xref: /openbmc/linux/drivers/misc/echo/echo.h (revision e3b9f1e8)
1 /*
2  * SpanDSP - a series of DSP components for telephony
3  *
4  * echo.c - A line echo canceller.  This code is being developed
5  *          against and partially complies with G168.
6  *
7  * Written by Steve Underwood <steveu@coppice.org>
8  *         and David Rowe <david_at_rowetel_dot_com>
9  *
10  * Copyright (C) 2001 Steve Underwood and 2007 David Rowe
11  *
12  * All rights reserved.
13  *
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.
17  *
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.
22  *
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.
26  */
27 
28 #ifndef __ECHO_H
29 #define __ECHO_H
30 
31 /*
32 Line echo cancellation for voice
33 
34 What does it do?
35 
36 This module aims to provide G.168-2002 compliant echo cancellation, to remove
37 electrical echoes (e.g. from 2-4 wire hybrids) from voice calls.
38 
39 How does it work?
40 
41 The heart of the echo cancellor is FIR filter. This is adapted to match the
42 echo impulse response of the telephone line. It must be long enough to
43 adequately cover the duration of that impulse response. The signal transmitted
44 to the telephone line is passed through the FIR filter. Once the FIR is
45 properly adapted, the resulting output is an estimate of the echo signal
46 received from the line. This is subtracted from the received signal. The result
47 is an estimate of the signal which originated at the far end of the line, free
48 from echos of our own transmitted signal.
49 
50 The least mean squares (LMS) algorithm is attributed to Widrow and Hoff, and
51 was introduced in 1960. It is the commonest form of filter adaption used in
52 things like modem line equalisers and line echo cancellers. There it works very
53 well.  However, it only works well for signals of constant amplitude. It works
54 very poorly for things like speech echo cancellation, where the signal level
55 varies widely.  This is quite easy to fix. If the signal level is normalised -
56 similar to applying AGC - LMS can work as well for a signal of varying
57 amplitude as it does for a modem signal. This normalised least mean squares
58 (NLMS) algorithm is the commonest one used for speech echo cancellation. Many
59 other algorithms exist - e.g. RLS (essentially the same as Kalman filtering),
60 FAP, etc. Some perform significantly better than NLMS.  However, factors such
61 as computational complexity and patents favour the use of NLMS.
62 
63 A simple refinement to NLMS can improve its performance with speech. NLMS tends
64 to adapt best to the strongest parts of a signal. If the signal is white noise,
65 the NLMS algorithm works very well. However, speech has more low frequency than
66 high frequency content. Pre-whitening (i.e. filtering the signal to flatten its
67 spectrum) the echo signal improves the adapt rate for speech, and ensures the
68 final residual signal is not heavily biased towards high frequencies. A very
69 low complexity filter is adequate for this, so pre-whitening adds little to the
70 compute requirements of the echo canceller.
71 
72 An FIR filter adapted using pre-whitened NLMS performs well, provided certain
73 conditions are met:
74 
75     - The transmitted signal has poor self-correlation.
76     - There is no signal being generated within the environment being
77       cancelled.
78 
79 The difficulty is that neither of these can be guaranteed.
80 
81 If the adaption is performed while transmitting noise (or something fairly
82 noise like, such as voice) the adaption works very well. If the adaption is
83 performed while transmitting something highly correlative (typically narrow
84 band energy such as signalling tones or DTMF), the adaption can go seriously
85 wrong. The reason is there is only one solution for the adaption on a near
86 random signal - the impulse response of the line. For a repetitive signal,
87 there are any number of solutions which converge the adaption, and nothing
88 guides the adaption to choose the generalised one. Allowing an untrained
89 canceller to converge on this kind of narrowband energy probably a good thing,
90 since at least it cancels the tones. Allowing a well converged canceller to
91 continue converging on such energy is just a way to ruin its generalised
92 adaption. A narrowband detector is needed, so adapation can be suspended at
93 appropriate times.
94 
95 The adaption process is based on trying to eliminate the received signal. When
96 there is any signal from within the environment being cancelled it may upset
97 the adaption process. Similarly, if the signal we are transmitting is small,
98 noise may dominate and disturb the adaption process. If we can ensure that the
99 adaption is only performed when we are transmitting a significant signal level,
100 and the environment is not, things will be OK. Clearly, it is easy to tell when
101 we are sending a significant signal. Telling, if the environment is generating
102 a significant signal, and doing it with sufficient speed that the adaption will
103 not have diverged too much more we stop it, is a little harder.
104 
105 The key problem in detecting when the environment is sourcing significant
106 energy is that we must do this very quickly. Given a reasonably long sample of
107 the received signal, there are a number of strategies which may be used to
108 assess whether that signal contains a strong far end component. However, by the
109 time that assessment is complete the far end signal will have already caused
110 major mis-convergence in the adaption process. An assessment algorithm is
111 needed which produces a fairly accurate result from a very short burst of far
112 end energy.
113 
114 How do I use it?
115 
116 The echo cancellor processes both the transmit and receive streams sample by
117 sample. The processing function is not declared inline. Unfortunately,
118 cancellation requires many operations per sample, so the call overhead is only
119 a minor burden.
120 */
121 
122 #include "fir.h"
123 #include "oslec.h"
124 
125 /*
126     G.168 echo canceller descriptor. This defines the working state for a line
127     echo canceller.
128 */
129 struct oslec_state {
130 	int16_t tx;
131 	int16_t rx;
132 	int16_t clean;
133 	int16_t clean_nlp;
134 
135 	int nonupdate_dwell;
136 	int curr_pos;
137 	int taps;
138 	int log2taps;
139 	int adaption_mode;
140 
141 	int cond_met;
142 	int32_t pstates;
143 	int16_t adapt;
144 	int32_t factor;
145 	int16_t shift;
146 
147 	/* Average levels and averaging filter states */
148 	int ltxacc;
149 	int lrxacc;
150 	int lcleanacc;
151 	int lclean_bgacc;
152 	int ltx;
153 	int lrx;
154 	int lclean;
155 	int lclean_bg;
156 	int lbgn;
157 	int lbgn_acc;
158 	int lbgn_upper;
159 	int lbgn_upper_acc;
160 
161 	/* foreground and background filter states */
162 	struct fir16_state_t fir_state;
163 	struct fir16_state_t fir_state_bg;
164 	int16_t *fir_taps16[2];
165 
166 	/* DC blocking filter states */
167 	int tx_1;
168 	int tx_2;
169 	int rx_1;
170 	int rx_2;
171 
172 	/* optional High Pass Filter states */
173 	int32_t xvtx[5];
174 	int32_t yvtx[5];
175 	int32_t xvrx[5];
176 	int32_t yvrx[5];
177 
178 	/* Parameters for the optional Hoth noise generator */
179 	int cng_level;
180 	int cng_rndnum;
181 	int cng_filter;
182 
183 	/* snapshot sample of coeffs used for development */
184 	int16_t *snapshot;
185 };
186 
187 #endif /* __ECHO_H */
188