Estimation of Noise Spectrum and its Application to SNR-
`Estimation and Speech Enhancement
`
`H.Günter Hirsch
`
`Technical Report TR-93-012
`
`International Computer Science Institute,
`Berkeley, California, USA
`
`1
`
`IPR No. 2017-00627
`Apple Inc. v. Andrea Electronics Inc. - Ex. 1032, p. 1
`
`

`

`Estimation of Noise Spectra and its Application to SNR-Estimation and
`Speech Enhancement
`
`H. Günter Hirsch
`
`Contents
`
`1. Introduction
`
`2. Principal Idea
`
`3. Practical Realization
`
`4. Signal-to-Noise Ratio (SNR) Estimation
`
`5. Speech Enhancement
`
`6. Conclusions
`
`7. References
`
`2
`
`IPR No. 2017-00627
`Apple Inc. v. Andrea Electronics Inc. - Ex. 1032, p. 2
`
`

`

`1. Introduction
`
`At the time of this writing, some experiments in robust speech recognition had already
`been done at ICSI.
`The original goal of this work was improved recognition of speech recorded with different
`microphones and transmitted over channels with different frequency characteristics. One
`practical application of this is the recognition of speech recorded via telephone lines
`where you have microphones and channels with different transmission characteristics. It
`could be shown that the recognition rates can be improved when introducing a high-pass
`filtering of the logarithmic spectral envelopes in subbands /1/.
`This idea is based on the fact that a frequency characteristic corresponds to a multiplica-
`tion of the speech spectrum with the frequency response of the transmission channel.
`The result would be a constant additive component in the logarithmic spectral envelopes
`in subbands (assuming a nearly constant transmission characteristic). Because of this, a
`high-pass filtering leads to a suppression of these constant components.
`
`Another aspect is the superposition of noise in many applications of speech recognizers
`in real environments, e.g. voice dialing in a car or serving any kind of machines on the
`street or in workshops. This noise would result in a nearly constant additive component
`to the magnitude spectral envelopes in subbands (assuming a nearly stationary noise) .
`It could be shown that recognition rates can be improved by high-pass filtering the mag-
`nitude spectral envelopes /2/.
`
`Additive noise as well as a certain frequency characteristic are present in many real situ-
`ations. One way to handle both effects could be to use a combination of processing in
`the magnitude as well as in the logarithmic spectral domain. Another possibility could be
`a processing anywhere between the magnitude and the logarithmic domain dependent
`of the amount of noise in the specific situation. This would presuppose an estimation of
`the signal-to-noise ratio (SNR).
`
`Looking at the first possibility several processing techniques are well known to reduce
`the noise in the magnitude spectral domain. One could be the already mentioned high-
`pass filtering. A disadvantage of this method is the suppression of certain spectral fea-
`tures in speech segments. Introducing a high-pass filter with a total suppression of the
`DC component, not only the constant noise components are suppressed but also the
`constant component of the speech. Because of this just the spectral features of the pho-
`nemes with less energy are reduced in the case of a preceding phoneme with higher
`energy and spectral components in the same subbands.
`
`One solution could be a kind of nonlinear filtering with the goal of preserving the spectral
`features of the phonemes with less energy on one hand but suppressing the noise com-
`ponents on the other hand. Another method to reduce the noise is the well known spec-
`tral subtraction technique /3/,/4/. This technique is based on the estimation of the noise
`spectrum during speech pauses and an adaptive filtering with the estimated noise spec-
`trum. A major disadvantage is the necessity of the detection of speech pauses to esti-
`mate the noise spectrum. This is a very difficult and ultimately unsolved problem for
`
`3
`
`IPR No. 2017-00627
`Apple Inc. v. Andrea Electronics Inc. - Ex. 1032, p. 3
`
`

`

`realistic situations with a varying noise level. Another disadvantage is the fact that the
`algorithm cannot adapt to a varying noise level during segments of speech. The adaptive
`filtering is always based on the estimated noise spectrum of the preceding speech
`pause.
`
`An improvement of the spectral subtraction technique would be an estimation of the
`noise spectrum without the necessity of a speech pause detection. A method is pre-
`sented in this report to estimate the noise spectrum without a speech pause detection.
`
`One application for this method presented in this report is the estimation of the actual
`SNR of a noisy signal. Furthermore the technique is applied to speech enhancement
`based on a spectral subtraction respectively on a nonlinear high-pass filtering of the
`spectral envelopes dependent on the actual SNR.
`
`4
`
`IPR No. 2017-00627
`Apple Inc. v. Andrea Electronics Inc. - Ex. 1032, p. 4
`
`

