Vector adaptive coding method for speech and audio.

This invention relates a real-time coder for compression of digitally encoded speech or audio signals for transmission or storage, and more particularly to a real-time vector adaptive predictive coding system.

In the past few years, most research in speech coding has focused on bit rates from 16 kb/s down to 150 bits/s. At the high end of this range, it is generally accepted that toll quality can be achieved at 16 kb/s by sophisticated waveform coders which are based on scalar quantization. N.S. Jayant and P. Noll, <u>Digital Coding of Waveforms</u>, Prentice-Hall Inc., Englewood Cliffs, N.J., 1984. At the other end, coders (such as linear-predictive coders) operating at 2400 bits/s or below only give synthetic-quality speech. For bit rates between these two extremes, particularly between 4.8 kb/s and 9.6 kb/s, neither type of coder can achieve high-quality speech. Part of the reason is that scalar quantization tends to break down at a bit rate of 1 bit/sample. Vector quantization (VQ), through its theoretical optimality and its capability of operating at a fraction of one bit per sample, offers the potential of achieving high-quality speech at 9.6 kb/s or even at 4.8 kb/s. J. Makhoul, S. Roucos, and H. Gish, "Vector Quantization in Speech Coding," Proc. IEEE, Vol. 73, No. 11, November 1985.

Vector quantization (VQ) can achieve a performance arbitrarily close to the ultimate rate-distortion bound if the vector dimension is large enough. T. Berger, <u>Rate Distortion Theory</u>, Prentice-Hall Inc., Englewood Cliffs,N.J., 1971. However, only small vector dimensions can be used in practical systems due to complexity considerations, and unfortunately, direct waveform VQ using small dimensions does not give adequate performance. One possible way to improve the performance is to combine VQ with other data compression techniques which have been used successfully in scalar coding schemes.

In speech coding below 16 kb/s, one of the most successful scalar coding schemes is Adaptive Predictive Coding (APC) developed by Atal and Schroeder [B.S. Atal and M.R. Schroeder, "Adaptive Predictive Coding of Speech Signals," Bell Syst. Tech. J., Vol. 49, pp. 1973-1986, October 1970; B.S. Atal and M.R. Schroeder, "Predictive Coding of Speech Signals and Subjective Error Criteria," IEEE Trans. Acoust., Speech, Signal Proc., Vol. ASSP-27, No. 3, June 1979; and B.S. Atal, "Predictive Coding of Speech at Low Bit Rates," IEEE Trans. Comm., Vol. COM-30, No. 4, April 1982]. It is the combined power of VQ and APC that led to the development of the present invention, a Vector Adaptive Predictive Coder (VAPC). Such a combination of VQ and APC will provide high-quality speech at bit rates between 4.8 and 9.6 kb/s, thus bridging the gap between scalar coders and VQ coders.

The basic idea of APC is to first remove the redundancy in speech waveforms using adaptive linear predictors, and then quantize the prediction residual using a scalar quantizer. In VAPC, the scalar quantizer in APC is replaced by a vector quantizer VQ. The motivation for using VQ is two-fold. First, although linear dependency between adjacent speech samples is essentially removed by linear prediction, adjacent prediction residual samples may still have nonlinear dependency which can be exploited by VQ. Secondly, VQ can operate at rates below one bit per sample. This is not achievable by scalar quantization, but it is essential for speech coding at low bit rates.

The vector adaptive predictive coder (VAPC) has evolved from APC and the vector predictive coder introduced by V. Cuperman and A. Gersho, "Vector Predictive Coding of Speech at 16 kb/s," IEEE Trans. Comm., Vol. COM-33, pp. 685-696, July 1985. VAPC contains some features that are somewhat similar to the Code-Excited Linear Prediction (CELP) coder by M.R. Schroeder, B.S. Atal, "Code-Excited Linear Prediction (CELP): High-Quality Speech at Very Low Bit Rates," Proc. Int'l. Conf. Acoustics, Speech, Signal Proc., Tampa, March 1985, but with much less computational complexity.

In computer simulations, VAPC gives very good speech quality at 9.6 kb/s, achieving 18 dB of signal-to-noise ratio (SNR) and 16 dB of segmental SNR. At 4.8- kb/s, VAPC also achieves reasonably good speech quality, and the SNR and segmental SNR are about 13 dB and 11.5 dB, respectively. The computations required to achieve these results are only in the order of 2 to 4 million flops per second (one flop, a floating point operation, is defined as one multiplication, one addition, plus the associated indexing), well within the capability of today's advanced digital signal processor chips. VAPC may become a low-complexity alternative to CELP, which is known to have achieved excellent speech quality at an expected bit rate around 4.8 kb/s but is not presently capable of being implemented in real-time due to its astronomical complexity. It requires over 400 million flops per second to implement the coder. In terms of the CPU time of a supercomputer CRAY-1, CELP requires 125 seconds of CPU time to encode one second of speech. There is currently a great need for a real-time, high-quality speech coder operating at encoding rates ranging from 4.8 to 9.6 kb/s. In this range of encoding rates, the two coders mentioned above (APC and CELP) are either unable to achieve high quality or too complex to implement. In contrast, the present invention, which combines Vector Quantization (VQ) with the advantages of both APC and CELP, is able to achieve high-quality speech with sufficiently low complexity for real-time coding.

