/* Copyright (C) 2003-2006 Jean-Marc Valin
File: mdf.c
Echo canceller based on the MDF algorithm (see below)
Redistribution and use in source and binary forms, with or without
modification, are permitted provided that the following conditions are
met:
1. Redistributions of source code must retain the above copyright notice,
this list of conditions and the following disclaimer.
2. Redistributions in binary form must reproduce the above copyright
notice, this list of conditions and the following disclaimer in the
documentation and/or other materials provided with the distribution.
3. The name of the author may not be used to endorse or promote products
derived from this software without specific prior written permission.
THIS SOFTWARE IS PROVIDED BY THE AUTHOR ``AS IS'' AND ANY EXPRESS OR
IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES
OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE
DISCLAIMED. IN NO EVENT SHALL THE AUTHOR BE LIABLE FOR ANY DIRECT,
INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES
(INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR
SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT,
STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN
ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE
POSSIBILITY OF SUCH DAMAGE.
*/
/*
The echo canceller is based on the MDF algorithm described in:
J. S. Soo, K. K. Pang Multidelay block frequency adaptive filter,
IEEE Trans. Acoust. Speech Signal Process., Vol. ASSP-38, No. 2,
February 1990.
We use the Alternatively Updated MDF (AUMDF) variant. Robustness to
double-talk is achieved using a variable learning rate as described in:
Valin, J.-M., On Adjusting the Learning Rate in Frequency Domain Echo
Cancellation With Double-Talk. Submitted to IEEE Transactions on Speech
and Audio Processing, 2006.
There is no explicit double-talk detection, but a continuous variation
in the learning rate based on residual echo, double-talk and background
noise.
About the fixed-point version:
All the signals are represented with 16-bit words. The filter weights
are represented with 32-bit words, but only the top 16 bits are used
in most cases. The lower 16 bits are completely unreliable (due to the
fact that the update is done only on the top bits), but help in the
adaptation -- probably by removing a "threshold effect" due to
quantization (rounding going to zero) when the gradient is small.
Another kludge that seems to work good: when performing the weight
update, we only move half the way toward the "goal" this seems to
reduce the effect of quantization noise in the update phase. This
can be seen as applying a gradient descent on a "soft constraint"
instead of having a hard constraint.
*/
#ifdef HAVE_CONFIG_H
#include "config.h"
#endif
#include "misc.h"
#include "speex/speex_echo.h"
#include "fftwrap.h"
#include "pseudofloat.h"
#include "math_approx.h"
#ifndef M_PI
#define M_PI 3.14159265358979323846
#endif
#define min(a,b) ((a)<(b) ? (a) : (b))
#define max(a,b) ((a)>(b) ? (a) : (b))
#ifdef FIXED_POINT
#define WEIGHT_SHIFT 11
#define NORMALIZE_SCALEDOWN 5
#define NORMALIZE_SCALEUP 3
#else
#define WEIGHT_SHIFT 0
#endif
#ifdef FIXED_POINT
static const spx_float_t MIN_LEAK = ((spx_float_t){16777, -24});
#define TOP16(x) ((x)>>16)
#else
static const spx_float_t MIN_LEAK = .001f;
#define TOP16(x) (x)
#endif
/** Speex echo cancellation state. */
struct SpeexEchoState_ {
int frame_size; /**< Number of samples processed each time */
int window_size;
int M;
int cancel_count;
int adapted;
spx_int32_t sampling_rate;
spx_word16_t spec_average;
spx_word16_t beta0;
