 |
Description  |
|
|
CROSS-REFERENCE TO RELATED APPLICATIONS AND MATERIALS
The following U.S. patent applications filed concurrently with the present application and assigned to the assignee of the present application are related to the present application and each is hereby incorporated herein as if set forth in its
entirety: "RATE LOOP PROCESSOR FOR PERCEPTUAL ENCODER/DECODER", by J. D. Johnston; "A METHOD AND APPARATUS FOR CODING AUDIO SIGNALS BASED ON PERCEPTUAL MODEL," by J. D. Johnston; and "AN ENTROPY CODER," by J. D. Johnston and J. A. Reeds.
FIELD OF THE INVENTION
The present invention relates to processing of information signals, and more particularly, to the efficient encoding and decoding of monophonic and stereophonic audio signals, including signals representative of voice and music information, for
storage or transmission.
BACKGROUND OF THE INVENTION
Consumer, industrial, studio and laboratory products for storing, processing and communicating high quality audio signals are in great demand. For example, so-called compact disc ("CD") and digital audio tape ("DAT") recordings for music have
largely replaced the long-popular phonograph record and cassette tape. Likewise, recently available digital audio tape ("DAT") recordings promise to provide greater flexibility and high storage density for high quality audio signals. See, also, Tan and
Vermeulen, "Digital audio tape for data storage", IEEE Spectrum, pp. 34-38 (Oct. 1989). A demand is also arising for broadcast applications of digital technology that offer CD-like quality.
While these emerging digital techniques are capable of producing high quality signals, such performance is often achieved only at the expense of considerable data storage capacity or transmission bandwidth. Accordingly, much work has been done
in an attempt to compress high quality audio signals for storage and transmission.
Most of the prior work directed to compressing signals for transmission and storage has sought to reduce the redundancies that the source of the signals places on the signal. Thus, such techniques as ADPCM, sub-band coding and transform coding
described, e.g., in N. S. Jayant and P. Noll, "Digital Coding of Waveforms," Prentice-Hall, Inc. 1984, have sought to eliminate redundancies that otherwise would exist in the source signals.
In other approaches, the irrelevant information in source signals is sought to be eliminated using techniques based on models of the human perceptual system. Such techniques are described, e.g., in E. F. Schroeder and J. J. Platte, "`MSC`:
Stereo Audio Coding with CD-Quality and 256 kBIT/SEC," IEEE Trans. on Consumer Electronics, Vol. CE-33, No. 4, November 1987; and Johnston, Transform Coding of Audio Signals Using Noise Criteria, Vol. 6, No. 2, IEEE J.S.C.A. (Feb. 1988).
Perceptual coding, as described, e.g., in the Johnston paper relates to a technique for lowering required bitrates (or reapportioning available bits) or total number of bits in representing audio signals. In this form of coding, a masking
threshold for unwanted signals is identified as a function of frequency of the desired signal. Then, inter alia, the coarseness of quantizing used to represent a signal component of the desired signal is selected such that the quantizing noise
introduced by the coding does not rise above the noise threshold, though it may be quite near this threshold. The introduced noise is therefore masked in the perception process. While traditional signal-to-noise ratios for such perceptually coded
signals may be relatively low, the quality of these signals upon decoding, as perceived by a human listener, is nevertheless high.
Brandenburg et at, U.S. Pat. No. 5,040,217, issued Aug. 13, 1991, describes a system for efficiently coding and decoding high quality audio signals using such perceptual considerations. In particular, using a measure of the "noise-like" or
"tone-like" quality of the input signals, the embodiments described in the lauer system provides a very efficient coding for monophonic audio signals.
It is, of course, important that the coding techniques used to compress audio signals do not themselves introduce offensive components or artifacts. This is especially important when coding stereophonic audio information where coded information
corresponding to one stereo channel, when decoded for reproduction, can interfere or interact with coding information corresponding to the other stereo channel. Implementation choices for coding two stereo channels include so-called "dual mono" coders
using two independent coders operating at fixed bit rates. By contrast, "joint mono" coders use two monophonic coders but share one combined bit rate, i.e., the bit rate for the two coders is constrained to be less than or equal to a fixed rate, but
trade-offs can be made between the bit rates for individual coders. "Joint stereo" coders are those that attempt to use interchannel properties for the stereo pair for realizing additional coding gain.