`

`2. Principal Idea
`
`The principal idea to estimate the noise level in a certain subband is based on a statisti-
`cal analysis of a segment of the magnitude spectral envelope.
`
`Looking at such a spectral envelope in figure 2.1 and the corresponding distribution den-
`sity function in figure 2.2 the most commonly occurring spectral magnitude value is zero.
`The spectral envelope was calculated for a clean speech signal with a duration of about
`4.5 s and in a subband of about 500 Hz. The distribution density function was calculated
`for the whole duration of 4.5 s with an accuracy of about 1 percent in regard to the maxi-
`mum spectral value inside this subband. The function is shown for the range of 0 to 50
`percent of the maximum. Only a few values occur which are higher than 50 percent.
`
`spectral magnitude
`
`time/s
`
`Figure 2.1: Spectral envelope in a band with a centre frequency of 500 Hz
`
`Noise was added artificially to this speech signal to produce different SNRs. The results
`can be seen in figure 2.3. The noise was a bandpass limited Gaussian noise with a cen-
`tre frequency of 500 Hz and a bandwidth of 200 Hz.
`
`An increase of the maximum value in the distribution function can be observed for a
`decreasing SNR. This most frequently occurring value can be taken as an estimation for
`the noise level inside this band.
`Also, an increasing variance of the spectral magnitude values of the noise can be seen
`for an decreasing SNR. Because of a broad distribution the estimation isn’t so accurate
`
`5
`
`IPR No. 2017-00627
`Apple Inc. v. Andrea Electronics Inc. - Ex. 1032, p. 5
`
`

`

`distribution
`density
`
`Figure 2.2: Distribution density function for the spectral envelope in figure 2.1
`
`magnitude/maximum
`
`for channels with a low SNR. A reduction of the accuracy for low SNRs has the effect of
`smoothing the distribution function and improving the maximum detection. On the other
`hand the accuracy has to be high for channels with less noise to get a reasonable
`estimation for the amount of noise at all. Because of this the accuracy for the calculation
`of the distribution density function is made dependent on the actual SNR inside one
`channel. The accuracy is less for channels with a bad SNR and higher for channels
`with a better SNR.
`
`6
`
`IPR No. 2017-00627
`Apple Inc. v. Andrea Electronics Inc. - Ex. 1032, p. 6
`
`

`

`SNR = 15 dB
`
`SNR = 5 dB
`
`SNR = -5 dB
`
`Figure 2.3: Spectral envelopes and distribution functions for different SNRs
`
`7
`
`IPR No. 2017-00627
`Apple Inc. v. Andrea Electronics Inc. - Ex. 1032, p. 7
`
`

`

`Some channels don’t show the good behaviour seen in figure 2.3. One result can be
`seen in figure 2.4 for a channel with a high centre frequency of about 2950 Hz. There is
`nearly no speech energy but a significant noise energy in this channel. The noise was a
`Gaussian noise in this case.
`
`Figure 2.4: Spectral envelope and the corresponding distribution density function of a fre-
`quency band with a centre frequency of 2590 Hz
`
`Sometimes a maximum respectively a noise level is calculated with an unrealistic high
`value in these cases. Because of this the possible estimated noise level is limited to the
`average spectral value. This is related to the fact that no noise energy can occur which is
`higher than the total amount of energy inside a band.
`
`Another channel with a nonideal behaviour is shown in figure 2.5. This is a channel with
`a centre frequency of about 219 Hz where the spectral magnitude of the signal often
`takes a very high value.
`
`8
`
`IPR No. 2017-00627
`Apple Inc. v. Andrea Electronics Inc. - Ex. 1032, p. 8
`
`

`

`spectral
`magnitude
`
`time/s
`
`Figure 2.5: Spectral envelope in a subband with a centre frequency of 219 Hz
`
`One problem for this envelope occurs with a statistical analysis in a short time window of
`e.g. 500 ms around the time of 2s. The signal takes nearly only high values inside this
`window so that it will be impossible to estimate a reasonable noise level. To avoid this
`problem we must detect channels with such a behaviour and to extend the analysis win-
`dow in these cases. Usually these are only channels with a centre frequency less than
`500 Hz and with high energy. Both these criteria are used for the detection of channels
`with a possible behaviour similar to the spectral envelopes of figure 2.5.
`
`9
`
`IPR No. 2017-00627
`Apple Inc. v. Andrea Electronics Inc. - Ex. 1032, p. 9
`
`

