Echo reduction system

Abstract

The present invention relates to a method for reducing an echo in a microphone signal generated by a microphone, comprising echo compensating the microphone signal by subtracting an estimated echo signal from the microphone signal to generate an echo compensated signal, detecting a speech activity of a local speaker on the basis of the microphone signal and the estimated echo signal and suppressing a residual echo in the echo compensated signal on the basis of the detected speech activity to obtain an output signal. The invention further relates to a system for processing a microphone signal generated by a microphone, comprising echo compensation filtering means configured to receive and echo compensate the microphone signal to output an echo compensated signal based on the received microphone signal, a speech activity detection means configured to detect speech activity of a local speaker by receiving and analyzing the echo compensated signal and to output a detection signal and a residual echo suppressing means configured to receive the detection signal and to receive and filter the echo compensated signal on the basis of the detection signal to output an output signal.

Description

FIELD OF THE INVENTION

The present invention relates to a system and a method for signal processing, in particular, speech signal processing, with acoustic echo suppression. The invention particularly relates to echo suppression in loudspeaker-room-microphone systems exhibiting impulse responses that are time-dependent.

BACKGROUND OF THE INVENTION

Echo compensation is a basic topic in audio signal processing in communication systems comprising microphones that detect not only the desired signal, e.g., a speech signal of a user of a speech recognition system or a hands-free set, but also disturbing signals output by loudspeakers of the same communication system. In case of a hands-free set, e.g., it is not desired that signals received from a remote party and output by loudspeakers at the near end are fed again in the system by microphones at the near end and transmitted back to the remote party. Detection of signals by the microphones that are output by the loudspeakers can result in annoying acoustic echoes that even may cause a complete breakdown of the communication, if the acoustic echoes are not significantly attenuated or substantially removed.

In the case of a speech recognition system or a speech dialog system used in a noisy environment a similar problem occurs. It has to be prevented that signals different from the speech signals of a user are supplied to the recognition unit. The microphone(s) of the speech dialog system, however, might detect loudspeaker outputs representing, e.g., synthesized speech signals output by a speech dialog system or audio signals reproduced by audio devices as CD or DVD player or a radio. If these signals were not sufficiently suppressed in the microphone signal, the wanted signal representing the utterance of a user could be deteriorated to a degree that renders appropriate speech recognition impossible.

Echo suppression is particularly difficult, if the speaker using a microphone for communication with a remote communication party is moving as, e.g., a driver using a hands-free set who steers a wheel while communicating with a remote party by the hands-free telephone set. In this case, the impulse response of the loudspeaker-room-microphone (LRM) system is time-variant. Usually residual echoes are still present in the processed audio signals to be provided to a remote communication party. These residual echoes, e.g., result in so-called echo blips in hands-free telephone systems thereby deteriorating the microphone signal significantly, in particular, due to the huge delay of current mobile phone connections.

Several methods for echo compensation have been proposed and implemented in communication systems in recent years. Adaptive filters are employed for echo compensation of acoustic signals (see, e.g., Acoustic Echo and Noise Control, E. Hänsler and G. Schmidt, John Wiley & Sons, New York, 2004) that are used to model the transfer function of the LRM system by means of an adaptive finite impulse response (FIR) filter. If multiple loudspeaker signals are output by a number of loudspeakers separately, one filter has to be employed for each loudspeaker.

In present echo compensation processing an adaptive filter is used to model the impulse response of the LRM system to generate an estimate for the echo signal that can be subtracted from the microphone signal. The adaptation of the echo compensation filtering means is usually carried out by the normalized least mean square (NLMS) algorithm.

However, the echo compensation is a rather time-consuming and processor intensive procedure and usually is restricted to some portion of the impulse response of the LRM system. Thus, echo compensation is often supplemented by suppression of residual echoes by means of filtering the microphone signal after subtraction of the estimated echo signal with an appropriate time-varying impulse response. This supplementary filtering is usually performed in a restricted sub-band or some restricted Fast Fourier Transform (FFT) range by some version of a Wiener filter.

However, current echo reduction processing is still not reliable, in particular, in LRM systems that show time-varying impulse responses. Thus, despite the engineering process in recent years there is still a problem in satisfying echo reduction of audio signal, in particular, speech signals in communication system, e.g., in hands-free telephone sets and speech dialog systems.

DESCRIPTION OF THE INVENTION

The above mentioned problem is solved by a method according to claim 1 for enhancing the quality of a microphone signal, in particular, for reducing an echo in a microphone signal generated by a microphone, comprising

echo compensating the microphone signal by subtracting an estimated echo signal from the microphone signal to generate an echo compensated signal;

detecting a speech activity of a local speaker on the basis of the microphone signal and the estimated echo signal; and

suppressing a residual echo in the echo compensated signal on the basis of the detected speech activity to obtain an output signal.

Echo compensating is carried out by an adaptive echo compensation filtering means that models the loudspeaker-room-microphone system transfer by an impulse response. The impulse response given by N_ĥ filter coefficients ĥ_i(n), where n is the discrete time index, is folded with the audio signal x(n) to obtain an estimated echo signal $\hat{d}\left(n\right)=\sum _{i=0}^{N_{\hat{h}}-1}x\left(n-i\right){\hat{h}}_i\left(n\right)$ that is to be subtracted from the microphone signal. This microphone signal may, in general, include a speech signal from a local speaker using the microphone and background noise in addition to the echo resulting from a loudspeaker output, e.g., based on a speech signal received from a remote speaker. The adaptation of the adaptive echo compensation filtering means can be carried out by the normalized least mean square (NLMS) algorithm.

After echo compensation some residual echo is still present in the echo compensated microphone signal. According to the present invention this residual echo is suppressed in dependence on the result of a detection of speech activity of a local speaker that uses the microphone for the communication.

Speech activity is detected by analyzing the microphone signal and the estimated echo signal. If, e.g., the local speaker is silent, a relatively strong residual echo suppression can be carried out by a residual echo suppressing means. On the other hand, a different characteristic of the residual echo suppressing means is preferred when utterances of the local speaker are detected. Therefore, a very satisfying echo suppression of audio signals, in particular, speech signals in communication system, e.g., in hands-free telephone sets, speech recognition systems and speech dialog systems, is achieved.

It is noted that the herein disclosed signal processing can be carried out in the sub-band or the Fourier transform regime. In the following description of aspects of the present invention signal processing in the sub-band regime is described. It is understood that a corresponding processing in the Fourier regime after Fourier transform may alternatively be carried out.

For processing in the sub-band regime the microphone signal is converted by filter banks to sub-band microphone signals and the estimated echo signal comprises estimated sub-band echo signals. In this case, the echo compensating, detecting of speech activity (speaker is silent or is speaking) and suppressing of residual echo is carried out in the sub-band regime. After suppression of the residual echo in the sub-bands a synthesizing filter bank can be used to synthesize the desired output signal, which, e.g., is to be transmitted to a remote communication party, from the output sub-band signals.

