• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
1 /***********************************************************************
2 Copyright (c) 2006-2011, Skype Limited. All rights reserved.
3 Redistribution and use in source and binary forms, with or without
4 modification, are permitted provided that the following conditions
5 are met:
6 - Redistributions of source code must retain the above copyright notice,
7 this list of conditions and the following disclaimer.
8 - Redistributions in binary form must reproduce the above copyright
9 notice, this list of conditions and the following disclaimer in the
10 documentation and/or other materials provided with the distribution.
11 - Neither the name of Internet Society, IETF or IETF Trust, nor the
12 names of specific contributors, may be used to endorse or promote
13 products derived from this software without specific prior written
14 permission.
15 THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
16 AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
17 IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
18 ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER OR CONTRIBUTORS BE
19 LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR
20 CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF
21 SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS
22 INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN
23 CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE)
24 ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE
25 POSSIBILITY OF SUCH DAMAGE.
26 ***********************************************************************/
27 
28 #ifdef HAVE_CONFIG_H
29 #include "config.h"
30 #endif
31 
32 #include "main_FLP.h"
33 #include "tuning_parameters.h"
34 
35 /* Compute gain to make warped filter coefficients have a zero mean log frequency response on a   */
36 /* non-warped frequency scale. (So that it can be implemented with a minimum-phase monic filter.) */
37 /* Note: A monic filter is one with the first coefficient equal to 1.0. In Silk we omit the first */
38 /* coefficient in an array of coefficients, for monic filters.                                    */
warped_gain(const silk_float * coefs,silk_float lambda,opus_int order)39 static OPUS_INLINE silk_float warped_gain(
40     const silk_float     *coefs,
41     silk_float           lambda,
42     opus_int             order
43 ) {
44     opus_int   i;
45     silk_float gain;
46 
47     lambda = -lambda;
48     gain = coefs[ order - 1 ];
49     for( i = order - 2; i >= 0; i-- ) {
50         gain = lambda * gain + coefs[ i ];
51     }
52     return (silk_float)( 1.0f / ( 1.0f - lambda * gain ) );
53 }
54 
55 /* Convert warped filter coefficients to monic pseudo-warped coefficients and limit maximum     */
56 /* amplitude of monic warped coefficients by using bandwidth expansion on the true coefficients */
warped_true2monic_coefs(silk_float * coefs,silk_float lambda,silk_float limit,opus_int order)57 static OPUS_INLINE void warped_true2monic_coefs(
58     silk_float           *coefs,
59     silk_float           lambda,
60     silk_float           limit,
61     opus_int             order
62 ) {
63     opus_int   i, iter, ind = 0;
64     silk_float tmp, maxabs, chirp, gain;
65 
66     /* Convert to monic coefficients */
67     for( i = order - 1; i > 0; i-- ) {
68         coefs[ i - 1 ] -= lambda * coefs[ i ];
69     }
70     gain = ( 1.0f - lambda * lambda ) / ( 1.0f + lambda * coefs[ 0 ] );
71     for( i = 0; i < order; i++ ) {
72         coefs[ i ] *= gain;
73     }
74 
75     /* Limit */
76     for( iter = 0; iter < 10; iter++ ) {
77         /* Find maximum absolute value */
78         maxabs = -1.0f;
79         for( i = 0; i < order; i++ ) {
80             tmp = silk_abs_float( coefs[ i ] );
81             if( tmp > maxabs ) {
82                 maxabs = tmp;
83                 ind = i;
84             }
85         }
86         if( maxabs <= limit ) {
87             /* Coefficients are within range - done */
88             return;
89         }
90 
91         /* Convert back to true warped coefficients */
92         for( i = 1; i < order; i++ ) {
93             coefs[ i - 1 ] += lambda * coefs[ i ];
