Code-Excited Linear Predictive Coding With Low Delay for Speech or Audio Signals

This invention relates to digital communications, and more particularly to digital coding of speech or audio signals with low coding delay and high-fidelity at reduced bit-rates.

This application is related to subject matter disclosed in U.S. patent application Ser. No. 07/298451, by J-H Chen, filed Jan. 17, 1989, now abandoned, and copending U.S. patent application Ser. No. 07/757,168 by J-H Chen, filed Sep. 10, 1991, assigned to the assignee of the present application. Also related to the subject matter of this application is a copending application Ser. No., filed Feb. 18, 1992 by J-H Chen, R. Cox and N. Jayant entitled "Low Delay Code-Excited Linear Predictive Coder For Speech Or Audio Signals," which application is assigned to the assignee of the present application. Each of these patent applications is incorporated by reference in the present application as if set forth in its entirety herein.

The International Telegraph and Telephone Consultative Committee (CCITT), an international communications standards organization, has been developing a standard for 16 kb/s speech coding and decoding for universal applications. The standardization process included the issuance by the CCITT of a document entitled "Terms of Reference" prepared by the ad hoc group on 16 kbit/s speech coding (Annex 1 to question 21/XV), June 1988.

Presently, the candidate being considered for the standard is Low-Delay Code Excited Linear Predictive Coding (hereinafter, LD-CELP) described in substantial part in the incorporated application Ser. No. 07/298451. Aspects of this coder are also described in J-H Chen, "A robust low-delay CELP speech coder at 16 kbit/s, "Proc. GLOBECOM, pp. 1237-1241 (Nov. 1989); J-H Chen, "High-quality 16 kb/s speech coding with a one-way delay less than 2 ms, "Proc. ICASSP, pp. 453-456 (April 1990); J-H Chen, M. J. Melchner, R. V. Cox and D. O. Bowker, "Realtime implementation of a 16 kb/s low-delay CELP speech coder, "Proc. ICASSP, pp. 181-184 (April 1990); all of which papers are hereby incorporated herein by reference as if set forth in their entirety. The patent application Ser. No. 07/298,451 and the cited papers incorporated by reference describe aspects of the LD-CELP system as evaluated in Phase 1. Accordingly, the system described in these papers and the application Ser. No. 07/298,451 will be referred to generally as the Phase 1 System.

A document further describing the LD-CELP candidate standard system was presented in a document entitled "Draft Recommendation on 16 kbit/s Voice Coding," submitted to the CCITT Study Group XV in its meeting in Geneva, Switzerland during Nov. 11-22, 1991 (hereinafter, "Draft Recommendation"), which document is incorporated herein by reference in its entirety. For convenience, and subject to deletion as may appear desirable, part or all of the Draft Recommendation is also attached to this application as Appendix 1. The system described in the Draft Recommendation has been evaluated during Phase 2 of the CCITT standardization process, and will accordingly be referred to as the Phase 2 System. Other aspects of the Phase 2 System are also described in a document entitled "A fixed-point Architecture for the 16 kb/s LD-CELP Algorithm" (hereinafter, "Architecture Document") submitted by the assignee of the present application to a meeting of Study Group XV of the CCITT held in Geneva, Switzerland on Feb. 18 through Mar. 1, 1991. The Architecture Document is hereby incorporated by reference as if set forth in its entirety herein and a copy of that document is attached to this application for convenience as Appendix 2. Also incorporated by reference as descriptive of the Phase 2 System and J. H. Chen, Y. C. Lin, and R. V. Cox, "A fixed point 16 kb/s LD-CELP Algorithm," Proc. ICASSP, pp. 21-24, (May 1991).

In many signal processing applications, including speech and audio signal coding, it proves convenient to use part of a sequence of signals for selective processing. For example, a sequence of time signals, such as samples of a speech signal, will be processed in groups or subsequences. For this purpose, the notion of a "window" is typically used to define a current (or past) subsequence, with the particular values changing as the window is allowed to shift with evolving time. In a similar way, the notion of a spectral window is conveniently used for processing in the frequency domain. Other kinds of windows are used in different domains and for particular kinds of signal processing. Some of the commonly used windows are described in R. B. Blackman and J. W. Tukey, The Measurement of Power Spectra, Dover: New York, 1958; and N. C. Geckinli and D. Yavuz, "Some Novel Windows and a Concise Tutorial Comparison of Window Families," IEEE Trans. Acoustics, Speech and Signal Processing, Vol. ASSP-26, No. 6, December 1978, pp. 501-507. The application of spectral windows in the context of a speech synthesis system is described in Y. Tohkura and F. Itakura, "Spectral Smoothing Techniques in PARCOR Speech Analysis-Synthesis," IEEE Trans. on Acoustics, Speech, and Signal Processing, Vol. ASSP-26, No. 6, December 1978. Also attached as Appendix 3 is a descriptive of the Phase 2 system as updated in accordance with the present invention.