According to one embodiment of the herein disclosed method the detecting of the speech activity of the local speaker comprises the steps of:

smoothing in frequency the microphone sub-band signals, in particular, by first order recursive filtering;

smoothing in frequency the estimated sub-band echo signals, in particular, by first order recursive filtering;

determining in each sub-band of a predetermined range of sub-bands a distance between the smoothed microphone sub-band signals and the smoothed estimated sub-band echo signals;

and wherein

the suppressing of the residual echo in the echo compensated signal is based on the determined distances in each sub-band of the predetermined range of sub-bands.

According to this embodiment the sub-band microphone signals Y(e^jΩμ,n) as well as the estimate for the echo signal in each sub-band {circumflex over (D)}(e^jΩμ,n) are smoothed in frequency over a predetermined range M of sub-bands μ=0, M−1: S_{{circumflex over (d)}{circumflex over (d)},smooth}(Ω_μ,n)=smooth└{circumflex over (D)}(e^jΩ0,n), {circumflex over (D)}(e^jΩ1,n), . . . , {circumflex over (D)}(e^jΩM-1,n)┘S_yy,smooth(Ω_μ,n)=smooth└Y(e^jΩ0,n), Y(e^jΩ1,n), . . . , Y(e^jΩM-1,n)┘ where Ω_μ denotes the mid-frequency of the sub-band μ and “smooth” indicates some kind of a proper smoothing function. The pre-determined range of sub-bands may preferably cover 200 Hz to 3500 Hz. This range, generally, shows a significant power for speech signals. For example, the magnitude or the square of the magnitude of the sub-band microphone signals and of the estimated echo sub-band signals may be smoothed in both the positive (Ω₀ to Ω_M-1) and negative (Ω_M-1 to Ω₀) direction in frequency.

Suppressing the residual echo in the echo compensated signal in dependence on some distance measures or differences of the smoothed microphone sub-band signals and the estimated echo sub-band signals, e.g., differences of the respective magnitudes, in each sub-band of the predetermined range of sub-bands provides an efficient and satisfying manner to suppress residual echoes on the basis of the detected speech activity to obtain an output signal with an enhanced quality.

According to one advantageous embodiment of the inventive method for reducing an echo in a microphone signal the power density spectrum of some background noise present in the microphone signal is estimated in sub-bands; and

smoothing in frequency of the microphone sub-band signals comprises recursive filtering the power density spectrum of the sub-band microphone signals to obtain a smoothed power density spectrum of the microphone sub-band signals;

smoothing in frequency the estimated sub-band echo signals comprises recursive filtering the power density spectrum of the estimated sub-band echo signals to obtain a smoothed power density spectrum of the estimated sub-band echo signals;

and wherein

determining in each sub-band a distance between the smoothed microphone sub-band signals and the smoothed sub-band echo signals comprises

determining in each sub-band the maximum of the smoothed power density spectrum of the microphone sub-band signals and the estimated background noise power spectrum enhanced by a first predetermined noise overestimate factor to obtain a modified microphone power density spectrum;

determining in each sub-band the maximum of the smoothed power density spectrum of the estimated sub-band echo signals and the estimated background noise power spectrum enhanced by a second predetermined noise overestimate factor, which may be the same as the first one, to obtain a modified echo power density spectrum;

comparing the modified microphone power density spectrum and the modified echo power density spectrum to obtain a spectrum distance measure;

and wherein

the suppressing of the residual echo in the echo compensated sub-band signals is based on the spectrum distance measure.

In this embodiment smoothing is carried out by ${\tilde{S}}_{\hat{d}\hat{d},\mathrm{smooth}}\left({\Omega }_0,n\right)=\frac{1}{2}\left[{\hat{D}\left(e^{j{\Omega }_0},n\right)}^2+{\hat{D}\left(e^{j{\Omega }_1},n\right)}^2\right]$$\mathrm{and}$${\tilde{S}}_{\hat{d}\hat{d},\mathrm{smooth}}\left({\Omega }_{\mu },n\right)={\lambda }_{\mathrm{Fre}}{\tilde{S}}_{\hat{d}\hat{d},\mathrm{smooth}}\left({\Omega }_{\mu -1},n\right)+\left(1-{\lambda }_{\mathrm{Fre}}\right){\hat{D}\left(e^{j{\Omega }_{\mu }},n\right)}^2,\mathrm{for}0<\mu <M.$ in the positive direction and $S_{\hat{d}\hat{d},\mathrm{smooth}}\left({\Omega }_{M-1},n\right)=\frac{1}{2}\left[{\tilde{S}}_{\hat{d}\hat{d},\mathrm{smooth}}\left({\Omega }_{M-2},n\right)+{\tilde{S}}_{\hat{d}\hat{d},\mathrm{smooth}}\left({\Omega }_{M-1},n\right)\right]$$S_{\hat{d}\hat{d},\mathrm{smooth}}\left({\Omega }_{\mu },n\right)={\lambda }_{\mathrm{Fre}}{S}_{\hat{d}\hat{d},\mathrm{smooth}}\left({\Omega }_{\mu +1},n\right)+\left(1-{\lambda }_{\mathrm{Fre}}\right){\tilde{S}}_{\hat{d}\hat{d},\mathrm{smooth}}\left({\Omega }_{\mu },n\right),\mathrm{for}0\le \mu <M-1$ in the negative direction. S_{{circumflex over (d)}{circumflex over (d)},smooth}(Ω_μ,n) is the smoothed power density spectrum of the estimated sub-band echo signals. Smoothing of the sub-band microphone signals Y(e^jΩμ,n) is performed accordingly to obtain the smoothed power density spectrum of the sub-band microphone signals S_yy,mod(Ω_μ,n). Experiments have shown that, e.g., for a typical sampling rate of 11025 Hz and M=256 sub-bands the smoothing parameter λ_Fre is advantageously chosen as 0.2≦λ_Fre≦0.8.

After the smoothing process the maximum values of the respective smoothed spectra and the power density spectrum of background noise (in sub-bands) are calculated for the sub-bands to obtain a modified microphone power density spectrum and a modified echo power density spectrum S_{{circumflex over (d)}{circumflex over (d)},mod}(Ω_μ,n)=max{S_{{circumflex over (d)}{circumflex over (d)},smooth}(Ω_μ,n),K_bŜ_bb(Ω_μ,n)}and S_yy,mod(Ω_μ,n)=max{S_yy,smooth(Ω_μ,n),K_bŜ_bb(Ω_μ,n)}.

Experiments have shown that the noise overestimate factor K_b may, e.g., be chosen as 2≦K_b≦16. It might be preferred that different noise overestimate factors are chosen for the modified microphone power density and the modified echo power density. Comparison of the modified spectra provides some spectrum distance measure that can advantageously be used for the controlling of the suppressing of the residual echo. As a result, an output signal with a previously unknown echo reduction can be achieved. The method of this embodiment has proven to be particularly reliable and efficient for echo reduction in a time-variant loudspeaker-room-microphone (LRM) system. One specific example for the comparison of the modified spectra is given in the detailed description of the embodiment below.

