HK1031421B

HK1031421B - Methods and apparatus for measuring signal level and delay at multiple sensors

Info

Publication number: HK1031421B
Application number: HK01102280.4A
Authority: HK
Inventors: K‧P‧海恩德尔; J‧A‧J‧拉斯姆松
Original assignee: 艾利森电话股份有限公司
Priority date: 1997-07-11
Filing date: 1998-07-03
Publication date: 2004-07-09

Description

Method and apparatus for measuring signal levels and delays at multiple sensors

Technical Field

The present invention relates to signal processing, and more particularly, to measuring signal levels and delays at multiple sensors.

Background

In many signal processing applications, it is often desirable to determine the relative sensitivity of multiple signal sensors with respect to a particular signal source. For example, in a hands-free mobile phone environment, dual microphones may be used in conjunction with a beamforming method to reduce the effects of background noise and echo in a car. To achieve this, information about the relative sensitivities of the microphones related to different acoustic sources is used, for example, to form a spatial beam towards a particular user and/or to form a spatial notch that blocks another user or speaker. This approach requires that dynamic information about the microphone sensitivity be obtained quickly and accurately.

Fig. 1 depicts a prior art system 100 for measuring the relative sensitivity of two microphones to different signal sources in a hands-free mobile telephone environment. As shown, the prior art system 100 includes a first microphone 115, a second microphone 125, an adaptive filter 135, and a summing device 140. Output y of the first microphone 115₁(k) Connected to the positive input of the summing device 140, the output y of the second microphone 125₂(k) Is connected to the input of the adaptive filter 135. Output of adaptive filter 135Is connected to the negative input of the summing device 140 and the output e (k) of the summing device 140 is used as a feedback signal to the adaptive filter 135.

As shown, the first microphone 115 is positioned closer to the first signal source 110 and the second microphone 125 is positioned closer to the second signal source 120. For example, the first microphone 115 may be a hands-free microphone attached to a visor in close proximity to the driver of the automobile, and the second microphone 125 may be an in-line microphone provided in a mobile unit in the vicinity of the passenger in the automobile. Although not shown in fig. 1, those skilled in the art will appreciate that analog pre-processing circuitry and analog-to-digital conversion circuitry may be included at the output of each of the first and second microphones 115, 125 such that the digital signals are processed by the adaptive filter 135 and the summing device 140. The output e (k) of the summing device 140 represents the output y of the first microphone 115₁(k) And the output of adaptive filter 135The difference between them and is referred to herein as the error signal.

In operation, the filter coefficients of the adaptive filter 135 are adjusted using a least squares algorithm such that the error signal e (k) is minimized. In other words, the adaptive filter 135 is adjusted such that the output of the adaptive filter 135As close as possible to the output (i.e., an estimate thereof) y of the first microphone 115₁(k) In that respect In this way, the adaptive filter 135 attempts to simulate the effects of the signals resulting from physically separating the microphones 115, 125. For example, when the passenger 120 speaks, his or her voice arrives at the second microphone 125 slightly earlier than the first microphone 115, and the corresponding speech signal level received at the first microphone 115 is attenuated somewhat compared to the level received at the second microphone 125. In this way, the adaptive filter 135 is adjusted to give the same delay and attenuation effect.

As a result, the relative time delay and Signal attenuation at the microphone associated with each user may be calculated based on the coefficients of the adaptive filter 135, such as described in y.t. chan, j.m. riley, and j.b. plant, "a parameter estimation process time delay and Signal detection," ieee transmission on Acoustics Speech and Signal Processing, volASSP-28, Feb, 1980, which is incorporated herein by reference. One disadvantage of the system of fig. 1, however, is that its performance deteriorates significantly when noise is present. As a result, the system of fig. 1 is not useful in most practical applications where significant background noise (e.g., road noise and traffic noise) is common. Thus, it is desirable to measure the relative signal levels and time delays at multiple sensors.

Disclosure of Invention

The present invention fulfills the above-described and other needs by providing a system in which a fixed filter and an adaptive filter are applied in combination to provide accurate and robust estimates of signal levels and time delays for multiple sensors. In an example embodiment, the fixed filter includes at least one relatively narrow pass band that is used to distinguish the signal source of interest from the broadband background noise. In an embodiment, the fixed filter is connected to the reference sensor and the adaptive filter is connected to the second sensor. The error signal derived from the outputs of the fixed filter and the adaptive filter is used to adjust the coefficients of the adaptive filter according to a suitable least squares algorithm. The coefficients of the fixed filter and the adaptive filter are used to calculate an estimate of the time delay and relative level between the two sensors. This estimate can then be used to make decisions regarding sensor selection and beamforming.

