EXPORT void speex_echo_playback(SpeexEchoState *st, const spx_int16_t *play)
{
/*speex_warning_int("playback with fill level ", st->play_buf_pos/st->frame_size);*/
if (!st->play_buf_started)
{
speex_warning("discarded first playback frame");
return;
}
if (st->play_buf_pos<=PLAYBACK_DELAY*st->frame_size)
{
int i;
for (i=0;i<st->frame_size;i++)
st->play_buf[st->play_buf_pos+i] = play[i];
st->play_buf_pos += st->frame_size;
if (st->play_buf_pos <= (PLAYBACK_DELAY-1)*st->frame_size)
{
speex_warning("Auto-filling the buffer (your application is buggy and/or got xruns)");
for (i=0;i<st->frame_size;i++)
st->play_buf[st->play_buf_pos+i] = play[i];
st->play_buf_pos += st->frame_size;
}
} else {
speex_warning("Had to discard a playback frame (your application is buggy and/or got xruns)");
}
}
void speex_init_header(SpeexHeader *header, int rate, int nb_channels, const SpeexMode *m)
{
int i;
const char *h="Speex ";
/*
strncpy(header->speex_string, "Speex ", 8);
strncpy(header->speex_version, SPEEX_VERSION, SPEEX_HEADER_VERSION_LENGTH-1);
header->speex_version[SPEEX_HEADER_VERSION_LENGTH-1]=0;
*/
for (i=0;i<8;i++)
header->speex_string[i]=h[i];
for (i=0;i<SPEEX_HEADER_VERSION_LENGTH-1 && SPEEX_VERSION[i];i++)
header->speex_version[i]=SPEEX_VERSION[i];
for (;i<SPEEX_HEADER_VERSION_LENGTH;i++)
header->speex_version[i]=0;
header->speex_version_id = 1;
header->header_size = sizeof(SpeexHeader);
header->rate = rate;
header->mode = m->modeID;
header->mode_bitstream_version = m->bitstream_version;
if (m->modeID<0)
speex_warning("This mode is meant to be used alone");
header->nb_channels = nb_channels;
header->bitrate = -1;
speex_mode_query(m, SPEEX_MODE_FRAME_SIZE, &header->frame_size);
header->vbr = 0;
header->frames_per_packet = 0;
header->extra_headers = 0;
header->reserved1 = 0;
header->reserved2 = 0;
}
SpeexHeader *speex_packet_to_header(char *packet, int size)
{
int i;
SpeexHeader *le_header;
const char *h = "Speex ";
for (i=0;i<8;i++)
if (packet[i]!=h[i])
{
speex_warning ("This doesn't look like a Speex file");
return NULL;
}
/*FIXME: Do we allow larger headers?*/
if (size < (int)sizeof(SpeexHeader))
{
speex_warning("Speex header too small");
return NULL;
}
le_header = (SpeexHeader*)speex_alloc(sizeof(SpeexHeader));
speex_move(le_header, packet, sizeof(SpeexHeader));
/*Make sure everything is converted correctly from little-endian*/
ENDIAN_SWITCH(le_header->speex_version_id);
ENDIAN_SWITCH(le_header->header_size);
ENDIAN_SWITCH(le_header->rate);
ENDIAN_SWITCH(le_header->mode);
ENDIAN_SWITCH(le_header->mode_bitstream_version);
ENDIAN_SWITCH(le_header->nb_channels);
ENDIAN_SWITCH(le_header->bitrate);
ENDIAN_SWITCH(le_header->frame_size);
ENDIAN_SWITCH(le_header->vbr);
ENDIAN_SWITCH(le_header->frames_per_packet);
ENDIAN_SWITCH(le_header->extra_headers);
return le_header;
}
EXPORT void speex_echo_capture(SpeexEchoState *st, const spx_int16_t *rec, spx_int16_t *out)
{
int i;
/*speex_warning_int("capture with fill level ", st->play_buf_pos/st->frame_size);*/
st->play_buf_started = 1;
if (st->play_buf_pos>=st->frame_size)
{
speex_echo_cancellation(st, rec, st->play_buf, out);
st->play_buf_pos -= st->frame_size;
for (i=0;i<st->play_buf_pos;i++)