In the past, the CCITT has only standardized fixed-point speech encodings. One principle reason for this was that floating-point processors were either unnecessary or unavailable at the time the standards were proposed. Another reason is that it is relatively easy to fully specify an algorithm with fixed-point arithmetic, a so-called bit-exact specification. By contrast, a floating-point specification may have difficulty with specific arithmetic precision, especially as implemented on a variety of hardware platforms. Therefore, with a fixed-point specification, test vectors can be used to verify conformance of a particular codec with the standard, while this would be much more difficult for floating-point specifications. A third reason is that fixed-point implementations usually result in lower cost and lower power consumption than floating-point implementations. In addition, a fixed-point specification facilitates VLSI implementations.

The LD-CELP system, in common with many linear predictive coding (LPC) arrangements, uses sets of autocorrelation coefficients to derive the LPC predictor coefficients used in updating the various adaptive elements of the system (i.e., gain predictor and LPC synthesis filter). See the documents describing the Phase 1 System cited above. The autocorrelation coefficients, in turn, are formed using windowed values of respective Phase 1 System signal sequences. In particular, the recursive windowing method described in T. P. Barnwell, III, "Recursive windowing for generating autocorrelation coefficients for LPC analysis," IEEE Trans. Acoust., Speech, Signal Processing, Vol. ASSP-29(5), pp. 1062-1066, October 1981, is advantageously employed in forming the autocorrelation coefficients of the Phase 1 System.

For the reasons given above, it proves advantageous to implement a 16-bit fixed-point version of the LD-CELP algorithm. However, implementation of Barnwell's recursive windowing techniques proves difficult when using fixed-point processing. In part, this is because 16-bit fixed-point arithmetic generally does not provide enough precision for the 50-th order Durbin's recursion used in the Phase 1 System, nor does it have a sufficient dynamic range to handle the recursive windowing method used in the Phase 1 System in performing the autocorrelation functionality.

Another problem arising in the context of the Phase 1 System (and the Phase 2 System described in Appendices 1 and 2) is one related to decoding certain sustained speech patterns, such as sustained vowel sounds. While such troublesome speech patterns are rare, they can occur with some regularity when coding and decoding certain machine-generated speech having little of the natural variation with time that human speech typically possesses. In particular, it has been found that such sustained sounds can cause the adaptive LPC synthesis filter at a decoder to fail to accurately track the LPC synthesis filter at the encoder. This can cause temporary unsatisfactory reception of the decoded speech.

In accordance with aspects of illustrative embodiments of the present invention, a method and corresponding system are provided which effectively avoid impairments or limitations of prior coders and decoders and produces improved performance. These improvements and distinctions are all achieved in an illustrative embodiment featuring fixed-point processing within the low delay constraints sought in the CCITT standardization process.

Briefly, it has proven advantageous to replace the Barnwell recursive windowing method by a new hybrid windowing method which is partially recursive and partially non-recursive. This new method avoids the dynamic range problem and the more complex double-precision arithmetic that would otherwise have been required. In particular, the recursive window of the Phase 1 System is advantageously replaced by a novel hybrid window comprising a recursively decaying tail and a section of non-recursive samples at the beginning.

In accordance with another aspect of the present invention, the above-noted problem arising from some sustained vowel sounds has been avoided in an improved Phase 2 System by introducing a simple additional processing step before the 50th order Durbin's recursion employed in both the Phase 1 and Phase 2 Systems. Thus by modifying the magnitude of the autocorrelation coefficients developed from the modified windowed signals, the LPC coefficients developed by the Durbin recursion are found to avoid the narrow spectral peaks that contribute to the occasional anomalous behavior of the Phase 2 System when presented with the sometimes troublesome sustained vowel signals. The modifying of the autocorrelation coefficients conveniently forms a simple postprocessing step to the normal window processing. In fact, the modifying of the autocorrelation coefficients can advantageously accompany the prior modification of the power-related autocorrelation coefficient, r(0). That is, previously, the value of f(0) has been modified by a factor slightly greater than 1, e.g., 1.00390625, to, in effect, add white noise at a level well below the speech power to add stability to certain of the LD-CELP processes as described in the Draft Recommendation, for example. This multiplying then is then extended in accordance with the present invention to others of the correlation coefficients prior to deriving the LPC coefficients using Durbin's recursion or other suitable means.