In an exemplary embodiment, the functionality of the system is implemented by an activity detector that indicates when there is no signal of interest. In the activity detector, the accumulated energy in the adaptive filter is compared with an expected minimum derived from the fixed filter coefficients. When the accumulated energy is less than the expected value, indicating the absence of any signal of interest (i.e., the presence of only background noise), then the time delay and relative level estimates are set to appropriate values to ensure that the system will operate properly even during periods when no signal of interest is present.

In another embodiment, more than two signal sensors are employed. In this embodiment, one sensor is considered a reference sensor and is connected to a fixed filter, while the other sensors are connected to an adaptive filter. For each additional sensor, the error signal derived from the outputs of the fixed filter and the corresponding adaptive filter is used to update the coefficients of the corresponding adaptive filter. In this way, robust estimates of the time delay and relative signal level between the reference sensor and each additional sensor can be calculated and sophisticated decisions can be made regarding sensor selection and beamforming.

In general, the present invention presents a computationally simple but accurate and powerful method for estimating time delays and relative signal levels at multiple sensors. The teachings of the present invention are applicable in a wide variety of signal processing environments. For example, the invention may be used in other acoustic applications, such as teleconferencing, in addition to the hands-free mobile phone application described above. Furthermore, the invention is also applicable in radio communication applications, where the signal of interest is a radio frequency transmission (e.g. from a mobile unit and/or a base station in a cellular radio system) and the sensor is an antenna element sensitive to radio frequencies. All features and advantages of the invention may be explained with reference to the examples given in the drawings.

Drawings

Fig. 1 depicts a prior art signal level and delay measurement system as described above.

Fig. 2 depicts a signal level and delay measurement system constructed in accordance with the present invention.

Fig. 3 depicts the relative signal levels and time delays of two signals detected at two signal sensors.

Fig. 4 depicts another signal level and delay measurement system constructed in accordance with the present invention.

Fig. 5 depicts the magnitude and phase response of an example signal filter that may be applied in the example systems of fig. 2 and 4.

FIG. 6 depicts example speech and noise signals used to demonstrate the operation of example embodiments of the present invention.

Fig. 7 depicts signal level and delay estimates generated by an exemplary embodiment of the present invention based on the signal of fig. 6.

Detailed Description

Fig. 2 depicts a level and delay measurement system 200 constructed in accordance with the teachings of the present invention. As shown, system 200 includes a first sensor 215, a second sensor 225, a fixed FIR filter 230, an adaptive FIR filter 235, and a summing device 240. Output y of the first sensor 215₁(k) Is connected to the input of the fixed filter 230, the output y of the fixed filter 230_F(k) Is connected to the positive input of the summing device 240. Output y of second sensor 225₂(k) Is connected to the input of the adaptive filter 235, the output of the adaptive filter 235Is connected to the negative input of the summing device 240. The output of the summing device 240, the error signal e (k), is fed back to the adaptive filter 235.

As shown, the first sensor 215 is positioned closer to the first signal source 210, while the second sensor 225 is positioned closer to the second signal source 220. For example, the first sensor 215 may be a hands-free microphone located on a visor in close proximity to the driver of the automobile, and the second sensor 225 may be an embedded microphone located in a mobile unit in the automobile near the passenger. Alternatively, the first and second sensors 215 and 225 may be antenna elements positioned closer to the first and second radio frequency signal sources, respectively. Although not shown in fig. 2, those skilled in the art will appreciate that analog pre-processing and analog-to-digital conversion circuitry may be included at the output of each of the first and second sensors 215, 225 so that the digital signal is processed by the fixed filter 230, the adaptive filter 235, and the summing device 240.

The fixed filter 230 is designed to include at least one relatively narrow passband of interest. For example, in a mobile phone environment, the passband may correspond to the 300-600Hz frequency band where most of the energy of human speech is concentrated. In radio communication applications, the passband may correspond to a band allocated for radio frequency transmission. In any case, the coefficients of the fixed filter 230 are adjusted as needed to compensate for changes in application requirements or environmental conditions. For example, in a hands-free mobile phone application, the fixed filter 230 may be set to optimize the received signal-to-noise ratio for a particular automotive device. Furthermore, the coefficients of the filter 230 may be dynamically adjusted, for example, based on the measured signal-to-noise ratio.

In accordance with the present invention, the fixed filter 230 is designed to provide unity gain and zero phase within each pass band. In addition, the noise gain of the fixed filter 230 is minimized to ensure maximum cut-off band attenuation. As described in detail below, the prior information provided by the fixed filter (i.e., the narrowband nature of the output signal of the fixed filter 230) is used to make the system more robust against noise.