According to one example of the inventive method estimating the power density of the echo compensated signal and estimating the power density of the residual echo is carried out. In this case, the suppressing of the residual echo in the echo compensated signal is based on the estimated power density of the echo compensated signal and the estimated power density of the residual echo. The power densities are given by the squares of the magnitudes of the respective signals.

If, e.g., the estimated power density of the echo compensated signal greatly exceeds the estimated power density of the residual echo, suppression of the residual echo might be very faint in order not to modify the already intelligible microphone signal too much. If, on the other hand, the estimated power density of the residual echo exceeds the estimated power density of the echo compensated signal a more aggressive filtering of the echo compensated signal is necessary.

The suppressing of the residual echo in the echo compensated signal may comprise filtering the echo compensated signal by a filter with the filter characteristic (frequency response) $G\left(e^{j{\Omega }_{\mu }},n\right)=1-\beta \left(n\right)\frac{{\hat{S}}_{\mathrm{\varepsilon \varepsilon }}\left({\Omega }_{\mu },n\right)}{{\hat{S}}_{\mathrm{ee}}\left({\Omega }_{\mu },n\right)}$ where Ŝ_ee(Ω_μ,n) and Ŝ_εε(Ω_μ,n) denote the estimated power density of the echo compensated signal and the estimated power density of the residual echo and β(n) is a filter parameter depending on the detected speech activity. If speech activity, e.g., measured by a spectrum distance measure as mentioned above exceeding some pre-determined detection threshold, is detected, β(n) is high, e.g., about β(n)=1000 and otherwise it is low, e.g., β(n)=1. In particular, the β(n) is a time-dependent parameter to account for a time-variant LRM system.

One may also prefer to limit the suppression to a pre-determined value given by G_min, i.e. $G\left(e^{j{\Omega }_{\mu }},n\right)=\max \left\{G_{\min },1-\beta \left(n\right)\frac{{\hat{S}}_{\mathrm{\varepsilon \varepsilon }}\left({\Omega }_{\mu },n\right)}{{\hat{S}}_{\mathrm{ee}}\left({\Omega }_{\mu },n\right)}\right\}.$

Experiments have proven that the shown filter characteristic is very efficient in suppressing residual echoes in already echo compensated microphone signals in dependence on detected speech activity of a local speaker using the microphone that generates the microphone signal that is to be improved in quality before transmission to a remote communication party.

According to an even more efficient but somewhat more complicated filtering of the echo compensated signal for suppressing a residual echo in the echo compensated signal the filter characteristic might be chosen as $G\left(e^{j{\Omega }_{\mu }},n\right)=\{\begin{array}{cc}1-\beta \left(n\right)\frac{{\hat{S}}_{\mathrm{\varepsilon \varepsilon }}\left({\Omega }_{\mu },n\right)}{{\hat{S}}_{\mathrm{ee}}\left({\Omega }_{\mu },n\right)},& \mathrm{if}\stackrel{\_}{C}\left(n\right)>C_{\mathrm{thres}}\\ 1-\beta \left(n\right)\frac{\max \left\{{\hat{S}}_{\mathrm{\varepsilon \varepsilon }}\left({\Omega }_{\mu },n\right),{\hat{D}\left(e^{j{\Omega }_{\mu }},n\right)}^2\right\}}{{\hat{S}}_{\mathrm{ee}}\left({\Omega }_{\mu },n\right)},& \mathrm{else}\end{array}\mathrm{or}\mathrm{as}G\left(e^{j{\Omega }_{\mu }},n\right)=\{\begin{array}{cc}1-\beta \left(n\right)\frac{{\hat{S}}_{\mathrm{\varepsilon \varepsilon }}\left({\Omega }_{\mu },n\right)}{{\hat{S}}_{\mathrm{ee}}\left({\Omega }_{\mu },n\right)},& \mathrm{if}\stackrel{\_}{C}\left(n\right)>C_{\mathrm{thres}}\\ 1-\beta \left(n\right)\frac{{\hat{D}\left(e^{j{\Omega }_{\mu }},n\right)}^2}{{\hat{S}}_{\mathrm{ee}}\left({\Omega }_{\mu },n\right)},& \mathrm{else}\end{array}$ where C(n) is a measure for the detected speech activity and C_thres is a predetermined threshold. For example, C(n) may be a time-averaged spectrum distance measure. Again suppression may be limited to a pre-determined value G_min as mentioned above.

In a further example of the inventive method the power density of the output signal after echo compensation and echo suppression processing as described above and the power density of background noise present in the microphone signal are determined and compared. If the power density of the background noise exceeds the power density of the output signal, artificial noise (so-called comfort noise) is transmitted to a remote communication party instead of the output signal that may usually be transmitted to the remote party. By this, it is avoided that residual echo suppression even suppresses background noise detected by the microphone which would result in annoying abrupt changes in the background noise level received by the remote communication party.

The present invention also provides a computer program product, comprising one or more computer readable media having computer-executable instructions for performing the steps of the herein disclosed method according to one of the above described examples.

The above mentioned problems are also solved by the system for processing a microphone signal generated by a microphone according to claim 11, comprising

echo compensation filtering means configured to receive and echo compensate the microphone signal to output an echo compensated signal based on the received microphone signal;

speech activity detection means configured to detect speech activity of a local speaker by receiving and analyzing the echo compensated signal and to output a detection signal in accordance with the result of the speech detection; and

residual echo suppressing means configured to receive the detection signal and to receive and filter the echo compensated signal on the basis of the detection signal to suppress a residual echo and to output an output signal.

The system further comprising filter banks configured to convert the microphone signal and another audio signal to be output by at least one loudspeaker installed in the same room as the microphone or Fourier transform means configured to Fourier transform the microphone signal and the other audio signal.

According to one embodiment the herein disclosed system further comprises a background noise estimation means configured to estimate the background noise power spectrum of background noise present in the microphone signal, and wherein the speech activity detection means comprises

recursive filtering means configured to smooth the power density spectrum of sub-band microphone signals to obtain a smoothed power density spectrum of the microphone sub-band signals;

recursive filtering means configured to smooth the power density spectrum of estimated sub-band echo signals to obtain a smoothed power density spectrum of the estimated sub-band echo signals;

determining means configured to determine in each sub-band the maximum of the smoothed power density spectrum of the microphone sub-band signals and the estimated background noise power spectrum enhanced by a predetermined noise overestimate factor and to generate a modified microphone power density spectrum of the determined maximum values;

determining means configured to determine in each sub-band the maximum of the smoothed power density spectrum of the estimated sub-band echo signals and the estimated background noise power spectrum enhanced by the predetermined noise overestimate factor and to generate a modified echo power density spectrum of the determined maximum values; and

comparing means configured to compare the modified microphone power density spectrum and the modified echo power density spectrum and to generate a spectrum distance signal;

and wherein

the residual echo suppressing means is configured to receive the spectrum distance signal and to receive and filter the echo compensated signal on the basis of the spectrum distance signal.