spx_word16_t beta_max;
spx_word32_t sum_adapt;
spx_word16_t *e;
spx_word16_t *x;
spx_word16_t *X;
spx_word16_t *d;
spx_word16_t *y;
spx_word16_t *last_y;
spx_word32_t *Yps;
spx_word16_t *Y;
spx_word16_t *E;
spx_word32_t *PHI;
spx_word32_t *W;
spx_word32_t *power;
spx_float_t *power_1;
spx_word16_t *wtmp;
#ifdef FIXED_POINT
spx_word16_t *wtmp2;
#endif
spx_word32_t *Rf;
spx_word32_t *Yf;
spx_word32_t *Xf;
spx_word32_t *Eh;
spx_word32_t *Yh;
spx_float_t Pey;
spx_float_t Pyy;
spx_word16_t *window;
void *fft_table;
spx_word16_t memX, memD, memE;
spx_word16_t preemph;
spx_word16_t notch_radius;
spx_mem_t notch_mem[2];
};
static inline void filter_dc_notch16(spx_int16_t *in, spx_word16_t radius, spx_word16_t *out, int len, spx_mem_t *mem)
{
int i;
spx_word16_t den2;
#ifdef FIXED_POINT
den2 = MULT16_16_Q15(radius,radius) + MULT16_16_Q15(QCONST16(.7,15),MULT16_16_Q15(32767-radius,32767-radius));
#else
den2 = radius*radius + .7*(1-radius)*(1-radius);
#endif
/*printf ("%d %d %d %d %d %d\n", num[0], num[1], num[2], den[0], den[1], den[2]);*/
for (i=0;i<len;i++)
{
spx_word16_t vin = in[i];
spx_word32_t vout = mem[0] + SHL32(EXTEND32(vin),15);
#ifdef FIXED_POINT
mem[0] = mem[1] + SHL32(SHL32(-EXTEND32(vin),15) + MULT16_32_Q15(radius,vout),1);
#else
mem[0] = mem[1] + 2*(-vin + radius*vout);
#endif
mem[1] = SHL32(EXTEND32(vin),15) - MULT16_32_Q15(den2,vout);
out[i] = SATURATE32(PSHR32(MULT16_32_Q15(radius,vout),15),32767);
}
}
static inline spx_word32_t inner_prod(const spx_word16_t *x, const spx_word16_t *y, int len)
{
spx_word32_t sum=0;
len >>= 2;
while(len--)
{
spx_word32_t part=0;
part = MAC16_16(part,*x++,*y++);
part = MAC16_16(part,*x++,*y++);
part = MAC16_16(part,*x++,*y++);
part = MAC16_16(part,*x++,*y++);
/* HINT: If you had a 40-bit accumulator, you could shift only at the end */
sum = ADD32(sum,SHR32(part,6));
}
return sum;
}
/** Compute power spectrum of a half-complex (packed) vector */
static inline void power_spectrum(spx_word16_t *X, spx_word32_t *ps, int N)
{
int i, j;
ps[0]=MULT16_16(X[0],X[0]);
for (i=1,j=1;i<N-1;i+=2,j++)
{
ps[j] = MULT16_16(X[i],X[i]) + MULT16_16(X[i+1],X[i+1]);
}
ps[j]=MULT16_16(X[i],X[i]);
}
/** Compute cross-power spectrum of a half-complex (packed) vectors and add to acc */
#ifdef FIXED_POINT
static inline void spectral_mul_accum(spx_word16_t *X, spx_word32_t *Y, spx_word16_t *acc, int N, int M)
{
int i,j;
spx_word32_t tmp1=0,tmp2=0;
for (j=0;j<M;j++)
{
tmp1 = MAC16_16(tmp1, X[j*N],TOP16(Y[j*N]));
}
acc[0] = PSHR32(tmp1,WEIGHT_SHIFT);
for (i=1;i<N-1;i+=2)
{
tmp1 = tmp2 = 0;
for (j=0;j<M;j++)
{
tmp1 = SUB32(MAC16_16(tmp1, X[j*N+i],TOP16(Y[j*N+i])), MULT16_16(X[j*N+i+1],TOP16(Y[j*N+i+1])));
tmp2 = MAC16_16(MAC16_16(tmp2, X[j*N+i+1],TOP16(Y[j*N+i])), X[j*N+i], TOP16(Y[j*N+i+1]));
}
acc[i] = PSHR32(tmp1,WEIGHT_SHIFT);
acc[i+1] = PSHR32(tmp2,WEIGHT_SHIFT);
}
tmp1 = tmp2 = 0;
for (j=0;j<M;j++)
{
tmp1 = MAC16_16(tmp1, X[(j+1)*N-1],TOP16(Y[(j+1)*N-1]));
}
acc[N-1] = PSHR32(tmp1,WEIGHT_SHIFT);
}
#else
static inline void spectral_mul_accum(spx_word16_t *X, spx_word32_t *Y, spx_word16_t *acc, int N, int M)
{
int i,j;
for (i=0;i<N;i++)
acc[i] = 0;
for (j=0;j<M;j++)
{
acc[0] += X[0]*Y[0];
for (i=1;i<N-1;i+=2)
{
acc[i] += (X[i]*Y[i] - X[i+1]*Y[i+1]);
acc[i+1] += (X[i+1]*Y[i] + X[i]*Y[i+1]);
}
acc[i] += X[i]*Y[i];
X += N;
Y += N;
}
}
#endif
/** Compute weighted cross-power spectrum of a half-complex (packed) vector with conjugate */
static inline void weighted_spectral_mul_conj(spx_float_t *w, spx_word16_t *X, spx_word16_t *Y, spx_word32_t *prod, int N)
{
int i, j;
prod[0] = FLOAT_MUL32(w[0],MULT16_16(X[0],Y[0]));
for (i=1,j=1;i<N