A SPECTRO-TEMPORALVARYING APPROACH FOR SPEECH

ENHANCEMENT .

INVENTORS: PHIL HETHERINGTON

XUEMAN Ll

BACKGROUND OF THE SYSTEM

1. Related Applications.

[0001] This application claims priority to U.S. Provisional Patent Application Serial No. 60/883,507, entitled "A Spectro-Temporal-Varying Approach For Speech Enhancement" filed on January 4, 2007, and is incorporated herein in its entirety by reference.

2. Technical Field.

[0002] The system is directed to the field of sound processing. More particularly, this system provides a way to enhance speech using spectro-temporal varying technique to computer suppression gain.

BACKGROUND OF THE INVENTION

[0003] Speech enhancement often involves the removal of noise from a speech signal. It has been a challenging topic of research to enhance a speech signal by removing extraneous noise from the signal so that the speech may be recognized by a speech processor or by a listener. Various approaches have been developed in the prior art. Among these approaches the spectral subtraction methods are the most widely used in real-time applications. In the spectral subtraction method, an average noise spectrum is estimated and subtracted from the noisy signal spectrum, so that average signal-to- noise ratio (SNR) is improved. It is assumed that when the signal is distorted by a broadband, stationary, additive noise, the noise estimate is the same during the analysis and the restoration and that the phase is the same in the original and restored signal.

[0004] Subtraction-type methods have a disadvantage in that the enhanced speech is often accompanied by a musical tone artifact that is annoying to human listeners. There are a number of distortion sources in the subtraction type scheme, but the dominant distortion is a random distribution of tones at different frequencies which produces a metallic sounding noise, known as "musical noise" due to its narrow-band spectrum and the tin-like sound.

[0005] This problem becomes more serious when there are high levels of noise, such as wind, fan, road, or engine noise, in the environment. Not only does the noise sound musical, the remaining voice left unmasked by the noise often sounds "thin", "tinny", or musical too. In fact, the musical noise has limited the performance of speech enhancement algorithms to a great extent.

[0006] Various solutions have been proposed to overcome the musical noise problem. Most of them are directed toward finding an improved estimate of the SNR using constant or adaptive time-averaging factors. The time-averaging based methods are effective in removing music noise, however at a cost of degrading the speech signal and also introducing unwanted delay to the system.

[0007] Another method of removing music noise is by overestimating the noise, which causes the musical tones to also be subtracted out. Unfortunately, speech that is close in spectral magnitude to the noise is also subtracted out producing even thinner sounding speech.

[0008] A classical speech enhancement system relies on the estimation of a short- time suppression gain which is a function of the a priori Signal-to-Noise Ratio (SNR) and/or the a posteriori SNR. Many approaches have been proposed over the years on how to estimate the a priori SNR when only the noisy speech is available. Examples of such prior art approaches include Ephraim, Y.; Malah, D.; Speech Enhancement Using A Minimum-Mean Square Error Short-Time Spectral Amplitude Estimator. IEEE Trans, on Acoustics, Speech, and Signal Processing Volume 32, Issue 6, Dec 1984 Pages: 1109 - 1121 and Linhard, K, Haulick, T; Spectral Noise Subtraction With Recursive Gain

Curves , 5 ^th International Conference on Spoken Language Processing, Sydney, Australia, November 30 - December 4, 1998.

[0009] In Ephraim, Y.; Malah, D.; Speech Enhancement Using A Minimum Mean- Square Error Log-Spectral Amplitude Estimator. IEEE Trans on Acoustics, Speech, and Signal Processing, Volume 33, Issue 2, Apr 1985 Pages:443 - 445, Ephraim and Malah proposed a decision-directed approach which is widely used for speech enhancement. The a priori SNR calculated based on this approach follows the shape of a posteriori SNR. However, this approach introduces delay because it uses the previous speech estimation to compute the current a priori SNR. Since the suppression gain depends on the a priori SNR, it does not match with the current frame and therefore degrades the performance of the speech enhancement system. This approach is described below.

[0010] Classical noise reduction algorithm

[0011] In the classical additive noise model, the noisy speech is given by

[0012] y(0 = x(t) + d(t)

[0013] Where ^X V' and "' denote the speech and the noise signal, respectively.

[0014] Let Y _{n k} , X _nJi , and D _{n k} designate the short-time Fourier transform of noisy speech, speech and noise at ⁿ th frame and ^th frequency bin. The noise reduction process consists in the application of a spectral gain " ^• * to each short-time spectrum value. An estimate of the clean speech X _{n k} can be obtained as:

[0015] X _{n k} = G _{ll k} .Y _n>k

[0016] The spectral suppression gain "•* is dependent on the a posteriori SNR defined by

[0017]

[0018] and the a priori SNR is defined by

[0019]

[0020] Since speech and noise power are not available, the two SNRs have to be estimated. The a posteriori SNR is usually estimated by:

[0021]

[0022] Here, is the noise estimate.

