Echo Cancellation for VoIP
Here are some techniques for dealing with echo in applications that bypass the conventional long distance telephone network, such as voice over IP, voice over Ethernet, and fixed wireless local loop.
By John C. Gammel
In a conventional voice circuit on an analog line card serving a
two-wire interface, a portion of the signal sent toward the subscriber’s telephone is subtracted from the signal on the two-wire telephone loop to obtain the signal sent from the telephone towards the line card. The fraction of the signal that should be subtracted depends on the divider formed by the actual loop impedance and the termination impedance of the line card.
Ideally, the subtracted signal would match the signal being sent to the telephone on the two-wire interface exactly, so that only the
signal coming from the telephone would be detected. If the subtracted signal does not match the signal being sent to the telephone on the two-wire loop exactly, then some of the signal being sent to the telephone appears as signal from the telephone. The term hybrid balance refers to the ratio (expressed in dB) of the signal being sent to the telephone to the apparent (reflected) signal due to imperfect subtraction. If the hybrid balance is not perfect, then a fraction of the signal being sent is reflected
back to the sender. If a delay in the network occurs, then the sender hears an echo, which can be objectionable.
The impedance of the two-wire loop depends on the length of the loop, the type of cable used, whether or not loading coils are used, the impedance of the customer premises equipment (such as telephone sets and modems) and the number of telephones in use (off hook). All these variables make it impossible to know the impedance characteristic exactly. If a constant fraction of the signal
being sent is subtracted from the signal on the two-wire loop, the hybrid balance will vary depending on the actual loop impedance, resulting in poor hybrid balance and echo for some loops.
The degree to which echo is objectionable depends on echo loudness and total delay. The total delay is associated with the process of digitally encoding the voice, delay in digital processing on both ends (such as time slot assignment or packetizing, in the case of data-oriented circuits), and two times any delay
in the long distance circuitry. Delay from transmission is in the range of 30 ms for transcontinental domestic calls, from 50 to 100 ms for international calls, and 600 ms for satellite calls. This delay is what affects the customer’s perception of echo. If the delay is small (less than 10 or 20 ms), the customer hears nothing or a reverberant sounding side-tone at most. Larger delays lead to a subjective annoyance perceived as echo. The larger the delay, the less masking there is by the direct speech
and the more annoying the echo is.
Figure 1
from ITU-T G.131 (control of talker echo) shows the variation in acceptable perceived echo loudness versus total delay, where:
1
- TELR = talker echo loudness rating = SLR + RLR + R + T + Lr
- SLR = speaker loudness rating = 7dB nom, 2dB min for most telephones
- RLR = receiver loudness rating = 3 dB nom, 1 dB min for most telephones
- R = receive loss
- T =
transmit loss
- R+T = 6 dB is introduced in most calls in the US for echo control
- Lr = return loss or hybrid balance = 14 dB nom, 8 dB min for line cards that subtract a constant fraction of the signal being sent due to the variation in telephone and loop impedance.
Adding all of these variations gives a 17-dB minimum TELR and 30-dB nominal. The worst case TELR of 17 dB would be objectionable for most subscribers at 8 ms of delay, and objectionable to some subscribers at less
than 5-ms delay. The average TELR of 30 dB would be objectionable to most subscribers at 35 ms of delay, and could be objectionable to some subscribers at 18 ms of delay.
As previously noted, both echo annoyance and echo loudness increase with delay. One way of dealing with echo is to intentionally introduce equal loss in the TX and RX circuits so the desired signal sees this loss once and the echoed signals see it twice. Hence the signal-to-echo ratio improves by the amount of loss inserted. The
downside, of course, is that the desired signal is made weaker, so this can’t be done too often. Around 1920, the via-net-loss (VNL) plan was devised. It intentionally introduced loss into trunk circuits, and the loss was roughly proportional to length. The idea was to wind up with lots of loss on long circuits where echo would be a problem, and less so on shorter circuits. Long-distance volume levels were lower because of this plan.
