Robust Method of Echo Suppressor

Abstract

This invention proposed an Echo Suppressor which can efficiently suppress both echoes and background noise without introducing “choppiness”. The Echo Suppressor System includes said two adaptive gains Gr(RSR) and Gn(NSR), said one adaptive zeros-filter A1(z) and said one adaptive poles-filter A2(z); wherein, thr gain Gr(RSR) is controlled by RSR (Residual echo level to Signal level Ratio); the gain Gn(NSR) is controlled by NSR (Noise signal level to current Signal (Tx) level Ratio); the filter A1(z) is converted from LSF1 obtained from the first modification of LSFTx (Line Spectral Frequencies of Tx signal); the filter A2(z) is converted from LSF2 obtained from the second modification of LSFTx.

Description

BACKGROUND OF THE INVENTION

1. Field of the Invention

The invention relates to telecommunication equipment, and in particular, to a method for echo cancellation and suppression.

2. Background Art

Echo Cancellers

Echo Canceller (EC) is a device or software which could cancel echo signals by using some reference signal. The reference signal is also called received signal (Rx signal). The echo signal is also called echo return signal which is mixed in the transmission signal (Tx signal). There are two major types of echo cancellers: one is called Acoustic Echo Canceller (AEC); another one is named as Line Echo Canceller (LEC). Obviously, AEC is used to cancel acoustic echoes and LEC is employed to cancel line echoes.

Line echoes result from imperfect impedance matching by hybrids in two-to-four wire signal conversion. Acoustic echoes happen in the field of telecommunication equipment where acoustically coupled echoes are transmitted back to the receiving party while also providing a full duplex connection; it is also a problem in the filed of audio conferencing to prevent adding noise of the inactive talkers into the output while also providing a conference bridge where two or more conferees can talk at one. Some echo is acceptable in voice conversations, however, users are annoyed by listening to their own speech delayed by the round-trip time of the system.

Echo canceller often consists of an adaptive filter with Least Mean Square (LMS) algorithm which generates an echo replica signal similar to the echo signal. A subtraction between the echo return signal and the echo replica signal is conducted in order to cancel the echo return signal. Due to the fact that the replica signal could not be perfect enough to duplicate the echo return signal for various reasons, there is always some residual echo left in the transmission signal. Echo Suppression is an apparatus which could perform echo reduction and/or echo elimination, in particular for reducing or eliminating residual echo signals left after echo canceller. Due to the inherent problems associated with echo cancellers, many solutions rely significantly on an additional echo suppression stage.

Echo Suppression

Echo Suppressor is a device or software which could dramatically reduce the (residual) echo energy without significantly distorting the non-echo speech signal. Although Echo Suppressor can work alone without combining with echo canceller, it often works as a compensation of echo canceller. Not only Echo Suppressor can dramatically suppress the residual echo energy but also reduce background noise energy by taking the advantage of available parameters. Echo Suppressor could be viewed as an independent function or just a part of Echo Canceller system.

Existing Echo Suppression methods are mostly based on the following approaches:

• • Spectrum Echo Suppression is one of the popular approaches. This approach requires FFT operation to transform time domain signals into frequency domain; so, the complexity is not low. This approach needs to carefully tune all the parameters in order to avoid possible “music noise” before doing Inverse FFT of the modified spectrum back into time domain.
  • Another popular way is called NLP (Non-Linear Processor) which replaces the residual echo signal with random noise or turns off transmission in the reverse direction while one person is talking and double talk does not exist. This method is simple but requires very robust double talk detection and residual echo detection. Due to the fact that it is very difficult to have a precise double talk detector or perfect residual echo detection, this approach often introduces “choppiness”, especially in noisy environments.

This invention proposed an Echo Suppression approach which can achieve robust performance with low complexity level.

SUMMARY OF THE INVENTION

In accordance with the purpose of the present invention as broadly described herein, there is provided model and system for Echo Suppressor.

The invention proposed an Echo Suppressor which can efficiently suppress both echoes and background noise without introducing “choppiness”. The Echo Suppressor System includes said two adaptive gains G_r(RSR) and G_n(NSR), said one adaptive zeros-filter A₁(z) and said one adaptive poles-filter A₂(z); wherein, thr gain G_r(RSR) is controlled by RSR (Residual echo level to Signal level Ratio); the gain G_n(NSR) is controlled by NSR (Noise signal level to current Signal (Tx) level Ratio); the filter A₁(z) is converted from LSF₁ obtained from first modification of LSF_Tx (Line Spectral Frequencies of Tx signal); the filter A₂(z) is converted from LSF₂ obtained from second modification of LSF_Tx.