[0023] The a priori SNR can be estimated in many different ways according to the prior art. The standard estimation without recursion has the form:

[0024]

[0025] Another approach for a priori SNR estimation is known as a "decision- directed" recursive version and is proposed in the prior art as:

[0026]

[0027] A simpler recursive version is proposed in another approach as:

[0028]

[0029] Where ) is the so-called Wiener suppression gain calculated by:

[0030]

[0031] In general, the suppression gain is a function of the two estimated SNRs.

[0032]

[0033] As noted above, because the suppression gain depends on the a priori SNR, it does not match with the current frame and therefore degrades the performance of the speech enhancement system.

BRIEF SUMMARY OF THE INVENTION

[0034] The present system proposes a technique called the spectro-temporal varying technique to compute the suppression gain. This method is motivated by the perceptual properties of human auditory system; specifically, that the human ear has better frequency resolution in the lower frequencies band and less frequency resolution in the higher frequencies, and also that the important speech information in the high frequencies are consonants which usually have random noise spectral shape. A second property of the human auditory system is that the human ear has lower temporal resolution in the lower frequencies and higher temporal resolution in the higher frequencies. Based on that, the system uses a spectro-temporal varying method which introduces the concept of frequency-smoothing by modifying the estimation of the a posteriori SNR. In addition, the system also makes the a priori SNR time-smoothing factor depend on frequency. As a result, the present method has better performance in reducing the amount of musical noise and preserves the naturalness of speech especially in very noisy conditions than do conventional methods.

[0035] Other systems, methods, features and advantages of the invention will be, or will become, apparent to one with skill in the art upon examination of the following figures and detailed description. It is intended that all such additional systems, methods, features and advantages be included within this description, be within the scope of the invention, and be protected by the following claims.

BRIEF DESCRIPTION OF THE DRAWINGS

[0036] The invention can be better understood with reference to the following drawings and description. The components in the Figures are not necessarily to scale, emphasis instead being placed upon illustrating the principles of the invention. Moreover, in the Figures, like reference numerals designate corresponding parts throughout the different views.

[0037] Figure 1 is an example of a filter bank in one embodiment of the system.

(0038] Figure 2 illustrates a smoothed spectrum after applying an asymmetric HR filter.

[0039] Figure 3 is an example of a decay curve.

[0040] Figure 4 is a flow diagram of an embodiment of the system.

[0041] Figure 5 is a flow diagram illustrating one embodiment for calculating a posteriori SNR.

[0042] Figure 6 is a flow diagram illustrating another embodiment for calculating a posteriori SNR.

DETAILED DESCRIPTION OF THE SYSTEM

[0043] The classic noise reduction methods use a uniform bandwidth filter bank and treats each band independently. This does not match with the human auditory filter bank where low frequencies tend to have narrower bandwidth (higher frequency resolution) and higher frequencies tend to have wider bandwidth (lower frequency resolution). In the present approach, we first modify the a posteriori SNR in general accordance with an auditory filter bank in two different ways by calculating the a posteriori SNR using a non-uniform filter bank and using an asymmetric HR filter. The noisy signal is divided into filter bands where the filter bands at lower frequencies are narrower to coincide with the better frequency resolution of the human ear while the filter bands at higher frequencies are wider because of less frequency resolution of the human ear. Each filter sub-band is then broken up into a plurality of frequency bins. Using broader filter bands at the higher frequencies reduces processing since there is no improvement at those frequencies by having narrower filter bands. The system focuses processing only where it can do the most good.

[0044] Figure 4 is a flow diagram illustrating the operation of an embodiment of the system. At step 401 a noisy signal is received. This signal is comprised of voice and noise data. At step 402 the a posteriori SNR is calculated. At step 403 the a pirori SNR is calculated using the previously calculated a posteriori SNR value of the same signal sample. With both a priori and a posteriori SNR values available, a suppression gain factor can be calculated at step 404.

[0045] The system proposes a number of methods of calculating a posteriori SNR. In one method, a non-uniform filter bank is used. In another embodiment, an asymmetric HR filter is used to generate a posteriori SNR. In a subsequent step, the resulting a posteriori SNR generated from either embodiment is used to generate a priori SNR. A suppression gain factor can then be calculated and used to clean up the noisy signal.

[0046] 1. Calculate the a posteriori SNR using a non-uniform filter bank

[0047] In one embodiment, the a posteriori SNR is calculated using non-uniform filter bands and is calculated for each band and each bin. Figure 5 is a flow diagram illustrating this embodiment. At step 501 the noisy signal is received. At step 502 the signal is divided into filter bands and each filter band is divided into frequency bins. At step 503 the a posteriori SNR for a filter band is calculated. At step 504 the a posteriori SNR for each frequency bin in that filter band is calculated. At decision block 505 it is determined if all filter bands have been analyzed. If so, the system exits at step 506. If not, the system returns to step 503 and calculates a posteriori SNR for the next filter band. The calculation scheme used in this embodiment are as follows:

[0048] The a posteriori SNR at each sub-band is estimated by:

[0049]

[0050] And the a posteriori SNR at each frequency bin is calculated by

[0051]

[0052] Here denotes the coefficient of ^m th filter band at ^k th bin. These filter bands have the properties that lower frequency bands cover a narrower range and higher frequency bands cover a wider range. Figure 1 is an example of a filter bank for use with an embodiment of the system. As can be seen, the lower frequency bands, such as bands 1 and m-1, are narrower than the later frequency bands such as m and m+1. This is because the human ear has better discrimination at lower frequencies and less discrimination at higher frequencies. is a normalization factor. It can be seen that the filters are non-uniform, and that their band-width may be calculated according to a MEL, Bark, or ERP scale (ref). The MEL scale is described in S.S. Stevens and J. Volkman (1940) "The Relation Of Pitch To Frequency: A Revised Scale" Am. J. Psychol.

