Recent developments in technology have induced new wireless applications designed for enhancing voice quality while providing customers with expanded data-communications choices. Echo cancellers haven't escaped this technological transformation. These voice-quality-enhancement systems have triggered much of the momentum, driving new network architectures and propelling mandatory features into existence.
Demand for higher speed and the remarkable growth in Internet applications have transformed echo-canceller technology into a more intelligent resource, capable of interworking with the new breed of emerging applications.
Network planners who focus on voice quality in wireless communications face new technical challenges today. Many of these challenges can be overcome through the latest innovations in signal processing designed to remedy shortcomings (electric echo, acoustic echo, noise, speech level, distortion) and enhance quality of speech over wireless networks.
TRANSFORMING PRESENT TECHNOLOGY Echo cancellers in wireless networks are no longer synonymous with plain electric echo-cancellation technology, but have become instruments for voice-quality enhancement. They optimize gain level, improve signal-to-noise ratio, cancel acoustic echo and work with a variety of new data architectures -- all in addition to a significantly superior electric echo cancellation. Echo cancellers were never regarded as perfect devices; the fresh breed is required to provide faultless performance. They provide for unclipped, uninterrupted background music during speech pauses, unclipped and echo-free doubletalk performance and exceedingly swift initial convergence -- all for less than half the cost and at more than quadrupled density since one year ago.
Major issues surrounding this transformation include the following developments:
* Recent improvements in cellular coverage and vocoder performance stirred focus to less significant -- yet imperative -- areas of speech quality: echo (electric and acoustic), clippings, optimal speech level and noise reduction.
* Recent advances in DSP MIPS and memory technology made algorithms more economical and, therefore, feasible.
* The smaller size of wireless devices contributed to acoustic echo. The short distance between the receive and transmit parts on the phone increase signal crosstalk and echo.
* The emergence of new data applications has imposed notable interworking requirements.
* The emergence of tandem-free operations standards has transformed the architecture and the nature of echo cancellers.
HYBRID ECHO-CANCELLATION PERFORMANCE The last two years have seen several noteworthy developments molding the performance of electric echo cancellation:
* Initial Convergence. The race for excellence started with a quest for speed of initial convergence. Some echo-canceller vendors introduced a more intelligent non-linear processor (NLP), which has not always proved beneficial since reliability on NLP may result in excessive clipping. Moreover, some implementations placed higher reliance on heuristics, which resulted in poor performance. Other solutions use the concept of dual-mode algorithm, which only can perform as well as the accuracy of its mechanism to detect path change. When this mechanism is implemented via triggers, the success rate may be less than perfect. Consequently, the echo-canceller-detection performance depends on the algorithm knack.
Recent breakthroughs have driven down the need for a mechanism to detect path change. A new algorithm introduced in early 1999 departs from the conventional algorithm, bringing about a markedly smarter, more accurate and faster methodology. The latest (ECP/R5) echo canceller ASIC puts into action this revolutionary approach.
* Near-End Speech Detection. Most textbook algorithms assume that signal levels on both ends of a call are equal, and electrical hybrids provide x-dB loss at the minimum. Under these postulations, a signal is marked "echo" whenever its average level is at least x-dB below the signal. Wireless networks stage a detour around the "equal level" supposition. The wireless end of an average call is normally 6dB to 8dB, or louder than its analog counterpart. These conditions would confuse most of the older echo-canceller algorithms and would set off excessive clipping of the public switched telephone network source signal.
Recent developments take into account the high likelihood of having unequal signal levels, which have led to more intelligent and dynamically adaptive threshold settings. One methodology consists of multiple algorithms, selectively triggered by an array of parameter state values. The latest product line operates efficiently without tradeoffs between speed of convergence and clipping performance.
* Hot Signals. Hot source signals present a challenge because of the corresponding hot echo and the fact that peaks of the hot signal get clipped. Consequently, the average echo-signal level may be high in relation to the clipped peaks and may be misclassified as doubletalk. A false doubletalk classification would freeze adaptation and result in divergence and bursts of echo.
Latest advances have taken special care of hot signals and have incorporated new algorithms designed to detect and adjust for the advent of these episodes.
