Processing of Two Noise Contaminated, Substantially Identical Signals to Improve Signal-to-Noise Ratio

1. Field of the Invention

This invention relates to processing a pair of substantially identical signals contaminated by uncorrelated noise to produce a like signal with an improved signal-to-noise ratio.

2. Description of the Prior Art

When two substantially identical signals contaminated by uncorrelated noise are received over two channels, a like signal with an improved signal-to-noise ratio (SNR) can be realized by linearly combining the contaminated signals. This may be appreciated by considering the following "left" and "right" contaminated signals

where s is the signal amplitude (with power S), n.sub.l is a "left" noise amplitude (with power L), and n.sub.r is a "right" noise amplitude (with power R). The SNRs of these two contaminated signals are

The average of the two contaminated signals (l + r)/2 has a signal power S and, if the two noises are uncorrelated, a noise power of (R + L)/4. The SNR of the average channel output is, therefore,

if the two uncorrelated noises have equal power (L = R), then

a situation usually referred to as a "3-dB improvement in SNR." This 3-dB gain in SNR is in fact the maximum achievable through linear processing.

By contrast, a human listener seems to be able to do much better when processing his two-channel ear inputs. If human performance is measured by speech intelligibility, for example, binaural performance may exceed monaural performance (of the "better" ear if they are different) by as much as 12 dB or more. This has often been referred to colloquially as the "cocktail-party" effect which derives its name from the ability to pick out a single talker from a loud babble of spatially dispersed voices -- a situation frequently encountered at cocktail parties. The fact that a human can understand speech under such circumstances, implies binaural signal-processing capabilities which go far beyond the 3-dB improvement obtainable from linear processing. Technically this phenomenon is often referred to as "binaural release from masking" and is measured by the "binaural masking level difference" (BMLD). What nonlinear processes a human uses to accomplish this feat, however, is not known in any detail, although "contra-lateral" neural inhibition and excitation processes in the auditory pathways between the ears and the human auditory cortex are probably involved.

An oject of the present invention is to process two substantially identical signals, which are each contaminated by uncorrelated noise, to produce an output having a signal-to-noise ratio improvement greater than that achievable through linear processing.

This and other objects of the invention are achieved by summing weighted values of the two contaminated signals. In accordance with the invention, the weighting factors comprise combinations of moving averages of magnitudes related to the contaminated signals. In particular, the instantaneous amplitude r of a first contaminated signal is multiplied by a weighting factor equal to ##EQU1## where l is the second contaminated signal instantaneous amplitude and the bars indicate moving averages of the values under them. Similarly, the instantaneous amplitude of the second contaminated signal is multiplied by a weighting factor equal to ##EQU2## These two weighting factors are functions of r and l and consequently are continuously changing in value. This process, which is not a linear one, results in a signal-to-noise ratio greater than that achievable with linear processing.

One possible use for an embodiment of the invention is in a speakerphone. Echo and other noises in a room are generally picked up by a speakerphone and are very objectionable to the listener at the other end of the line. By using a pair of microphones with the user of the speakerphone substantially equally spaced between them and furthermore, an embodiment of the invention to process the microphone outputs, a substantial signal-to-noise improvement can be realized before transmitting to the listener. When the user is not equally spaced between the microphones, delay and amplifying devices may be employed in a conventional manner to produce substantially identical input signals to the embodiment of the invention. These devices may be constructed in accordance with known techniques to produce a desired delay or amplifying versus frequency characteristic.

Embodiments of the invention are not restricted to use in the audio frequency band but may be used at frequencies below and above this band. Furthermore, such embodiments may be of either an analog or a digital form so as to be operative on either analog or digitally encoded signals.

In the drawings:

FIGS. 1 and 2 are block diagrams representing embodiments of the invention,

FIGS. 3 and 4 are block diagrams of structure which may be used in the embodiment depicted by FIG. 2, and

FIGS. 5, 6 and 7 show various ways in which embodiments of the invention may be combined.

The symbols and bars used in the following discussion have been defined in the previous discussion. In brief, s = signal amplitude, S = signal power, n.sub.l and n.sub.r = left and right channel noise amplitudes, respectively, L and R = left and right channel noise powers, respectively, l = s + n.sub.l, r = s + n.sub.r and the bars indicate moving averages of the values under them.