In the above examples of the inventive system the speech activity detection means can be configured to estimate the power density of the echo compensated signal and the power density of a residual echo present in the echo compensated signal; and the residual echo suppressing means can be configured to suppress the residual echo in the echo compensated signal based on the estimated power density of the echo compensated signal and the estimated power density of the residual echo.

The residual echo suppressing means may advantageously comprise a filtering means with one of the following filter characteristic $G\left(e^{j{\Omega }_{\mu }},n\right)=1-\beta \left(n\right)\frac{{\hat{S}}_{\mathrm{\varepsilon \varepsilon }}\left({\Omega }_{\mu },n\right)}{{\hat{S}}_{\mathrm{ee}}\left({\Omega }_{\mu },n\right)}$$\mathrm{or}$$G\left(e^{j{\Omega }_{\mu }},n\right)=\{\begin{array}{cc}1-\beta \left(n\right)\frac{{\hat{S}}_{\mathrm{\varepsilon \varepsilon }}\left({\Omega }_{\mu },n\right)}{{\hat{S}}_{\mathrm{ee}}\left({\Omega }_{\mu },n\right)},& \mathrm{if}\stackrel{\_}{C}\left(n\right)>C_{\mathrm{thres}}\\ 1-\beta \left(n\right)\frac{\max \left\{{\hat{S}}_{\mathrm{\varepsilon \varepsilon }}\left({\Omega }_{\mu },n\right),{\hat{D}\left(e^{j{\Omega }_{\mu }},n\right)}^2\right\}}{{\hat{S}}_{\mathrm{ee}}\left({\Omega }_{\mu },n\right)},& \mathrm{else}\end{array}\mathrm{or}G\left(e^{j{\Omega }_{\mu }},n\right)=\{\begin{array}{cc}1-\beta \left(n\right)\frac{{\hat{S}}_{\mathrm{\varepsilon \varepsilon }}\left({\Omega }_{\mu },n\right)}{{\hat{S}}_{\mathrm{ee}}\left({\Omega }_{\mu },n\right)},& \mathrm{if}\stackrel{\_}{C}\left(n\right)>C_{\mathrm{thres}}\\ 1-\beta \left(n\right)\frac{{\hat{D}\left(e^{j{\Omega }_{\mu }},n\right)}^2}{{\hat{S}}_{\mathrm{ee}}\left({\Omega }_{\mu },n\right)},& \mathrm{else}\end{array}$ where the mid-frequency of the sub-band μ is denoted by Ω_μ, Ŝ_ee(Ω_μ,n), Ŝ_εε(Ω_μ,n) denote the estimated power density of the echo compensated signal and the estimated power density of the residual echo, n is the discrete time index and β(n) is a filter parameter depending on the detected speech activity, and where C(n) is a measure for the detected speech activity and C_thres is a predetermined threshold.

The filter characteristic of the residual echo filtering means my alternatively be limited by replacing one of the above mentioned characteristics by max[G_min, G(e^jΩμ,n)] with some pre-determined value G_min.

The system may further comprise a noise generator configured to generate artificial noise;

a means configured to determine the power density of the output signal and of background noise present in the microphone signal;

a means configured to compare the power density of the output signal with the power density of the background noise; and

a control means configured to cause transmission of the output signal to a remote party, if the power density of the output signal exceeds the power density of the background, and to cause transmission of the generated artificial noise, noise if the power density of the background noise exceeds the power density of the output signal.

Thus, the desired signal that is, e.g., to be transmitted to the remote communication party or to be recognized by a speech recognition system, may be composed of the sub-band signals $\hat{S}\left(e^{j{\Omega }_{\mu }},n\right)=\{\begin{array}{cc}E\left(e^{j{\Omega }_{\mu }},n\right)G\left(e^{j{\Omega }_{\mu }},n\right),& \mathrm{if}\mathrm{no}\mathrm{comfort}\mathrm{noise}\mathrm{is}\mathrm{to}\mathrm{be}\mathrm{output}\\ B\left(e^{j{\Omega }_{\mu }},n\right),& \mathrm{if}\mathrm{comfort}\mathrm{noise}\mathrm{only}\mathrm{is}\mathrm{to}\mathrm{be}\mathrm{output}\end{array}$ where E(e^jΩμ,n) and B(e^jΩμ,n) denote the echo compensated signal and the artificial noise signal (in sub-bands).

It is also provided a hands-free telephone set comprising one of the above described examples of a system for processing a microphone signal. The above described examples are particular useful for enhancing the quality of a speech signal transmitted by a hands-free set to a remote communication party. In particular, the intelligibility of the transmitted signal processed for echo compensation and echo suppression as disclosed above is enhanced with respect to the prior art when the LRM system temporarily changes. In this context, a vehicle communication system is provided comprising at least one microphone, at least one loudspeaker and the system according to one of the examples above or the mentioned hands-free telephone set. The disclosed system is particularly efficient in reducing echoes in situations in which a speaker driving a car is moving, e.g., due to steering a wheel.

The present invention furthermore provides a speech recognition system or a speech dialog system comprising the system according to one of the above examples. The reliability of recognition results of speech inputs processed by an example of the inventive system is greatly enhanced as compared to the art.

Additional features and advantages of the invention will be described in detail with reference to the drawings. In the description, reference is made to the accompanying figures that are meant to illustrate preferred embodiments of the invention. It is understood that such embodiments do not represent the full scope of the invention that is defined by the claims given below.

FIG. 1 shows the structure of an echo reduction system for enhancing the quality of a microphone signal generated on the basis of a speech signal, a loudspeaker signal and background noise detected by a microphone.

FIG. 2 illustrates an example of the detection of speech activity that represents an important element of the disclosed echo reduction processing.

FIG. 3 shows basic elements of the suppression of a residual echo in an echo compensated microphone signal comprising detection speech activity and adaptation of a filter characteristic of a residual echo suppressing means.

FIG. 4 shows a result of a simulation experiment for the disclosed echo reduction processing of a microphone signal.

In the following, an example of the signal processing (system) disclosed in this application is described in detail with respect to FIGS. 1 to 3. As shown in FIG. 1 a telephony hands-free set comprises a microphone 1 and a loudspeaker 2. Utterances of a local speaker are detected by the microphone 1 and the loudspeaker 2 generates a loudspeaker signal based on an audio signal x(n) provided by a remote communication party.

The microphone not only detects the speech signal s(n) of a locate speaker but also a background noise signal b(n) and the loudspeaker-room-microphone (LRM) transfer signal d(n) based on the impulse response of the LRM system h(n). The microphone signal y(n), thus, includes contributions of the speech signal s(n), the background noise signal b(n) and the echo signal d(n).

By n the discrete time index is denoted. In this example, echo compensation and residual echo suppression performed by signal processing in sub-bands is described. Alternatively, processing in the frequency range (Fast Fourier Transform range) can be performed after Fourier transforming the respective audio signal x(n) and the microphone signal y(n).