It has been found that the independent coding of the two channels of a stereo pair, especially at low bit-rates, can lead to a number of undesirable psychoacoustic artifacts. Among them are those related to the localization of coding noise that
does not match the localization of the dynamically imaged signal. Thus the human stereophonic perception process appears to add constraints to the encoding process if such mismatched localization is to be avoided. This finding is consistent with
reports on binaural masking-level differences that appear to exist, at least for low frequencies, such that noise may be isolated spatially. Such binaural masking-level differences are considered to unmask a noise component that would be masked in a
monophonic system. See, for example, B. C. J. Morre, "An Introduction to the Psychology of Hearing, Second Edition," especially chapter 5, Academic Press, Orlando, Fla., 1982.
One technique for reducing psychoacoustic artifacts in the stereophonic context employs the ISO-WG11-MPEG-Audio Psychoacoustic II [ISO] Model. In this model, a second limit of signal-to-noise ratio "SNR") is applied to signal-to-noise ratios
inside the psychoacoustic model. However, such additional SNR constraints typically require the expenditure of additional channel capacity or (in storage applications) the use of additional storage capacity, at low frequencies, while also degrading the
monophonic performance of the coding.
SUMMARY OF THE INVENTION
Limitations of the prior art are overcome and a technical advance is made in a method and apparatus for coding a stereo pair of high quality audio channels in accordance with aspects of the present invention. Interchannel redundancy and
irrelevancy are exploited to achieve lower bit-rates while maintaining high quality reproduction after decoding. While particularly appropriate to stereophonic coding and decoding, the advantages of the present invention may also be realized in
conventional dual monophonic stereo coders.
An illustrative embodiment of the present invention employs a filter bank architecture using a Modified Discrete Cosine Transform (MDCT). In order to code the full range of signals that may be presented to the system, the illustrative embodiment
advantageously uses both L/R (Left and Right) and M/S (Sum/Difference) coding, switched in both frequency and time in a signal dependent fashion. A new stereophonic noise masking model advantageously detects and avoids binaural artifacts in the coded
stereophonic signal. Interchannel redundancy is exploited to provide enhanced compression for without degrading audio quality.
The time behavior of both Right and Left audio channels is advantageously accurately monitored and the results used to control the temporal resolution of the coding process. Thus, in one aspect, an illustrative embodiment of the present
invention, provides processing of input signals in terms of either a normal MDCT window, or, when signal conditions indicate, shorter windows. Further, dynamic switching between RIGHT/LEFt or SUM/DIFFERENCE coding modes is provided both in time and
frequency to control unwanted binaural noise localization, to prevent the need for overcoding of SUM/DIFFERENCE signals, and to maximize the global coding gain.
A typical bitstream definition and rate control loop are described which provide useful flexibility in forming the coder output. Interchannel irrelevancies, are advantageously eliminated and stereophonic noise masking improved, thereby to
achieve improved reproduced audio quality in jointly coded stereophonic pairs. The rate control method used in an illustrative embodiment uses an interpolation between absolute thresholds and masking threshold for signals below the rate-limit of the
coder, and a threshold elevation strategy under rate-limited conditions.
In accordance with an overall coder/decoder system aspect of the present invention, it proves advantageously to employ an improved Huffman-like entropy coder/decoder to further reduce the channel bit rate requirements, or storage capacity for
storage applications. The noiseless compression method illustratively used employs Huffman coding along with a frequency-partitioning scheme to efficiently code the frequency samples for L, R, M and S, as may be dictated by the perceptual threshold.
The present invention provides a mechanism for determining the scale factors to be used in quantizing the audio signal (i.e., the MDCT coefficients output from the analysis filter bank) by using an approach different from the prior art, and while
avoiding many of the restrictions and costs of prior quantizer/rate-loops. The audio signals quantized pursuant to the present invention introduce less noise and encode into fewer bits than the prior art.
These results are obtained in an illustrative embodiment of the present invention whereby the utilized scale factor, is iteratively derived by interpolating between a scale factor derived from a calculated threshold of hearing at the frequency
corresponding to the frequency of the respective spectral coefficient to be quantized and a scale factor derived from the absolute threshold of hearing at said frequency until the quantized spectral coefficients can be encoded within permissible limits.
