KR100460159B1 - Audio signal encoding method and apparatus - Google Patents

Abstract

The audio coding technique of the present invention utilizes a signal adaptive switched filterbank having a first filter bank and a wavelet filter bank. The filter bank switches between the first filter bank and the wavelet filter bank to filter the input signal as a function of the stationarity of the input signal. The first filter bank is used to filter stationary signal components. The wavelet filter bank is used to filter non-stationary signal components (eg, attachments).

Description

Audio signal encoding method and apparatus

This application claims the priority of US Provisional Application No. 60 / 014,725, filed March 19, 1996.

TECHNICAL FIELD The present invention relates to the processing of signals, and in particular, to the encoding of an audio signal using subband coding designs, such as perceptual audio coding.

Consumers, producers, studios and laboratories have high demands for products for storing, processing and communicating high quality audio signals. Compression of audio signals at very low bit rates is very desirable for many modern digital audio applications such as digital audio tapes, compact discs and multimedia applications. The compression techniques used in these digital applications can process good signals. However, such performance is often achieved through large amounts of data storage capacity and transmission bandwidth.

Large workloads in the compression domain have sought to reduce data storage and transmission bandwidth requirements in the coding of digital audio. Such compression techniques remove the extraneous information of the source signals by using a model of the human perceptual system. This perceptual audio coding ("PAC") technique, for example, is entitled "Method and Apparatus for Coding Audio Signals Based on Perceptual Model". Jay, published February 8, 1994. Dee. J. D. Johnston, US Pat. No. 5,285,498, which is incorporated herein by reference (hereinafter referred to as the "Johnston" patent).

For example, perceptual audio coding as described in the Johnston patent is a technique for lowering the number of bit rates or total bits required to represent audio signals. PAC technology uses short-term energy distribution as a function of frequency. From this energy distribution, it is known that a set of thresholds representing immediately recognizable noise levels can be calculated. At this time, among others, the coarseness of the quantization used to represent the signal component of the desired signal is selected so that the quantization noise induced by the coding itself does not rise above the noise thresholds. Therefore, the induced noise is masked in the perceptual process. Masking occurs because of the inability of human perceptual mechanisms to distinguish between two signal components (one belonging to the signal and one belonging to the noise) in the same spectrum, temporal or spatial place.

Recently, a number of perceptual audio coders have been developed which are required to provide transmission compression in the range of 128 to 256 kbps (ie 6 to 12 compression factors). Typically, such coders use analysis filter banks that divide the input signal into its frequency components. These components are then quantized using a perceptual model based on the masking characteristics of the human auditory, as described above. In the Johnston patent, for example, a PAC approach in which a high frequency resolution filterbank, known as a Modified Discrete cosine Transform ("MST") filter bank, is used to split the signal into frequency components, Are described. This high frequency decomposition MDCT filter bank (e.g. with 1024 subbands or frequency lines) is a very compact representation of so-called stationary signals (e.g. playing music and most vocal music). to the representation. However, non-stationary audio signals, including so-called momentary or sharp attachments (e.g. castanets or triangles), cannot be closely represented using a high frequency decomposition MDCT filter bank. . This is due to the higher time resolution required at higher frequencies for dense representations. In addition, using MDCT for abnormal signal components leads to poor quality coded signals.

Other techniques have been developed to address the filtering problem encountered when coding abnormal signals. For example, such a technique described in the Johnston patent uses a so-called "window switching" design. This PAC design uses so-called "long" and "short" MDCT windows to handle sharp utterances of abnormal signals. In "window switching", the normality of the signal is monitored at two levels. First, long MDCT windows (e.g., a window with 1024 subbands) are used for normal signal components, and if necessary, short windows (e.g., a window with 128 subbands) are periods of abnormality. Is used during. However, a drawback of this approach is that short MDCT windows increase the time resolution evenly for all frequencies. In other words, in order to increase the time resolution to the desired degree at higher frequencies, this technique must also increase the time resolution at lower frequencies.

