Copyright © 1981 American Telephone and Telegraph Company
`T H E B E L L S Y S T E M T E C H N I C A L J O U R N A L
`Vol. 60, No. 8, October 1981
`Printed in U.S.A.
`
`Improving the Quality of a Noisy Speech Signal
`
`By M. M. SONDHI, C. E. SCHMIDT, and L. R. RABINER
`
`(Manuscript received December 18, 1980)
`
`In this paper we discuss the problem of reducing the noise level of
`a noisy speech signal. Several variants of the well-known class of
`"spectral subtraction" techniques are described. The basic implemen­
`tation consists of a channel vocoder in which both the noise spectral
`level and the overall (signal + noise) spectral level are estimated in
`each channel, and the gain of each channel is adjusted on the basis
`of the relative noise level in that channel. Two improvements over
`previously known techniques have been studied. One is a noise level
`estimator based on a slowly varying, adaptive noise-level histogram.
`The other is a nonlinear smoother based on inter-channel continuity
`constraints for eliminating the so-called "musical tones" (i.e., narrow­
`band noise bursts of varying pitch). Informal listening indicates that
`for modest signal-to-noise ratios (greater than about 8 dB) substan­
`tial noise reduction is achieved with little degradation of the speech
`quality.
`
`I.
`
`INTRODUCTION
`
`The idea that a vocoder may be used to improve the quality of a
`noisy speech signal, has been around for about twenty years. T o the
`best of our knowledge the first such proposal was made in 1960 by M .
`R. Schroeder.1 The basic idea of this proposal can be explained with
`the help of Fig. 1, as follows:
`Figure la shows a typical short-term magnitude spectrum of a voiced
`portion of a noisy speech signal. Let S (ů) denote the envelope of this
`spectrum. (Recall that the "channel gains" of a vocoder are estimates
`of this envelope at the center frequencies of the channels. The fine
`structure of the spectrum is attributed to the harmonics of the fun­
`damental voice frequency.)
`, of the envelope.
`Figure lb shows a "formant equalized" version, S (ω)
`The peaks in S and S occur at the same frequencies but the peaks of
`S (unlike those of S) are all of the same amplitude.
`
`1847
`
`Exhibit 1022
`Page 01 of 13
`
`

`

`D
`ι -
`
`α <
`
`FREQUENCY
`Fig. 1—Illustration of noise stripping by increasing the dynamic range between
`formant peaks and noise valleys, (a) Original spectral envelope and fine structure, (b)
`Formant-level equalized spectral envelope, (c) The product spectrum S s(to)S(u) in
`which the ratios between formant peaks and valleys is larger than in the original
`spectrum.
`
`The proposal is, essentially, to generate a signal with a fine structure
`as close as possible to that of the original speech signal, but with an
`envelope given by SnS, where ç is some intetger, say, 1 or 2. Except for
`a scale factor, the spectral envelope of the resulting signal is the same
`as that of the original signal at the formant peaks, but is considerably
`reduced in the valleys. As shown in Fig. lc this processing effectively
`reduces the overall noise level. Of course, the formant peaks also
`become sharper, i.e., the formant bandwidths get reduced.
`Reference 1 describes two implementations of this idea: a frequency
`domain method in which the envelope is modified by modifying the
`channel gains of a self-excited channel vocoder, and a time domain
`method in which the same effect is achieved by repeated convolution.
`In many practical cases of interest, the noise is additive and uncor-
`related with the speech signal. In such a situation, if it were possible
`to estimate the spectral level of the noise as a function of frequency,
`then the noise reduction could be achieved in a somewhat different
`manner. Suppose the noisy speech is applied to the input of a channel
`vocoder (see Section I I for a detailed description). Let the output of
`the kth channel be y* = s* + ra*, where s* is in the speech signal and
`the noise signal in that channel. Let Nl be the average power of the
`noise and Si that of the speech signal. Then, assuming that the noise
`and speech are uncorrected, the average power of the noisy speech is
`given by
`
`Now Y\ can be estimated directly from the output signal yk. If an
`
`Υ ٠ = SI + Nl
`
`(1)
`
`1848 THE BELL SYSTEM TECHNICAL JOURNAL, OCTOBER 1981
`
`Exhibit 1022
`Page 02 of 13
`
`

