2 * SpanDSP - a series of DSP components for telephony
4 * echo.c - A line echo canceller. This code is being developed
5 * against and partially complies with G168.
7 * Written by Steve Underwood <steveu@coppice.org>
8 * and David Rowe <david_at_rowetel_dot_com>
10 * Copyright (C) 2001 Steve Underwood and 2007 David Rowe
12 * All rights reserved.
14 * This program is free software; you can redistribute it and/or modify
15 * it under the terms of the GNU General Public License version 2, as
16 * published by the Free Software Foundation.
18 * This program is distributed in the hope that it will be useful,
19 * but WITHOUT ANY WARRANTY; without even the implied warranty of
20 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
21 * GNU General Public License for more details.
23 * You should have received a copy of the GNU General Public License
24 * along with this program; if not, write to the Free Software
25 * Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA.
31 /*! \page echo_can_page Line echo cancellation for voice
33 \section echo_can_page_sec_1 What does it do?
34 This module aims to provide G.168-2002 compliant echo cancellation, to remove
35 electrical echoes (e.g. from 2-4 wire hybrids) from voice calls.
37 \section echo_can_page_sec_2 How does it work?
38 The heart of the echo cancellor is FIR filter. This is adapted to match the
39 echo impulse response of the telephone line. It must be long enough to
40 adequately cover the duration of that impulse response. The signal transmitted
41 to the telephone line is passed through the FIR filter. Once the FIR is
42 properly adapted, the resulting output is an estimate of the echo signal
43 received from the line. This is subtracted from the received signal. The result
44 is an estimate of the signal which originated at the far end of the line, free
45 from echos of our own transmitted signal.
47 The least mean squares (LMS) algorithm is attributed to Widrow and Hoff, and
48 was introduced in 1960. It is the commonest form of filter adaption used in
49 things like modem line equalisers and line echo cancellers. There it works very
50 well. However, it only works well for signals of constant amplitude. It works
51 very poorly for things like speech echo cancellation, where the signal level
52 varies widely. This is quite easy to fix. If the signal level is normalised -
53 similar to applying AGC - LMS can work as well for a signal of varying
54 amplitude as it does for a modem signal. This normalised least mean squares
55 (NLMS) algorithm is the commonest one used for speech echo cancellation. Many
56 other algorithms exist - e.g. RLS (essentially the same as Kalman filtering),
57 FAP, etc. Some perform significantly better than NLMS. However, factors such
58 as computational complexity and patents favour the use of NLMS.
60 A simple refinement to NLMS can improve its performance with speech. NLMS tends
61 to adapt best to the strongest parts of a signal. If the signal is white noise,
62 the NLMS algorithm works very well. However, speech has more low frequency than
63 high frequency content. Pre-whitening (i.e. filtering the signal to flatten its
64 spectrum) the echo signal improves the adapt rate for speech, and ensures the
65 final residual signal is not heavily biased towards high frequencies. A very
66 low complexity filter is adequate for this, so pre-whitening adds little to the
67 compute requirements of the echo canceller.
69 An FIR filter adapted using pre-whitened NLMS performs well, provided certain
72 - The transmitted signal has poor self-correlation.
73 - There is no signal being generated within the environment being
76 The difficulty is that neither of these can be guaranteed.
78 If the adaption is performed while transmitting noise (or something fairly
79 noise like, such as voice) the adaption works very well. If the adaption is
80 performed while transmitting something highly correlative (typically narrow
81 band energy such as signalling tones or DTMF), the adaption can go seriously
82 wrong. The reason is there is only one solution for the adaption on a near
83 random signal - the impulse response of the line. For a repetitive signal,
84 there are any number of solutions which converge the adaption, and nothing
85 guides the adaption to choose the generalised one. Allowing an untrained
86 canceller to converge on this kind of narrowband energy probably a good thing,
87 since at least it cancels the tones. Allowing a well converged canceller to
88 continue converging on such energy is just a way to ruin its generalised
89 adaption. A narrowband detector is needed, so adapation can be suspended at
92 The adaption process is based on trying to eliminate the received signal. When
93 there is any signal from within the environment being cancelled it may upset
94 the adaption process. Similarly, if the signal we are transmitting is small,
95 noise may dominate and disturb the adaption process. If we can ensure that the
96 adaption is only performed when we are transmitting a significant signal level,
97 and the environment is not, things will be OK. Clearly, it is easy to tell when
98 we are sending a significant signal. Telling, if the environment is generating
99 a significant signal, and doing it with sufficient speed that the adaption will
100 not have diverged too much more we stop it, is a little harder.
102 The key problem in detecting when the environment is sourcing significant
103 energy is that we must do this very quickly. Given a reasonably long sample of
104 the received signal, there are a number of strategies which may be used to
105 assess whether that signal contains a strong far end component. However, by the
106 time that assessment is complete the far end signal will have already caused
107 major mis-convergence in the adaption process. An assessment algorithm is
108 needed which produces a fairly accurate result from a very short burst of far
111 \section echo_can_page_sec_3 How do I use it?
112 The echo cancellor processes both the transmit and receive streams sample by
113 sample. The processing function is not declared inline. Unfortunately,
114 cancellation requires many operations per sample, so the call overhead is only
122 G.168 echo canceller descriptor. This defines the working state for a line
142 /* Average levels and averaging filter states */
143 int Ltxacc
, Lrxacc
, Lcleanacc
, Lclean_bgacc
;
147 int Lbgn
, Lbgn_acc
, Lbgn_upper
, Lbgn_upper_acc
;
149 /* foreground and background filter states */
150 struct fir16_state_t fir_state
;
151 struct fir16_state_t fir_state_bg
;
152 int16_t *fir_taps16
[2];
154 /* DC blocking filter states */
155 int tx_1
, tx_2
, rx_1
, rx_2
;
157 /* optional High Pass Filter states */
158 int32_t xvtx
[5], yvtx
[5];
159 int32_t xvrx
[5], yvrx
[5];
161 /* Parameters for the optional Hoth noise generator */
166 /* snapshot sample of coeffs used for development */
170 #endif /* __ECHO_H */