Echo in the public network: What is the worst case?

Echo cancellers are widely used today in the public switched telephone network (PSTN) to remove echo, which otherwise would cause a catastrophic quality problem for voice calls. The human ear is increasingly sensitive to echo as the delay of the echo is increased, therefore echo cancellers traditionally were deployed on circuits with more than a certain amount of round-trip delay, for example 20 ms (milliseconds). In addition, an echo transfer function itself has a dispersion time: a decaying function in amplitude with time. This is commonly called an "echo tail." Echo canceller devices are usually designed to deal with a maximum amount of echo tail length, and thus have an "echo tail capacity".

Two questions are often asked: "What is the worst case echo tail length seen in the PSTN?" and "What is the worst-case dispersion time of an echo path?" These questions are important to network providers, echo canceller designers and manufacturers.

The key source of echo in the PSTN is the 2-wire/4-wire (2W/4W) hybrid circuit used on subscriber access line cards. Network architects often assume that the echo dispersion introduced by these circuits is fairly limited, rarely exceeding about 12 ms. As a result, many of today's echo cancellation chips, particularly those based on commercial digital signal processors (DSPs), have only enough computational resources to handle limited dispersions.

However, many common situations--such as multi-party conference calls and other PBX call configurations--generate longer echo tails. In these instances, low-cost DSP-based echo cancellers can fail to do the job, and voice quality will deteriorate.

The only way to truly guarantee voice quality in the variable global PSTN environment is by using echo cancellers equipped to handle a full-length echo tail. These devices have fully populated adaptive filter coefficients and are capable of canceling echo tails with long dispersions. Called "full-tail" or "full-band" echo cancellation chips, these devices do not exploit opportunities for complexity reduction, i.e. limited dispersion length capacity.

Cutting corners: Windowed adaptive filtering

Most echo cancellers with reduced computational capabilities use a technique called windowed adaptive filtering, basically the adaptation of reduced regions of a long FIR (finite impulse response) filter to reduce computational load. To understand why some designers assume that windowed adaptive filtering is adequate, let's review the operation of a basic echo canceller, illustrated in Figure 1.

 

Figure 1: Typical echo canceller and hybrid circuit

The source of the echo is the hybrid circuit in the 2W/4W interface. The impulse response of a typical hybrid has a dispersion time of only 4 to 5 ms. The echo canceller uses a FIR filter to synthesize a replica of the echo signal and subtract it from the "send" input. A convergence algorithm is applied to adapt the filter coefficients to optimally cancel echo. Ideally, the tap weights of the echo canceller match the impulse response of the true echo path. The main assumption behind windowed adaptive filtering is that most echo paths require very few of the filter coefficients to be non-zero.

The PSTN reality is somewhat different. Tandem connections of hybrid circuits in 2W/4W circuits often form an impulse response with a longer echo tail--a convolution of multiple echo transfer functions. Additionally, network delays, the duration of which are largely unknown before each call, add to the length of the echo tail that must be cancelled. Large delays between hybrid circuits can cause multiple reflectors that are distinctly separate, although this is rare. Multi-party conference calls, and calls transferred through multiple PBXs without automatic re-routing functions, can also easily generate longer echo tails.

All conditions considered, the worst-case echo tail capacity needed for a network echo canceller is approximately 128 ms; however, this is network dependent, and many networks do not require more than 64 ms.

 

Figure 2: Example of an echo path response
and corresponding regions of active filter coefficients

Figure 2 shows an echo path response with two distinct reflectors, and the corresponding regions where active tap weights have been identified. Clearly, a large percentage of the coefficients should be zero-valued.

There is a temptation to make a design trade-off, where for any given channel, only a fixed percentage of the maximum number of possible tap weights is active. For example, a seemingly reasonable choice might be 20%, or about 200 of 1,024 coefficients. The simplified echo canceller design would then be equipped with the means to identify the locations of the required active tap values, followed by the adaptation of these coefficients as quickly as possible.