94         }
95         gain = 1.0f / gain;
96         for( i = 0; i < order; i++ ) {
97             coefs[ i ] *= gain;
98         }
99 
100         /* Apply bandwidth expansion */
101         chirp = 0.99f - ( 0.8f + 0.1f * iter ) * ( maxabs - limit ) / ( maxabs * ( ind + 1 ) );
102         silk_bwexpander_FLP( coefs, order, chirp );
103 
104         /* Convert to monic warped coefficients */
105         for( i = order - 1; i > 0; i-- ) {
106             coefs[ i - 1 ] -= lambda * coefs[ i ];
107         }
108         gain = ( 1.0f - lambda * lambda ) / ( 1.0f + lambda * coefs[ 0 ] );
109         for( i = 0; i < order; i++ ) {
110             coefs[ i ] *= gain;
111         }
112     }
113     silk_assert( 0 );
114 }
115 
limit_coefs(silk_float * coefs,silk_float limit,opus_int order)116 static OPUS_INLINE void limit_coefs(
117     silk_float           *coefs,
118     silk_float           limit,
119     opus_int             order
120 ) {
121     opus_int   i, iter, ind = 0;
122     silk_float tmp, maxabs, chirp;
123 
124     for( iter = 0; iter < 10; iter++ ) {
125         /* Find maximum absolute value */
126         maxabs = -1.0f;
127         for( i = 0; i < order; i++ ) {
128             tmp = silk_abs_float( coefs[ i ] );
129             if( tmp > maxabs ) {
130                 maxabs = tmp;
131                 ind = i;
132             }
133         }
134         if( maxabs <= limit ) {
135             /* Coefficients are within range - done */
136             return;
137         }
138 
139         /* Apply bandwidth expansion */
140         chirp = 0.99f - ( 0.8f + 0.1f * iter ) * ( maxabs - limit ) / ( maxabs * ( ind + 1 ) );
141         silk_bwexpander_FLP( coefs, order, chirp );
142     }
143     silk_assert( 0 );
144 }
145 
146 /* Compute noise shaping coefficients and initial gain values */
silk_noise_shape_analysis_FLP(silk_encoder_state_FLP * psEnc,silk_encoder_control_FLP * psEncCtrl,const silk_float * pitch_res,const silk_float * x)147 void silk_noise_shape_analysis_FLP(
148     silk_encoder_state_FLP          *psEnc,                             /* I/O  Encoder state FLP                           */
149     silk_encoder_control_FLP        *psEncCtrl,                         /* I/O  Encoder control FLP                         */
150     const silk_float                *pitch_res,                         /* I    LPC residual from pitch analysis            */
151     const silk_float                *x                                  /* I    Input signal [frame_length + la_shape]      */
152 )
153 {
154     silk_shape_state_FLP *psShapeSt = &psEnc->sShape;
155     opus_int     k, nSamples, nSegs;
156     silk_float   SNR_adj_dB, HarmShapeGain, Tilt;
157     silk_float   nrg, log_energy, log_energy_prev, energy_variation;
158     silk_float   BWExp, gain_mult, gain_add, strength, b, warping;
159     silk_float   x_windowed[ SHAPE_LPC_WIN_MAX ];
160     silk_float   auto_corr[ MAX_SHAPE_LPC_ORDER + 1 ];
161     silk_float   rc[ MAX_SHAPE_LPC_ORDER + 1 ];
162     const silk_float *x_ptr, *pitch_res_ptr;
163 
164     /* Point to start of first LPC analysis block */
165     x_ptr = x - psEnc->sCmn.la_shape;
166 
167     /****************/
168     /* GAIN CONTROL */
169     /****************/
170     SNR_adj_dB = psEnc->sCmn.SNR_dB_Q7 * ( 1 / 128.0f );
171 
172     /* Input quality is the average of the quality in the lowest two VAD bands */
173     psEncCtrl->input_quality = 0.5f * ( psEnc->sCmn.input_quality_bands_Q15[ 0 ] + psEnc->sCmn.input_quality_bands_Q15[ 1 ] ) * ( 1.0f / 32768.0f );
174 
175     /* Coding quality level, between 0.0 and 1.0 */
176     psEncCtrl->coding_quality = silk_sigmoid( 0.25f * ( SNR_adj_dB - 20.0f ) );
177 
178     if( psEnc->sCmn.useCBR == 0 ) {
179         /* Reduce coding SNR during low speech activity */
180         b = 1.0f - psEnc->sCmn.speech_activity_Q8 * ( 1.0f /  256.0f );
181         SNR_adj_dB -= BG_SNR_DECR_dB * psEncCtrl->coding_quality * ( 0.5f + 0.5f * psEncCtrl->input_quality ) * b * b;
182     }
183 
184     if( psEnc->sCmn.indices.signalType == TYPE_VOICED ) {
185         /* Reduce gains for periodic signals */
186         SNR_adj_dB += HARM_SNR_INCR_dB * psEnc->LTPCorr;