An object of this invention is to provide adaptive post-filtering of a speech or audio signal that has been corrupted by noise resulting from a coding system or other sources of degradation so as to enhance the perceived quality of said speech or audio signal.

According to the invention there is provided an adaptive postfiltering method for enhancing digitally processed speech or audio signals comprising the steps of    buffering said speech or audio signals into frames of vectors, each vector having K successive samples,    performing analysis of said speech signals in predetermined blocks to compute linear-predictive coefficient LPC predictor, pitch and pitch predictor parameters, and    filtering each vector with long-delay and short-delay filtering in cascade, said long-delay filtering being controlled by pitch and pitch predictor parameters and said short-delay filtering being controlled by said LPC predictor parameters.

The preferred embodiment provides postfiltering for use with a system which approximates each vector of K speech samples by using each of M fixed vectors stored in a VQ codebook to excite a time-varying synthesis filter and picking the best synthesized vector that minimizes a perceptually meaningful distortion measure. The original sampled speech is first buffered and partitioned into vectors and frames of vectors, where each frame is partitioned into N vectors, each vector having K speech samples. Predictive analysis of pitch-filtering parameters (P) linear-predictive coefficient filtering parameters (LPC), perceptual weighting filter parameters (W) and residual gain scaling factor (G) for each of successive frames of speech is then performed. The parameters determined in the analyses are quantized and reset every frame for processing each input vector s_n in the frame, except the perceptual weighting parameter. A perceptual weighting filter responsive to the parameters W is used to help select the VQ vector that minimizes the perceptual distortion between the coded speech and the original speech. Although not quantized, the perceptual weighting filter parameters are also reset every frame.

After each frame is buffered and the above analysis is completed at the beginning of each frame, M zero-state response vectors are computed and stored in a zero-state response codebook. These M zero-state response vectors are obtained by setting to zero the memory of an LPC synthesis filter and a perceptual weighting filter in cascade after a scaling unit controlled by the factor G, and controlling the respective filters with the quantized LPC filter parameters and the unquantized perceptual weighting filter parameters, and exciting the cascaded filters using one predetermined and fixed codebook vector at a time. The output vector of the cascaded filters for each VQ codebook vector is then stored in the corresponding address, i.e., is assigned the same index of a temporary zero-state response codebook as of the VQ codebook. In encoding each input speech vector s_n within a frame, a pitch prediction ŝ_n of the vector is determined by processing the last vector encoded as an index code through a scaling unit, LPC synthesis filter and pitch predictor filter controlled by the parameters QG, QLPC, QP and QPP for the frame. In addition, the zero-input response of the cascaded filters (the ringing from excitation of a previous vector) is first set in a filter. Once the pitch-predicted vector ŝ_n is subtracted from the input signal vector s_n, and a difference vector d_n is passed through the perceptual weighting filter to produce a filtered difference vector f_n, the zero-input response vector in the aforesaid filter is subtracted from the perceptual weight filtered difference vector f_n, and the resulting vector v_n is compared with each of the M stored zero-state response vectors in search of the one having a minimum difference Δ or distortion.

The index (address) of the zero-state response vector that produces the smallest distortion, i.e., that is closest to v_n, identifies the best vector in the permanent codebook. Its index (address) is transmitted as the compressed code for the vector, and used by a receiver which has an identical VQ codebook as the transmitter to find the best-match vector. In the transmitter, that best-match vector is used at the time of transmission of its index to excite the LPC synthesis filter and pitch prediction filter to generate an estimate ŝ_n of the next speech vector. The best-match vector is also used to excite the zero-input response filter to set it for the next speech vector s_n as described above. The indices of the best-match vector for a frame of vectors are combined in a multiplexer with the frame analysis information hereinafter referred to as "side information," comprised of the indices of parameters which control pitch, pitch predictor and LPC predictor filtering and the gain used in the coding process, in order that it may be used by the receiver in decoding the vector indices of a frame into vectors using a codebook identical to the permanent codebook at the transmitter. This side information is preferably transmitted through the multiplexer first, once for each frame of VQ indices that follow, but it would be possible to first transmit a frame of vector indices, and then transmit the side information since the frames of vector indices will require some buffering in either case; the difference is only in some initial delay at the beginning of speech or audio frames transmitted in succession. The resulting stream of multiplexed indices are transmitted over a communication channel to a decoder, or stored for later decoding.