However, design simplifications of this nature can result in un-cancelled echo becoming audible to subscribers, and thus may not be a good choice.

Echo challenge: The multi-party conference call

Figure 3 illustrates one scenario--a six-party PBX conference bridge--where the windowed adaptive filtering technique would prove inadequate. 

Figure 3:  Network connections in a six-party conference call

The PBX on the right side of Figure 3 serves as the conference bridge. Party #6 is located on an internal line of that PBX, and the other five parties are linked in through the PSTN. Parties #1, #2, and #4 are connected through PBXs at other sites, party #3 is on a subscriber loop, and party #5, calling from a cell phone, is connected through a cellular network.

An echo canceller must be installed at the mobile switching center (MSC) to prevent PSTN echo from entering the mobile network. The PBXs in this example are traditional TDM switches, not IP PBXs, and do not have echo cancellers.

The PBX providing the conference bridge digitally sums together the signals from the six parties, subtracting each party's own signal from the return path. This is typical. All phones in the conference except the cell phone are analog sets connected to a 2-wire analog line circuit. The PBX serving party #1 is also connected to the PSTN via an analog trunk, also typical. The echo return loss (ERL) of the 2W/4W circuits are representative of those seen in many PBXs. All PSTN circuits are assumed to be 4-wire digital circuits, with losses of about 6 dB, and end-to-end delays of between 2 ms and 18 ms.

 

Figure 4: Echo response from six-party conference call,
measured at MSC interface

Figure 4 shows the resulting echo tail produced by the conferenced parties, measured at the interface to the MSC. The echo response shows an initial flat delay of about 10 ms, due to delay between the MSC and the PBX with the conference bridge. The peaks after that reflect several delayed and combined echo responses from the different party's hybrids. The response is thus highly dependent upon the ERL values of the hybrid circuits, as well as on the delay through the PSTN. An echo canceller that is not equipped to handle this many active filter coefficients would be unable to properly cancel this echo response.

Full-tail capacity echo cancellers deliver better quality

In conclusion, common, everyday call configurations - such as PBX-bridged audio conferences--can generate echo path responses with high numbers of active filter coefficients. Moreover, if there are large network delays for one or more parties through the PSTN, echo path responses can span up to 128 ms.

Next-generation networks (packet-switched networks, and wireless networks) that are coupled to the PSTN via Gateways and other equipment, add significant extra delays to phone calls. This is due to packetization and compression functions, forcing all connections to need echo cancellation. Echo canceller performance has thus become more important than ever, as echo removal is critical before large packet network delays are added to the round trip delay of a voice call. Echo cancellation is a key function performed by Gateway equipment, which removes echo in the circuit-switched network before transporting over the packet network. The increased reliance upon Gateway echo canceller performance makes the question of worst-case PSTN echo more significant than ever.

The most commonly referenced International standard for echo cancellation is ITU-T G.168. This standard provides recommendations for echo canceller performance for application of network echo cancellers to network equipment such as MSCs and Access or Trunking Gateways.

However, ITU G.168 contains no guideline for the requirement of Echo Tail Capacity. Individual service providers usually provision this feature for the particular needs for their network. ITU-T G.168 also does not offer any particular recommendation for the ability of any network echo canceller to deal with echo paths that have long dispersion, such as those presented in this article. G.168 does recommend that up to 3 separate echo tail reflections should be provided, and it details a set of 8 hybrid path models to test against. G.168-2002 Appendix II.5 concludes that the worst-case dispersion time of a hybrid will usually be less than 12ms. In G.168-2002 appendix III, there is a discussion of multiple-tail circuits due to bridged PBX calls, but the worst case echo path estimates are not provided for these configurations.

To ensure that next-generation networks provide quality of service comparable to today's PSTN, and that subscribers do not hear annoying echoes when making calls, network equipment manufacturers should use echo cancellers with sufficient echo tail capacity, and these cancellers should provide "full-tail" dispersion capability.

Gordon J. Reesor is System Architect for Zarlink Semiconductor.

Visit Zarlink Semiconductor online.