187     } else {
188         /* For unvoiced signals and low-quality input, adjust the quality slower than SNR_dB setting */
189         SNR_adj_dB += ( -0.4f * psEnc->sCmn.SNR_dB_Q7 * ( 1 / 128.0f ) + 6.0f ) * ( 1.0f - psEncCtrl->input_quality );
190     }
191 
192     /*************************/
193     /* SPARSENESS PROCESSING */
194     /*************************/
195     /* Set quantizer offset */
196     if( psEnc->sCmn.indices.signalType == TYPE_VOICED ) {
197         /* Initially set to 0; may be overruled in process_gains(..) */
198         psEnc->sCmn.indices.quantOffsetType = 0;
199     } else {
200         /* Sparseness measure, based on relative fluctuations of energy per 2 milliseconds */
201         nSamples = 2 * psEnc->sCmn.fs_kHz;
202         energy_variation = 0.0f;
203         log_energy_prev  = 0.0f;
204         pitch_res_ptr = pitch_res;
205         nSegs = silk_SMULBB( SUB_FRAME_LENGTH_MS, psEnc->sCmn.nb_subfr ) / 2;
206         for( k = 0; k < nSegs; k++ ) {
207             nrg = ( silk_float )nSamples + ( silk_float )silk_energy_FLP( pitch_res_ptr, nSamples );
208             log_energy = silk_log2( nrg );
209             if( k > 0 ) {
210                 energy_variation += silk_abs_float( log_energy - log_energy_prev );
211             }
212             log_energy_prev = log_energy;
213             pitch_res_ptr += nSamples;
214         }
215 
216         /* Set quantization offset depending on sparseness measure */
217         if( energy_variation > ENERGY_VARIATION_THRESHOLD_QNT_OFFSET * (nSegs-1) ) {
218             psEnc->sCmn.indices.quantOffsetType = 0;
219         } else {
220             psEnc->sCmn.indices.quantOffsetType = 1;
221         }
222     }
223 
224     /*******************************/
225     /* Control bandwidth expansion */
226     /*******************************/
227     /* More BWE for signals with high prediction gain */
228     strength = FIND_PITCH_WHITE_NOISE_FRACTION * psEncCtrl->predGain;           /* between 0.0 and 1.0 */
229     BWExp = BANDWIDTH_EXPANSION / ( 1.0f + strength * strength );
230 
231     /* Slightly more warping in analysis will move quantization noise up in frequency, where it's better masked */
232     warping = (silk_float)psEnc->sCmn.warping_Q16 / 65536.0f + 0.01f * psEncCtrl->coding_quality;
233 
234     /********************************************/
235     /* Compute noise shaping AR coefs and gains */
236     /********************************************/
237     for( k = 0; k < psEnc->sCmn.nb_subfr; k++ ) {
238         /* Apply window: sine slope followed by flat part followed by cosine slope */
239         opus_int shift, slope_part, flat_part;
240         flat_part = psEnc->sCmn.fs_kHz * 3;
241         slope_part = ( psEnc->sCmn.shapeWinLength - flat_part ) / 2;
242 
243         silk_apply_sine_window_FLP( x_windowed, x_ptr, 1, slope_part );
244         shift = slope_part;
245         silk_memcpy( x_windowed + shift, x_ptr + shift, flat_part * sizeof(silk_float) );
246         shift += flat_part;
247         silk_apply_sine_window_FLP( x_windowed + shift, x_ptr + shift, 2, slope_part );
248 
249         /* Update pointer: next LPC analysis block */
250         x_ptr += psEnc->sCmn.subfr_length;
251 
252         if( psEnc->sCmn.warping_Q16 > 0 ) {
253             /* Calculate warped auto correlation */
254             silk_warped_autocorrelation_FLP( auto_corr, x_windowed, warping,
255                 psEnc->sCmn.shapeWinLength, psEnc->sCmn.shapingLPCOrder );
256         } else {
257             /* Calculate regular auto correlation */
258             silk_autocorrelation_FLP( auto_corr, x_windowed, psEnc->sCmn.shapeWinLength, psEnc->sCmn.shapingLPCOrder + 1 );
259         }
260 
261         /* Add white noise, as a fraction of energy */
262         auto_corr[ 0 ] += auto_corr[ 0 ] * SHAPE_WHITE_NOISE_FRACTION + 1.0f;
263 
264         /* Convert correlations to prediction coefficients, and compute residual energy */
265         nrg = silk_schur_FLP( rc, auto_corr, psEnc->sCmn.shapingLPCOrder );
266         silk_k2a_FLP( &psEncCtrl->AR[ k * MAX_SHAPE_LPC_ORDER ], rc, psEnc->sCmn.shapingLPCOrder );
