1/* 2 * ReplayGainAnalysis - analyzes input samples and give the recommended dB change 3 * Copyright (C) 2001 David Robinson and Glen Sawyer 4 * 5 * This library is free software; you can redistribute it and/or 6 * modify it under the terms of the GNU Lesser General Public 7 * License as published by the Free Software Foundation; either 8 * version 2.1 of the License, or (at your option) any later version. 9 * 10 * This library is distributed in the hope that it will be useful, 11 * but WITHOUT ANY WARRANTY; without even the implied warranty of 12 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU 13 * Lesser General Public License for more details. 14 * 15 * You should have received a copy of the GNU Lesser General Public 16 * License along with this library; if not, write to the Free Software 17 * Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA 02111-1307 USA 18 * 19 * concept and filter values by David Robinson (David@Robinson.org) 20 * -- blame him if you think the idea is flawed 21 * original coding by Glen Sawyer (glensawyer@hotmail.com) 22 * -- blame him if you think this runs too slowly, or the coding is otherwise flawed 23 * 24 * lots of code improvements by Frank Klemm ( http://www.uni-jena.de/~pfk/mpp/ ) 25 * -- credit him for all the _good_ programming ;) 26 * 27 * minor cosmetic tweaks to integrate with FLAC by Josh Coalson 28 * 29 * 30 * For an explanation of the concepts and the basic algorithms involved, go to: 31 * http://www.replaygain.org/ 32 */ 33 34/* 35 * Here's the deal. Call 36 * 37 * InitGainAnalysis ( long samplefreq ); 38 * 39 * to initialize everything. Call 40 * 41 * AnalyzeSamples ( const Float_t* left_samples, 42 * const Float_t* right_samples, 43 * size_t num_samples, 44 * int num_channels ); 45 * 46 * as many times as you want, with as many or as few samples as you want. 47 * If mono, pass the sample buffer in through left_samples, leave 48 * right_samples NULL, and make sure num_channels = 1. 49 * 50 * GetTitleGain() 51 * 52 * will return the recommended dB level change for all samples analyzed 53 * SINCE THE LAST TIME you called GetTitleGain() OR InitGainAnalysis(). 54 * 55 * GetAlbumGain() 56 * 57 * will return the recommended dB level change for all samples analyzed 58 * since InitGainAnalysis() was called and finalized with GetTitleGain(). 59 * 60 * Pseudo-code to process an album: 61 * 62 * Float_t l_samples [4096]; 63 * Float_t r_samples [4096]; 64 * size_t num_samples; 65 * unsigned int num_songs; 66 * unsigned int i; 67 * 68 * InitGainAnalysis ( 44100 ); 69 * for ( i = 1; i <= num_songs; i++ ) { 70 * while ( ( num_samples = getSongSamples ( song[i], left_samples, right_samples ) ) > 0 ) 71 * AnalyzeSamples ( left_samples, right_samples, num_samples, 2 ); 72 * fprintf ("Recommended dB change for song %2d: %+6.2f dB\n", i, GetTitleGain() ); 73 * } 74 * fprintf ("Recommended dB change for whole album: %+6.2f dB\n", GetAlbumGain() ); 75 */ 76 77/* 78 * So here's the main source of potential code confusion: 79 * 80 * The filters applied to the incoming samples are IIR filters, 81 * meaning they rely on up to <filter order> number of previous samples 82 * AND up to <filter order> number of previous filtered samples. 