/* Compute noise shaping coefficients and initial gain values */
void silk_noise_shape_analysis_FLP(
silk_encoder_state_FLP *psEnc, /* I/O Encoder state FLP */
silk_encoder_control_FLP *psEncCtrl, /* I/O Encoder control FLP */
const silk_float *pitch_res, /* I LPC residual from pitch analysis */
const silk_float *x /* I Input signal [frame_length + la_shape] */
)
{
silk_shape_state_FLP *psShapeSt = &psEnc->sShape;
opus_int k, nSamples;
silk_float SNR_adj_dB, HarmBoost, HarmShapeGain, Tilt;
silk_float nrg, pre_nrg, log_energy, log_energy_prev, energy_variation;
silk_float delta, BWExp1, BWExp2, gain_mult, gain_add, strength, b, warping;
silk_float x_windowed[ SHAPE_LPC_WIN_MAX ];
silk_float auto_corr[ MAX_SHAPE_LPC_ORDER + 1 ];
const silk_float *x_ptr, *pitch_res_ptr;
/* Point to start of first LPC analysis block */
x_ptr = x - psEnc->sCmn.la_shape;
/****************/
/* GAIN CONTROL */
/****************/
SNR_adj_dB = psEnc->sCmn.SNR_dB_Q7 * ( 1 / 128.0f );
/* Input quality is the average of the quality in the lowest two VAD bands */
psEncCtrl->input_quality = 0.5f * ( psEnc->sCmn.input_quality_bands_Q15[ 0 ] + psEnc->sCmn.input_quality_bands_Q15[ 1 ] ) * ( 1.0f / 32768.0f );
/* Coding quality level, between 0.0 and 1.0 */
psEncCtrl->coding_quality = silk_sigmoid( 0.25f * ( SNR_adj_dB - 20.0f ) );
if( psEnc->sCmn.useCBR == 0 ) {
/* Reduce coding SNR during low speech activity */
b = 1.0f - psEnc->sCmn.speech_activity_Q8 * ( 1.0f / 256.0f );
SNR_adj_dB -= BG_SNR_DECR_dB * psEncCtrl->coding_quality * ( 0.5f + 0.5f * psEncCtrl->input_quality ) * b * b;
}
if( psEnc->sCmn.indices.signalType == TYPE_VOICED ) {
/* Reduce gains for periodic signals */
SNR_adj_dB += HARM_SNR_INCR_dB * psEnc->LTPCorr;
} else {
/* For unvoiced signals and low-quality input, adjust the quality slower than SNR_dB setting */
SNR_adj_dB += ( -0.4f * psEnc->sCmn.SNR_dB_Q7 * ( 1 / 128.0f ) + 6.0f ) * ( 1.0f - psEncCtrl->input_quality );
}
/*************************/
/* SPARSENESS PROCESSING */
/*************************/
/* Set quantizer offset */
if( psEnc->sCmn.indices.signalType == TYPE_VOICED ) {
/* Initially set to 0; may be overruled in process_gains(..) */
psEnc->sCmn.indices.quantOffsetType = 0;
psEncCtrl->sparseness = 0.0f;
} else {
/* Sparseness measure, based on relative fluctuations of energy per 2 milliseconds */
nSamples = 2 * psEnc->sCmn.fs_kHz;
energy_variation = 0.0f;
log_energy_prev = 0.0f;
pitch_res_ptr = pitch_res;
for( k = 0; k < silk_SMULBB( SUB_FRAME_LENGTH_MS, psEnc->sCmn.nb_subfr ) / 2; k++ ) {
nrg = ( silk_float )nSamples + ( silk_float )silk_energy_FLP( pitch_res_ptr, nSamples );
log_energy = silk_log2( nrg );
if( k > 0 ) {
energy_variation += silk_abs_float( log_energy - log_energy_prev );
}
log_energy_prev = log_energy;
pitch_res_ptr += nSamples;
}
psEncCtrl->sparseness = silk_sigmoid( 0.4f * ( energy_variation - 5.0f ) );
/* Set quantization offset depending on sparseness measure */
if( psEncCtrl->sparseness > SPARSENESS_THRESHOLD_QNT_OFFSET ) {
psEnc->sCmn.indices.quantOffsetType = 0;
} else {
psEnc->sCmn.indices.quantOffsetType = 1;
}
/* Increase coding SNR for sparse signals */