BRIEF DESCRIPTION OF THE DRAWINGS
FIG. 1 presents an illustrative prior art audio communication/storage system of a type in which aspects of the present invention find application, and provides improvement and extension.
FIG. 2 presents an illustrative perceptual audio coder (PAC) in which the advances and teachings of the present invention find application, and provide improvement and extension.
FIG. 3 shows a representation of a useful masking level difference factor used in threshold calculations.
FIG. 4 presents an illustrative analysis filter bank according to an aspect of the present invention.
FIG. 5(a) through 5(e) illustrate the operation of various window functions.
FIG. 6 is a flow chart illustrating window switching functionality.
FIG. 7 is a block/flow diagram illustrating the overall processing of input signals to derive the output bitstream.
FIG. 8 illustrates certain threshold variations.
FIG. 9 is a flowchart representation of certain bit allocation functionality.
FIG. 10 shows bitstream organization.
FIGS. 11a through 11c illustrate certain Huffman coding operations.
FIG. 12 shows operations at a decoder that are complementary to those for an encoder.
FIG. 13 is a flowchart illustrating certain quantization operations in accordance with an aspect of the present invention.
FIG. 14(a) through 14(g) are illustrative windows for use with the filter bank of FIG. 4.
DETAILED DESCRIPTION
1. Overview
To simplify the present disclosure, the following patents, patent applications and publications are hereby incorporated by reference in the present disclosure as if fully set forth herein: U.S. Pat. No. 5,040,217, issued Aug. 13, 1991 by K.
Brandenburg et al, U.S. patent application Ser. No. 07/292,598, entitled Perceptual Coding of Audio Signals, filed Dec. 30, 1988; J. D. Johnston, Transform Coding of Audio Signals Using Perceptual Noise Criteria, IEEE Journal on Selected Areas in
Communications, Vol. 6, No. 2 (Feb. 1988); International Patent Application (PCT) WO 88/01811, filed Mar. 10, 1988; U.S. patent application Ser. No. 07/491,373, entitled Hybrid Perceptual Coding, filed Mar. 9, 1990, Brandenburg et al, Aspec:
Adaptive Spectral Entropy Coding of High Quality Music Signals, AES 90th Convention (1991); Johnston, J., Estimation of Perceptual Entropy Using Noise Masking Criteria, ICASSP, (1988); J. D. Johnston, Perceptual Transform Coding of Wideband Stereo
Signals, ICASSP (1989); E. F. Schroeder and J. J. Platte, "`MSC`: Stereo Audio Coding with CD-Quality and 256 kBIT/SEC,"IEEE Trans. on Consumer Electronics, Vol. CE-33, No. 4, November 1987; and Johnston, Transform Coding of Audio Signals Using Noise
Criteria, Vol. 6, No. 2, IEEE J.S.C.A. (February 1988).
For clarity of explanation, the illustrative embodiment of the present invention is presented as comprising individual functional blocks (including functional blocks labeled as "processors"). The functions these blocks represent may be provided
through the use of either shared or dedicated hardware, including, but not limited to, hardware capable of executing software. (Use of the term "processor" should not be construed to refer exclusively to hardware capable of executing software.)
Illustrative embodiments may comprise digital signal processor (DSP) hardware, such as the AT&T DSP16 or DSP32C, and software performing the operations discussed below. Very large scale integration (VLSI) hardware embodiments of the present invention,
as well as hybrid DSP/VLSI embodiments, may also be provided.
FIG. 1 is an overall block diagram of a system useful for incorporating an illustrative embodiment of the present invention. At the level shown, the system of FIG. 1 illustrates systems known in the prior an, but modifications, and extensions
described herein will make clear the contributions of the present invention. In FIG. 1, an analog audio signal 101 is fed into a preprocessor 102 where it is sampled (typically at 48 KHz) and convened into a digital pulse code modulation ("PCM") signal
103 (typically 16 bits) in standard fashion. The PCM signal 103 is fed into a perceptual audio coder 104 ("PAC") which compresses the PCM signal and outputs the compressed PAC signal to a communications channel/storage medium 105. From the
communications channel/storage medium the compressed PAC signal as fed into a perceptual audio decoder 107 which decompresses the compressed PAC signal and outputs a PCM signal 108 which is representative of the compressed PAC signal. From the
perceptual audio decoder, the PCM signal 108 is fed into a post-processor 109 which creates an analog representation of the PCM signal 108.