A more preferred filter bank for filtering sharp vocals is to have a non-uniform structure with subbands that match the critical band division of the frequency axis (ie, the subbands are uniform on the bark scale). In addition, it is much desirable that the high frequency filters of the filter bank be proportionally shorter. One coding scheme that meets these objectives is a hybrid or cascaded structure (e.g., K. Brandenburg et al. "ISO-MPEG Audio Codec: General Standard for Coding of High Quality Digital Audio" The ISO-MPEG-Audio Codec: A Generic Standard for Coding of High Quality Digital Audio. "Journal of Audio Engineering Society, Vol. 42, No. 10, October, 1994, and J. Princen and Jay. JD Johnston's "Audio Coding with Signal Adaptive Filterbanks", Proceedings of IEEE, ICASSP, Detroit, 1995). This coding technique consists of a first stage having a uniform or non-uniform filter bank. Each subband may be further split using uniform filter banks. However, a drawback of this approach compared to MDCT filter banks is that the hybrid / cascade structure must be used for both normal and abnormal signals leading to poorer frequency response of the filters as well as increased implementation costs.

Therefore, there is a need in the art for a filter bank that overcomes the drawbacks of prior art filtering arrangements for handling abnormal signals in subband coding.

Signal compression techniques using the principles of the present invention switch between a first filter bank and a wavelet filterbank to code audio signals using perceptual audio coding or similar subband type coding.

In a preferred embodiment, the switching between the two filter banks is based on the time-varying characteristics of the signal, preferably its perceptual entropy level. Also in preferred embodiments, the first filter bank is a high frequency decomposition MDCT filter bank. In general, a high frequency decomposition MDCT filter bank is used to filter the input signal, but in the case of anomalies a wavelet filter bank is used. Advantageously, the present invention achieves a more dense representation of the signal when it contains abnormal components. According to a preferred embodiment, the wavelet filter bank is a filter bank of non-uniform tree structure.

The present invention provides a signal adaptive switch filter bank for switching between a first filter bank (preferably a high frequency decomposition MDCT filter bank) and a wavelet filter to handle abnormal signals coded using perceptual audio coding or similar subband type coding. It relates to an audio signal compression technique using.

Exemplary embodiments of the invention are provided in functional blocks for clarity. The functions represented by these blocks may be provided through the use of either shared or dedicated hardware, including but not limited to hardware capable of executing software. In addition, the use of the term "processor" should not be interpreted as exclusively referring to hardware capable of executing software. Some embodiments may include a digital signal processor ("DSP") hardware, such as ATT & T DSP16 or DSP32, and software for performing the operations discussed below. In addition to hybrid DSP / VLSI embodiments, large scale integrated (“VLSI”) hardware embodiments of the present invention may also be provided.

1 is an overall block diagram of an exemplary system in which the present invention is implemented. In FIG. 1, analog audio signal 101 is provided to preprocessor 102, sampled (typically at 48 kHz), and 16-bit digital pulse code modulation per sample on lead 103 in a conventional manner (hereinafter PCM signal is supplied to a perceptual audio coder 200 which compresses the PCM signal and outputs a compressed PAC signal on the lead wire 105 on either the communication channel or the storage medium. For example, a magnetic tape, a compact disc, or other storage medium .. From a communication channel or storage medium, a compressed PAC encoded signal on a wire 107 decompresses the compressed PAC encoded signal and the original audio signal ( Supplied to a perceptual audio decoder 108 that outputs a PCM signal on a lead 109, which is a digital representation of 101. From the perceptual audio decoder, a PCM signal on a lead 108 is used to produce an analog representation of the signal. It is supplied to the bit processor (post-processor).

An exemplary embodiment of the perceptual audio coder 200 is shown in block diagram in FIG. The perceptual audio coder 200 has the observed benefits as it includes a signal adaptive switch filter bank 202, a perceptual model processor 210, a quantizer / speed loop processor 212, and an entropy coder 214. The structure and operation of the perceptual model processor 210, the quantizer / speed loop processor 212, and the entropy coder 214 are generally similar to the structure and operation of components as identified in the Johnston patent for processing audio signals. Do. However, the signal adaptive switch filter bank 202 is described in detail in connection with the switching between the first filter bank (preferably high frequency decomposition MDCT filter bank) and the wavelet filter bank. In combination with the other elements of FIG. 2, the characteristics of the switch filter bank 202 provide the advantages of the present invention.