In operation, the coefficients of the adaptive filter 235 are adjusted using a suitable least squares algorithm such that the error signal e (k) is minimized, and the output of the adaptive filter 235As close as possible to the output y of the fixed filter 230_F(k) In that respect As described below, the relative time delays and signal attenuations at the first and second sensors 215, 225 with respect to each signal source 210, 220 are calculated based on the coefficients of the adaptive filter 235 and previous information related to the fixed filter 230. Although not explicitly shown in fig. 2, those skilled in the art will appreciate that a suitable digital signal processor may be integrated with system 200 to perform the least squares update of adaptive filter 235 and calculate the time delay and signal level estimate.

To clarify the operation of the system 200 of fig. 2, a rigorous mathematical analysis will be performed with reference to fig. 3 and 4. Although the analysis is clearly performed for two sensors and two signal sources, one skilled in the art will appreciate that the methods described are still useful for applications involving any number of signal sources and sensors. Furthermore, although reference is sometimes made to the acoustic hands-free mobile phone application described above, it will be appreciated by those skilled in the art that the above described method is also applicable to many other signal processing environments including the aforementioned radio communication.

Fig. 3 depicts a typical example of signal sources and sensors placed in a two-dimensional space. In the figure, the first and second sensors 215, 225 are positioned adjacent to the two signal sources 210, 220. As shown, a signal (represented by a first dashed arc 315) from the first signal source 210 will be input to the first signal sensor 215 before being input to the second signal sensor 225. Thus, the signal received at the second sensor 225 by the first signal source 210 will be a delayed and attenuated version of the signal received at the first sensor 215 by the same signal source 210. In addition, a signal (represented by a second dashed arc 325) from the second signal source 220 will enter the second sensor 225 before entering the first sensor 215. The signal received at the first sensor 215 by the second signal source 220 will be a delayed and attenuated version of the signal received at the second sensor 225 by the same signal source 220. The spatial separation (and thus the corresponding time delay and level decay) of the sensors 215, 225 relative to the first and second signal sources 210, 220 is represented in fig. 3 by second and first straight line segments 320, 310, respectively.

If the first and second sensor inputs at time k (after analog preprocessing and analog-to-digital conversion), respectively, are represented by x₁(k) And x₂(k) Indicates that then the second sensor inputs x₂(k) Is the first sensor input x₁(k) Delayed and scaled versions of (a). In other words, x₂(k)＝1/S·x₁(k-D) wherein the scaling factor S is greater than 0, the delay D may be either positive or negative. Strictly speaking, for D < 0 (e.g., for a signal emanating from second signal source 220), first input x₁(k) Is the second input x₂(k) Delayed and scaled versions of (a). However, to simplify the representation, the second input x is input without loss of generality₂(k) Is represented as a delayed signal for all D values.

To address the problem of causal filtering, a fixed delay Δ may be introduced in the signal path after the first sensor. Although in most applications this is a natural approach, it is not a prerequisite for the feasibility of the invention. This will be described in detail below.

With the introduction of the additional delay Δ, the first and second intermediate signals y may be defined as follows₁(k)，y₂(k)：

y₁(k)＝x₁(k-Δ) (1)

Where q denotes the well-known delay operator (i.e., qy (k) ═ y (k +1), q^-1y (k) ═ y (k-1), etc.), D is defined as Δ -D. Note that for causal filtering, Δ > D.

To aid discussion, FIG. 4 illustrates an input signal x in a level and delay measurement system₁(k)，x₂(k) And intermediate signal y₁(k)，y₂(k) In that respect The system 400 of fig. 4 is the same as the system 200 of fig. 2, except that a delay block 410 (corresponding to the fixed delay delta described above) is positioned between the first sensor 215 and the fixed sensor 230. In the following discussion, it is assumed that the coefficients of the fixed filter 230 are stored in the first coefficient vector c₀In, the time-varying coefficients of the adaptive filter 235 are storedStored in the second coefficient vectorIn (1).

In general, the invention provides a sensor input x based on measurements₁(k) And x₂(k) A computationally simple but accurate method for estimating the delay D and the scaling factor S. Advantageously, the method is very resistant to background noise so that the method can be successfully used in a hands-free mobile phone environment such as that described above. Estimated magnitude, i.e. D_kAnd(where D and S are calculated using sensor inputs at and before time k) may be used to improve system performance. For example, in a mobile phone environment, estimate D_kAndmay be used in conjunction with well-known beam forming techniques to electronically enhance or reduce the sensitivity of the sensors 215, 225 relative to the first and second signal sources 210, 220. For example, when a particular signal source is activated, (e.g., when the driver speaks), a beam is formed in the direction of that source to optimize its reception. In addition, spatial filtering can be employed to eliminate the sensitivity of a sensor related to a signal source when the source is providing a signal that should be blocked at the sensor (e.g., when the source is a speaker that may cause objectionable feedback or echo).

