/* Find pitch lags */
void SKP_Silk_find_pitch_lags_FIX(
SKP_Silk_encoder_state_FIX *psEnc, /* I/O encoder state */
SKP_Silk_encoder_control_FIX *psEncCtrl, /* I/O encoder control */
SKP_int16 res[], /* O residual */
const SKP_int16 x[] /* I Speech signal */
)
{
SKP_Silk_predict_state_FIX *psPredSt = &psEnc->sPred;
SKP_int buf_len, i, scale;
SKP_int32 thrhld_Q15, res_nrg;
const SKP_int16 *x_buf, *x_buf_ptr;
SKP_int16 Wsig[ FIND_PITCH_LPC_WIN_MAX ], *Wsig_ptr;
SKP_int32 auto_corr[ MAX_FIND_PITCH_LPC_ORDER + 1 ];
SKP_int16 rc_Q15[ MAX_FIND_PITCH_LPC_ORDER ];
SKP_int32 A_Q24[ MAX_FIND_PITCH_LPC_ORDER ];
SKP_int32 FiltState[ MAX_FIND_PITCH_LPC_ORDER ];
SKP_int16 A_Q12[ MAX_FIND_PITCH_LPC_ORDER ];
/******************************************/
/* Setup buffer lengths etc based on Fs */
/******************************************/
buf_len = SKP_ADD_LSHIFT( psEnc->sCmn.la_pitch, psEnc->sCmn.frame_length, 1 );
/* Safty check */
SKP_assert( buf_len >= psPredSt->pitch_LPC_win_length );
x_buf = x - psEnc->sCmn.frame_length;
/*************************************/
/* Estimate LPC AR coefficients */
/*************************************/
/* Calculate windowed signal */
/* First LA_LTP samples */
x_buf_ptr = x_buf + buf_len - psPredSt->pitch_LPC_win_length;
Wsig_ptr = Wsig;
SKP_Silk_apply_sine_window_new( Wsig_ptr, x_buf_ptr, 1, psEnc->sCmn.la_pitch );
/* Middle un - windowed samples */
Wsig_ptr += psEnc->sCmn.la_pitch;
x_buf_ptr += psEnc->sCmn.la_pitch;
SKP_memcpy( Wsig_ptr, x_buf_ptr, ( psPredSt->pitch_LPC_win_length - SKP_LSHIFT( psEnc->sCmn.la_pitch, 1 ) ) * sizeof( SKP_int16 ) );
/* Last LA_LTP samples */
Wsig_ptr += psPredSt->pitch_LPC_win_length - SKP_LSHIFT( psEnc->sCmn.la_pitch, 1 );
x_buf_ptr += psPredSt->pitch_LPC_win_length - SKP_LSHIFT( psEnc->sCmn.la_pitch, 1 );
SKP_Silk_apply_sine_window_new( Wsig_ptr, x_buf_ptr, 2, psEnc->sCmn.la_pitch );
/* Calculate autocorrelation sequence */
SKP_Silk_autocorr( auto_corr, &scale, Wsig, psPredSt->pitch_LPC_win_length, psEnc->sCmn.pitchEstimationLPCOrder + 1 );
/* Add white noise, as fraction of energy */
auto_corr[ 0 ] = SKP_SMLAWB( auto_corr[ 0 ], auto_corr[ 0 ], SKP_FIX_CONST( FIND_PITCH_WHITE_NOISE_FRACTION, 16 ) );
/* Calculate the reflection coefficients using schur */
res_nrg = SKP_Silk_schur( rc_Q15, auto_corr, psEnc->sCmn.pitchEstimationLPCOrder );
/* Prediction gain */
psEncCtrl->predGain_Q16 = SKP_DIV32_varQ( auto_corr[ 0 ], SKP_max_int( res_nrg, 1 ), 16 );
/* Convert reflection coefficients to prediction coefficients */
SKP_Silk_k2a( A_Q24, rc_Q15, psEnc->sCmn.pitchEstimationLPCOrder );
/* Convert From 32 bit Q24 to 16 bit Q12 coefs */
for( i = 0; i < psEnc->sCmn.pitchEstimationLPCOrder; i++ ) {
A_Q12[ i ] = ( SKP_int16 )SKP_SAT16( SKP_RSHIFT( A_Q24[ i ], 12 ) );
}
/* Do BWE */
SKP_Silk_bwexpander( A_Q12, psEnc->sCmn.pitchEstimationLPCOrder, SKP_FIX_CONST( FIND_PITCH_BANDWITH_EXPANSION, 16 ) );
/*****************************************/
/* LPC analysis filtering */
/*****************************************/
SKP_memset( FiltState, 0, psEnc->sCmn.pitchEstimationLPCOrder * sizeof( SKP_int32 ) ); /* Not really necessary, but Valgrind will complain otherwise */