2, the signal adaptive switch filter bank 202 is used for switching 206 between the high frequency decomposition MDCT filter bank 204 and the two filter banks during encoding of the signal in a predetermined form as described above. A wavelet filter bank 208 for use. As mentioned above, the use of high frequency decomposition MDCT (eg, 1024 subbands or frequency lines within a PAC) in the encoding process is useful for the MDCT to derive a very dense representation of a normal signal. For the purpose of the PAC, MDCT has the following characteristics: (i) critical sampling characteristics (ie, for all n samples into the filter bank, n samples are obtained); (Ii) In general, MDCT is a half-overlap that provides a good way to handle the control of noise injected independently into the filter bank (i.e. the strain length is exactly twice the number n of samples shifted into the filter bank). (I) MDCT provides accurate reconstruction of the input sample only with a delay of integer samples. Known MDCTs are, for example, Jay. Bloody. J. P. Princen and A. A. B. Bradley, "Analysis / Synthesis filterbank Design Based on Time Domain Aliasing Cancellation", IEEE Yrans. ASSP, Vol. 34, No. 5, October 1986. Known applications of MDCT for use of PACs and the functions performed by the high frequency decomposition MDCT filter bank 204 herein are fully described, for example, in the Johnston patent.

However, while used to represent the high frequency decomposition MDCT 204 filter bank steady signal as described above, MDCT does not provide a dense representation of an abnormal signal (ie, a signal that includes instantaneous or sharp speech). However, we have realized a technique that provides the advantage of using a high frequency decomposition MDCT filter bank 204 that can improve the audio compression characteristics of the audio coder 200.

Thus, according to the present invention, a signal suitable for the switch filter bank 202 uses both a high frequency decomposition MDCT filter bank 204 and a wavelet filter bank 208 to encode the audio signal 101, for example. According to a preferred embodiment, the high frequency decomposition MDCT filter bank 204 uses a high frequency decomposition MDCT for encoding purposes. That is, when an abnormal signal is supplied, the filter bank 204 uses only so-called long windows (i.e., 1024 subbands) and "switches" to so-called short windows (i.e. 128 sub-bands opposing the 1024 sub-bands). "I never do that. Of course, this is described in the Johnston patent as the conventional window switching technique mentioned previously. According to the present invention, the switch filter bank 202 uses the wavelet filter bank 208 rather than switching to a short MDCT window during the abnormal period.

In particular, the wavelet filter bank 208 uses wavelet modifications to effectively filter input signals with abnormal components. Wavelets serve to provide a complete orthogonal regime to the space of finite energy signals through various deformation and expansion characteristics. The overall coding of the audio signal using the optimal wavelet modification is, for example, Dei. D. Sinha and A. H. A. H. Tewfik, "Low Bit Rate Transparent Audio Compression Adapted Using Adapted Wavelets," IEEE Transactions on Signal Processing, Vol. 41, No. 12, PP. 3463-3479, Dec. Described in 1993. In accordance with an embodiment of the present invention, we use constant frequency and time characteristics to use a psychoacoustic model based on a PAC, and according to key criteria in the drawings of the wavelets illustrated herein. Wavelet variants have been employed for this purpose.

It is known that the time-frequency decomposition of the psychocore acoustic analysis can be combined with the time-frequency decomposition of the auditory system. The decomposition characteristic is reflected in the reference band scale which instructs the frequency decomposition in the Psycorecoustic model to vary from a low frequency of 100 Hz to a high frequency of about 4 kHz (ie, a change of 40: 1 upon decomposition). As such, the instantaneous decomposition in the PAC coder can be increased at a ratio of about 40: 1 at low to high frequencies. It is known that most psycho acoustic models use very low uniform temporal decomposition. The lack of temporal decomposition at high frequencies has little effect on the calculated threshold for steady signals. However, the threshold calculated for the abnormal signal may be inaccurate and cause audible distortions. The operation can be improved by using the signal adaptive switch filter bank of the present invention.