83 * 84 * I set up the AnalyzeSamples routine to minimize memory usage and interface 85 * complexity. The speed isn't compromised too much (I don't think), but the 86 * internal complexity is higher than it should be for such a relatively 87 * simple routine. 88 * 89 * Optimization/clarity suggestions are welcome. 90 */ 91 92#if HAVE_CONFIG_H 93# include <config.h> 94#endif 95 96#include <stdio.h> 97#include <stdlib.h> 98#include <string.h> 99#include <math.h> 100 101#include "replaygain_analysis.h" 102 103Float_t ReplayGainReferenceLoudness = 89.0; /* in dB SPL */ 104 105typedef unsigned short Uint16_t; 106typedef signed short Int16_t; 107typedef unsigned int Uint32_t; 108typedef signed int Int32_t; 109 110#define YULE_ORDER 10 111#define BUTTER_ORDER 2 112#define RMS_PERCENTILE 0.95 /* percentile which is louder than the proposed level */ 113#define MAX_SAMP_FREQ 48000. /* maximum allowed sample frequency [Hz] */ 114#define RMS_WINDOW_TIME 0.050 /* Time slice size [s] */ 115#define STEPS_per_dB 100. /* Table entries per dB */ 116#define MAX_dB 120. /* Table entries for 0...MAX_dB (normal max. values are 70...80 dB) */ 117 118#define MAX_ORDER (BUTTER_ORDER > YULE_ORDER ? BUTTER_ORDER : YULE_ORDER) 119/* [JEC] the following was originally #defined as: 120 * (size_t) (MAX_SAMP_FREQ * RMS_WINDOW_TIME) 121 * but that seemed to fail to take into account the ceil() part of the 122 * sampleWindow calculation in ResetSampleFrequency(), and was causing 123 * buffer overflows for 48kHz analysis, hence the +1. 124 */ 125#ifndef __sun 126 #define MAX_SAMPLES_PER_WINDOW (size_t) (MAX_SAMP_FREQ * RMS_WINDOW_TIME + 1.) /* max. Samples per Time slice */ 127#else 128 /* [JEC] Solaris Forte compiler doesn't like float calc in array indices */ 129 #define MAX_SAMPLES_PER_WINDOW (size_t) (2401) 130#endif 131#define PINK_REF 64.82 /* 298640883795 */ /* calibration value */ 132 133static Float_t linprebuf [MAX_ORDER * 2]; 134static Float_t* linpre; /* left input samples, with pre-buffer */ 135static Float_t lstepbuf [MAX_SAMPLES_PER_WINDOW + MAX_ORDER]; 136static Float_t* lstep; /* left "first step" (i.e. post first filter) samples */ 137static Float_t loutbuf [MAX_SAMPLES_PER_WINDOW + MAX_ORDER]; 138static Float_t* lout; /* left "out" (i.e. post second filter) samples */ 139static Float_t rinprebuf [MAX_ORDER * 2]; 140static Float_t* rinpre; /* right input samples ... */ 141static Float_t rstepbuf [MAX_SAMPLES_PER_WINDOW + MAX_ORDER]; 142static Float_t* rstep; 143static Float_t routbuf [MAX_SAMPLES_PER_WINDOW + MAX_ORDER]; 144static Float_t* rout; 145static unsigned int sampleWindow; /* number of samples required to reach number of milliseconds required for RMS window */ 146static unsigned long totsamp; 147static double lsum; 148static double rsum; 149static int freqindex; 150#ifndef __sun 151static Uint32_t A [(size_t)(STEPS_per_dB * MAX_dB)]; 152static Uint32_t B [(size_t)(STEPS_per_dB * MAX_dB)]; 153#else 154/* [JEC] Solaris Forte compiler doesn't like float calc in array indices */ 155static Uint32_t A [12000]; 156static Uint32_t B [12000]; 157#endif 158 159/* for each filter: 160 [0] 48 kHz, [1] 44.1 kHz, [2] 32 kHz, [3] 24 kHz, [4] 22050 Hz, [5] 16 kHz, [6] 12 kHz, [7] is 11025 Hz, [8] 8 kHz */ 161 162#ifdef WIN32 163#pragma warning ( disable : 4305 ) 164#endif 165 166static const Float_t AYule [9] [11] = { 167 { 1., -3.84664617118067, 7.81501653005538,-11.34170355132042, 13.05504219327545,-12.28759895145294, 9.48293806319790, -5.87257861775999, 2.75465861874613, -0.86984376593551, 0.13919314567432 }, 168 { 1., -3.47845948550071, 6.36317777566148, -8.54751527471874, 9.47693607801280, -8.81498681370155, 6.85401540936998, -4.39470996079559, 2.19611684890774, -0.75104302451432, 0.13149317958808 }, 169 { 1., -2.37898834973084, 2.84868151156327, -2.64577170229825, 2.23697657451713, -1.67148153367602, 1.00595954808547, -0.45953458054983, 0.16378164858596, -0.05032077717131, 0.02347897407020 }, 170 { 1., -1.61273165137247, 1.07977492259970, -0.25656257754070, -0.16276719120440, -0.22638893773906, 0.39120800788284, -0.22138138954925, 0.04500235387352, 0.02005851806501, 0.00302439095741 }, 171 { 1., -1.49858979367799, 0.87350271418188, 0.12205022308084, -0.80774944671438, 0.47854794562326, -0.12453458140019, -0.04067510197014, 0.08333755284107, -0.04237348025746, 0.02977207319925 }, 172 { 1., -0.62820619233671, 0.29661783706366, -0.37256372942400, 0.00213767857124, -0.42029820170918, 0.22199650564824, 0.00613424350682, 0.06747620744683, 0.05784820375801, 0.03222754072173 }, 173 { 1., -1.04800335126349, 0.29156311971249, -0.26806001042947, 0.00819999645858, 0.45054734505008, -0.33032403314006, 0.06739368333110, -0.04784254229033, 0.01639907836189, 0.01807364323573 }, 174 { 1., -0.51035327095184, -0.31863563325245, -0.20256413484477, 0.14728154134330, 0.38952639978999, -0.23313271880868, -0.05246019024463, -0.02505961724053, 0.02442357316099, 0.01818801111503 }, 175 { 1., -0.25049871956020, -0.43193942311114, -0.03424681017675, -0.04678328784242, 0.26408300200955, 0.15113130533216, -0.17556493366449, -0.18823009262115, 0.05477720428674, 0.04704409688120 } 176}; 177 178static const Float_t BYule [9] [11] = { 179 { 0.03857599435200, -0.02160367184185, -0.00123395316851, -0.00009291677959, -0.01655260341619, 0.02161526843274, -0.02074045215285, 0.00594298065125, 0.00306428023191, 0.00012025322027, 0.00288463683916 }, 180 { 0.05418656406430, -0.02911007808948, -0.00848709379851, -0.00851165645469, -0.00834990904936, 0.02245293253339, -0.02596338512915, 0.01624864962975, -0.00240879051584, 0.00674613682247, -0.00187763777362 }, 181 { 0.15457299681924, -0.09331049056315, -0.06247880153653, 0.02163541888798, -0.05588393329856, 0.04781476674921, 0.00222312597743, 0.03174092540049, -0.01390589421898, 0.00651420667831, -0.00881362733839 }, 182 { 0.30296907319327, -0.22613988682123, -0.08587323730772, 