BRIEF DESCRIPTION OF THE DRAWINGS

The features and advantages of the present invention will become more readily apparent to those ordinarily skilled in the art after reviewing the following detailed description and accompanying drawings, wherein:

FIG. 1 shows an example of AEC system 112.

FIG. 2 shows an example of LEC system 209.

FIG. 3 shows an example of a double talker's spectrum.

FIG. 4 explains an example case with 3 signals of Rx and Tx.

DETAILED DESCRIPTION OF THE INVENTION

The present invention discloses an Echo Suppressor system which can suppress residual echoes and background noise even at double talker case, without introducing “choppiness”. The following description contains specific information pertaining to Echo Suppressor. However, one skilled in the art will recognize that the present invention may be practiced in conjunction with various algorithms different from those specifically discussed in the present application. Moreover, some of the specific details, which are within the knowledge of a person of ordinary skill in the art, are not discussed to avoid obscuring the present invention.

The drawings in the present application and their accompanying detailed description are directed to merely example embodiments of the invention. To maintain brevity, other embodiments of the invention which use the principles of the present invention are not specifically described in the present application and are not specifically illustrated by the present drawings.

FIG. 1 shows an example of AEC system 112. The echo signal 111 comes from speaker 110 and returns to micro phone 101. The signal which is sent to speaker is called received signal 109 or reference signal to AEC 103 and Echo Suppressor 105. The signal which goes into AEC 103 and comes from MIC 101 is called the transmission signal 102 which includes the echo return signal (originally from speaker). This echo return signal needs to be cancelled or suppressed by AEC 103 and Echo Suppressor 105. AEC usually works by using the reference signal to generate some replica signal which is similar to the echo return signal; and then subtracting or suppressing the echo return signal. The residual echo signal 104 will be further suppressed through a post-processing block 105 which could be Echo Suppressor or Non-Linear Processor (NLP).

FIG. 2 shows an example of LEC system 209. Typically the echo signal 211 returns from the telephone hybrid 210. This echo return signal 211 needs to be cancelled or suppressed by LEC system 209. LEC usually works in a similar way as AEC. One of the major differences between them is that the echo path is different. The echo delay range could be also different. The residual echo signal 203 needs to be further suppressed through a post-processing block 204 which could be Echo Suppressor or Non-Linear Processor (NLP).

FIG. 3 shows a double talker's case that the residual echo signal is mixed with the non echo speech signal. 302 shows the spectral envelop of the mixed signal. 301 assumes the spectral envelop of the speech signal without echo signal. In most cases, residual echo formant 303 is much smaller than speech formant 304. It is much more difficult to suppress the residual echo in double talker area when it is mixed into speech signal than just the residual echo signal alone. This is because we want to suppress the residual echo signal without distorting the speech signal.

FIG. 4 explains an example case with 3 signals: Rx signal (reference signal) 401, Tx signal 402 before going to (residual) Echo Suppressor and Tx signal 406 as the output of (residual) Echo Suppressor. As shown in the figure, a successful (residual) Echo Suppressor could clean up both residual echo signal 404 and background noise 405 if it runs properly.

This invention proposed an Echo Suppression approach which can achieve robust performance with low complexity level. It works by gain-controlled filtering processing which can be defined by using the following filtering model:

$\begin{array}{cc}F\left(z\right)=G_n\left(\mathrm{NSR}\right)·G_r\left(\mathrm{RSR}\right)·\frac{A_1\left(z\right)}{A_2\left(z\right)}& \left(1\right)\end{array}$

where G_n( ) is a gain which is a function of NSR (or SNR); NSR is defined as background Noise level to Signal level Ratio. NSR is measured by analyzing the Tx signal before the echo suppressor and utilizing VAD (Voice Activity Detection) information. G_r( ) is a gain which is a function of RSR; RSR is defined as an estimate of Residual echo level to Signal level Ratio in the Tx signal; RSR estimation is more complicated as explained later. A₁(z) and A₂(z) are linear predictors consisting of LPC coefficients, which are converted from LSF₁ and LSF₂ where LSF means Line Spectral Frequencies. Both LPC coefficients and LSF are well-known parameters in speech processing domain, that are often used to represent spectral envelope. LSF₁ is obtained by first modification of LSF_Tx, wherein LSF_Tx is calculated from doing LPC analysis on the Tx signal (input signal of the echo suppressor); LSF₂ is obtained from second modification of LSF_Tx. The LSF modifications are controlled by the parameters SNR, RSR, and another set of LSF_Rx which is calculated from the LPC analysis on the Rx signal.

