Double talk, NLP and comfort noise

Abstract

A state machine for attenuating the transition into and out of NLP state to reduce voice clipping and to reduce echo leak in a voice over packet signal transmission. The state machine interposes two additional transitional states between the NLP active state and the NLP inactive state to eliminate the sharp transition of NLP activity. An NLP entering state is used to gradually reduce the mixing ratio of echo and voice to attenuate the transition from passage of echo in the NLP inactive state to the suppression of echo in the NLP active state. An NLP exiting state is used to gradually increase the mixing ratio of echo and voice to attenuate the transition from suppression of echo in the NLP active state to passage of echo in the NLP inactive state. The rate of increase/decrease is determined based upon the signal parameters, including the presence of double talk.

Description

FIELD OF THE INVENTION

[0001] The invention relates to double talk detection and the application and/or disengagement of NLP logic for the enhanced processing of voice over packet speech including echo cancellation. Specifically, the present invention relates to enhanced differentiation of echo and double talk and the introduction of intermediate states to reduce echo leak and speech clipping.

BACKGROUND OF THE INVENTION

[0002] FIG. 1 illustrates a simplified block diagram of an equipment configuration for one terminal of a communication link which includes a near end hybrid. The communication link has a near-end 8 comprising a telephone 2, a four-to-two wire hybrid circuit 3, an echo canceller circuit 4, a filter 5 and a NLP 6. A far-end connected to communication network 23, can be similarly configured but is not illustrated in FIG. 1. During a conversation between the near-end and far-end users, the far end signal, x, which contains both the far-end user's speech and incidental background noise, enters the near-end 8 as signal x at node 9.

[0003] FIG. 2 illustrates a representative communication link 20 between two telephones 24 and 25. The link is comprised of a near-end 21, a far-end 22, and a communication network 23 that interconnects the near-end 21 and far-end 22. The near-end 21 has a user telephone 24, a hybrid circuit 26, and an echo canceller circuit 28. Similarly, the far-end 22 has a user telephone 25, a hybrid circuit 27, and an echo canceller circuit 29. Far-end signal power, X, is received by the near-end. Signal Y is the coupled echo signal from the far-end signal as well as the near-end signal produced by telephone 24. This near-end signal contains both the speech of the near-end telephone user and the background noise of the user's environment. Together, the near-end signal and far-end echo signal are represented by Y.

[0004] The far-end signal is provided to the four-to-two wire hybrid circuit 3 and then to near-end telephone 2. Due to the unavoidable non-linearities present in the hybrid circuit 3, some portion of the far-end signal power is coupled onto the output 7 of the hybrid circuit 3 as an echo. A composite signal y exists at node 7 containing the echo signal and the combined speech of the near-end user and any incidental background noise from the near-end user's environment. A filter having a filter length period selected and designed to be longer than the hybrid dispersion time is used prior to power level measurements at 7 to allow the echo canceller 4 to operate properly.

[0005] Echo canceller 4 synthesizes the expected value e of the echo signal in adaptive filter 5, and subtracts this value at 10 from the composite signal y existing at node 7. The resulting difference signal, d, existing at node 14, is intended to contain only the near-end signal s originating from telephone 2. Ultimately, difference signal, d, is provided to the far-end telephone through the communications network 23.

[0006] Methods of measuring the echo return loss typically measure a signal at node 9, where the signal power from the far-end would normally exist. A measurement of the signal power, x, at node 9 is made. Additionally, the power level of the composite signal y, comprised of the coupled echo signal and any signal s generated by the near-end telephone 2, is measured at node 7. The measurement can be made when little-to-no signal is being generated at near end telephone 2. Assuming the signal power of any signal generated by the near-end telephone is very small in comparison to the coupled echo signal power, the ratio of the measured test signal power x to the measured power level y provides an estimate of the echo return loss (ERL) for the near-end 8. The magnitude of echo return loss is usually measured as a difference in dB between signal x and signal y. As described in the co-pending application Ser. No. 10/029,669, incorporated herein in its entirety by reference, echo return loss may be measured dynamically during the course of a telephone conversation.