The VNL plan was a disaster for digital telephony because you
could not introduce loss without requantizing to µ-law after going through a digital gain. Cascaded requantizations would leave the SNR unacceptable. Instead, all of the incremental VNLs were eliminated and 6 db of loss were added to all trunks coming into the end office from the toll plant. (6 dB is a US number; other countries are slightly different. )
Another technique for reducing echo is the use of loop segregation. In one loop segregation technique, the impedance versus frequency of copper
wire loops is measured during telephone on-hook periods to determine the loop length and type (loaded or unloaded) and choose an appropriate compromise balance. Alternatively, a look-up table can be used when a physical loop is connected to a given line-card port. These techniques are imperfect because the impedance of the telephone on-hook is not the same as the impedance of a telephone off-hook, and because a subscriber can have multiple telephone sets with different off-hook impedance.
Introduction of network loss and loop segregation provides adequate voice quality for regional (inter-LATA) calls. For calls completed out of the LATA, other means of echo control must be used.
Other sources of echo
The dominant source of echo in telephone networks is imperfect hybrid balance. If the loop connected from a line card to a telephone were extremely long, then reflections from the far end of the loop could be thought of as echoes. Generally, two wire loops are short enough so that the
round-trip delay through a loop is negligible. A worst case scenario is a loaded loop where the velocity of propagation is 14,000 meters/second.
2
Even in this case, a 7-mile loop is required for a 1-ms round-trip delay. Generally, the impedance of the two wire loops is lumped with the telephone and the effects of far-end reflections are included as part of the hybrid balance.
Another source of echo is acoustic echo which is echo due to coupling from the speaker to the microphone of a
telephone set. In a conventional telephone set, this coupling path has a large amount of loss (over 20 dB) and a small delay (less than 1 ms). This path can once again be lumped with hybrid balance.
Speakerphones have the potential of much longer delay paths and low loss, which can create echo. Due to the potential of highly time-variant echo paths, long delay, and multiple echo paths, echo control for speakerphones is extremely difficult. Echo control is most commonly achieved by turning off the TX
path when RX is active or vice versa. Echo control for these types of devices is normally done at the device itself.
Adaptive hybrid balance and echo cancellation
It’s possible to correlate the signal sent in one direction to the signal being received from the other direction, and determine the amount of echo. It is also possible to dynamically adjust the tap weights in a digital filter to subtract a delayed portion of the signal being sent from the signal being received, so as to
minimize the echo signal. This technique is known as adaptive hybrid balance, or adaptive echo cancellation.
This technique could be used in two possible ways. One is to eliminate the need for loop segregation. The degree of balance (or echo return loss [ERL]) that can be achieved with loop segregation is 10 to 20 dB, so 20-dB ERL is adequate for this purpose. This would not eliminate the need for echo control on long distance calls, but it would alleviate the need for loop segregation.
In small
line-count systems or systems with remote line cards, it may be impractical or too costly to implement loop segregation. In these situations, adaptive balance in the line card codec provides value.
In the case of long delays (for example, more than 20 ms), better echo performance is required (over 35 dB of ERL). Some examples of this are extremely long distance calls (over 1,000 km), VoIP, voice over Ethernet, and wireless local calls. In these situations, fixed hybrid balance with loop segregation
generally provides unacceptable performance and echo cancellation at the line-card level or network echo cancellation is required.
Figure 2
illustrates the echo path and some possible places (noted by the large arrows) where echo cancellation can be done in the VoIP environment.
In the conventional telephone network, digital echo cancellation devices are customarily placed within 500 miles of the echo source. This allows reasonable echo control
without excessive tail length (the number of taps in the digital filter that subtracts a delayed portion of the signal being sent from the signal being received, minimizing the echo signal). These types of devices can control echo of 32 to 64 ms (256 to 512 taps at 8,000 samples per second). The limitations of echo cancellation in the network are briefly discussed later.
