As broadband operators begin delivering voice capabilities, echo cancellation is becoming a bigger concern. In part 1 of this series, we detailed the basics of echo cancellation for delivering VoP services over broadband. We also explored acoustic echo cancellation (AEC) techniques. In this part, we'll further the discussion by covering the second primary echo technique required in VoP-enabled broadband designs--line echo cancellation (LEC).
LEC is used in a host of VoP designs, including high-density gateways, voice-over-cable and voice-over-DSL (VoDSL) architectures. Each of these systems come with their own LEC requirements.
The main difference in each of these applications revolves around the selection of filter length. If an LEC has to deal with echoes created by the PSTN, designers should consider using at least 32-ms filter length. Sometimes, the network providers may use cheaper equipment that would result in longer echo delays. Delays of up to 40-50ms could be encountered as well for relatively short distances.
Unless linking in with the PSTN, private networks typically require very short filters. The same is true for the CPE products. In most cases a CPE product would be terminated with an analog hybrid that should not generate large delay. If the electrical properties of a hybrid are good, one may get away with just an 8-ms adaptive filter.
In order to use an 8-ms tail in CPE products, a good echo return loss (ERL) is required. Specifically, designers will need to implement hybrids featuring an ERL of 20-dB or better. The problem is that these hybrids are more expensive. Therefore, designers must choose between investing a bit more into high quality hybrids or using 16-ms filter length that would be safe for wider range of hybrids. The longer filter length requires more memory as well as more processing power, which could affect channel density and power consumptions of the final product.
While LEC requirements vary from application to application, some principles cut across all LEC designs. These include:
- Adaptive filter length: 8 to 128 ms. CPE products, including voice over cable and voice over DSL (VoDSL) typically require only 8 ms filter length, although service providers may require 32 ms, 64 ms, or 128 ms. filter lengths. Small gateways typically require 32 ms filter length, although this may increase to 128 ms in some applications. High density gateways typically require a 128 ms filter length, although some applications may use 32 ms and 64 ms tail lengths.
- Full duplex performance.
- Consistent performance over large number of echo paths including digital terminations.
- Consistent performance for a wide range of speech signals.
- Consistent performance for a wide range of background noises.
- LEC performance in accordance with the ITU G.168 specification.
- The convergence on speech should be as fast and as deep as possible without affecting overall performance under wide range of conditions.
When working with LEC's, one of the big questions designers must struggle with is how to achieve fast and deep convergence on speech signals. The simple answer is: it depends. One should not push too hard to achieve fast and deep convergence on speech signals since that could create very bad effects during double-talk or in noisy environments.
In general, a good LEC has to sound good within a variety of conditions, such as various echo paths, noise environments, speech signals, and conversational patterns. To best understand LEC performance, designers must consider a variety of measurements ranging from standard ITU G.168 tests to specially designed tests that use speech signals and a variety of noise sources. Various echo paths should be studied that are constructed from hybrids that may have different frequency characteristics. The test results should be evaluated based on probability of encountering test scenario in practice. Those that are more likely to occur within a particular type of network should be given more weight.
General LEC Design Issues
Figure 1 illustrates a block diagram of a typical digital echo canceller terminated by a hybrid. In general, the hybrid may be substituted for an arbitrary echo path. The nonlinear processor (NLP) is used only in a send direction for most LEC's.
Figure 1: Block diagram of an echo canceller.
The designers of a VoP system that use LEC's must provide an answer to the following 7 questions:
1. Software or hardware implementation? Some software solutions use only a small percentage of hardware. This may result in system solutions that take too much space and consume too much power. Hardware solutions may also burn too much power simply because they are performing exactly the same functions without much flexibility to adapt the algorithm execution based on environmental conditions. Upgrading hardware solutions is more expensive than upgrading software. As the channel density is considerably increased, e.g. high-density gateways, a hardware approach may tend to require more space and electrical power then a clever software approach on a low power general purpose digital signal processor (DSP).
