Designers who worked on the plain old telephone systems (POTS) never dreamed of the challenges facing those working on next generation telecommunications equipment. Today's engineers live in the digital world where equipment designs are required to codify, compress, cancel echo, control jitter and loss, packetize, switch, route, and bill, and do it all more quickly and efficiently than their competitors. It is easy to lose sight of the need to provide a quality voice connection when the principal design goal is to maximize the volume of data the device can handle.
Given these concerns, the length of the circuit tail delay in an echo-cancellation algorithm may be low on the design requirements list. But failure to appreciate tail delay will adversely affect voice quality. Voice quality assumes satisfied customers and therefore positively affects revenues.
While data may have eclipsed voice in volume, most networks carrying both generate more money from voice traffic. As voice users now have a choice among many service providers, carriers must pay close attention to the voice quality in their networks.
What does it represent?
What does tail represent in the world of digital telephony? Simply put, tail is the distance between an echo cancellation device (or standalone embedded cancellation algorithm) and a signal reflector measured in milliseconds. In the PSTN, this reflector is commonly a digital-to-analog converter (DAC), called a hybrid, which, because the conversion is not perfect, presents an impedance mismatch to the signal and reflects some of the signal back to the source in a linear fashion.
It is appropriate to examine the classic case of where and why echo cancellers are placed in the PSTN, as the principals described are directly transferable to next generation network topologies.
In the PSTN, an inter-exchange carrier (IXC) typically deploys an echo cancellers at both ends of their network with the echo cancellers on the near end protecting the far end caller from echo and visa versa. The reason this scenario occurs is because the near end hybrid generates an echo of the far end caller and reflects it back to the far end, and the near end caller's voice is reflected by the far end hybrid in a similar fashion.
Each of these reflections is commonly referred to as network echo and generally have only linear waveforms. Echo control, therefore, needs to be at each end of the call, and the echo canceller installed at the far end carrier actually protects the other carrier's customer. That is why when using a US-based carrier to call a country with a poor voice network, an echo is sometimes heard.
The far-end company may not have installed an echo canceller (or deployed one with an inadequate tail length). Because the cancellers are installed on the US side protect their customers, the far-end company may be unaware of the US customers experience.
Proper installation in traditional circuit-switched calls require the EC be positioned between a class 4 switch (toll) and the long haul side of the network, with the tail pointed toward the drop or terminal side -- usually an analog phone or PBX. Usually, the reflector on the hybrid occurs between the local class 5 switch and the local service loop.
Round-trip latency
When dealing with echo cancellation, it is important to understand round-trip latency. Latency is the total length of time (round-trip delay) it takes the far end caller's signal to reach the near end hybrid and back. Latency and echo control are coupled (Figure 1), therefore without proper echo control, the longer the latency or path delay, the worse the perceived echo and voice quality.
Click here for Figure 1
Figure 1: The impact of delay and echo on voice quality.
When long distance calling was less common, and local networks less complex, total end-to-end latency was shorter. Provisioning for 32-ms maximum tail delays was considered adequate, and provisioning for 64 ms insured an echo free call.
A certain measure of this expanding local loop is the fact that today, 64 ms is considered the minimum and 96 ms is typical with the new guarantee level set at 128 ms. Yet, as networks increase in complexity, even this level of protection is being exceeded.
Why the long tail?
The local loop is not really local anymore - it is getting longer and more complex. In the past, a single IXC may have performed all of the echo control deep in the network at their class 4 switch points, before passing it off through the local loop where the hybrids occur. Data traffic moved over dedicated lines at relatively low speeds and in low volumes.
Seldom did data and voice mix except on low speed dial-up modem or fax calls. Echo control may have been installed every 1000 km within the long haul network to account for path delay and unanticipated terminations. Voice quality, especially in the US, became a standard.
Currently, there are over 50 significant IXCs in the US. The local exchange carriers (LECs) have merged, creating conglomerates which span the continent. Data-only networks are used for back-hauling voice by all kinds of voice carriers (IXC, CLEC, VPN and wireless), while the core of the network has become increasingly complex.
The local cloud is a complex matrix of interconnections, protocols, hand-offs, and terminations(Figure 2). Consequently, both processing and propagation times are increasing. Accurately predicting the tail requirement on all calls has become difficult to impossible. For network equipment designers, it is better to be safe and choose to deploy the longest adaptive tail available on the echo cancellation device.
Click here for Figure 2
Figure 2: Universal echo canceller deployment topology.
To ensure the highest quality, LECs may consider deploying echo cancellers in thir network due to next-generation topologies. Additionally, calls terminated in the LEC, but originated from another network must have echo control applied at the point of interface with the LEC by the alternative carrier, and that EC must have its tail pointed toward the LEC. If not, the alternative carrier's customers receive poor call quality.
Dynamic reflectors
Designers concerned about echo cancellation and voice quality will face a new phenomenon, called the dynamic reflector, when building systems to operate within modern network designs. However, telecommunications is clearly migrating to a common transport fabric -- IP.
In this new environment, call routing is almost always dynamic, as each packet takes a different path. Reflectors along a path create momentary echoes and then disappear. Since echo control is usually a shared resource for many circuits or paths, the solution to this challenge is an echo canceller capable of dynamically locating and canceling echoes with extended amounts of delay to guarantee the highest levels of voice quality.
To accomplish superior voice quality, an echo canceller must have several characteristics. It must have enough memory for each DS0 channel to cancel across 128 ms of path delay; converge within 50 ms; provide a wide enough double talk range that speech clipping is not noticeable in a normal conversation; and be an echo canceller as opposed to an echo suppressor.
