine the minimum value, either the actually occurring code words of each code table or simply the longest code words of all code tables used in one audio frame can be used. This choice can also be used for non-missing tables, or "basic" Huffman tables, but is not as efficient as missed tables.

  Transmitting the length of the longest codeword in a spectral part or spectral block creates another advantageous side effect. The decoder can detect from the already generated maximum length whether there is a longer codeword in the bitstream that may have been disturbed. Long code words usually represent high energy spectral values. If a very long codeword occurs due to a transmission error, this will be a very audible disturbance. Transmitting the maximum length will in most cases provide a means of sensing such errors and taking action. The countermeasure against the error may be to simply leave the overly long codeword blank or to do something more complicated.

  It is important to note that as many raster points as possible are desirable for efficient coding as well as error tolerance. However, the number of raster points is limited by the total length of the bitstream. This should, of course, not be lengthened as a result of rastering, because it leaves unused places in the bitstream, which contradicts the theory of overall data compression. However, it must also be pointed out that in some applications the extension of the bitstream may be acceptable for a high degree of error tolerance. Another point to consider is that the raster is preferably constructed so that as many code words as possible start at the raster point. Thus, the present invention has more flexibility in selecting the distance between raster points than in the prior art. In the very ideal case, this flexibility places all code words on raster points, but this requires a great deal of technical effort. The above-described method of arranging raster points, that is, a method of determining the distance between raster points in each spectral portion based on an associated code table, can be very close to this optimal case. However, this is especially the case when the code words that are less important psychoacoustically are not rasterized in the bit stream, because not all code words are psychoacoustically significant and leave no unused places in the bit stream. This is because it is inserted between the psychoacoustically important secret words.

  According to a third aspect of the invention, the code words are no longer linearly arranged in the bitstream in order of increasing frequency, but code words of different spectral values are "scrambled". In FIG. 1, there is some alternating linear arrangement related to the codeword frequency. This is because priority passwords indicated by oblique lines are arranged in descending order of frequency, and non-priority passwords that are not indicated by oblique lines are also inserted into the bit stream in descending order of frequency. If a so-called "sudden" error occurs in the bit stream shown in FIG. 1, i.e., if a disturbance occurs that leads to the collapse of several subsequent cryptograms, for example, the cryptograms 6,7a, 2,3,7b You will be affected at the same time.

  Corresponding decoded audio signals will have clearly audible disturbances in the spectral band indicated by the preferred code words 2, 3 because they are relatively wide spectrally. The problem of catastrophic errors is less clear from the very simple example of FIG. However, in reality, there are five or more raster points, and it is considered that a catastrophic error often occurs over a plurality of raster points. In such a case, data loss occurs in a relatively wide frequency band. obtain. For this reason, according to the third aspect of the present invention, it is preferable that the priority code words of the spectrum values are not arranged in the order of increasing frequency, and the scramble is made so as to be arranged in a random or random manner with respect to frequency. Is more preferable. Non-priority codewords may be handled as well. In the case of a random-like arrangement, this distribution need not be transmitted as side information as this distribution can be set in advance in the decoder. As a result, the loss of consecutive code words in the bitstream does not result in a complete loss of one frequency band, but only a very small loss in some frequency bands. This disturbance is rarely audible and can be masked more effectively than the loss of an entire frequency band.

  According to a fourth feature of the present invention, instead of a linear arrangement in which the priority of the priority word and the non-priority word are higher in frequency, for example, every nth word is raster-arranged, and the remaining word is interposed between them. Can be used. As described above, the number of raster points for one bit stream is limited by the total length and the distance between raster points. For example, considering the case of sampling at low bandwidth, the majority of code words are psycho-acoustically important code words. If a sampling rate of 16 kHz is used, all signals have a logically usable bandwidth of 8 kHz. Empirically, only 30% of the code words must be arranged on raster points and the remaining 70% must be arranged to completely fill the raster. However, this means that important frequency regions, e.g. 0-4 kHz for speech signals, cannot be covered or "protected" by code words located on raster points. Therefore, instead of placing all priority code words on a raster point, only the second, third, fourth, etc. priority code words are used to properly block error transmission in important frequency domains. The other priority code words are not arranged in a straight line, but arranged to fill the raster. For example, if every second, every third, etc. spectral value is found to be in the low frequency region, and the interspersed codewords collapse during transmission, the error concealment techniques such as prediction may be used in the decoder. It is also possible to reconstruct these code words.