SKP_Silk_MA_Prediction( x_buf, A_Q12, FiltState, res, buf_len, psEnc->sCmn.pitchEstimationLPCOrder );
SKP_memset( res, 0, psEnc->sCmn.pitchEstimationLPCOrder * sizeof( SKP_int16 ) );
/* Threshold for pitch estimator */
thrhld_Q15 = SKP_FIX_CONST( 0.45, 15 );
thrhld_Q15 = SKP_SMLABB( thrhld_Q15, SKP_FIX_CONST( -0.004, 15 ), psEnc->sCmn.pitchEstimationLPCOrder );
thrhld_Q15 = SKP_SMLABB( thrhld_Q15, SKP_FIX_CONST( -0.1, 7 ), psEnc->speech_activity_Q8 );
thrhld_Q15 = SKP_SMLABB( thrhld_Q15, SKP_FIX_CONST( 0.15, 15 ), psEnc->sCmn.prev_sigtype );
thrhld_Q15 = SKP_SMLAWB( thrhld_Q15, SKP_FIX_CONST( -0.1, 16 ), psEncCtrl->input_tilt_Q15 );
thrhld_Q15 = SKP_SAT16( thrhld_Q15 );
/*****************************************/
/* Call pitch estimator */
/*****************************************/
psEncCtrl->sCmn.sigtype = SKP_Silk_pitch_analysis_core( res, psEncCtrl->sCmn.pitchL, &psEncCtrl->sCmn.lagIndex,
&psEncCtrl->sCmn.contourIndex, &psEnc->LTPCorr_Q15, psEnc->sCmn.prevLag, psEnc->sCmn.pitchEstimationThreshold_Q16,
( SKP_int16 )thrhld_Q15, psEnc->sCmn.fs_kHz, psEnc->sCmn.pitchEstimationComplexity, SKP_FALSE );
}
/* Find pitch lags */
void SKP_Silk_find_pitch_lags_FIX(
SKP_Silk_encoder_state_FIX *psEnc, /* I/O encoder state */
SKP_Silk_encoder_control_FIX *psEncCtrl, /* I/O encoder control */
SKP_int16 res[], /* O residual */
const SKP_int16 x[] /* I Speech signal */
)
{
SKP_Silk_predict_state_FIX *psPredSt = &psEnc->sPred;
SKP_int buf_len, i;
SKP_int32 scale;
SKP_int32 thrhld_Q15;
const SKP_int16 *x_buf, *x_buf_ptr;
SKP_int16 Wsig[ FIND_PITCH_LPC_WIN_MAX ], *Wsig_ptr;
SKP_int32 auto_corr[ FIND_PITCH_LPC_ORDER_MAX + 1 ];
SKP_int16 rc_Q15[ FIND_PITCH_LPC_ORDER_MAX ];
SKP_int32 A_Q24[ FIND_PITCH_LPC_ORDER_MAX ];
SKP_int32 FiltState[ FIND_PITCH_LPC_ORDER_MAX ];
SKP_int16 A_Q12[ FIND_PITCH_LPC_ORDER_MAX ];
/******************************************/
/* Setup buffer lengths etc based of Fs. */
/******************************************/
buf_len = SKP_ADD_LSHIFT( psEnc->sCmn.la_pitch, psEnc->sCmn.frame_length, 1 );
/* Safty check */
SKP_assert( buf_len >= psPredSt->pitch_LPC_win_length );
x_buf = x - psEnc->sCmn.frame_length;
/*************************************/
/* Estimate LPC AR coeficients */
/*************************************/
/* Calculate windowed signal */
/* First LA_LTP samples */
x_buf_ptr = x_buf + buf_len - psPredSt->pitch_LPC_win_length;
Wsig_ptr = Wsig;
SKP_Silk_apply_sine_window( Wsig_ptr, x_buf_ptr, 1, psEnc->sCmn.la_pitch );
/* Middle un - windowed samples */
Wsig_ptr += psEnc->sCmn.la_pitch;
x_buf_ptr += psEnc->sCmn.la_pitch;
SKP_memcpy( Wsig_ptr, x_buf_ptr, ( psPredSt->pitch_LPC_win_length - SKP_LSHIFT( psEnc->sCmn.la_pitch, 1 ) ) * sizeof( SKP_int16 ) );
/* Last LA_LTP samples */
Wsig_ptr += psPredSt->pitch_LPC_win_length - SKP_LSHIFT( psEnc->sCmn.la_pitch, 1 );
x_buf_ptr += psPredSt->pitch_LPC_win_length - SKP_LSHIFT( psEnc->sCmn.la_pitch, 1 );
SKP_Silk_apply_sine_window( Wsig_ptr, x_buf_ptr, 2, psEnc->sCmn.la_pitch );
/* Calculate autocorrelation sequence */
SKP_Silk_autocorr( auto_corr, &scale, Wsig, psPredSt->pitch_LPC_win_length, psEnc->sCmn.pitchEstimationLPCOrder + 1 );
/* add white noise, as fraction of energy */
auto_corr[ 0 ] = SKP_SMLAWB( auto_corr[ 0 ], auto_corr[ 0 ], FIND_PITCH_WHITE_NOISE_FRACTION_Q16 );
