DEDTMF - Tool to remove stationary DTMF-like tones from audio

dedtmf is a Matlab script that attempts to suppress stationary tones an audio file, while leaving more dynamic components unchanged. This could be useful, for instance, to suppress DTMF tones mixed in with speech.

Principle
Example usage
Limitations
Installation
Python Port
Changelog
Acknowledgment

Principle

Stationary tones (i.e., locally periodic signals) will lead to LPC fits with very sharp poles (i.e., z-domain magnitudes close to the unit circle). The routine fits LPC models to fairly long (e.g., 250ms) windows of the signal. A signal whose periodicity is stable over that window is detected by its large-radius poles; these poles are then converted to zeros on the unit circle along with compensatory poles just inside them, leading to a set of narrow notches, which are used to filter the tone out of that segment. A small segment from the middle of the block is used in overlap-add to reconstruct the original signal with the steady tones attenuated by at least 40 dB (by eye).

Example usage

The example waveform includes dial tone, the DTMF tones during dialing, then some speech mixed with a repeating tone.

[d,sr] = wavread('phonexample.wav');
soundsc(d,sr); 
% Now we suppress it using a 40-pole fit, over 4096 point windows
% every 256 samples
H = 256;
W = 4096;
[y,E,F,R,T] = dedtmf(d, 40, W, H);
% Listening to result, tones are largely suppressed
soundsc(y,sr); 
% Compare spectrograms
tlim = 3.0;
axs(1) = subplot(311);
specgram(d,256,sr);
caxis([-50 30]);
axis([0 tlim 0 4000]);
title('Original');
axs(2) = subplot(312);
specgram(y,256,sr);
caxis([-50 30]);
axis([0 tlim 0 4000]);
title('Tones suppressed')
% Plot energy ratio
axs(3) = subplot(313);
plot(([0:length(E)-1]*H+W/2)/sr, E);
%axis([0 length(d)/sr 0 1.1]);
axis([0 tlim 0 1.1]);
title('Frame energy ratio - filtered / original')
linkaxes(axs, 'x');
% Show where the main detected tones were
mainpoles = find(R>.99);
subplot(311)
hold on;
plot(T(mainpoles)/sr, F(mainpoles)/pi*sr/2, '.w');
hold off;
% Notice how the DTMF tones are completely eliminated, leading to
% energy ratios close to zero.  The speech, however, is largely
% untouched.  Some of the other tones have complex AM that makes
% some of their components not be detected as steady tones, and
% hence not removed.

Limitations

The tones are removed by stationary notches, albeit applied on short blocks (256 samples = 16 ms in this eample). However, if the tones are not entirely stable, they will not be completely eliminated. You can hear that with the ascending tones in this example, which have a bit of tape flutter and thus are not exactly fixed frequency.

Installation

The Matlab code is available at

 <http://labrosa.ee.columbia.edu/projects/dedtmf/>

All sources are in the package dedtmf.zip.

Feel free to contact me with any problems.

Python Port

I have ported this code to Python. See dedtmf.py.

Changelog

% 2013-11-26 v0.1 Initial release

Acknowledgment

This work was supported by DARPA under the RATS program via a subcontract from the SRI-led team SCENIC. My work was on behalf of ICSI.

$Header: /Users/dpwe/docs/grants/2010-01-DARPA-RATS/code/dedtmf/RCS/demo_dedtmf.m,v 1.2 2011/05/03 20:32:37 dpwe Exp dpwe $