267         psEncCtrl->Gains[ k ] = ( silk_float )sqrt( nrg );
268 
269         if( psEnc->sCmn.warping_Q16 > 0 ) {
270             /* Adjust gain for warping */
271             psEncCtrl->Gains[ k ] *= warped_gain( &psEncCtrl->AR[ k * MAX_SHAPE_LPC_ORDER ], warping, psEnc->sCmn.shapingLPCOrder );
272         }
273 
274         /* Bandwidth expansion for synthesis filter shaping */
275         silk_bwexpander_FLP( &psEncCtrl->AR[ k * MAX_SHAPE_LPC_ORDER ], psEnc->sCmn.shapingLPCOrder, BWExp );
276 
277         if( psEnc->sCmn.warping_Q16 > 0 ) {
278             /* Convert to monic warped prediction coefficients and limit absolute values */
279             warped_true2monic_coefs( &psEncCtrl->AR[ k * MAX_SHAPE_LPC_ORDER ], warping, 3.999f, psEnc->sCmn.shapingLPCOrder );
280         } else {
281             /* Limit absolute values */
282             limit_coefs( &psEncCtrl->AR[ k * MAX_SHAPE_LPC_ORDER ], 3.999f, psEnc->sCmn.shapingLPCOrder );
283         }
284     }
285 
286     /*****************/
287     /* Gain tweaking */
288     /*****************/
289     /* Increase gains during low speech activity */
290     gain_mult = (silk_float)pow( 2.0f, -0.16f * SNR_adj_dB );
291     gain_add  = (silk_float)pow( 2.0f,  0.16f * MIN_QGAIN_DB );
292     for( k = 0; k < psEnc->sCmn.nb_subfr; k++ ) {
293         psEncCtrl->Gains[ k ] *= gain_mult;
294         psEncCtrl->Gains[ k ] += gain_add;
295     }
296 
297     /************************************************/
298     /* Control low-frequency shaping and noise tilt */
299     /************************************************/
300     /* Less low frequency shaping for noisy inputs */
301     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 ) );
302     strength *= psEnc->sCmn.speech_activity_Q8 * ( 1.0f /  256.0f );
303     if( psEnc->sCmn.indices.signalType == TYPE_VOICED ) {
304         /* Reduce low frequencies quantization noise for periodic signals, depending on pitch lag */
305         /*f = 400; freqz([1, -0.98 + 2e-4 * f], [1, -0.97 + 7e-4 * f], 2^12, Fs); axis([0, 1000, -10, 1])*/
306         for( k = 0; k < psEnc->sCmn.nb_subfr; k++ ) {
307             b = 0.2f / psEnc->sCmn.fs_kHz + 3.0f / psEncCtrl->pitchL[ k ];
308             psEncCtrl->LF_MA_shp[ k ] = -1.0f + b;
309             psEncCtrl->LF_AR_shp[ k ] =  1.0f - b - b * strength;
310         }
311         Tilt = - HP_NOISE_COEF -
312             (1 - HP_NOISE_COEF) * HARM_HP_NOISE_COEF * psEnc->sCmn.speech_activity_Q8 * ( 1.0f /  256.0f );
313     } else {
314         b = 1.3f / psEnc->sCmn.fs_kHz;
315         psEncCtrl->LF_MA_shp[ 0 ] = -1.0f + b;
316         psEncCtrl->LF_AR_shp[ 0 ] =  1.0f - b - b * strength * 0.6f;
317         for( k = 1; k < psEnc->sCmn.nb_subfr; k++ ) {
318             psEncCtrl->LF_MA_shp[ k ] = psEncCtrl->LF_MA_shp[ 0 ];
319             psEncCtrl->LF_AR_shp[ k ] = psEncCtrl->LF_AR_shp[ 0 ];
320         }
321         Tilt = -HP_NOISE_COEF;
322     }
323 
324     /****************************/
325     /* HARMONIC SHAPING CONTROL */
326     /****************************/
327     if( USE_HARM_SHAPING && psEnc->sCmn.indices.signalType == TYPE_VOICED ) {
328         /* Harmonic noise shaping */
329         HarmShapeGain = HARMONIC_SHAPING;
330 
331         /* More harmonic noise shaping for high bitrates or noisy input */
332         HarmShapeGain += HIGH_RATE_OR_LOW_QUALITY_HARMONIC_SHAPING *
333             ( 1.0f - ( 1.0f - psEncCtrl->coding_quality ) * psEncCtrl->input_quality );
334 
335         /* Less harmonic noise shaping for less periodic signals */
336         HarmShapeGain *= ( silk_float )sqrt( psEnc->LTPCorr );
337     } else {
338         HarmShapeGain = 0.0f;
339     }
340 
341     /*************************/
342     /* Smooth over subframes */
343     /*************************/
344     for( k = 0; k < psEnc->sCmn.nb_subfr; k++ ) {
345         psShapeSt->HarmShapeGain_smth += SUBFR_SMTH_COEF * ( HarmShapeGain - psShapeSt->HarmShapeGain_smth );
346         psEncCtrl->HarmShapeGain[ k ]  = psShapeSt->HarmShapeGain_smth;
347         psShapeSt->Tilt_smth          += SUBFR_SMTH_COEF * ( Tilt - psShapeSt->Tilt_smth );
348         psEncCtrl->Tilt[ k ]           = psShapeSt->Tilt_smth;
349     }
350 }
351