Mercurial > hg > camir-aes2014

annotate core/magnatagatune/MTTAudioFeatureSlaney08.m @ 0:e9a9cd732c1e tip

Find changesets by keywords (author, files, the commit message), revision number or hash, or revset expression.
first hg version after svn
author wolffd
date Tue, 10 Feb 2015 15:05:51 +0000
parents
children
rev   line source
wolffd@0 1 classdef MTTAudioFeatureSlaney08 < MTTAudioFeature & handle
wolffd@0 2 % ---
wolffd@0 3 % This Class contains
wolffd@0 4 % a basic summary of MTT features complementary to those in
wolffd@0 5 % MTTAudioFeatureBasicSm, features are extracted
wolffd@0 6 % as described in Slaney 08 - LEARNING A METRIC FOR MUSIC SIMILARITY
wolffd@0 7 %
wolffd@0 8 % The usual workflow for these features constist of three steps
wolffd@0 9 % 1. extract: extracts the basic single-file dependent features
wolffd@0 10 % 2. define_global_transform: calculates the global feature
wolffd@0 11 % transformation parameters
wolffd@0 12 % 3. finalise: applies the common transformations to a specific feature
wolffd@0 13 % ---
wolffd@0 14
wolffd@0 15 properties(Constant = true)
wolffd@0 16
wolffd@0 17 % svn hook
wolffd@0 18 my_revision = str2double(substr('$Rev$', 5, -1));
wolffd@0 19 end
wolffd@0 20
wolffd@0 21 properties
wolffd@0 22 % ---
wolffd@0 23 % Set default parameters
wolffd@0 24 % ---
wolffd@0 25 my_params = struct(...
wolffd@0 26 'norm_mttstats', 1, ... %
wolffd@0 27 'whiten_mttstats', 0, ... % NOTE: whitening as in slaney??
wolffd@0 28 'select_mttstats', 1 ...% TODO: way to select certain features
wolffd@0 29 );
wolffd@0 30 end
wolffd@0 31
wolffd@0 32 % ---
wolffd@0 33 % member functions
wolffd@0 34 % ---
wolffd@0 35 methods
wolffd@0 36
wolffd@0 37 % ---
wolffd@0 38 % constructor: pointer to feature in database
wolffd@0 39 % ---
wolffd@0 40 function feature = MTTAudioFeatureSlaney08(varargin)
wolffd@0 41
wolffd@0 42 feature = feature@MTTAudioFeature(varargin{:});
wolffd@0 43
wolffd@0 44 end
wolffd@0 45 % ---
wolffd@0 46 % extract feature data from raw audio features
wolffd@0 47 % ---
wolffd@0 48 function data = extract(feature, clip)
wolffd@0 49 % ---
wolffd@0 50 % get features. this includes possible
wolffd@0 51 % local normalisations
wolffd@0 52 % ---
wolffd@0 53
wolffd@0 54 global globalvars;
wolffd@0 55
wolffd@0 56 % ---
wolffd@0 57 % get casimir child clip if available
wolffd@0 58 % ---
wolffd@0 59 if isa(clip, 'CASIMIRClip')
wolffd@0 60 baseclip = clip.child_clip();
wolffd@0 61 else
wolffd@0 62 baseclip = clip;
wolffd@0 63 end
wolffd@0 64 if isa(baseclip, 'MTTClip')
wolffd@0 65 rawf = baseclip.audio_features_raw();
wolffd@0 66 elseif isa(baseclip, 'MSDClip')
wolffd@0 67 rawf = baseclip.features('MSDAudioFeatureRAW');
wolffd@0 68 end
wolffd@0 69
wolffd@0 70
wolffd@0 71 % ---
wolffd@0 72 % TODO: implement time_weighted version of the statistical
wolffd@0 73 % evaluations below
wolffd@0 74 % ---
wolffd@0 75
wolffd@0 76 % segmentDurationMean: mean segment duration (sec.).
wolffd@0 77 data.mttstats.segmentDurationMean = mean(rawf.data.segments_duration);
wolffd@0 78
wolffd@0 79 % segmentDurationVariance: variance of the segment duration
wolffd@0 80 data.mttstats.segmentDurationVariance = var(rawf.data.segments_duration);
wolffd@0 81
wolffd@0 82 % timeLoudnessMaxMean: mean time to the segment maximum, or attack duration (sec.).
wolffd@0 83 data.mttstats.timeLoudnessMaxMean = mean(rawf.data.segments_loudness_max_time);
wolffd@0 84
wolffd@0 85 % loudnessMaxMean: mean of segments’ maximum loudness(dB).
wolffd@0 86 data.mttstats.loudnessMaxMean = mean(rawf.data.segments_loudness_max);
wolffd@0 87
wolffd@0 88 % loudnessMaxVariance: variance of the segments’ maximum loudness (dB).