The use of the signal adaptive switch filter bank of the present invention provides several advantages over the prior art for the coding of abnormal signal segments or transients. For example, it leads to a simpler representation of the abnormal signal component. It also leads to more accurate psychocore modeling of abnormal segments of the signal. It can be understood that the feature significantly saves the overall bit rate component for representing transitions. In addition, the use of the present signal adaptive switch filter bank protects the benefits of the known performance of the high frequency decomposition MDCT filter bank for the compression of normal signal segments.

In particular, according to a preferred embodiment of the present invention a tree-shaped wavelet filter bank is used. As mentioned above, the accuracy of the psychocore model is important because the frequency split used is quite close to the critical bank division of the frequency axis. The wavelet filter bank provides good frequency selectivity (ie, small overlap between adjacent subbands). In addition, the wavelet filter bank provides good temporal characteristics (also known to be closely located) in which the impulse response of the higher frequency subbands quickly disappears. The closely located higher frequency subbands lead to an effective indication of the abnormal signal segment. The tree structure used in the preferred embodiment of the present invention is beneficial to provide the desired wavelet filter bank characteristics described above. The tree structure provides a relatively short advantage of filters for higher frequency subbands because the critical band is wider at higher frequencies, thus enabling the desired frequency decomposition with fewer stages in the overall tree structure. In addition, control of the temporal characteristics of the filter bank of the tree structure is supplied using a moment condition. To closely combine the tree structure with the critical band division, the tree structure wavelet filter bank of the preferred embodiment uses three sets of filter banks. One set of filter banks provides four subband splits, while the other two sets each provide two subband splits as described below.

3 illustrates a decompression tree 300 for a wavelet filter bank of a tree structure used for the switch filter bank 202. According to a preferred embodiment, the three sets of filter banks used in the exemplary tree structure of the wavelet filter bank 208 provide sufficient design flexibility for the three configurations to be close to the critical band portion. In particular, the first filter bank set 310 provides four band splits (ie, 311-314) of the signal. By way of example, the four band filter splits increase the frequency from filters 311-314, and each filter has 64 supports (length). Further, by way of example, the second filter bank 320 provides two band splits (i.e., 321 and 322) with 40 supports, while the third filter bank 330 also has two supports with 20 supports. Band splits (ie, 331 and 332). As will be apparent to those skilled in the art, the application of the filter bank 310 at any node of the decompression tree 300 depends on the decision by four factors. Similarly, each application of filter banks 320 and 330 is subject to a determination by two factors. By way of example, with an input block of N samples, subband 331 has N / 64 filtered samples, while subband 322 has N / 4 filtered samples. The three filter banks used by wavelet filter bank 208 are optimized, for example, by using a known parameterized parity filter bank and by providing standard optimization means. The optimization provided by the tree filter filters described above allows each of the three filter banks as well as the entire filter bank to provide excellent frequency selectivity on its own. The optimization criteria used to optimize the wavelet filter bank 208 are based on known weight stop band energy criteria (eg, P. Vaidyanathan's "Multi-rate Digital Filters, Filter Banks"). , Multiphase Digital Filters, filterbanks, Polyphase Networks, and Applications: A Tutorial, "Proceedings of the IEEE, Vol. 78, No. 1, pp. 56-92, January 1990. Able to know). The optimization provided by the tree structure filter bank described above allows each of the three filter banks as well as the entire filter bank to provide excellent frequency selectivity on its own.