53: 329-353. The Bark scale is described in Zwicker, E. (1961), "Subdivision Of The Audible Frequency Range Into Critical Bands." The Journal of the Acoustical Society of America, 33, Feb., 1961, and the ERP scale is described in B.C.J. Moore and B.R. Glasberg (1983) "Suggested Formulae For Calculating Auditory-Filter Bandwidths And Excitation Patterns" J. Acoust. Soc. Am. 74: 750-753.

[0053] 2. Calculate the a posteriori SNR using an asymmetric IIR filter

[0054] In an alternate embodiment we apply an asymmetric HR filter to the short- time Fourier spectrum to achieve a smoothed spectrum. Figure 6 is a flow diagram illustrating the operation of this embodiment. At step 601 the noisy signal at a frequency bin is retrieved. At step 602 this value is compared to the noisy signal value at the prior frequency bin. At decision block 603 it is determined if the current value is greater than or equal to the prior value. If so, then a first smoothing function is applied at step 604. If not, then a second smoothing function is applied at step 605. At step 606, the calculated smoothed value is used to generate the a posteriori SNR for that frequency bin.

[0055] In this embodiment, a smoothed value Y(k) is generated by applying one or the other of two smoothing functions depending on the comparison of the current bins signal value to the prior bins signal value as shown below.

[0056] when

[0057] when

[0058] Here ^β 1 ^(k) and ^β 2 ^(k) are two parameters in the range between 0 and 1 that are used to adjust the rise and fall adaptation rate. For example, when a new value is encountered that is higher than the filtered output, it is smoothed more or less than if it is lower than the filtered output. When the rise and fall adaptation rates are the same then the smoothing may be a simple IIR. When we choose different values for the rise and fall adaptation rates and also make them vary across frequency bins, the smoothed spectrum

has interesting qualities that match an auditory filter bank. For example when we set Pi and ^ ¹ to be close to 1 at bin zero and decay as the frequency bin number increases, the smoothed spectrum follows closely to the original spectrum at low frequencies and begins to rise and follow the peak envelop at high frequencies.

[0059] The same filter can be run through the noise spectrum in forward or reverse direction to achieve better result. Figure 2 shows a simulation result of applying this filter on a modulated Cosine signal. The smoothing curves are comparing two asymmetric HR filters applied to a cosine series. One is a constant rate and the other is a variable rate. The cosine series represents the log spectral energy across a harmonic series. The rise factors in each case are 1, with no decay across frequencies. The fall factor in the constant rate HR is 0.3 and did not decay. The fall factor in the variable rate IIR is 0.7 and it decayed at a rate of 0.4 across frequencies.

[0060] This smoothed spectrum is then used to calculate the a posteriori SNR

[0061]

[0062] 3. Calculate the a priori SNR using the computed a posteriori SNR

[0063] The a posteriori SNR generated using either embodiment above can then be used to calculate the a priori SNR using equation (1), (2), and (3) with some modifications as noted below:

[0064] We modify the " decision-directed" method in equation (2) as follows:

[0066] Instead of using a constant averaging factor for all frequency bins, we introduce a frequency- varying averaging factor ^a ' 'which decays as frequency increases. Figure 3 shows an example of such a decay curve. This matches up with the

need for greater temporal fidelity in the higher frequencies and less temporal fidelity in the lower frequencies. Other suitable curves may be used without departing from the scope and spirit of the system. Finally, ^α W may be asymmetric to differentially smooth onsets and decays, which is also a characteristic of the human auditory system (e.g., pre- masking, post-masking). For example ^a ^ ' may be 1 for all rises and 0.5 for all falls, and both may decay independently across frequencies.

(0067] Similarly, we modify the recursive version in equation (3) to as:

[0068]

(0069] Here ^δ(k) is a frequency varying floor which increases from a minimum value (e.g., 0) to a maximum value (e.g., 1) over frequencies.

[0070] 4. Generate Suppression Gain Factor and Apply Noise ReductionAfter the a priori SNR is generated, a suppression gain factor can be generated as noted in equation (4) above. The suppression gain factor can then be applied to the signal ^as: Noise reduction methods based on the above a priori SNR are successful in reducing musical noise and preserving the naturalness of speech quality.The illustrations have been discussed with reference to functional blocks identified as modules and components that are not intended to represent discrete structures and may be combined or further sub-divided. In addition, while various embodiments of the invention have been described, it will be apparent to those of ordinary skill in the art that other embodiments and implementations are possible that are within the scope of this invention. Accordingly, the invention is not restricted except in light of the attached claims and their equivalents.