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 ];
SKP_int32 sqrt_nrg[ MAX_NB_SUBFR ], Qnrg_vec[ MAX_NB_SUBFR ];
const SKP_int16 *x_ptr, *pitch_res_ptr;
/* 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_SMULBB( psEnc->BufferedInChannel_ms, SKP_FIX_CONST( 0.1, 7 ) );
/* Reduce SNR_dB because of any inband FEC used */
psEncCtrl->current_SNR_dB_Q7 -= psEnc->inBandFEC_SNR_comp_Q7;
/****************/
/* 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 )psEnc->sCmn.input_quality_bands_Q15[ 0 ]
+ psEnc->sCmn.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 */
SNR_adj_dB_Q7 = psEncCtrl->current_SNR_dB_Q7;
if( psEnc->sCmn.useCBR == 0 ) {
b_Q8 = SKP_FIX_CONST( 1.0, 8 ) - psEnc->sCmn.speech_activity_Q8;
b_Q8 = SKP_SMULWB( SKP_LSHIFT( b_Q8, 8 ), b_Q8 );
SNR_adj_dB_Q7 = SKP_SMLAWB( SNR_adj_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
}
void SKP_Silk_LTP_scale_ctrl_FIX(
SKP_Silk_encoder_state_FIX *psEnc, /* I/O encoder state FIX */
SKP_Silk_encoder_control_FIX *psEncCtrl /* I/O encoder control FIX */
)
{
SKP_int round_loss, frames_per_packet;
SKP_int g_out_Q5, g_limit_Q15, thrld1_Q15, thrld2_Q15;
/* 1st order high-pass filter */
psEnc->HPLTPredCodGain_Q7 = SKP_max_int( psEncCtrl->LTPredCodGain_Q7 - psEnc->prevLTPredCodGain_Q7, 0 )
+ SKP_RSHIFT_ROUND( psEnc->HPLTPredCodGain_Q7, 1 );
psEnc->prevLTPredCodGain_Q7 = psEncCtrl->LTPredCodGain_Q7;
/* combine input and filtered input */
g_out_Q5 = SKP_RSHIFT_ROUND( SKP_RSHIFT( psEncCtrl->LTPredCodGain_Q7, 1 ) + SKP_RSHIFT( psEnc->HPLTPredCodGain_Q7, 1 ), 3 );
g_limit_Q15 = SKP_Silk_sigm_Q15( g_out_Q5 - ( 3 << 5 ) );
/* Default is minimum scaling */
psEncCtrl->sCmn.LTP_scaleIndex = 0;
/* Round the loss measure to whole pct */
round_loss = ( SKP_int )psEnc->sCmn.PacketLoss_perc;
/* Only scale if first frame in packet 0% */
if( psEnc->sCmn.nFramesInPayloadBuf == 0 ) {
frames_per_packet = SKP_DIV32_16( psEnc->sCmn.PacketSize_ms, FRAME_LENGTH_MS );
round_loss += frames_per_packet - 1;
thrld1_Q15 = LTPScaleThresholds_Q15[ SKP_min_int( round_loss, NB_THRESHOLDS - 1 ) ];
thrld2_Q15 = LTPScaleThresholds_Q15[ SKP_min_int( round_loss + 1, NB_THRESHOLDS - 1 ) ];
if( g_limit_Q15 > thrld1_Q15 ) {
/* Maximum scaling */
psEncCtrl->sCmn.LTP_scaleIndex = 2;
} else if( g_limit_Q15 > thrld2_Q15 ) {
/* Medium scaling */
psEncCtrl->sCmn.LTP_scaleIndex = 1;
}
}
psEncCtrl->LTP_scale_Q14 = SKP_Silk_LTPScales_table_Q14[ psEncCtrl->sCmn.LTP_scaleIndex ];
}
SKP_int SKP_Silk_VAD_GetSA_Q8( /* O Return value, 0 if success */