`

`estimate of Ni is available, as postulated, then ( Y * — TV*) 1 / 2 provides
`an estimate of the magnitude of the signal alone in the kth channel.
`Thus, if the level of the channel signal is multiplied by the ratio of this
`estimated signal power to overall power, then a noise reduction is
`achieved.
`In 1964, at the suggestion of M . R. Schroeder, this "spectral sub­
`traction" idea was implemented as a BLODI language computer pro­
`gram by one of us (MMS) in collaboration with Sally Sievers.2 Besides
`spectral subtraction, one other feature was incorporated into this
`implementation. It had been recently demonstrated that autocorrela­
`tion and cepstrum pitch extraction are quite accurate and reliable for
`noisy speech signals with signal-to-noise ratio (s/n) as low as 6 dB. 3 , 4
`Such extractors provide a clean excitation signal even from a highly
`noisy speech signal. Therefore, the self-excitation described in Ref. 1
`was replaced by a voiced-unvoiced (buzz-hiss) signal derived from an
`autocorrelation pitch extractor.
`Although this implementation demonstrated the feasibility of the
`basic idea, the computer facilities available at that time did not allow
`a thorough investigation of the effects of changing various parameters
`and configurations. Also, since digital hardware was not yet readily
`available, it did not appear likely that such noise-stripping techniques
`would find application in the immediate future. For these reasons
`these techniques were not actively pursued at that time.
`Since the mid-seventies, presumably due to the vastly improved
`digital technology and renewed military interest, noise-stripping has
`again attracted considerable attention. The renewed interest in this
`problem appears to have started in 1974, when Weiss et al. independ­
`ently discovered the spectral subtraction method.5 Except for the fact
`that the filter bank of the channel vocoder was replaced by short-term
`Fourier analysis, the implementation of Weiss et al. was quite similar
`to the one described above. During the past five or six years several
`studies have explored this and other methods for noise removal.
`Notable among these is the work of Boll, Berouti et al., and McAulay
`and Malpass. 6 , 7 , 8 A review of these and other studies is given in a recent
`paper by Lim and Oppenheim.9
`In view of the current interest in noise removal, we have recently
`been experimenting with the spectral subtraction method by computer
`simulation. Subsequent sections of this paper describe the results of
`our experiments.
`From the brief description given above, it is clear that spectral
`subtraction is expected to be useful only in cases when the noise is
`additive. With this constraint, there are basically two types of situa­
`tions in which this method might find application:
`(i) The speech may be produced in a noisy environment, e.g., in
`
`S P E E C H S I G N A L
`
`1 8 4 9
`
`Exhibit 1022
`Page 03 of 13
`
`