[0007] Echo is an important factor in communications which include a hybrid between a four wire communication network 23 and the end terminals 24 and 25 as illustrated in FIG. 2. When echo is present, it is preferable to eliminate the echo. To eliminate the echo, the magnitude of the echo must be determined. One way of determining the magnitude of the echo is through echo return loss (ERL) estimation. A high echo return loss means that there is very little echo because most of the energy from the far end has been lost when the near end signal combined with echo is measured.

[0008] A typical echo canceller, as illustrated in FIG. 1, includes an adaptive finite impulse filter FIR 5. Under the control of an adaptation algorithm, FIR filter 5 models the impulse response of the echo path. A non-linear processor (NLP) 6 can be used to remove residual echo that may remain after linear processing of the input signal. The echo canceller may also typically include a double talk detector 11. Double talk occurs when both far end and near end speech are present at the same time. A double talk detector 11 can also be used to control and inhibit the adaptation process of the FIR 5 and/or the NLP 6 when double talk is present and it may be undesirable to cancel or suppress echo because double talk will be suppressed.

[0009] In the echo canceller, the signal y is the perceived near end signal. Signal y is a combination of the actual near end signal s and the echo from the far end signal x which comes through hybrid 3. The output signal d is the signal y less the echo estimate e generated by the adaptive filter 5. The adaptive filter 5 is programmed to generate an output signal e that is as close to the echo as possible so that the echo is largely cancelled by the echo estimate e and the difference signal d closely resembles the generated near end signal s. The NLP 6 controls the amount of signal d that is transmitted to the far end. When there is no near end signal s, or a large echo over riding near end signal is present, NLP 6 can provide comfort noise to the far end instead of near end signal so as to prevent any possible uncancelled echo from being transmitted. When a valid s exists, NLP opens so as to let the far end hear the signal. False detection of a lack of near end signal s can cause clipping of speech and failure to detect echo can result in echo leak through the NLP. The NLP as an on/off switch can result in abrupt audible changes which are undesirable in speech communications.

SUMMARY OF THE INVENTION

[0010] The present invention teaches the attenuation of state change of the NLP so that speech is faded instead of clipped and echo leak is minimized and does not appear as a sudden signal change. The present invention teaches the implementation of a state machine to guide the transition into and out of NLP active state and the inclusion of transition states for entry and exit of NLP active and inactive states.

[0011] FIG. 3 illustrates a situation where there is no near end signal s and only a far end echo. A portion of an exemplary typical echo signal 30 without near end speech, is illustrated. If the echo has been properly detected and the NLP 6 is functioning properly, the NLP will be Active and no echo will leak through to the far end. However, if the NLP is in error for a portion of the duration of the echo signal, as illustrated by the state change at 32 in NLP state line 31, then echo leak, as illustrated by echo leak line 33, will result. Because of the on/off nature of the prior art NLP operation, the entire echo signal will leak through causing a noticeable audible disruption. The present invention NLP, as detailed below, attenuates leak caused by NLP error resulting in the minimized echo leak illustrated as line 34.

[0012] FIG. 4 illustrates the situation where there is a near end signal s. A portion of an exemplary typical near end signal s, is illustrated. This signal s will typically represent speech from the near end speaker. When the NLP 6 is functioning properly, the NLP will be Inactive and the entire speech signal will pass through to the far end. However, if the NLP is in error for a portion of the duration of the echo signal, as illustrated by the state change at 42 in NLP state line 41, then abrupt clipping of the speech signal 43 will result. Because of the nature of the prior art NLP operation, the clipping of speech is sharp, causing a noticeable audible disruption. The present invention NLP, as detailed below, attenuates the clipping caused by NLP error resulting in smooth loss of the speech signal illustrated as line 44.