In a preferred embodiment, the moment condition plays an important role in implementing the desired temporal characteristics of the high frequency filter (ie, the filter corresponding to the subband in the compression decomposition tree 300 with higher frequency) at the center frequency at the moment condition. Determine the flatness (order of differentiablility) of the higher subband frequency response closer. As can be seen below, the greater flatness near the center frequency leads to a correspondingly located corresponding impulse response. In particular, the M-band parunitary filterbank with subband filter {H _i } _{i = 1} to M is H (e ^iw ) for i = 2, 3, ... M with w = 0 at w = 0 ^{If th} order is zero, the P ^th order moment condition is met. At this time, the filter makes P a moment with a value of zero. In the illustrated wavelet filter bank 208 configuration, with a predetermined support K for the filter, which should be P > 1, the effective support makes a filter that reduces the increase in P. That is, most of the energy is concentrated in the interval KT, where KT is smaller than the higher P.

It can be seen that the improvement in the instantaneous response of the filter is generally in an increased transition band in the amplitude frequency response (eg, P. Vaidyanathan's "Multi-Speed Digital Filters, Filters" Banks, Multiphase Networks and Applications: Multirate Digital Filters, filterbanks, Polyphase Networks, and Applications: A Tutorial ", Proceedings of the IEEE, Vol. 78, No. 1, pp. 56-92, January, 1990 The filter bank of the tree structure has two zero-valued moments (ie, P = 2) for each of the three sets of filter banks to have a desired position in the temporal characteristics of the filter. For example, the impulse response 410 of the highest frequency subband of the wavelet filter bank 208 (e.g., the subband 314 shown in FIG. 3) is derived from a cosine modulated filter bank having the same frequency characteristics. Compare the response of the filter 420 As shown, the response 410 from the wavelet filter bank constructed in accordance with the preferred embodiment is best positioned in time as evidenced by the impulse response 410 of the high frequency wavelet filter 314. The high frequency wavelet filter 314 concentrates most of its energy between n = 10 and n = 40. In comparison, the response 420 of the cosine modulated filter bank is in the full range n = 1 to n. Energy diffused to = 64.

In accordance with the principles of the present invention, the high frequency decomposition MDCT filter bank 204 is used to code a normal signal, and the filter bank 208 is used to code an abnormal signal. The threshold of efficiency using two filter banks is a mechanism for switching between those based on special signal components (i.e., normal versus abnormal signals). For this purpose, MDCT should be an overlapping orthogonal deformation. That is, there is a 50 percent overlap between adjacent blocks unlike conventional block variations. Thus, switching between the high frequency decomposition MDCT filter bank 204 and the wavelet filter bank 204 requires orthogonalization in the overlap region between the MDCT block and the wavelet block. On the other hand, there is a problem with known general orthogonal design methods (e.g. C. Herley et al., "Tiling in the Time-Frequency Plane: Arbitrary Orthogonal Foundations and Fast Tiling Algorithms. of the Time-Frequency Plane: Construction of Arbitrary Orthogonal Bases and Fast Tiling Algorithm. ", IEEE Transactions on Signal Processing, Vol. 41, No. 12, December, 1993." The matrix is inefficient at the point of execution. In other words, any configuration defect in the resulting filter makes the calculation of wavelet deformation very difficult.

Thus, the MDCT operation on a block of 2N samples matches the symmetrical operation in the data windowed along the N point orthogonal block variant Q on the N sample (i.e., an outer N / 2 sample from one end of the window is the inner N / 2 of the window). Simplification can be realized by overlapping with samples). It can be seen that the complete reconstruction of the signal is independent of the special block orthogonal deformation (Q). Thus, Q may be a MDCT for one block and a wavelet variant for the next block. The matrix Q corresponding to MDCT is known and will not be described further. The matrix Q used by the wavelet filter bank is described below. When using a wavelet variant, an orthogonal matrix (Q) filter bank (hereinafter referred to as Q ^WFB ) is an N × N matrix based on the three filter banks of the wavelet of the tree structure described above. The matrix Q ^WFB is composed of a plurality of blocks having each block corresponding to a node (ie, a subband) remaining in the decompression tree 300 of FIG. 3. As will be apparent to one of ordinary skill in the art, the matrix for decompression tree 300 may determine the filters and finite block sizes (ie boundary conditions) in three filter banks 310, 320, 330. Completely classified by the technique for manipulation. For clarity, the handling of the boundary conditions for the four band splits 310 of the decompression tree 300 shown in FIG. 3 in the preferred embodiment is described below. This expansion of the overall tree structure is apparent to those skilled in the art.