In the decoder, the bit stream is first demultiplexed to separate the side information from the indices that follow. Each index is used at the receiver to extract the corresponding vector from the duplicate codebook. The extracted vector is first scaled by the gain parameter, using a table to convert the gain index to the appropriate scaling factor, and then used to excite cascaded LPC synthesis and pitch synthesis filters controlled by the same side information used in selecting the best-match index utilizing the zero-state response codebook in the transmitter. The output of the pitch synthesis filter is the coded speech, which is perceptually close to the original speech. All of the side information, except the gain information, is used in an adaptive postfilter to enhance the quality of the speech synthesized. This postfiltering technique may be used to enhance any voice or audio signal. All that would be required is an analysis section to produce the parameters used to make the postfilter adaptive.

Other modifications and variation to this invention may occur to those skilled in the art, such as variable-frame-rate coding, fast codebook searching, reversal of the order of pitch prediction and LPC prediction, and use of alternative perceptual weighting techniques. Consequently, the claims which define the present invention are intended to encompass such modifications and variations.

Although the purpose of this invention is to encode for transmission and/or storage of analog speech or audio waveforms for subsequent reconstruction of the waveforms upon reproduction of the speech or audio program, reference is made hereinafter only to speech, but the invention described and claimed is applicable to audio waveforms or to sub-band filtered speech or audio waveforms.

The present application is a divisional application from European Patent Application 88303038.9 filed on 6 April 1988.

An example of the invention will now be described with reference to the accompanying drawings.

_{<b>FIG. 1a</b>} is a block diagram of a Vector Adaptive Predictive Coding (VAPC) processor embodying the present invention, and <b>FIG. 1b</b> is a block diagram of a receiver for the encoded speech transmitted by the system of <b>FIG. 1a.</b>

<b>FIG. 2</b> is a schematic diagram that illustrates the adaptive computation of vectors for a zero-state response codebook in the system of <b>FIG. 1a.</b>

<b>FIG. 3</b> is a block diagram of an analysis processor in the system of <b>FIG. 1a.</b>

<b>FIG. 4</b> is a block diagram of an adaptive post filter of <b>FIG. 1b.</b>

<b>FIG. 5</b> illustrates the LPC spectrum and the corresponding frequency response of an all-pole post-filter 1/[1̂-P(z/ α)] for different values of α. The offset between adjacent plots is 20 dB.

<b>FIG. 6</b> illustrates the frequency responses of the postfilter [1-µz⁻¹][1-P̂(z/β)]/[1-P̂(z/ α)] corresponding to the LPC spectrum shown in <b>FIG. 5.</b> In both plots, α=0.8 and β=0.5. The offset between the two plots is 20 dB.

The preferred mode of implementation contemplates using programmable digital signal processing chips, such as one or two AT&amp;T DSP32 chips, and auxiliary chips for the necessary memory and controllers for such functions as input sampling, buffering and multiplexing. Since the system is digital, it is synchronized throughout with the samples. For simplicity of illustration and explanation, the synchronizing logic is not shown in the drawings. Also for simplification, at each point where a signal vector is subtracted from another, the subtraction function is symbolically indicated by an adder represented by a plus sign within a circle. The vector being subtracted is on the input labeled with a minus sign. In practice, the two's complement of the subtrahend is formed and added to the minuend. However, although the preferred implementation contemplates programmable digital signal processors, it would be possible to design and fabricate special integrated circuits using VLSI techniques to implement the present invention as a special purpose, dedicated digital signal processor once the quantities needed would justify the initial cost of design.

Referring to <b>FIG. 1a,</b> original speech samples, s_n in digital form from sampling analog-to-digital converter <b>10</b> are received by an analysis processor <b>11</b> which partitions them into vectors s_n of K samples per vector, and into frames of N vectors per frame. The analysis processor stores the samples in a dual buffer memory which has the capacity for storing more than one frame of vectors, for example two frames of 8 vectors per frame, each vector consisting of 20 samples, so that the analysis processor may compute parameters used for coding the following frame. As each frame is being processed out of one buffer, a new frame coming in is stored in the other buffer so that when processing of a frame has been completed, there is a new frame buffered and ready to be processed.

