/* Processing of gains */
void silk_process_gains_FIX(
silk_encoder_state_FIX *psEnc, /* I/O Encoder state */
silk_encoder_control_FIX *psEncCtrl, /* I/O Encoder control */
opus_int condCoding /* I The type of conditional coding to use */
)
{
silk_shape_state_FIX *psShapeSt = &psEnc->sShape;
opus_int k;
opus_int32 s_Q16, InvMaxSqrVal_Q16, gain, gain_squared, ResNrg, ResNrgPart, quant_offset_Q10;
/* Gain reduction when LTP coding gain is high */
if( psEnc->sCmn.indices.signalType == TYPE_VOICED ) {
/*s = -0.5f * silk_sigmoid( 0.25f * ( psEncCtrl->LTPredCodGain - 12.0f ) ); */
s_Q16 = -silk_sigm_Q15( silk_RSHIFT_ROUND( psEncCtrl->LTPredCodGain_Q7 - SILK_FIX_CONST( 12.0, 7 ), 4 ) );
for( k = 0; k < psEnc->sCmn.nb_subfr; k++ ) {
psEncCtrl->Gains_Q16[ k ] = silk_SMLAWB( psEncCtrl->Gains_Q16[ k ], psEncCtrl->Gains_Q16[ k ], s_Q16 );
}
}
/* Limit the quantized signal */
/* InvMaxSqrVal = pow( 2.0f, 0.33f * ( 21.0f - SNR_dB ) ) / subfr_length; */
InvMaxSqrVal_Q16 = silk_DIV32_16( silk_log2lin(
silk_SMULWB( SILK_FIX_CONST( 21 + 16 / 0.33, 7 ) - psEnc->sCmn.SNR_dB_Q7, SILK_FIX_CONST( 0.33, 16 ) ) ), psEnc->sCmn.subfr_length );
for( k = 0; k < psEnc->sCmn.nb_subfr; k++ ) {
/* Soft limit on ratio residual energy and squared gains */
ResNrg = psEncCtrl->ResNrg[ k ];
ResNrgPart = silk_SMULWW( ResNrg, InvMaxSqrVal_Q16 );
if( psEncCtrl->ResNrgQ[ k ] > 0 ) {
ResNrgPart = silk_RSHIFT_ROUND( ResNrgPart, psEncCtrl->ResNrgQ[ k ] );
} else {
if( ResNrgPart >= silk_RSHIFT( silk_int32_MAX, -psEncCtrl->ResNrgQ[ k ] ) ) {
ResNrgPart = silk_int32_MAX;
} else {
ResNrgPart = silk_LSHIFT( ResNrgPart, -psEncCtrl->ResNrgQ[ k ] );
}
}
gain = psEncCtrl->Gains_Q16[ k ];
gain_squared = silk_ADD_SAT32( ResNrgPart, silk_SMMUL( gain, gain ) );
if( gain_squared < silk_int16_MAX ) {
/* recalculate with higher precision */
gain_squared = silk_SMLAWW( silk_LSHIFT( ResNrgPart, 16 ), gain, gain );
silk_assert( gain_squared > 0 );
gain = silk_SQRT_APPROX( gain_squared ); /* Q8 */
gain = silk_min( gain, silk_int32_MAX >> 8 );
psEncCtrl->Gains_Q16[ k ] = silk_LSHIFT_SAT32( gain, 8 ); /* Q16 */
} else {
void silk_noise_shape_analysis_FIX(
silk_encoder_state_FIX *psEnc, /* I/O Encoder state FIX */
silk_encoder_control_FIX *psEncCtrl, /* I/O Encoder control FIX */
const opus_int16 *pitch_res, /* I LPC residual from pitch analysis */
const opus_int16 *x, /* I Input signal [ frame_length + la_shape ] */
int arch /* I Run-time architecture */
)
{
silk_shape_state_FIX *psShapeSt = &psEnc->sShape;
opus_int k, i, nSamples, Qnrg, b_Q14, warping_Q16, scale = 0;
opus_int32 SNR_adj_dB_Q7, HarmBoost_Q16, HarmShapeGain_Q16, Tilt_Q16, tmp32;
opus_int32 nrg, pre_nrg_Q30, log_energy_Q7, log_energy_prev_Q7, energy_variation_Q7;
opus_int32 delta_Q16, BWExp1_Q16, BWExp2_Q16, gain_mult_Q16, gain_add_Q16, strength_Q16, b_Q8;
opus_int32 auto_corr[ MAX_SHAPE_LPC_ORDER + 1 ];
opus_int32 refl_coef_Q16[ MAX_SHAPE_LPC_ORDER ];
opus_int32 AR1_Q24[ MAX_SHAPE_LPC_ORDER ];
opus_int32 AR2_Q24[ MAX_SHAPE_LPC_ORDER ];
VARDECL( opus_int16, x_windowed );
const opus_int16 *x_ptr, *pitch_res_ptr;
SAVE_STACK;
/* Point to start of first LPC analysis block */
x_ptr = x - psEnc->sCmn.la_shape;
/****************/
/* GAIN CONTROL */
/****************/
