1 /* SPDX-License-Identifier: GPL-2.0-only */
2 /*
3  * SpanDSP - a series of DSP components for telephony
4  *
5  * echo.c - A line echo canceller.  This code is being developed
6  *          against and partially complies with G168.
7  *
8  * Written by Steve Underwood <steveu@coppice.org>
9  *         and David Rowe <david_at_rowetel_dot_com>
10  *
11  * Copyright (C) 2001 Steve Underwood and 2007 David Rowe
12  *
13  * All rights reserved.
14  */
15 
16 #ifndef __ECHO_H
17 #define __ECHO_H
18 
19 /*
20 Line echo cancellation for voice
21 
22 What does it do?
23 
24 This module aims to provide G.168-2002 compliant echo cancellation, to remove
25 electrical echoes (e.g. from 2-4 wire hybrids) from voice calls.
26 
27 How does it work?
28 
29 The heart of the echo cancellor is FIR filter. This is adapted to match the
30 echo impulse response of the telephone line. It must be long enough to
31 adequately cover the duration of that impulse response. The signal transmitted
32 to the telephone line is passed through the FIR filter. Once the FIR is
33 properly adapted, the resulting output is an estimate of the echo signal
34 received from the line. This is subtracted from the received signal. The result
35 is an estimate of the signal which originated at the far end of the line, free
36 from echos of our own transmitted signal.
37 
38 The least mean squares (LMS) algorithm is attributed to Widrow and Hoff, and
39 was introduced in 1960. It is the commonest form of filter adaption used in
40 things like modem line equalisers and line echo cancellers. There it works very
41 well.  However, it only works well for signals of constant amplitude. It works
42 very poorly for things like speech echo cancellation, where the signal level
43 varies widely.  This is quite easy to fix. If the signal level is normalised -
44 similar to applying AGC - LMS can work as well for a signal of varying
45 amplitude as it does for a modem signal. This normalised least mean squares
46 (NLMS) algorithm is the commonest one used for speech echo cancellation. Many
47 other algorithms exist - e.g. RLS (essentially the same as Kalman filtering),
48 FAP, etc. Some perform significantly better than NLMS.  However, factors such
49 as computational complexity and patents favour the use of NLMS.
50 
51 A simple refinement to NLMS can improve its performance with speech. NLMS tends
52 to adapt best to the strongest parts of a signal. If the signal is white noise,
53 the NLMS algorithm works very well. However, speech has more low frequency than
54 high frequency content. Pre-whitening (i.e. filtering the signal to flatten its
55 spectrum) the echo signal improves the adapt rate for speech, and ensures the
56 final residual signal is not heavily biased towards high frequencies. A very
57 low complexity filter is adequate for this, so pre-whitening adds little to the
58 compute requirements of the echo canceller.
59 
60 An FIR filter adapted using pre-whitened NLMS performs well, provided certain
61 conditions are met:
62 
63     - The transmitted signal has poor self-correlation.
64     - There is no signal being generated within the environment being
65       cancelled.
66 
67 The difficulty is that neither of these can be guaranteed.
68 
69 If the adaption is performed while transmitting noise (or something fairly
70 noise like, such as voice) the adaption works very well. If the adaption is
71 performed while transmitting something highly correlative (typically narrow
72 band energy such as signalling tones or DTMF), the adaption can go seriously
73 wrong. The reason is there is only one solution for the adaption on a near
74 random signal - the impulse response of the line. For a repetitive signal,
75 there are any number of solutions which converge the adaption, and nothing
76 guides the adaption to choose the generalised one. Allowing an untrained
77 canceller to converge on this kind of narrowband energy probably a good thing,
78 since at least it cancels the tones. Allowing a well converged canceller to
79 continue converging on such energy is just a way to ruin its generalised
80 adaption. A narrowband detector is needed, so adapation can be suspended at
81 appropriate times.
82 
83 The adaption process is based on trying to eliminate the received signal. When
84 there is any signal from within the environment being cancelled it may upset
85 the adaption process. Similarly, if the signal we are transmitting is small,
86 noise may dominate and disturb the adaption process. If we can ensure that the
87 adaption is only performed when we are transmitting a significant signal level,
88 and the environment is not, things will be OK. Clearly, it is easy to tell when
89 we are sending a significant signal. Telling, if the environment is generating
90 a significant signal, and doing it with sufficient speed that the adaption will
91 not have diverged too much more we stop it, is a little harder.
92 
93 The key problem in detecting when the environment is sourcing significant
94 energy is that we must do this very quickly. Given a reasonably long sample of
95 the received signal, there are a number of strategies which may be used to
96 assess whether that signal contains a strong far end component. However, by the
97 time that assessment is complete the far end signal will have already caused
98 major mis-convergence in the adaption process. An assessment algorithm is
99 needed which produces a fairly accurate result from a very short burst of far
100 end energy.
101 
102 How do I use it?
103 
104 The echo cancellor processes both the transmit and receive streams sample by
105 sample. The processing function is not declared inline. Unfortunately,
106 cancellation requires many operations per sample, so the call overhead is only
107 a minor burden.
108 */
109 
110 #include "fir.h"
111 #include "oslec.h"
112 
113 /*
114     G.168 echo canceller descriptor. This defines the working state for a line
115     echo canceller.
116 */
117 struct oslec_state {
118 	int16_t tx;
119 	int16_t rx;
120 	int16_t clean;
121 	int16_t clean_nlp;
122 
123 	int nonupdate_dwell;
124 	int curr_pos;
125 	int taps;
126 	int log2taps;
127 	int adaption_mode;
128 
129 	int cond_met;
130 	int32_t pstates;
131 	int16_t adapt;
132 	int32_t factor;
133 	int16_t shift;
134 
135 	/* Average levels and averaging filter states */
136 	int ltxacc;
137 	int lrxacc;
138 	int lcleanacc;
139 	int lclean_bgacc;
140 	int ltx;
141 	int lrx;
142 	int lclean;
143 	int lclean_bg;
144 	int lbgn;
145 	int lbgn_acc;
146 	int lbgn_upper;
147 	int lbgn_upper_acc;
148 
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];
153 
154 	/* DC blocking filter states */
155 	int tx_1;
156 	int tx_2;
157 	int rx_1;
158 	int rx_2;
159 
160 	/* optional High Pass Filter states */
161 	int32_t xvtx[5];
162 	int32_t yvtx[5];
163 	int32_t xvrx[5];
164 	int32_t yvrx[5];
165 
166 	/* Parameters for the optional Hoth noise generator */
167 	int cng_level;
168 	int cng_rndnum;
169 	int cng_filter;
170 
171 	/* snapshot sample of coeffs used for development */
172 	int16_t *snapshot;
173 };
174 
175 #endif /* __ECHO_H */
176