An illustrative embodiment of the perceptual audio coder 104 is shown in block diagram form in FIG. 2. As in the case of the system illustrated in FIG. 1, the system of FIG. 2, without more, may equally describe certain prior an systems, e.g.,
the system disclosed in the Brandenburg, et al U.S. Pat. No. 5,040,217. However, with the extensions and modifications described herein, important new results are obtained. The perceptual audio coder of FIG. 2 may advantageously be viewed as
comprising an analysis filter bank 202, a perceptual model processor 204, a quantizer/rate-loop processor 206 and an entropy coder 208.
The filter bank 202 in FIG. 2 advantageously transforms an input audio signal in time/frequency in such manner as to provide both some measure of signal processing gain (i.e. redundancy extraction) and a mapping of the filter bank inputs in a way
that is meaningful in light of the human perceptual system. Advantageously, the well-known Modified Discrete Cosine Transform (MDCT) described, e.g., in J. P. Princen and A. B. Bradley, "Analysis/Synthesis Filter Bank Design Based on Time Domain
Aliasing Cancellation," IEEE Trans. ASSP, Vol. 34, No. 5, October, 1986, may be adapted to perform such transforming of the input signals.
Features of the MDCT that make it useful in the present context include its critical sampling characteristic, i.e. for every n samples into the filter bank, n samples are obtained from the filter bank. Additionally, the MDCT typically provides
half-overlap, i.e. the transform length is exactly twice the length of the number of samples, n, shifted into the filterbank. The half-overlap provides a good method of dealing with the control of noise injected independently into each filter tap as
well as providing a good analysis window frequency response. In addition, in the absence of quantization, the MDCT provides exact reconstruction of the input samples, subject only to a delay of an integral number of samples.
One aspect in which the MDCT is advantageously modified for use in connection with a highly efficient stereophonic audio coder is the provision of the ability to switch the length of the analysis window for signal sections which have strongly
non-stationary components in such a fashion that it retains the critically sampled and exact reconstruction properties. describes a filter bank appropriate for performing the functions of element 202 in FIG. 2.
The perceptual model processor 204 shown in FIG. 2 calculates an estimate of the perceptual importance, noise masking properties, or just noticeable noise floor of the various signal components in the analysis bank. Signals representative of
these quantifies are then provided to other system elements to provide improved control of the filtering operations and organizing of the data to be sent to the channel or storage medium. Rather than using the critical band by critical band analysis
described in J. D. Johnston, "Transform Coding of Audio Signals Using Perceptual Noise Criteria," IEEE J. on Selected Areas in Communications, February 1988, an illustrative embodiment of the present invention advantageously uses finer frequency
resolution in the calculation of thresholds. Thus instead of using an overall tonality metric as in the last-cited Johnston paper, a tonality method, e.g. one based on that mentioned in K. Brandenburg and J. D. Johnston, "Second Generation Perceptual
Audio Coding: The Hybrid Coder," AES 89th Convention, 1990 provides a tonality estimate that varies over frequency, thus providing a better fit for complex signals.
The psychoacoustic analysis performed in the perceptual model processor 204 provides a noise threshold for the L (Left), R (Right), M (Sum) and S (Difference) channels, as may be appropriate, for both the normal MDCT window and the shorter
windows. Use of the shorter windows is advantageously controlled entirely by the psychoacoustic model processor.
In operation, an illustrative embodiment of the perceptual model processor 204 evaluates thresholds for the left and fight channels, denoted THR.sub.i and THR.sub.r. The two thresholds are then compared in each of the illustrative 35 coder
frequency partitions (56 partitions in the case of an active window-switched block). In each partition where the two thresholds vary between left and right by less than some amount, typically 2dB, the coder is switched into M/S mode. That is, the left
signal for that band of frequencies is replaced by M=(L+R)/2, and the right signal is replaced by S=(L-R)/2. The actual amount of difference that triggers the last-mentioned substitution will vary with bitrate constraints and other system parameters.