wolffd@0 89 data.mttstats.loudnessMaxVariance = var(rawf.data.segments_loudness_max);
wolffd@0 90
wolffd@0 91 % loudnessBeginMean: average loudness at the start of segments (dB)
wolffd@0 92 data.mttstats.loudnessBeginMean = mean(rawf.data.segments_loudness);
wolffd@0 93
wolffd@0 94 % loudnessBeginVariance: variance of the loudness at the startof segments (dB2). Correlated with loudnessMaxVariance
wolffd@0 95 data.mttstats.loudnessBeginVariance = var(rawf.data.segments_loudness);
wolffd@0 96
wolffd@0 97 % loudnessDynamicsMean: average of overall dynamic rangein the segments (dB).
wolffd@0 98 % loudnessDynamicsVariance: segment dynamic range variance
wolffd@0 99 % (dB). Higher variances suggest more dynamics ineach segment.
wolffd@0 100 % ---
wolffd@0 101 % NOTE: the above information cannot be extracted from the MTT
wolffd@0 102 % Features, maybe more recent echonest features allow for this
wolffd@0 103 % ---
wolffd@0 104
wolffd@0 105 % loudness: overall loudness estimate of the track (dB).
wolffd@0 106 data.mttstats.loudness = rawf.data.loudness;
wolffd@0 107
wolffd@0 108 % ---
wolffd@0 109 % TODO: get these from the beat loundesses?
wolffd@0 110 % ---
wolffd@0 111
wolffd@0 112 % tempo: overall track tempo estimate (in beat per minute,BPM). Doubling and halving errors are possible.
wolffd@0 113 data.mttstats.tempo = rawf.data.tempo;
wolffd@0 114
wolffd@0 115 % tempoConfidence: a measure of the con?dence of the tempo estimate (beween 0 and 1).
wolffd@0 116 %data.mttstats.tempoConfidence = rawf.data.tempoConfidence;
wolffd@0 117
wolffd@0 118 beats = rawf.data.beats;
wolffd@0 119 tatums = rawf.data.tatums;
wolffd@0 120
wolffd@0 121 % beatVariance: ameasure of the regularity of the beat (secs).
wolffd@0 122 if numel(beats) > 0
wolffd@0 123 bdiff = diff(beats(1,:));
wolffd@0 124 data.mttstats.beatVariance = var(bdiff);
wolffd@0 125 else
wolffd@0 126
wolffd@0 127 % ---
wolffd@0 128 % This is a facke repolacement variance
wolffd@0 129 % ---
wolffd@0 130 data.mttstats.beatVariance = 0;
wolffd@0 131 end
wolffd@0 132
wolffd@0 133
wolffd@0 134 % tatum: estimated overall tatum duration (in seconds). Tatums are subdivisions of the beat.
wolffd@0 135 % ---
wolffd@0 136 % note: the tatum length could be also
wolffd@0 137 % accessed by comparison with the global bpm estimate
wolffd@0 138 % ---
wolffd@0 139 if numel(tatums) > 0
wolffd@0 140 tdiff = diff(tatums(1,:));
wolffd@0 141 data.mttstats.tatum = median(tdiff);
wolffd@0 142
wolffd@0 143 % tatumConfidence: a measure of the con?dence of the tatum estimate (beween 0 and 1).
wolffd@0 144 data.mttstats.tatumConfidence = mean(tatums(2,:));
wolffd@0 145
wolffd@0 146 % numTatumsPerBeat: number of tatums per beat
wolffd@0 147 data.mttstats.numTatumsPerBeat = median(bdiff) / data.mttstats.tatum;
wolffd@0 148 else
wolffd@0 149 % ---
wolffd@0 150 % This is a facke replacement tatum
wolffd@0 151 % TODO: maybe set confidence to -1?
wolffd@0 152 % ---
wolffd@0 153
wolffd@0 154 data.mttstats.tatum = 0;
wolffd@0 155
wolffd@0 156 % tatumConfidence: a measure of the con?dence of the tatum estimate (beween 0 and 1).