The analysis processor determines the parameters of filters employed in the Vector Adaptive Predictive Coding technique that is the subject of this invention. These parameters are transmitted through a multiplexer <b>12</b> as side information just ahead of the frame of vector codes generated with the use of a vector quantized (VQ) permanent codebook <b>13</b> and a zero-state response (ZSR) codebook <b>14.</b> The side information conditions the receiver to properly filter decoded vectors of the frame. The analysis processor <b>11</b> also computes other parameters used in the encoding process. The latter are represented in <b>FIG. 1a</b> by dashed lines, and consist of sets of parameters which are designated W for a perceptual weighting filter <b>18</b>, a quantized LPC predictor QLPC for an LPC synthesis filter <b>15</b>, and quantized pitch QP and pitch predictor QPP for a pitch synthesis filter <b>16.</b> Also computed by the analysis processor is a scaling factor G for control of a scaling unit <b>17.</b> The four quantized parameters transmitted as side information are encoded using a quantizing table as the quantized pitch index, pitch predictor index, LPC predictor index and gain index. The manner in which the analysis processor computes all of these parameters will be described with reference to <b>FIG. 3.</b>

The multiplexer <b>12</b> preferably transmits the side information as soon as it is available, although it could follow the frame of encoded input vectors, and while that is being done, M zero-state response vectors are computed for the zero-state response (ZSR) codebook <b>14</b> in a manner illustrated in <b>FIG. 2,</b> which is to process each vector in the VQ codebook, 13 e.g., 128 vectors, through a gain scaling unit <b>17',</b> an LPC synthesis filter <b>15',</b> and perceptual weighting filters <b>18'</b> corresponding to the gain scaling unit <b>17,</b> the LPC synthesis filter <b>15,</b> and perceptual weighting filter <b>18</b> in the transmitter <b>(FIG. 1a).</b> Ganged commutating switches S₁ and S₂ are shown to signify that each fixed VQ vector processed is stored in memory locations of the same index (address) in the ZSR codebook.

At the beginning of each vector processing, the initial conditions of the cascaded filters <b>15'</b> and <b>18'</b> are set to zero. This simulates what the cascaded filters <b>15'</b> and <b>18'</b> will do with no previous vector present from its corresponding VQ codebook. Thus, if the output of a zero-input response filter <b>19</b> in the transmitter <b>(FIG. 1a)</b> is held or stored,    at each step of computing the VQ code index (to transmit for each vector of a frame), it is possible to simplify encoding the speech vectors by subtracting the zero-state response output from the vector f_n. In other words, assuming M=128, there are 128 different vectors permanently stored in the VQ codebook to use in coding the original speech vectors s_n. Then every one of the 128 VQ vectors is read out in sequence, fed through the scaling unit <b>17',</b> the LPC synthesis filter <b>15',</b> and the perceptual weighting filter <b>18'</b> without any history of previous vector inputs by resetting those filters at each step. The resulting filter output vector is then stored in a corresponding location in the zero-state response codebook. Later, while encoding input signal vectors s_n by finding the best match between a vector v_n and all of the zero state response vector codes, it is necessary to subtract from a vector f_n derived from the perceptual weighting filter a value that corresponds to the effect of the previously selected VQ vector. That is done through the zero-input response filter <b>19.</b> The index (address) of the best match is used as the compressed vector code transmitted for the vector s_n. Of the 128 zero-state response vectors, there will be only one that provides the best match, i.e., least distortion. Assume it is in location 38 of the zero-state response codebook as determined by a computer <b>20</b> labeled "compute norm." An address register <b>20a</b> will store the index <b>38.</b> It is that index that is then transmitted as a VQ index to the receiver shown in <b>FIG. 1b.</b>

In the receiver, a demultiplexer <b>21</b> separates the side information which conditions the receiver with the same parameters as corresponding filters and scaling unit of the transmitter. The receiver uses a decoder <b>22</b> to translate the parameters indices to parameter values. The VQ index for each successive vector in the frame addresses a VQ codebook <b>23</b> which is identical to the fixed VQ codebook <b>13</b> of the transmitter. The LPC synthesis filter <b>24,</b> pitch synthesis filter <b>25,</b> and scaling unit <b>26</b> are conditioned by the same parameters which were used in computing the zero-state codebook values, and which were in turn used in the process of selecting the encoding index for each input vector. At each step of finding and transmitting an encoding index, the zero-input response filter 19 computes from the VQ vector at the location of the index transmitted a value to be subtracted from the input vector f_n to present a zero-input response to be used in the best-match search.

There are various procedures that may be used to determine the best match for an input vector s_n. The simplest is to store the resulting distortion between each zero-state response vectorcode output and the vector v_n with the index of that. Assuming there are 128 vectorcodes stored in the codebook <b>14,</b> there would then be 128 resulting distortions stored in a best address computer <b>20.</b> Then, after all have been stored, a search is made in the computer <b>20</b> for the lowest distortion value. Its index is then transmitted to the receiver as an encoded vector via the multiplexer <b>12,</b> and to the VQ codebook for reading the corresponding VQ vector to be used in the processing of the next input vector s_n.