/* calculate the reflection coefficients using schur */
SKP_Silk_schur( rc_Q15, auto_corr, psEnc->sCmn.pitchEstimationLPCOrder );
/* convert reflection coefficients to prediction coefficients */
SKP_Silk_k2a( A_Q24, rc_Q15, psEnc->sCmn.pitchEstimationLPCOrder );
/* Convert From 32 bit Q24 to 16 bit Q12 coefs */
for( i = 0; i < psEnc->sCmn.pitchEstimationLPCOrder; i++ ) {
A_Q12[ i ] = ( SKP_int16 )SKP_SAT16( SKP_RSHIFT( A_Q24[ i ], 12 ) );
}
/* Do BWE */
SKP_Silk_bwexpander( A_Q12, psEnc->sCmn.pitchEstimationLPCOrder, FIND_PITCH_BANDWITH_EXPANSION_Q16 );
/*****************************************/
/* LPC analysis filtering */
/*****************************************/
SKP_memset( FiltState, 0, psEnc->sCmn.pitchEstimationLPCOrder * sizeof( SKP_int16 ) );
SKP_Silk_MA_Prediction( x_buf, A_Q12, FiltState, res, buf_len, psEnc->sCmn.pitchEstimationLPCOrder );
SKP_memset( res, 0, psEnc->sCmn.pitchEstimationLPCOrder * sizeof( SKP_int16 ) );
/* Threshold for pitch estimator */
thrhld_Q15 = ( 1 << 14 ); // 0.5f in Q15
thrhld_Q15 = SKP_SMLABB( thrhld_Q15, -131, psEnc->sCmn.pitchEstimationLPCOrder );
thrhld_Q15 = SKP_SMLABB( thrhld_Q15, -13, ( SKP_int16 )SKP_Silk_SQRT_APPROX( SKP_LSHIFT( ( SKP_int32 )psEnc->speech_activity_Q8, 8 ) ) );
thrhld_Q15 = SKP_SMLABB( thrhld_Q15, 4587, psEnc->sCmn.prev_sigtype );
thrhld_Q15 = SKP_MLA( thrhld_Q15, -31, SKP_RSHIFT( psEncCtrl->input_tilt_Q15, 8 ) );
thrhld_Q15 = SKP_SAT16( thrhld_Q15 );
/*****************************************/
/* Call Pitch estimator */
/*****************************************/
psEncCtrl->sCmn.sigtype = SKP_Silk_pitch_analysis_core( res, psEncCtrl->sCmn.pitchL, &psEncCtrl->sCmn.lagIndex,
&psEncCtrl->sCmn.contourIndex, &psEnc->LTPCorr_Q15, psEnc->sCmn.prevLag, psEnc->pitchEstimationThreshold_Q16,
( SKP_int16 )thrhld_Q15, psEnc->sCmn.fs_kHz, psEnc->sCmn.pitchEstimationComplexity );
}
void SKP_Silk_noise_shape_analysis_FIX(
SKP_Silk_encoder_state_FIX *psEnc, /* I/O Encoder state FIX */
SKP_Silk_encoder_control_FIX *psEncCtrl, /* I/O Encoder control FIX */
const SKP_int16 *pitch_res, /* I LPC residual from pitch analysis */
const SKP_int16 *x /* I Input signal [ frame_length + la_shape ] */
)
{
SKP_Silk_shape_state_FIX *psShapeSt = &psEnc->sShape;
SKP_int k, i, nSamples, Qnrg, b_Q14, warping_Q16, scale = 0;
SKP_int32 SNR_adj_dB_Q7, HarmBoost_Q16, HarmShapeGain_Q16, Tilt_Q16, tmp32;
SKP_int32 nrg, pre_nrg_Q30, log_energy_Q7, log_energy_prev_Q7, energy_variation_Q7;
SKP_int32 delta_Q16, BWExp1_Q16, BWExp2_Q16, gain_mult_Q16, gain_add_Q16, strength_Q16, b_Q8;
SKP_int32 auto_corr[ MAX_SHAPE_LPC_ORDER + 1 ];
SKP_int32 refl_coef_Q16[ MAX_SHAPE_LPC_ORDER ];
SKP_int32 AR1_Q24[ MAX_SHAPE_LPC_ORDER ];
SKP_int32 AR2_Q24[ MAX_SHAPE_LPC_ORDER ];
SKP_int16 x_windowed[ SHAPE_LPC_WIN_MAX ];
const SKP_int16 *x_ptr, *pitch_res_ptr;
SKP_int32 sqrt_nrg[ NB_SUBFR ], Qnrg_vec[ NB_SUBFR ];
/* Point to start of first LPC analysis block */
x_ptr = x - psEnc->sCmn.la_shape;
/****************/
/* CONTROL SNR */
/****************/
/* Reduce SNR_dB values if recent bitstream has exceeded TargetRate */
psEncCtrl->current_SNR_dB_Q7 = psEnc->SNR_dB_Q7 - SKP_SMULWB( SKP_LSHIFT( ( SKP_int32 )psEnc->BufferedInChannel_ms, 7 ),