  A method and apparatus for decoding a bitstream serves to reflect the encoding described above.

  The encoded bitstream has rasters with raster points (10,12,14) equidistant from codewords of different lengths in one code table, and these codewords are psychological compared to other spectral values. A common method for decoding a bitstream represented by an encoded audio signal that includes a preferred codeword that represents some acoustically significant spectral value, wherein the preferred codeword is arranged by raster points, (A) The distance D1 between two raster points is determined. If the distance between the two raster points is known, (b) the preferred code words in the encoded bitstream arranged at the raster points are in linear order with respect to frequency, and each priority code word begins with the raster point. It is reclassified to have a matching sequence. This results in the preferred code words being in a linear array with respect to the general frequencies shown in FIG. 2, and (c) decoding these code words using the associated code table to obtain decoded spectral values. Can be. (D) transforming the decoded spectral values back into the time domain to obtain a decoded audio signal, which can be processed in any well-known way, for example to be sent to a loudspeaker .

  If the bitstream is encoded using only one code table, the distance between raster points can be very easily determined by finding out which table was used for encoding from the side information of the bitstream. Can be set. Depending on the encoding, this distance may be the length of the longest codeword in this table, which can be set permanently to its coder. If that distance is the length of the longest codeword that actually occurs in a portion of the given bitstream, this is signaled to the decoder, such as in side information associated with the bitstream.

  The decoder performs the reclassification of the preferred and non-preferred code words, for example, by pointing a pointer to the encoded bitstream. If the decoder is aware of the raster distance and the preferred code words are linearly arranged in frequency, the decoder can jump to the raster point and read the code words starting there. After reading one codeword, the pointer jumps to the next raster point and repeats this process. Even if all preferred code words have been read, the bitstream still contains non-preferred code words. If a linear arrangement of priority and non-priority code words in the bitstream is selected, the non-priority code words are already linearly arranged in frequency and can be decoded and returned to their original state without further classification. Is performed.

  If the coding according to the third or fourth aspect of the invention is selected, the scrambling information is sent as side information or the distribution of the scrambling state is fixed in advance, so that from the beginning the decoder can do this. Understand. The same considerations apply to the fourth feature point. It is always possible to define a constant distribution or to select a variable distribution that is communicated to the decoder as side information.

  An advantageous method for determining and manipulating the preferred secret word will now be described. Determine the raster for the encoded bitstream by setting a single raster distance when using only one code table, or multiple raster distances when using multiple code tables Later, the preferred code words must be placed on the raster such that each coincides with a raster point.

  In a preferred embodiment of the invention, this arrangement is achieved by inserting the codewords from a sort table into an otherwise empty raster. You can start with the first codeword in the table. Thus, the preferred codeword is affected by the sequence of the codewords in the table, but the codeword is always the codeword at the location of the raster, that is, the codeword for which raster points are available. For code words in the table that no longer have raster points, there is no choice but to insert them into the remaining space in the bitstream. These passwords are therefore not preferred passwords in the sense of the present invention.

  The number of preferred code words is not predetermined. Priority code words are written until the memory for the encoded bitstream is full, that is, until the priority code words can no longer be written. The size of this memory is the same as the sum of bits previously used for that spectral data. That is, no more bits are required by rastering. The memory is limited by the number of code words in order to prevent the efficiency of the encoding from dropping as a result of the raster processing. Of course, all code words may be placed on raster points so that errors can be tolerated. However, in this case, empty bits remain unused between raster points, leading to a significant reduction in coding efficiency.