SKP_Silk_encoder_state *psEncC, /* I/O Encoder state */
const SKP_int16 pIn[] /* I PCM input */
)
{
SKP_int SA_Q15, pSNR_dB_Q7, input_tilt;
SKP_int decimated_framelength, dec_subframe_length, dec_subframe_offset, SNR_Q7, i, b, s;
SKP_int32 sumSquared, smooth_coef_Q16;
SKP_int16 HPstateTmp;
SKP_int16 X[ VAD_N_BANDS ][ MAX_FRAME_LENGTH / 2 ];
SKP_int32 Xnrg[ VAD_N_BANDS ];
SKP_int32 NrgToNoiseRatio_Q8[ VAD_N_BANDS ];
SKP_int32 speech_nrg, x_tmp;
SKP_int ret = 0;
SKP_Silk_VAD_state *psSilk_VAD = &psEncC->sVAD;
/* Safety checks */
SKP_assert( VAD_N_BANDS == 4 );
SKP_assert( MAX_FRAME_LENGTH >= psEncC->frame_length );
SKP_assert( psEncC->frame_length <= 512 );
SKP_assert( psEncC->frame_length == 8 * SKP_RSHIFT( psEncC->frame_length, 3 ) );
/***********************/
/* Filter and Decimate */
/***********************/
/* 0-8 kHz to 0-4 kHz and 4-8 kHz */
SKP_Silk_ana_filt_bank_1( pIn, &psSilk_VAD->AnaState[ 0 ], &X[ 0 ][ 0 ], &X[ 3 ][ 0 ], psEncC->frame_length );
/* 0-4 kHz to 0-2 kHz and 2-4 kHz */
SKP_Silk_ana_filt_bank_1( &X[ 0 ][ 0 ], &psSilk_VAD->AnaState1[ 0 ], &X[ 0 ][ 0 ], &X[ 2 ][ 0 ], SKP_RSHIFT( psEncC->frame_length, 1 ) );
/* 0-2 kHz to 0-1 kHz and 1-2 kHz */
SKP_Silk_ana_filt_bank_1( &X[ 0 ][ 0 ], &psSilk_VAD->AnaState2[ 0 ], &X[ 0 ][ 0 ], &X[ 1 ][ 0 ], SKP_RSHIFT( psEncC->frame_length, 2 ) );
/*********************************************/
/* HP filter on lowest band (differentiator) */
/*********************************************/
decimated_framelength = SKP_RSHIFT( psEncC->frame_length, 3 );
X[ 0 ][ decimated_framelength - 1 ] = SKP_RSHIFT( X[ 0 ][ decimated_framelength - 1 ], 1 );
HPstateTmp = X[ 0 ][ decimated_framelength - 1 ];
for( i = decimated_framelength - 1; i > 0; i-- ) {
X[ 0 ][ i - 1 ] = SKP_RSHIFT( X[ 0 ][ i - 1 ], 1 );
X[ 0 ][ i ] -= X[ 0 ][ i - 1 ];
}
X[ 0 ][ 0 ] -= psSilk_VAD->HPstate;
psSilk_VAD->HPstate = HPstateTmp;
/*************************************/
/* Calculate the energy in each band */
/*************************************/
for( b = 0; b < VAD_N_BANDS; b++ ) {
/* Find the decimated framelength in the non-uniformly divided bands */
decimated_framelength = SKP_RSHIFT( psEncC->frame_length, SKP_min_int( VAD_N_BANDS - b, VAD_N_BANDS - 1 ) );
/* Split length into subframe lengths */
dec_subframe_length = SKP_RSHIFT( decimated_framelength, VAD_INTERNAL_SUBFRAMES_LOG2 );
dec_subframe_offset = 0;
/* Compute energy per sub-frame */
/* initialize with summed energy of last subframe */
