Echo Cancellation

Echo cancellation is a critical application of adaptive filters in telecommunications that removes acoustic or electrical echoes from voice communication systems. Without echo cancellation, a speaker hears their own voice reflected back with a delay, severely degrading communication quality. Adaptive filtering solves this by continuously learning the echo path and generating a synthetic echo to subtract from the microphone signal.

Core Concept of Echo Cancellation

In a hands-free telephone or conferencing system, the far-end speaker's voice x(n) is played through a loudspeaker. Part of this sound travels across the room, reflects off walls and objects, and is picked up by the microphone as an acoustic echo. This echo reaches the far-end listener with a delay, making the conversation confusing. The delay can be 50 to several hundred milliseconds depending on room size and loudspeaker placement.

The microphone signal d(n) contains both the near-end talker's voice s(n) and the echo component. Since the far-end signal x(n) is known at the near-end processor, it can be used as the reference input to an adaptive filter. The filter learns the echo path impulse response h(n), which characterizes how sound propagates from the loudspeaker to the microphone through the room. The filter output y(n) is then subtracted from d(n) to obtain the echo-cancelled signal e(n).

The challenge is that the echo path changes whenever someone moves, a door opens, or the loudspeaker volume changes. The adaptive filter must continuously track these changes, making the LMS algorithm an ideal choice due to its low complexity and continuous adaptation capability.

Mathematical Expression

The echo component at the microphone is modeled as echo(n) = h^T * x_vec(n), where h is the echo path impulse response vector and x_vec(n) contains the M most recent samples of x(n). The adaptive filter estimate is y(n) = w^T(n) * x_vec(n). The residual signal is e(n) = s(n) + echo(n) - y(n). When w(n) converges to h, e(n) converges to s(n).

The echo return loss enhancement (ERLE) is the key performance metric, defined as ERLE = 10 * log10(E[d^2(n)] / E[e^2(n)]). It quantifies how much the echo power is reduced. A good echo canceller achieves 30 to 40 dB ERLE. The filter order M must be large enough to span the entire echo tail, which requires M > fs * T_echo, where fs is the sampling rate and T_echo is the echo delay spread.

Practical Understanding

A major practical problem in echo cancellation is double talk, which occurs when both the near-end and far-end talkers speak simultaneously. During double talk, the error signal e(n) contains both s(n) and residual echo, so using e(n) to update the adaptive filter will corrupt the filter coefficients. Double-talk detectors (DTD) monitor the signal and freeze the weight update when double talk is detected.

Electrical echo cancellation is needed in telephone line hybrids, where a 2-wire to 4-wire conversion produces an impedance mismatch that reflects part of the transmitted signal back toward the transmitter. This type of echo has a well-defined and relatively stable echo path, making it easier to cancel than acoustic echo.

Given:
Echo path length T_echo = 200 ms, Sampling rate fs = 8000 Hz
LMS step size mu = 0.001

Why this formula applies:
Filter order must cover the entire echo path

Formula:
M_min = fs * T_echo

Substitution:
M_min = 8000 * 0.200 = 1600 taps

Calculation:
Filter order M = 1600
With mu = 0.001, convergence condition: mu < 1/(M * P_x)
If P_x = 0.5, mu_max = 1/(1600*0.5) = 0.00125
Chosen mu = 0.001 < 0.00125, so convergence is guaranteed.

Final Answer:
Echo canceller requires at least 1600-tap filter for 200 ms echo path at 8 kHz sampling.

Exam Tip: For GATE, remember ERLE = 10*log10(input echo power / residual echo power). A higher ERLE means better cancellation. Double-talk is the main reason adaptive weight updates are paused in practical echo cancellers. Filter order M must satisfy M >= fs * T_echo.

• Echo path h(n) is unknown and time-varying; adaptive filter learns it from the far-end signal.
• Filter order M must satisfy M >= fs * T_echo to cover the full echo tail.
• ERLE in dB measures cancellation quality; 30 to 40 dB is typical in good systems.
• Double talk causes weight corruption; DTD freezes updates during simultaneous speech.
• Electrical echo from hybrid transformers is easier to cancel than acoustic echo.

Quick Revision

• d(n) = s(n) + echo(n); adaptive filter estimates echo path h(n) using far-end signal x(n).
• ERLE = 10*log10(E[d^2(n)] / E[e^2(n)]), higher is better.
• Filter order M_min = fs * T_echo; for 200 ms at 8 kHz, M = 1600 taps.
• Double talk = both ends speaking; solution is double-talk detector (DTD).
• LMS preferred for echo cancellation due to low complexity and continuous adaptation.
• GATE trap: echo cancellation error signal converges to s(n), not to zero.
• Acoustic echo path changes with room conditions; electrical echo path is more stable.