In summary, it should be noted that the VQ codebook is used (accessed) in two different steps: first, to compute vector codes for the zero-state response codebook at the beginning of each frame, using the LPC synthesis and perceptual weighting filter parameters determined for the frame; and second, to excite the filters <b>15</b> and <b>16</b> through the scaling unit <b>17</b> while searching for the index of the best-match vector, during which the estimate ŝ_n thus produced is subtracted from the input vector s_n. The difference d_n is used in the best-match search.

As the best match for each input vector s_n is found, the corresponding predetermined and fixed vector from the VQ codebook is used to reset the zero input response filter <b>19</b> for the next vector of the frame. The function of the zero-input response filter <b>19</b> is thus to find the residual response of the gain scaling unit <b>17'</b> and filters <b>15'</b> and <b>18'</b> to previously selected vectors from the VQ codebook. Thus, the selected vector is not transmitted; only its index is transmitted. At the receiver its index is used to read out the selected vector from a VQ codebook <b>23</b> identical to the VQ codebook <b>13</b> in the transmitter.

The zero-input response filter <b>19</b> is the same filtering operation that is used to generate the ZSR codebook, namely the combination of a gain G, an LPC synthesis filter and a weighting filter, as shown in <b>FIG. 2.</b> Once a best codebook vector match is determined, the best-match vector is applied as an input to this filter (sample by sample, sequentially). An input switch s_i is closed and an output switch s_o is open during this time so that the first K output samples are ignored. (K is the dimension of the vector and a typical value is 20.) As soon as all K samples have been applied as input to the filter, the filter input switch s_i is opened and the output switch s_o is closed. The next K samples of the vector f_n, the output of the perceptual weighting filter, begin to arrive and are subtracted from the samples of the vector f_n. The difference so generated is a set of K samples forming the vector v_n which is stored in a static register for use in the ZSR codebook search procedure. In the ZSR codebook search procedure, the vector v_n is subtracted from each vector stored in the ZSR codebook, and the difference vector Δ is fed to the computer <b>20</b> together with the index (or stored in the same order), thereby to imply the index of the vector out of the ZSR codebook. The computer <b>20</b> then determines which difference is the smallest, i.e., which is the best match between the vector v_n and each vector stored temporarily (for one frame of input vectors s_n). The index of that best-match vector is stored in a register <b>20a.</b> That index is transmitted as a vectorcode and used to address the VQ codebook to read the vector stored there into the scaling unit <b>17,</b> as noted above. This search process is repeated for each vector in the ZSR codebook, each time using the same vector v_n. Then the best vector is determined.

Referring now to <b>FIG. 1b;</b> it should be noted that the output of the VQ codebook <b>23,</b> which precisely duplicates the VQ codebook <b>13</b> of the transmitter, is identical to the vector extracted from the best-match index applied as an address to the VQ codebook <b>13;</b> the gain unit <b>26</b> is identical to the gain unit <b>17</b> in the transmitter, and filters <b>24</b> and <b>25</b> exactly duplicate the filters <b>15</b> and <b>16,</b> respectively, except that at the receiver, the approximation <img file="EP0503684A2_D0001.tif" />_n rather than the prediction ŝ_n is taken as the output of the pitch synthesis filter <b>25.</b> The result, after converting from digital to analog form, is synthesized speech that reproduces the original speech with very good quality.

It has been found that by applying an adaptive postfilter <b>30</b> to the synthesized speech before converting it from digital to analog form, the perceived coding noise may be greatly reduced without introducing significant distortion in the filtered speech. <b>FIG. 4</b> illustrates the organization of the adaptive postfilter as a long-delay filter <b>31</b> and a short-delay filter <b>32.</b> Both filters are adaptive in that the parameters used in them are those received as side information from the transmitter, except for the gain parameter, G. The basic idea of adaptive post-filtering is to attenuate the frequency components of the coded speech in spectral valley regions. At low bit rates, a considerable amount of perceived coding noise comes from spectral valley regions where there are no strong resonances to mask the noise. The postfilter attenuates the noise components in spectral valley regions to make the coding noise less perceivable. However, such filtering operation inevitably introduces some distortion to the shape of the speech spectrum. Fortunately, our ears are not very sensitive to distortion in spectral valley regions; therefore, adaptive postfiltering only introduces very slight distortion in perceived speech, but it significantly reduces the perceived noise level. The adaptive postfilter will be described in greater detail after first describing in more detail the analysis of a frame of vectors to determine the side information.

Referring now to <b>FIG. 3,</b> it shows the organization of the initial analysis of block <b>11</b> in <b>FIG. 1a.</b> The input speech samples s_n are first stored in a buffer <b>40</b> capable of storing, for example, more than one frame of 8 vectors, each vector having 20 samples.