`

`the cockpit of an airplane. In such a situation the spectrum of the noise
`is unknown a priori. This information must be estimated from the
`noisy speech signal itself, e.g., during intervals of silence between
`speech bursts. The algorithm for estimating the noise spectrum is,
`therefore, one of the most important parts of the simulations described
`later.
`(ii) The speech itself may be generated in a quiet environment but
`might be transformed to a noisy signal because of the action of a
`coding device. Examples where such noise may be modelled as additive
`are pulse-code modulation (PCM) coders, and delta modulators whose
`step size is chosen such that granular noise predominates over the
`slope-overload noise. In such cases, both the level of the noise and its
`spectral composition might be known a priori. Use of this a priori
`information simplifies the system and improves its performance.
`There is a third way in which noise may enter the communication
`channel additively. The speech signal may be generated in a quiet
`environment but the listener may be in a noisy environment. A
`message sent over the public address system at a busy railway station
`is such an example. In this case, the problem is to preprocess the
`speech signal in such a way that its intelligibility is least impaired by
`the noise. Some work on this problem has been reported in the
`literature;10 however, we will not deal with this problem.
`Before turning to a description of our simulations, it is worth
`emphasizing that we deliberately used the word "quality" rather than
`"intelligibility" in the title of this paper. Ideally, of course, one would
`Like the intelligibility also to be increased. However, this is not abso­
`lutely essential. It is quite annoying and fatiguing to have to listen to
`a noisy speech signal for any length of time. Therefore, a device that
`reduces or eliminates the noise can be quite useful even if the cleaner
`signal is no more intelligible than the noisy one.
`
`II. THE BASIC STRUCTURES
`
`Two basic channel vocoder configurations for implementing spectral
`subtraction were simulated. For reasons that will become apparent
`from the following descriptions, we call these configurations self-ex­
`cited and pitch-excited, respectively.
`
`2.1 The self-excited
`
`configuration
`
`A block diagram of the self-excited method of noise removal is
`shown in Fig. 2. The noisy speech, sampled 10,000 times per second is
`first passed through a bank of Í equispaced bandpass filters that span
`the telephone channel bandwidth (approximately 200 to 3200 Hz). The
`processing of the output of the bandpass filter is identical for each
`
`1850 THE BELL SYSTEM TECHNICAL JOURNAL, OCTOBER 1981
`
`Exhibit 1022
`Page 04 of 13
`
`

`

`ESTIMATE
`NOISE
`LEVEL
`
`yk
`
`ESTIMATE
`O V E R A L L
`SIGNAL LEVEL
`
`Yk
`
`ESTIMATE
`CORRECTED
`GAIN
`
`Sk
`
`L-4
`
`Yk'Yk
`
`Fig. 2—Block diagram of the self-excited channel bank noise stripper consisting of a
`bank of Ν
` FIR bandpass filters with gain estimation and correction within each channel.
`
`channel. In the kth channel, the following operations are performed
`on the output yk'-
`(i) The level (magnitude) of the noisy speech signal, Y*, is esti­
`mated.
`( « ) In a parallel path the level of the noise, Nk, is estimated.
`(iii) The estimates Nk and Yk are used to derive an estimate § k of
`the level of the uncorrupted speech signal in the kth channel.
`(iv) The adjusted channel signal is computed by the relation
`
`Sk = y k ^ .
`Ik
`
`( 2 )
`
`Clearly s* has the desired estimated magnitude § k . The sum s = ÓΑ'-
`sk then provides the final processed output.
`
`Ι
`
`2.2 The pitch-excited
`
`configuration
`
`A block diagram of the pitch-excited method is shown in Fig. 3. The
`estimates §k, k = 1, 2, · · · N, are obtained exactly as in the case of the
`self-excited configuration. However, the adjusted channel signals are
`obtained differently.
`(t) The noisy speech signal is first processed by a pitch extractor
`which also provides the voiced/unvoiced classification. The particular
`pitch extractor used is described in Ref. 11.
`(ii) The output of the pitch extractor is used to provide a clean
`excitation signal which consists of a Gaussian noise during unvoiced
`
`SPEECH SIGNAL 1851
`
`Exhibit 1022
`Page 05 of 13
`
`

`

`Fig. 3—Block diagram of the pitch-excited channel bank noise stripper in which a
`voiced/unvoiced excitation is used in place of the bandpass channel signals.
`
`portions and a train of impulses at the pitch rate during voiced
`segments.
`(iii) This clean excitation signal is passed through a bank of band­
`pass filters, identical to the ones shown in Fig. 2, to give channel
`signals, s*, which are approximately equal in magnitude.
`(iv) The adjusted channel signal is computed as
`
`sk = Sk-§k-
`
`(3)
`
`As before, s* has the correct magnitude and, as before, the sum of
`these adjusted channel signals gives the final processed output.
`As discussed in the next section, the estimates of •§* are computed
`every 0.01 s (i.e., 100 times a second). In our initial experiments the
`channel gains were held constant between estimates. In this case, the
`gain jumps in value every 0.01 s, producing annoying audible clicks.
`These clicks were eliminated by replacing each jump by a linear
`interpolation of the channel gains over 6 speech samples (i.e., over 0.6
`ms).
`
`III. ALTERNATIVE CONFIGURATIONS SIMULATED
`
`Several modified versions of the basic configurations of Figs. 2 and
`3 have been simulated, and several sentences processed with these
`simulations. The alternatives that we have studied in some detail are
`two choices for the number of channels; two methods of estimating
`
`1852 THE BELL SYSTEM TECHNICAL JOURNAL, O C T O B E R 1981
`
`Exhibit 1022
`Page 06 of 13
`
`

