/* Convert NLSF parameters to stable AR prediction filter coefficients */ void SKP_Silk_NLSF2A_stable( SKP_int16 pAR_Q12[ MAX_LPC_ORDER ], /* O Stabilized AR coefs [LPC_order] */ const SKP_int pNLSF[ MAX_LPC_ORDER ], /* I NLSF vector [LPC_order] */ const SKP_int LPC_order /* I LPC/LSF order */ ) { SKP_int i; SKP_int32 invGain_Q30; SKP_Silk_NLSF2A( pAR_Q12, pNLSF, LPC_order ); /* Ensure stable LPCs */ for( i = 0; i < MAX_LPC_STABILIZE_ITERATIONS; i++ ) { if( SKP_Silk_LPC_inverse_pred_gain( &invGain_Q30, pAR_Q12, LPC_order ) == 1 ) { SKP_Silk_bwexpander( pAR_Q12, LPC_order, 65536 - SKP_SMULBB( 10 + i, i ) ); /* 10_Q16 = 0.00015 */ } else { break; } } /* Reached the last iteration */ if( i == MAX_LPC_STABILIZE_ITERATIONS ) { SKP_assert( 0 ); for( i = 0; i < LPC_order; i++ ) { pAR_Q12[ i ] = 0; } } }
/* 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 ); }
void SKP_Silk_PLC_conceal( SKP_Silk_decoder_state *psDec, /* I/O Decoder state */ SKP_Silk_decoder_control *psDecCtrl, /* I/O Decoder control */ SKP_int16 signal[], /* O concealed signal */ SKP_int length /* I length of residual */ ) { SKP_int i, j, k; SKP_int16 *B_Q14, exc_buf[ MAX_FRAME_LENGTH ], *exc_buf_ptr; SKP_int16 rand_scale_Q14, A_Q12_tmp[ MAX_LPC_ORDER ]; SKP_int32 rand_seed, harm_Gain_Q15, rand_Gain_Q15; SKP_int lag, idx, shift1, shift2; SKP_int32 energy1, energy2, *rand_ptr, *pred_lag_ptr, Atmp; SKP_int32 sig_Q10[ MAX_FRAME_LENGTH ], *sig_Q10_ptr, LPC_exc_Q10, LPC_pred_Q10, LTP_pred_Q14; SKP_Silk_PLC_struct *psPLC; psPLC = &psDec->sPLC; /* Update LTP buffer */ SKP_memcpy( psDec->sLTP_Q16, &psDec->sLTP_Q16[ psDec->frame_length ], psDec->frame_length * sizeof( SKP_int32 ) ); /* LPC concealment. Apply BWE to previous LPC */ SKP_Silk_bwexpander( psPLC->prevLPC_Q12, psDec->LPC_order, BWE_COEF_Q16 ); /* Find random noise component */ /* Scale previous excitation signal */ exc_buf_ptr = exc_buf; for( k = ( NB_SUBFR >> 1 ); k < NB_SUBFR; k++ ) { for( i = 0; i < psDec->subfr_length; i++ ) { exc_buf_ptr[ i ] = ( SKP_int16 )SKP_RSHIFT( SKP_SMULWW( psDec->exc_Q10[ i + k * psDec->subfr_length ], psPLC->prevGain_Q16[ k ] ), 10 ); } exc_buf_ptr += psDec->subfr_length; } /* Find the subframe with lowest energy of the last two and use that as random noise generator */ SKP_Silk_sum_sqr_shift( &energy1, &shift1, exc_buf, psDec->subfr_length ); SKP_Silk_sum_sqr_shift( &energy2, &shift2, &exc_buf[ psDec->subfr_length ], psDec->subfr_length ); if( SKP_RSHIFT( energy1, shift2 ) < SKP_RSHIFT( energy1, shift2 ) ) { /* First sub-frame has lowest energy */ rand_ptr = &psDec->exc_Q10[ SKP_max_int( 0, 3 * psDec->subfr_length - RAND_BUF_SIZE ) ]; } else { /* Second sub-frame has lowest energy */ rand_ptr = &psDec->exc_Q10[ SKP_max_int( 0, psDec->frame_length - RAND_BUF_SIZE ) ]; } /* Setup Gain to random noise component */ B_Q14 = psPLC->LTPCoef_Q14; rand_scale_Q14 = psPLC->randScale_Q14; /* Setup attenuation gains */ harm_Gain_Q15 = HARM_ATT_Q15[ SKP_min_int( NB_ATT - 