In the equation (1), the gains, which usually do not go down to zero but could be small enough in pure echo area so that pure echoes can not be heard, mainly contribute to non-double talk area to significantly reduce the energy of pure echo or noise. The gain factors could be also smaller than 1 in speech or double talk areas, which depend on the parameters of NSR and RSR. Because NSR and RSR are made changing smoothly or slowly, the gain factors are also changed smoothly in time to avoid any “choppiness”.

The LPC filters A₁(z) and A₂(z) in the equation (1) mainly contribute to suppress the residual echo spectral formants (see FIG. 3) in double talk area or to reduce noise spectrum magnitude in low SNR speech area. Since the parameters of LPC filters A₁(z) and A₂(z) change smoothly and slowly, no “choppiness” is introduced and no obvious speech distortion would be heard.

The basic concept has been summarized in the above. The details will be explained later.

• • Estimate of NSR (or SNR), which is defined as background Noise level to current Tx Signal level Ratio. Well-known approach can be used to determine this parameter. The background noise means the recent average background noise level when only the background noise exists in Tx signal. The signal level means the current frame or subframe signal level of Tx signal. When only the background noise exists, the NSR value should be around 1 in direct ratio domain or around 0 in dB domain. In speech area, NSR value should be smaller than 1 in direct ratio domain.
  • Residual Echo Signal Detector, which is designed to detect the most residual echo signal areas where only residual echo signal and noise exist. This detector does not need to be accurate because it will be only used to estimate the average echo energy loss comparing to the Rx signal energy. After the delay between the Rx signal and the echo return signal is detected, both the Rx signal and the residual echo signal have been already made synchronized in echo canceller. When double talk does not exist, the energy loss of the residual echo signal should be very significant after the basic echo canceller, comparing to the original Rx signal energy. This information can be used to detect most residual echo signal.
  • Estimate of RSR, which is defined as the Residual echo level to the current Tx Signal level Ratio. The signal level still means the current frame or subframe Tx signal level. It is more complicated to evaluate the residual echo level. If there is no residual echo, the residual echo level is zero. First, an average residual echo energy loss is estimated in the areas where only residual echoes exist, which is defined as a Ratio of the Residual echo level to the corresponding Received signal level (the Rx signal level). The energy level can be in direct domain or in dB domain. Let's note the average ratio (or the running mean of the ratio) is

$\begin{array}{cc}\mathrm{RRR}=\text{Average}\left\{\frac{\text{Residual Echo Energy Level}}{\mathrm{Rx}\text{Signal Energy Level}}\right\}& \left(2\right)\end{array}$

So, the current residual echo energy level can be estimates as

Current Residial Echo Level=(RRR) (Current Rx Signal Level)   (3)

With the above formula, the current residual echo level can also be estimated during even double talk areas. Finally, RSR is calculated by

$\begin{array}{cc}\mathrm{RSR}=\frac{\text{Current Residual Echo Level}}{\mathrm{Current}\mathrm{Tx}\mathrm{Signal}\mathrm{Level}}& \left(4\right)\end{array}$

According to the definition of (4), RSR is around the value of 1 in pure residual echo areas and it should be smaller than 1 in double talk areas.

Gain G_n(NSR) could be linear function or non-linear function of the parameter NSR. An example of the linear function can be

G _n(NSR)=1−NSR   (5)

where C_n is a constant: 0<C_n<1

Gain G_r(RSR) could be linear function or non-linear function of the parameter RSR. An example of the linear function can be

G _r(RSR)=1−C_r·RSR   (6)

where C_r is a constant: 0<C_r<1

Estimate of LSF of Tx signal, LSF_Tx(i), i=0, 1, . . . , Order-1, is based on LPC-analyzing the Tx signal. Typical number of Order is around 10 for narrow band signal at the sampling rate of 8 kHz.

Estimate of LSF of noise signal, LSF_nois(i), i=0, 1, . . . , Order-1, is based on the average (or running mean) of LSF_Tx(i) in background noise areas of the Tx signal.

Estimate of LSF of Rx signal, LSF_Rx(i), i=0, 1, . . . , Order-1, is based on LPC-analyzing the Rx signal.