`

`Y*; two methods of estimating Nk; and two methods of estimating §/,.
`These will now be described.
`
`3.1 The filter bank
`
`Two designs were simulated, each with equispaced filters. In one
`design 16 channels (200-Hz wide) were used, and in the other 32
`channels (100-Hz wide). The filter responses and the sum of the
`responses for each design are shown in Fig. 4. (Each filter was a linear
`phase, finite impulse response (FIR) filter of duration 88 samples in the
`16-channel filter bank and 176 samples in the 32-channel filter bank.)
`
`3.2 Estimating Y„
`
`The two methods of estimating the magnitude, Y*, of the noisy
`channel signal are shown in Fig. 5. Either \yk | or y\ is low-pass-filtered
`to 30 Hz. In the second case, the square-root of the output of the low-
`pass filter is computed. The impulse and frequency responses of the
`low-pass filter [a 3rd order infinite impulse response (IIR) Bessel filter]
`are shown in Fig. 6.
`The choice of bandwidth of the low-pass filter is governed by a
`compromise between the following two requirements: For accurate
`estimation of Y* the averaging time should be as large as possible, i.e.,
`
`0
`
`2500
`FREQUENCY IN HERTZ
`
`5000
`
`Fig. 4—(a) Frequency responses of individual filters of the 16-channel filter bank,
`(b) Composite responses for 16-channel filter bank, (c) Frequency responses of individ­
`ual filters of the 32-channel filter bank, ( d ) Composite responses for 32-channel filter
`bank.
`
`SPEECH SIGNAL 1853
`
`Exhibit 1022
`Page 07 of 13
`
`

`

`
`
`MAGNITUDE MAGNITUDE
`
`1". I
`
`LP
`30 Hz
`
`
`
`SQUARER SQUARER
`
`2
`
`yk
`
`( a )
`
`
`LP LP
`
`30 Hz 30 Hz
`
`(b)
`
`Fig. 5—Signal processing for estimating the overall signal level using either a mag­
`nitude (a) or a squaring (b) nonlinearity, followed by a low-pass filter. In the case of the
`squaring nonlinearity, the low-pass filter is followed by a square root box.
`
`the filter bandwidth should be as small as possible. On the other hand,
`the spectrum of speech varies with time so the bandwidth should be as
`large as possible to track these variations. The usual compromise cut­
`off frequency in channel vocoders is about 30 Hz.
`Note that the outputs of the low-pass filters need be sampled only
`60 times/s. T o allow for the roll-off of the filters, the sampling rate was
`chosen as 100/s. Somewhat surprisingly, a much higher sampling rate
`was found to degrade performance. We will explain this paradox in
`Section IV (The Musical Tones).
`
`3.3 Estimating Nk
`
`During intervals of silence in the speech, the input signal consists of
`noise alone. Therefore, one possible estimate for Nk is the smallest
`value attained by Yk. However, because of statistical fluctuations, Yk
`quite rapidly takes on an unrealistically low value. Therefore, this
`estimate is quite unsatisfactory. In order to avoid such problems with
`outliers, the method schematized in Figure 7 has been simulated.
`As a first step, the magnitude of yk is estimated by a procedure
`identical to that of Fig. 5, except that the low-pass filter has a cut-off
`frequency of 10 Hz instead of 30 Hz. (The impulse response of the 10-
`Hz filter is quite similar to that of the 30-Hz filter with the time axis
`scaled by a factor of 3.)
`As before, the cut-off frequency of the low-pass filter should be
`chosen no larger than that necessary to follow the time-variations of
`the noise spectrum. Our choice of 10 Hz is an extremely conservative
`value. For most applications a cut-off frequency of 1 Hz or less should
`suffice.
`Analogously to the estimation of Y*, we have two ways of estimating
`
`1854 THE BELL SYSTEM TECHNICAL JOURNAL, OCTOBER 1981
`
`Exhibit 1022
`Page 08 of 13
`
`