0.03282930172664, -0.00915702933434, -0.02364141202522, -0.00584456039913, 0.06276101321749, -0.00000828086748, 0.00205861885564, -0.02950134983287 }, 183 { 0.33642304856132, -0.25572241425570, -0.11828570177555, 0.11921148675203, -0.07834489609479, -0.00469977914380, -0.00589500224440, 0.05724228140351, 0.00832043980773, -0.01635381384540, -0.01760176568150 }, 184 { 0.44915256608450, -0.14351757464547, -0.22784394429749, -0.01419140100551, 0.04078262797139, -0.12398163381748, 0.04097565135648, 0.10478503600251, -0.01863887810927, -0.03193428438915, 0.00541907748707 }, 185 { 0.56619470757641, -0.75464456939302, 0.16242137742230, 0.16744243493672, -0.18901604199609, 0.30931782841830, -0.27562961986224, 0.00647310677246, 0.08647503780351, -0.03788984554840, -0.00588215443421 }, 186 { 0.58100494960553, -0.53174909058578, -0.14289799034253, 0.17520704835522, 0.02377945217615, 0.15558449135573, -0.25344790059353, 0.01628462406333, 0.06920467763959, -0.03721611395801, -0.00749618797172 }, 187 { 0.53648789255105, -0.42163034350696, -0.00275953611929, 0.04267842219415, -0.10214864179676, 0.14590772289388, -0.02459864859345, -0.11202315195388, -0.04060034127000, 0.04788665548180, -0.02217936801134 } 188}; 189 190static const Float_t AButter [9] [3] = { 191 { 1., -1.97223372919527, 0.97261396931306 }, 192 { 1., -1.96977855582618, 0.97022847566350 }, 193 { 1., -1.95835380975398, 0.95920349965459 }, 194 { 1., -1.95002759149878, 0.95124613669835 }, 195 { 1., -1.94561023566527, 0.94705070426118 }, 196 { 1., -1.92783286977036, 0.93034775234268 }, 197 { 1., -1.91858953033784, 0.92177618768381 }, 198 { 1., -1.91542108074780, 0.91885558323625 }, 199 { 1., -1.88903307939452, 0.89487434461664 } 200}; 201 202static const Float_t BButter [9] [3] = { 203 { 0.98621192462708, -1.97242384925416, 0.98621192462708 }, 204 { 0.98500175787242, -1.97000351574484, 0.98500175787242 }, 205 { 0.97938932735214, -1.95877865470428, 0.97938932735214 }, 206 { 0.97531843204928, -1.95063686409857, 0.97531843204928 }, 207 { 0.97316523498161, -1.94633046996323, 0.97316523498161 }, 208 { 0.96454515552826, -1.92909031105652, 0.96454515552826 }, 209 { 0.96009142950541, -1.92018285901082, 0.96009142950541 }, 210 { 0.95856916599601, -1.91713833199203, 0.95856916599601 }, 211 { 0.94597685600279, -1.89195371200558, 0.94597685600279 } 212}; 213 214#ifdef WIN32 215#pragma warning ( default : 4305 ) 216#endif 217 218/* When calling this procedure, make sure that ip[-order] and op[-order] point to real data! */ 219 220static void 221filter ( const Float_t* input, Float_t* output, size_t nSamples, const Float_t* a, const Float_t* b, size_t order ) 222{ 223 double y; 224 size_t i; 225 size_t k; 226 227 for ( i = 0; i < nSamples; i++ ) { 228 y = input[i] * b[0]; 229 for ( k = 1; k <= order; k++ ) 230 y += input[i-k] * b[k] - output[i-k] * a[k]; 231 output[i] = (Float_t)y; 232 } 233} 234 235/* returns a INIT_GAIN_ANALYSIS_OK if successful, INIT_GAIN_ANALYSIS_ERROR if not */ 236 237int 238ResetSampleFrequency ( long samplefreq ) { 239 