SNR_adj_dB_Q7 = psEnc->sCmn.SNR_dB_Q7;
/* Input quality is the average of the quality in the lowest two VAD bands */
psEncCtrl->input_quality_Q14 = ( opus_int )silk_RSHIFT( (opus_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 = silk_RSHIFT( silk_sigm_Q15( silk_RSHIFT_ROUND( SNR_adj_dB_Q7 -
SILK_FIX_CONST( 20.0, 7 ), 4 ) ), 1 );
/* Reduce coding SNR during low speech activity */
if( psEnc->sCmn.useCBR == 0 ) {
b_Q8 = SILK_FIX_CONST( 1.0, 8 ) - psEnc->sCmn.speech_activity_Q8;
b_Q8 = silk_SMULWB( silk_LSHIFT( b_Q8, 8 ), b_Q8 );
SNR_adj_dB_Q7 = silk_SMLAWB( SNR_adj_dB_Q7,
silk_SMULBB( SILK_FIX_CONST( -BG_SNR_DECR_dB, 7 ) >> ( 4 + 1 ), b_Q8 ), /* Q11*/
silk_SMULWB( SILK_FIX_CONST( 1.0, 14 ) + psEncCtrl->input_quality_Q14, psEncCtrl->coding_quality_Q14 ) ); /* Q12*/
}
opus_int silk_VAD_GetSA_Q8( /* O Return value, 0 if success */
silk_encoder_state *psEncC, /* I/O Encoder state */
const opus_int16 pIn[] /* I PCM input */
)
{
opus_int SA_Q15, pSNR_dB_Q7, input_tilt;
opus_int decimated_framelength1, decimated_framelength2;
opus_int decimated_framelength;
opus_int dec_subframe_length, dec_subframe_offset, SNR_Q7, i, b, s;
opus_int32 sumSquared, smooth_coef_Q16;
opus_int16 HPstateTmp;
VARDECL( opus_int16, X );
opus_int32 Xnrg[ VAD_N_BANDS ];
opus_int32 NrgToNoiseRatio_Q8[ VAD_N_BANDS ];
opus_int32 speech_nrg, x_tmp;
opus_int X_offset[ VAD_N_BANDS ];
opus_int ret = 0;
silk_VAD_state *psSilk_VAD = &psEncC->sVAD;
SAVE_STACK;
/* Safety checks */
silk_assert( VAD_N_BANDS == 4 );
silk_assert( MAX_FRAME_LENGTH >= psEncC->frame_length );
silk_assert( psEncC->frame_length <= 512 );
silk_assert( psEncC->frame_length == 8 * silk_RSHIFT( psEncC->frame_length, 3 ) );
/***********************/
/* Filter and Decimate */
/***********************/
decimated_framelength1 = silk_RSHIFT( psEncC->frame_length, 1 );
decimated_framelength2 = silk_RSHIFT( psEncC->frame_length, 2 );
decimated_framelength = silk_RSHIFT( psEncC->frame_length, 3 );
/* Decimate into 4 bands:
0 L 3L L 3L 5L
- -- - -- --
8 8 2 4 4
[0-1 kHz| temp. |1-2 kHz| 2-4 kHz | 4-8 kHz |
They're arranged to allow the minimal ( frame_length / 4 ) extra
scratch space during the downsampling process */
X_offset[ 0 ] = 0;
X_offset[ 1 ] = decimated_framelength + decimated_framelength2;
X_offset[ 2 ] = X_offset[ 1 ] + decimated_framelength;
X_offset[ 3 ] = X_offset[ 2 ] + decimated_framelength2;
ALLOC( X, X_offset[ 3 ] + decimated_framelength1, opus_int16 );
/* 0-8 kHz to 0-4 kHz and 4-8 kHz */
silk_ana_filt_bank_1( pIn, &psSilk_VAD->AnaState[ 0 ],
X, &X[ X_offset[ 3 ] ], psEncC->frame_length );
/* 0-4 kHz to 0-2 kHz and 2-4 kHz */
silk_ana_filt_bank_1( X, &psSilk_VAD->AnaState1[ 0 ],
X, &X[ X_offset[ 2 ] ], decimated_framelength1 );
/* 0-2 kHz to 0-1 kHz and 1-2 kHz */
silk_ana_filt_bank_1( X, &psSilk_VAD->AnaState2[ 0 ],
X, &X[ X_offset[ 1 ] ], decimated_framelength2 );
/*********************************************/
/* HP filter on lowest band (differentiator) */
/*********************************************/
X[ decimated_framelength - 1 ] = silk_RSHIFT( X[ decimated_framelength - 1 ], 1 );
HPstateTmp = X[ decimated_framelength - 1 ];
for( i = decimated_framelength - 1; i > 0; i-- ) {
X[ i - 1 ] = silk_RSHIFT( X[ i - 1 ], 1 );
X[ i ] -= X[ i - 1 ];
}