st->play_buf[i] = st->play_buf[i+st->frame_size];
} else {
speex_warning("No playback frame available (your application is buggy and/or got xruns)");
if (st->play_buf_pos!=0)
{
speex_warning("internal playback buffer corruption?");
st->play_buf_pos = 0;
}
for (i=0;i<st->frame_size;i++)
out[i] = rec[i];
}
}
/** Get one packet from the jitter buffer */
int jitter_buffer_get(JitterBuffer *jitter, JitterBufferPacket *packet, spx_uint32_t *start_offset)
{
int i, j;
float late_ratio_short;
float late_ratio_long;
float ontime_ratio_short;
float ontime_ratio_long;
float early_ratio_short;
float early_ratio_long;
int chunk_size;
int incomplete = 0;
if (LT32(jitter->current_timestamp+jitter->tick_size, jitter->pointer_timestamp))
{
jitter->current_timestamp = jitter->pointer_timestamp;
speex_warning("did you forget to call jitter_buffer_tick() by any chance?");
}
/*fprintf (stderr, "get packet %d %d\n", jitter->pointer_timestamp, jitter->current_timestamp);*/
/* FIXME: This should be only what remaining of the current tick */
chunk_size = jitter->tick_size;
/* Compiling arrival statistics */
late_ratio_short = 0;
late_ratio_long = 0;
for (i=0;i<LATE_BINS;i++)
{
late_ratio_short += jitter->shortterm_margin[i];
late_ratio_long += jitter->longterm_margin[i];
}
ontime_ratio_short = jitter->shortterm_margin[LATE_BINS];
ontime_ratio_long = jitter->longterm_margin[LATE_BINS];
early_ratio_short = early_ratio_long = 0;
for (i=LATE_BINS+1;i<MAX_MARGIN;i++)
{
early_ratio_short += jitter->shortterm_margin[i];
early_ratio_long += jitter->longterm_margin[i];
}
if (0&&jitter->pointer_timestamp%1000==0)
{
/*fprintf (stderr, "%f %f %f %f %f %f\n", early_ratio_short, early_ratio_long, ontime_ratio_short, ontime_ratio_long, late_ratio_short, late_ratio_long);*/
/*fprintf (stderr, "%f %f\n", early_ratio_short + ontime_ratio_short + late_ratio_short, early_ratio_long + ontime_ratio_long + late_ratio_long);*/
}
/* Adjusting the buffering */
if (late_ratio_short > .1 || late_ratio_long > .03)
{
/* If too many packets are arriving late */
jitter->shortterm_margin[MAX_MARGIN-1] += jitter->shortterm_margin[MAX_MARGIN-2];
jitter->longterm_margin[MAX_MARGIN-1] += jitter->longterm_margin[MAX_MARGIN-2];
for (i=MAX_MARGIN-3;i>=0;i--)
{
jitter->shortterm_margin[i+1] = jitter->shortterm_margin[i];
jitter->longterm_margin[i+1] = jitter->longterm_margin[i];
}
jitter->shortterm_margin[0] = 0;
jitter->longterm_margin[0] = 0;
jitter->pointer_timestamp -= jitter->tick_size;
jitter->current_timestamp -= jitter->tick_size;
/*fprintf (stderr, "i");*/
/*fprintf (stderr, "interpolate (getting some slack)\n");*/
} else if (late_ratio_short + ontime_ratio_short < .005 && late_ratio_long + ontime_ratio_long < .01 && early_ratio_short > .8)
{
/* Many frames arriving early */
jitter->shortterm_margin[0] += jitter->shortterm_margin[1];
jitter->longterm_margin[0] += jitter->longterm_margin[1];
for (i=1;i<MAX_MARGIN-1;i++)
{
jitter->shortterm_margin[i] = jitter->shortterm_margin[i+1];
jitter->longterm_margin[i] = jitter->longterm_margin[i+1];
}
jitter->shortterm_margin[MAX_MARGIN-1] = 0;
jitter->longterm_margin[MAX_MARGIN-1] = 0;
/*fprintf (stderr, "drop frame\n");*/
/*fprintf (stderr, "d");*/