  A first feature of the present invention relates to the determination of a priority code word, that is, a code word representing a spectral value that is more psychoacoustical than other spectral values. The psychoacoustically important spectral line is, for example, a spectral line containing more energy than other spectral lines. Generally speaking, the higher the energy, the more important the spectral line is. It is therefore important that the spectral lines with high energy are not disturbed and, similarly, that the spectral lines with high energy do not occur as a result of errors.

  So far, the description has been made on the assumption that the spectral line with the higher energy is located mainly in the lower part of the spectrum. This is often the case, but not all. The present invention provides an implied indicator for estimating the energy of spectral lines encoded in one codeword or, if multiple spectral lines are encoded in one codeword, these multiple spectral lines. Use ignores this hypothesis.

  This indicator is a code book or code table, such as the Huffman code table used. For example, in the AAC standard, eleven tables are used. The ranges in these tables are quite different. The maximum absolute values in Tables 1 to 11 are as follows.

  1; 1; 2; 2; 4; 4; 7; 7; 12; 12; 8191.

  As a result of these different ranges, the maximum error is tabulated. Considering the display of each table, this can be confirmed from within the table or transmitted from outside the table, but the maximum error is twice the absolute value. According to the invention, the determination of the preferred codeword is made on the basis of the code table used, the indicator of which is the code table number implicitly with the maximum absolute value. First, consider a code word whose code table has the largest range. Next, the code table is followed by a code word having the second largest value. Thus, for the AAC standard, Table 11 is considered first, followed by Tables 9 and 10, and Tables 1 and 2 are the last in priority. The preferred codeword placed at the raster point is thus the codeword in the sort table for which the raster point can be used.

  The advantage of this method of determining the secret word is that the table used is transmitted in side information, from which the decoder can determine the sequence of secret words used during the transmission, so that other additional information can be passed to the decoder. No need to send to.

  A second feature of the invention is that a short (sampling) is used as opposed to a long window to transform discrete time samples of the audio signal in its frequency domain to obtain spectral values representing the audio signal. ) Regarding using windows. Short windows are defined in the AAC standard and standard layer 3. For short windows, multiple short MDCTs are used instead of one long MDCT.

  In the AAC standard, a group of eight MDCTs each having 128 output values is used instead of one MDCT having, for example, 1024 output values. This will increase the temporal resolution at the expense of the coder's frequency resolution. Generally, short windows are used for temporary signals. For example, if a short window is used with AAC, eight consecutive complete spectra are obtained, i.e., eight sets of spectral values, each set containing the full spectrum, but the distance between the spectral values is also eight times as large. It is. This represents a decrease in frequency resolution, and this is accompanied by an increase in temporal resolution.

  Grouping is performed in the AAC standard. That is, a group is formed from eight spectra. There is one set of magnifications for these groups. In the simplest case, each group contains one window. In this case, eight sets of magnifications must be transmitted. To achieve stronger compression, multiple windows are grouped into one group of the AAC standard, generally taking into account psychoacoustic requirements. This reduces the number of magnifications to be transmitted and results in better data compression. The spectrum data is described and transmitted in a bit stream that is sequentially coded for each group. Within the group, the magnification bands are alternately arranged.

  This is illustrated by the following example. Here, it is divided into three groups. The first group includes two windows, the second group includes three windows, and the third group includes three windows. Each spectrum has 12 magnification bands. The grouping is as follows.

1st group, 1st window, 1st magnification band 1st group, 2nd window, 1st magnification band 1st group, 1st window, 2nd magnification band 1st group, 2nd window, 2nd magnification band. . .
1st group, 2nd window, 12th magnification band 2nd group, 3rd window, 1st magnification band 2nd group, 4th window, 1st magnification band 2nd group, 5th window, 1st magnification band 2nd Group, third window, second magnification band. . .