* Divergence During Doubletalk. Progress in performance during doubletalk is one of the most dramatic recent enhancements. Whereas past and textbook algorithms placed their most weighty reliance on signal-relative level, the new approach employs an additional innovation. The criterion used for determining whether a detected signal is echo or doubletalk tags on a comparison of residual echo to the expected reading under a doubletalk condition before freezing the h register. Contrasting the results, the algorithm then makes the determination as to whether the detected signal is near-end speech or plain echo. An assessment in favor of speech would replace the convoluted signal by the one containing the frozen h register.
The latest G.168 ITU standard confirms the superior performance employing the latest methodology. Test 3b in the G.168 standard document runs under the conditions specified in the G.168 appendix. Echo cancellers employing the latest approach don't exhibit divergence during doubletalk.
* Background Music/Music on Hold. Let's suppose that Beethoven were coloring the far-end background while you were trying to make your point over the telephone. Did you ever interrupt (clip) the crescendo when uttering a "Hmm" after a brief pause? If so, then you must have experienced the working of an echo canceller operating with an old-fashioned on/off type NLP.
Recent developments in NLP tolerate low-level background music sneaking in while speech is present on the other end. It makes for a more natural interaction and contributes to a mood of "being there."
* Nonlinear Transitions. Wireless vocoders use compression algorithms that rely on linear prediction, which perform better when operating within a linear landscape. Traditional echo cancellers provide for on/off NLP and on/off noise-matching functions, giving off discontinuities affecting the linear prediction process. This operation leads to distortion, noise bursts and garbled speech.
Recent developments in NLP and noise-matching algorithms replace the on/off functionality with a smooth transition. Specifically the NLP function has been transformed from a switching function into a gradual loss-insertion maneuver. The loss-insertion rate and its associated step size have been optimized to provide for the best trade-off between speech vocoder and residual echo-elimination performance.
ACOUSTIC ECHO CONTROL Acoustic echo is created by sound waves leaking through a telephone handset from the receiver to the speaker and/or from a handset or a speaker phone receiver via reflections bouncing back from solid objects in the sound path.
Until recently, digital wireless applications implemented echo cancellers to nullify echo generated on the PSTN end only. At the same time, the PSTN end was not supposed to experience any echo because mobile-phone standards require equipment manufacturers to engineer sufficient loss in the acoustic echo path, and the 2-to-4-wire hybrid present in the PSTN is absent in the digital wireless environment.
If wireless-device vendors do not comply with official standards requiring adequate isolation between the receiver and the microphone, acoustic echo can become even more of a concern for wireless service providers. The acoustic-echo problem is much more evident in the case of digital wireless due to the long processing delay (>200ms round-trip) introduced by the speech-compression techniques and the non-linearities introduced by these algorithms.
* Acoustic echo in digital wireless applications is non-linear. The acoustic echo injected into the handset microphone enters a vocoder, which processes it with speech-compression techniques, modifying the signal, while fashioning a non-linear association between the signal and its acoustic echo.
Non-linearity implies that employing the relatively simple mathematical model for eliminating or reducing acoustic echo in the digital wireless environment is analogous to taking antibiotics for the common cold. It is a wrong medicine, and it might cause additional unwanted side effects including distortion, more echo and noise.
Consequently, products employing standard linear convolution on the wireless acoustic echo are expensive and ineffective.
* Acoustic echo in digital wireless applications is non-stationary. Because acoustic echo comes from sound waves bouncing off solid objects, changes in the positions of these objects relative to the microphone change the echo.
Being non-stationary implies that the signal-impulse response is a moving target, and the mathematical model created inside the echo canceller would have to perpetually chase it throughout a call. This makes training on a frequency response useless. By the time the echo canceller has been trained, the solution induces more harm than good. This is one more reason why products employing standard linear convolution on wireless acoustic echo are expensive and ineffective.
The pre-eminent technique for controlling the non-linear non-stationary acoustic echo is via a non-linear operation. However, this method may be subject to biting side effects, including clippingand disturbing background variations. The proper non-linear procedure must take meaningful steps to minimize side effects while identifying acoustic echo and separating it from signals.
Another important parameter in controlling acoustic echo is timing. The traditional non-linear processor operates over a relatively short echo-path delay averaging between 10ms and 64ms. Timing errors could result in troublesome clipping after-effects.