Further, the system may selectively send only signals detected at a particular sensor when a particular signal source is activated. For example, if a sensor is more sensitive to a passenger than to a driver (e.g., because it is physically closer to the passenger), it may be desirable to only transmit the signal received at that sensor when only the passenger is speaking.

Returning to FIG. 4, the signal y output by the fixed filter 230_F(k) (i.e., the first intermediate signal y)₁(k) Filtered version of (d) is given by:

y_F(k)＝y₁(k)^Tc₀ (3)

y₁(k)＝(y₁(k)…y₁(k-L))^T (4)

c₀＝(c₀…c_L)^T(5) where L is the order of the fixed filter 230, { c_l0, … L is the coefficient of the fixed filter. In addition, the signal output by the adaptive filter 235(i.e., the second intermediate signal y)₂(k) Filtered version of (d) is given by:

y₂(k)＝(y₂(k)…y₂(k-L))^T (7)

wherein the vectorIncluding the time-varying filter coefficients of adaptive filter 235. VectorUpdated based on the error signal e (k), as follows:

where μ is a gain factor (constant or time-varying) in the range 0 μ < 2,representing the squared euclidean vector norm. The adaptive algorithm described by equations (9) and (10) is the well-known normalized least mean squares algorithm (N-LMS). Alternative adaptive schemes, e.g. Recursive Least Squares (RLS) or Least Mean Squares (LMS)The method (LMS) may also be used. For a more detailed description of Adaptive algorithms see, for example, b.widrow and s.d.steads, Adaptive Signal Processing, preface Hall, Englewood Cliffs, NJ, 1985, and l.ljung and t.soderstrom, Theory and Practice of curative Identification, m.i.t.press, Cambridge, MA 1983, which are incorporated herein by reference. Advantageously, each of the above defined quantities may be calculated using standard digital signal processing components.

For a broadband signal source encountering a sensor 215, 225, the coefficients of the adaptive filter 235 converge towards a delayed and scaled version of the coefficients of the fixed filter 230. In particular if the vector c is fixed₀Is 1 (i.e., if no fixed filter is actually used), then the time-varying vectorConverge towards a scaled delay approximation (i.e., sq)^D-Δ＝sq^-D). This result has been used in prior art systems to estimate the time delay. See, for example, Y.T. Chan, J.M.F.Riley and J.B.plant, "Modeling of time delay and applications to estimation of probability delays", IEEE Transactions on Acoustics, Speech, and Signal Processing, Vol, ASSP-29, No.3, pp.577-581, June 1981, which are incorporated herein by reference. One disadvantage associated with such systems is that the overall performance of the system deteriorates significantly when background noise is present, making such systems impractical for most real-world applications.

Advantageously, the present invention indicates that by introducing a priori knowledge to distinguish background noise from the source signal, system performance can be significantly improved. To ensure improved overall performance, previous techniques should be correct in all circumstances. For example, the present invention teaches that such a priori information is available when the energy of the source signal of interest is concentrated near one or more center frequencies, while the background noise has relatively flat and broad frequency or power spectral densities. In this case, the present invention indicates that fixed FIR filter 230 may be designed as a bandpass filter with one or several passbands.

For example, for speech signals in a mobile hands-free environment, it is reasonable to assume that the energy of the speech signal is concentrated between 100-250 Hz. More specifically, the fundamental frequency of a male speaker is typically around 100Hz, while the fundamental frequency of a female speaker is around 250 Hz. From the information point of view, the invention gives several alternative designs possible for the fixed filter 230. For example, the fixed filter 230 may be designed to include two passbands, the first and second passbands having center frequencies of 100Hz and 250Hz, respectively. In addition, the fixed filter 230 may be designed to include a single pass band having a pass band centered at 200Hz and spanning a frequency band including the fundamental frequency of female voice and the first harmonic frequency of male voice.