A first filter bank means 3 generates sub-band signals X(e^jΩμ,n) from the audio signal x(n) and sub-band signals Y(e^jΩμ,n) from the microphone signal y(n). The mid-frequency of the sub-band μ is denoted by Ω_μ. The audio sub-band signals X(e^jΩμ,n) are filtered by an adaptive echo compensation filtering means 5. The filter coefficients of the echo compensation filtering means 5 are determined in order to model the impulse response h(n) in the sub-bands. An estimate for the echo signal in each sub-band {circumflex over (D)}(e^jΩμ,n) is obtained by folding the audio sub-band signals X(e^jΩμ,n) with the impulse response of the echo compensation filtering means in each sub-band Ĥ(e^jΩμ,n).

By subtracting the estimated echo {circumflex over (D)}(e^jΩμ,n) from the sub-band microphone signals Y(e^jΩμ,n) echo compensated signals in each sub-band E(e^jΩμ,n) are obtained. The echo compensated signals E(e^jΩμ,n) are further processed for suppression of residual echoes by a residual echo reduction means 6. The residual echo reduction means 6 may comprise or be supplemented with a noise reduction means for substantially removing the background noise contribution of the echo compensated signals E(e^jΩμ,n). The sub-band output signals Ŝ(e^jΩμ,n) obtained by the residual echo reduction means 6 can subsequently by synthesized to obtain a desired signal composed of the sub-band signals that can be transmitted to a remote communication party.

The present invention is mainly concerned with the realization of the residual echo reduction means 6. It is known in the art to simply employ some variant of a Wiener filter making use of the estimated power density spectrum of the residual echo Ŝ_εε(Ω_μ,n) and of the echo compensated sub-band signals Ŝ_ee(Ω_μ,n). The Wiener filter may exhibit the following filter characteristic $G\left(e^{j{\Omega }_{\mu }},n\right)=\max \left\{G_{\min },1-\beta \frac{{\hat{S}}_{\mathrm{\varepsilon \varepsilon }}\left({\Omega }_{\mu },n\right)}{{\hat{S}}_{\mathrm{ee}}\left({\Omega }_{\mu },n\right)}\right\}$ wherein the maximum damping can be pre-determined by the parameter G_min and the sensibility of the filter is controlled by the parameter β. If β>1 the damping might by too strong, thereby damping also the desired signal below an appropriate level.

According to the present invention the filter characteristic (frequency response) of an employed residual echo filtering means can be adjusted such that a very sensitive (“aggressive”) damping is carried out when no speech activity of the local speaker is detected. The inventive method secures that the detection of speech activity is satisfying even when the speaker is moving. In particular, the herein disclosed method is able to distinguish between speech signals of a local speaker and output by a loudspeaker (i.e. provided by a remote speaker) in a time-variant LRM system. The filter characteristic of a residual echo reduction means can then be adapted on the basis of the detected speech activity of the local speaker.

An example for the detection of speech activity according to the present invention is illustrated in FIG. 2. A noise estimating means 7 receives the sub-band microphone signals Y(e^jΩμ,n) in order to estimate the power density spectrum Ŝ_bb(Ω_μ,n) (in sub-bands) of the background noise b(n) detected by the microphone 1. The noise estimating means 7 may, e.g., make use of a minimum statistics in order to estimate the power density spectrum Ŝ_bb(Ω_μ,n) (see, e.g., An Efficient Algorithm to Estimate the Instantaneous SNR of Speech Signals, R. Martin, EUROSPEECH 1993, Berlin, Conf. Proceed., p. 1093-1096, September 1993).

The sub-band microphone signals Y(e^jΩμ,n) as well as the estimate for the echo signal in each sub-band {circumflex over (D)}(e^jΩμ,n) are smoothed 8, 9 in frequency over the predetermined M sub-bands μ=0, M−1.

A particular efficient smoothing function can be realized by a recursive filter of 1^st order for smoothing the magnitudes or the squares of the magnitudes in positive an negative direction of the frequency range. According to the present example smoothing is carried according to ${\tilde{S}}_{\hat{d}\hat{d},\mathrm{smooth}}\left({\Omega }_0,n\right)=\frac{1}{2}\left[{\hat{D}\left(e^{j{\Omega }_0},n\right)}^2+{\hat{D}\left(e^{j{\Omega }_1},n\right)}^2\right]$$\mathrm{and}$${\tilde{S}}_{\hat{d}\hat{d},\mathrm{smooth}}\left({\Omega }_{\mu },n\right)={\lambda }_{\mathrm{Fre}}{\tilde{S}}_{\hat{d}\hat{d},\mathrm{smooth}}\left({\Omega }_{\mu -1},n\right)+\left(1-{\lambda }_{\mathrm{Fre}}\right){\hat{D}\left(e^{j{\Omega }_{\mu }},n\right)}^2,\mathrm{for}0<\mu <M.$ in the positive direction and $S_{\hat{d}\hat{d},\mathrm{smooth}}\left({\Omega }_{M-1},n\right)=\frac{1}{2}\left[{\tilde{S}}_{\hat{d}\hat{d},\mathrm{smooth}}\left({\Omega }_{M-2},n\right)+{\tilde{S}}_{\hat{d}\hat{d},\mathrm{smooth}}\left({\Omega }_{M-1},n\right)\right]$$S_{\hat{d}\hat{d},\mathrm{smooth}}\left({\Omega }_{\mu },n\right)={\lambda }_{\mathrm{Fre}}{S}_{\hat{d}\hat{d},\mathrm{smooth}}\left({\Omega }_{\mu +1},n\right)+\left(1-{\lambda }_{\mathrm{Fre}}\right){\tilde{S}}_{\hat{d}\hat{d},\mathrm{smooth}}\left({\Omega }_{\mu },n\right),\mathrm{for}0\le \mu <M-1$ in the negative direction. Experiments have shown that for a typical sampling rate of 11025 Hz and M=256 sub-bands the smoothing parameter is advantageously chosen as 0.2≦λ_Fre≦0.8. Smoothing of the sub-band microphone signals Y(e^jΩμ,n) is performed accordingly in order to obtain the smoothed microphone spectrum S_yy,mod(Ω_μ,n).

The processing means 8 and 9 are also configured to receive the output of the noise estimating means 7 and to determine the maximum of the smoothed estimated echo spectrum given above and the corresponding microphone spectrum and the estimate for the power density spectrum Ŝ_bb(Ω_μ,n) of the background noise, respectively S_{{circumflex over (d)}{circumflex over (d)},mod}(Ω_μ,n)=max{S_{{circumflex over (d)}{circumflex over (d)},smooth}(Ω_μ,n),K_bŜ_bb(Ω_μ,n)}and S_yy,mod(Ω_μ,n)=max{S_yy,smooth(Ω_μ,n),K_bŜ_bb(Ω_μ,n)}. where the background noise is overestimated by K_b. Experiments have shown that K_b may be chosen from 2≦K_b≦16 for satisfying results for the echo suppression. Some distance (difference) w₁S_{{circumflex over (d)}{circumflex over (d)},mod}(Ω_μ,n)−w₂S_yy,mod(Ω_μ,n), where w₁ and w₂ are properly chosen weight functions (e.g., depending on Ω_μ) or constants, can be used to determine a distance measure indicative for speech activity by which the filter characteristic for suppressing residual echoes can be controlled.