`

`3. Practical realization
`
`In a practical application the calculation must be done online with running speech. The
`length of a window has to be defined for the calculation of the distribution density func-
`tion. On one hand, the length should be as high as possible to increase the accuracy of
`the noise level estimation. On the other hand, the length may not exceed a certain dura-
`tion if a signal with a varying noise level should be analyzed. The spectral analysis is
`done with a universal program for a short-term spectral analysis.
`
`The FFT length used was 256 so that the centre frequencies of the estimated spectrum
`have a distance of 31.25 Hz at a sampling frequency of 8000 Hz. The window for weight-
`ing the speech samples was a Kaiser window multiplied by a sinc function. The influence
`of different window types was not examined but it could be assumed that it is not very
`high. The spectrum is calculated every 8 ms so that the magnitude spectral values inside
`one band are given with a sampling frequency of 125 Hz.
`
`The first estimation of the noise spectrum used the average of the first ten magnitude
`spectral values, calculated within each band. It is assumed that the first incoming speech
`samples of a recording are related to the noise. This noise estimation is used up to a
`time of 250 ms after starting the recording. Afterwards a window with an increasing
`length is used up to the final window length for the calculation of the distribution density
`function. We finally considered window lengths of 250ms, 500ms, 1s and 2s.
`
`The distribution density function is calculated for the magnitude spectral values inside
`the window for each band. The function is computed in a range from 0 to the maximum
`spectral value which was found in this band up to this time. The accuracy is 0.25 per cent
`of the maximum which corresponds to dividing the whole range into 400 intervals. A
`search for the maximum peak value is done where at first an accuracy of 2 % for the dis-
`tribution function is used. This is done by summing 8 neighboring values of the function,
`corresponding to a smoothing of the distribution density function.
`
`If the maximum value is higher than 10 % of the maximum spectral value inside this band
`the estimated value is taken directly. If the detected maximum is anywhere in the range
`of 5 to 10 per cent an accuracy of 1 % is used, in the range of 2.5 to 5 % an accuracy of
`0.5 % and under 2.5 % the highest available accuracy of 0.25 % is used for a more accu-
`rate noise level estimation. This kind of analysis is related to the fact that on one hand a
`more smoothed version of the distribution function should be used for the noise level
`estimation in cases of a low SNR inside a band. On the other hand the resolution has to
`be higher for high SNRs where you have only a small amount of noise in one channel.
`The fixing of the analysis intervals and the accuracy was done empirically.
`
`The following processing is done to avoid a poor estimation of the noise level inside
`some low frequency bands where spectral magnitude values very often occur with a high
`value (already mentioned in the preceding section). The five channels with the highest
`amount of energy are calculated by looking back 1 s of time to the past. An upper limit for
`
`10
`
`IPR No. 2017-00627
`Apple Inc. v. Andrea Electronics Inc. - Ex. 1032, p. 10
`
`

`

`the noise level inside each of these bands is estimated by taking an average of a certain
`amount of the smallest spectral values of the last second. Then a distribution density
`function is calculated with a variable window length. Up to 50 values of the past are con-
`sidered which take a value between 0 and the upper limit of the noise level. The maxi-
`mum of the distribution density function is calculated as mentioned before.
`
`The estimated magnitude spectral values are smoothed in all cases by computing a
`weighted sum of the actual estimated noise level and all estimated values of the past.
`The weighting is done with an exponentially decaying curve.
`
`A result of this kind of noise estimation can be seen in figure 3.2. There is shown the
`average estimated spectrum of a noise signal. The estimation was done after adding the
`noise signal to a nearly clean speech signal. The length of the analysis window for the
`distribution density function was 500 ms. The average spectrum of the noise itself is
`shown in figure 3.1. The averaging was done for the whole noise signal.
`
`The estimation of the noise spectrum appears to work well. However, some overestima-
`tion can be seen for a frequency component at about 750 Hz with high energy.
`
`Some further estimated noise spectra are shown in figure 3.3 for additive noise resulting
`in different SNRs. The estimation seems to be nearly independent of the SNR. Some
`small differences can be seen for the case of a high SNR of 25 dB. The reason for this is
`the influence of the “clean” speech signal itself. The speech was recorded at a SNR of
`about 35 to 40 dB. This noise of the “clean” speech can already be seen for example in
`the region of higher frequencies.
`
`A spectrum of another noise signal and some corresponding noise spectral estimates
`are shown in figure 3.4.
`
`11
`
`IPR No. 2017-00627
`Apple Inc. v. Andrea Electronics Inc. - Ex. 1032, p. 11
`
`