SNR_adj_dB += SPARSE_SNR_INCR_dB * ( psEncCtrl->sparseness - 0.5f );
}
/*******************************/
/* Control bandwidth expansion */
/*******************************/
/* More BWE for signals with high prediction gain */
strength = FIND_PITCH_WHITE_NOISE_FRACTION * psEncCtrl->predGain; /* between 0.0 and 1.0 */
BWExp1 = BWExp2 = BANDWIDTH_EXPANSION / ( 1.0f + strength * strength );
delta = LOW_RATE_BANDWIDTH_EXPANSION_DELTA * ( 1.0f - 0.75f * psEncCtrl->coding_quality );
BWExp1 -= delta;
BWExp2 += delta;
/* BWExp1 will be applied after BWExp2, so make it relative */
BWExp1 /= BWExp2;
if( psEnc->sCmn.warping_Q16 > 0 ) {
/* Slightly more warping in analysis will move quantization noise up in frequency, where it's better masked */
warping = (silk_float)psEnc->sCmn.warping_Q16 / 65536.0f + 0.01f * psEncCtrl->coding_quality;
} else {
warping = 0.0f;
}
/********************************************/
/* Compute noise shaping AR coefs and gains */
/********************************************/
for( k = 0; k < psEnc->sCmn.nb_subfr; k++ ) {
/* Apply window: sine slope followed by flat part followed by cosine slope */
opus_int shift, slope_part, flat_part;
flat_part = psEnc->sCmn.fs_kHz * 3;
slope_part = ( psEnc->sCmn.shapeWinLength - flat_part ) / 2;
silk_apply_sine_window_FLP( x_windowed, x_ptr, 1, slope_part );
shift = slope_part;
silk_memcpy( x_windowed + shift, x_ptr + shift, flat_part * sizeof(silk_float) );
shift += flat_part;
silk_apply_sine_window_FLP( 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 */
silk_warped_autocorrelation_FLP( auto_corr, x_windowed, warping,
psEnc->sCmn.shapeWinLength, psEnc->sCmn.shapingLPCOrder );
} else {
/* Calculate regular auto correlation */
silk_autocorrelation_FLP( auto_corr, x_windowed, psEnc->sCmn.shapeWinLength, psEnc->sCmn.shapingLPCOrder + 1 );
}
/* Add white noise, as a fraction of energy */
auto_corr[ 0 ] += auto_corr[ 0 ] * SHAPE_WHITE_NOISE_FRACTION;
/* Convert correlations to prediction coefficients, and compute residual energy */
nrg = silk_levinsondurbin_FLP( &psEncCtrl->AR2[ k * MAX_SHAPE_LPC_ORDER ], auto_corr, psEnc->sCmn.shapingLPCOrder );
psEncCtrl->Gains[ k ] = ( silk_float )sqrt( nrg );
if( psEnc->sCmn.warping_Q16 > 0 ) {
/* Adjust gain for warping */
psEncCtrl->Gains[ k ] *= warped_gain( &psEncCtrl->AR2[ k * MAX_SHAPE_LPC_ORDER ], warping, psEnc->sCmn.shapingLPCOrder );
}
/* Bandwidth expansion for synthesis filter shaping */
silk_bwexpander_FLP( &psEncCtrl->AR2[ k * MAX_SHAPE_LPC_ORDER ], psEnc->sCmn.shapingLPCOrder, BWExp2 );
/* Compute noise shaping filter coefficients */
silk_memcpy(
&psEncCtrl->AR1[ k * MAX_SHAPE_LPC_ORDER ],
&psEncCtrl->AR2[ k * MAX_SHAPE_LPC_ORDER ],
psEnc->sCmn.shapingLPCOrder * sizeof( silk_float ) );
/* Bandwidth expansion for analysis filter shaping */
silk_bwexpander_FLP( &psEncCtrl->AR1[ k * MAX_SHAPE_LPC_ORDER ], psEnc->sCmn.shapingLPCOrder, BWExp1 );
/* Ratio of prediction gains, in energy domain */
pre_nrg = silk_LPC_inverse_pred_gain_FLP( &psEncCtrl->AR2[ k * MAX_SHAPE_LPC_ORDER ], psEnc->sCmn.shapingLPCOrder );
nrg = silk_LPC_inverse_pred_gain_FLP( &psEncCtrl->AR1[ k * MAX_SHAPE_LPC_ORDER ], psEnc->sCmn.shapingLPCOrder );