In fact, the former method requires the use of a higher order filter than the latter method. Generally, if the number of design frequencies is doubled, the order L of the filter is also doubled. In the following discussion, fixed filters 230 having m different passbands are considered. At the center frequency { omega_l1, …, the m-up filter is designed to give unity gain and zero phase. Further, the fixed filter 230 is designed to provide maximum attenuation in the cutoff band by minimizing the filter Noise Gain (NG) defined as follows:

wherein C is₀(z^-1)＝c₀+c₁z^-1+…+c_Lz^-LAnd integration is performed around the unit circle. By the Parseval relationship, the noise gain of the FIR filter is given by:

to design the fixed filter 230, consider a signal y comprising the sum of several sinusoidal terms_in(k) The following are:

where ω is_l1, …, m being the center frequency of the desired fixed FIR filter, ω_l∈(0，π)，ω_l≠ω_j，l≠j，{α_lIs an unknown constant, α_l> 0, l ═ 1, …, m, and { φ -_lIs a uniformly distributed random variable phi_l∈(-π，π]1, … m. Consider now having a coefficient vector c_dAnd gives an output y_out(k) The fixed FIR filter 230, wherein the output y_out(k) Is an input of y_in(k) For any value of d, - ∞ < d < ∞, the value after d-step prediction (after any initial transients have decayed) is performed as follows:

y_out(k+d)＝y_in(k)^Tc_d (14)

y_in(k)＝(y_in(k)…y_in(k-L))^T (15)

c_d＝(c₀…c_L)^T (16)

for d equal to 0 (i.e. for coefficient vector c)₀) The fixed FIR filter 230 is at the center frequency { omega_l1, … m gives unity gain and zero phase. Furthermore, if the sensitivity to wideband noise is minimized (i.e., if the value in equation (12) is minimized) then (for filter length L, such that L > 2m-1) the following result is true:

c_d＝L^T(LL^T)^-1p(d) (17)

where L is a 2m × (L +1) matrix:where p (d) is a 2m prediction vector

The first and second curves 510, 520 of fig. 5 respectively represent the curves designed using the method described above, and d is 0, L is 32, m is 1, ω is₁200Hz, with a sampling rate of 8000Hz, the amplitude response and phase response of fixed filter 230. As indicated by the dashed line in fig. 5, the fixed filter 230 gives unity gain and zero phase at a center frequency of 200Hz, as desired.

As described above, the adaptation algorithm used to update the adaptive filter 235 will cause the adaptive filter to converge towards the delayed and scaled replica of the fixed filter 230. In particular, for a fixed FIR filter 230 with d-0 (i.e., coefficient c)₀) The coefficients of adaptive filter 235 will converge as follows:

where S and D are the scaling factor and time delay, respectively, due to the separation of the sensors 215, 225 in physical location. Thus, the present invention indicates that the estimated values of the scaling factor S and the time delay D may be based onThe vector relationship given by equation (20) is calculated. For example, ifRepresenting coefficients based on adaptive filter 235For the estimated value of S, that estimated valueCan be calculated from equation (20) as follows:furthermore, S can be estimated without prior information about D. To see this, it is first noted that equation (21) is a vector magnitude, where c is_-DAndare all vectors of size L + 1. Thus, according to equation (17), there is Lc_-DP (-D). Thus, the matrix L of 2m × (L +1) defined in equation (18) is multiplied on both sides of equation (21) to give the following result:

both sides of equation (22) are 2m vectors. Thus, assume p (D)^Tp (D) ═ m (see equation (19)), equation (22) can be rewritten as follows:

according to equation (23), and according to the fact that S > 0, an estimate of the scaling factor S at time kCan be calculated as follows:

estimate of a given scaling factor SEstimated value D of time delay D_kCan be calculated using a least squares algorithm as follows:

equivalently, the estimated value D of the time delay D_kCan be calculated as follows:

advantageously, in practice, the estimated value D is_kMay be calculated iteratively. Note that the delay gradient dp (D)/dD follows equation (19).

Thus, the invention indicates that the estimated values of the scaling factor S and the time delay D can be calculated in a simple manner. Advantageously, each of the above described calculations may be implemented using well known digital signal processing components. These estimates are valid even in the presence of background noise due to consistent prior information provided by the fixed filter 230.

The system can be further enhanced by the addition of an activity detector that ensures correct system performance even when all signal sources are not activated. For example, when neither source 210, 220 is active, signal x is received at sensor 215, 225₁(k) And x₂(k) Only uncorrelated noise will be included. In this case, the adaptive filter coefficientsWill converge towards the zero vector, which means that the scale factor estimate isWill tend towards zero and the time delay estimate D_kAny value may be taken. To demonstrate this, the estimate is made when the activity detector detects that no signal of interest is present，D_kMay be explicitly set to an appropriate value.

