Nonlinear processor for acoustic echo canceller with...

Details Nonlinear processor for acoustic echo canceller with... Nonlinear processor for acoustic echo canceller with...

1998-08-31

2001-08-28

{haeck over (S)}mits, T{overscore (a)}livaldis Ivars (Department: 2641)

Telephonic communications

Subscriber line or transmission line interface

Network interface device

C455S570000, C704S233000

Reexamination Certificate

active

06282286

ABSTRACT:

FIELD OF THE INVENTION
This invention pertains to the field of adaptive, speech echo cancellation, and more particularly to acoustic echo cancellation for speaker-phones and voice conferencing systems utilizing a nonlinear processor (NLP).
BACKGROUND
Nonlinear processors (NLPs) are used in echo cancellation generally, and in particular for echo cancellation of acoustic speech signals. Speech echo cancellation can be grouped into two major categories: network echo cancellation and acoustic echo cancellation. The primary difference between acoustic echo signals and network echo signals is that an acoustic echo channel includes both loudspeaker and microphone transducers that convert signals to and from audible (acoustic) sound signals, as opposed to network echo signals that are generated by electric circuits (hybrids). The acoustic type typically has high background noise signals present from the surrounding environment that makes application of prior art nonlinear processors unfavorable.
PRIOR ART
The term “nonlinear processor” or NLP can be used to describe a signal processing circuit or algorithm that is placed in the speech path after echo cancellation, so as to provide further attenuation or removal of residual echo signals that cannot be cancelled completely by an echo canceller. A non-linearity, a distortion, or an added noise signal are examples of signals that can not be fully cancelled by an echo canceller, and these signals are typically removed or attenuated by a nonlinear processor. One example of a prior art NLP is a “center clipper” in which all signal samples with amplitude less than a threshold value are set to zero. This method has been used for network echo cancellation for many years by many different equipment suppliers. A description of the operation of such an NLP has been included in the appendix of the ITU-T G.165 recommendation as a reference design for an NLP. A known problem with this type of NLP is the so called “noise gating” phenomena wherein a party listening to the resulting speech signals, after a center clipping NLP, hears the background noise signals disappearing and then reappearing during periods of activation and de-activation of the NLP.
Improvements upon this center clipper method that reduce or eliminate the “noise gating” problem have been introduced in recent years. These improvements are primarily used for network type echo cancelers in which background noise levels are typically very low in comparison to the noise levels experienced with acoustic echo signals. An example of a prior art NLP improvement is a center clipper method combined with the injection of a matched artificial noise source to mask the removal of noise signals by the center clipper. Yet another example is a variable attenuator that provides a soft-switched transition between on/off states of signal attenuation with complementary soft-switched injection of artificial noise. U.S. Pat. No. 5,274,705, which issued Dec. 28, 1993 to Younce et al, describes another example of an improved NLP using dual thresholds in the NLP transfer function which allows transparent transfer of low level noise signals if below the low threshold, and transparent transfer of large signals if above an upper threshold while removing or modifying any signals in-between the two thresholds.
Problems with all of the aforementioned methods arise when dealing with signals from an acoustic environment because of the higher noise levels. Noise injection methods are not typically used because the character of the background noise changes very noticeably if an artificial noise is injected in place of the original noise. Variable attenuation methods without noise injection appear to be most commonly used for the control of residual echo in acoustic echo cancelers. This appears to be an extension of methods used previously by half-duplex speakerphones and network echo suppressors which used complementary attenuators to provide switched loss to control echo. The use of echo cancellation for a “full duplex” hands-free telephone appears to also make use of prior art complementary attenuators with reduced attenuation “depth” to make the connection close to full duplex, or perhaps, subjectively, “full duplex”. Some other implementations appear to allow complete full-duplex communication some of the time (e.g. during double-talk periods), while providing some extra attenuation control of echo residual during other periods of time (e.g. single talk periods). All of these methods cause audible changes in background noise signals producing some degradation of overall subjective performance.
The prior art dual threshold method when applied to acoustic background noise signals, produces noticeable levels of extra signal distortion. This distortion is caused by the changes made to signals when the NLP is on. This audible distortion changes the character of the background noise during speech from the far end side, and can best be described as a raspy type noise with some high frequency components that sound different than a typical background noise. Note as used in this description the far end talker is the party who is also listening to the resulting signal after the NLP.
Another problem with the prior art NLP is that it has no control over a long echo path environment. To save cost most echo cancelers can only deal with a short echo length (e.g. 128 ms or less). In some acoustic environments, the echo can last for about 0.5 to 1 sec. Although, in most cases, the echo residual is very small after 128 ms, when both sides of telephone line are quiet, even a very small echo residual is noticeable. After the loudspeaker has been quiet for over ½ sec, the echo may still be present at the microphone input. The echo residual is treated as near-end single talk by the speaker-phone, and therefore the NLP will not attenuate this signal.
SUMMARY OF THE INVENTION
The method used in the present invention builds upon the dual threshold method. The NLP turns on only if both a double talk condition and an echo suppression requirement are met.
The present invention further relates to a method of reducing the level of extra signal distortion by processing signals in a different manner than the methods described in prior art NLP designs. The signal will be transparent if it is detected to be noise, otherwise a noise prediction value is sent out.
In the present invention, the long echo residual is dealt with by the new NLP structure. In lab tests, the echo residual is significantly reduced with the new NLP structure, even in the case when echo signals last up to 1 sec. and the adaptation algorithm can only deal with 100 ms echo length.
Briefly, the NLP structure of the present invention determines whether the residual signal from the echo canceller is greater or less than an estimated noise level. If it is less than the estimated noise level the residual signal is passed through the NLP substantially unchanged. If the residual signal is greater than the estimated noise level it is further evaluated to determine whether or not it represents a near-end speech signal. If it is speech as in near-end single talk or double talk the residual signal is again passed through the NLP unchanged. If, however the incoming signal is echo residual or long term echo the NLP outputs a low level noise signal which represents a prediction based on previous noise samples.
Therefore in accordance with a first aspect of the present invention there is provided a non linear processor (NLP) for use with an acoustic echo canceller associated with a telephone terminal to selectively reduce residual signals therefrom. The NLP comprises: a first input to receive the residual signal; a second input to receive a reference signal representing a signal from a far end user; a third input for receiving a near end signal from a microphone in the terminal; an output for delivering a NLP output to a far end user; a NLP switch, switchable between a first position wherein the residual signal is passed directly to the output and a second position wherein a signal repre

Affiliated with

Also associated with

Rating

Nonlinear processor for acoustic echo canceller with... does not yet have a rating. At this time, there are no reviews or comments for this patent.

If you have personal experience with Nonlinear processor for acoustic echo canceller with..., we encourage you to share that experience with our LandOfFree.com community. Your opinion is very important and Nonlinear processor for acoustic echo canceller with... will most certainly appreciate the feedback.

Rate now

Comments { 0 }

Profile ID: LFUS-PAI-O-2442304