SKP_FIX_CONST( 0.05, 16 ) );
/* Reduce SNR_dB if inband FEC used */
if( psEnc->speech_activity_Q8 > SKP_FIX_CONST( LBRR_SPEECH_ACTIVITY_THRES, 8 ) ) {
psEncCtrl->current_SNR_dB_Q7 -= SKP_RSHIFT( psEnc->inBandFEC_SNR_comp_Q8, 1 );
}
/****************/
/* GAIN CONTROL */
/****************/
/* Input quality is the average of the quality in the lowest two VAD bands */
psEncCtrl->input_quality_Q14 = ( SKP_int )SKP_RSHIFT( ( SKP_int32 )psEncCtrl->input_quality_bands_Q15[ 0 ]
+ psEncCtrl->input_quality_bands_Q15[ 1 ], 2 );
/* Coding quality level, between 0.0_Q0 and 1.0_Q0, but in Q14 */
psEncCtrl->coding_quality_Q14 = SKP_RSHIFT( SKP_Silk_sigm_Q15( SKP_RSHIFT_ROUND( psEncCtrl->current_SNR_dB_Q7 -
SKP_FIX_CONST( 18.0, 7 ), 4 ) ), 1 );
/* Reduce coding SNR during low speech activity */
b_Q8 = SKP_FIX_CONST( 1.0, 8 ) - psEnc->speech_activity_Q8;
b_Q8 = SKP_SMULWB( SKP_LSHIFT( b_Q8, 8 ), b_Q8 );
SNR_adj_dB_Q7 = SKP_SMLAWB( psEncCtrl->current_SNR_dB_Q7,
SKP_SMULBB( SKP_FIX_CONST( -BG_SNR_DECR_dB, 7 ) >> ( 4 + 1 ), b_Q8 ), // Q11
SKP_SMULWB( SKP_FIX_CONST( 1.0, 14 ) + psEncCtrl->input_quality_Q14, psEncCtrl->coding_quality_Q14 ) ); // Q12
if( psEncCtrl->sCmn.sigtype == SIG_TYPE_VOICED ) {
/* Reduce gains for periodic signals */
SNR_adj_dB_Q7 = SKP_SMLAWB( SNR_adj_dB_Q7, SKP_FIX_CONST( HARM_SNR_INCR_dB, 8 ), psEnc->LTPCorr_Q15 );
} else {
/* For unvoiced signals and low-quality input, adjust the quality slower than SNR_dB setting */
SNR_adj_dB_Q7 = SKP_SMLAWB( SNR_adj_dB_Q7,
SKP_SMLAWB( SKP_FIX_CONST( 6.0, 9 ), -SKP_FIX_CONST( 0.4, 18 ), psEncCtrl->current_SNR_dB_Q7 ),
SKP_FIX_CONST( 1.0, 14 ) - psEncCtrl->input_quality_Q14 );
}
/*************************/
/* SPARSENESS PROCESSING */
/*************************/
/* Set quantizer offset */
if( psEncCtrl->sCmn.sigtype == SIG_TYPE_VOICED ) {
/* Initally set to 0; may be overruled in process_gains(..) */
psEncCtrl->sCmn.QuantOffsetType = 0;
psEncCtrl->sparseness_Q8 = 0;
} else {
/* Sparseness measure, based on relative fluctuations of energy per 2 milliseconds */
nSamples = SKP_LSHIFT( psEnc->sCmn.fs_kHz, 1 );
energy_variation_Q7 = 0;
log_energy_prev_Q7 = 0;
pitch_res_ptr = pitch_res;
for( k = 0; k < FRAME_LENGTH_MS / 2; k++ ) {
SKP_Silk_sum_sqr_shift( &nrg, &scale, pitch_res_ptr, nSamples );
nrg += SKP_RSHIFT( nSamples, scale ); // Q(-scale)
log_energy_Q7 = SKP_Silk_lin2log( nrg );
if( k > 0 ) {
energy_variation_Q7 += SKP_abs( log_energy_Q7 - log_energy_prev_Q7 );
}
log_energy_prev_Q7 = log_energy_Q7;
pitch_res_ptr += nSamples;
}
psEncCtrl->sparseness_Q8 = SKP_RSHIFT( SKP_Silk_sigm_Q15( SKP_SMULWB( energy_variation_Q7 -
SKP_FIX_CONST( 5.0, 7 ), SKP_FIX_CONST( 0.1, 16 ) ) ), 7 );
/* Set quantization offset depending on sparseness measure */
if( psEncCtrl->sparseness_Q8 > SKP_FIX_CONST( SPARSENESS_THRESHOLD_QNT_OFFSET, 8 ) ) {
psEncCtrl->sCmn.QuantOffsetType = 0;
} else {
psEncCtrl->sCmn.QuantOffsetType = 1;
}
/* Increase coding SNR for sparse signals */
SNR_adj_dB_Q7 = SKP_SMLAWB( SNR_adj_dB_Q7, SKP_FIX_CONST( SPARSE_SNR_INCR_dB, 15 ), psEncCtrl->sparseness_Q8 - SKP_FIX_CONST( 0.5, 8 ) );
}
/*******************************/
/* Control bandwidth expansion */