The example activity detector compares the estimated value of the filter noise gain to a predetermined threshold (i.e., a desired noise gain value.) an appropriate threshold may be derived from equation (12) as follows:(27)

in operation, the activity detector calculates an estimate of the filter noise gain, NG, as the sum of the squares of the adaptive filter taps (i.e.,). If the estimated value NG is much smaller than the predetermined threshold value, then the estimated value D is delayed_kIs set to 0, scale factor estimateIs set to a unity value to ensure proper system operation. It is sufficient to note that only one threshold value is stored, since the value of the noise gain NG is independent of the value of the delay D.

An example system may be implemented with the following pseudo code. Those skilled in the art will appreciate that such pseudo code is readily adjustable when implemented using standard digital signal processing elements.

Scaling factor and time delay estimation subroutine

Filtering: the outputs from the fixed FIR filter and the adaptive FIR filter are computed (k denotes the run-time index). Y1 ═ Y1 (k: -1: k-L); y2 ═ Y2 (k: -1: k-L); y1hat (k) ═ y 2' × C; y1fil (k) ═ Y1' × C0; err (k) ═ y1fil (k) -y1hat (k);

energy calculation and gain control: a simple gain control scheme is used to set the gain mu to 0 when there is low energy in the input. The instantaneous energy is compared to the long-term average. emom (k) ═ sum (y1hat (k: -1: k-L) · 2); effective (k) ═ 0.999 × effective (k-1) +0.001 × emom (k); if (emom (k) >. 5 away (k)) g (k) ═ mu; else g (k) ═ 0; end

N-LMS update: the adaptive filter coefficients are updated using the N-LMS algorithm. c + g (k) Y2 err (k)/((Y2' Y2) + 0.01);

updating of estimated values of S and D: the scale estimate is first order recursively smoothed, while D is estimated by an iterative gradient method. delta denotes the fixed time delay in channel 1.

LLC＝LL*C；

PPD＝[cos(warr*(1-Dhat+delta))；sin(warr*(1-Dhat+delta))]；

DPD＝[sin(warr*(1-Dhat+delta))；-cos(warr*(1-Dhat+delta))]；

shat＝(1-mu)*Shat+mu*sqrt((LLC′*LLC)/m)；Dhat＝Dhat+mu*DPD′*(LLC-Shat*PPD)；

Activity detection: if the estimated sum of the squares of the filter taps is 20dB less than the desired sum of the squares of the filter taps, then the gain is forced to unity and the delay estimate is brought closer to 0. C ═ C' — C; if (eC < 0.01 × eC0) shad ═ 1; dhat is 0; end

The estimates of S and D are further smoothed by a first order running average. Sh (k) ═ rho Sh (k-1) + (1-rho) × shade; dh (k) ═ rho Dh (k-1) + (1-rho) × Dhat; sh (k) ═ rho Sh (k-1) + (1-rho) × shade; dh (k) ═ rho Dh (k-1) + (1-rho) × Dhat;

to further illustrate the operation of the example embodiments, numerical examples using the above pseudo-code are given. In the example, an acoustic scenario is considered in which the sensor is assumed to be a microphone and the signal source is assumed to be a human speaker or a speaker conveying human speech. As described above, such an acoustic scenario may arise in the case of hands-free mobile phones used in an automotive environment. Although the example is limited to two sensors and signal sources, one skilled in the art will appreciate that the method may be used with any number of signal sources and sensors.

When the distance between the first signal source 210 and the first sensor 215 is 0.5 meters and the first sensor 215 is considered as a reference sensor, the actual time delay at the second sensor 225 with respect to the first signal source 210 is 2.25 samples D for a sampling rate of 8 kHz. With the same assumption, the actual time delay at the second sensor 225 relative to the second signal source 220 is-8.75 samples. For example, these assumptions are reasonable for a car compartment with a mobile phone (including second sensor 225) placed in a cradle that is close to the passenger and another sticker microphone (first sensor 215) placed on a visor in front of the driver (first signal source 210).

In such a vehicle cabin, there is typically severe background noise (e.g., from ac fans, vehicle engines, road surfaces, wind, etc.). For the purpose of numerical example, the sensitivities of the microphones in different directions are assumed to be given in table 1. TABLE 1 microphone sensitivity for diffuse background noise and signal sources in different locations

Signal source	First sensor 215 (e.g., visor microphone)	Second sensor 225 (e.g., an in-line microphone)
Signal source	First sensor 215 (e.g., visor microphone)	Second sensor 225 (e.g., an in-line microphone)	Diffuse background noise	0dB	0dB
First signal source 210 (e.g., driver)	+3dB	0dB	Diffuse background noise	0dB	0dB
First signal source 210 (e.g., driver)	+3dB	0dB	Second signal source 220 (e.g., passenger)	-10dB	0dB

In addition, a composite two-channel measurement is created that combines a male speaker located at the first signal source 210 and a female speaker located at the second signal source 220. These files are connected so that there is no speaker activity for the first second, then the male speaker is activated for 7 seconds, then the activity is stopped for 3 seconds, and then the female speaker is activated for 10 seconds. The signal-to-noise ratio for the second sensor 225 is 8dB for male speakers and 7dB for female speakers (measured over the entire active period). The speech signals detected at the first and second sensors 215, 225 are depicted in the first and second curves 610, 620 of fig. 6, respectively.