  This arrangement is not suitable for pre-sorting or inserting dark words from a sorted table into a raster. This is because, when successive insertions are made, the entire spectrum of the first group is protected but the spectrum of the last group is not. For this reason, the pre-sorting according to the second aspect of the invention is performed for short windows. In the case of the AAC standard, the grouping and scaling band approaches are abandoned. The new presorting is now in the form of spectral line units.

  In a preferred embodiment of the present invention, each unit contains four spectral lines. According to the AAC standard, each window thus contains 32 units, corresponding to 128 spectral lines. The spectrum data is as follows.

First window, first unit Second window, first unit. . .
Eighth window, first unit First window, second unit Second window, second unit,. . .
Eighth window, second unit First window, third unit,. . .

  This pre-sorting is such that the individual spectral regions of all windows are located close to each other, i.e., lower spectral values are based on frequency and a shorter table from the individual set of spectral values before spectral values with higher frequencies. Is written in front of. The spectral values in the lower spectral range are particularly important psychoacoustically, and the aforementioned presorting of the sort table provides a basis for inserting spectral values from the sort table into a raster. By pre-sorting this codeword, i.e. determining the preferred codeword, it is not necessary to send any additional information. Because the decoder knows from the side information that a short window has been used in this block or frame, the coder's classification algorithm for generating units is always constant, so that it is permanently programmed into the decoder. This is because that.

  It is important to keep in mind that pre-sorting code words into a sort table is equivalent to determining a preferred code word. Because the codeword located at the raster point, that is, the preferred codeword, is likely to be at the beginning, that is, before or above the sort table, the table itself is highly likely, and which codeword is at the raster point. It is because it determines whether it can be written.

  Other than in this preferred embodiment, this pre-sorting is not performed by the sort table, but by pointing to individual code words and determining the order in which the indicated code words are written into the bitstream.

  It is known from the AAC standard that some code tables are two-dimensional or four-dimensional, that is, one codeword encodes two or four spectral values. Therefore, it is advantageous to group the four spectral lines or their multiples into one unit. In this way, the code words encoding the same frequency region are classified so as to be directly continuous with each other. Preferably, the number of spectral lines in a unit is divisible by the different dimensions of the code table. That is, the number of lines per unit is a common multiple of the number of lines per codeword, and the least common multiple is optimal.

  The present invention is particularly effective when the first and second features are combined. Reclassification into units according to the invention is performed for short windows, and then a determination of the preferred codeword is made using the indicators in the code table, where a high code is used to achieve a high degree of error protection. The results of the reclassification into units are reclassified again to ensure that the code words from the table are located on certain raster points. This combination is not required, but will lead to the best results.

FIG. 9 shows an example of rastering of an encoded bitstream including a cryptographic word according to the second aspect of the present invention. FIG. 1 shows an arrangement of code words in order of linearly increasing frequency with respect to the prior art.

Claims (8)