X[ 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 = silk_RSHIFT( psEncC->frame_length, silk_min_int( VAD_N_BANDS - b, VAD_N_BANDS - 1 ) );
/* Split length into subframe lengths */
dec_subframe_length = silk_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 * ( silk_int16_MIN / 8 ) ^ 2. */
/* Therefore we can accumulate with no risk of overflow (unless dec_subframe_length > 128) */
x_tmp = silk_RSHIFT(
X[ X_offset[ b ] + i + dec_subframe_offset ], 3 );
sumSquared = silk_SMLABB( sumSquared, x_tmp, x_tmp );
/* Safety check */
silk_assert( sumSquared >= 0 );
}
/* Add/saturate summed energy of current subframe */
if( s < VAD_INTERNAL_SUBFRAMES - 1 ) {
Xnrg[ b ] = silk_ADD_POS_SAT32( Xnrg[ b ], sumSquared );
} else {
/* Look-ahead subframe */
Xnrg[ b ] = silk_ADD_POS_SAT32( Xnrg[ b ], silk_RSHIFT( sumSquared, 1 ) );
}
dec_subframe_offset += dec_subframe_length;
}
psSilk_VAD->XnrgSubfr[ b ] = sumSquared;
}
/********************/
/* Noise estimation */
/********************/
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 ] = silk_DIV32( silk_LSHIFT( Xnrg[ b ], 8 ), psSilk_VAD->NL[ b ] + 1 );
} else {
NrgToNoiseRatio_Q8[ b ] = silk_DIV32( Xnrg[ b ], silk_RSHIFT( psSilk_VAD->NL[ b ], 8 ) + 1 );
}
/* Convert to log domain */
SNR_Q7 = silk_lin2log( NrgToNoiseRatio_Q8[ b ] ) - 8 * 128;
/* Sum-of-squares */
sumSquared = silk_SMLABB( sumSquared, SNR_Q7, SNR_Q7 ); /* Q14 */
/* Tilt measure */
if( speech_nrg < ( (opus_int32)1 << 20 ) ) {
/* Scale down SNR value for small subband speech energies */
SNR_Q7 = silk_SMULWB( silk_LSHIFT( silk_SQRT_APPROX( speech_nrg ), 6 ), SNR_Q7 );
}
input_tilt = silk_SMLAWB( input_tilt, tiltWeights[ b ], SNR_Q7 );
} else {
NrgToNoiseRatio_Q8[ b ] = 256;
}
}
/* Mean-of-squares */
sumSquared = silk_DIV32_16( sumSquared, VAD_N_BANDS ); /* Q14 */
/* Root-mean-square approximation, scale to dBs, and write to output pointer */
pSNR_dB_Q7 = (opus_int16)( 3 * silk_SQRT_APPROX( sumSquared ) ); /* Q7 */
/*********************************/
/* Speech Probability Estimation */
/*********************************/
SA_Q15 = silk_sigm_Q15( silk_SMULWB( VAD_SNR_FACTOR_Q16, pSNR_dB_Q7 ) - VAD_NEGATIVE_OFFSET_Q5 );
/**************************/
/* Frequency Tilt Measure */
/**************************/
psEncC->input_tilt_Q15 = silk_LSHIFT( 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 ) * silk_RSHIFT( Xnrg[ b ] - psSilk_VAD->NL[ b ], 4 );
}
/* Power scaling */
if( speech_nrg <= 0 ) {
SA_Q15 = silk_RSHIFT( SA_Q15, 1 );
} else if( speech_nrg < 32768 ) {
if( psEncC->frame_length == 10 * psEncC->fs_kHz ) {
speech_nrg = silk_LSHIFT_SAT32( speech_nrg, 16 );
} else {
speech_nrg = silk_LSHIFT_SAT32( speech_nrg, 15 );
}
/* square-root */
speech_nrg = silk_SQRT_APPROX( speech_nrg );
SA_Q15 = silk_SMULWB( 32768 + speech_nrg, SA_Q15 );
}
/* Copy the resulting speech activity in Q8 */
psEncC->speech_activity_Q8 = silk_min_int( silk_RSHIFT( SA_Q15, 7 ), silk_uint8_MAX );
/***********************************/
/* Energy Level and SNR estimation */
/***********************************/
/* Smoothing coefficient */
smooth_coef_Q16 = silk_SMULWB( VAD_SNR_SMOOTH_COEF_Q18, silk_SMULWB( (opus_int32)SA_Q15, SA_Q15 ) );
if( psEncC->frame_length == 10 * psEncC->fs_kHz ) {
smooth_coef_Q16 >>= 1;
}