jitter->pointer_timestamp += jitter->tick_size;
jitter->current_timestamp += jitter->tick_size;
/*fprintf (stderr, "dropping packet (getting more aggressive)\n");*/
}
/* Searching for the packet that fits best */
/* Search the buffer for a packet with the right timestamp and spanning the whole current chunk */
for (i=0;i<SPEEX_JITTER_MAX_BUFFER_SIZE;i++)
{
if (jitter->buf[i] && jitter->timestamp[i]==jitter->pointer_timestamp && GE32(jitter->timestamp[i]+jitter->span[i],jitter->pointer_timestamp+chunk_size))
break;
}
/* If no match, try for an "older" packet that still spans (fully) the current chunk */
if (i==SPEEX_JITTER_MAX_BUFFER_SIZE)
{
for (i=0;i<SPEEX_JITTER_MAX_BUFFER_SIZE;i++)
{
if (jitter->buf[i] && jitter->timestamp[i]<=jitter->pointer_timestamp && GE32(jitter->timestamp[i]+jitter->span[i],jitter->pointer_timestamp+chunk_size))
break;
}
}
/* If still no match, try for an "older" packet that spans part of the current chunk */
if (i==SPEEX_JITTER_MAX_BUFFER_SIZE)
{
for (i=0;i<SPEEX_JITTER_MAX_BUFFER_SIZE;i++)
{
if (jitter->buf[i] && jitter->timestamp[i]<=jitter->pointer_timestamp && GT32(jitter->timestamp[i]+jitter->span[i],jitter->pointer_timestamp))
break;
}
}
/* If still no match, try for earliest packet possible */
if (i==SPEEX_JITTER_MAX_BUFFER_SIZE)
{
int found = 0;
spx_uint32_t best_time=0;
int best_span=0;
int besti=0;
for (i=0;i<SPEEX_JITTER_MAX_BUFFER_SIZE;i++)
{
/* check if packet starts within current chunk */
if (jitter->buf[i] && LT32(jitter->timestamp[i],jitter->pointer_timestamp+chunk_size) && GE32(jitter->timestamp[i],jitter->pointer_timestamp))
{
if (!found || LT32(jitter->timestamp[i],best_time) || (jitter->timestamp[i]==best_time && GT32(jitter->span[i],best_span)))
{
best_time = jitter->timestamp[i];
best_span = jitter->span[i];
besti = i;
found = 1;
}
}
}
if (found)
{
i=besti;
incomplete = 1;
/*fprintf (stderr, "incomplete: %d %d %d %d\n", jitter->timestamp[i], jitter->pointer_timestamp, chunk_size, jitter->span[i]);*/
}
}
/* If we find something */
if (i!=SPEEX_JITTER_MAX_BUFFER_SIZE)
{
/* We (obviously) haven't lost this packet */
jitter->lost_count = 0;
jitter->loss_rate = .999*jitter->loss_rate;
/* Check for potential overflow */
packet->len = jitter->len[i];
/* Copy packet */
for (j=0;j<(int)packet->len;j++)
packet->data[j] = jitter->buf[i][j];
/* Remove packet */
speex_free(jitter->buf[i]);
jitter->buf[i] = NULL;
/* Set timestamp and span (if requested) */
if (start_offset)
*start_offset = jitter->timestamp[i]-jitter->pointer_timestamp;
packet->timestamp = jitter->timestamp[i];
packet->span = jitter->span[i];
/* Point at the end of the current packet */
jitter->pointer_timestamp = jitter->timestamp[i]+jitter->span[i];
if (incomplete)
return JITTER_BUFFER_INCOMPLETE;
else
return JITTER_BUFFER_OK;
}
/* If we haven't found anything worth returning */
/*fprintf (stderr, "not found\n");*/
jitter->lost_count++;
/*fprintf (stderr, "m");*/
/*fprintf (stderr, "lost_count = %d\n", jitter->lost_count);*/
jitter->loss_rate = .999*jitter->loss_rate + .001;
if (start_offset)
*start_offset = 0;
packet->timestamp = jitter->pointer_timestamp;
packet->span = jitter->tick_size;
jitter->pointer_timestamp += chunk_size;