psEncCtrl->GainsPre[ k ] = 1.0f - 0.7f * ( 1.0f - pre_nrg / nrg );
/* Convert to monic warped prediction coefficients and limit absolute values */
warped_true2monic_coefs( &psEncCtrl->AR2[ k * MAX_SHAPE_LPC_ORDER ], &psEncCtrl->AR1[ k * MAX_SHAPE_LPC_ORDER ],
warping, 3.999f, psEnc->sCmn.shapingLPCOrder );
}
/*****************/
/* Gain tweaking */
/*****************/
/* Increase gains during low speech activity */
gain_mult = (silk_float)pow( 2.0f, -0.16f * SNR_adj_dB );
gain_add = (silk_float)pow( 2.0f, 0.16f * MIN_QGAIN_DB );
for( k = 0; k < psEnc->sCmn.nb_subfr; k++ ) {
psEncCtrl->Gains[ k ] *= gain_mult;
psEncCtrl->Gains[ k ] += gain_add;
}
gain_mult = 1.0f + INPUT_TILT + psEncCtrl->coding_quality * HIGH_RATE_INPUT_TILT;
for( k = 0; k < psEnc->sCmn.nb_subfr; k++ ) {
psEncCtrl->GainsPre[ k ] *= gain_mult;
}
/************************************************/
/* Control low-frequency shaping and noise tilt */
/************************************************/
/* Less low frequency shaping for noisy inputs */
strength = LOW_FREQ_SHAPING * ( 1.0f + LOW_QUALITY_LOW_FREQ_SHAPING_DECR * ( psEnc->sCmn.input_quality_bands_Q15[ 0 ] * ( 1.0f / 32768.0f ) - 1.0f ) );
strength *= psEnc->sCmn.speech_activity_Q8 * ( 1.0f / 256.0f );
if( psEnc->sCmn.indices.signalType == TYPE_VOICED ) {
/* Reduce low frequencies quantization noise for periodic signals, depending on pitch lag */
/*f = 400; freqz([1, -0.98 + 2e-4 * f], [1, -0.97 + 7e-4 * f], 2^12, Fs); axis([0, 1000, -10, 1])*/
for( k = 0; k < psEnc->sCmn.nb_subfr; k++ ) {
b = 0.2f / psEnc->sCmn.fs_kHz + 3.0f / psEncCtrl->pitchL[ k ];
psEncCtrl->LF_MA_shp[ k ] = -1.0f + b;
psEncCtrl->LF_AR_shp[ k ] = 1.0f - b - b * strength;
}
Tilt = - HP_NOISE_COEF -
(1 - HP_NOISE_COEF) * HARM_HP_NOISE_COEF * psEnc->sCmn.speech_activity_Q8 * ( 1.0f / 256.0f );
} else {
b = 1.3f / psEnc->sCmn.fs_kHz;
psEncCtrl->LF_MA_shp[ 0 ] = -1.0f + b;
psEncCtrl->LF_AR_shp[ 0 ] = 1.0f - b - b * strength * 0.6f;
for( k = 1; k < psEnc->sCmn.nb_subfr; k++ ) {
psEncCtrl->LF_MA_shp[ k ] = psEncCtrl->LF_MA_shp[ 0 ];
psEncCtrl->LF_AR_shp[ k ] = psEncCtrl->LF_AR_shp[ 0 ];
}
Tilt = -HP_NOISE_COEF;
}
/****************************/
/* HARMONIC SHAPING CONTROL */
/****************************/
/* Control boosting of harmonic frequencies */
HarmBoost = LOW_RATE_HARMONIC_BOOST * ( 1.0f - psEncCtrl->coding_quality ) * psEnc->LTPCorr;
/* More harmonic boost for noisy input signals */
HarmBoost += LOW_INPUT_QUALITY_HARMONIC_BOOST * ( 1.0f - psEncCtrl->input_quality );
if( USE_HARM_SHAPING && psEnc->sCmn.indices.signalType == TYPE_VOICED ) {
/* Harmonic noise shaping */
HarmShapeGain = HARMONIC_SHAPING;
/* More harmonic noise shaping for high bitrates or noisy input */
HarmShapeGain += HIGH_RATE_OR_LOW_QUALITY_HARMONIC_SHAPING *
( 1.0f - ( 1.0f - psEncCtrl->coding_quality ) * psEncCtrl->input_quality );
/* Less harmonic noise shaping for less periodic signals */
HarmShapeGain *= ( silk_float )sqrt( psEnc->LTPCorr );
} else {
HarmShapeGain = 0.0f;
}
/*************************/
/* Smooth over subframes */
/*************************/
for( k = 0; k < psEnc->sCmn.nb_subfr; k++ ) {