`

`Figure 3.1: Average magnitude spectrum of a noise signal
`
`frequency/Hz
`
`Figure 3.2: Average estimated magnitude noise spectrum of a noisy speech signal with a
`SNR of 5 dB
`
`frequency/Hz
`
`12
`
`IPR No. 2017-00627
`Apple Inc. v. Andrea Electronics Inc. - Ex. 1032, p. 12
`
`

`

`SNR = -5 dB
`
`SNR = 15 dB
`
`SNR = 25 dB
`
`frequency/Hz
`
`Figure 3.3: Estimated magnitude noise spectra for different SNRs
`
`13
`
`IPR No. 2017-00627
`Apple Inc. v. Andrea Electronics Inc. - Ex. 1032, p. 13
`
`

`

`noise spectrum
`
`SNR = -5 dB
`
`SNR = 5 dB
`
`SNR = 15 dB
`
`SNR = 25 dB
`
`Figure 3.4: Average noise and estimated noise spectra for different SNRs
`
`14
`
`IPR No. 2017-00627
`Apple Inc. v. Andrea Electronics Inc. - Ex. 1032, p. 14
`
`

`

`4. SNR (signal-to-noise ratio) estimation
`
`The estimation of the noise spectrum can be used for an estimation of the SNR (signal-
`to-noise ratio). Despite the well-known engineering meaning of the acronym, SNR must
`first be defined in further detail.
`
`In the field of speech coding the term SNR is often used as a so called segmental SNR
`where the SNR is actually only a description of the ratio of speech energy to noise
`energy for a short time period, e.g. 20 ms. Often the expression is really interpreted as a
`long-term measurement, especially for stationary noise situations. In this case the aver-
`age energy of the speech is usually calculated over a longer time period, at least some
`seconds and including speech pauses. The noise energy is calculated for the same time
`period assuming that the noise is nearly stationary during this time. This can be ideally
`done when artificially mixing a speech and a noise signal with an optional SNR.
`
`In this application the term SNR can apply to any of a range of temporal scales.
`On one hand it should be more related to the long-term SNR. But on the other hand it
`should be possible to follow slowly changing noise situations as they occur in real situa-
`tions.
`
`Given a real noisy signal it is difficult to directly estimate the SNR without any specific
`knowledge about the speech or noise energy. Because of this the noise to signal-plus-
`noise ratio N/(S+N) was considered instead of the signal to noise ratio itself.
`
`An estimation of the short-term energy Nenergy of the noise at a specific time t is calcu-
`lated with Parseval’s relation..
`
`Nenergy t( )
`
`=
`
`nfft∑
`
`1
`nfft
`
`
`
`Nspec iΔf t,(
`
`)
`
`0=
`i
`where Nspec(iΔf) is the estimated spectral magnitude of the noise in a subband with a
`centre frequency of iΔf where Δf = <sampling frequency> divided by nfft and nfft = <FFT-
`length>.
`
`The short term energy of the noisy signal x is calculated in the same way.
`
`Xenergy t( )
`
`=
`
`nfft∑
`
`1
`nfft
`
`
`
`Xspec iΔf t,(
`
`)
`
`0=
`i
`where Xspec is the spectral magnitude of the noisy speech signal.
`The average energy of the noisy signal can not be calculated over a longer period of
`speech, (e.g. several seconds), in this application because the interactive character of
`
`15
`
`IPR No. 2017-00627
`Apple Inc. v. Andrea Electronics Inc. - Ex. 1032, p. 15
`
`