In addition, attention must be paid to real world conditions. Failure to test designs under real world, dynamic conditions can result in additional problems such as echo control drop out, non-linear echoes caused by compression, acoustic feedback, or network echo.
Non-linear echoes
Non-linear echoes (variance in time) have multiple causes. The most common are acoustic echoes.
There are two main classifications of acoustic echoes. The first type of acoustic echo is generated by reception of multi-path audio information from a speaker placing a call in a highly reflective environment. The second type of acoustic echo results when supersensitive microphones detect unwanted feedback from earpiece speakers.
Both kinds of acoustic echo are possible when using small wireless handsets. With these non-linear echoes, the causes are usually poor isolation between the speaker and the earpiece in the handset.
An additional form of non-linear echo caused by compression schemes has emerged in these next-generation networks. This new form of non-linear echo is generated when the near-end decompression algorithm does not perfectly match the compression sequence on the far end. This creates a non-linear echo, which is reflected back by the analog to digital mismatch. If the path length is long, as it often is in packetized circuits, (Figure 3), these low level echoes become apparent. The echo canceller at the far end must then be a sophisticated device capable of handling all types of non-linear echoes.
Click here for Figure 3
Figure 3: Delay and Propagation tables.
Echo cancellation advancements
A modern voice enhancement device would consist of both linear and non-linear processors where a model is made of the outgoing audio stream, saved in memory and then compared with the return path. A sophisticated echo cancellation system must not only work very fast, it must have enough memory to hold all the information for a long tail.
There are several methods for implementing echo cancellation, One is a single-chip methodology. A common fault with this approach is to have the echo canceller supported by a relatively slow processor (in terms of instructions per second) that is also tasked with other functions such as packetization, compression, and tone detection.
This scheme works fine until the processor reaches some threshold of call volume vs. design limit (say 90%). At that point, the echo cancellation/suppression algorithm receives a lower priority in the processing hierarchy, or the memory capacity is exceeded. Either way, echoes are not properly cancelled and the far end listener is not pleased with the voice quality.
Tail Length
The importance of an adequate tail delay cannot be overstated. In reviewing actual designs, it appears that too many manufacturers are referencing unenforceable standards instead of proven conditions to make their design choices for tail delay. In a number of cases, designs for core network equipment tail delay capacity are based on a reference to ITU standards (G.114) for maximum transmission path delay of 50 ms for call transmitted over either the continental US or the Atlantic.
The existence of this specification has initiated an unfortunate trend in the equipment manufacturing community that delay and echo do not exist in digital networks. The results are equipment designs insufficiently robust to accommodate actual conditions.
Today's telecommunications networks are a patchwork of voice calls traversing many different carriers' networks. Often calls are transported on TDM converted to VoIP transported over ATM on a SONET backbone and back again. In this scenario, no one is responsible for the total delay; yet it routinely exceeds 50 ms and 70 to 90 ms especially for long-range packet based backbone VPN telephony in North America. Given the unpredictable routes of every voice connection, echo cancellers that handle long-tail circuit delays would be an asset in next generation networks.
Wrap Up
Drawn from extensive experience with multiple customers in multiple networks across the globe, robust echo cancellation devices with long tail capacities will dramatically affect voice call quality in next generation networks. Any impedance mismatch will generate an echo, even ones at far distances. These reflectors are often other devices with improper echo control (impedance mismatch), or non-linear acoustic echo. In addition, craftsmanship errors are suspected in more of these non-hybrid echo generators, but other devices are inherently voice reflectors.
In reality, the all-digital call is often not purely digital from end-to-end, hence it will contain unanticipated reflectors. Moreover the circuit-switched telecom model will persist because analog phones will be somewhere in the network for a long time to come.
As long as a network connects to the PSTN in any way, hybrid echo will occur -- unless echo cancellers are utilized. In next generation networks, tail path provisioning of 64 ms is now considered inadequate to guarantee carrier class voice quality. A new standard of 128 ms is suggested.
Furthermore, there are cases where voice carriers are using data networks to backhaul voice calls, where reflectors in excess of 172 ms are being measured. For example, carriers using the public Internet for voice calls are measuring very long end-to-end circuit delays depending on the terrestrial distance and routing path employed. Extreme round trip delays of more than 780 ms have been measured on public Internet telephony calls. Consequently planning must be done for maximum delay.
Echo cancellation can occur at several points. First, voice gateways that sit between the PSTN and a packet cloud must cancel echo through the LEC-administered PSTN. However, to prevent echo, these packet-to-TDM gateways at either end must be provisioned with at least 96 ms. Echo cancellation is crucial because some VoP carriers who utilize integrated access devices (IADs) at the far end of the packet cloud, are experiencing low level echoes generated by the non-linear effects of compression, poor echo algorithms in the IAD and extremely long return paths.
A second opportunity to fix the long tail delay problem is with the IXCs. IXCs have long provisioned their networks with echo cancellers; these cancellers are pointed at the LEC. Keep in mind, echo cancellers do not work well on compressed signals, including the G.7xx As a result echo cancellers must be deployed on either side of the compressed signal path.
Overall, the variability in tail delays of all voice calls undoubtedly will require designers to build systems with some level of echo cancellation. To achieve the highest level of voice quality in next generation networks, understanding tail length will be of paramount importance.
Author's Note:
The author would like to thanks Tim Hult for his help on this article
About the Author
Conne Skidanenko is the director of engineering at Ditech Communications Corp. He can be reached at cskidanenko@ditechcom.com.