For the four band splits shown in FIG. 3, the corresponding deformation matrix Q is four subs of size N / 4 × N with one subblock corresponding to each of the filters 311, 312, 313, 314. It is composed of blocks. By way of example, it defines the length of the filter according to K, and also defines another constant K1 = (K / 4) -1. Respective subband filters and (N / 4-K1-1) modifications of the subband filters correspond to each of the four subblocks except the N / 4-K1 row of the subblock. To avoid cyclic rotation, the remaining K1 rows of sub-blocks are deformation filters designed to operate near the end of the block. In particular, Q1, Q2, Q3, and Q4 are defined according to K1 × N corresponding to four different rows. Next, Q1 to Q4 are selected such that the matrix in orthogonal regime to the subspace is collectively orthogonal to the predefined 4x (N / 4-K1) of Q. Further, Q1 to Q4 are selected to maximize the cost function with the form of

Cost = Trace (Q1WTD1WQ1T + Q2WTD2WQ2T + Q3WTD3WQ3T + Q4WTD4WQ4T), where W is a strain matrix in N × N glass, and D1 through D4 are diagonal matrices with N / 4 and 1 of non-zero N diagonal elements. A nonzero N / 4 nonzero device for a particular subband matches the location of the particular subband on the frequency axis. As will be apparent to those skilled in the art, this is a subspace limited optimization problem that can be solved using, for example, standard optimization means. For each subband, the distortion filter is placed in the Q ^WFB in order of increased group delay so that the subband coefficients can be accurately instantaneously interpreted.

It can also be seen that the orthogonality described above has the effect of introducing the wavelet filter's expansion and / or discontinuities in the wavelet filter itself within time. Any possible damage to the wavelet filter bank 208 can be mitigated for the following reasons: (i) The instantaneous START and STOP windows (eg, as described in the Johnston patent) may be used for the high frequency decomposition MDCT filter bank ( 204 and wavelet filter bank 208 are used depending on the transition during use; (Ii) Use a family of so-called flat windows to reduce the effective overlap between the transition window and the wavelet window. An exemplary switching sequence between the high frequency decomposition MDCT filter bank 204 and the wavelet filter bank 208 using the techniques described above is shown in FIG. 5. As shown in FIG. 5, the START window 502 is used for the transition between the high frequency decomposition MDCT filter bank window 501 and the wavelet filter bank window 503. STOP window 504 is also used for the transition between wavelet filter bank window 504 and high frequency decomposition MDCT filter bank window 505.

The so-called flat window is used for the overlap area between the START window 502 and the wavelet window 503, and again between the overlap area between the wavelet window 503 and the STOP window 504. The flat window is useful as a base band filter and is located densely in time (ie, most of the energy in the window is concentrated around the center). The flat window is generated using the following equation: h (n) = h (t) | t = (n + 1/2) (1 / N), n = 0, 1, ..... N- 1, where h (t) is not zero in the interval [0, 1] and zero otherwise.

Returning to FIG. 2, the perceptual model processor 210 uses psychocore analysis to calculate the perceptual importance assessment within the switch analysis filter bank 202 and to calculate the noise masking properties of the various signal components. Psychoacoustic analysis occurring in the processor 210 is known, for example Johnston Patent and J. Dee. J. D. Johnston, "Transform Coding of Audio Signals Using Perceptual Noise Criteria," IEEE Journal on Selected Areas in Communication, Vol. 6, pp. 319-323, February, 1988. On the other hand, the threshold for quantization of the coefficients in the MDCT block is provided directly in a known method from psychocore analysis, and the threshold used by the wavelet block requires additional processing.