In this situation, it can’t be guaranteed that an echo control device will be in the call path if the call is placed locally (<
500 miles), and the point of entry into the network does not have echo cancellation. For this reason, placing echo control at point B to control echo from subscriber B is recommended. This will prevent the echo from subscriber B being objectionable to subscriber A because of the delays associated with VoIP. It’s possible that the call will be a long distance call and route through the network to a second echo cancellation device at point C. This is not a problem because these types of echo
cancellation devices operate nicely in cascade.
While echo control from subscriber A may be put at the head end after vocoding (if required) and packetizing, putting the echo control at point A (at the line card) will result in better performance. Echo control is required for all calls (even within the network). More echo-control devices will be required when providing this on a per-line basis (as opposed to the head end) because for large systems, it is acceptable to design for only a fraction of subscribers
on the IP network to be off hook (talking) at any given time. On the other hand, when putting the echo control at the head end, this device must deal with voice compression (vocoding or companding) and with long delay paths, so it will be more expensive. Based on performance and cost effectiveness, it is recommended that the echo control be placed on the line card at point A.
Filter requirements and implementation
When doing adaptive hybrid balance (as an alternate to loop
segregation), balance on the order of 20 dB is required. At the line card level, this requires approximately an 8-tap FIR and single-tap IIR filter.
When doing echo control for long delay paths at the line-card level, more than 35 dB of balance is required. This can be achieved with a 62-tap FIR and 2-tap IIR filter.
6
The larger number of taps does not deal with long echo tail length (which is only about 1 ms if the echo control is at the line card); it deals with the lower audio frequencies. The
IIR taps are required for good performance, but do complicate the algorithms for convergence (IIR taps can cause local minimums in echo versus tap strength).
If echo control is done in the network, a sufficient number of taps must be provided to deal with the round-trip delay from the echo control device to the line card and back. As previously noted, this is normally 256 to 512 taps for an echo path of 32 to 64 ms. A shorter filter (such as 64 taps) can be provided, and only control echo in a
certain window.
Advantages of echo cancellation on voice circuits
Network echo cancellation, when present, is imperfect for several reasons:
- The quantization of the signals with µ-law or A-law companding places a fundamental limit on the accuracy of subtracting the echo. This limit is signal level-dependent but is on the order of 30 dB of ERL enhancement (ERLE). ERLE is the ratio of ERL with and without echo cancellation.
- When about 30 dB of ERLE
offered by network echo cancellation is coupled with a typical 10-dB natural ERL provided by hybrid balance, the echo is reduced to a level that is just slightly hotter than what is required to keep users happy (not complaining about echo) on long distance circuits with echo delays greater than 30 ms. Hence, there is the need to resort to a residual echo control device (REC) in conventional network echo cancellation. The REC attempts to decide when the TX-side customer is talking, and the RX-side is quiet.
Whenever that occurs, it opens the TX path to prevent any echo from coming back. This introduces a whole host of impairments. One is the break up of honest TX-side speech because of the inevitable delay in detecting its beginning and getting the TX path closed again. Another is noise modulation, where the customer at the end of the TX path hears any background noise accompanying the customer at the other end of the TX path whenever he or she talks, but then does not hear it when he or she is quiet. Noise
matching (NM) is used to combat this, but it is also imperfect, and brings along another set of impairments.
The advantages of putting echo cancellation at the line-card level are:
- Cancellation can be done before companding so the achievable cancellation is much more precise.
- At least some of the 6-dB loss normally put into all calls can be eliminated if the line cards on both ends of the call had echo cancellation.
- The REC and NM impairments associated with network
echo cancellation can be avoided.
- Circuits that do not route through network echo cancellation (such as local calls placed through a remote line card) will have at least as good a quality as calls that do go through network echo cancellation, if the line cards on both ends of the call echo cancellation.