packet->len = 0;
return JITTER_BUFFER_MISSING;
}
/** Performs echo cancellation on a frame */
EXPORT void speex_echo_cancellation(SpeexEchoState *st, const spx_int16_t *in, const spx_int16_t *far_end, spx_int16_t *out)
{
int i,j, chan, speak;
int N,M, C, K;
spx_word32_t Syy,See,Sxx,Sdd, Sff;
#ifdef TWO_PATH
spx_word32_t Dbf;
int update_foreground;
#endif
spx_word32_t Sey;
spx_word16_t ss, ss_1;
spx_float_t Pey = FLOAT_ONE, Pyy=FLOAT_ONE;
spx_float_t alpha, alpha_1;
spx_word16_t RER;
spx_word32_t tmp32;
N = st->window_size;
M = st->M;
C = st->C;
K = st->K;
st->cancel_count++;
#ifdef FIXED_POINT
ss=DIV32_16(11469,M);
ss_1 = SUB16(32767,ss);
#else
ss=.35/M;
ss_1 = 1-ss;
#endif
for (chan = 0; chan < C; chan++)
{
/* Apply a notch filter to make sure DC doesn't end up causing problems */
filter_dc_notch16(in+chan, st->notch_radius, st->input+chan*st->frame_size, st->frame_size, st->notch_mem+2*chan, C);
/* Copy input data to buffer and apply pre-emphasis */
/* Copy input data to buffer */
for (i=0;i<st->frame_size;i++)
{
spx_word32_t tmp32;
/* FIXME: This core has changed a bit, need to merge properly */
tmp32 = SUB32(EXTEND32(st->input[chan*st->frame_size+i]), EXTEND32(MULT16_16_P15(st->preemph, st->memD[chan])));
#ifdef FIXED_POINT
if (tmp32 > 32767)
{
tmp32 = 32767;
if (st->saturated == 0)
st->saturated = 1;
}
if (tmp32 < -32767)
{
tmp32 = -32767;
if (st->saturated == 0)
st->saturated = 1;
}
#endif
st->memD[chan] = st->input[chan*st->frame_size+i];
st->input[chan*st->frame_size+i] = EXTRACT16(tmp32);
}
}
for (speak = 0; speak < K; speak++)
{
for (i=0;i<st->frame_size;i++)
{
spx_word32_t tmp32;
st->x[speak*N+i] = st->x[speak*N+i+st->frame_size];
tmp32 = SUB32(EXTEND32(far_end[i*K+speak]), EXTEND32(MULT16_16_P15(st->preemph, st->memX[speak])));
#ifdef FIXED_POINT
/*FIXME: If saturation occurs here, we need to freeze adaptation for M frames (not just one) */
if (tmp32 > 32767)
{
tmp32 = 32767;
st->saturated = M+1;
}
if (tmp32 < -32767)
{
tmp32 = -32767;
st->saturated = M+1;
}
#endif
st->x[speak*N+i+st->frame_size] = EXTRACT16(tmp32);
st->memX[speak] = far_end[i*K+speak];
}
}
for (speak = 0; speak < K; speak++)
{
/* Shift memory: this could be optimized eventually*/
for (j=M-1;j>=0;j--)
{
for (i=0;i<N;i++)
st->X[(j+1)*N*K+speak*N+i] = st->X[j*N*K+speak*N+i];
}
/* Convert x (echo input) to frequency domain */
spx_fft(st->fft_table, st->x+speak*N, &st->X[speak*N]);
}
Sxx = 0;
for (speak = 0; speak < K; speak++)
{
Sxx += mdf_inner_prod(st->x+speak*N+st->frame_size, st->x+speak*N+st->frame_size, st->frame_size);
power_spectrum_accum(st->X+speak*N, st->Xf, N);
}
Sff = 0;
for (chan = 0; chan < C; chan++)
{
#ifdef TWO_PATH
/* Compute foreground filter */
spectral_mul_accum16(st->X, st->foreground+chan*N*K*M, st->Y+chan*N, N, M*K);
spx_ifft(st->fft_table, st->Y+chan*N, st->e+chan*N);
for (i=0;i<st->frame_size;i++)
st->e[chan*N+i] = SUB16(st->input[chan*st->frame_size+i], st->e[chan*N+i+st->frame_size]);
Sff += mdf_inner_prod(st->e+chan*N, st->e+chan*N, st->frame_size);
#endif
}
/* Adjust proportional adaption rate */
/* FIXME: Adjust that for C, K*/
if (st->adapted)
mdf_adjust_prop (st->W, N, M, C*K, st->prop);
/* Compute weight gradient */