If s, n.sub.l, and n.sub.r are uncorrelated, have zero-mean Gaussian values defined by

and are of known powers

then the signal amplitude s and the sum .sigma. = r+l and difference .delta. = r-l have the following covariance matrix: ##EQU3## When its inverse is called .mu..sub.ik, the joint distribution of (s, .sigma., .delta.) is given by the probability density function

The corresponding conditional probability distribution p(s .alpha.,.delta.) for s, given .sigma. and .delta., differs from expression (2) only by a factor which does not depend on s. With this in mind the expression ##EQU4## may be written. By setting expression (3) equal to zero, the most likely value of s is found to be

With the .mu..sub.ik computed from expression (1), expression (4) becomes ##EQU5##

When the two noise powers are equal to one another so that R = L, expression (5) becomes ##EQU6##

Three things should be noted, namely:

a. when the noise power L .noteq. 0, .sigma./2 is multiplied by less than one,

b. when the noise power L = 0, .sigma./2 is multiplied by one, and

c. when the signal power S .fwdarw. 0, the resulting signal s.sub.o .fwdarw. 0.

When R .noteq. L, the difference signal .delta. enters into s.sub.o. The effect of this is best seen by reintroducing the left and right channel components for .sigma. and .delta. in expression (5) to produce ##EQU7## Three things should again be noted:

a. when noise power R = 0, then s.sub.o = r, which equals s,

b. when noise power L = 0, then s.sub.o = l, which equals s, and

c. when noise power S.fwdarw.0, then s.sub.o .fwdarw.0.

This may be better appreciated by rewriting expression (6) as follows: ##EQU8##

In accordance with the present invention, values for S, L and R are estimated from the noise contaminated signals as follows:

S + r = the infinite-time average of r.sup.2,

S + l = the infinite-time average of l.sup.2, and

S = the infinite-time average of rl.

Obtaining the infinite-time averages of these values is, of course, not physically possible. It is, however, physically possible to obtain short-time moving averages of the values.

In accordance with the invention, short-time moving averages are used. Notwithstanding the fact that some degradation in the final signal-to-noise ratio was anticipated because of this "compromise", investigations with the short-time averages resulted in ratios greater than those expected for infinite-time averages. This is discussed later.

When the short-term moving average values are applied to expression (6), that expression becomes ##EQU9## Expression (7) may also be written as ##EQU10## Expression (7a) is in a form which immediately demonstrates that the most likely value s.sub.o is the sum of weighted values of the two contaminated signals. Furthermore, it immediately demonstrates that the weighting factors are nonfixed quantities made up of the contaminated signals.

FIG. 1 shows a block diagram representative of an embodiment of the invention constructed in accordance with expressions (7) and (7a). A pair of input ports 10 and 11 are provided for receiving a pair of left and right channel signals, respectively, which signals are contaminated by uncorrelated noise. The contaminated left and right channel signals have instantaneous amplitudes l and r, respectively. The left channel contaminated signal is applied to both inputs of a multiplier 12 which produces an output whose amplitude is l.sup.2. Similarly, the right channel contaminated signal is applied to both inputs of a multiplier 13 to produce an output whose amplitude is r.sup.2. The contaminated left and right channel signals are also applied to inputs of a multiplier 14 which produces an output having an amplitude rl. These three multiplier outputs are passed through low pass filters 15, 16 and 17, respectively, to produce moving average outputs having amplitudes l.sup.2, r.sup.2 and rl, respectively.

The next step is the subtraction of the output of filter 17 from the outputs of filters 15 and 16 in substractors 18 and 19 to produce a pair of outputs having amplitudes equal to (l.sup.2 - rl) and (r.sup.2 - rl), respectively. The outputs of subtractors 18 and 19 are then multiplied, in multipliers 20 and 21, by the amplitudes of delayed contaminated right and left channel signals, respectively. Delay devices 22 and 23 provide delays to compensate for delays introduced by filters 15, 16 and 17. The outputs of multipliers 20 and 21 are then summed in summer 24 to produce (l.sup.2 - rl)l + (r.sup.2 - rl) r, which is then multiplied, in a multiplier 25, by rl. This results in the numerator portion of expression (7).

The denominator portion of expression (7) is achieved by first multiplying l.sup.2 by r.sup.2 in a multiplier 26 and squaring rl in a multiplier 27. The output of multiplier 27 is then subtracted in a subtractor 28 from the output of multiplier 26 to produce the denominator portion.

Finally, the output of subtractor 28 is divided into the output of multiplier 25 by divider 29 to produce the output s.sub.o.

Other embodiments of the invention are also possible. Consider, for example, the order of the processing as set forth in expression (7a). This may be implemented as shown in FIG. 2. In FIG. 2, the values l.sup.2, r.sup.2 and rl are all derived as in FIG. 1. The outputs of delay device 23 and low pass filters 15, 16 and 17 are all applied to block 30 while the outputs of delay device 22 and low pass filters 15, 16 and 17 are all applied to block 31. The outputs of blocks 30 and 31 are summed in a summer 32 whose output comprises s.sub.o. Blocks 30 and 31 perform the two fractional processes, respectively, set forth in expression (7a).