wolffd@0 157
wolffd@0 158 data.mttstats.tatumConfidence = 0;
wolffd@0 159
wolffd@0 160 % numTatumsPerBeat: number of tatums per beat
wolffd@0 161 data.mttstats.numTatumsPerBeat = 2;
wolffd@0 162 end
wolffd@0 163
wolffd@0 164
wolffd@0 165 % ---
wolffd@0 166 % TODO: beat analysis
wolffd@0 167 % ---
wolffd@0 168
wolffd@0 169 % timeSignature: estimated time signature (number of beats per measure). (0-7 / 7)
wolffd@0 170 data.mttstats.timeSignature = rawf.data.timeSignature;
wolffd@0 171
wolffd@0 172 % timeSignatureStability: a rough estimate of the stability of the time signature throughout the track
wolffd@0 173 data.mttstats.timeSignatureStability = rawf.data.timeSignatureConfidence;
wolffd@0 174
wolffd@0 175 % ---
wolffd@0 176 % prepare field for final features
wolffd@0 177 % ---
wolffd@0 178 data.final.vector = [];
wolffd@0 179 data.final.vector_info = struct();
wolffd@0 180 data.final.dim = 0;
wolffd@0 181
wolffd@0 182 % save info data
wolffd@0 183 data.info.type = 'MTTAudioFeatureSlaney08';
wolffd@0 184 data.info.owner_id = clip.id;
wolffd@0 185 data.info.creatorrev = feature.my_revision;
wolffd@0 186
wolffd@0 187 % save parameters
wolffd@0 188 data.info.params = feature.my_params;
wolffd@0 189 end
wolffd@0 190
wolffd@0 191 function define_global_transform(features)
wolffd@0 192 % calculate and set normalization factors from the group of
wolffd@0 193 % input features. These features will be set for the full database
wolffd@0 194
wolffd@0 195 for i = 1:numel(features)
wolffd@0 196 data = features(i).data.mttstats;
wolffd@0 197
wolffd@0 198 final(:,i) = [data.segmentDurationMean; ...
wolffd@0 199 data.segmentDurationVariance; ...
wolffd@0 200 data.timeLoudnessMaxMean; ...
wolffd@0 201 data.loudnessMaxMean; ...
wolffd@0 202 data.loudnessMaxVariance; ...
wolffd@0 203 data.loudnessBeginMean; ...
wolffd@0 204 data.loudnessBeginVariance; ...
wolffd@0 205 data.loudness; ...
wolffd@0 206 data.tempo; ...
wolffd@0 207 ... % data.tempoConfidence; ...
wolffd@0 208 data.beatVariance; ...
wolffd@0 209 data.tatum; ...
wolffd@0 210 data.tatumConfidence; ...
wolffd@0 211 data.numTatumsPerBeat; ...
wolffd@0 212 data.timeSignature; ...
wolffd@0 213 data.timeSignatureStability];
wolffd@0 214 end
wolffd@0 215
wolffd@0 216 if features(1).my_params.norm_mttstats
wolffd@0 217 if numel(features) == 1
wolffd@0 218 error ('Insert feature array for this method, or set normalisation to 0');
wolffd@0 219 end
wolffd@0 220
wolffd@0 221 % ---
wolffd@0 222 % here, we only need to define the post-normalisation
wolffd@0 223 % ---
wolffd@0 224 [final, pstd] = mapminmax(final,0,1);
wolffd@0 225 common.mttstats.pre_norm = pstd;
wolffd@0 226
wolffd@0 227 % ---
wolffd@0 228 % NOTE: whitening as in slaney??
wolffd@0 229 % Would make reading the
wolffd@0 230 % mahal matrices really hard
wolffd@0 231 % ---
wolffd@0 232
wolffd@0 233 features(1).my_db.set_common(common);
wolffd@0 234
wolffd@0 235 else
wolffd@0 236
wolffd@0 237 features(1).my_db.set_common([1]);
wolffd@0 238 end
wolffd@0 239
wolffd@0 240 % save the normalised features straight away!
wolffd@0 241 features.finalise(final);
wolffd@0 242 end
wolffd@0 243
wolffd@0 244
wolffd@0 245 function finalise(features, final)
wolffd@0 246 % applies a final transformation and
wolffd@0 247 % collects the information of this feature within a single vector
wolffd@0 248 % see info for types in specific dimensions
wolffd@0 249 % check if features have been finalised already
wolffd@0 250
wolffd@0 251 % ---
wolffd@0 252 % check for dummy feature
wolffd@0 253 % ---
wolffd@0 254 if isfield(features(1).my_params,'select_mttstats') && ...
wolffd@0 255 isnumeric(features(1).my_params.select_mttstats) && ...
wolffd@0 256 features(1).my_params.select_mttstats == 0
wolffd@0 257
wolffd@0 258 % if no information needed just fill everything 0
wolffd@0 259 for i = 1:numel(features)
wolffd@0 260 features(i).data.final.vector = [];
wolffd@0 261 features(i).data.final.dim = 0;
wolffd@0 262
wolffd@0 263 % fill up info struct and append to feature
wolffd@0 264 features(i).data.final.vector_info.labels = {};
wolffd@0 265 end
wolffd@0 266
wolffd@0 267 return;
wolffd@0 268 end
wolffd@0 269
wolffd@0 270 % ---
wolffd@0 271 % set feature labelling
wolffd@0 272 % ---
wolffd@0 273 info = {'segmentDurationMean', ...
