00001 /* 00002 * SpanDSP - a series of DSP components for telephony 00003 * 00004 * echo.h - An echo cancellor, suitable for electrical and acoustic 00005 * cancellation. This code does not currently comply with 00006 * any relevant standards (e.g. G.164/5/7/8). 00007 * 00008 * Written by Steve Underwood <steveu@coppice.org> 00009 * 00010 * Copyright (C) 2001 Steve Underwood 00011 * 00012 * Based on a bit from here, a bit from there, eye of toad, 00013 * ear of bat, etc - plus, of course, my own 2 cents. 00014 * 00015 * All rights reserved. 00016 * 00017 * This program is free software; you can redistribute it and/or modify 00018 * it under the terms of the GNU General Public License version 2, as 00019 * published by the Free Software Foundation. 00020 * 00021 * This program is distributed in the hope that it will be useful, 00022 * but WITHOUT ANY WARRANTY; without even the implied warranty of 00023 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the 00024 * GNU General Public License for more details. 00025 * 00026 * You should have received a copy of the GNU General Public License 00027 * along with this program; if not, write to the Free Software 00028 * Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA. 00029 * 00030 * $Id: echo.h,v 1.10 2007/04/05 19:20:49 steveu Exp $ 00031 */ 00032 00033 /*! \file */ 00034 00035 #if !defined(_SPANDSP_ECHO_H_) 00036 #define _SPANDSP_ECHO_H_ 00037 00038 /*! \page echo_can_page Line echo cancellation for voice 00039 00040 \section echo_can_page_sec_1 What does it do? 00041 This module aims to provide G.168-2002 compliant echo cancellation, to remove 00042 electrical echoes (e.g. from 2-4 wire hybrids) from voice calls. 00043 00044 \section echo_can_page_sec_2 How does it work? 00045 The heart of the echo cancellor is FIR filter. This is adapted to match the echo 00046 impulse response of the telephone line. It must be long enough to adequately cover 00047 the duration of that impulse response. The signal transmitted to the telephone line 00048 is passed through the FIR filter. Once the FIR is properly adapted, the resulting 00049 output is an estimate of the echo signal received from the line. This is subtracted 00050 from the received signal. The result is an estimate of the signal which originated 00051 at the far end of the line, free from echos of our own transmitted signal. 00052 00053 The least mean squares (LMS) algorithm is attributed to Widrow and Hoff, and was 00054 introduced in 1960. It is the commonest form of filter adaption used in things 00055 like modem line equalisers and line echo cancellers. There it works very well. 00056 However, it only works well for signals of constant amplitude. It works very poorly 00057 for things like speech echo cancellation, where the signal level varies widely. 00058 This is quite easy to fix. If the signal level is normalised - similar to applying 00059 AGC - LMS can work as well for a signal of varying amplitude as it does for a modem 00060 signal. This normalised least mean squares (NLMS) algorithm is the commonest one used 00061 for speech echo cancellation. Many other algorithms exist - e.g. RLS (essentially 00062 the same as Kalman filtering), FAP, etc. Some perform significantly better than NLMS. 00063 However, factors such as computational complexity and patents favour the use of NLMS. 00064 00065 A simple refinement to NLMS can improve its performance with speech. NLMS tends 00066 to adapt best to the strongest parts of a signal. If the signal is white noise, 00067 the NLMS algorithm works very well. However, speech has more low frequency than 00068 high frequency content. Pre-whitening (i.e. filtering the signal to flatten 00069 its spectrum) the echo signal improves the adapt rate for speech, and ensures the 00070 final residual signal is not heavily biased towards high frequencies. A very low 00071 complexity filter is adequate for this, so pre-whitening adds little to the 00072 compute requirements of the echo canceller. 00073 00074 An FIR filter adapted using pre-whitened NLMS performs well, provided certain 00075 conditions are met: 00076 00077 - The transmitted signal has poor self-correlation. 00078 - There is no signal being generated within the environment being cancelled. 00079 00080 The difficulty is that neither of these can be guaranteed. 00081 00082 If the adaption is performed while transmitting noise (or something fairly noise 00083 like, such as voice) the adaption works very well. If the adaption is performed 00084 while transmitting something highly correlative (typically narrow band energy 00085 such as signalling tones or DTMF), the adaption can go seriously wrong. The reason 00086 is there is only one solution for the adaption on a near random signal - the impulse 00087 response of the line. For a repetitive signal, there are any number of solutions 00088 which converge the adaption, and nothing guides the adaption to choose the generalised 00089 one. Allowing an untrained canceller to converge on this kind of narrowband 00090 energy probably a good thing, since at least it cancels the tones. Allowing a well 00091 converged canceller to continue converging on such energy is just a way to ruin 00092 its generalised adaption. A narrowband detector is needed, so adapation can be 00093 suspended at appropriate times. 