2. Time or frequency domain adaptive filter? Software implementations tend to favor time domain adaptive filters. Frequency domain solutions often require more memory and may add substantial delay to the send out signal.
3. Least mean square (LMS) or recursive least square (RLS) based adaptation? The LMS algorithm is almost always used, sometimes in combination with the RLS. The RLS may be numerically unstable and generally requires more processing power.
4. Filter structure (FIR or some other)? The FIR filter structure is the most common.
5. Full-filter or multi-segment (MSEG) approach? Hardware implementations tend to favor full-filter approaches. Since the hardware implementations tend to use special purpose architectural components there is an opportunity to spend more hardware on more complex algorithms and potentially more memory within the same physical space. However, this often results in brute force approaches that end up burning more electrical power. The full-filter approach may considerably degrade performance in terms of depth and speed of convergence for longer filter lengths.
6. Type of NLP and comfort noise generation (CNG)? The NLP may use attenuation, center clipping, comfort noise generation (CNG), or a combination to deal with the echo residuals.
7. Step size control? Better LEC designs may use adaptive step size control to improve convergence properties for speech and to better deal with noisy environments.
The biggest challenge of LEC design is to find a good balance between cancellation and nonlinear processing. In theory, this may look very simple. One should consider activating NLP only when there is not enough echo cancellation (ERL+ERLE). The NLP should not be used during double-talk. One should always use the minimum amount of NLP based on achieved convergence and noise properties. The NLP should not be used for digital terminations, unless a minor acoustic or electrical echo leakage is detected (most likely from a phone handset). In practice, it is very hard to reliably detect and quantify these situations in real-time when dealing with various speech signals, network configurations and background noises.
Problems with LECs
Designing an LEC would be much simpler if most of the following statements would be true:
- Echo path has only one reflection with low initial delay, very short dispersion (flat frequency response), and ERL around 24 dB.
- Both sides of conversation are in a quiet room environment (background noises below --60 dBm).
- Speech levels of both sides are well balanced and average around --15 dBm.
People are polite and do not talk at the same time.
- The LEC filter is reset for every off-hook event unless it is within the CPE product for which it may keep the same coefficients between the calls and keep improving over time.
Unfortunately, most of the above conditions cannot be met in real-life design situations. Nevertheless, a good echo canceller should have a perfect performance under perfect conditions like these. The problems occur when one or more of the above conditions are not met. Even in case of digital terminations, designers may suffer from acoustic or electrical leakage within the handset or handset cable.
When balancing the cancellation with the NLP action, designers have to be aware of the problems that occur during an average phone call on regular basis. A good starting point is to create an LEC that uses minimum of NLP action under perfect conditions. Clearly, this assumes one can reliably quantify and detect the perfect conditions or any other conditions for that matter. The next step would be to balance the NLP action based on particular dynamic properties of the adaptive algorithm and for various working conditions, such as different echo paths, speech signals, and noise conditions.
Some implementation problems are tied to the specific design. A good example is the full-filter approach vs. MSEG. When exposed to speech signals, the full-filter approach tends to result in slower convergence. The steady state depth for speech may not be as good as for the MSEG solution. The full-filter solution may also be more sensitive to double-talk and noise conditions.
MSEG solutions have problems of their own. The biggest one is the robust localization of hybrid reflections. If a hybrid reflection is not detected or only partially covered with active filter coefficients, the LEC will not be able to provide enough cancellation. This would result in either stronger residual echo or frequent talk-over based on the particular NLP logic.
Dealing with the problems and design challenges comes at the expense of algorithm complexity. This translates into memory and processing requirements for software solutions or memory and hardware complexity for hardware solutions. Both hardware and software solutions would tend to use more electrical power to perform more work.
High-Density Gateway
While the LEC algorithm designers need to find the optimal balance between cancellation and NLP action, the system designers that use LECs are faced with a different set of problems some of which can be solved through a proper choice of algorithmic components that comprise an LEC. As the channel density is increased the system designers will be facing issues of physical size and electrical power consumption.