/*******************************/
/* More BWE for signals with high prediction gain */
strength_Q16 = SKP_SMULWB( psEncCtrl->predGain_Q16, SKP_FIX_CONST( FIND_PITCH_WHITE_NOISE_FRACTION, 16 ) );
BWExp1_Q16 = BWExp2_Q16 = SKP_DIV32_varQ( SKP_FIX_CONST( BANDWIDTH_EXPANSION, 16 ),
SKP_SMLAWW( SKP_FIX_CONST( 1.0, 16 ), strength_Q16, strength_Q16 ), 16 );
delta_Q16 = SKP_SMULWB( SKP_FIX_CONST( 1.0, 16 ) - SKP_SMULBB( 3, psEncCtrl->coding_quality_Q14 ),
SKP_FIX_CONST( LOW_RATE_BANDWIDTH_EXPANSION_DELTA, 16 ) );
BWExp1_Q16 = SKP_SUB32( BWExp1_Q16, delta_Q16 );
BWExp2_Q16 = SKP_ADD32( BWExp2_Q16, delta_Q16 );
/* BWExp1 will be applied after BWExp2, so make it relative */
BWExp1_Q16 = SKP_DIV32_16( SKP_LSHIFT( BWExp1_Q16, 14 ), SKP_RSHIFT( BWExp2_Q16, 2 ) );
if( psEnc->sCmn.warping_Q16 > 0 ) {
/* Slightly more warping in analysis will move quantization noise up in frequency, where it's better masked */
warping_Q16 = SKP_SMLAWB( psEnc->sCmn.warping_Q16, psEncCtrl->coding_quality_Q14, SKP_FIX_CONST( 0.01, 18 ) );
} else {
warping_Q16 = 0;
}
/********************************************/
/* Compute noise shaping AR coefs and gains */
/********************************************/
for( k = 0; k < NB_SUBFR; k++ ) {
/* Apply window: sine slope followed by flat part followed by cosine slope */
SKP_int shift, slope_part, flat_part;
flat_part = psEnc->sCmn.fs_kHz * 5;
slope_part = SKP_RSHIFT( psEnc->sCmn.shapeWinLength - flat_part, 1 );
SKP_Silk_apply_sine_window_new( x_windowed, x_ptr, 1, slope_part );
shift = slope_part;
SKP_memcpy( x_windowed + shift, x_ptr + shift, flat_part * sizeof(SKP_int16) );
shift += flat_part;
SKP_Silk_apply_sine_window_new( x_windowed + shift, x_ptr + shift, 2, slope_part );
/* Update pointer: next LPC analysis block */
x_ptr += psEnc->sCmn.subfr_length;
if( psEnc->sCmn.warping_Q16 > 0 ) {
/* Calculate warped auto correlation */
SKP_Silk_warped_autocorrelation_FIX( auto_corr, &scale, x_windowed, warping_Q16, psEnc->sCmn.shapeWinLength, psEnc->sCmn.shapingLPCOrder );
} else {
/* Calculate regular auto correlation */
SKP_Silk_autocorr( auto_corr, &scale, x_windowed, psEnc->sCmn.shapeWinLength, psEnc->sCmn.shapingLPCOrder + 1 );
}
/* Add white noise, as a fraction of energy */
auto_corr[0] = SKP_ADD32( auto_corr[0], SKP_max_32( SKP_SMULWB( SKP_RSHIFT( auto_corr[ 0 ], 4 ),
SKP_FIX_CONST( SHAPE_WHITE_NOISE_FRACTION, 20 ) ), 1 ) );
/* Calculate the reflection coefficients using schur */
nrg = SKP_Silk_schur64( refl_coef_Q16, auto_corr, psEnc->sCmn.shapingLPCOrder );
SKP_assert( nrg >= 0 );
/* Convert reflection coefficients to prediction coefficients */
SKP_Silk_k2a_Q16( AR2_Q24, refl_coef_Q16, psEnc->sCmn.shapingLPCOrder );
Qnrg = -scale; // range: -12...30
SKP_assert( Qnrg >= -12 );
SKP_assert( Qnrg <= 30 );
/* Make sure that Qnrg is an even number */
if( Qnrg & 1 ) {
Qnrg -= 1;
nrg >>= 1;
}
tmp32 = SKP_Silk_SQRT_APPROX( nrg );
Qnrg >>= 1; // range: -6...15
sqrt_nrg[ k ] = tmp32;
Qnrg_vec[ k ] = Qnrg;
psEncCtrl->Gains_Q16[ k ] = SKP_LSHIFT_SAT32( tmp32, 16 - Qnrg );
if( psEnc->sCmn.warping_Q16 > 0 ) {
/* Adjust gain for warping */
gain_mult_Q16 = warped_gain( AR2_Q24, warping_Q16, psEnc->sCmn.shapingLPCOrder );
SKP_assert( psEncCtrl->Gains_Q16[ k ] >= 0 );
