cannam@154: /*********************************************************************** cannam@154: Copyright (c) 2006-2011, Skype Limited. All rights reserved. cannam@154: Redistribution and use in source and binary forms, with or without cannam@154: modification, are permitted provided that the following conditions cannam@154: are met: cannam@154: - Redistributions of source code must retain the above copyright notice, cannam@154: this list of conditions and the following disclaimer. cannam@154: - Redistributions in binary form must reproduce the above copyright cannam@154: notice, this list of conditions and the following disclaimer in the cannam@154: documentation and/or other materials provided with the distribution. cannam@154: - Neither the name of Internet Society, IETF or IETF Trust, nor the cannam@154: names of specific contributors, may be used to endorse or promote cannam@154: products derived from this software without specific prior written cannam@154: permission. cannam@154: THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS" cannam@154: AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE cannam@154: IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE cannam@154: ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER OR CONTRIBUTORS BE cannam@154: LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR cannam@154: CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF cannam@154: SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS cannam@154: INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN cannam@154: CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) cannam@154: ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE cannam@154: POSSIBILITY OF SUCH DAMAGE. cannam@154: ***********************************************************************/ cannam@154: cannam@154: #ifdef HAVE_CONFIG_H cannam@154: #include "config.h" cannam@154: #endif cannam@154: cannam@154: #include "main_FLP.h" cannam@154: #include "tuning_parameters.h" cannam@154: cannam@154: /* Compute gain to make warped filter coefficients have a zero mean log frequency response on a */ cannam@154: /* non-warped frequency scale. (So that it can be implemented with a minimum-phase monic filter.) */ cannam@154: /* Note: A monic filter is one with the first coefficient equal to 1.0. In Silk we omit the first */ cannam@154: /* coefficient in an array of coefficients, for monic filters. */ cannam@154: static OPUS_INLINE silk_float warped_gain( cannam@154: const silk_float *coefs, cannam@154: silk_float lambda, cannam@154: opus_int order cannam@154: ) { cannam@154: opus_int i; cannam@154: silk_float gain; cannam@154: cannam@154: lambda = -lambda; cannam@154: gain = coefs[ order - 1 ]; cannam@154: for( i = order - 2; i >= 0; i-- ) { cannam@154: gain = lambda * gain + coefs[ i ]; cannam@154: } cannam@154: return (silk_float)( 1.0f / ( 1.0f - lambda * gain ) ); cannam@154: } cannam@154: cannam@154: /* Convert warped filter coefficients to monic pseudo-warped coefficients and limit maximum */ cannam@154: /* amplitude of monic warped coefficients by using bandwidth expansion on the true coefficients */ cannam@154: static OPUS_INLINE void warped_true2monic_coefs( cannam@154: silk_float *coefs, cannam@154: silk_float lambda, cannam@154: silk_float limit, cannam@154: opus_int order cannam@154: ) { cannam@154: opus_int i, iter, ind = 0; cannam@154: silk_float tmp, maxabs, chirp, gain; cannam@154: cannam@154: /* Convert to monic coefficients */ cannam@154: for( i = order - 1; i > 0; i-- ) { cannam@154: coefs[ i - 1 ] -= lambda * coefs[ i ]; cannam@154: } cannam@154: gain = ( 1.0f - lambda * lambda ) / ( 1.0f + lambda * coefs[ 0 ] ); cannam@154: for( i = 0; i < order; i++ ) { cannam@154: coefs[ i ] *= gain; cannam@154: } cannam@154: cannam@154: /* Limit */ cannam@154: for( iter = 0; iter < 10; iter++ ) { cannam@154: /* Find maximum absolute value */ cannam@154: maxabs = -1.0f; cannam@154: for( i = 0; i < order; i++ ) { cannam@154: tmp = silk_abs_float( coefs[ i ] ); cannam@154: if( tmp > maxabs ) { cannam@154: maxabs = tmp; cannam@154: ind = i; cannam@154: } cannam@154: } cannam@154: if( maxabs <= limit ) { cannam@154: /* Coefficients are within range - done */ cannam@154: return; cannam@154: } cannam@154: cannam@154: /* Convert back to true warped coefficients */ cannam@154: for( i = 1; i < order; i++ ) { cannam@154: coefs[ i - 1 ] += lambda * coefs[ i ]; cannam@154: } cannam@154: gain = 1.0f / gain; cannam@154: for( i = 0; i < order; i++ ) { cannam@154: coefs[ i ] *= gain; cannam@154: } cannam@154: cannam@154: /* Apply bandwidth expansion */ cannam@154: chirp = 0.99f - ( 0.8f + 0.1f * iter ) * ( maxabs - limit ) / ( maxabs * ( ind + 1 ) ); cannam@154: silk_bwexpander_FLP( coefs, order, chirp ); cannam@154: cannam@154: /* Convert to monic warped coefficients */ cannam@154: for( i = order - 1; i > 0; i-- ) { cannam@154: coefs[ i - 1 ] -= lambda * coefs[ i ]; cannam@154: } cannam@154: gain = ( 1.0f - lambda * lambda ) / ( 1.0f + lambda * coefs[ 0 ] ); cannam@154: for( i = 0; i < order; i++ ) { cannam@154: coefs[ i ] *= gain; cannam@154: } cannam@154: } cannam@154: silk_assert( 0 ); cannam@154: } cannam@154: cannam@154: static OPUS_INLINE void limit_coefs( cannam@154: silk_float *coefs, cannam@154: silk_float limit, cannam@154: opus_int order cannam@154: ) { cannam@154: opus_int i, iter, ind = 0; cannam@154: silk_float tmp, maxabs, chirp; cannam@154: cannam@154: for( iter = 0; iter < 10; iter++ ) { cannam@154: /* Find maximum absolute value */ cannam@154: maxabs = -1.0f; cannam@154: for( i = 0; i < order; i++ ) { cannam@154: tmp = silk_abs_float( coefs[ i ] ); cannam@154: if( tmp > maxabs ) { cannam@154: maxabs = tmp; cannam@154: ind = i; cannam@154: } cannam@154: } cannam@154: if( maxabs <= limit ) { cannam@154: /* Coefficients are within range - done */ cannam@154: return; cannam@154: } cannam@154: cannam@154: /* Apply bandwidth expansion */ cannam@154: chirp = 0.99f - ( 0.8f + 0.1f * iter ) * ( maxabs - limit ) / ( maxabs * ( ind + 1 ) ); cannam@154: silk_bwexpander_FLP( coefs, order, chirp ); cannam@154: } cannam@154: silk_assert( 0 ); cannam@154: } cannam@154: cannam@154: /* Compute noise shaping coefficients and initial gain values */ cannam@154: void silk_noise_shape_analysis_FLP( cannam@154: silk_encoder_state_FLP *psEnc, /* I/O Encoder state FLP */ cannam@154: silk_encoder_control_FLP *psEncCtrl, /* I/O Encoder control FLP */ cannam@154: const silk_float *pitch_res, /* I LPC residual from pitch analysis */ cannam@154: const silk_float *x /* I Input signal [frame_length + la_shape] */ cannam@154: ) cannam@154: { cannam@154: silk_shape_state_FLP *psShapeSt = &psEnc->sShape; cannam@154: opus_int k, nSamples, nSegs; cannam@154: silk_float SNR_adj_dB, HarmShapeGain, Tilt; cannam@154: silk_float nrg, log_energy, log_energy_prev, energy_variation; cannam@154: silk_float BWExp, gain_mult, gain_add, strength, b, warping; cannam@154: silk_float x_windowed[ SHAPE_LPC_WIN_MAX ]; cannam@154: silk_float auto_corr[ MAX_SHAPE_LPC_ORDER + 1 ]; cannam@154: silk_float rc[ MAX_SHAPE_LPC_ORDER + 1 ]; cannam@154: const silk_float *x_ptr, *pitch_res_ptr; cannam@154: cannam@154: /* Point to start of first LPC analysis block */ cannam@154: x_ptr = x - psEnc->sCmn.la_shape; cannam@154: cannam@154: /****************/ cannam@154: /* GAIN CONTROL */ cannam@154: /****************/ cannam@154: SNR_adj_dB = psEnc->sCmn.SNR_dB_Q7 * ( 1 / 128.0f ); cannam@154: cannam@154: /* Input quality is the average of the quality in the lowest two VAD bands */ cannam@154: psEncCtrl->input_quality = 0.5f * ( psEnc->sCmn.input_quality_bands_Q15[ 0 ] + psEnc->sCmn.input_quality_bands_Q15[ 1 ] ) * ( 1.0f / 32768.0f ); cannam@154: cannam@154: /* Coding quality level, between 0.0 and 1.0 */ cannam@154: psEncCtrl->coding_quality = silk_sigmoid( 0.25f * ( SNR_adj_dB - 20.0f ) ); cannam@154: cannam@154: if( psEnc->sCmn.useCBR == 0 ) { cannam@154: /* Reduce coding SNR during low speech activity */ cannam@154: b = 1.0f - psEnc->sCmn.speech_activity_Q8 * ( 1.0f / 256.0f ); cannam@154: SNR_adj_dB -= BG_SNR_DECR_dB * psEncCtrl->coding_quality * ( 0.5f + 0.5f * psEncCtrl->input_quality ) * b * b; cannam@154: } cannam@154: cannam@154: if( psEnc->sCmn.indices.signalType == TYPE_VOICED ) { cannam@154: /* Reduce gains for periodic signals */ cannam@154: SNR_adj_dB += HARM_SNR_INCR_dB * psEnc->LTPCorr; cannam@154: } else { cannam@154: /* For unvoiced signals and low-quality input, adjust the quality slower than SNR_dB setting */ cannam@154: SNR_adj_dB += ( -0.4f * psEnc->sCmn.SNR_dB_Q7 * ( 1 / 128.0f ) + 6.0f ) * ( 1.0f - psEncCtrl->input_quality ); cannam@154: } cannam@154: cannam@154: /*************************/ cannam@154: /* SPARSENESS PROCESSING */ cannam@154: /*************************/ cannam@154: /* Set quantizer offset */ cannam@154: if( psEnc->sCmn.indices.signalType == TYPE_VOICED ) { cannam@154: /* Initially set to 0; may be overruled in process_gains(..) */ cannam@154: psEnc->sCmn.indices.quantOffsetType = 0; cannam@154: } else { cannam@154: /* Sparseness measure, based on relative fluctuations of energy per 2 milliseconds */ cannam@154: nSamples = 2 * psEnc->sCmn.fs_kHz; cannam@154: energy_variation = 0.0f; cannam@154: log_energy_prev = 0.0f; cannam@154: pitch_res_ptr = pitch_res; cannam@154: nSegs = silk_SMULBB( SUB_FRAME_LENGTH_MS, psEnc->sCmn.nb_subfr ) / 2; cannam@154: for( k = 0; k < nSegs; k++ ) { cannam@154: nrg = ( silk_float )nSamples + ( silk_float )silk_energy_FLP( pitch_res_ptr, nSamples ); cannam@154: log_energy = silk_log2( nrg ); cannam@154: if( k > 0 ) { cannam@154: energy_variation += silk_abs_float( log_energy - log_energy_prev ); cannam@154: } cannam@154: log_energy_prev = log_energy; cannam@154: pitch_res_ptr += nSamples; cannam@154: } cannam@154: cannam@154: /* Set quantization offset depending on sparseness measure */ cannam@154: if( energy_variation > ENERGY_VARIATION_THRESHOLD_QNT_OFFSET * (nSegs-1) ) { cannam@154: psEnc->sCmn.indices.quantOffsetType = 0; cannam@154: } else { cannam@154: psEnc->sCmn.indices.quantOffsetType = 1; cannam@154: } cannam@154: } cannam@154: cannam@154: /*******************************/ cannam@154: /* Control bandwidth expansion */ cannam@154: /*******************************/ cannam@154: /* More BWE for signals with high prediction gain */ cannam@154: strength = FIND_PITCH_WHITE_NOISE_FRACTION * psEncCtrl->predGain; /* between 0.0 and 1.0 */ cannam@154: BWExp = BANDWIDTH_EXPANSION / ( 1.0f + strength * strength ); cannam@154: cannam@154: /* Slightly more warping in analysis will move quantization noise up in frequency, where it's better masked */ cannam@154: warping = (silk_float)psEnc->sCmn.warping_Q16 / 65536.0f + 0.01f * psEncCtrl->coding_quality; cannam@154: cannam@154: /********************************************/ cannam@154: /* Compute noise shaping AR coefs and gains */ cannam@154: /********************************************/ cannam@154: for( k = 0; k < psEnc->sCmn.nb_subfr; k++ ) { cannam@154: /* Apply window: sine slope followed by flat part followed by cosine slope */ cannam@154: opus_int shift, slope_part, flat_part; cannam@154: flat_part = psEnc->sCmn.fs_kHz * 3; cannam@154: slope_part = ( psEnc->sCmn.shapeWinLength - flat_part ) / 2; cannam@154: cannam@154: silk_apply_sine_window_FLP( x_windowed, x_ptr, 1, slope_part ); cannam@154: shift = slope_part; cannam@154: silk_memcpy( x_windowed + shift, x_ptr + shift, flat_part * sizeof(silk_float) ); cannam@154: shift += flat_part; cannam@154: silk_apply_sine_window_FLP( x_windowed + shift, x_ptr + shift, 2, slope_part ); cannam@154: cannam@154: /* Update pointer: next LPC analysis block */ cannam@154: x_ptr += psEnc->sCmn.subfr_length; cannam@154: cannam@154: if( psEnc->sCmn.warping_Q16 > 0 ) { cannam@154: /* Calculate warped auto correlation */ cannam@154: silk_warped_autocorrelation_FLP( auto_corr, x_windowed, warping, cannam@154: psEnc->sCmn.shapeWinLength, psEnc->sCmn.shapingLPCOrder ); cannam@154: } else { cannam@154: /* Calculate regular auto correlation */ cannam@154: silk_autocorrelation_FLP( auto_corr, x_windowed, psEnc->sCmn.shapeWinLength, psEnc->sCmn.shapingLPCOrder + 1 ); cannam@154: } cannam@154: cannam@154: /* Add white noise, as a fraction of energy */ cannam@154: auto_corr[ 0 ] += auto_corr[ 0 ] * SHAPE_WHITE_NOISE_FRACTION + 1.0f; cannam@154: cannam@154: /* Convert correlations to prediction coefficients, and compute residual energy */ cannam@154: nrg = silk_schur_FLP( rc, auto_corr, psEnc->sCmn.shapingLPCOrder ); cannam@154: silk_k2a_FLP( &psEncCtrl->AR[ k * MAX_SHAPE_LPC_ORDER ], rc, psEnc->sCmn.shapingLPCOrder ); cannam@154: psEncCtrl->Gains[ k ] = ( silk_float )sqrt( nrg ); cannam@154: cannam@154: if( psEnc->sCmn.warping_Q16 > 0 ) { cannam@154: /* Adjust gain for warping */ cannam@154: psEncCtrl->Gains[ k ] *= warped_gain( &psEncCtrl->AR[ k * MAX_SHAPE_LPC_ORDER ], warping, psEnc->sCmn.shapingLPCOrder ); cannam@154: } cannam@154: cannam@154: /* Bandwidth expansion for synthesis filter shaping */ cannam@154: silk_bwexpander_FLP( &psEncCtrl->AR[ k * MAX_SHAPE_LPC_ORDER ], psEnc->sCmn.shapingLPCOrder, BWExp ); cannam@154: cannam@154: if( psEnc->sCmn.warping_Q16 > 0 ) { cannam@154: /* Convert to monic warped prediction coefficients and limit absolute values */ cannam@154: warped_true2monic_coefs( &psEncCtrl->AR[ k * MAX_SHAPE_LPC_ORDER ], warping, 3.999f, psEnc->sCmn.shapingLPCOrder ); cannam@154: } else { cannam@154: /* Limit absolute values */ cannam@154: limit_coefs( &psEncCtrl->AR[ k * MAX_SHAPE_LPC_ORDER ], 3.999f, psEnc->sCmn.shapingLPCOrder ); cannam@154: } cannam@154: } cannam@154: cannam@154: /*****************/ cannam@154: /* Gain tweaking */ cannam@154: /*****************/ cannam@154: /* Increase gains during low speech activity */ cannam@154: gain_mult = (silk_float)pow( 2.0f, -0.16f * SNR_adj_dB ); cannam@154: gain_add = (silk_float)pow( 2.0f, 0.16f * MIN_QGAIN_DB ); cannam@154: for( k = 0; k < psEnc->sCmn.nb_subfr; k++ ) { cannam@154: psEncCtrl->Gains[ k ] *= gain_mult; cannam@154: psEncCtrl->Gains[ k ] += gain_add; cannam@154: } cannam@154: cannam@154: /************************************************/ cannam@154: /* Control low-frequency shaping and noise tilt */ cannam@154: /************************************************/ cannam@154: /* Less low frequency shaping for noisy inputs */ cannam@154: 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 ) ); cannam@154: strength *= psEnc->sCmn.speech_activity_Q8 * ( 1.0f / 256.0f ); cannam@154: if( psEnc->sCmn.indices.signalType == TYPE_VOICED ) { cannam@154: /* Reduce low frequencies quantization noise for periodic signals, depending on pitch lag */ cannam@154: /*f = 400; freqz([1, -0.98 + 2e-4 * f], [1, -0.97 + 7e-4 * f], 2^12, Fs); axis([0, 1000, -10, 1])*/ cannam@154: for( k = 0; k < psEnc->sCmn.nb_subfr; k++ ) { cannam@154: b = 0.2f / psEnc->sCmn.fs_kHz + 3.0f / psEncCtrl->pitchL[ k ]; cannam@154: psEncCtrl->LF_MA_shp[ k ] = -1.0f + b; cannam@154: psEncCtrl->LF_AR_shp[ k ] = 1.0f - b - b * strength; cannam@154: } cannam@154: Tilt = - HP_NOISE_COEF - cannam@154: (1 - HP_NOISE_COEF) * HARM_HP_NOISE_COEF * psEnc->sCmn.speech_activity_Q8 * ( 1.0f / 256.0f ); cannam@154: } else { cannam@154: b = 1.3f / psEnc->sCmn.fs_kHz; cannam@154: psEncCtrl->LF_MA_shp[ 0 ] = -1.0f + b; cannam@154: psEncCtrl->LF_AR_shp[ 0 ] = 1.0f - b - b * strength * 0.6f; cannam@154: for( k = 1; k < psEnc->sCmn.nb_subfr; k++ ) { cannam@154: psEncCtrl->LF_MA_shp[ k ] = psEncCtrl->LF_MA_shp[ 0 ]; cannam@154: psEncCtrl->LF_AR_shp[ k ] = psEncCtrl->LF_AR_shp[ 0 ]; cannam@154: } cannam@154: Tilt = -HP_NOISE_COEF; cannam@154: } cannam@154: cannam@154: /****************************/ cannam@154: /* HARMONIC SHAPING CONTROL */ cannam@154: /****************************/ cannam@154: if( USE_HARM_SHAPING && psEnc->sCmn.indices.signalType == TYPE_VOICED ) { cannam@154: /* Harmonic noise shaping */ cannam@154: HarmShapeGain = HARMONIC_SHAPING; cannam@154: cannam@154: /* More harmonic noise shaping for high bitrates or noisy input */ cannam@154: HarmShapeGain += HIGH_RATE_OR_LOW_QUALITY_HARMONIC_SHAPING * cannam@154: ( 1.0f - ( 1.0f - psEncCtrl->coding_quality ) * psEncCtrl->input_quality ); cannam@154: cannam@154: /* Less harmonic noise shaping for less periodic signals */ cannam@154: HarmShapeGain *= ( silk_float )sqrt( psEnc->LTPCorr ); cannam@154: } else { cannam@154: HarmShapeGain = 0.0f; cannam@154: } cannam@154: cannam@154: /*************************/ cannam@154: /* Smooth over subframes */ cannam@154: /*************************/ cannam@154: for( k = 0; k < psEnc->sCmn.nb_subfr; k++ ) { cannam@154: psShapeSt->HarmShapeGain_smth += SUBFR_SMTH_COEF * ( HarmShapeGain - psShapeSt->HarmShapeGain_smth ); cannam@154: psEncCtrl->HarmShapeGain[ k ] = psShapeSt->HarmShapeGain_smth; cannam@154: psShapeSt->Tilt_smth += SUBFR_SMTH_COEF * ( Tilt - psShapeSt->Tilt_smth ); cannam@154: psEncCtrl->Tilt[ k ] = psShapeSt->Tilt_smth; cannam@154: } cannam@154: }