Xnrg[ b ] = psSilk_VAD->XnrgSubfr[ b ];
for( s = 0; s < VAD_INTERNAL_SUBFRAMES; s++ ) {
sumSquared = 0;
for( i = 0; i < dec_subframe_length; i++ ) {
/* The energy will be less than dec_subframe_length * ( SKP_int16_MIN / 8 ) ^ 2. */
/* Therefore we can accumulate with no risk of overflow (unless dec_subframe_length > 128) */
x_tmp = SKP_RSHIFT( X[ b ][ i + dec_subframe_offset ], 3 );
sumSquared = SKP_SMLABB( sumSquared, x_tmp, x_tmp );
/* Safety check */
SKP_assert( sumSquared >= 0 );
}
/* Add/saturate summed energy of current subframe */
if( s < VAD_INTERNAL_SUBFRAMES - 1 ) {
Xnrg[ b ] = SKP_ADD_POS_SAT32( Xnrg[ b ], sumSquared );
} else {
/* Look-ahead subframe */
Xnrg[ b ] = SKP_ADD_POS_SAT32( Xnrg[ b ], SKP_RSHIFT( sumSquared, 1 ) );
}
dec_subframe_offset += dec_subframe_length;
}
psSilk_VAD->XnrgSubfr[ b ] = sumSquared;
}
/********************/
/* Noise estimation */
/********************/
SKP_Silk_VAD_GetNoiseLevels( &Xnrg[ 0 ], psSilk_VAD );
/***********************************************/
/* Signal-plus-noise to noise ratio estimation */
/***********************************************/
sumSquared = 0;
input_tilt = 0;
for( b = 0; b < VAD_N_BANDS; b++ ) {
speech_nrg = Xnrg[ b ] - psSilk_VAD->NL[ b ];
if( speech_nrg > 0 ) {
/* Divide, with sufficient resolution */
if( ( Xnrg[ b ] & 0xFF800000 ) == 0 ) {
NrgToNoiseRatio_Q8[ b ] = SKP_DIV32( SKP_LSHIFT( Xnrg[ b ], 8 ), psSilk_VAD->NL[ b ] + 1 );
} else {
NrgToNoiseRatio_Q8[ b ] = SKP_DIV32( Xnrg[ b ], SKP_RSHIFT( psSilk_VAD->NL[ b ], 8 ) + 1 );
}
/* Convert to log domain */
SNR_Q7 = SKP_Silk_lin2log( NrgToNoiseRatio_Q8[ b ] ) - 8 * 128;
/* Sum-of-squares */
sumSquared = SKP_SMLABB( sumSquared, SNR_Q7, SNR_Q7 ); /* Q14 */
/* Tilt measure */
if( speech_nrg < ( 1 << 20 ) ) {
/* Scale down SNR value for small subband speech energies */
SNR_Q7 = SKP_SMULWB( SKP_LSHIFT( SKP_Silk_SQRT_APPROX( speech_nrg ), 6 ), SNR_Q7 );
}
input_tilt = SKP_SMLAWB( input_tilt, tiltWeights[ b ], SNR_Q7 );
} else {
NrgToNoiseRatio_Q8[ b ] = 256;
}
}
/* Mean-of-squares */
sumSquared = SKP_DIV32_16( sumSquared, VAD_N_BANDS ); /* Q14 */
/* Root-mean-square approximation, scale to dBs, and write to output pointer */
pSNR_dB_Q7 = ( SKP_int16 )( 3 * SKP_Silk_SQRT_APPROX( sumSquared ) ); /* Q7 */
/*********************************/
/* Speech Probability Estimation */
/*********************************/
SA_Q15 = SKP_Silk_sigm_Q15( SKP_SMULWB( VAD_SNR_FACTOR_Q16, pSNR_dB_Q7 ) - VAD_NEGATIVE_OFFSET_Q5 );
/**************************/