The additional background noise is modeled as white gaussian noise. The noise signals detected at the first and second sensors 215, 225 are depicted by the third and fourth curves 630, 640, respectively, in fig. 6. The combined speech and noise signals measured at the first and second sensors are depicted in the fifth and sixth curves 650, 660, respectively, in fig. 6.

In the simulation, the parameters used were: l-32, Δ -10, ω -2 pi 200/8000, m-1, μ -0.01 and rho-0.99. The results are given in figure 7. In particular, the delay estimate D_kDepicted in a first curve 710, a scale factor estimateDepicted in a second curve 720. In curves 710, 720, every 50 th sample is displayed. The horizontal dashed lines represent delays of-3, 0 and 9 samples and gains of-10 dB, 0dB, 3 dB. As given, when the driver speaks, the system correctly gives a scaling factor and time delay estimate of 0dB and-3 samples, respectively, and when the passenger speaks, the system correctly gives a scaling of-10 dB and 9 samples, respectivelyFactor and time delay estimate. Furthermore, the activity detector correctly sets the scaling factor and the time delay estimate to 0dB and 0 samples, respectively, while both the driver and the passenger remain silent.

Although these embodiments are described in the context of causal filtering (i.e., Δ > 0), the inventive idea is equally applicable in the context of non-causal filtering. In particular, for Δ ═ 0, the adaptive scheme includes an adaptive block that acts as a signal smoother, a backward predictor (D < 0) and/or a forward predictor (D > 0). Thus, it is not necessary to add a fixed delay to the signal stream (e.g., by delay block 410), and an adaptive scheme with minimal inherent delay may also be implemented. This feature is fundamentally practical in many real-time applications. However, because the accuracy of the estimated value will be somewhat worse in a non-causal approach (because of the narrower passband required to fix the FIR filter 230), the exact value of Δ may be set based on system design conditions. For example, Δ may be set to cover "most cases" rather than "all possible cases" because the system will provide reasonable results even in few extreme cases.

Those skilled in the art will appreciate that the present invention is not limited to the specific exemplary embodiments described herein for illustrative purposes. The scope of the invention is, therefore, indicated by the appended claims rather than by the foregoing description. Equivalents consistent with the meaning of the claims are also included.

Claims

1. A signal processing apparatus comprising:

a first signal sensor;

a first filter having an input connected to an output of the first sensor;

a second signal sensor;

a second filter having an input connected to an output of the second sensor and having an adjustable filter characteristic;

a summing device having a first input connected to the output of the first filter and a second input connected to the output of the second filter, wherein the adjustable filtering characteristic of the second filter is adjusted in dependence on the output of the summing device;

a processor for calculating an estimate of at least one parameter related to said first and second sensors, wherein said estimate is calculated as a function of the filter characteristic of said first filter and as a function of the adjustable filter characteristic of said second filter.

2. A signal processing apparatus according to claim 1, wherein said processor calculates an estimate of a relative time delay and a relative scale factor between said first and second sensors relative to a signal source.

3. The signal processing apparatus of claim 1, wherein the first filter, the second filter, the summing apparatus and the processor are implemented using a Digital Signal Processor (DSP) Integrated Circuit (IC).

4. The signal processing apparatus of claim 1, wherein the first filter, the second filter, the summing device and the processor are implemented with an Application Specific Integrated Circuit (ASIC).

5. A signal processing device according to claim 1, wherein said signal processing device is a telephone, and wherein said first and second sensors are microphones.

6. A signal processing device according to claim 1, wherein the signal processing device is a radio transceiver, wherein the first and second sensors are antenna elements.

7. The signal processing apparatus of claim 1, wherein said first filter is a Finite Impulse Response (FIR) filter having a fixed filtering characteristic.

8. The signal processing apparatus according to claim 1, wherein the filter characteristics of said first filter include at least one pass band giving a unity gain and a zero phase at a center frequency of the pass band.

9. The signal processing apparatus according to claim 1, wherein the filter characteristic of the first filter includes a coefficient set so that a noise gain of the first filter is minimum.