Once a frame of input vectors s_n has been stored, the parameters to be used, and their indices to be transmitted as side information, are determined from that frame and at least a part of the previous frame in order to perform analysis with information from more than the frame of interest. The analysis is carried out as shown using a pitch detector <b>41,</b> pitch quantizer <b>42</b> and a pitch predictor coefficient quantizer <b>43.</b> What is referred to as "pitch" applies to any observed periodicity in the input signal, which may not necessarily correspond to the classical use of "pitch" corresponding to vibrations in the human vocal folds. The direct output of the speech is also used in the pitch predictor coefficient quantizer <b>43.</b> The quantized pitch (QP) and quantized pitch predictor (QPP) are used to compute a pitch-prediction residual in block <b>44,</b> and as control parameters for the pitch synthesis filter <b>16-</b> used as a predictor in <b>FIG. 1a.</b> Only a pitch index and a pitch prediction index are included in the side information to minimize the number of bits transmitted. At the receiver, the decoder <b>22</b> will use each index to produce the corresponding control parameters for the pitch synthesis filter <b>25</b>.

The pitch-prediction residual is stored in a buffer <b>45</b> for LPC analysis in block <b>46.</b> The LPC predictor from the LPC analysis is quantized in block <b>47.</b> The index of the quantized LPC predictor is transmitted as a third one of four pieces of side information, while the quantized LPC predictor is used as a parameter for control of the LPC synthesis filter <b>15,</b> and in block <b>48</b> to compute the rms value of the LPC predictive residual. This value (unquantized residual gain) is then quantized in block <b>49</b> to provide gain control G in the scaling unit <b>17</b> of <b>FIG. 1a.</b> The index of the quantized residual gain is the fourth part of the side information transmitted.

In addition to the foregoing, the analysis section provides LPC analysis in block <b>50</b> to produce an LPC predictor from which the set of parameters W for the perceptual weighting filter <b>18 (FIG. 1a)</b> is computed in block <b>51.</b>

The adaptive postfilter <b>30</b> in <b>FIG. 1b</b> will now be described with reference to <b>FIG. 4.</b> It consists of a long-delay filter <b>31</b> and a short-delay filter 32 in cascade. The long-delay filter is derived from the decoded pitch-predictor information available at the receiver. It attenuates frequency components between pitch harmonic frequencies. The short-delay filter is derived from LPC predictor information, and it attenuates the frequency components between formant frequencies.

The noise masking effect of human auditory perception, recognized by M.R. Schroeder, B.S. Atal, and J.L. Hall, "Optimizing Digital Speech Coders by Exploiting Masking Properties of the Human Ear," J. Acoust. Soc. Am., Vol. 66, No. 6, pp. 1647-1652, December 1979, is exploited in VAPC by using noise spectral shaping. However, in noise spectral shaping, lowering noise components at certain frequencies can only be achieved at the price of increased noise components at other frequencies. [B.S. Atal and M.R. Schroeder, "Predictive Coding of Speech Signals and Subjective Error Criteria," IEEE Trans. Acoust., Speech, and Signal Processing, Vol. ASSP-27, No. 3, pp. 247-254, June 1979] Therefore, at bit rates as low as 4800 bps, where the average noise level is quite high, it is very difficult, if not impossible, to force noise below the masking threshold at all frequencies. Since speech formants are much more important to perception than spectral valleys, the approach of the present invention is to preserve the formant information by keeping the noise in the formant regions as low as is practical during encoding. Of course, in this case, the noise components in spectral valleys may exceed the threshold; however, these noise components can be attenuated later by the postfilter <b>30.</b> In performing such postfiltering, the speech components in spectral valleys will also be attenuated. Fortunately, the limen, or "just noticeable difference," for the intensity of spectral valleys can be quite large [J.L. Flanagan, Speech Analysis, Synthesis, and Perception, Academic Press, New York, 1972]. Therefore, by attenuating the components in spectral valleys, the postfilter only introduces minimal distortion in the speech signal, but it achieves a substantial noise reduction.

Adaptive postfiltering has been used successfully in enhancing ADPCM-coded speech. See V. Ramamoorthy and J.S. Jayant, "Enhancement of ADPCM Speech by Adaptive Postfiltering," AT&amp;T Bell Labs Tech. J., pp. 1465-1475, October 1984; and N.S. Jayant and V. Ramamoorthy, "Adaptive Postfiltering of 16 kb/s-ADPCM Speech," Proc. ICASSP, pp. 829-832, Tokyo, Japan, April 1986. The postfilter used by Ramamoorthy, et al., supra, is derived from the two-pole six-zero ADPCM, synthesis filter by moving the poles and zeros radially toward the origin. If this idea is extended directly to an all-pole LPC synthesis filter 1/[1-P̂(z)], the result is l/[1-P̂(z/α)] as the corresponding postfilter, where 0&lt;α&lt;1. Such an all-pole postfilter indeed reduces the perceived noise level; however, sufficient noise reduction can only be achieved with severe muffling in the filtered speech. This is due to the fact that the frequency response of this all-pole postfilter generally has a lowpass spectral tilt for voiced speech.