`

`Nk, which differ only in the type of nonlinearity used. Figure 7 shows
`the front end of the alternate noise estimator that we have simulated.
`Let Zk(n) be the estimates of the magnitude of y* obtained by one of
`these methods, sampled every 0.01 s. Then the algorithm for finding
`the noise level is as follows:
`(i) Store Z*(n), ç = 1, • · • , Q in a buffer of size Q.
`(ii) Find the smallest value such that the next higher value is
`within 6 dB of it. Call this smallest value M I N .
`(Hi) Make a histogram with 1-dB bins of all the values that lie in
`the range M I N to M A X = M I N + 15 dB.
`(vi) Declare Ę times the magnitude corresponding to the peak of
`the histogram, as the noise level.
`(v) Get next sample.
`(vi) If this sample is greater than M A X , discard it and go to step
`(v).
`(vii) If the sample is less than M A X replace the oldest sample in
`the buffer by the new sample and go to step (ii).
`
`IMPULSE RESPONSE
`
`0
`
`100
`
`300
`200
`SAMPLE NUMBER
`
`400
`
`500
`
`FREQUENCY RESPONSE
`
`(b)
`
`0
`
`2500
`FREQUENCY IN HERTZ
`
`5000
`
`Fig. 6—Impulse response (a) and frequency response (b) of the 30-Hz, 3rd order,
`Bessel IIR filter used in estimating overall signal level.
`
`S P E E C H S I G N A L
`
`1855
`
`Exhibit 1022
`Page 09 of 13
`
`

`

`M A G N I T U D E
`M A G N I T U D E
`
`L P
`1 0 H z
`
`zk
`
`N O I S E E S T I M A T I O N N O I S E E S T I M A T I O N
`
`B Y P R O C E D U R E
`D E S C R I B E D
`I N
`S E C T I O N I I I
`
`(a)
`
`"k
`
`S Q U A R E R
`S Q U A R E R
`
`L P
`1 0 H z
`
`/
`V
`
`zk
`
`N O I S E E S T I M A T I O N
`B Y P R O C E D U R E
`D E S C R I B E D
`I N
`S E C T I O N I I I
`
`( b )
`
`Fig. 7—Signal processing for estimating noise level. In (a) and ( b ) the estimates of
`channel levels are obtained exactly as in ( a ) and (b) of Fig. 6, except that the low-pass
`filters have a bandwidth of 10 H z instead of 30 Hz. T h e final step in both (a) and ( b ) is
`an adaptive noise estimation procedure based on a time-varying noise histogram.
`
`After some experimentation, Q = 100 and Κ
` = 3 or 3.5 were found to
`be most satisfactory for the range of s/n's considered. All experiments
`to be described later were performed with these values of Q and K.
`Careful considerations of the above algorithm should convince the
`reader that this procedure ignores occasional low values of Zk\ it guards
`against sudden increased in Zk because of the onset of speech; and
`finally, it allows adaptation to a slowly varying noise level.
`
`3.4 Estimating S\
`
`As mentioned in the introduction, under the assumption that s* and
`rik are uncorrected, § k should be estimated as S* = (Yl —
`Nl)l/2.
`However, there is statistical fluctuation because of the finite averaging
`time even if the assumption is strictly valid. Therefore, sometimes the
`estimated value of Y* is less than that of iV*. In such cases, at is set to
`zero. Thus, our first procedure for estimating
`is
`
`& = VY£ - Nl,
`
`Yh > Nk
`
`= 0,
`
`Yk < Nk.
`
`A second estimate that we have tried is
`
`Sk=Yk-
`
`Nk,
`
`Yk > Nk
`
`= 0,
`
`Yk < Nk.
`
`(4a)
`
`(4b)
`
`(5a)
`
`(5b)
`
`IV. THE MUSICAL TONES
`
`We have processed several speech signals through a variety of noise-
`stripping algorithms obtained by selecting from the alternatives listed
`above. The results will be discussed in detail in the next section.
`
`1856 T H E B E L L S Y S T E M T E C H N I C A L J O U R N A L , O C T O B E R 1 9 8 1
`
`Exhibit 1022
`Page 10 of 13
`
`