The same threshold calculation used for L and R thresholds is also used for M and S thresholds, with the threshold calculated on the actual M and S signals. First, the basic thresholds, denoted BTHR.sub.m and MLD.sub.s are calculated. Then, the
following steps are used to calculate the stereo masking contribution of the M and S signals.
1. An additional factor is calculated for each of the M and S thresholds. This factor, called MLD.sub.m, and MLD.sub.s, is calculated by multiplying the spread signal energy, (as derived, e.g., in J. D. Johnston, "Transform Coding of Audio
Signals Using Perceptual Noise Criteria," IEEE J. on Selected Areas in Communications, February 1988; K. Brandenburg and J. D. Johnston, "Second Generation Perceptual Audio Coding: The Hybrid Coder," AES 89th Convention, 1990; and Brandenburg, et al U.S. Pat. No. 5,040,217) by a masking level difference factor shown illustratively in FIG. 3. This calculates a second level of detectability of noise across frequency in the M and S channels, based on the masking level differences shown in various sources.
2. The actual threshold for M (THR.sub.m) is calculated as THR.sub.m =max(BTHR.sub.m, min(BTHR.sub.s,MLD.sub.s)) and the threshold m=max(BTHR.sub.m, min(BTHR.sub.s,MLD.sub.s)) and the threshold for S is calculated as THR.sub.s =max
(BTHR.sub.s,min(BTHR.sub.m, MLD.sub.m)).
In effect, the MLD signal substitutes for the BTHR signal in cases where there is a chance of stereo unmasking. It is not necessary to consider the issue of M and S threshold depression due to unequal L and R thresholds, because of the fact that
L and R thresholds are known to be equal.
The quantizer and rate control processor 206 used in the illustrative coder of FIG. 2 takes the outputs from the analysis bank and the perceptual model, and allocates bits, noise, and controls other system parameters so as to meet the required
bit rate for the given application. In some example coders this may consist of nothing more than quantization so that the just noticeable difference of the perceptual model is never exceeded, with no (explicit) attention to bit rate; in some coders this
may be a complex set of iteration loops that adjusts distortion and bitrate in order to achieve a balance between bit rate and coding noise. A particularly useful quantizer and rate control processor is described in incorporated U.S. patent application
by J. D. Johnston, entitled "RATE LOOP PROCESSOR FOR PERCEPTUAL ENCODER/DECODER," (hereinafter referred to as the "rate loop application") filed of even date with the present application. Also desirably performed by the rate loop processor 206, and
described in the rate loop application, is the function of receiving information from the quantized analyzed signal and any requisite side information, inserting synchronization and framing information. Again, these same functions are broadly described
in the incorporated Brandenburg, et al, U.S. Pat. No. 5,040,217.
Entropy coder 208 is used to achieve a further noiseless compression in cooperation with the rate control processor 206. In particular, entropy coder 208, in accordance with another aspect of the present invention, advantageously receives inputs
including a quantized audio signal output from quantizer/rate-loop 206, performs a lossless encoding on the quantized audio signal, and outputs a compressed audio signal to the communications channel/storage medium 106.
Illustrative entropy coder 208 advantageously comprises a novel variation of the minimum-redundancy Huffman coding technique to encode each quantized audio signal. The Huffman codes are described, e.g., in D. A. Huffman, "A Method for the
Construction of Minimum Redundancy Codes", Proc. IRE, 40:1098-1101 (1952) and T. M. Cover and J. A. Thomas, .us Elements of Information Theory, pp. 92-101 (1991). The useful adaptations of the Huffman codes advantageously used in the context of the
coder of FIG. 2 are described in more detail in the incorporated U.S. Pat. No. 5,227,788 by J. D. Johnston and J. Reeds (hereinafter the "entropy coder application") filed of even date with the present application and assigned to the assignee of this
application. Those skilled in the data communications arts will readily perceive how to implement alternative embodiments of entropy coder 208 using other noiseless data compression techniques, including the well-known Lempel-Ziv compression methods.
The use of each of the elements shown in FIG. 2 will be described in greater detail in the context of the overall system functionality; details of operation will be provided for the perceptual model processor 204.