[0013] Proper identification of double talk also provides for smooth transitions. If there is double talk or an idle in the near end signal s, the NLP will be inactive, otherwise use of active NLP will be evaluated depending on power measurements. The state machine, detailed below, will pass some decisions, for example double talk true or false; NLP=active or inactive. The prior art used NLP as on/off switch. If NLP was active, the near end signal was replaced with comfort noise otherwise the near end signal was left unmodified. The present invention does not implement NLP as a hard switch but instead as a smooth transition. The NLP logic an still make a mistake and be inactive or active at the wrong time. However, the logic state described below smooths the transition by attenuating the signal when there is a NLP transition.

[0014] The logic of the state machine of the present invention attenuates the echo instead of sharp removal. Instead of removing echo and all near end noise, the signal level is attenuated down to a level where echo is no longer audible and near end noise and/or signal is preserved and passed to the far end, as illustrated in FIG. 5, which keeps the noise section similar to non-NLP state.

[0015] The amount of the calculated attenuation 51 depends on far end power x and combined loss. Echo reduced to −80 dBm is essentially silence level. The calculated attenuation will first apply and then an additional amount of user selectable attenuation 53 can be added towards the target of −80 dBm. The target level can change depending on noise level estimated for near end noise level. If the near end noise level is high the echo signal does not need to be attenuated down to −80 dBm. The echo signal can just be attenuated to a level below the noise level so that echo is just as bad as noise and will not cause problems because the echo will not be distinguishable by the listener. The present invention also provides a user adjustable level 53 to provide a user adjustment. The user adjustment will not allow for amplification of the signal above the original power level x nor will it need to allow for adjustment below −80 dBm.

[0016] The user adjustment is provided to allow some control over NLP to provide a selection between non-aggressive NLP to allow the leakage of some minor echo while always having near end noise come through. Other users do not care about near end noise they just want to eliminate all echo so these users desire a more aggressive NLP. The user can achieve zero echo level by setting 53 to −80 dBm. Less attenuation will bring near end noise audible.

[0017] Double talk is treated differently by the state machine. If double talk is true than NLP is false. If double talk is false then could still be some weak double talk detection and NLP may be true or false depending on parameters of the state machine. NLP is immediately set inactive if double talk is detected.

[0018] The slope of the transition ramp depends upon the length of time in the double talk state. The logic is that the longer the time in the double talk state, the farther away the double talk is and therefore the steeper the slope of the ramp for quicker entry into NLP active state. If double talk detection is recent then double talk may still exist so ramp to NLP is slower, less steep.

BRIEF DESCRIPTION OF THE DRAWINGS

[0019] Preferred embodiments of the invention are discussed hereinafter in reference to the drawings, in which:

[0020] FIG. 1 is a simplified block diagram of one end of a communication link.

[0021] FIG. 2 is a simplified block diagram of a representative communication link between two two wire telephones each attached to the communication link via a hybrid.

[0022] FIG. 3 is a signal diagram of an echo signal and corresponding NLP activation.

[0023] FIG. 4 is a signal diagram of a voice signal and corresponding NLP activation.

[0024] FIG. 5 illustrates power attenuation.

[0025] FIG. 6 is a state diagram illustrating the NLP and double talk logic states.

DETAILED DESCRIPTION OF THE INVENTION

[0026] The activity of the NLP is regulated by the state machine illustrated in FIG. 6. If the state machine is in NLP inactive state 61, the state machine must transition through NLP entering state 62 prior to NLP active state 63. Similarly, the state machine must pass through NLP exiting state prior 64 to returning to NLP inactive state 61. Unless double talk is detected. If double talk is detected, the state machine will first pass through double talk state 65. When in the NLP entering state 62, the NLP will idle in this state until the NLP minimum mixing ratio has been reached. This allows the engagement of the NLP to be attenuated instead of abrupt, to avoid sudden voice clipping, such as illustrated at 42 in FIG. 4.

[0027] When in the NLP existing state, the NLP will idle until the NLP maximum mixing ratio has been reached, allowing the disengagement of NLP to be attenuated instead of abrupt. This reduces echo leak as illustrated in FIG. 3.