`

`However, one general observation that can be made is that although
`the noise can be eliminated even from severely noisy speech signals, it
`gets replaced by "musical tones." These are short bursts of more or
`less sinusoidal tones with varying pitch. The explanation of the origin
`of these tones is as follows: The algorithm for estimating S* will, in
`general, set several consecutive channels to zero (in the valleys between
`formant peaks). If because of statistical fluctuation a single channel
`escapes elimination, (i.e., is above the noise threshold) it will appear as
`a narrow band signal much like a tone-burst with the center frequency
`of the channel. Every time such a tone-burst appears, its pitch will be
`determined by the particular isolated channel that gives rise to it. (At
`this point, the paradox mentioned in Section 3.2 can be explained.
`Consider a channel where the speech energy is very low, i.e., Y* » Nk.
`If Yk is oversampled, the number of times it crosses N/, increases and,
`therefore, the number of spurious noise bursts also increases.)
`We have found one simple procedure to combat this phenomenon.
`Every time the channel gains S* are updated, the new values are
`scanned across channels (i.e., the array S*(n), k = 1, · · · Í is examined
`at the time instant n). A nonzero value which is flanked by zero on
`both sides, is set equal to zero.
`For male voices, this removal of isolated channels works extremely
`well. However, the method does not work well for high-pitched female
`voices when the noise level is high. The reason is that in the latter
`case there may be only one or two pitch harmonics in a formant peak.
`Thus, the noise stripping algorithm might create several isolated
`channels in formant regions as well. Therefore, removal of isolated
`channels removes a large part of the speech signal, along with the
`musical tones. We do not have a good method of dealing with this
`problem for high-pitched voices at high noise levels.
`Suggestions for combatting these musical tones have also been made
`by Boll and Berouti et al. 6 , 7 We have compared our method to these
`other methods and find that except in the case of high-pitched voices
`at very low s/n our method performs better.
`
`V. EXPERIMENTS
`
`We have processed several sentences spoken by male and female
`speakers through noise-strippers obtained by selecting most of the
`possible combinations of alternatives listed in Section II. Uncorrelated
`Gaussian noise was added to provide the noisy test samples. The
`variance of the Gaussian distribution was selected so as to provide
`several s/n's in the range of about 4 to 16 dB. We have not conducted
`formal listening tests on the outputs. However, informal listening
`
`SPEECH SIGNAL 1857
`
`Exhibit 1022
`Page 11 of 13
`
`