1, psDec->lossCnt ) ]; if( psDec->prev_sigtype == SIG_TYPE_VOICED ) { rand_Gain_Q15 = PLC_RAND_ATTENUATE_V_Q15[ SKP_min_int( NB_ATT - 1, psDec->lossCnt ) ]; } else { rand_Gain_Q15 = PLC_RAND_ATTENUATE_UV_Q15[ SKP_min_int( NB_ATT - 1, psDec->lossCnt ) ]; } /* First Lost frame */ if( psDec->lossCnt == 0 ) { rand_scale_Q14 = (1 << 14 ); /* Reduce random noise Gain for voiced frames */ if( psDec->prev_sigtype == SIG_TYPE_VOICED ) { for( i = 0; i < LTP_ORDER; i++ ) { rand_scale_Q14 -= B_Q14[ i ]; } rand_scale_Q14 = SKP_max_16( 3277, rand_scale_Q14 ); /* 0.2 */ rand_scale_Q14 = ( SKP_int16 )SKP_RSHIFT( SKP_SMULBB( rand_scale_Q14, psPLC->prevLTP_scale_Q14 ), 14 ); } /* Reduce random noise for unvoiced frames with high LPC gain */ if( psDec->prev_sigtype == SIG_TYPE_UNVOICED ) { SKP_int32 invGain_Q30, down_scale_Q30; SKP_Silk_LPC_inverse_pred_gain( &invGain_Q30, psPLC->prevLPC_Q12, psDec->LPC_order ); down_scale_Q30 = SKP_min_32( SKP_RSHIFT( ( 1 << 30 ), LOG2_INV_LPC_GAIN_HIGH_THRES ), invGain_Q30 ); down_scale_Q30 = SKP_max_32( SKP_RSHIFT( ( 1 << 30 ), LOG2_INV_LPC_GAIN_LOW_THRES ), down_scale_Q30 ); down_scale_Q30 = SKP_LSHIFT( down_scale_Q30, LOG2_INV_LPC_GAIN_HIGH_THRES ); rand_Gain_Q15 = SKP_RSHIFT( SKP_SMULWB( down_scale_Q30, rand_Gain_Q15 ), 14 ); } } rand_seed = psPLC->rand_seed; lag = SKP_RSHIFT_ROUND( psPLC->pitchL_Q8, 8 ); psDec->sLTP_buf_idx = psDec->frame_length; /***************************/ /* LTP synthesis filtering */ /***************************/ sig_Q10_ptr = sig_Q10; for( k = 0; k < NB_SUBFR; k++ ) { /* Setup pointer */ pred_lag_ptr = &psDec->sLTP_Q16[ psDec->sLTP_buf_idx - lag + LTP_ORDER / 2 ]; for( i = 0; i < psDec->subfr_length; i++ ) { rand_seed = SKP_RAND( rand_seed ); idx = SKP_RSHIFT( rand_seed, 25 ) & RAND_BUF_MASK; /* Unrolled loop */ LTP_pred_Q14 = SKP_SMULWB( pred_lag_ptr[ 0 ], B_Q14[ 0 ] ); LTP_pred_Q14 = SKP_SMLAWB( LTP_pred_Q14, pred_lag_ptr[ -1 ], B_Q14[ 1 ] ); LTP_pred_Q14 = SKP_SMLAWB( LTP_pred_Q14, pred_lag_ptr[ -2 ], B_Q14[ 2 ] ); LTP_pred_Q14 = SKP_SMLAWB( LTP_pred_Q14, pred_lag_ptr[ -3 ], B_Q14[ 3 ] ); LTP_pred_Q14 = SKP_SMLAWB( LTP_pred_Q14, pred_lag_ptr[ -4 ], B_Q14[ 4 ] ); pred_lag_ptr++; /* Generate LPC residual */ LPC_exc_Q10 = SKP_LSHIFT( SKP_SMULWB( rand_ptr[ idx ], rand_scale_Q14 ), 2 ); /* Random noise part */ LPC_exc_Q10 = SKP_ADD32( LPC_exc_Q10, SKP_RSHIFT_ROUND( LTP_pred_Q14, 4 ) ); /* Harmonic part */ /* Update states */ psDec->sLTP_Q16[ psDec->sLTP_buf_idx ] = SKP_LSHIFT( LPC_exc_Q10, 6 ); psDec->sLTP_buf_idx++; /* Save LPC residual */ sig_Q10_ptr[ i ] = LPC_exc_Q10; } sig_Q10_ptr += psDec->subfr_length; /* Gradually reduce LTP gain */ for( j = 0; j < LTP_ORDER; j++ ) { B_Q14[ j ] = SKP_RSHIFT( SKP_SMULBB( harm_Gain_Q15, B_Q14[ j ] ), 15 ); } /* Gradually reduce excitation gain */ rand_scale_Q14 = SKP_RSHIFT( SKP_SMULBB( rand_scale_Q14, rand_Gain_Q15 ), 15 ); /* Slowly increase pitch lag */ psPLC->pitchL_Q8 += SKP_SMULWB( psPLC->pitchL_Q8, PITCH_DRIFT_FAC_Q16 ); psPLC->pitchL_Q8 = SKP_min_32( psPLC->pitchL_Q8, SKP_LSHIFT( SKP_SMULBB( MAX_PITCH_LAG_MS, psDec->fs_kHz ), 8 ) ); lag = SKP_RSHIFT_ROUND( psPLC->pitchL_Q8, 8 ); } /***************************/ /* LPC synthesis filtering */ /***************************/ sig_Q10_ptr = sig_Q10; /* Preload LPC coeficients to array on stack. Gives small performance gain */ SKP_memcpy( A_Q12_tmp, psPLC->prevLPC_Q12, psDec->LPC_order * sizeof( SKP_int16 ) ); SKP_assert( psDec->LPC_order >= 10 ); /* check that unrolling works */ for( k = 0; k < NB_SUBFR; k++ ) { for( i = 0; i < psDec->subfr_length; i++ ){ /* unrolled */ Atmp = *( ( SKP_int32* )&A_Q12_tmp[ 0 ] ); /* read two coefficients at once */ LPC_pred_Q10 = SKP_SMULWB( psDec->sLPC_Q14[ MAX_LPC_ORDER + i - 1 ], Atmp ); LPC_pred_Q10 = SKP_SMLAWT( LPC_pred_Q10, psDec->sLPC_Q14[ MAX_LPC_ORDER + i - 2 ], Atmp ); Atmp = *( ( SKP_int32* )&A_Q12_tmp[ 2 ] ); LPC_pred_Q10 = SKP_SMLAWB( LPC_pred_Q10, psDec->sLPC_Q14[ MAX_LPC_ORDER + i - 3 ], Atmp ); LPC_pred_Q10 = SKP_SMLAWT( LPC_pred_Q10, psDec->sLPC_Q14[ MAX_LPC_ORDER + i - 4 ], Atmp ); Atmp = *( ( SKP_int32* )&A_Q12_tmp[ 4 ] ); LPC_pred_Q10 = SKP_SMLAWB( LPC_pred_Q10, psDec->sLPC_Q14[ MAX_LPC_ORDER + i - 5 ], Atmp ); LPC_pred_Q10 = SKP_SMLAWT( LPC_pred_Q10, psDec->sLPC_Q14[ MAX_LPC_ORDER + i - 6 ], Atmp ); Atmp = *( ( SKP_int32* )&A_Q12_tmp[ 6 ] ); LPC_pred_Q10 = SKP_SMLAWB( LPC_pred_Q10, psDec->sLPC_Q14[ MAX_LPC_ORDER + i - 7 ], Atmp ); LPC_pred_Q10 = SKP_SMLAWT( LPC_pred_Q10, psDec->sLPC_Q14[ MAX_LPC_ORDER + i - 8 ], Atmp ); Atmp = *( ( SKP_int32* )&A_Q12_tmp[ 8 ] ); LPC_pred_Q10 = SKP_SMLAWB( LPC_pred_Q10, psDec->sLPC_Q14[ MAX_LPC_ORDER + i - 9 ], Atmp ); LPC_pred_Q10 = SKP_SMLAWT( LPC_pred_Q10, psDec->sLPC_Q14[ MAX_LPC_ORDER + i - 10 ], Atmp ); for( j = 10 ; j < psDec->LPC_order ; j+=2 ) { Atmp = *( ( SKP_int32* )&A_Q12_tmp[ j ] ); LPC_pred_Q10 = SKP_SMLAWB( LPC_pred_Q10, psDec->sLPC_Q14[ MAX_LPC_ORDER + i - 1 - j ], Atmp ); LPC_pred_Q10 = SKP_SMLAWT( LPC_pred_Q10, psDec->sLPC_Q14[ MAX_LPC_ORDER + i - 2 - j ], Atmp ); } /* Add prediction to LPC residual */ sig_Q10_ptr[ i ] = SKP_ADD32( sig_Q10_ptr[ i ], LPC_pred_Q10 ); /* Update states */ psDec->sLPC_Q14[ MAX_LPC_ORDER + i ] = SKP_LSHIFT( sig_Q10_ptr[ i ], 4 ); } sig_Q10_ptr += psDec->subfr_length; /* Update LPC filter state */ SKP_memcpy( psDec->sLPC_Q14, &psDec->sLPC_Q14[ psDec->subfr_length ], MAX_LPC_ORDER * sizeof( SKP_int32 ) ); } /* Scale with Gain */ for( i = 0; i < psDec->frame_length; i++ ) { signal[ i ] = ( SKP_int16 )SKP_SAT16( SKP_RSHIFT_ROUND( SKP_SMULWW( sig_Q10[ i ], psPLC->prevGain_Q16[ NB_SUBFR - 1 ] ), 10 ) ); } /**************************************/ /* Update states */ /**************************************/ psPLC->rand_seed = rand_seed; psPLC->randScale_Q14 = rand_scale_Q14; for( i = 0; i < NB_SUBFR; i++ ) { psDecCtrl->pitchL[ i ] = lag; } }
/* 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 [ 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 ); }