[0028] In the NLP active state, the NLP prevents echo as in the prior art. The minimum mixing ratio and mixing at minimum ratio are updated. If double talk is detected, the state machine transitions to NLP double talk out state, and a counter is initiated to monitor the time in double talk state. If no double talk is detected, the state machine transitions to NLP going out state

[0029] In NLP entering state 62, the NLP is engaged and the mixing ratio is gradually decreased. By decreasing the mixing ratio of noise to voice, the suppression of echo/noise is gradual or attenuated, not sharp as in the prior art. The, mixing ratio continues to be decreased while in the NLP entering state until a condition exists to shift the state machine into a different state. Then, if double talk is detected, the state machine shifts to Double Talk state 65 and the double talk counter is initialized. If no double talk is detected and the NLP is not engaged, the state machine instead shifts to NLP exiting state 64 and the mixing ratio is gradually increased to allow more echo/noise into the voice stream. If however, the NLP minimum mixing ratio is reached while the state machine is still in NLP entering state 62, the state machine switches to NLP active state 63.

[0030] When in NLP exiting state 64, the NLP is engaged and the mixing ratio of noise to voice is increased. This allows for an attenuated transition back from an NLP active echo suppression to a non-active NLP pass through. If double talk is detected, the state machine shifts to the double talk state 65 and the double talk counter is initialized. If a new condition to engage NLP is received, then the state machine shifts back to NLP entering state 62. When the NLP maximum mixing ratio is reached and the state machine is still in NLP exiting sates 64, the state machine will shift to NLP inactive state.

[0031] When the state machine is shifted into the double talk state, the NLP is disengaged and the double talk hangover count is decreased. The double talk hangover count sets the amount of time that the state machine should remain in double talk state 65 after receipt of a double talk signal. As long as double talk is detected, the state machine will remain in the double talk state 65 and the double talk counter is reset. When the state machine is in double talk state and no double talk is detected, the double talk hangover counter counts down during the elapsing time period. The double talk counter indicates the elapsed time since the last double talk detection. If an NLP engage event is detected, the state machine shifts to NLP entering state and recalculates a ramping rate for the decrease of the mixing ratio based upon the relationship between the current double talk hangover and the initial double talk hangover. The ramping rate is used in NLP entering state 62 and NLP exiting state 64 to determine the speed of decrease or increase in mixing ratio. If the double talk counter reaches zero the state machine will exit from the double talk state 65 and shift to NLP inactive state.

[0032] In the NLP inactive state 61, the NLP is disengaged. if double talk is detected, the state machine shifts to double talk state 65 and an initial double talk hangover countdown counter is set. If an NLP event, such as echo detection, occurs, the state machine shifts to NLP entering state and begins decreasing the mixing ratio at a ramping rate set to Ri=Ri0. If no double talk, NLP or echo is detected, the state machine remains in NLP inactive state.

[0033] Because many varying and different embodiments may be made within the scope of the inventive concept herein taught, and because many modifications may be made in the embodiments herein detailed in accordance with the descriptive requirements of the law, it is to be understood that the details herein are to be interpreted as illustrative and not in a limiting sense.

Claims

1. A method of echo cancellation in a communication link, comprising the steps of: ramping the change in the mixing ratio of echo to voice during a transition into echo suppression to attenuate the suppression of echo.

2. The method of claim 1, further comprising: ramping the change in the mixing ratio of echo to voice during a transition out of echo suppression to attenuate the resumption of voice signal.

3. An state machine for enhanced NLP control, comprising: an NLP active state; an NLP inactive state; an NLP entering state interposed between said NLP inactive state and said NLP active state, wherein the activation of said NLP is gradual during said NLP entering state, attenuating the reduction of echo; an NLP exiting state interposed between said NLP active state and said NLP inactive state, wherein the deactivation of said NLP is gradual during said NLP exiting state, attenuating the resumption of signal previously depressed as echo.

4. The state machine of claim 3, further including: a double talk state wherein said NLP is inactive and a timer is initialized upon the detection of double talk; an NLP engage transition rate indicator for determining the rate of engagement of said NLP upon resumption of echo cancellation after loss of detection of double talk.

5. The state machine of claim 4, wherein the value of said transition rate indicator is dependant upon the time lapsed since entry into te double talk state as reflected by said timer.