In the present example, however, it is determined whether a strong level increase or decrease of the microphone signal and/or the estimate echo signal is detected. Strong temporary level jumps would probably result in artifacts when the distance measure for determining the speech activity of a local speaker is calculated as follows. If no abrupt level changes are present, i.e. temporarily relatively homogeneous signals are present, the smoothed output signals of the processing means 8 and 9, i.e. S_yy,smooth(Ω_μ,n) and S_{{circumflex over (d)}{circumflex over (d)},mod}(Ω_μ,n), respectively, are used for signal flank detection 10$\Delta \left({\Omega }_{\mu },n\right)=\{\begin{array}{cc}0,\mathrm{if}& \left(S_{\hat{d}\hat{d},\mathrm{smooth}}\left({\Omega }_{\mu },n\right)>K_{\Delta }{S}_{\hat{d}\hat{d},\mathrm{smooth}}\left({\Omega }_{\mu },n-1\right)\right)\\ & ⩔\left(S_{\hat{d}\hat{d},\mathrm{smooth}}\left({\Omega }_{\mu },n\right)<\frac{1}{K_{\Delta }}{S}_{\hat{d}\hat{d},\mathrm{smooth}}\left({\Omega }_{\mu },n-1\right)\right)\\ & ⩔\left(S_{\mathrm{yy},\mathrm{smooth}}\left({\Omega }_{\mu },n\right)>K_{\Delta }{S}_{\mathrm{yy},\mathrm{smooth}}\left({\Omega }_{\mu },n-1\right)\right)\\ & ⩔\left(S_{\mathrm{yy},\mathrm{smooth}}\left({\Omega }_{\mu },n\right)<\frac{1}{K_{\Delta }}{S}_{\mathrm{yy},\mathrm{smooth}}\left({\Omega }_{\mu },n-1\right)\right)\\ 1,& \mathrm{else}\end{array}$ with a detection threshold of typically 4≦K_Δ≦100.

By a distance detection means 11 a spectrum distance measure can be determined on the basis of Δ(Ω_μ,n) and the modified spectra S_{{circumflex over (d)}{circumflex over (d)},mod}(Ω_μ,n) and S_yy,mod(Ω_μ,n) calculated by the processing means 8 and 9: $C\left({\Omega }_{\mu },n\right)=\{\begin{array}{cc}\Delta \left({\Omega }_{\mu },n\right)C_{1,},& \mathrm{if}\lfloor S_{\hat{d}\hat{d},\mathrm{mod}}\left({\Omega }_{\mu },n\right)>K_1{S}_{\mathrm{yy},\mathrm{mod}}\left({\Omega }_{\mu },n\right)\rfloor ,\\ \Delta \left({\Omega }_{\mu },n\right)C_2,& \mathrm{if}\left[K_2{S}_{\mathrm{yy},\mathrm{mod}}\left({\Omega }_{\mu },n\right)\le S_{\hat{d}\hat{d},\mathrm{mod}}\left({\Omega }_{\mu },n\right)\le K_1{S}_{\mathrm{yy},\mathrm{mod}}\left({\Omega }_{\mu },n\right)\right],\\ \Delta \left({\Omega }_{\mu },n\right)C_3,& \mathrm{if}\left[K_3{S}_{\hat{d}\hat{d},\mathrm{mod}}\left({\Omega }_{\mu },n\right)\le S_{\mathrm{yy},\mathrm{mod}}\left({\Omega }_{\mu },n\right)\le K_4{S}_{\hat{d}\hat{d},\mathrm{mod}}\left({\Omega }_{\mu },n\right)\right],\\ \Delta \left({\Omega }_{\mu },n\right)C_4,& \mathrm{if}\left[S_{\mathrm{yy},\mathrm{mod}}\left({\Omega }_{\mu },n\right)>K_4{S}_{\hat{d}\hat{d},\mathrm{mod}}\left({\Omega }_{\mu },n\right)\right],\\ 0,& \mathrm{else}.\end{array}$

Suitable choices for the detection thresholds are, e.g., K₁=16, K₂=4, K₃=4, and K₄=16. Distances C(Ω_μ,n) are specified, e.g., by detection parameters C₁=−0.4, C₂=0.1, C₃=0.1, and C₄=0.6. Large positive values of the spectrum distance measure C(Ω_μ,n) indicate that the power density of the microphone spectrum dominates the power density of the estimated echo. If the power density of the estimated echo significantly exceeds the power density of the microphone spectrum, strong changes in the LRM system are detected, characterized by a high negative cost parameter C₁.

The determination of both Δ(Ω_μ,n) and C(Ω_μ,n) are restricted to those sub-bands that show a significant power of speech signals. This implies that the sub-bands are restricted to με[μ_start,μ_end] where μ_start and μ_end are chosen corresponding to a frequency range coverage of about 200 Hz to about 3500 Hz.

An adding means 12 sums up the results C(Ω_μ,n) for the individual sub-bands $C\left(n\right)=\sum _{\mu ={\mu }_{\mathrm{start}}}^{{\mu }_{\mathrm{end}}}C\left({\Omega }_{\mu },n\right).$

Subsequently, smoothing over a pre-determined time interval is preferably performed to obtain a smoothed distance measure C(n).

The detection result obtained by the distance detection means 12 can now be used for an efficient adaptation of the filter characteristic of a residual echo suppressing means 6 (see FIG. 1) in dependence of the detected speech activity as measured by C(n). If no speech activity is detected on the basis of the smoothed distance measure C(n), the residual echo filtering can be performed by means of the square of the magnitude of the estimated echo {circumflex over (D)}(e^jΩμ,n) or the maximum of |{circumflex over (D)}(e^jΩμ,n)|² and the power density of the residual echo Ŝ_εε(Ω_μ,n). If, however, a double talk situation is present, i.e. both the local and the remote speaker are speaking, i.e. significant speech activity is detected, it is preferred that a Wiener filter characteristic but with a time-dependent filter parameter β(n) is used. In conclusion, residual echo filtering according to the present example is carried out by means of the filter characteristic $G_{\mathrm{mod}}\left(e^{j{\Omega }_{\mu }},n\right)=\max \left\{G_{\min },{\tilde{G}}_{\mathrm{mod}}\left(e^{j{\Omega }_{\mu }},n\right)\right\}$$\mathrm{with}$$\tilde{G}\left(e^{j{\Omega }_{\mu }},n\right)=\{\begin{array}{cc}1-\beta \left(n\right)\frac{{\hat{S}}_{\mathrm{\varepsilon \varepsilon }}\left({\Omega }_{\mu },n\right)}{{\hat{S}}_{\mathrm{ee}}\left({\Omega }_{\mu },n\right)},& \mathrm{if}\stackrel{\_}{C}\left(n\right)>C_{\mathrm{thres}}\\ 1-\beta \left(n\right)\frac{\max \left\{{\hat{S}}_{\mathrm{\varepsilon \varepsilon }}\left({\Omega }_{\mu },n\right),{\hat{D}\left(e^{j{\Omega }_{\mu }},n\right)}^2\right\}}{{\hat{S}}_{\mathrm{ee}}\left({\Omega }_{\mu },n\right)},& \mathrm{else}.\end{array}$