if (st->saturated == 0)
{
for (chan = 0; chan < C; chan++)
{
for (speak = 0; speak < K; speak++)
{
for (j=M-1;j>=0;j--)
{
weighted_spectral_mul_conj(st->power_1, FLOAT_SHL(PSEUDOFLOAT(st->prop[j]),-15), &st->X[(j+1)*N*K+speak*N], st->E+chan*N, st->PHI, N);
for (i=0;i<N;i++)
st->W[chan*N*K*M + j*N*K + speak*N + i] += st->PHI[i];
}
}
}
} else {
st->saturated--;
}
/* FIXME: MC conversion required */
/* Update weight to prevent circular convolution (MDF / AUMDF) */
for (chan = 0; chan < C; chan++)
{
for (speak = 0; speak < K; speak++)
{
for (j=0;j<M;j++)
{
/* This is a variant of the Alternatively Updated MDF (AUMDF) */
/* Remove the "if" to make this an MDF filter */
if (j==0 || st->cancel_count%(M-1) == j-1)
{
#ifdef FIXED_POINT
for (i=0;i<N;i++)
st->wtmp2[i] = EXTRACT16(PSHR32(st->W[chan*N*K*M + j*N*K + speak*N + i],NORMALIZE_SCALEDOWN+16));
spx_ifft(st->fft_table, st->wtmp2, st->wtmp);
for (i=0;i<st->frame_size;i++)
{
st->wtmp[i]=0;
}
for (i=st->frame_size;i<N;i++)
{
st->wtmp[i]=SHL16(st->wtmp[i],NORMALIZE_SCALEUP);
}
spx_fft(st->fft_table, st->wtmp, st->wtmp2);
/* The "-1" in the shift is a sort of kludge that trades less efficient update speed for decrease noise */
for (i=0;i<N;i++)
st->W[chan*N*K*M + j*N*K + speak*N + i] -= SHL32(EXTEND32(st->wtmp2[i]),16+NORMALIZE_SCALEDOWN-NORMALIZE_SCALEUP-1);
#else
spx_ifft(st->fft_table, &st->W[chan*N*K*M + j*N*K + speak*N], st->wtmp);
for (i=st->frame_size;i<N;i++)
{
st->wtmp[i]=0;
}
spx_fft(st->fft_table, st->wtmp, &st->W[chan*N*K*M + j*N*K + speak*N]);
#endif
}
}
}
}
/* So we can use power_spectrum_accum */
for (i=0;i<=st->frame_size;i++)
st->Rf[i] = st->Yf[i] = st->Xf[i] = 0;
Dbf = 0;
See = 0;
#ifdef TWO_PATH
/* Difference in response, this is used to estimate the variance of our residual power estimate */
for (chan = 0; chan < C; chan++)
{
spectral_mul_accum(st->X, st->W+chan*N*K*M, st->Y+chan*N, N, M*K);
spx_ifft(st->fft_table, st->Y+chan*N, st->y+chan*N);
for (i=0;i<st->frame_size;i++)
st->e[chan*N+i] = SUB16(st->e[chan*N+i+st->frame_size], st->y[chan*N+i+st->frame_size]);
Dbf += 10+mdf_inner_prod(st->e+chan*N, st->e+chan*N, st->frame_size);
for (i=0;i<st->frame_size;i++)
st->e[chan*N+i] = SUB16(st->input[chan*st->frame_size+i], st->y[chan*N+i+st->frame_size]);
See += mdf_inner_prod(st->e+chan*N, st->e+chan*N, st->frame_size);
}
#endif
#ifndef TWO_PATH
Sff = See;
#endif
#ifdef TWO_PATH
/* Logic for updating the foreground filter */
/* For two time windows, compute the mean of the energy difference, as well as the variance */
st->Davg1 = ADD32(MULT16_32_Q15(QCONST16(.6f,15),st->Davg1), MULT16_32_Q15(QCONST16(.4f,15),SUB32(Sff,See)));
st->Davg2 = ADD32(MULT16_32_Q15(QCONST16(.85f,15),st->Davg2), MULT16_32_Q15(QCONST16(.15f,15),SUB32(Sff,See)));
st->Dvar1 = FLOAT_ADD(FLOAT_MULT(VAR1_SMOOTH, st->Dvar1), FLOAT_MUL32U(MULT16_32_Q15(QCONST16(.4f,15),Sff), MULT16_32_Q15(QCONST16(.4f,15),Dbf)));
st->Dvar2 = FLOAT_ADD(FLOAT_MULT(VAR2_SMOOTH, st->Dvar2), FLOAT_MUL32U(MULT16_32_Q15(QCONST16(.15f,15),Sff), MULT16_32_Q15(QCONST16(.15f,15),Dbf)));
/* Equivalent float code:
st->Davg1 = .6*st->Davg1 + .4*(Sff-See);
st->Davg2 = .85*st->Davg2 + .15*(Sff-See);
st->Dvar1 = .36*st->Dvar1 + .16*Sff*Dbf;