psEncCtrl->Gains_Q16[ k ] = SKP_SMULWW( psEncCtrl->Gains_Q16[ k ], gain_mult_Q16 );
if( psEncCtrl->Gains_Q16[ k ] < 0 ) {
psEncCtrl->Gains_Q16[ k ] = SKP_int32_MAX;
}
}
/* Bandwidth expansion for synthesis filter shaping */
SKP_Silk_bwexpander_32( AR2_Q24, psEnc->sCmn.shapingLPCOrder, BWExp2_Q16 );
/* Compute noise shaping filter coefficients */
SKP_memcpy( AR1_Q24, AR2_Q24, psEnc->sCmn.shapingLPCOrder * sizeof( SKP_int32 ) );
/* Bandwidth expansion for analysis filter shaping */
SKP_assert( BWExp1_Q16 <= SKP_FIX_CONST( 1.0, 16 ) );
SKP_Silk_bwexpander_32( AR1_Q24, psEnc->sCmn.shapingLPCOrder, BWExp1_Q16 );
/* Ratio of prediction gains, in energy domain */
SKP_Silk_LPC_inverse_pred_gain_Q24( &pre_nrg_Q30, AR2_Q24, psEnc->sCmn.shapingLPCOrder );
SKP_Silk_LPC_inverse_pred_gain_Q24( &nrg, AR1_Q24, psEnc->sCmn.shapingLPCOrder );
//psEncCtrl->GainsPre[ k ] = 1.0f - 0.7f * ( 1.0f - pre_nrg / nrg ) = 0.3f + 0.7f * pre_nrg / nrg;
pre_nrg_Q30 = SKP_LSHIFT32( SKP_SMULWB( pre_nrg_Q30, SKP_FIX_CONST( 0.7, 15 ) ), 1 );
psEncCtrl->GainsPre_Q14[ k ] = ( SKP_int ) SKP_FIX_CONST( 0.3, 14 ) + SKP_DIV32_varQ( pre_nrg_Q30, nrg, 14 );
/* Convert to monic warped prediction coefficients and limit absolute values */
limit_warped_coefs( AR2_Q24, AR1_Q24, warping_Q16, SKP_FIX_CONST( 3.999, 24 ), psEnc->sCmn.shapingLPCOrder );
/* Convert from Q24 to Q13 and store in int16 */
for( i = 0; i < psEnc->sCmn.shapingLPCOrder; i++ ) {
psEncCtrl->AR1_Q13[ k * MAX_SHAPE_LPC_ORDER + i ] = (SKP_int16)SKP_SAT16( SKP_RSHIFT_ROUND( AR1_Q24[ i ], 11 ) );
psEncCtrl->AR2_Q13[ k * MAX_SHAPE_LPC_ORDER + i ] = (SKP_int16)SKP_SAT16( SKP_RSHIFT_ROUND( AR2_Q24[ i ], 11 ) );
}
}
void SKP_Silk_noise_shape_analysis_FIX(
SKP_Silk_encoder_state_FIX *psEnc, /* I/O Encoder state FIX */
SKP_Silk_encoder_control_FIX *psEncCtrl, /* I/O Encoder control FIX */
const SKP_int16 *pitch_res, /* I LPC residual from pitch analysis */
const SKP_int16 *x /* I Input signal [ 2 * frame_length + la_shape ]*/
)
{
SKP_Silk_shape_state_FIX *psShapeSt = &psEnc->sShape;
SKP_int k, nSamples, lz, Qnrg, b_Q14, scale = 0, sz;
SKP_int32 SNR_adj_dB_Q7, HarmBoost_Q16, HarmShapeGain_Q16, Tilt_Q16, tmp32;
SKP_int32 nrg, pre_nrg_Q30, log_energy_Q7, log_energy_prev_Q7, energy_variation_Q7;
SKP_int32 delta_Q16, BWExp1_Q16, BWExp2_Q16, gain_mult_Q16, gain_add_Q16, strength_Q16, b_Q8;
SKP_int32 auto_corr[ SHAPE_LPC_ORDER_MAX + 1 ];
SKP_int32 refl_coef_Q16[ SHAPE_LPC_ORDER_MAX ];
SKP_int32 AR_Q24[ SHAPE_LPC_ORDER_MAX ];
SKP_int16 x_windowed[ SHAPE_LPC_WIN_MAX ];
const SKP_int16 *x_ptr, *pitch_res_ptr;
SKP_int32 sqrt_nrg[ NB_SUBFR ], Qnrg_vec[ NB_SUBFR ];
/* Point to start of first LPC analysis block */
x_ptr = x + psEnc->sCmn.la_shape - SKP_SMULBB( SHAPE_LPC_WIN_MS, psEnc->sCmn.fs_kHz ) + psEnc->sCmn.frame_length / NB_SUBFR;
/****************/
/* CONTROL SNR */
/****************/
/* Reduce SNR_dB values if recent bitstream has exceeded TargetRate */
psEncCtrl->current_SNR_dB_Q7 = psEnc->SNR_dB_Q7 - SKP_SMULWB( SKP_LSHIFT( ( SKP_int32 )psEnc->BufferedInChannel_ms, 7 ), 3277 );
/* Reduce SNR_dB if inband FEC used */
if( psEnc->speech_activity_Q8 > LBRR_SPEECH_ACTIVITY_THRES_Q8 ) {
psEncCtrl->current_SNR_dB_Q7 -= SKP_RSHIFT( psEnc->inBandFEC_SNR_comp_Q8, 1 );