These parameters are mapped to computational complexity of the algorithms as well as their memory requirements. Let C1 be the amount of cycles spent by an LEC algorithm on a single channel. The total amount of cycles CN can be calculated as CN = C1 x N, where N is the number of channels.
Using this approach, a designer would spend all of the available cycles relatively soon. However, it is possible to implement a software control of the LEC algorithm on a large number of channels. Specifically, designers can assume that not all of the channels would require achieving the same improvement in performance all the time. Typically, most of useful work is done at the beginning of a call or immediately upon echo path change. Once certain level of convergence is achieved, the available CPU cycles can be distributed to other channels that may need to further improve their performance. In this way, scheduling algorithms can be implemented that limit the total consumption of CPU cycles for the system as a whole.
Figure 2 illustrates how software scheduling can reduce the total computational complexity on a system level. The straight green line represents the simple approach where the computational complexity of an LEC system changes linearly with the number of channels. The red line has logarithmic properties and illustrates how software scheduling may reduce the computational complexity of the same system for the same number of channels.
Figure 2: CPU cycle consumption as a function of channel density.
Clearly, as Figure 2 points out, by using the same amount of CPU cycles, designers may be able to achieve rather different channel densities based on a type of software scheduling. Conversely, for the same channel density, designers may be able to spend lower amount of CPU cycles if software scheduling is done right and hence reduce the amount of electrical power that is required.
Most hardware implementations use a brute force approach as captured by the straight green line. Software implementations that use the same approach will lose the battle in a long run since they will tend to occupy more board space as well as consume even more electrical power.
The software implementations that are captured by the red line have better chances for success. To judge success, designers must ensure high quality of the scheduling algorithms. In order to implement good LEC scheduling algorithms, a designer has to rely on statistical properties of speech signals over large number of channels. The non-stationary characteristics of speech signal, certain conversational patterns, network conditions, and various types of background noise may work against such algorithms. Moreover, catastrophic equipment failures may prompt a large number of channels to require considerable amount of processing power in order to restart convergence at the same time following a "system glitch".
Scheduling and control algorithms can differentiate on how well they handle these real-life situations. Algorithms that result in a graceful and predictable degradation of performance are preferred to those that act in unpredictable ways.
A good algorithm should also guarantee certain level of minimum performance that is not affected by changes in the environment. Ultimately, if a human observer cannot tell a difference between systems that have linear complexity vs. logarithmic, one should choose the system that provides smaller board space and consumes less electric power for the same channel density.
Future Trends
Future work on VoP systems incorporating acoustic echo canceling (AEC) will move toward further integration of hardware components. Ultimately, a single chip solution may integrate an analog front-end (AFE), packet network switch, DSP engine, and control and configuration unit. Additional features will include: conferencing support, additional channels for connecting fax machine or providing multi-channel voice gateway functions. Video will become a must very soon.
Future work on VoP systems using LEC's focus on providing highest possible channel densities keeping superb voice quality that is not affected much by changes in system load. At the same time, the board space and electric power consumption should be minimized.
It is hard to tell which LEC implementation will be more successful. But, based on analysis, it appears that hybrid solutions provide the best approaches to handling line echo in VoP architectures. Specifically, the solutions that win out appear to be implementations using sophisticated software control and scheduling running on hardware platforms with acceleration of specific DSP functions. The software solutions that implement the load control on a general purpose DSP's appear to be ahead of the game at this point.
Future research in DSP algorithms for line echo canceling will most likely focus on providing solutions that are suitable for efficient scheduling within high-density systems and that have reduced memory requirements, still keeping the highest possible performance, i.e. depth and speed of convergence as well as stability.
Editor's Note: To view part 1 of this article, Click Here
About the Author
Bogdan Kosanovic is the manager of echo cancellation technology at Telogy Networks, a Texas Instruments Company. He received a Dipl.Eng. degree in electrical engineering from the University of Belgrade, Serbia, and M.S. and Ph.D. degrees in electrical engineering from the University of Pittsburgh. Bogdan can be reached at bogdank@ti.com.