2.1. The Analysis Filter Bank
The analysis filter bank 202 of the perceptual audio coder 104 receives as input pulse code modulated ("PCM") digital audio signals (typically 16-bit signals sampled at 48 KHz), and outputs a representation of the input signal which identifies
the individual frequency components of the input signal. For example, an output of the analysis filter bank 202 comprises a Modified Discrete Cosine Transform ("MDCT") of the input signal. See, J. Princen et at, "Sub-band Transform Coding Using Filter
Bank Designs Based on Time Domain Aliasing Cancellation," IEEE ICASSP, pp. 2161-2164 (1987).
An illustrative analysis filter bank 202 according to one aspect of the present invention is presented in FIG. 4. Analysis filter bank 202 comprises an input signal buffer 302, a window multiplier 304, a window memory 306, an FFT processor 308,
an MDCT processor 310, a concatenator 311, a delay memory 312 and a data selector 132.
The analysis filter bank 202 operates on frames. A frame is conveniently chosen as the 2N PCM input audio signal samples held by input signal buffer 302. As stated above, each PCM input audio signal sample is represented by M bits.
Illustratively, N=512 and M=16.
Input signal buffer 302 comprises two sections: a first section comprising N samples in buffer locations 1 to N, and a second section comprising N samples in buffer locations N+1 to 2N. Each frame to be coded by the perceptual audio coder 104 is
defined by shifting N consecutive samples of the input audio signal into the input signal buffer 302. Older samples are located at higher buffer locations than newer samples.
Assuming that, at a given time, the input signal buffer 302 contains a frame of 2N audio signal samples, the succeeding frame is obtained by (1) shifting the N audio signal samples in buffer locations 1 to N into buffer locations N+1 to 2N,
respectively, (the previous audio signal samples in locations N+1 to 2N may be either overwritten or deleted), and (2) by shifting into the input signal buffer 302, at buffer locations 1 to N, N new audio signal samples from preprocessor 102. Therefore,
it can be seen that consecutive frames contain N samples in common: the first of the consecutive frames having the common samples in buffer locations 1 to N, and the second of the consecutive frames having the common samples in buffer locations N+1 to
2N. Analysis filter bank 202 is a critically sampled system (i.e., for every N audio signal samples received by the input signal buffer 302, the analysts filter bank 202 outputs a vector of N scalers to the quantizer/rate-loop 206).
Each frame of the input audio signal is provided to the window multiplier 304 by the input signal buffer 302 so that the window multiplier 304 may apply seven distinct data windows to the frame. Each data window is a vector of scalers called
"coefficients". While all seven of the data windows have 2N coefficients (i.e., the same number as there are audio signal samples in the frame), four of the seven only have N/2 non-zero coefficients (i.e., one-fourth the number of audio signal samples
in the frame). As is discussed below, the data window coefficients may be advantageously chosen to reduce the perceptual entropy of the output of the MDCT processor 310.
The information for the data window coefficients is stored in the window memory 306. The window memory 306 may illustratively comprise a random access memory ("RAM"), read only memory ("ROM"), or other magnetic or optical media. Drawings of
seven illustrative data windows, as applied by window multiplier 304, are presented in FIG. 4. Typical vectors of coefficients for each of the seven data windows presented in FIG. 4 are presented in Appendix A. As may be seen in both FIG. 4 and in
Appendix A, some of the data window coefficients may be equal to zero.
Keeping in mind that the data window is a vector of 2N scalers and that the audio signal frame is also a vector of 2N scalers, the data window coefficients are applied to the audio signal frame scalers through point-to-point multiplication (i.e.,
the first audio signal frame scaler is multiplied by the first data window coefficient, the second audio signal frame scaler is multiplied by the second data window coefficient, etc.). Window multiplier 304 may therefore comprise seven microprocessors
operating in parallel, each performing 2N multiplications in order to apply one of the seven data window to the audio signal frame held by the input signal buffer 302. The output of the window multiplier 304 is seven vectors of 2N scalers to be referred
to as "windowed frame vectors".