`

`the task precludes a long initial delay.
`
`The average energy X(t) is calculated for a past segment where the length of this window
`corresponds to the window length for the calculation of the distribution density function
`for the noise level estimation.
`
`If the actual value of this average energy X(t) would always be used, the result of the
`relation N(t)/X(t) would be a kind of segmental N/(S+N) ratio. Instead, the maximum of
`the energy X(t) up to this time is used. This maximum is slowly decreased with an expo-
`nential decay to adapt to an overall change of the signal level so as not to use a local
`peak value of X over a long time.
`
`A result for the estimation of N(t)/X(t) can be seen in figure 4.3. A sentence with a dura-
`tion of about 7s was used as a speech signal. A Gaussian noise was artificially added
`with a SNR of 5 dB. The time signal is shown in figure 4.1. The estimation of the noise
`level N(t) itself can be seen in figure 4.2. The noise estimation as well as the calculation
`of X(t) were done within a window of 1 s.
`
`amplitude
`
`Figure 4.1: Time signal of a noisy sentence
`
`time/s
`
`16
`
`IPR No. 2017-00627
`Apple Inc. v. Andrea Electronics Inc. - Ex. 1032, p. 16
`
`

`

`N(t)
`
`Figure 4.2: Estimation of the noise level N(t) for the signal of figure 4.1
`
`time/s
`
`N(t)/X(t)
`
`time/s
`
`Figure 4.3: Estimation of the relation N(t)/X(t) for the signal of figure 4.1
`
`The noise level estimation seems to be a little bit too high at the beginning of the speech.
`One reason for this is a noise inside the speech signal itself caused by breathing of the
`speaker. The relation N(t)/X(t) takes a high value of about 0.7 at the beginning because
`there is nearly only noise. Then the curve rapidly slopes when the first high values for
`X(t) are calculated at the beginning of speech activity. Later on the relation takes a nearly
`constant value of about 0.07.
`
`17
`
`IPR No. 2017-00627
`Apple Inc. v. Andrea Electronics Inc. - Ex. 1032, p. 17
`
`

`

`N/X
`
`N/X
`
`N/X
`
`SNR = -5 dB
`
`time/s
`
`SNR = 15 dB
`
`time/s
`
`SNR = 25 dB
`
`time/s
`
`Figure 4.4: Estimation of the relation N(t)/X(t) for different SNRs
`
`The calculation of the maximum of the average energy Xenergy(t) is actually not the same
`as computing the long term average of X. The value of Xenergy(t) is higher than a long
`term average. Because of this the relation doesn’t take the value of 0.24 in the constant
`
`18
`
`IPR No. 2017-00627
`Apple Inc. v. Andrea Electronics Inc. - Ex. 1032, p. 18
`
`

`

`part of the curve in figure 4.3 which it ideally should take for a SNR of 5 dB.
`Some further results using the signal of figure 4.1 are shown in figure 4.4 for different
`SNRs.
`
`N/X
`
`SNR/dB
`
`SNR/dB
`
`Figure 4.5: Average estimation of the relation N/X for different noise signals
`
`N/X
`
`Figure 4.6: Average estimation of the relation N/X for different noise signals
`
`19
`
`IPR No. 2017-00627
`Apple Inc. v. Andrea Electronics Inc. - Ex. 1032, p. 19
`
`

`

`An average of these estimated values is calculated for the last 4 s of the signal corre-
`sponding to the nearly constant part of the curve. The result of this averaging is shown in
`figures 4.4 and 4.5 for different SNRs and different noise signals where a window length
`of 500 ms was used for the noise estimation.
`
`Stationary segments of two naturally recorded signals and four artificially generated sig-
`nals were used as noise signals. The results are shown for a 40 dB range in dynamics.
`The noise signals were
`
`1) car noise
`2) computer room noise
`3) white Gaussian noise within a bandwidth of 0 to 4 kHz
`4) white Gaussian noise within a bandwidth of 0 to 0.5 kHz
`5) white Gaussian noise within a bandwidth of 0 to1 kHz
`6) white Gaussian noise within a bandwidth of 2 to 4 kHz
`
`The accuracy of the estimation slightly decreases for high SNRs. But overall a high cor-
`relation can be seen for the different noise signals. The accuracy is in a range of about 1
`to 2 dB. The results can be used to realize a mapping from the estimated Nenergy/Xenergy
`to the real SNR.
`
`The results when using different window lengths for the noise level estimation are shown
`in figure 4.7.
`
`The scaling of the ordinates is different for the different window lengths because of the
`calculation of X in the corresponding window. A much higher maximum of X occurs for
`the short window of 250 ms which is nearly comparable with the estimation of the energy
`of a vowel. No big difference can be seen for the different lengths of the window when
`comparing the correlations of SNR to the computed N/(S+N). However, these curves are
`the result of an averaging over 4 s so that the influence of the temporal fluctuations can
`not be seen.
`
`20
`
`IPR No. 2017-00627
`Apple Inc. v. Andrea Electronics Inc. - Ex. 1032, p. 20
`
`