The spectral tilt of the all-pole postfilter 1/[1-P̂(z/α)] can be easily reduced by adding zeros having the same phase angles as the poles but with smaller radii. The transfer function of the resulting pole-zero postfilter <b>32a</b> has the form<maths id="math0001" num=""><img file="EP0503684A2_D0002.tif" /></maths> where α and β are coefficients empirically determined, with some tradeoff between spectral peaks being so sharp as to produce chirping and being so low as to not achieve any noise reduction. The frequency response of H(z) can be expressed as<maths id="math0002" num=""><img file="EP0503684A2_D0003.tif" /></maths><maths id="math0003" num=""><img file="EP0503684A2_D0004.tif" /></maths> Therefore, in logarithmic scale, the frequency response of the pole-zero postfilter H(z) is simply the difference between the frequency responses of two all-pole postfilters.

Typical values of α and β are 0.8 and 0.5, respectively. From <b>FIG. 5,</b> it is seen that the response for α=0.8 has both formant peaks and spectral tilt, while the response for α=0.5 has spectral tilt only. Thus, with α=0.8 and β=0.5 in Equation 2, we can at least partially remove the spectral tilt by subtracting the response for α=0.5 from the response for α=0.8. The resulting frequency response of H(z) is shown in the upper plot of <b>FIG. 6.</b>

In informal listening tests, it has been found that the muffling effect was significantly reduced after the numerator term [1-P̂(z/β)] was included in the transfer function H(z). However, the filtered speech remained slightly muffled even with the spectral-tilt compensating term [1-P̂(z/β)]. To further reduce the muffling effect, a first-order filter <b>32b</b> was added which has a transfer function of [1-µz⁻¹], where µ is typically 0.5. Such a filter provides a slightly highpassed spectral tilt and thus helps to reduce muffling. This first-order filter is used in cascade with H(z), and a combined frequency response with µ=0.5 is shown in the lower plot of <b>FIG. 6.</b>

The short-delay postfilter <b>32</b> just described basically amplifies speech formants and attenuates inter-formant valleys. To obtain the ideal postfilter frequency response, we also have to amplify the pitch harmonics and attenuate the valleys between harmonics. Such a characteristic of frequency response can be achieved with a long-delay postfilter using the information in the pitch predictor.

In VAPC, we use a three-tap pitch predictor; the pitch synthesis filter corresponding to such a pitch predictor is not guaranteed to be stable. Since the poles of such a synthesis filter may be outside the unit circle, moving the poles toward the origin may not have the same effect as in a stable LPC synthesis filter. Even if the three-tap pitch synthesis filter is stabilized, its frequency response may have an undesirable spectral tilt. Thus, it is not suitable to obtain the long-delay postfilter by scaling down the three tap weights of the pitch synthesis filter.

With both poles and zeroes, the long-delay postfilter can be chosen as<maths id="math0004" num=""><img file="EP0503684A2_D0005.tif" /></maths> where p is determined by pitch analysis, and C_g is an adaptive scaling factor.

Knowing the information provided by a single or three-tap pitch predictor as the value b₂ or the sum of b₁+b₂+b₃, the factors γ and λ are determined according to the following formulas: <maths id="math0005" num=""><math display="inline"><mrow><msub><mrow><mtext>γ = C</mtext></mrow><mrow><mtext>z</mtext></mrow></msub><msub><mrow><mtext>f(x), λ = C</mtext></mrow><mrow><mtext>p</mtext></mrow></msub><msub><mrow><mtext>f(x), 0 &lt; C</mtext></mrow><mrow><mtext>z</mtext></mrow></msub><msub><mrow><mtext>, C</mtext></mrow><mrow><mtext>p</mtext></mrow></msub><mtext> &lt; 1   (4)</mtext></mrow></math><img file="EP0503684A2_D0006.tif" /></maths> where<maths id="math0006" num=""><img file="EP0503684A2_D0007.tif" /></maths> where Uth is a threshold value (typically 0.6) determined empirically, and x can be either b₂ or b₁+b₂+b₃ depending on whether a one-tap or a three-tap pitch predictor is used. Since a quantized three-tap pitch predictor is preferred and therefore already available at the VAPC receiver, x is chosen as<maths id="math0007" num=""><img file="EP0503684A2_D0008.tif" /></maths> in VAPC postfiltering. On the other hand, if the postfilter is used elsewhere to enhance noisy input speech, a separate pitch analysis is needed, and x may be chosen as a single value b₂ since a one-tap pitch predictor suffices. (The value b₂ when used alone indicates a value from a single-tap predictor, which in practice would be the same as a three-tap predictor when b₁ and b₃ are set to zero.)