Block 30 and 31 may take any one of a number of forms, including those disclosed in FIGS. 3 and 4. The structure of FIG. 3 may be made up, for example, of elements 18, 20, 25, 27, 29, 28 and 26 of FIG. 1. The arrangement of FIG. 4 differs from that of FIG. 3 in that the organization of the three uppermost elements has been changed as if the rl and parenthetical portions of expression (7a) were multiplied together.

The embodiment of FIG. 1 may be preferred to that of FIG. 2 because it requires less elements than used in the embodiment of FIG. 2. This occurs because FIG. 1 combines blocks 30 and 31 of FIG. 2. On the other hand, the embodiment of FIG. 2 may be preferable in some instances for other reasons, such as manufacturing.

The various circuits represented by the blocks in the FIGS. may be either of an analog or digital nature. When the contaminated signals are of an analog nature and it is desired to use digital circuits, the contaminated signals are first converted to a digital form by passing them through conventional analog-to-digital converters prior to application to ports 10 and 11. In this case, the output s.sub.o is in a digital form but may be converted back into an analog form through the use of a conventional digital-to-analog converter. Such techniques are appreciated by those skilled in the art.

The moving average outputs produced by filters 15, 16 and 17 are relatively short term values. The following discussion relating to the processing of speech signals serves as an example.

Speech is a nonstationary process. In particular, it is a process whose power fluctuates rapidly at syllabic rates (ca. 10 Hz). Thus, by using an averaging time of about 20 msec, a better estimate of the syllabic signal power is obtained than by either much longer or shorter averaging times. The same is true for the estimates of the noise powers, particularly, if the noise is actually interfering speech or a babble of voices.

In order to get a rough idea of the effect of nonstationarity, one may assume that a speech signal of long-time average power S has actually zero amplitude two-thirds of the time and power 3S the remaining one-third of the time. The noise power in the averaged microphone signals .sigma.2, when the speech signal is on, equals (R+L)/4. When the signal is off, the noise is also essentially switched off. Thus, the average noise power is

the corresponding SNR is

compared to 4S/(R+L) for the stationary case. Thus, there is an additional improvement in SNR of about 5dB due to nonstationarity. The total SNR improvement is 8 dB compared to the SNR of a single microphone and R = L.

The above averaging time of 20 msec was used in digital simulations in a computer. The advantage in simulating is that the noise and speech levels may be accurately controlled for testing purposes. In the performed simulations, equal noise levels and equal speech levels were applied to both channels. Because of the nonstationary character of speech signals, the SNR improvement obtained was actually larger than that predicted for stationary signals (3dB). Listening to these particular simulations suggests that the improvement is even greater than 8 dB, i.e., there is a subjective as well as an objective improvement. This is probably because there is little noise when it would be most audible -- namely during low-level or silent speech intervals during which the factor rl approaches zero. In fact, it is well known that for equal average noise powers, a noise whose power fluctuates in synchrony with the syllabic speech power -- as it does at the output of the processor described here -- is less subjectively objectionable ("sounds less noisy") than a stationary noise.

In some applications of the present invention, it may be advantageous to divide the channel signals into pluralities of signals residing in contiguous frequency bands, to use a like plurality of embodiments of the invention to process like frequency-band signals and to sum the resulting outputs for the embodiments. This is shown in FIG. 5 wherein channel signals on inputs 10' and 11' are divided into frequency bands A and B by bandpass filters 33, 34, 35 and 36. Those portions of the channel signals in frequency band A are applied to a processor 37 while those portions in frequency band B are applied to a processor 38. These processors are each constructed in accordance with the present invention. The outputs of processors 37 and 38 are combined in a combiner 39. This approach may be found, for example, to be a desirable way to introduce controlled delay and amplification to produce like signals prior to application to processors 37 and 38; i.e., once divided into such frequency bands, delay and/or amplification may be applied over limited frequency ranges to cause the signal contents in corresponding bands to be substantially equal prior to processing. This approach may also be found to be desirable in that the individual embodiments are not required to be operable over the frequency range of the channel signals.

Embodiments of the invention may also be used with more than two channel signals to achieve still greater signal-to-noise ratios. FIGS. 6 and 7 show combinations of processors being used to combine like signals over three and four channels, respectively. In FIG. 6, two channel inputs are processed in a first processor while its output and the third channel signal are processed in a second processor. In FIG. 7, two channel signals are processed in a first processor, the remaining two channel signals are processed in a second processor and the outputs of the first and second processors are processed in a third processor. Larger numbers of channel signals may similarly be accommodated. In these combinations, delay and amplification control of the inputs to the various processors may have to be introduced to produce like signal inputs to any given processor as previously discussed in detail.