st->Dvar2 = .7225*st->Dvar2 + .0225*Sff*Dbf;
*/
update_foreground = 0;
/* Check if we have a statistically significant reduction in the residual echo */
/* Note that this is *not* Gaussian, so we need to be careful about the longer tail */
if (FLOAT_GT(FLOAT_MUL32U(SUB32(Sff,See),ABS32(SUB32(Sff,See))), FLOAT_MUL32U(Sff,Dbf)))
update_foreground = 1;
else if (FLOAT_GT(FLOAT_MUL32U(st->Davg1, ABS32(st->Davg1)), FLOAT_MULT(VAR1_UPDATE,(st->Dvar1))))
update_foreground = 1;
else if (FLOAT_GT(FLOAT_MUL32U(st->Davg2, ABS32(st->Davg2)), FLOAT_MULT(VAR2_UPDATE,(st->Dvar2))))
update_foreground = 1;
/* Do we update? */
if (update_foreground)
{
st->Davg1 = st->Davg2 = 0;
st->Dvar1 = st->Dvar2 = FLOAT_ZERO;
/* Copy background filter to foreground filter */
for (i=0;i<N*M*C*K;i++)
st->foreground[i] = EXTRACT16(PSHR32(st->W[i],16));
/* Apply a smooth transition so as to not introduce blocking artifacts */
for (chan = 0; chan < C; chan++)
for (i=0;i<st->frame_size;i++)
st->e[chan*N+i+st->frame_size] = MULT16_16_Q15(st->window[i+st->frame_size],st->e[chan*N+i+st->frame_size]) + MULT16_16_Q15(st->window[i],st->y[chan*N+i+st->frame_size]);
} else {
int reset_background=0;
/* Otherwise, check if the background filter is significantly worse */
if (FLOAT_GT(FLOAT_MUL32U(NEG32(SUB32(Sff,See)),ABS32(SUB32(Sff,See))), FLOAT_MULT(VAR_BACKTRACK,FLOAT_MUL32U(Sff,Dbf))))
reset_background = 1;
if (FLOAT_GT(FLOAT_MUL32U(NEG32(st->Davg1), ABS32(st->Davg1)), FLOAT_MULT(VAR_BACKTRACK,st->Dvar1)))
reset_background = 1;
if (FLOAT_GT(FLOAT_MUL32U(NEG32(st->Davg2), ABS32(st->Davg2)), FLOAT_MULT(VAR_BACKTRACK,st->Dvar2)))
reset_background = 1;
if (reset_background)
{
/* Copy foreground filter to background filter */
for (i=0;i<N*M*C*K;i++)
st->W[i] = SHL32(EXTEND32(st->foreground[i]),16);
/* We also need to copy the output so as to get correct adaptation */
for (chan = 0; chan < C; chan++)
{
for (i=0;i<st->frame_size;i++)
st->y[chan*N+i+st->frame_size] = st->e[chan*N+i+st->frame_size];
for (i=0;i<st->frame_size;i++)
st->e[chan*N+i] = SUB16(st->input[chan*st->frame_size+i], st->y[chan*N+i+st->frame_size]);
}
See = Sff;
st->Davg1 = st->Davg2 = 0;
st->Dvar1 = st->Dvar2 = FLOAT_ZERO;
}
}
#endif
Sey = Syy = Sdd = 0;
for (chan = 0; chan < C; chan++)
{
/* Compute error signal (for the output with de-emphasis) */
for (i=0;i<st->frame_size;i++)
{
spx_word32_t tmp_out;
#ifdef TWO_PATH
tmp_out = SUB32(EXTEND32(st->input[chan*st->frame_size+i]), EXTEND32(st->e[chan*N+i+st->frame_size]));
#else
tmp_out = SUB32(EXTEND32(st->input[chan*st->frame_size+i]), EXTEND32(st->y[chan*N+i+st->frame_size]));
#endif
tmp_out = ADD32(tmp_out, EXTEND32(MULT16_16_P15(st->preemph, st->memE[chan])));
/* This is an arbitrary test for saturation in the microphone signal */
if (in[i*C+chan] <= -32000 || in[i*C+chan] >= 32000)
{
if (st->saturated == 0)
st->saturated = 1;
}
out[i*C+chan] = WORD2INT(tmp_out);
st->memE[chan] = tmp_out;
}
#ifdef DUMP_ECHO_CANCEL_DATA
dump_audio(in, far_end, out, st->frame_size);
#endif
/* Compute error signal (filter update version) */
for (i=0;i<st->frame_size;i++)
{
st->e[chan*N+i+st->frame_size] = st->e[chan*N+i];
st->e[chan*N+i] = 0;
}
/* Compute a bunch of correlations */