RECENT ADVANCES Reducing acoustic echo without side effects depends on the ability to characterize signals. Classification errors are harmful. The generic spectral characteristics of echo resemble speech, and non-linearities in the echo path give rise to mismatch. Consequently, potential level differences provide the information leading to distinction of echo from its ascendant speech.
Acoustic echo is intermittent, appearing in many sessions while absent in others. Most vendors provide acoustic echo-control capabilities, however effectiveness varies remarkably from one to another depending on detection. False detection may lead to excessive clipping, while lowering the threshold and softening the trigger may lead to excessive acoustic echo. A major enhancement to detection has been introduced via a lined-up noise-reduction capability. This feature contributes to a cleaning up of the signal-level reading before a discriminating algorithm employs it in determining whether or not it spots echo.
APPLICATION/METHODOLOGY Background noises such as street traffic, car engines and wind are transmitted and mixed with the speech signal. This nuisance may interfere with comprehension and cause dissatisfaction. The latest noise-reduction feature embedded in the latest echo cancellers separate the good from the bad.
By the time a signal lays a hand on the echo canceller, the speech and the noise are blended into a gourmet scalded dish, too bitter for a listener taste. Only a grand chemist (or a super chef) is able to either dilute the bitterness or cart some of it off. Diluting the colored noise by pumping white noise entails a masking strategy. It does not solve the problem since it worsens the overall signal-to-noise ratio. The only acceptable strategy is the identification and separation of the noise from the signal content.
Background noise targeted for elimination carries characteristics that can be identified and isolated, such as monotony accompanied by specific spectral properties. When they are found, the algorithm shifts to a more thorough analysis of signal auto-correlation, complemented by level and frequency-range examination. The noise is attenuated, while the lion's share of the signal is passed through.
Given the noise-reduction algorithm characteristics, its strengths would best be exhibited when the noise is monotonic, such as fan noise. At the same time, random, short bursts of noise such as police sirens would not be purged, and would pass through. Most prefer it this way, because the signal-processing product seeks to represent the background atmosphere without interfering with the ability to communicate.
AUTOMATIC LEVEL/GAIN CONTROL There are five basic approaches to automatic level/gain control (ALC or AGC respectively):
* Boosting a low-level signal and carrying it to a fixed target value
* Boosting a low-level signal while providing an equalizer effect by further amplifying base frequencies
* Boosting a low-level signal or attenuating a (too) hot signal and carrying either to a fixed target value
* Boosting a low-level signal to compensate for the listener's surrounding noise
* Boosting a low-level signal or attenuating a hot signal, and carrying either to a dynamically changing target value. The target value, in this case, moves to compensate for the listener's surrounding noise.
ALC/AGC corrects transmission deficiencies and/or compensates for noise. The average wireless signal is louder than the PSTN equivalent. The level inequality adds stress as explained earlier. At the same time, the loud wireless or subdued PSTN signals don't support optimal listening levels. Many phones have volume control, which may grant partial correction while amplifying sidetones. Adaptive adjustments -- compensating for noise -- cannot be duplicated on the handset. The new AGC features adjust signals from the network, eliminating the corresponding rise in noise levels and improving signal-to-noise ratio.
The latest innovation in this area combines the fixed and the adaptive level controls into a single feature. It either amplifies a low-level signal or attenuates a hot signal, carrying whichever to a dynamically changing target value. The target value, in this case, is set to compensate for the noise, ensuring an optimal listening level under any ambiance. This methodology has been rightly named intelligent level control.
BUYER AWARE Echo-cancellation technology has advanced significantly, shaping some of the latest standards requirements. The growth in new echo-canceller vendors creates a stark necessity for understanding the innovations and comprehending the richness in quality and performance features available to sophisticated buyers.
Anyone who says, "An echo canceller is just an echo canceller -- get me the least expensive one," is not dealing with reality. Although the echo canceller represents a minor feature within an average telecommunications offer, its impact on overall perception of quality is enormous. Network planners must realize echo cancellers have evolved to become voice-enhancement systems, and consumers must vote with their wallets to maximize satisfaction by appreciating those who provide them with a "just like being there" quality.
Want to use this article? Click here for options!
© 2014 Penton Media Inc.
From the Blog
Join the Discussion
Get more out of Connected Planet by visiting our related resources below:
Connected Planet highlights the next generation of service providers, as well as how their customers use services in new ways.Subscribe Now