}
/****************/
/* GAIN CONTROL */
/****************/
/* Input quality is the average of the quality in the lowest two VAD bands */
psEncCtrl->input_quality_Q14 = ( SKP_int )SKP_RSHIFT( ( SKP_int32 )psEncCtrl->input_quality_bands_Q15[ 0 ]
+ psEncCtrl->input_quality_bands_Q15[ 1 ], 2 );
/* Coding quality level, between 0.0_Q0 and 1.0_Q0, but in Q14 */
psEncCtrl->coding_quality_Q14 = SKP_RSHIFT( SKP_Silk_sigm_Q15( SKP_RSHIFT_ROUND( psEncCtrl->current_SNR_dB_Q7 - ( 18 << 7 ), 4 ) ), 1 );
/* Reduce coding SNR during low speech activity */
b_Q8 = ( 1 << 8 ) - psEnc->speech_activity_Q8;
b_Q8 = SKP_SMULWB( SKP_LSHIFT( b_Q8, 8 ), b_Q8 );
SNR_adj_dB_Q7 = SKP_SMLAWB( psEncCtrl->current_SNR_dB_Q7,
SKP_SMULBB( -BG_SNR_DECR_dB_Q7 >> ( 4 + 1 ), b_Q8 ), // Q11
SKP_SMULWB( ( 1 << 14 ) + psEncCtrl->input_quality_Q14, psEncCtrl->coding_quality_Q14 ) ); // Q12
if( psEncCtrl->sCmn.sigtype == SIG_TYPE_VOICED ) {
/* Reduce gains for periodic signals */
SNR_adj_dB_Q7 = SKP_SMLAWB( SNR_adj_dB_Q7, HARM_SNR_INCR_dB_Q7 << 1, psEnc->LTPCorr_Q15 );
} else {
/* For unvoiced signals and low-quality input, adjust the quality slower than SNR_dB setting */
SNR_adj_dB_Q7 = SKP_SMLAWB( SNR_adj_dB_Q7,
SKP_SMLAWB( 6 << ( 7 + 2 ), -104856, psEncCtrl->current_SNR_dB_Q7 ), //-104856_Q18 = -0.4_Q0, Q9
( 1 << 14 ) - psEncCtrl->input_quality_Q14 ); // Q14
}
/*************************/
/* SPARSENESS PROCESSING */
/*************************/
/* Set quantizer offset */
if( psEncCtrl->sCmn.sigtype == SIG_TYPE_VOICED ) {
/* Initally set to 0; may be overruled in process_gains(..) */
psEncCtrl->sCmn.QuantOffsetType = 0;
psEncCtrl->sparseness_Q8 = 0;
} else {
/* Sparseness measure, based on relative fluctuations of energy per 2 milliseconds */
nSamples = SKP_LSHIFT( psEnc->sCmn.fs_kHz, 1 );
energy_variation_Q7 = 0;
log_energy_prev_Q7 = 0;
pitch_res_ptr = pitch_res;
for( k = 0; k < FRAME_LENGTH_MS / 2; k++ ) {
SKP_Silk_sum_sqr_shift( &nrg, &scale, pitch_res_ptr, nSamples );
nrg += SKP_RSHIFT( nSamples, scale ); // Q(-scale)
log_energy_Q7 = SKP_Silk_lin2log( nrg );
if( k > 0 ) {
energy_variation_Q7 += SKP_abs( log_energy_Q7 - log_energy_prev_Q7 );
}
log_energy_prev_Q7 = log_energy_Q7;
pitch_res_ptr += nSamples;
}
psEncCtrl->sparseness_Q8 = SKP_RSHIFT( SKP_Silk_sigm_Q15( SKP_SMULWB( energy_variation_Q7 - ( 5 << 7 ), 6554 ) ), 7 ); // 6554_Q16 = 0.1_Q0
/* Set quantization offset depending on sparseness measure */
if( psEncCtrl->sparseness_Q8 > SPARSENESS_THRESHOLD_QNT_OFFSET_Q8 ) {
psEncCtrl->sCmn.QuantOffsetType = 0;
} else {
psEncCtrl->sCmn.QuantOffsetType = 1;
}
/* Increase coding SNR for sparse signals */
SNR_adj_dB_Q7 = SKP_SMLAWB( SNR_adj_dB_Q7, SPARSE_SNR_INCR_dB_Q7 << 8, psEncCtrl->sparseness_Q8 - ( 1 << 7 ) );
}
/*******************************/
/* Control bandwidth expansion */
/*******************************/
delta_Q16 = SKP_SMULWB( ( 1 << 16 ) - SKP_SMULBB( 3, psEncCtrl->coding_quality_Q14 ), LOW_RATE_BANDWIDTH_EXPANSION_DELTA_Q16 );
BWExp1_Q16 = BANDWIDTH_EXPANSION_Q16 - delta_Q16;
BWExp2_Q16 = BANDWIDTH_EXPANSION_Q16 + delta_Q16;
if( psEnc->sCmn.fs_kHz == 24 ) {
/* Less bandwidth expansion for super wideband */