/* Frequency Tilt Measure */
/**************************/
psEncC->input_tilt_Q15 = SKP_LSHIFT( SKP_Silk_sigm_Q15( input_tilt ) - 16384, 1 );
/**************************************************/
/* Scale the sigmoid output based on power levels */
/**************************************************/
speech_nrg = 0;
for( b = 0; b < VAD_N_BANDS; b++ ) {
/* Accumulate signal-without-noise energies, higher frequency bands have more weight */
speech_nrg += ( b + 1 ) * SKP_RSHIFT( Xnrg[ b ] - psSilk_VAD->NL[ b ], 4 );
}
/* Power scaling */
if( speech_nrg <= 0 ) {
SA_Q15 = SKP_RSHIFT( SA_Q15, 1 );
} else if( speech_nrg < 32768 ) {
if( psEncC->frame_length == 10 * psEncC->fs_kHz ) {
speech_nrg = SKP_LSHIFT_SAT32( speech_nrg, 16 );
} else {
speech_nrg = SKP_LSHIFT_SAT32( speech_nrg, 15 );
}
/* square-root */
speech_nrg = SKP_Silk_SQRT_APPROX( speech_nrg );
SA_Q15 = SKP_SMULWB( 32768 + speech_nrg, SA_Q15 );
}
/* Copy the resulting speech activity in Q8 */
psEncC->speech_activity_Q8 = SKP_min_int( SKP_RSHIFT( SA_Q15, 7 ), SKP_uint8_MAX );
/***********************************/
/* Energy Level and SNR estimation */
/***********************************/
/* Smoothing coefficient */
smooth_coef_Q16 = SKP_SMULWB( VAD_SNR_SMOOTH_COEF_Q18, SKP_SMULWB( SA_Q15, SA_Q15 ) );
if( psEncC->frame_length == 10 * psEncC->fs_kHz ) {
smooth_coef_Q16 >>= 1;
}
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 ) );
}
}
/* Processing of gains */
void SKP_Silk_process_gains_FIX(
SKP_Silk_encoder_state_FIX *psEnc, /* I/O Encoder state_FIX */
SKP_Silk_encoder_control_FIX *psEncCtrl /* I/O Encoder control_FIX */
)
{
SKP_Silk_shape_state_FIX *psShapeSt = &psEnc->sShape;
SKP_int k;
SKP_int32 s_Q16, InvMaxSqrVal_Q16, gain, gain_squared, ResNrg, ResNrgPart;
/* Gain reduction when LTP coding gain is high */
if( psEncCtrl->sCmn.sigtype == SIG_TYPE_VOICED ) {
/*s = -0.5f * SKP_sigmoid( 0.25f * ( psEncCtrl->LTPredCodGain - 12.0f ) ); */
s_Q16 = -SKP_Silk_sigm_Q15( SKP_RSHIFT_ROUND( psEncCtrl->LTPredCodGain_Q7 - (12 << 7), 4 ) );
for( k = 0; k < NB_SUBFR; k++ ) {
psEncCtrl->Gains_Q16[ k ] = SKP_SMLAWB( psEncCtrl->Gains_Q16[ k ], psEncCtrl->Gains_Q16[ k ], s_Q16 );
}
}
/* Limit the quantized signal */
/* 69 = 21.0f + 16/0.33 */
InvMaxSqrVal_Q16 = SKP_DIV32_16( SKP_Silk_log2lin(
SKP_SMULWB( (69 << 7) - psEncCtrl->current_SNR_dB_Q7, SKP_FIX_CONST( 0.33, 16 )) ), psEnc->sCmn.subfr_length );
for( k = 0; k < NB_SUBFR; k++ ) {
/* Soft limit on ratio residual energy and squared gains */