int i; 240 241 /* zero out initial values */ 242 for ( i = 0; i < MAX_ORDER; i++ ) 243 linprebuf[i] = lstepbuf[i] = loutbuf[i] = rinprebuf[i] = rstepbuf[i] = routbuf[i] = 0.; 244 245 switch ( (int)(samplefreq) ) { 246 case 48000: freqindex = 0; break; 247 case 44100: freqindex = 1; break; 248 case 32000: freqindex = 2; break; 249 case 24000: freqindex = 3; break; 250 case 22050: freqindex = 4; break; 251 case 16000: freqindex = 5; break; 252 case 12000: freqindex = 6; break; 253 case 11025: freqindex = 7; break; 254 case 8000: freqindex = 8; break; 255 default: return INIT_GAIN_ANALYSIS_ERROR; 256 } 257 258 sampleWindow = (int) ceil (samplefreq * RMS_WINDOW_TIME); 259 260 lsum = 0.; 261 rsum = 0.; 262 totsamp = 0; 263 264 memset ( A, 0, sizeof(A) ); 265 266 return INIT_GAIN_ANALYSIS_OK; 267} 268 269int 270InitGainAnalysis ( long samplefreq ) 271{ 272 if (ResetSampleFrequency(samplefreq) != INIT_GAIN_ANALYSIS_OK) { 273 return INIT_GAIN_ANALYSIS_ERROR; 274 } 275 276 linpre = linprebuf + MAX_ORDER; 277 rinpre = rinprebuf + MAX_ORDER; 278 lstep = lstepbuf + MAX_ORDER; 279 rstep = rstepbuf + MAX_ORDER; 280 lout = loutbuf + MAX_ORDER; 281 rout = routbuf + MAX_ORDER; 282 283 memset ( B, 0, sizeof(B) ); 284 285 return INIT_GAIN_ANALYSIS_OK; 286} 287 288/* returns GAIN_ANALYSIS_OK if successful, GAIN_ANALYSIS_ERROR if not */ 289 290int 291AnalyzeSamples ( const Float_t* left_samples, const Float_t* right_samples, size_t num_samples, int num_channels ) 292{ 293 const Float_t* curleft; 294 const Float_t* curright; 295 long batchsamples; 296 long cursamples; 297 long cursamplepos; 298 int i; 299 300 if ( num_samples == 0 ) 301 return GAIN_ANALYSIS_OK; 302 303 cursamplepos = 0; 304 batchsamples = num_samples; 305 306 switch ( num_channels) { 307 case 1: right_samples = left_samples; 308 case 2: break; 309 default: return GAIN_ANALYSIS_ERROR; 310 } 311 312 if ( num_samples < MAX_ORDER ) { 313 memcpy ( linprebuf + MAX_ORDER, left_samples , num_samples * sizeof(Float_t) ); 314 memcpy ( rinprebuf + MAX_ORDER, right_samples, num_samples * sizeof(Float_t) ); 315 } 316 else { 317 memcpy ( linprebuf + MAX_ORDER, left_samples, MAX_ORDER * sizeof(Float_t) ); 318 memcpy ( rinprebuf + MAX_ORDER, right_samples, MAX_ORDER * sizeof(Float_t) ); 319 } 320 321 while ( batchsamples > 0 ) { 322 cursamples = batchsamples > (long)(sampleWindow-totsamp) ? (long)(sampleWindow - totsamp) : batchsamples; 323 if ( cursamplepos < MAX_ORDER ) { 324 curleft = linpre+cursamplepos; 325 curright = rinpre+cursamplepos; 326 if (cursamples > MAX_ORDER - cursamplepos ) 327 cursamples = MAX_ORDER - cursamplepos; 328 } 329 else { 330 curleft = left_samples + cursamplepos; 331 curright = right_samples + cursamplepos; 332 } 333 334 filter ( curleft , lstep + totsamp, cursamples, AYule[freqindex], BYule[freqindex], YULE_ORDER ); 335 filter ( curright, rstep + totsamp, cursamples, AYule[freqindex], BYule[freqindex], YULE_ORDER ); 336 337 filter ( lstep + totsamp, lout + totsamp, cursamples, AButter[freqindex], BButter[freqindex], BUTTER_ORDER ); 338 filter ( rstep + totsamp, rout + totsamp, cursamples, AButter[freqindex], BButter[freqindex], BUTTER_ORDER ); 339 340 for ( i = 0; i < cursamples; i++ ) { /* Get the squared values */ 341 lsum += lout [totsamp+i] * lout [totsamp+i]; 342 rsum += rout [totsamp+i] * rout [totsamp+i]; 343 } 344 345 batchsamples -= cursamples; 346 cursamplepos += cursamples; 347 totsamp += cursamples; 348 if ( totsamp == sampleWindow ) { /* Get the Root Mean Square (RMS) for this set of samples */ 349 double val = STEPS_per_dB * 10. * log10 ( (lsum+rsum) / totsamp * 0.5 + 1.e-37 ); 350 int ival = (int) val; 351 if ( ival < 0 ) ival = 0; 352 if ( ival >= (int)(sizeof(A)/sizeof(*A)) ) ival = (int)(sizeof(A)/sizeof(*A)) - 1; 353 A [ival]++; 354 lsum = rsum = 0.; 355 memmove ( loutbuf , loutbuf + totsamp, MAX_ORDER * sizeof(Float_t) ); 356 memmove ( routbuf , routbuf + totsamp, MAX_ORDER * sizeof(Float_t) ); 357 memmove ( lstepbuf, lstepbuf + totsamp, MAX_ORDER * sizeof(Float_t) ); 358 memmove ( rstepbuf, rstepbuf + totsamp, MAX_ORDER * sizeof(Float_t) ); 359 totsamp = 0; 360 } 361 if ( totsamp > sampleWindow ) /* somehow I really screwed up: Error in programming! Contact author about totsamp > sampleWindow */ 362 return GAIN_ANALYSIS_ERROR; 363 } 364 if ( num_samples < MAX_ORDER ) { 365 memmove ( linprebuf, linprebuf + num_samples, (MAX_ORDER-num_samples) * sizeof(Float_t) ); 366 memmove ( rinprebuf, rinprebuf + num_samples, (MAX_ORDER-num_samples) * sizeof(Float_t) ); 367 memcpy ( linprebuf + MAX_ORDER - num_samples, left_samples, num_samples * sizeof(Float_t) ); 368 memcpy ( rinprebuf + MAX_ORDER - num_samples, right_samples, num_samples * sizeof(Float_t) ); 369 } 370 else { 371 memcpy ( linprebuf, left_samples + num_samples - MAX_ORDER, MAX_ORDER * sizeof(Float_t) ); 372 memcpy ( rinprebuf, right_samples + num_samples - MAX_ORDER, MAX_ORDER * sizeof(Float_t) ); 373 } 374 375 return GAIN_ANALYSIS_OK; 376} 377 378 379static Float_t 380analyzeResult ( Uint32_t* Array, size_t len ) 381{ 382 Uint32_t elems; 383 Int32_t upper; 384 size_t i; 385 386 elems = 0; 387 for ( i = 0; i < len; i++ ) 388 elems += Array[i]; 389 if ( elems == 0 ) 390 return GAIN_NOT_ENOUGH_SAMPLES; 391 392 upper = (Int32_t) ceil (elems * (1. - RMS_PERCENTILE)); 393 for ( i = len; i-- > 0; ) { 394 if ( (upper -= Array[i]) <= 0 ) 395 break; 396 } 397 398 return (Float_t) ((Float_t)PINK_REF - (Float_t)i / (Float_t)STEPS_per_dB); 399} 400 401 402Float_t 403GetTitleGain ( void ) 404{ 405 Float_t retval; 406 unsigned int i; 407 408 retval = analyzeResult ( A, sizeof(A)/sizeof(*A) ); 409 410 for ( i = 0; i < sizeof(A)/sizeof(*A); i++ ) { 411 B[i] += A[i]; 412 A[i] = 0; 413 } 414 415 for ( i = 0; i < MAX_ORDER; i++ ) 416 linprebuf[i] = lstepbuf[i] = loutbuf[i] = rinprebuf[i] = rstepbuf[i] = routbuf[i] = 0.f; 417 418 totsamp = 0; 419 lsum = rsum = 0.; 420 return retval; 421} 422 423 424Float_t 425GetAlbumGain ( void ) 426{ 427 return analyzeResult ( B, sizeof(B)/sizeof(*B) ); 428} 429 430/* end of replaygain_analysis.c */ 431