psShapeSt->HarmBoost_smth += SUBFR_SMTH_COEF * ( HarmBoost - psShapeSt->HarmBoost_smth );
psEncCtrl->HarmBoost[ k ] = psShapeSt->HarmBoost_smth;
psShapeSt->HarmShapeGain_smth += SUBFR_SMTH_COEF * ( HarmShapeGain - psShapeSt->HarmShapeGain_smth );
psEncCtrl->HarmShapeGain[ k ] = psShapeSt->HarmShapeGain_smth;
psShapeSt->Tilt_smth += SUBFR_SMTH_COEF * ( Tilt - psShapeSt->Tilt_smth );
psEncCtrl->Tilt[ k ] = psShapeSt->Tilt_smth;
}
}
/* Processing of gains */
void silk_process_gains_FLP(
silk_encoder_state_FLP *psEnc, /* I/O Encoder state FLP */
silk_encoder_control_FLP *psEncCtrl, /* I/O Encoder control FLP */
opus_int condCoding /* I The type of conditional coding to use */
) {
silk_shape_state_FLP *psShapeSt = &psEnc->sShape;
opus_int k;
opus_int32 pGains_Q16[MAX_NB_SUBFR];
silk_float s, InvMaxSqrVal, gain, quant_offset;
/* Gain reduction when LTP coding gain is high */
if (psEnc->sCmn.indices.signalType == TYPE_VOICED) {
s = 1.0f - 0.5f * silk_sigmoid(0.25f * (psEncCtrl->LTPredCodGain - 12.0f));
for (k = 0; k < psEnc->sCmn.nb_subfr; k++) {
psEncCtrl->Gains[k] *= s;
}
}
/* Limit the quantized signal */
InvMaxSqrVal = (silk_float)(pow(2.0f, 0.33f * (21.0f - psEnc->sCmn.SNR_dB_Q7 * (1 / 128.0f))) /
psEnc->sCmn.subfr_length);
for (k = 0; k < psEnc->sCmn.nb_subfr; k++) {
/* Soft limit on ratio residual energy and squared gains */
gain = psEncCtrl->Gains[k];
gain = (silk_float) sqrt(gain * gain + psEncCtrl->ResNrg[k] * InvMaxSqrVal);
psEncCtrl->Gains[k] = silk_min_float(gain, 32767.0f);
}
/* Prepare gains for noise shaping quantization */
for (k = 0; k < psEnc->sCmn.nb_subfr; k++) {
pGains_Q16[k] = (opus_int32)(psEncCtrl->Gains[k] * 65536.0f);
}
/* Save unquantized gains and gain Index */
silk_memcpy(psEncCtrl->GainsUnq_Q16, pGains_Q16, psEnc->sCmn.nb_subfr * sizeof(opus_int32));
psEncCtrl->lastGainIndexPrev = psShapeSt->LastGainIndex;
/* Quantize gains */
silk_gains_quant(psEnc->sCmn.indices.GainsIndices, pGains_Q16,
&psShapeSt->LastGainIndex, condCoding == CODE_CONDITIONALLY,
psEnc->sCmn.nb_subfr);
/* Overwrite unquantized gains with quantized gains and convert back to Q0 from Q16 */
for (k = 0; k < psEnc->sCmn.nb_subfr; k++) {
psEncCtrl->Gains[k] = pGains_Q16[k] / 65536.0f;
}
/* Set quantizer offset for voiced signals. Larger offset when LTP coding gain is low or tilt is high (ie low-pass) */
if (psEnc->sCmn.indices.signalType == TYPE_VOICED) {
if (psEncCtrl->LTPredCodGain + psEnc->sCmn.input_tilt_Q15 * (1.0f / 32768.0f) > 1.0f) {
psEnc->sCmn.indices.quantOffsetType = 0;
} else {
psEnc->sCmn.indices.quantOffsetType = 1;
}
}
/* Quantizer boundary adjustment */
quant_offset = silk_Quantization_Offsets_Q10[psEnc->sCmn.indices.signalType >>
1][psEnc->sCmn.indices.quantOffsetType] / 1024.0f;
psEncCtrl->Lambda = LAMBDA_OFFSET
+ LAMBDA_DELAYED_DECISIONS * psEnc->sCmn.nStatesDelayedDecision
+ LAMBDA_SPEECH_ACT * psEnc->sCmn.speech_activity_Q8 * (1.0f / 256.0f)
+ LAMBDA_INPUT_QUALITY * psEncCtrl->input_quality
+ LAMBDA_CODING_QUALITY * psEncCtrl->coding_quality
+ LAMBDA_QUANT_OFFSET * quant_offset;
silk_assert(psEncCtrl->Lambda > 0.0f);
silk_assert(psEncCtrl->Lambda < 2.0f);
}