ResNrg = psEncCtrl->ResNrg[ k ];
ResNrgPart = SKP_SMULWW( ResNrg, InvMaxSqrVal_Q16 );
if( psEncCtrl->ResNrgQ[ k ] > 0 ) {
if( psEncCtrl->ResNrgQ[ k ] < 32 ) {
ResNrgPart = SKP_RSHIFT_ROUND( ResNrgPart, psEncCtrl->ResNrgQ[ k ] );
} else {
ResNrgPart = 0;
}
} else if( psEncCtrl->ResNrgQ[k] != 0 ) {
if( ResNrgPart > SKP_RSHIFT( SKP_int32_MAX, -psEncCtrl->ResNrgQ[ k ] ) ) {
ResNrgPart = SKP_int32_MAX;
} else {
ResNrgPart = SKP_LSHIFT( ResNrgPart, -psEncCtrl->ResNrgQ[ k ] );
}
}
gain = psEncCtrl->Gains_Q16[ k ];
gain_squared = SKP_ADD_SAT32( ResNrgPart, SKP_SMMUL( gain, gain ) );
if( gain_squared < SKP_int16_MAX ) {
/* recalculate with higher precision */
gain_squared = SKP_SMLAWW( SKP_LSHIFT( ResNrgPart, 16 ), gain, gain );
SKP_assert( gain_squared > 0 );
gain = SKP_Silk_SQRT_APPROX( gain_squared ); /* Q8 */
psEncCtrl->Gains_Q16[ k ] = SKP_LSHIFT_SAT32( gain, 8 ); /* Q16 */
} else {
gain = SKP_Silk_SQRT_APPROX( gain_squared ); /* Q0 */
psEncCtrl->Gains_Q16[ k ] = SKP_LSHIFT_SAT32( gain, 16 ); /* Q16 */
}
}
/* Noise shaping quantization */
SKP_Silk_gains_quant( psEncCtrl->sCmn.GainsIndices, psEncCtrl->Gains_Q16,
&psShapeSt->LastGainIndex, psEnc->sCmn.nFramesInPayloadBuf );
/* Set quantizer offset for voiced signals. Larger offset when LTP coding gain is low or tilt is high (ie low-pass) */
if( psEncCtrl->sCmn.sigtype == SIG_TYPE_VOICED ) {
if( psEncCtrl->LTPredCodGain_Q7 + SKP_RSHIFT( psEncCtrl->input_tilt_Q15, 8 ) > ( 1 << 7 ) ) {
psEncCtrl->sCmn.QuantOffsetType = 0;
} else {
psEncCtrl->sCmn.QuantOffsetType = 1;
}
}
/* Quantizer boundary adjustment */
if( psEncCtrl->sCmn.sigtype == SIG_TYPE_VOICED ) {
psEncCtrl->Lambda_Q10 = SKP_FIX_CONST( 1.3, 10 )
- SKP_SMULWB( SKP_FIX_CONST( 0.5, 18 ), psEnc->speech_activity_Q8 )
- SKP_SMULWB( SKP_FIX_CONST( 0.3, 12 ), psEncCtrl->input_quality_Q14 )
+ SKP_SMULBB( SKP_FIX_CONST( 0.2, 10 ), psEncCtrl->sCmn.QuantOffsetType )
- SKP_SMULWB( SKP_FIX_CONST( 0.1, 12 ), psEncCtrl->coding_quality_Q14 );
} else {
psEncCtrl->Lambda_Q10 = SKP_FIX_CONST( 1.3, 10 )
- SKP_SMULWB( SKP_FIX_CONST( 0.5, 18 ), psEnc->speech_activity_Q8 )
- SKP_SMULWB( SKP_FIX_CONST( 0.4, 12 ), psEncCtrl->input_quality_Q14 )
+ SKP_SMULBB( SKP_FIX_CONST( 0.4, 10 ), psEncCtrl->sCmn.QuantOffsetType )
- SKP_SMULWB( SKP_FIX_CONST( 0.1, 12 ), psEncCtrl->coding_quality_Q14 );
}
SKP_assert( psEncCtrl->Lambda_Q10 >= 0 );
SKP_assert( psEncCtrl->Lambda_Q10 < SKP_FIX_CONST( 2, 10 ) );
}
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 );
}