- Because echo annoyance is dependent on total circuit delay and because no one component of a circuit (such as an end office) knows what else will be cascaded into the connection (which was
the case before common channel signaling), a scheme exists to allocate delay to various network components. Under this scheme, the end office gets a 2-ms delay. This low allocation is due to the fact that when the network was set up, all end offices were just metallic analog switches (which did not introduce delay), so nobody complained. This reduces remoting distances for CO. Providing echo cancellation at the line level would allow virtually unlimited remoting distance if it were provided on both ends of
a call.
- When calls are contained within a network that does not have network echo cancellation available (such as wireless local loop, IP, or ATM telephony), the cost and complexity of providing echo cancellation in the network can be avoided by having echo cancellation at the line card on both ends of the call.
Fax and modem calls
For voice circuits with long delays, an echo as low as -35 dB is objectionable. Because modems try to cancel the echo path anyway, the main
problem with having echo in a data call is that the echo is quantized and contributes additional quantization noise which cannot be canceled. Echo starts degrading modem performance at a level of about -20 dB.
Just as loop segregation provides an acceptable level of balance for voice calls with network echo cancellation, it also provides acceptable balance for data calls with adaptation disabled in CO situations. The improvement that would be achieved by doing adaptive hybrid balance at a line-card
level would not provide a noticeable increase in average modem connect speed in situations where loop segregation and careful network engineering already reduces the echo path to -20 dB.
Adaptation and/or echo cancellation must be disabled during data calls because modems also try to detect and suppress echoes. Once modem training occurs, there is an assumption that echo paths will remain constant. It’s possible to try and have an echo cancellation device train during a modem call setup and
freeze (or train) on a previous voice call, or use the same cancellation for a data call. Both of these techniques are imperfect, so echo cancellation does not work well on data calls.
In the case of digital loop carrier systems that have universal adapters where analog lines are converted back to digital for transport to a remote collection of line cards, modem connect speeds are fundamentally reduced due to multiple analog-to-digital and digital-to-analog conversions. In these situations, there has
been speculation that the use of adaptation at the line-card level could compensate for this loss in performance. Actual experiments have shown that adaptation at the line card provides almost no measurable improvement in modem connect speeds when compared to loop segregation.
As in the case of voice calls, in small line-count systems or systems with remote line cards (such as DLC or Pairgain), it may be impractical or costly to implement loop segregation. In these situations, adaptive balance in the
line-card codec will provide higher modem connect speeds.
John C. Gammel is technical manager for line card integrated circuit design at Lucent Technologies Microelectronics Group. He can be reached at
jgammel@lucent.com
.
| References
|
- This figure has been reproduced with the permission of the International Telecommunications Union
(ITU), the original copyright holder. The sole responsibility for selection of this figure lies with Lucent Technologies, Inc. alone and can in no way be attributed to the ITU. The complete volume of ITU G.131 can be obtained from: International Telecommunications Union Sales and Marketing Service Place des Nations – CH-1211 GENEVA 20 (Switzerland)
Telephone: 41.22.730.61.41 (English), 41.22.730.61.42 (French), 41.22.730.61.43 (Spanish), Telex: 421.000.u1t.ch Fax: 41.22.730.51.94.
X.400 : S = Sales;
A= 400net; P= itu; C=ch. E-mail:
sales@itu.int
http://www.itu.int/publications
- Bell Telephone Laboratories, Inc., Transmission Systems for Communications, Western Electric Company, Winston-Salem, N.C., pp. 34, 75, 1970.
- “Voice Over Packet White Paper,” Edward B. Morgan Telogy Networks, Inc. can be found on the web at
http://www.telogy.com
- ITU-T G.131, “Control of Talker Echo.”
- ITU-T G.165, “Echo Cancellers.”
- Lucent Technologies, Microelectronics Group T8533/34 Quad Programmable Line Card Signal Processor Data sheet, June 1999.
|
Return to the
Table of Contents
Puzzled by a network processing design issue?