The seven windowed frame vectors are provided by window multiplier 304 to FFT processor 308. The FFF processor 308 performs an odd-frequency FFT on each of the seven windowed frame vectors. The odd-frequency FFT is an Discrete Fourier Transform
evaluated at frequencies: ##EQU1## where k=1, 3, 5, . . . ,2N, and f.sub.H equals one half the sampling rate. The illustrative FFT processor 308 may comprise seven conventional decimation-in-time FFT processors operating in parallel, each operating on
a different windowed frame vector. An output of the FFT processor 308 is seven vectors of 2N complex elements, to be referred to collectively as "FFT vectors".
FFT processor 308 provides the seven FFT vectors to both the perceptual model processor 204 and the MDCT processor 310. The perceptual model processor 204 uses the FFT vectors to direct the operation of the data selector 314 and the
quantizer/rate-loop processor 206. Details regarding the operation of data selector 314 and perceptual model processor 204 are presented below.
MDCT processor 310 performs an MDCT based on the real components of each of the seven FFF vectors received from FFF processor 308. .P MDCT processor 310 may comprise seven microprocessors operating in parallel. Each such microprocessor
determines one of the seven "MDCT vectors" of N real scalars based on one of the seven respective FFT vectors. For each FFT vector, F(k), the resulting MDCT vector, X(k), is formed as follows: ##EQU2## The procedure need run k only to N, not 2N, because
of redundancy in the result. To wit, for N<k.ltoreq.2N:
MDCT processor 310 provides the seven MDCT vectors to concatenator 311 and delay memory 312.
As discussed above with reference to window multiplier 304, four of the seven data windows have N/2 non-zero coefficients (see FIG. 14c-f). This means that four of the windowed frame vectors contain only N/2 non-zero values. Therefore, the
non-zero values of these four vectors may be concatenated into a single vector of length 2N by concatenator 311 upon output from MDCF processor 310. The resulting concatenation of these vectors is handled as a single vector for subsequent purposes.
Thus, delay memory 312 is presented with four MDCT vectors, rather than seven.
Delay memory 312 receives the four MDCT vectors from MDCT processor 310 and concatenator 311 for the purpose of providing temporary storage. Delay memory 312 provides a delay of one audio signal frame (as defined by input signal buffer 302) on
the flow of the four MDCT vectors through the filter bank 202. The delay is provided by (i) storing the two most recent consecutive sets of MDCT vectors representing consecutive audio signal frames and (ii) presenting as input to data selector 314 the
older of the consecutive sets of vectors. Delay memory 312 may comprise random access memory (RAM) of size:
where 2 is the number of consecutive sets of vectors, 4 is the number of vectors in a set, N is the number of elements in an MDCT vector, and M is the number of bits used to represent an MDCT vector element.
Data selector 314 selects one of the four MDCT vectors provided by delay memory 312 to be output from the filter bank 202 to quantizer/rate-loop 206. As mentioned above, the perceptual model processor 204 directs the operation of data selector
314 based on the FFT vectors provided by the FFT processor 308. Due to the operation of delay memory 312, the seven FFT vectors provided to the perceptual model processor 204 and the four MDCT vectors concurrently provided to data selector 314 are not
based on the same audio input frame, but rather on two consecutive input signal frames--the MDCT vectors based on the earlier of the frames, and the FFT vectors based on the later of the frames. Thus, the selection of a specific MDCT vector is based on
information contained in the next successive audio signal frame. The criteria according to which the perceptual model processor 204 directs the selection of an MDCT vector is described in Section 2.2, below.
For purposes of an illustrative stereo embodiment, the above analysis filterbank 202 is provided for each of the left and right channels.
2.2. The Perceptual Model Processor
A perceptual coder achieves success in reducing the number of bits required to accurately represent high quality audio signals, in part, by introducing noise associated with quantization of information bearing signals, such as the MDCT
information from the filter bank 202. The goal is, of course, to introduce this noise in an imperceptible or benign way. This noise shaping is primarily a frequency analysis instrument, so it is convenient to convert a signal into a spectral
representation (e.g., the MDCT vectors provided by filter bank 202), compute the shape and amount of the noise that will be masked by these signals and injecting it by quantizing the spectral values. These and other basic operations are represented in
the structure of the perceptual coder shown in FIG. 2.