An apparatus for decoding a bit stream, the bit stream including code words having different lengths extracted from a code table, and including information regarding the length of the longest code word that actually occurs as side information. Wherein the device comprises:
A decoding unit for decoding a bitstream using a code table, which decoding unit is longer than the length of the longest codeword in which the codeword extracted from the bitstream occurs, and thus whether it is the wrong codeword; Can be detected, and if such an incorrect codeword is detected, countermeasures can be taken.   2. The apparatus according to claim 1, wherein the decoding unit can perform blanking or concealment of an incorrect codeword as the countermeasure.   Apparatus according to claim 1 or 2, wherein the bit stream represents an encoded audio signal, wherein long code words correspond to spectral values of the audio signal with high energy and relatively short code words are relatively low. It corresponds to the spectral value of an audio signal having energy. 4. The apparatus according to claim 1, wherein the bit stream comprises code words of different lengths derived from a code table, the raster points being arranged at equidistant distances. 12, 14), wherein the code word comprises a priority code word representing a group of spectrally-acceptable spectral values compared to other spectral values. The value group indicates a spectrum of a temporal sample group of the audio signal, wherein the priority code word has a start point of the priority code word representing one spectrum value of the spectrum value group coincides with one raster point, The apparatus is arranged on a raster such that the beginning of a preferred codeword representing another spectral value of corresponds to another raster point. Characterized in that it comprises the following in al,
The length (D1) between two adjacent raster points is detected by using the length of the longest actually occurring code word transmitted as side information in the bit stream, and the distance is made equal to that length. Unit to set,
A unit for reading a priority codeword located at a raster point in the bitstream such that its start coincides with the raster point, whereby the priority codeword is decoded using the associated code table and the decoded spectral value Which is transformed to the time domain to obtain a decoded audio signal. 5. The apparatus according to claim 4, wherein the bit stream comprises codewords of different lengths derived from at least two code tables and equidistantly arranged raster points (10,12,14 and 14,16,14). 18) a raster having at least two groups, including, for one part, the length of the longest code word actually occurring in that spectral part as side information, wherein the device is associated with the spectral part. It further includes a unit that recognizes the code table,
The distance sensing unit sets a distance between two raster points for a cryptographic word associated with a spectral portion to be equal to the length of the longest cryptographic word occurring in that spectral portion;
In decoding, the preferred code words of the spectral part are decoded by the associated code table,
The decoding unit is able to detect in one part whether the codeword associated with that part is longer than the longest codeword that occurs. A method for decoding a bitstream, wherein the bitstream includes codewords having different lengths extracted from a code table, and includes information regarding the length of the longest codeword that actually occurs as side information. The method comprises the following steps:
Decoding the bitstream using the code table, the decoding step includes the following sub-steps:
Detecting whether the codeword extracted from the bitstream is longer than the longest codeword occurring, and thus whether it is a wrong codeword,
If such an incorrect codeword is detected, take measures. 7. The method of claim 6, wherein
The encoded bitstream is encoded with rasters having raster points (10,12,14) equidistantly aligned with each other, including codewords of different lengths derived from a code table. An audio signal, wherein the secret word includes a priority secret word representing a special spectral value of a group of spectral values that are psychoacoustically significant compared to other secret words, and the spectral value group includes a temporal code of the audio signal. 5 shows a spectrum of a group of samples, wherein the priority codeword is a different priority value representing the first spectral value of one of the spectral value groups coincides with one raster point and representing another spectral value of the spectral value group. Are arranged by raster points so that the beginning of the codeword coincides with another raster point,
The method further comprises the following steps:
The distance (D1) between two adjacent raster points is detected based on the information regarding the length of the longest code word actually generated and transmitted as the side information in the bit stream, and the distance becomes equal to the length. To set,
In the encoded bitstream, the priority codewords arranged by raster points such that the beginning of the priority codeword coincides with the raster point are re-classified in a linear arrangement with respect to frequency, whereby The codeword is decoded by an associated code table to obtain a decoded spectral value, and further converted to a time domain to obtain a decoded audio signal. A method according to claim 7, wherein
The bit stream contains codewords of different lengths derived from at least two code tables and has at least two groups of equidistantly arranged raster points (10,12,14 and 14,16,18). Having a raster and, for one part, including, as side information, the length of the longest codeword actually occurring in that spectral part, the apparatus further comprises the step of recognizing a code table associated with the spectral part. ,
In the distance detecting step, a distance between two raster points for a secret word associated with one spectral part is set to be equal to a length of the longest secret word occurring in the spectral part;
In decoding, the preferred codeword of one spectral part is decoded by its associated code table,
The detection and countermeasure sub-steps are performed in each part using the length of the longest code word occurring in each part included in the bitstream as side information.

Images