10. The signal processing apparatus according to claim 1, wherein the filter characteristic of the first filter includes a coefficient adjusted so that a signal-to-noise ratio of the first filter is optimal.

11. The signal processing apparatus of claim 1, wherein the adjustable filtering characteristic of the second filter is adjusted using a Normalized Least Mean Squares (NLMS) algorithm.

12. A signal processing device according to claim 1, wherein the adjustable filtering characteristic of said second filter is adjusted using a Least Mean Square (LMS) algorithm.

13. The signal processing apparatus of claim 1, wherein the adjustable filtering characteristic of the second filter is adjusted using a Recursive Least Squares (RLS) algorithm.

14. A signal processing apparatus according to claim 2, further comprising a beamformer for adjusting the beam patterns provided by said first and second transducers in accordance with relative time delays between said first and second transducers and estimates of relative scaling factors.

15. The signal processing apparatus of claim 14 wherein the beam pattern comprises a spatial beam directed at a particular signal source.

16. The signal processing apparatus of claim 14 wherein the beam pattern comprises a spatial notch pattern aimed at a particular signal source.

17. A signal processing apparatus according to claim 2, wherein said processor selects for transmission a signal detected by a particular one of said first and second sensors based on an estimate of a relative time delay between said first and second sensors and a relative scaling factor.

18. A signal processing device according to claim 1, further comprising at least one additional sensor and at least one additional filter having an adjustable filter characteristic,

wherein the adjustable filtering characteristic of the additional filter is adjusted in dependence on the difference between the outputs of the first filter and the additional filter,

wherein the processor calculates an estimate of at least one parameter related to the first sensor and the additional sensor based on the filter characteristic of the first filter and the adjustable filter characteristic of the additional filter.

19. A signal processing apparatus according to claim 18, wherein the processor calculates an estimate of a relative time delay and a relative scale factor between the first sensor and the additional sensor relative to the signal source.

20. A signal processing apparatus according to claim 2, further comprising an activity detector for detecting when a signal source of interest is activated, wherein said processor sets the relative time delay and the estimate of the relative scaling factor to predetermined values when said activity detector indicates that no signal source of interest is activated.

21. A signal processing apparatus according to claim 20, wherein the processor sets the estimate of the relative time delay to 0 and sets the estimate of the scaling factor to 1 when the activity detector indicates that no signal source of interest is activated.

22. The signal processing apparatus of claim 1, further comprising a fixed delay block located in a signal flow path corresponding to the first sensor.

23. The signal processing device according to claim 1,

wherein the filter characteristics of said first and second filters each comprise L filter coefficients,

wherein the filter characteristic of said first filter comprises m pass bands, each pass band i, l e (1, m) having a center frequency ω_l，

Wherein at time k, an estimate D of the relative time delay D between the first and second sensors_kAnd an estimate of a relative scaling factor S between said first and second sensorsIs based on the adjustable filtering characteristic of the second filterThe matrix L and the prediction vector p (D) are calculated as follows:where the matrix L and the prediction vector p (D) are calculated as follows:

24. method for processing a signal, comprising the steps of:

detecting a first signal with a first signal sensor;

filtering the first signal with a first filter to give a first filtered signal;

detecting a second signal with a second signal sensor;

filtering the second signal with a second filter to give a second filtered signal;

calculating a difference between the first filtered signal and the second filtered signal;

adjusting the filter characteristic of the second filter according to the difference obtained in the calculating step;

at least one parameter relating to the first and second sensors is estimated as a function of the filter characteristic of the first filter and as a function of the filter characteristic of the second filter.

25. The method of claim 24, wherein said step of estimating at least one parameter comprises the step of estimating a relative time delay and a relative scale factor between said first and second sensors related to the signal source.

26. The method of claim 24, wherein the filter characteristics of the first filter include at least one pass band that gives unity gain and zero phase delay at its center frequency.

27. The method of claim 24, wherein the filtering characteristics of the second filter are adjusted using a Normalized Least Mean Squares (NLMS) algorithm.

28. The method of claim 25, further comprising the step of adjusting the shape of the beam pattern given by said first and second sensors in accordance with the relative time delays and relative scale factors obtained as a result of said estimating step.

29. The method of claim 25, further comprising the step of selecting and transmitting a signal detected by a particular one of the first and second sensors based on a relative time delay and a relative scaling factor obtained as a result of the estimating step.

30. The method of claim 24, further comprising detecting whether the signal source of interest is activated and setting the relative time delay and the estimate of the scaling factor to predetermined values when said detecting step indicates that no signal source of interest is activated.