These and other advances provided by the present invention are achieved, in an illustrative embodiment, in a speech coder in a low delay code excited linear predictive coding (LD-CELP) system of the type characterized above as the Phase 2 System.

FIGS. 1A and 1B are simplified block diagrams of a Phase 2 LD-CELP encoder and decoder, respectively, in accordance with an illustrative embodiment of the present invention.

FIG. 2 is a schematic block diagram of a Phase 2 LD-CELP encoder in accordance with an illustrative embodiment of the present invention.

FIG. 3 is a schematic block diagram of a Phase 2 LD-CELP decoder in accordance with an illustrative embodiment of the present invention.

FIG. 4A is a schematic block diagram of a perceptual weighting filter adapter for use in a Phase 2 System in accordance with an illustrative embodiment of the present invention.

FIG. 4B illustrates a hybrid window used in a Phase 2 System in accordance with an illustrative embodiment of the present invention.

FIG. 5 is a schematic block diagram of a backward synthesis filter adapter for use in a Phase 2 System in accordance with an illustrative embodiment of the present invention.

FIG. 6 is a schematic block diagram of a backward vector gain adapter for use in a Phase 2 System in accordance with an illustrative embodiment of the present invention.

FIG. 7 is a schematic block diagram of a postfilter for use in a Phase 2 System in accordance with an illustrative embodiment of the present invention.

FIG. 8 is a schematic block diagram of a postfilter adapter for use in a Phase 2 System in accordance with an illustrative embodiment of the present invention.

FIG. 9 is a schematic block diagram of a preprocessor to the Durbin recursion functionality of a Phase 2 System to avoid certain adverse affects arising from particular sustained speech or speech-like signals.

1. The above-cited Draft Recommendation describes the Phase 2 system in detail and should be referred to for additional information in making and using the present invention. FIGS. 1A and 1B correspond to FIG. 1 of the Draft Recommendation and FIGS. 2 through 8 correspond to identically numbered figures in the Draft Recommendation.

2. Review of floating-point LD-CELP

The original floating-point LD-CELP coder is shown in FIG. 1A. More details about this coder can be found in the Phase 1 documents identified above, including U.S. patent application Ser. No. 07/298451. Here only its main features are reviewed.

In this coder, both the gain 101 and the 50-th order LPC predictor 102 are backward-adaptive based on previously quantized signals, and only the excitation is coded and transmitted forward to the decoder. The input speech is coded vector-by-vector, where each vector illustratively contains 5 samples. Vector quantization (VQ) is used to encode each 5-dimensional excitation vector into 10 bits, resulting in a total bit-rate of 2 bits/sample, or 16 kb/s with a sampling rate of 8 kHz. The codebook search is done in a closed-loop, or "analysis-by-sythesis" manner typical to all CELP coders. See, e.g., M. R. Schroeder and B. S. Atal, "Code Excited Linear Prediction (CELP); high quality speech at very low bit rates, "Proc. ICASSP, pp. 937-940 (1985). The 50-th order LPC predictor is implemented as a direct-form transversal filter. The filter coefficients are backward adapted once every 4 vectors (20 samples) by performing LPC analysis on previously coded speech. The LD-CELP decoder performs the same LPC analysis as the encoder does, so there is no need to transmit LPC parameters. Similarly, the gain is also backward-adaptive. It is updated once every vector by using a 10-th order adaptive linear predictor in the logarithmic gain domain. The coefficients of this log-gain predictor are also updated once every 4 vectors by performing a similar LPC analysis on the logarithmic gains of previously quantized and scaled excitation vectors. The perceptual weighting filter is also of order 10, and its coefficients are also updated once every 4 vectors by LPC analysis, although the analysis is based on the input speech rather than the coded speech. The time period between predictor updates is considered a "frame" of LD-CELP. Thus, the "frame size" of LD-CELP is 20 samples, although the actual speech buffer size is only 5 samples.

In all three LPC analyses mentioned above, a modified version of Barnwell's recursive windowing method is first used to calculate the autocorrelation coefficients. Durbin's recursion (see, L. R. Rabiner and R. W. Shafer, Digital Processing of Speech Signals, Prentice-Hall, Inc., Englewood Cliffs, N.J. (1978)) is then used to convert the autocorrelation coefficients to LPC predictor coefficients.

3. Overview of fixed-point LD-CELP algorithm

The newly created fixed-point LD-CELP coder (the Phase 2 coder) is shown in FIG. 2. This coder is mostly the same as the original LD-CELP coder in FIG. 1 except that the recursive windowing method has been replaced by a hybrid windowing method. The changes will be described in detail in the following two sections.