The parameter β(n) controlling the sensibility of the filter depends on the smoothed distance measure $\beta \left(n\right)=\{\begin{array}{cc}{\beta }_1,& \mathrm{if}\stackrel{\_}{C}\left(n\right)>C_{\mathrm{thres}}\\ {\beta }_2>{\beta }_1,& \mathrm{else}.\end{array}$

Experiments have shown that suitable values for the β—parameters are, e.g., β₁=1 and β₂=1000. By C_thres a predetermined threshold is given above which significant speech activity of the local speaker is considered to be present. Suppression is limited by G_min, e.g., G_min=0.1.

The power density spectrum Ŝ_ee(Ω_μ,n) of the echo compensated sub-band microphone signals E(e^jΩμ,n) can be recursively determined by Ŝ_ee(Ω_μ,n)=λ_eŜ_ee(Ω_μ,n−1)+(1−λ_e)|E(e^jΩμ,n)|² with a smoothing parameter chosen as 0≦λ_e≦1.

In the present example the power density spectrum of the residual echo Ŝ_εε(Ω_μ,n) is determined by Ŝ_εε(Ω_μ,n)=Ŝ_xx(Ω_μ,n)|Ĥ_Δ(e^jΩμ,n)|² wherein the estimated echo compensation obtained by the echo compensation filtering means 5 Ĥ_Δ(e^jΩμ,n) can be calculated by methods known in the art (see, e.g., Step-Size Control in Subband Echo Cancellation Systems, G. Schmidt, IWANEC 1999, Pocona Manor, Pa., USA, Conf. Proceed., p. 116-119, 1999). The estimated power density spectrum of the audio signal output by the loudspeaker Ŝ_xx(Ω_μ,n) is calculated similar to the power density spectrum Ŝ_εε(Ω_μ,n) of the echo compensated sub-band microphone signals: Ŝ_xx(Ω_μ,n)=λ_xŜ_xx(Ω_μ,n−1)+(1−λ_x)|X(e^jΩμ,n)|² with the smoothing parameter chosen as 0≦λ_x≦1.

FIG. 3 shows an overview of the residual echo suppression according to an example of the present invention. The echo compensated sub-band microphone signals E(e^jΩμ,n) are filtered by a residual echo suppression means 6. For this filtering a filter characteristic G_mod(e^jΩμ,n) has to be adapted. Adaptation of the filter characteristic G_mod(e^jΩμ,n) is carried out on the basis of the result of the speech activity detection 13 (relating to the speech activity of a local speaker) C(n) or the smoothed distance measure C(n), to be more specific. In particular, the above mentioned β—parameter is controlled 14 by the detected speech activity.

As explained above a means for the detection of speech activity of the local speaker 13 receives the estimate for the echo sub-band signals {circumflex over (D)}(e^jΩμ,n) as well as the microphone sub-band signals Y(e^jΩμ,n) and, in addition, the output of a noise estimation means 7. The power density spectrum of the background noise Ŝ_bb(Ω_μ,n) obtained by the noise estimation means 7 is input in an artificial noise generator 15.

The noise generator is configured to generate artificial noise with substantially the same statistical power distribution as determined for the background noise by the noise estimation means 7. The artificially generated background noise sub-band signals B(e^jΩμ,n) are output, if the output signal output by the residual echo suppressing means 6 would have a power density below the remaining background noise in order to avoid annoying abrupt changes of the background noise transmitted to the remote communication party. In the present example the residual echo suppressing means 6 also controls whether this so-called comfort noise shall be output. Thus, the desired signal to be transmitted to the remote communication party is eventually achieved by synthesizing the sub-band signals $\hat{S}\left(e^{j{\Omega }_{\mu }},n\right)=\{\begin{array}{cc}E\left(e^{j{\Omega }_{\mu }},n\right)G_{\mathrm{mod}}\left(e^{j{\Omega }_{\mu }},n\right),& \mathrm{if}\mathrm{no}\mathrm{comfort}\mathrm{noise}\mathrm{is}\mathrm{to}\mathrm{be}\mathrm{output}\\ B\left(e^{j{\Omega }_{\mu }},n\right),& \mathrm{if}\mathrm{comfort}\mathrm{noise}\mathrm{only}\mathrm{is}\mathrm{to}\mathrm{be}\mathrm{output}.\end{array}$

FIG. 4 illustrates the efficiency of the herein disclosed signal processing for echo reduction of a microphone signal by means of a computer simulation study. In all panels the abscissa show the amplitude of a signal in arbitrary units and the ordinate shows the time in seconds.

The upper panel shows a simulated speech signal of a local speaker and the second panel a corresponding microphone signal. The microphone signal comprises the speech signal as well as an echo contribution. During the simulation run the LRM system was changed by simulating speaker's movements after about 5 second and about 15 seconds. After about 20.5 seconds a double talk situation was simulated for a period of about 7 seconds, i.e. the simulated speech signal and echo contributions are detected by the microphone.

The smoothed distance measure C(n) as a result of the speech activity detection is displayed in the third panel. As can be clearly seen from FIG. 4, the smoothed distance measure C(n) does not or only very slightly respond to the speakers movements. In particular, C(n) lies below the threshold C_thres=1 and, thus, significant suppressing of residual echoes is carried out when the speaker moves. Contrary, during the period characterized by double-talk the smoothed distance measure C(n) indicates speech activity of the local speaker. As a result, the eventually obtained desired output signal Ŝ(e^jΩμ,n) shows a satisfying concordance with the simulated speech signal of the local speaker.

It is to be understood that some or all of the above described features can also be combined in different ways. Whereas in the discussed example hands-free telephony is considered, the disclosed algorithms can be applied for reducing the echoes in microphone signals, in general.

Claims

1. A method for reducing an echo and/or a residual echo in a microphone output signal, the method comprising: generating an estimated echo signal based on received audio input signals; subtracting the estimated echo signal from the microphone output signal to generate an echo compensated signal; detecting speech activity of a local speaker based on the microphone output signal and the estimated echo signal; and suppressing the residual echo in the echo compensated signal based on the detected speech activity to obtain an echo suppressed output signal.

2. The method according to claim 1, comprising separating the microphone output signal into a plurality of sub-band microphone signals, and separating the audio input signal into a plurality of estimated sub-band echo signals.

3. The method according to claim 2, where generating the echo compensated signal, detecting speech activity, and suppressing the residual echo are performed for the respective sub-band ranges.