00094 00095 The adaption process is based on trying to eliminate the received signal. When 00096 there is any signal from within the environment being cancelled it may upset the 00097 adaption process. Similarly, if the signal we are transmitting is small, noise 00098 may dominate and disturb the adaption process. If we can ensure that the 00099 adaption is only performed when we are transmitting a significant signal level, 00100 and the environment is not, things will be OK. Clearly, it is easy to tell when 00101 we are sending a significant signal. Telling, if the environment is generating a 00102 significant signal, and doing it with sufficient speed that the adaption will 00103 not have diverged too much more we stop it, is a little harder. 00104 00105 The key problem in detecting when the environment is sourcing significant energy 00106 is that we must do this very quickly. Given a reasonably long sample of the 00107 received signal, there are a number of strategies which may be used to assess 00108 whether that signal contains a strong far end component. However, by the time 00109 that assessment is complete the far end signal will have already caused major 00110 mis-convergence in the adaption process. An assessment algorithm is needed which 00111 produces a fairly accurate result from a very short burst of far end energy. 00112 00113 \section echo_can_page_sec_3 How do I use it? 00114 The echo cancellor processes both the transmit and receive streams sample by 00115 sample. The processing function is not declared inline. Unfortunately, 00116 cancellation requires many operations per sample, so the call overhead is only a 00117 minor burden. 00118 */ 00119 00120 #include "fir.h" 00121 00122 #define NONUPDATE_DWELL_TIME 600 /* 600 samples, or 75ms */ 00123 00124 /* Mask bits for the adaption mode */ 00125 #define ECHO_CAN_USE_NLP 0x01 00126 #define ECHO_CAN_USE_SUPPRESSOR 0x02 00127 #define ECHO_CAN_USE_CNG 0x04 00128 #define ECHO_CAN_USE_ADAPTION 0x08 00129 00130 /*! 00131 G.168 echo canceller descriptor. This defines the working state for a line 00132 echo canceller. 00133 */ 00134 typedef struct 00135 { 00136 int tx_power[4]; 00137 int rx_power[3]; 00138 int clean_rx_power; 00139 00140 int rx_power_threshold; 00141 int nonupdate_dwell; 00142 00143 fir16_state_t fir_state; 00144 /*! Echo FIR taps (16 bit version) */ 00145 int16_t *fir_taps16[4]; 00146 /*! Echo FIR taps (32 bit version) */ 00147 int32_t *fir_taps32; 00148 00149 int curr_pos; 00150 00151 int taps; 00152 int tap_mask; 00153 int adaption_mode; 00154 00155 int32_t supp_test1; 00156 int32_t supp_test2; 00157 int32_t supp1; 00158 int32_t supp2; 00159 int vad; 00160 int cng; 00161 /* Parameters for the Hoth noise generator */ 00162 int cng_level; 00163 int cng_rndnum; 00164 int cng_filter; 00165 00166 int16_t geigel_max; 00167 int geigel_lag; 00168 int dtd_onset; 00169 int tap_set; 00170 int tap_rotate_counter; 00171 00172 int32_t latest_correction; /* Indication of the magnitude of the latest 00173 adaption, or a code to indicate why adaption 00174 was skipped, for test purposes */ 00175 int32_t last_acf[28]; 00176 int narrowband_count; 00177 int narrowband_score; 00178 } echo_can_state_t; 00179 00180 /*! Create a voice echo canceller context. 00181 \param len The length of the canceller, in samples. 00182 \return The new canceller context, or NULL if the canceller could not be created. 00183 */ 00184 echo_can_state_t *echo_can_create(int len, int adaption_mode); 00185 00186 /*! Free a voice echo canceller context. 00187 \param ec The echo canceller context. 00188 */ 00189 void echo_can_free(echo_can_state_t *ec); 00190 00191 /*! Flush (reinitialise) a voice echo canceller context. 00192 \param ec The echo canceller context. 00193 */ 00194 void echo_can_flush(echo_can_state_t *ec); 00195 00196 /*! Set the adaption mode of a voice echo canceller context. 00197 \param ec The echo canceller context. 00198 \param adapt The mode. 00199 */ 00200 void echo_can_adaption_mode(echo_can_state_t *ec, int adaption_mode); 00201 00202 /*! Process a sample through a voice echo canceller. 00203 \param ec The echo canceller context. 00204 \param tx The transmitted audio sample. 00205 \param rx The received audio sample. 00206 \return The clean (echo cancelled) received sample. 00207 */ 00208 int16_t echo_can_update(echo_can_state_t *ec, int16_t tx, int16_t rx); 00209 00210 #endif 00211 /*- End of file ------------------------------------------------------------*/