`

`windowlength = 250 ms
`
`windowlength = 250 ms
`
`windowlength = 1 s
`
`windowlength = 1 s
`
`windowlength = 2 s
`
`windowlength = 2 s
`
`Figure 4.7: Average estimation of the relation N/X for different window lengths
`
`21
`
`IPR No. 2017-00627
`Apple Inc. v. Andrea Electronics Inc. - Ex. 1032, p. 21
`
`

`

`Some experiments were done adding noise with a varying SNR. A modulated Gaussian
`noise was added to the speech signal shown in figure 4.1 with an overall SNR of 10 dB.
`The modulation signal itself can be seen in figure 4.8. The result for the estimation of
`N(t)/X(t) is shown in figure 4.9 using an analysis window of 500 ms.
`
`Figure 4.8: Time signal for the modulation of a Gaussian noise
`
`N(t)/X(t)
`
`window length: 500 ms
`
`time/s
`
`time/s
`
`Figure 4.9: Estimated N(t)/X(t) for a signal disturbed by a modulated Gaussian noise
`
`The estimation of N/X follows this artificial modulation characteristic quite good. A delay
`of about 500 ms can be considered because of the analysis window in the past.
`
`22
`
`IPR No. 2017-00627
`Apple Inc. v. Andrea Electronics Inc. - Ex. 1032, p. 22
`
`

`

`The results for different window lengths are shown in figure 4.10.
`
`N(t)/X(t)
`
`window length: 250 ms
`
`N(t)/X(t)
`
`window length: 1s
`
`N(t)/X(t)
`
`window llength: 2s
`
`Figure 4.10: Estimated N(t)/X(t) for different window lengths
`
`time/s
`
`23
`
`IPR No. 2017-00627
`Apple Inc. v. Andrea Electronics Inc. - Ex. 1032, p. 23
`
`

`

`The delay is smaller for an analysis window of 250 ms but some errors occur for the
`noise level estimation. In the case of a 1 s window the curve doesnot fit the modulation
`characteristic as good as in the case of 500 ms. The length of 2 s is too high to follow the
`varying noise level.
`
`The length of the analysis window should be chosen for the particular noisy situations to
`which the processing is applied. A length of 500 ms seems to be a good compromise for
`the cases we have examined.
`
`Experiments with naturally recorded noisy speech signals have shown a good agree-
`ment of the estimated SNRs with the expected curves.
`
`24
`
`IPR No. 2017-00627
`Apple Inc. v. Andrea Electronics Inc. - Ex. 1032, p. 24
`
`

`

`5. Speech Enhancement
`
`The estimation of noise spectra can also be used for speech enhancement. Two applica-
`tions were considered in this study.One is the well-known spectral subtraction technique.
`The other one is based on a modified high-pass filtering of the spectral envelopes in sub-
`bands.
`
`Only the magnitude spectral values are processed in both cases. The phase of the noisy
`speech is used for the resynthesis. An existing program called “synthese” was used for
`the resynthesis, in which time signals are generated from processed spectra with the
`overlap-add method.
`
`The noise spectrum estimation described above (section 3) is used for the spectral sub-
`traction. The subtraction is applied in 128 bands. The estimated magnitude noise spec-
`tral value N(iΔf) is used for an adaptive weighting of X(iΔf), the spectral magnitude of the
`noisy speech, in each subband with a centre frequency iΔf.
`
`An estimation of the magnitude component of the speech is calculated as
`
`Sˆ
`
`(
`
`iΔf
`
`)
`
`=
`
`
`1 N iΔf(
`−(
`X iΔf(
`
`
`)
`)
`
`
`
`) X iΔf(
`
`)
`
`The contour of X(iΔf) is usually smoothed with an exponentially decaying weighting of
`past values.
`
`Various modifications of the weighting function 1 - (N(iΔf)/X(iΔf)) are possible /4/, /5/, e.g.
`the realization as a Wiener filter.
`
`It is possible that negative weighting factors occur. One solution is to set these factors to
`zero. The time signals of clean speech, of noisy speech where a car noise was added
`with a SNR of 5 dB and of the processed noisy speech are shown in figure 5.1.
`
`A considerable improvement of the SNR can be seen. However, listening to the resyn-
`thesized speech one can hear a new artificial noise that is often referred to as ‘‘musical
`tones’’. This degradation, common to spectral subtraction-based enhancement tech-
`niques, significantly disturbs the subjective impression. However, earlier experiments /6/
`had shown that this artificial noise has little influence on the recognition rates of an iso-
`lated word recognizer. Thus, recognition rates could be improved by the introduction of
`the spectral subtraction technique.
`
`25
`
`IPR No. 2017-00627
`Apple Inc. v. Andrea Electronics Inc. - Ex. 1032, p. 25
`
`