4. Hybrid windowing method

In the original recursive windowing method, the products of the current speech sample and previous samples are passed through a bank of third-order IIR filters, and the autocorrelation coefficients are obtained at the outputs of these IIR filters. Since each speech sample is represented by 16 bits, the product of two speech samples has a dynamic range of 32 bits. Thus, to filter this product term, 32-bit by 32-bit multiplication and addition is required to fully preserve the precision. Such computation requires double-precision arithmetic in a 16-bit fixed-point DSP device. Since double-precision arithmetic generally takes significantly more DSP instruction cycles than single-precision arithmetic, and since autocorrelation computation is a significant portion of the total complexity of LD-CELP, implementing recursive windowing in double precision results in very high complexity.

To avoid double-precision arithmetic, an alternative is to use a conventional block-by-block, non-recursive windowing method with, for instance, a Hamming window or half Hamming window. See, e.g., T. Moriya, "Medium-delay 8 kbit/s speech coder based on conditional pitch prediction", Proc. Int. Conf. Spoken Language Processing (Nov. 1990). However, since our frame size of 20 samples is much smaller than the typical window size of 160 to 200 samples, this means a very significant window overlap and a very high computational complexity. In addition, it was found that Hamming windowing gave poorer prediction gain and perceptual speech quality than recursive windowing in the context of backward-adaptive LPC analysis. Therefore, it is desirable to at least keep the window shape similar to that of the recursive window.

The present invention provides a novel hybrid window which consists of a recursively decaying tail and a section of non-recursive samples at the beginning (see FIG. 4B). The tail of the window is exponentially decaying with a decaying factor .alpha. slightly less than unity. The non-recursive part of the window is a section of the sine function and it makes the shape of the entire window similar to that of the original recursive window. An example of such a hybrid window is shown in FIG. 4B. In the following, it will first be shown how to determine the window parameters, and then the procedure to calculate autocorrelation coefficients using this hybrid window will be described.

Let s(n) denote the signal for which we want to calculate the autocorrelation coefficients. To be general, let us assume that the signal samples corresponding to the current LD-CELP frame are s(m),s s(m+1), s(m+2), . . . , s(m+L-1). Then, for backward-adaptive LPC analysis, the hybrid window is applied to all signal samples with a time index less than m (as shown in FIG. 3). Let there be N non-recursive samples in the hybrid window function. Then, the signal samples s(m-1), s(m-2), . . . , s(m-N) are all weighted by the non-recursive portion of the window. Starting with s(m-N-1), all signal samples to the left of (and including) this sample are weighted by the recursive portion of the window, which has values b, b.alpha., b.alpha..sup.2, . . . , where 0<b<1 and 0<.alpha.<1.

At time m, the hybrid window function w.sub.m (n) is defined as ##EQU1##

To suppress the sidelobe of the Fourier transform of the window, a smooth junction between the sine function and the exponential function at n=m-N-1 is desired. Therefore, the following two continuity conditions are imposed: (1) the functions f.sub.m (n) and g.sub.m (n) have the same value at n=m-N-1, and (2) the slopes of these two function curves are also the same at n=m-N-1. From the first condition and Eq. (1), we have

The second condition yields

Substituting Eq. (2) into Eq. (3) gives ##EQU2##

In designing the hybrid window, the decaying factor .alpha. is first determined, based on how long the effective length of the exponential tail is to be. Then, N, the number of non-recursive samples, is determined based on how the initial part of the window is to be shaped and how much computational complexity can be accommodated by the processing systems. (The larger the number N, the higher the complexity.) Once the parameters .alpha. and N are determined, the only unknown in Eq. (4) is the constant c.

Since Eq. (4) is a non-linear equation on c, it is not convenient to directly solve for c. However, a very accurate solution can be obtained by using iterative approximation techniques. From FIG. 4B and Eq. (2), it should be clear that the desired range for c(N+1) is between .pi./2 and .pi.. Note that -ccot[c(N+1)] is zero at c(N+1)=.pi./2, and its value monotonically increases and finally approaches infinity as c(N+1) increases and approach .pi.. Also note that -ln.alpha. is a small positive constant. Therefore, the two curves y(c)=-ccot[c(N+1)] and y(c)=-ln.alpha. always have a unique intersection in the range of .pi./2<c(N+1)<.pi.. It was found that for an initial step size of .pi./8 and an initial guess of 3.pi./4 for c(N+1), and if the step size is reduced by half every time the intersection point is "crossed over" while searching for it, then usually within 20 iterations the two sides of Eq. (4) to agree for at least 5 decimal digits. Once the value of c is found, the value of b is easily obtained by using Eq. (2). Note that this iterative method to find c and b is done only once during the coder design stage.