`

`(mostly by the three authors of this report) allows us to draw the
`following general conclusions:
`(i) The algorithm is capable of following slow variations of the
`noise spectrum. We tested this on noise with a flat spectrum but with
`a sudden jump of 6 dB in its amplitude. The algorithm attained the
`correct estimates of channel gains within 0.5 s.
`(ii) The implementation with the 32-channel filter bank performs
`better than the one with the 16-channel filter bank.
`(iii) In Fig. 5 the second alternative performs significantly better
`than the first, i.e., the square root of the average power is a better
`statistic to use than the average of the magnitude.
`(iv) Power subtraction feqs. (4a, 4b)] and spectral magnitude sub­
`traction [eqs. (5a, 5b)] appear to work about equally well even at the
`lowest s/n (about 5 dB) that we tried.
`(v) The factor Ę in the noise estimation procedure of Section III,
`should be set to about 3 or 3.5 for the range of s/n's considered in this
`paper.
`(vi) For male voices, if isolated channels are eliminated as discussed
`in Section III, then pitch excitation and self excitation both work about
`equally well.
`(vii) For female voices it is not possible to remove isolated channels
`at high noise levels (s/n's less than say 8 dB). In these situations, pitch
`excitation is superior to self excitation.
`
`VI. CONCLUSION
`
`We have described several algorithms based on spectral subtraction
`for removing noise from a noisy speech signal. Two noteworthy fea­
`tures of our simulations are the manner in which we estimate the noise
`level and the manner in which we deal with the narrow-band, time-
`varying noise bursts that commonly arise in spectral subtraction
`methods.
`Our simulations were arranged to provide flexibility to allow us to
`test various modifications. However, it should be possible to realize
`the final preferred version of our algorithm in digital hardware that
`runs in real time.
`The ultimate test of such a system is a large-scale statistical study
`of listeners' preference. We have not attempted such a study. However,
`on the basis of informal listening we can say that our method is quite
`successful in removing noise, and in most instances is superior to the
`other methods known to us.
`
`REFERENCES
`
`1. M. R. Schroeder, U.S. Patent No. 3,180,936 filed December 1960, issued April, 1965.
`
`1858 THE BELL SYSTEM TECHNICAL JOURNAL, OCTOBER 1981
`
`Exhibit 1022
`Page 12 of 13
`
`

`

`2. This simulation by M. M. Sondhi and S. Sievers is documented only in an unpub­
`lished internal report. The procedure is described briefly in a review paper by M.
`R. Schroeder and A. M. Noll, Paper A21, 5th Int. Congress on Acoust., Liege,
`Belgium, 1965. The device was also patented by M. R. Schroeder, U.S. Patent No.
`3,403,224, filed May 1965, issued September, 1968.
`3. M. M. Sondhi, "New Methods of Pitch Extraction," IEEE Trans. Audio, AU-6, No.
`2 (June 1968), pp. 262-6.
`4. A. M. Noll, "Cepstrum Pitch Determination," J. Acoust. Soc. Am, 41, No. 2
`(February 1967), pp. 293-309.
`5. M. R. Weiss, E. Aschkenasy, and T. W. Parsons, "Processing Speech Signals to
`Attenuate Interference," IEEE Symp. on Speech Recognition, Pittsburgh, April
`1974, Contributed Papers, pp. 292-3.
`6. S. F. Boll, "Suppression of Acoustic Noise in Speech Using Spectral Subtraction,"
`IEEE Trans. Acoust. Speech and Sig. Process., ASSP-29, No. 2 (April 1979), pp.
`113-20.
`7. M. Berouti, R. Schwartz, and J. Makhoul, "Enhancement of Speech Corrupted by
`Acoustic Noise," Proc. Int. Conf. Acoust. Speech, and Sig. Process. (April 1979),
`pp. 208-11.
`8. R. j . McAulay and M. L. Malpass, "Speech Enhancement Using a Soft-Decision
`Maximum Likelihood Noise Suppression Filter," Tech. Note 1979-31, MIT Lincoln
`Lab., Lexington, Ma., June 1979.
`9. J. S. Lim and Α
`. V. Oppenheim, "Enhancement and Bandwidth Compression of
`Noisy Speech," Proc. IEEE, 67, No. 12 (December 1979), pp. 1586-604.
`10. I. B. Thomas and R. J. Niederjohn, "Enhancement of Speech Intelligibility at High
`Noise Levels by Filtering and Clipping," J. Aud. Eng. Soc, 16, No. 10 (October
`1968), pp. 412-15.
`11. R. A. GiÛman, "A Fast Frequency Domain Pitch Algorithm," J. Acoust. Soc. Am.,
`58, Supplement No. 1 (Fall 1975), p. S62.
`
`SPEECH SIGNAL 1859
`
`Exhibit 1022
`Page 13 of 13
`
`