`

`8000
`
`6000
`
`4000
`
`2000
`
`0
`
`-2000
`
`-4000
`
`-6000
`0
`
`8000
`
`6000
`
`4000
`
`2000
`
`0
`
`-2000
`
`-4000
`
`-6000
`0
`
`8000
`
`6000
`
`4000
`
`2000
`
`0
`
`-2000
`
`-4000
`
`-6000
`0
`
`1
`
`2
`
`3
`
`4
`
`5
`
`6
`
`7
`
`time/s
`
`1
`
`2
`
`3
`
`4
`
`5
`
`6
`
`7
`
`time/s
`
`1
`
`2
`
`3
`
`4
`
`5
`
`6
`
`7
`
`time/s
`
`Figure 5.1: Time signals of clean and noisy speech and after processing with a spectral
`subtraction technique
`
`26
`
`IPR No. 2017-00627
`Apple Inc. v. Andrea Electronics Inc. - Ex. 1032, p. 26
`
`

`

`The second speech enhancement algorithm is based on a high-pass filtering of the spec-
`tral magnitude contours in subbands. The known filter function from /1/ is used in each
`subband. Applying this filtering directly to a spectral trajectory in one subband, many
`negative values occur as shown in figure 5.2.
`
`The envelope of clean speech, of noisy speech and the filtered envelope in a subband
`with a centre frequency of 500 Hz are shown. Setting all negative values to zero the
`noise is considerably reduced but also certain parts of the speech are suppressed.
`
`Because of this the filter structure is modified as shown in figure 5.3.
`
`spectral
`magnitude X(t,iΔf)
`
`1 - c(t,iΔf)
`
`Figure 5.3: Modified high-pass filtering
`
`c(t,iΔf)
`
`High-Pass
`Filter
`5
`
`f/Hz
`
`+
`
`The attenuation of the DC-component can be varied with this filter scheme with the factor
`c. In principal the attenuation should be high for noise segments and should be made
`dependent on the actual SNR for a speech segment. The weighting function 1 - N/X
`already used for spectral subtraction, is applied according to the additional weighting fac-
`tor c.
`
`The spectral trajectories of figure 5.2 are shown again in figure 5.4 but this time the pro-
`cessing is done with the modified filter structure.
`
`From this experiment, it appears that the suppression of speech segments is much less
`of a problem than in the case of a static high-pass filter. The result of processing a whole
`noisy speech sentence is shown in figure 5.5.
`
`The time signals of the clean, the noisy and the processed speech are plotted. A consid-
`erable improvement of the SNR can be obtained with this processing. The generation of
`musical tones seems to be a little bit less of a problem than in the case of spectral sub-
`traction.
`
`Some experiments with automatic recognition for a combination of this technique and
`Rasta-PLP are currently in progress and will be presented at ICASSP93 for signals cor-
`rupted by convolutional and additive noise.
`
`27
`
`IPR No. 2017-00627
`Apple Inc. v. Andrea Electronics Inc. - Ex. 1032, p. 27
`
`

`

`6
`
`5
`
`4
`
`3
`
`2
`
`1
`
`0
`
`0
`
`5
`
`4.5
`
`4
`
`3.5
`
`3
`
`2.5
`
`2
`
`1.5
`
`1
`
`0.5
`
`0
`
`0
`
`4
`
`3.5
`
`3
`
`2.5
`
`2
`
`1.5
`
`1
`
`0.5
`
`0
`
`-0.5
`
`0.5
`
`1
`
`1.5
`
`2
`
`2.5
`
`time/s
`
`0.5
`
`1
`
`1.5
`
`2
`
`2.5
`
`time/s
`
`-1
`0
`
`0.5
`
`1
`
`1.5
`
`time/s
`Figure 5.2: Spectral envelope of clean and noisy speech and after RASTA high-pass fil-
`tering
`
`2
`
`2.5
`
`28
`
`IPR No. 2017-00627
`Apple Inc. v. Andrea Electronics Inc. - Ex. 1032, p. 28
`
`