/* FIXME: bad merge */
Sey += mdf_inner_prod(st->e+chan*N+st->frame_size, st->y+chan*N+st->frame_size, st->frame_size);
Syy += mdf_inner_prod(st->y+chan*N+st->frame_size, st->y+chan*N+st->frame_size, st->frame_size);
Sdd += mdf_inner_prod(st->input+chan*st->frame_size, st->input+chan*st->frame_size, st->frame_size);
/* Convert error to frequency domain */
spx_fft(st->fft_table, st->e+chan*N, st->E+chan*N);
for (i=0;i<st->frame_size;i++)
st->y[i+chan*N] = 0;
spx_fft(st->fft_table, st->y+chan*N, st->Y+chan*N);
/* Compute power spectrum of echo (X), error (E) and filter response (Y) */
power_spectrum_accum(st->E+chan*N, st->Rf, N);
power_spectrum_accum(st->Y+chan*N, st->Yf, N);
}
/*printf ("%f %f %f %f\n", Sff, See, Syy, Sdd, st->update_cond);*/
/* Do some sanity check */
if (!(Syy>=0 && Sxx>=0 && See >= 0)
#ifndef FIXED_POINT
|| !(Sff < N*1e9 && Syy < N*1e9 && Sxx < N*1e9)
#endif
)
{
/* Things have gone really bad */
st->screwed_up += 50;
for (i=0;i<st->frame_size*C;i++)
out[i] = 0;
} else if (SHR32(Sff, 2) > ADD32(Sdd, SHR32(MULT16_16(N, 10000),6)))
{
/* AEC seems to add lots of echo instead of removing it, let's see if it will improve */
st->screwed_up++;
} else {
/* Everything's fine */
st->screwed_up=0;
}
if (st->screwed_up>=50)
{
speex_warning("The echo canceller started acting funny and got slapped (reset). It swears it will behave now.");
speex_echo_state_reset(st);
return;
}
/* Add a small noise floor to make sure not to have problems when dividing */
See = MAX32(See, SHR32(MULT16_16(N, 100),6));
for (speak = 0; speak < K; speak++)
{
Sxx += mdf_inner_prod(st->x+speak*N+st->frame_size, st->x+speak*N+st->frame_size, st->frame_size);
power_spectrum_accum(st->X+speak*N, st->Xf, N);
}
/* Smooth far end energy estimate over time */
for (j=0;j<=st->frame_size;j++)
st->power[j] = MULT16_32_Q15(ss_1,st->power[j]) + 1 + MULT16_32_Q15(ss,st->Xf[j]);
/* Compute filtered spectra and (cross-)correlations */
for (j=st->frame_size;j>=0;j--)
{
spx_float_t Eh, Yh;
Eh = PSEUDOFLOAT(st->Rf[j] - st->Eh[j]);
Yh = PSEUDOFLOAT(st->Yf[j] - st->Yh[j]);
Pey = FLOAT_ADD(Pey,FLOAT_MULT(Eh,Yh));
Pyy = FLOAT_ADD(Pyy,FLOAT_MULT(Yh,Yh));
#ifdef FIXED_POINT
st->Eh[j] = MAC16_32_Q15(MULT16_32_Q15(SUB16(32767,st->spec_average),st->Eh[j]), st->spec_average, st->Rf[j]);
st->Yh[j] = MAC16_32_Q15(MULT16_32_Q15(SUB16(32767,st->spec_average),st->Yh[j]), st->spec_average, st->Yf[j]);
#else
st->Eh[j] = (1-st->spec_average)*st->Eh[j] + st->spec_average*st->Rf[j];
st->Yh[j] = (1-st->spec_average)*st->Yh[j] + st->spec_average*st->Yf[j];
#endif
}
Pyy = FLOAT_SQRT(Pyy);
Pey = FLOAT_DIVU(Pey,Pyy);
/* Compute correlation updatete rate */
tmp32 = MULT16_32_Q15(st->beta0,Syy);
if (tmp32 > MULT16_32_Q15(st->beta_max,See))
tmp32 = MULT16_32_Q15(st->beta_max,See);
alpha = FLOAT_DIV32(tmp32, See);
alpha_1 = FLOAT_SUB(FLOAT_ONE, alpha);
/* Update correlations (recursive average) */
st->Pey = FLOAT_ADD(FLOAT_MULT(alpha_1,st->Pey) , FLOAT_MULT(alpha,Pey));
st->Pyy = FLOAT_ADD(FLOAT_MULT(alpha_1,st->Pyy) , FLOAT_MULT(alpha,Pyy));
if (FLOAT_LT(st->Pyy, FLOAT_ONE))
st->Pyy = FLOAT_ONE;
/* We don't really hope to get better than 33 dB (MIN_LEAK-3dB) attenuation anyway */