The thresholds for quantization of the wavelet coefficients are based on the evaluation of time-varing spread energy and the sound quality specification evaluated according to the PAC. The spread energy is calculated taking into account the spread of frequency cross-masking as well as time. That is, the internal frequency as well as the temporal spreading function are used. The form of the diffusion function is, for example, Jay. B. Derived from chochlear filter as described in JB Allen's "The ASA edition of Speech Hearing in Communications," Acoustical Society of America, New York, 1995. do. The temporal spread of the masking is frequency dependent and roughly determined by the inversion of the bandwidth of the superclear filter at a particular frequency. Fixed temporal spreading functions are preferably used for a range of frequencies or subbands. Thus, the shape of the diffusion function becomes narrower at higher frequencies. Coefficients in the subbands are grouped within the coder band, with one threshold per coder band used during quantization. By way of example, the coder band extends from 10 msec in the lowest frequency subband to about 2.5 msec in the highest frequency subband.

In addition, the quantization / velocity loop processor 212 described in the Johnston patent takes the output from the switch analysis filter bank 202, the perceptual model processor 210, the allocated bits, and the noise to determine the required bit rate for a given application. To control other system parameters. Entropy coder 214 is used in conjunction with loop processor 212 to implement further noiseless compression. As described, for example, in the Johnston patent, entropy coder 214 receives the quantized audio signal output from quantization / speed loop processor 212. At this time, the entropy coder 214 may perform lossless encoding of the quantized audio signal using, for example, a known minimum-redundancy Huffman coding technique. Huffman code is, for example, Dee. a. D. A. Huffman, "A Method for the Construction of Minimum Redundancy Codes", Proc. IRE, 40: 1090-1101, 1952, and tee. M. T. M. Cover and J. a. J. A. Thomas, "Elements of Information Theory", pp. 92-101, 1991. The Johnston patent also describes the use of Huffman coding in the PAC content of entropy coder 214. Those skilled in the art can readily recognize how to implement another embodiment of entropy coder 214 using other noiseless data compression techniques, including the known Lempel-Ziv compression method.

As a result, the switching criteria 206 is used to further perform effective switching between the high frequency decomposition NDCT filter bank 204 and the wavelet filter bank 208. In order to be efficient, the criteria must search for utterances accurately without any false alarms or missed utterances. For example, unsearched utterances cause significant distortion, especially at low bit rates, when encoded using the high frequency decomposition MDCT filter bank 204. In contrast, coding a relatively normal signal with wavelet filter bank 208 introduces a significant waste of output bits and processing power.

Thus, according to a preferred embodiment, perceptual entropy criteria are used. As mentioned above, perceptual entropy is a measure of a particular strained segment of a signal that clearly feeds the code segment a theoretically low boundary of bits per sample. A large increase in perceptual entropy from one segment to the next makes a valid indication of a strong anomaly (eg, speech) in the signal. According to the embodiment of FIG. 2, this form of perceptual entropy change is used by coder 202 to trigger switching from high frequency decomposition MDCT filter bank 204 to wavelet filter bank 208. By way of example, a decision regarding switching between high frequency decomposition MDCT filter bank 204 and wavelet filter bank 208 is made by coder 202 every 25 msec. In the end, the foregoing merely illustrates the principles of the invention. Persons of ordinary skill in the art, although not explicitly shown and described herein, may make various other arrangements to implement the above principles and the principles within the scope of the invention as defined in the appended claims. Can be.

The audio signal compression technique of the present invention can handle coded abnormal signals by using signal adaptive switch filter banks that switch between high frequency decomposition MDCT filter banks and wavelet filters using perceptual audio coding or similar subband type coding. .

1 is a block diagram of a system in which the present invention is illustratively implemented.

2 is a block diagram of an exemplary perceptual audio coder used in the system of FIG. 1 using the signal adaptive switch filter bank of the present invention.

3 illustrates a tree structured wave and filter bank used in the signal adaptive switch filter bank of FIG.

4 illustrates a comparison between a cosine modulation filter and a wavelet filter used in the signal adaptive switch filter bank of FIG. 2.

5 illustrates an example filter bank switching sequence generated using the signal adaptive switch filter bank of FIG. 2.