BWExp1_Q16 = ( 1 << 16 ) - SKP_SMULWB( SWB_BANDWIDTH_EXPANSION_REDUCTION_Q16, ( 1 << 16 ) - BWExp1_Q16 );
BWExp2_Q16 = ( 1 << 16 ) - SKP_SMULWB( SWB_BANDWIDTH_EXPANSION_REDUCTION_Q16, ( 1 << 16 ) - BWExp2_Q16 );
}
/* BWExp1 will be applied after BWExp2, so make it relative */
BWExp1_Q16 = SKP_DIV32_16( SKP_LSHIFT( BWExp1_Q16, 14 ), SKP_RSHIFT( BWExp2_Q16, 2 ) );
/********************************************/
/* Compute noise shaping AR coefs and gains */
/********************************************/
sz = ( SKP_int )SKP_SMULBB( SHAPE_LPC_WIN_MS, psEnc->sCmn.fs_kHz );
for( k = 0; k < NB_SUBFR; k++ ) {
/* Apply window */
SKP_Silk_apply_sine_window( x_windowed, x_ptr, 0, SHAPE_LPC_WIN_MS * psEnc->sCmn.fs_kHz );
/* Update pointer: next LPC analysis block */
x_ptr += psEnc->sCmn.frame_length / NB_SUBFR;
/* Calculate auto correlation */
SKP_Silk_autocorr( auto_corr, &scale, x_windowed, sz, psEnc->sCmn.shapingLPCOrder + 1 );
/* Add white noise, as a fraction of energy */
auto_corr[0] = SKP_ADD32( auto_corr[0], SKP_max_32( SKP_SMULWB( SKP_RSHIFT( auto_corr[ 0 ], 4 ), SHAPE_WHITE_NOISE_FRACTION_Q20 ), 1 ) );
/* Calculate the reflection coefficients using schur */
nrg = SKP_Silk_schur64( refl_coef_Q16, auto_corr, psEnc->sCmn.shapingLPCOrder );
/* Convert reflection coefficients to prediction coefficients */
SKP_Silk_k2a_Q16( AR_Q24, refl_coef_Q16, psEnc->sCmn.shapingLPCOrder );
/* Bandwidth expansion for synthesis filter shaping */
SKP_Silk_bwexpander_32( AR_Q24, psEnc->sCmn.shapingLPCOrder, BWExp2_Q16 );
/* Make sure to fit in Q13 SKP_int16 */
SKP_Silk_LPC_fit( &psEncCtrl->AR2_Q13[ k * SHAPE_LPC_ORDER_MAX ], AR_Q24, 13, psEnc->sCmn.shapingLPCOrder );
/* Compute noise shaping filter coefficients */
SKP_memcpy(
&psEncCtrl->AR1_Q13[ k * SHAPE_LPC_ORDER_MAX ],
&psEncCtrl->AR2_Q13[ k * SHAPE_LPC_ORDER_MAX ],
psEnc->sCmn.shapingLPCOrder * sizeof( SKP_int16 ) );
/* Bandwidth expansion for analysis filter shaping */
SKP_assert( BWExp1_Q16 <= ( 1 << 16 ) ); // If ever breaking, use LPC_stabilize() in these cases to stay within range
SKP_Silk_bwexpander( &psEncCtrl->AR1_Q13[ k * SHAPE_LPC_ORDER_MAX ], psEnc->sCmn.shapingLPCOrder, BWExp1_Q16 );
/* Increase residual energy */
nrg = SKP_SMLAWB( nrg, SKP_RSHIFT( auto_corr[ 0 ], 8 ), SHAPE_MIN_ENERGY_RATIO_Q24 );
Qnrg = -scale; // range: -12...30
SKP_assert( Qnrg >= -12 );
SKP_assert( Qnrg <= 30 );
/* Make sure that Qnrg is an even number */
if( Qnrg & 1 ) {
Qnrg -= 1;
nrg >>= 1;
}
tmp32 = SKP_Silk_SQRT_APPROX( nrg );
Qnrg >>= 1; // range: -6...15
sqrt_nrg[ k ] = tmp32;
Qnrg_vec[ k ] = Qnrg;
psEncCtrl->Gains_Q16[ k ] = SKP_LSHIFT_SAT32( tmp32, 16 - Qnrg );
/* Ratio of prediction gains, in energy domain */
SKP_Silk_LPC_inverse_pred_gain_Q13( &pre_nrg_Q30, &psEncCtrl->AR2_Q13[ k * SHAPE_LPC_ORDER_MAX ], psEnc->sCmn.shapingLPCOrder );
SKP_Silk_LPC_inverse_pred_gain_Q13( &nrg, &psEncCtrl->AR1_Q13[ k * SHAPE_LPC_ORDER_MAX ], psEnc->sCmn.shapingLPCOrder );
lz = SKP_min_32( SKP_Silk_CLZ32( pre_nrg_Q30 ) - 1, 19 );
pre_nrg_Q30 = SKP_DIV32( SKP_LSHIFT( pre_nrg_Q30, lz ), SKP_RSHIFT( nrg, 20 - lz ) + 1 ); // Q20
pre_nrg_Q30 = SKP_RSHIFT( SKP_LSHIFT_SAT32( pre_nrg_Q30, 9 ), 1 ); /* Q28 */
psEncCtrl->GainsPre_Q14[ k ] = ( SKP_int )SKP_Silk_SQRT_APPROX( pre_nrg_Q30 );
}