The goal is to make the power of [y(n)] about the same as that of [s(n)]. An appropriate scaling factor is chosen as<maths id="math0008" num=""><img file="EP0503684A2_D0009.tif" /></maths>

The first-order filter <b>32b</b> can also be made adaptive to better track the change in the spectral tilt of H(z). However, it has been found that even a fixed filter with µ=0.5 gives quite satisfactory results. A fixed value of µ may be determined empirically.

To avoid occasional large gain excursions, an automatic gain control (AGC) was added at the output of the adaptive postfilter. The purpose of AGC is to scale the enhanced speech such that it has roughly the same power as the unfiltered noisy speech. It is comprised of a gain (volume) estimator <b>33</b> operating on the speech input s(n), a gain (volume) estimator <b>34</b> operating on the postfiltered output r(n), and a circuit <b>35</b> to compute a scaling factor as the ratios of the two gains. The postfiltering output r(n) is then multiplied by this ratio in a multiplier <b>36</b>. AGC is thus achieved by estimating the power of the unfiltered and filtered speech separately and then using the ratio of the two values as the scaling factor. Let [s(n)] be the sequence of either unfiltered or filtered speech samples; then, the speech power σ²(n) is estimated by using <maths id="math0009" num=""><math display="inline"><mrow><mtext>σ²(n)=ξ σ²(n-1)+(1-ξ)s²(n), 0&lt;ξ&lt;1.   (7)</mtext></mrow></math><img file="EP0503684A2_D0010.tif" /></maths> A suitable value of ξ is 0.99.

The complexity of the postfilter described in this section is only a small fraction of the overall complexity of the rest of the VAPC system, or any other coding system that may be used. In simulations, this postfilter achieves significant noise reduction with almost negligible distortion in speech. To test for possible distorting effects, the adaptive postfiltering operation was applied to clean, uncoded speech and it was found that the unfiltered original and its filtered version sound essentially the same, indicating that the distortion introduced by this postfilter is negligible.

It should be noted that although this novel postfiltering technique was developed for use with the VAPC system , its applications are not restricted to use with it. In fact, this technique can be used not only to enhance the quality of any noisy digital speech signal but also to enhance the decoded speech of other speech coders when provided with a buffer and analysis section for determining the parameters.

What has been disclosed is a real-time Vector Adaptive Predictive Coder (VAPC) for speech or audio which may be implemented with software using the commercially available AT&amp;T DSP32 digital processing chip. In its newest version, this chip has a processing power of 6 million instructions per second (MIPS). To facilitate implementation for real-time speech coding, a simplified version of the 4800 bps VAPC is available. This simplified version has a much lower complexity, but gives nearly the same speech-quality as a full complexity version.

In the real-time implementation, an inner-product approach is used for computing the norm (smallest distortion) which is more efficient than the conventional difference-square approach of computing the mean square error (MSE) distortion. Given a test vector v and M ZSR codebook vectors, z_j, j=1,2, . . .,M, the j-th MSE distortion can be computed as <maths id="math0010" num=""><math display="block"><mrow><msub><mrow><mtext>∥ v-z</mtext></mrow><mrow><mtext>j</mtext></mrow></msub><msup><mrow><mtext> ∥² = ∥ v ∥² -2 [ v</mtext></mrow><mrow><mtext>T</mtext></mrow></msup><msub><mrow><mtext>z</mtext></mrow><mrow><mtext>j</mtext></mrow></msub><mtext>- </mtext><mfrac><mrow><mtext>1</mtext></mrow><mrow><mtext>2</mtext></mrow></mfrac><msub><mrow><mtext> ∥ z</mtext></mrow><mrow><mtext>j</mtext></mrow></msub><mtext> ∥² ]   (8)</mtext></mrow></math><img file="EP0503684A2_D0011.tif" /></maths> At the beginning of each frame, it is possible to compute and store 1/2∥ Z_j ∥² . With the DSP32 processor and for the dimension and codebook size used, the difference-square approach of the codebook search requires about 2.5 MIPS to implement, while the inner-product approach only requires about 1.5 MIPS.

The complexity of the VAPC is only about 3 million multiply-adds/second and 6 k words of data memory. However, due to the overhead in implementation, a single DSP32 chip was not sufficient for implementing the coder. Therefore, two DSP32 chips were used to implement the VAPC. With a faster-DSP32 chip now available, which has an instruction cycle time of 160 ns rather than 250 ns, it is expected that the VAPC can be implemented using only one DSP32 chip.