if (FLOAT_LT(st->Pey, FLOAT_MULT(MIN_LEAK,st->Pyy)))
st->Pey = FLOAT_MULT(MIN_LEAK,st->Pyy);
if (FLOAT_GT(st->Pey, st->Pyy))
st->Pey = st->Pyy;
/* leak_estimate is the linear regression result */
st->leak_estimate = FLOAT_EXTRACT16(FLOAT_SHL(FLOAT_DIVU(st->Pey, st->Pyy),14));
/* This looks like a stupid bug, but it's right (because we convert from Q14 to Q15) */
if (st->leak_estimate > 16383)
st->leak_estimate = 32767;
else
st->leak_estimate = SHL16(st->leak_estimate,1);
/*printf ("%f\n", st->leak_estimate);*/
/* Compute Residual to Error Ratio */
#ifdef FIXED_POINT
tmp32 = MULT16_32_Q15(st->leak_estimate,Syy);
tmp32 = ADD32(SHR32(Sxx,13), ADD32(tmp32, SHL32(tmp32,1)));
/* Check for y in e (lower bound on RER) */
{
spx_float_t bound = PSEUDOFLOAT(Sey);
bound = FLOAT_DIVU(FLOAT_MULT(bound, bound), PSEUDOFLOAT(ADD32(1,Syy)));
if (FLOAT_GT(bound, PSEUDOFLOAT(See)))
tmp32 = See;
else if (tmp32 < FLOAT_EXTRACT32(bound))
tmp32 = FLOAT_EXTRACT32(bound);
}
if (tmp32 > SHR32(See,1))
tmp32 = SHR32(See,1);
RER = FLOAT_EXTRACT16(FLOAT_SHL(FLOAT_DIV32(tmp32,See),15));
#else
RER = (.0001*Sxx + 3.*MULT16_32_Q15(st->leak_estimate,Syy)) / See;
/* Check for y in e (lower bound on RER) */
if (RER < Sey*Sey/(1+See*Syy))
RER = Sey*Sey/(1+See*Syy);
if (RER > .5)
RER = .5;
#endif
/* We consider that the filter has had minimal adaptation if the following is true*/
if (!st->adapted && st->sum_adapt > SHL32(EXTEND32(M),15) && MULT16_32_Q15(st->leak_estimate,Syy) > MULT16_32_Q15(QCONST16(.03f,15),Syy))
{
st->adapted = 1;
}
if (st->adapted)
{
/* Normal learning rate calculation once we're past the minimal adaptation phase */
for (i=0;i<=st->frame_size;i++)
{
spx_word32_t r, e;
/* Compute frequency-domain adaptation mask */
r = MULT16_32_Q15(st->leak_estimate,SHL32(st->Yf[i],3));
e = SHL32(st->Rf[i],3)+1;
#ifdef FIXED_POINT
if (r>SHR32(e,1))
r = SHR32(e,1);
#else
if (r>.5*e)
r = .5*e;
#endif
r = MULT16_32_Q15(QCONST16(.7,15),r) + MULT16_32_Q15(QCONST16(.3,15),(spx_word32_t)(MULT16_32_Q15(RER,e)));
/*st->power_1[i] = adapt_rate*r/(e*(1+st->power[i]));*/
st->power_1[i] = FLOAT_SHL(FLOAT_DIV32_FLOAT(r,FLOAT_MUL32U(e,st->power[i]+10)),WEIGHT_SHIFT+16);
}
} else {
/* Temporary adaption rate if filter is not yet adapted enough */
spx_word16_t adapt_rate=0;
if (Sxx > SHR32(MULT16_16(N, 1000),6))
{
tmp32 = MULT16_32_Q15(QCONST16(.25f, 15), Sxx);
#ifdef FIXED_POINT
if (tmp32 > SHR32(See,2))
tmp32 = SHR32(See,2);
#else
if (tmp32 > .25*See)
tmp32 = .25*See;
#endif
adapt_rate = FLOAT_EXTRACT16(FLOAT_SHL(FLOAT_DIV32(tmp32, See),15));
}
for (i=0;i<=st->frame_size;i++)
st->power_1[i] = FLOAT_SHL(FLOAT_DIV32(EXTEND32(adapt_rate),ADD32(st->power[i],10)),WEIGHT_SHIFT+1);
/* How much have we adapted so far? */
st->sum_adapt = ADD32(st->sum_adapt,adapt_rate);
}
/* FIXME: MC conversion required */
for (i=0;i<st->frame_size;i++)
st->last_y[i] = st->last_y[st->frame_size+i];
if (st->adapted)
{
/* If the filter is adapted, take the filtered echo */
for (i=0;i<st->frame_size;i++)
st->last_y[st->frame_size+i] = in[i]-out[i];
} else {
/* If filter isn't adapted yet, all we can do is take the far end signal directly */
/* moved earlier: for (i=0;i<N;i++)
st->last_y[i] = st->x[i];*/
}
}