wolffd@0 274 'segmentDurationVariance', ...
wolffd@0 275 'timeLoudnessMaxMean', ...
wolffd@0 276 'loudnessMaxMean', ...
wolffd@0 277 'loudnessMaxVariance', ...
wolffd@0 278 'loudnessBeginMean', ...
wolffd@0 279 'loudnessBeginVariance', ...
wolffd@0 280 'loudness', ...
wolffd@0 281 'tempo', ...
wolffd@0 282 ...% 'tempoConfidence', ...
wolffd@0 283 'beatVariance', ...
wolffd@0 284 'tatum', ...
wolffd@0 285 'tatumConfidence', ...
wolffd@0 286 'numTatumsPerBeat', ...
wolffd@0 287 'timeSignature', ...
wolffd@0 288 'timeSignatureStability'};
wolffd@0 289
wolffd@0 290 % ---
wolffd@0 291 % construct resulting feature vector out of features
wolffd@0 292 % ---
wolffd@0 293 if nargin == 2 && isempty(final)
wolffd@0 294
wolffd@0 295 % the final vector etc already are set to zero;
wolffd@0 296 return;
wolffd@0 297
wolffd@0 298 elseif nargin == 2 && (numel(features) == size(final, 2))
wolffd@0 299 for i = 1:numel(features)
wolffd@0 300
wolffd@0 301 % check for neccesary parameters
wolffd@0 302 if isempty(features(i).my_db.commondb)
wolffd@0 303
wolffd@0 304 error('Define the global transformation first')
wolffd@0 305 return;
wolffd@0 306 end
wolffd@0 307
wolffd@0 308 features(i).data.final.vector = final(:,i);
wolffd@0 309 features(i).data.final.dim = size(final,1);
wolffd@0 310
wolffd@0 311 % fill up info struct and append to feature
wolffd@0 312 features(i).data.final.vector_info.labels = info;
wolffd@0 313 end
wolffd@0 314 else
wolffd@0 315 % ---
wolffd@0 316 % if features have been added after gettin gnormalisation
wolffd@0 317 % parameters, ther should be still an option to include
wolffd@0 318 % them
wolffd@0 319 % ---
wolffd@0 320
wolffd@0 321 for i = 1:numel(features)
wolffd@0 322
wolffd@0 323 % check for neccesary parameters
wolffd@0 324 if isempty(features(i).my_db.commondb)
wolffd@0 325
wolffd@0 326 error('Define the global transformation first')
wolffd@0 327 return;
wolffd@0 328 end
wolffd@0 329
wolffd@0 330 data = features(i).data.mttstats;
wolffd@0 331 final = [data.segmentDurationMean; ...
wolffd@0 332 data.segmentDurationVariance; ...
wolffd@0 333 data.timeLoudnessMaxMean; ...
wolffd@0 334 data.loudnessMaxMean; ...
wolffd@0 335 data.loudnessMaxVariance; ...
wolffd@0 336 data.loudnessBeginMean; ...
wolffd@0 337 data.loudnessBeginVariance; ...
wolffd@0 338 data.loudness; ...
wolffd@0 339 data.tempo; ...
wolffd@0 340 ... % data.tempoConfidence; ...
wolffd@0 341 data.beatVariance; ...
wolffd@0 342 data.tatum; ...
wolffd@0 343 data.tatumConfidence; ...
wolffd@0 344 data.numTatumsPerBeat; ...
wolffd@0 345 data.timeSignature; ...
wolffd@0 346 data.timeSignatureStability];
wolffd@0 347
wolffd@0 348 if features(1).my_params.norm_mttstats == 1
wolffd@0 349
wolffd@0 350 [final] = mapminmax('apply', final, features(1).common.mttstats.pre_norm);
wolffd@0 351 end
wolffd@0 352
wolffd@0 353 features(i).data.final.vector = final;
wolffd@0 354 features(i).data.final.dim = size(final,1);
wolffd@0 355
wolffd@0 356 % fill up info struct and append to feature
wolffd@0 357 features(i).data.final.vector_info.labels = info;
wolffd@0 358 end
wolffd@0 359
wolffd@0 360 end
wolffd@0 361
wolffd@0 362 % ---
wolffd@0 363 % TODO: Maybe delete more basic features again at this point?
wolffd@0 364 % ---
wolffd@0 365 end
wolffd@0 366
wolffd@0 367 % ---
wolffd@0 368 % destructor: do we really want to remove this
wolffd@0 369 % from the database? No, but
wolffd@0 370 % TODO: create marker for unused objects in db, and a cleanup
wolffd@0 371 % function
wolffd@0 372 % ---
wolffd@0 373 function delete(feature)
wolffd@0 374
wolffd@0 375 end
wolffd@0 376 end
wolffd@0 377 end