4. The method according to claim 3, where detecting the speech activity of the local speaker comprises: smoothing the sub-band microphone signals in frequency to generate respective smoothed sub-band microphone signals; smoothing the estimated sub-band echo signals in frequency to generate respective smoothed estimated sub-band echo signals; and determining for the respective sub-bands, a distance between the smoothed sub-band microphone signals and the smoothed estimated sub-band echo signals.

5. The method according to claim 4, where suppressing the residual echo is based on the determined distance between the smoothed sub-band microphone signals and the smoothed estimated sub-band echo signals for the respective sub-bands.

6. The method according to claim 4, where smoothing the estimated sub-band echo signals is performed using a first order recursive filter.

7. The method according to claim 4, further comprising: estimating a power density spectrum of background noise present in the microphone output signal; where smoothing the sub-band microphone signals includes recursively filtering a power density spectrum of the respective sub-band microphone signals to generate a smoothed power density spectrum of the respective sub-band microphone signals; and where smoothing the estimated sub-band echo signals includes recursively filtering a power density spectrum of the respective estimated sub-band echo signals to obtain a smoothed power density spectrum of the respective estimated sub-band echo signals.

8. The method according to claim 7, where determining the distance between the smoothed sub-band microphone signals and the smoothed estimated sub-band echo signals further comprises: determining in each sub-band a maximum of the smoothed power density spectrum of the sub-band microphone signals and the estimated power density spectrum of the background noise enhanced by a first predetermined noise overestimate factor, to obtain a modified microphone power density spectrum; determining in each sub-band the maximum of the smoothed power density spectrum of the estimated sub-band echo signals and the estimated power density spectrum of the background noise enhanced by a second predetermined noise overestimate factor, to obtain a modified echo power density spectrum; comparing the modified microphone power density spectrum and the modified echo power density spectrum to obtain a spectrum distance measure; and suppressing the residual echo in the respective echo compensated sub-band signals based on the spectrum distance measure.

9. The method according to claim 1, further comprising: estimating a power density of the echo compensated signal; and estimating a power density of the residual echo, where suppressing the residual echo in the echo compensated signal is based on the estimated power density of the echo compensated signal and the estimated power density of the residual echo.

10. The method according to claim 9, where suppressing the residual echo in the echo compensated signal further comprises filtering the echo compensated signal with a filter having a frequency response of:

11. The method according to claim 9, wherein suppressing the residual echo in the echo compensated signal further comprises filtering the echo compensated signal with a filter having a frequency response of:

12. (canceled)

13. (canceled)

14. The method according to claim 1, further comprising: determining a power density of the echo-suppressed output signal; determining a power density of background noise present in the microphone output signal; comparing the power density of the echo-suppressed output signal with the power density of the background noise; transmitting the echo-suppressed output signal to a remote party if the power density of the echo-suppressed output signal exceeds the power density of the background noise; and transmitting artificial noise to the remote party if the power density of the echo-suppressed output signal does not exceed the power density of the background noise.

15. A computer-readable storage medium having processor executable instructions to reduce an echo and/or a residual echo in a microphone output signal, by performing the acts of: generating an estimated echo signal; subtracting the estimated echo signal from the microphone output signal to generate an echo compensated signal; detecting speech activity of a local speaker based on the microphone output signal and the estimated echo signal; and suppressing the residual echo in the echo compensated signal based on the detected speech activity to obtain an echo suppressed output signal.

16. The computer-readable storage medium of claim 15, further comprising processor executable instructions to cause a processor to perform the acts of separating the microphone output signal into a plurality of sub-band microphone signals, and separating the estimated echo signal into a plurality of estimated sub-band echo signals.

17. The computer-readable storage medium of claim 16, further comprising processor executable instructions to cause a processor to perform the acts of generating the echo compensated signal, detecting speech activity and suppressing the residual echo, all in the respective sub-band ranges.

18. The computer-readable storage medium of claim 17, further comprising processor executable instructions to cause a processor to perform the acts of detecting the speech activity of the local speaker by: smoothing the sub-band microphone signals in frequency to generate respective smoothed sub-band microphone signals; smoothing the estimated sub-band echo signals in frequency to generate respective smoothed estimated sub-band echo signals; and determining for the respective sub-bands, a distance between the smoothed sub-band microphone signals and the smoothed estimated sub-band echo signals.

19. (canceled)

20. (canceled)

21. (canceled)

22. (canceled)

23. (canceled)

24. (canceled)

25. A system for reducing an echo and/or a residual echo in a microphone output signal, the microphone receiving audio input signals, the system comprising: an echo compensation filter configured to receive the audio input signals and generate an estimated echo signal; a speech activity detector configured to detect speech of a local speaker; a combining circuit configured to generate an echo compensated signal by subtracting the estimated echo signal from the microphone output signal; and a residual echo reduction circuit configured to suppress the residual echo in the echo compensated signal based on the detected speech activity, and output an echo suppressed output signal.

26. The system according to claim 25, comprising: a microphone output filter bank configured to separate the microphone output signal into a plurality of sub-band microphone signals; and an audio input filter bank configured to separate the audio input signals into a plurality of sub-band audio signals to facilitate generation of a plurality of sub-band estimated echo signals.

27. The system according to claim 26, where the estimated echo signal, the echo compensated signal, the detected speech of the local speaker and suppression of the residual echo are processed for the respective sub-band ranges.

28. The system according to claim 27, where the speech activity detector further comprises: a first smoothing circuit configured to smooth the sub-band microphone signals in frequency to generate respective smoothed sub-band microphone signals; a second smoothing circuit configured to smooth the estimated sub-band echo signals in frequency to generate respective smoothed estimated sub-band echo signals; and a distance detection circuit configured to determine for the respective sub-bands, a distance between the smoothed sub-band microphone signals and the smoothed estimated sub-band echo signals.

29. The system according to claim 28, where the residual echo is suppressed based on the determined distance between the smoothed sub-band microphone signals and the smoothed estimated sub-band echo signals for the respective sub-bands.

30. The system according to claim 28, further comprising a first order recursive filter configured to smooth the estimated sub-band echo signals.

31. The system according to claim 25, further comprising: logic configured to estimate a power density of the echo compensated signal; and logic configured to estimate a power density of the residual echo, where the residual echo in the echo compensated signal is suppressed based on the estimated power density of the echo compensated signal and the estimated power density of the residual echo.

32. The system according to claim 31, where the residual echo reduction circuit includes a filter configured to suppress the residual echo in the echo compensated signal by filtering the echo compensated signal, the filter having a frequency response of:

33. (canceled)

34. (canceled)

35. (canceled)

36. The system according to claim 25, further comprising: logic configured to determine a power density of the echo-suppressed output signal; logic configured to determine a power density of background noise present in the microphone output signal; a noise generator configured to generate artificial noise; comparing logic configured to compare the power density of the echo-suppressed output signal with the power density of the background noise; and where the echo-suppressed output signal is transmitted to a remote party if the power density of the echo-suppressed output signal exceeds the power density of the background noise, and artificial noise is transmitted to the remote party if the power density of the echo-suppressed output signal does not exceed the power density of the background noise.

37. (canceled)