`

`6
`
`5
`
`4
`
`3
`
`2
`
`1
`
`0
`
`0
`
`5
`
`4.5
`
`4
`
`3.5
`
`3
`
`2.5
`
`2
`
`1.5
`
`1
`
`0.5
`
`0
`
`0
`
`5
`
`4
`
`3
`
`2
`
`1
`
`0
`
`0.5
`
`1
`
`1.5
`
`2
`
`2.5
`
`time/s
`
`0.5
`
`1
`
`1.5
`
`2
`
`2.5
`
`time/s
`
`-1
`0
`
`0.5
`
`1
`
`1.5
`
`time/s
`Figure 5.4: Spectral envelope of clean and noisy speech and after modified high-pass fil-
`tering
`
`2
`
`2.5
`
`29
`
`IPR No. 2017-00627
`Apple Inc. v. Andrea Electronics Inc. - Ex. 1032, p. 29
`
`

`

`8000
`
`6000
`
`4000
`
`2000
`
`0
`
`-2000
`
`-4000
`
`-6000
`0
`
`8000
`
`6000
`
`4000
`
`2000
`
`0
`
`-2000
`
`-4000
`
`-6000
`0
`
`8000
`
`6000
`
`4000
`
`2000
`
`0
`
`-2000
`
`-4000
`
`-6000
`0
`
`1
`
`2
`
`3
`
`4
`
`5
`
`6
`
`7
`
`time/s
`
`1
`
`2
`
`3
`
`4
`
`5
`
`6
`
`7
`
`time/s
`
`1
`
`2
`
`3
`
`4
`
`5
`
`6
`
`7
`
`time/s
`
`Figure 5.5: Time signals of clean and noisy speech and after processing with the modi-
`fied high-pass filtering
`
`30
`
`IPR No. 2017-00627
`Apple Inc. v. Andrea Electronics Inc. - Ex. 1032, p. 30
`
`

`

`6. Conclusions
`
`A method is presented in this report to estimate the noise spectrum of speech utterances
`which are disturbed by additive noise. One advantage is that no speech pause detection
`is required.
`
`The processing is based on a calculation of the distribution density function of spectral
`magnitude values in a subband. The histogram for one subband is calculated for a past
`segment with a defined duration. Good results were obtained for a segment of 0,5 s. In
`this case the noise spectrum can also follow a slowly changing noise.
`
`Two applications of this technique are described in this report. The first one is an estima-
`tion of the actual SNR of a speech segment. Good results were obtained for a wide
`range of SNRs and for different noise signals. The second one is the use for an enhance-
`ment of noisy speech. The enhancement techniques attempted were a new form of
`spectral subtraction, and a modified high-pass filtering of the spectral envelopes in sub-
`bands.
`
`31
`
`IPR No. 2017-00627
`Apple Inc. v. Andrea Electronics Inc. - Ex. 1032, p. 31
`
`

`

`7. References
`
`/1/
`
`/2/
`
`/3/
`
`/4/
`
`/5/
`
`/6/
`
`H. Hermansky, N. Morgan, A. Bayya, P. Kohn: Compensation for the effect of the
`communication channel in auditory-like analysis of speech (RASTA_PLP), Euro-
`speech, 1991, pp. 1367-1370
`
`H.G. Hirsch, P. Meyer, H.W. Rühl: Improved speech recognition using high-pass
`filtering of subband envelopes, Eurospeech, 1991, pp. 413-416
`
`S.F. Boll: Suppression of acoustic noise in speech using spectral subtraction,
`IEEE ASSP-28, No.2, 1979, pp.113-120
`
`P. Vary: Noise suppression by spectral magnitude estimation- Mechanism and
`theoretical limits, Signal Processing, 1985, pp. 387-400
`
`P. Lockwood, J. Boudy: Experiments with a nonlinear spectral subtractor, Hidden
`Markov Models and the projection, for robust speech recognition in cars, Speech
`Communication 11, 1992, pp. 215-228
`
`H.G. Hirsch, H.W. Rühl: Automatic speech recognition in a noisy environment,
`Eurospeech, 1989, pp. 652-655
`
`32
`
`IPR No. 2017-00627
`Apple Inc. v. Andrea Electronics Inc. - Ex. 1032, p. 32
`
`