Skip to content
Tutorials Center
My learning

Topics

Lesson 35 of 36 ยท Projects and Practice

Project: Echo, Reverb and Inverse Filtering

Model an echo as an LTI system, study its comb-filter frequency response, add feedback for repeated echoes, estimate the delay by autocorrelation and remove the echo.

Shout in a canyon, and a delayed, quieter copy of your voice comes back. That is an echo, and it is one of the cleanest real-world LTI systems: an output equal to the input plus a delayed, scaled copy. In this project you describe it with an impulse response, a transfer function and a frequency response; add repeated echoes with feedback; find an unknown delay from a recording; and then undo the echo with an inverse filter.

Model 1: a single echo (FIR comb)

With delay D samples and gain a (0 < a < 1):

y[n] = x[n] + a x[n - D].

The impulse response is h[n] = ฮด[n] + a ฮด[n - D], two spikes. The transfer function is H(z) = 1 + a z^(-D), which has D zeros spread on a circle of radius a^(1/D) inside the unit circle. On the unit circle:

|H(e^(jฯ‰))| = |1 + a e^(-jฯ‰D)|.

This ripples between 1 + a (when ฯ‰D is a multiple of 2ฯ€, echo adds in phase) and 1 - a (when ฯ‰D is an odd multiple of ฯ€, the echo cancels part of the signal). The picture of peaks and valleys looks like a comb, so this is called a comb filter. It explains the hollow, "phasey" sound of a short echo, and why a hard wall behind a speaker colors its sound.

Python The comb in numbers

Press Run (or Ctrl + Enter) to see the result here.

Program input ยท what you'd type while the program runs

Type the answers before you press Run, one per line. Each time the program reads a line (with input()), it gets the next line from this box.

The peaks repeat every fs / D = 200 Hz. A longer delay gives a denser comb.

Model 2: repeated echoes (IIR comb)

A cave or a hallway returns the echo of the echo. Feed the output back:

y[n] = x[n] + a y[n - D], with H(z) = 1 / (1 - a z^(-D)).

The impulse response is h[n] = ฮฃ aแต ฮด[n - kD], an infinite train of spikes with geometrically falling heights. This is the simplest building block of artificial reverb: the Schroeder reverberator uses several such combs with different delays, followed by all-pass filters to thicken the sound. The system is stable when |a| < 1 because all D poles then lie at radius a^(1/D) < 1.

Python Single versus repeated echoes

Press Run (or Ctrl + Enter) to see the result here.

Program input ยท what you'd type while the program runs

Type the answers before you press Run, one per line. Each time the program reads a line (with input()), it gets the next line from this box.

Estimating an unknown delay with autocorrelation

Suppose you only have the recording y and want to find D. For a signal x that is nearly random (like noise or speech over a long window), x is uncorrelated with itself shifted by a nonzero lag. The autocorrelation of y = x + a x[n - D] has a peak at lag 0 and another at lag D, of relative height about a / (1 + aยฒ). Computing the autocorrelation with the FFT makes this fast.

Python Finding the echo delay

Press Run (or Ctrl + Enter) to see the result here.

Program input ยท what you'd type while the program runs

Type the answers before you press Run, one per line. Each time the program reads a line (with input()), it gets the next line from this box.

Removing the echo: the inverse system

An echo with FIR transfer function 1 + a z^(-D) is undone by H_inv(z) = 1 / (1 + a z^(-D)), the inverse system, which is exactly a feedback comb with gain -a: x_hat[n] = y[n] - a x_hat[n - D]. Because |a| < 1 the inverse is causal and stable, so the original is recovered completely. If the echo were louder than the direct sound (|a| > 1), the zeros of H would lie outside the unit circle and the causal inverse would be unstable, a general rule: an inverse of a stable system is stable only if the system is minimum phase.

Python Dereverberate with the inverse filter

Press Run (or Ctrl + Enter) to see the result here.

Program input ยท what you'd type while the program runs

Type the answers before you press Run, one per line. Each time the program reads a line (with input()), it gets the next line from this box.

The exact recovery works because you know a and D exactly and the data has no noise. With noisy data the inverse filter amplifies noise at the frequencies where |H| is small, so real dereverberation uses regularized or adaptive methods.

Ideas to extend the project

  • Load a WAV file locally with scipy.io.wavfile, apply the echo and write it back.
  • Combine four IIR combs with different delays (for example 29.7, 37.1, 41.1 and 43.7 ms) and sum them to build a simple Schroeder reverb.
  • Use the Fourier transform of a measured room impulse response to see its room modes.

Check your understanding

Quick check ยท 4 questions

of 4 answered You scored / 4 ยท perfect!

  1. 1. What is the transfer function of the single-echo system y[n] = x[n] + a x[n - D]?

    A delay of D samples is z^(-D), so the output transform is X(z)(1 + a z^(-D)).
  2. 2. What is the largest gain of the comb filter 1 + a e^(-jฯ‰D)?

    When the echo is in phase with the direct sound, the two add to give 1 + a.
  3. 3. Why does the feedback comb y[n] = x[n] + a y[n - D] need |a| < 1?

    The poles satisfy z^D = a, so they lie at radius |a|^(1/D); only |a| < 1 puts them inside the unit circle.
  4. 4. How can you estimate an unknown echo delay from a recording of random-like sound?

    A delayed copy correlates with the original, so the autocorrelation has a second peak at lag D.

Summary

  • An echo is an LTI system; its impulse response has spikes at the delays.
  • The single echo gives a comb filter with peaks at 1 + a and valleys at 1 - a, spaced fs / D apart.
  • Feedback creates repeated echoes and the first building block of reverb; stability needs |a| < 1.
  • Autocorrelation reveals an unknown delay; a feedback inverse filter removes a known echo.

Next: Practice Problems for Signals and Systems