Estimate of LSF of (residual) echo signal, LSF_echo(i), i=0, 1, . . . , Order-1, is more difficult especially in double talk areas where residual echo signal is mixed in speech signal. As an example, LSF_echo(i) can be calculated by using prediction from LSF_Rx(i). First, the prediction factors, P(i), are evaluated by taking the recent average ratio (or running mean ratio) between LSF_Rx(i) and LSF_Tx(i) in the pure residual echo areas:

$\begin{array}{cc}P\left(i\right)=\text{Average}\left(\frac{{\mathrm{LSF}}_{\mathrm{Tx}}\left(i\right)}{{\mathrm{LSF}}_{\mathrm{Rx}}\left(i\right)}\right),0,1,\dots ,\text{Order}-1& \left(7\right)\end{array}$

Then, the current LSF_echo(i) of residual echo is estimated by

LSF _echo(i)=P(i)·LSF_Rx(i), i=0, 1, . . . , Order-1   (8)

where LSF_Rx(i) is the current Linear Spectral Frequencies of the Rx signal.

LPC predictors A₁(z) and A₂(z) are respectively converted from two sets of LSF, noted as LSF₁(i) and LSF₂(i), i=0, 1, . . . , Order-1. Both LSF₁(i) and LSF₂(i) are based on modifications of LSF_Tx(i). The modifications are mainly influenced by LSF_echo(i), LSF_nois(i), NSR, and RSR. As an example of the modifications, LSF₁(i) and LSF₂(i) are constructed as following:

LSF ₁(i)=λ₁·LSF_Tx(i)+β·LSF_echo(i)+α·LSF_nois(i), i=0,1, . . . ,Order-1   (9)

LSF ₂(i)=λ₂·[LSF_Tx(i)−β·LSF_echo(i)−α·LSF_nois(i)], i=0,1, . . . ,Order-1   (10)

where

β=C_β·RSR,   (11)

α=C_α·NSR,   (12)

C_α and C_β are constants. Their values should be larger than zero and much smaller than 1. λ₁ and λ₂ are determined in the following way:

$\begin{array}{cc}{\lambda }_1=1-\beta -\alpha ,& \left(13\right)\\ {\lambda }_2=\frac{1}{{\lambda }_1},& \left(14\right)\end{array}$

• • EC Filter Divergence Protection means a way to avoid the level of Tx signal after Echo Canceller and Echo Suppressor to be larger than the one before Echo Canceller (noted as Tx1 signal). The two gains G_n and G_r in the equation (1) are tuned in a simple way to make sure that the output energy of the Echo Suppressor is always smaller or equal to the level of the Tx1 signal.

The present invention may be embodied in other specific forms without departing from its spirit or essential characteristics. The described embodiments are to be considered in all respects only as illustrative and not restrictive. The scope of the invention is, therefore, indicated by the appended claims rather than the foregoing description. All changes which come within the meaning and range of equivalency of the claims are to be embraced within their scope.

Claims

1. An Echo Suppressor System includes said two adaptive gains Gr(RSR) and Gn(NSR), said one adaptive zeros-filter A1(z) and said one adaptive poles-filter A2(z); wherein, thr gain Gr(RSR) is controlled by RSR;the gain Gn(NSR) is controlled by NSR;the filter A1(z) is converted from LSF1 obtained from first modification of LSFTx;the filter A2(z) is converted from LSF2 obtained from second modification of LSFTx;

2. The method of claim 1, wherein RSR is defined as Residual echo level to Signal level Ratio; the Signal level means the current frame or subframe Tx signal level; the Residual echo level is calculated by multiplying RRR with the current received signal (Rx) level.

3. The method of claim 2, wherein RRR is defined as the average Ratio of Residual echo level to Received signal (Rx) level.

4. The method of claim 1, wherein NSR is defined as Noise signal level to current Signal (Tx) level Ratio.

5. The method of claim 1, wherein LSF means Line Spectral Frequencies.

6. The method of claim 1, wherein the initial LSFTx set before the modification is obtained from LPC-analyzing Tx signal.

7. The method of claim 1, wherein the first modification of LSFTx and the second modification of LSFTx are controlled by RSR, NSR, LSFecho and LSFnois.

8. The method of claim 7, wherein LSFecho of residual echo is predicted from recent estimated relationship between LSFRx and LSFTx in pure residual echo areas of Tx signal, where LSFRx is obtained from LPC-analyzing Rx signal.

9. The method of claim 7, wherein LSFnois of background noise is obtained by LPC-analyzing background noise areas of Tx signal.