* Description of the symbols for the main parts of the drawings *

101: analog audio 102: preprocessor

103: PCM signal 106: communication channel / storage medium

108: perceptual audio decoder 110: postprocessor

202: signal adaptive switch filter bank 300: decompression tree

Claims (17)

In the method for encoding an audio signal, Sampling the audio signal; Selectively filtering the sampled audio signal by switching between a first filter bank and a wavelet filter bank to produce a filtered signal, wherein the wavelet filter bank is a non-uniform filter bank of a tree structure and the first filter bank; The filter bank is independent of the wavelet filter bank, and the switching occurs in response to a stationarity of the audio signal; Encoding the filtered signal to provide a compressed output signal. The method of claim 1, And the first filter bank is a high frequency resolution MDCT filterbank. The method of claim 2, And wherein the wavelet filter bank uses a plurality of moment conditions to distinguish frequency response within the non-uniform filter bank. The method of claim 2, And in said filtering step said high frequency decomposition MDCT filter bank is used to filter normal components of said audio signal, and said wavelet filter bank is used to filter abnormal components of said audio signal. The method of claim 1, And wherein said encoding step comprises perceptual audio coding. In the method for encoding an audio signal, Generating a plurality of noise thresholds as a function of frequency characteristics of the audio signal; Selectively filtering the audio signal by switching between a first filter bank and a wavelet filter bank to produce a filtered signal, wherein the wave and filter bank are a non-uniform filter bank of a tree structure and the first filter bank Is independent from the wavelet filter bank, and the switching occurs in response to the normality of the audio signal; Quantizing the filtered signal, wherein the coarseness of the quantization is determined by the noise thresholds; And perceptually encoding the quantized signal. The method of claim 6, And the first filter bank is a high frequency decomposition MDCT filter bank. The method of claim 7, wherein And wherein the wavelet filter bank uses a plurality of moment conditions to distinguish frequency response within the non-uniform filter bank. The method of claim 7, wherein And in said filtering step said high frequency decomposition MDCT filter bank is used to filter normal components of said audio signal, and said wavelet filter bank is used to filter abnormal components of said audio signal. The method of claim 8, And said normality of said audio signal is determined using perceptual entropy. The method of claim 10, The first of the non-uniform filter banks provides four band splits of the audio signal and the second of the non-uniform filter banks provides two band splits of the signal Way. A method for encoding a digital audio signal to generate a compressed output signal, the method comprising: Generating a plurality of noise thresholds as a function of the frequency characteristic of the digital signal; Selectively filtering the digital signal by switching between a first filter bank and a wavelet filter bank to produce a filtered signal, wherein the wavelet filter bank is a non-uniform filter bank of a tree structure and the first filter bank Is independent from the wavelet filter bank, and the switching occurs in response to the normality of the audio signal; And perceptually encoding the filtered signal to provide the compressed output signal. The method of claim 12, And the first filter bank is a high frequency decomposition MDCT filter bank. An apparatus for encoding an audio signal, the apparatus comprising: The device, Means for sampling the audio; Means for selectively filtering the sampled audio signal by switching between a first filter bank and a wave and filter bank to produce a filtered signal; Means for encoding the filtered signal to produce a compressed output signal, wherein the wavelet filter bank is a non-uniform filter bank of a tree structure, the first filter bank is independent from the wavelet filter bank, And the switching occurs in response to the normality of the audio signal. The method of claim 14, And the first filter bank is a high frequency decomposition MDCT filter bank. The method of claim 15, And wherein said normality is determined as a function of perceptual entropy of said audio signal. An apparatus for encoding an audio signal, the apparatus comprising: Means for generating a plurality of noise thresholds as a function of the audio signal and frequency characteristics; Means for sampling the audio signal; Means for selectively filtering the sampled audio signal by switching between a first filter bank and a wavelet filter bank to produce a filtered signal, wherein the wavelet filter bank is a non-uniform filter bank of a tree structure and the first filter bank; Means for filtering, wherein one filter bank is independent of the wavelet filter bank and the switching occurs in response to a normality of the audio signal; Means for quantizing the filtered signal, wherein the coarseness of the quantization is controlled by the noise thresholds; Means for perceptually encoding the quantized signal.

Images