To describe the way to calculate autocorrelation coefficients using the hybrid window, let us define the window-weighted signal for the current frame (starting at time m) to be ##EQU3## For an M-th order LPC analysis, we need to calculate the autocorrelation coefficients R.sub.m (i) for i=0, 1, 2, . . . , M. The i-th autocorrelation coefficient for the current frame can be expressed as ##EQU4##

On the right-hand side of Eq. (6), the first term r.sub.m (i) is the "recursive component" of R.sub.m (i), while the second term is the "non-recursive component". The finite summation of the non-recursive component is calculated for each frame. On the other hand, we obviously cannot directly calculate the infinite summation of the recursive component; instead, we have to calculate it recursively. The following paragraphs explain how.

Suppose we have calculated and stored all r.sub.m (i)'s for the current frame and want to go on to the next frame, which starts at sample s(m+L). After the hybrid window is shifted to the right by L samples, the new window-weighted signal for the next frame becomes ##EQU5## The recursive component of R.sub.m+L (i) can be written as ##EQU6## Therefore, r.sub.m+L (i) can be calculated recursively from r.sub.m (i) using Eq. (10). This newly calculated r.sub.m+L (i) is stored back to memory for use in the following frame. The autocorrelation coefficient R.sub.m+L (i) is then obtained as ##EQU7##

Note that the autocorrelation calculation procedure described above does not depend on the shape of the non-recursive part of the hybrid window. In other words, any other function can be used for that part. The sine function we used may not be the best possible choice; We chose it only for its simplicity and for its similarity to the shape of Barnwell's recursive window.

With proper scaling, the second terms on the right-hand side of Eqs. (10) and (11) represents 16-bit by 16-bit multiply-accumulate, while the first term of Eq. (10) is a 16-bit by 32-bit multiplication if the constant .alpha..sup.2L is represented by 16 bits. Note that this 16-bit by 32-bit multiplication can be replaced by a k-bit accumulator shift followed by a subtraction if we choose .alpha..sup.2L =(2.sup.k -1)/2.sup.k, or by a single k-bit accumulator shift if we choose .alpha..sup.2L =1/2.sup.k for a large L. In any case, this hybrid windowing method can be implemented without using 32-bit by 32-bit double precision arithmetic. Furthermore, when compared with the original recursive windowing method, this hybrid windowing method saves about 20% to 30% of the number of multiply-adds required for calculating the autocorrelation coefficients.

Since the shapes of Barnwell's recursive window and the new hybrid window are quite similar, the two windows give quite comparable prediction gains.

FIG. 9 shows the arrangements for the weighting of the correlation coefficients R.sub.m (i) to avoid the prolonged vowel sound anomaly noted earlier.

In particular, the normal Phase 2 System processing indicated in FIG. 5, is modified in FIG. 9 to include the weighting in multiplier 150 of the autocorrelation coefficients provided in the manner described above by the hybrid windowing module 49. The weighting values are stored in a memory 149 after being calculated using any one of a number of weighting windows extending over the range of R(1) through R(50). Recall that the weight for R(0) had been previously determined as 257/256 for ease in modifying the power level and, in effect, introducing the desired level of white noise into the LPC spectrum. This weighting value is also included in the table memory 149 in FIG. 9. The other values, as noted, are conveniently calculated and stored in the same table. One convenient weighting function that has proved useful in determining the weighting values for R(1) through R(50) is that described in the above-referenced paper by Y. Tohkura, et al. In particular, the binomial or Gaussian window given by ##EQU8## have proved convenient. In operation the stored weight for a current frame are applied to the respective autocorrelation coefficients to form modified autocorrelation coefficient given by R'(i)=W(i)*R(i), i=0,1,2, . . . ,50. The Tohkura reference is incorporated by reference as if set forth in its entirety to avoid the need for a detailed description of the well-known methodology for populating the weight values of memory 149. While the above description has been presented in terms of the CCITT Phase 1 and Phase 2 Systems, it should be understood that the windowing functionality and associated methods described herein have applicability beyond such particular classes of systems. Further, though the emphasis has been on processing using fixed point processors, no such limitation is fundamental to the present invention. Likewise, while the particular program codes presented in the Draft Recommendation incorporated by reference and attached as Appendix 1, or any particular processors mentioned in the cited references or incorporated by reference may offer advantages in some implementations, those skilled in the art will recognize that other particular codes or processors will be useful in practicing the invention in accordance with the teachings of the overall disclosure. ##SPC1##