The perceptual model processor 204 of the perceptual audio coder 104 illustratively receives its input from the analysis filter bank 202 which operates on successive frames. The perceptual model processor inputs then typically comprise seven
Fast Fourier Transform (FFT) vectors from the analysis filter bank 202. These are the outputs of the FFT processor 308 in the form of seven vectors of 2N complex elements, each corresponding to one of the windowed frame vectors.
In order to mask the quantization noise by the signal, one must consider the spectral contents of the signal and the duration of a particular spectral pattern of the signal. These two aspects are related to masking in the frequency domain where
signal and noise are approximately steady state--given the integration period of the heating system--and also with masking in the time domain where signal and noise are subjected to different cochlear filters. The shape and length of these filters are
frequency dependent.
Masking in the frequency domain is described by the concept of simultaneous masking. Masking in the time domain is characterized by the concept of premasking and postmasking. These concepts are extensively explained in the literature; see, for
example, E. Zwicker and H. Fastl, "Psychoacoustics, Facts, and Models," Springer-Verlag, 1990. To make these concepts useful to perceptual coding, they are embodied in different ways.
Simultaneous masking is evaluated by using perceptual noise shaping models. Given the spectral contents of the signal and its description in terms of noise-like or tone-like behavior, these models produce an hypothetical masking threshold that
rules the quantization level of each spectral component. This noise shaping represents the maximum amount of noise that may be introduced in the original signal without causing any perceptible difference. A measure called the PERCEPTUAL ENTROPY (PE)
uses this hypothetical masking threshold to estimate the theoretical lower bound of the bitrate for transparent encoding. See, e.g., J. D. Johnston, Estimation of Perceptual Entropy Using Noise Masking Criteria,"ICASSP, 1989.
Premasking characterizes the (in)audibility of a noise that starts some time before the masker signal which is louder than the noise. The noise amplitude must be more attenuated as the delay increases. This attenuation level is also frequency
dependent. If the noise is the quantization noise attenuated by the first half of the synthesis window, experimental evidence indicates the maximum acceptable delay to be about 1 millisecond.
This problem is very sensitive and can conflict directly with achieving a good coding gain. Assuming stationary conditions--which is a false premiss--The coding gain is bigger for larger transforms, but, the quantization error spreads till the
beginning of the reconstructed time segment. So, if a transform length of 1024 points is used, with a digital signal sampled at a rate of 48000 Hz, the nose will appear at most 21 milliseconds before the signal. This scenario is particularly critical
when the signal takes the form of a sharp transient in the time domain commonly known as an "attack". In this case the quantization noise is audible before the attack. The effect is known as pre-echo.
Thus, a fixed length filter bank is a not a good perceptual solution nor a signal processing solution for non-stationary regions of the signal. It will be shown later that a possible way to circumvent this problem is to improve the temporal
resolution of the coder by reducing the analysis/synthesis window length. This is implemented as a window switching mechanism when conditions of attack are detected. In this way, the coding gain achieved by using a long analysis/synthesis window will
be affected only when such detection occurs with a consequent need to switch to a shorter analysis/synthesis window.
Postmasking characterizes the (in)audibility of a noise when it remains after the cessation of a stronger masker signal. In this case the acceptable delays are in the order of 20 milliseconds. Given that the bigger transformed time segment
lasts 21 milliseconds (1024 samples), no special care is needed to handle this situation.
WINDOW SWITCHING
The PERCEPTUAL ENTROPY (PE) measure of a particular transform segment gives the theoretical lower bound of bits/sample to code that segment transparently. Due to its memory properties, which are related to premasking protection, this measure
shows a significant increase of the PE value to its previous value--related with the previous segment--when some situations of strong non-stationarity of the signal (e.g. an attack) are presented. This important property is used to activate the window
switching mechanism in order to reduce pre-echo. This window switching mechanism is not a new strategy, having been used, e.g., in the ASPEC coder, described in the ISO/MPEG Audio Coding Report, 1990, but the decision technique behind it is new using
the PE information to accurately localize the non-stationarity and define the fight moment to operate the switch.
Two basic window lengths: 1024 samples and 256 samples are used. The former corresponds to a segment duration of about 21 milliseconds and the latter to a segment duration of about 5 milliseconds. Short windows are ass | | |
