Mercurial > hg > camir-aes2014
diff core/magnatagatune/MTTAudioFeatureHMM.m @ 0:e9a9cd732c1e tip
first hg version after svn
author | wolffd |
---|---|
date | Tue, 10 Feb 2015 15:05:51 +0000 |
parents | |
children |
line wrap: on
line diff
--- /dev/null Thu Jan 01 00:00:00 1970 +0000 +++ b/core/magnatagatune/MTTAudioFeatureHMM.m Tue Feb 10 15:05:51 2015 +0000 @@ -0,0 +1,452 @@ +classdef MTTAudioFeatureHMM < MTTAudioFeature & handle + % --- + % the MTTAudioFeatureBasicSm Class contains + % a basic summary of chroma, mfcc and tempo features + % a few common chroma and mfcc vectors are concatenated + % along with some clip-wide variance + % a metric / rhythm fingerprint is added + % + % The usual workflow for these features consists of three steps + % 1. extract: extracts the basic single-file dependent features + % 2. define_global_transform: calculates the global feature + % transformation parameters + % 3. finalise: applies the common transformations to a specific feature + % --- + + properties(Constant = true) + + % svn hook + my_revision = str2double(substr('$Rev: 2332 $', 5, -1)); + end + + properties + % --- + % Set default parameters + % --- + my_params = struct(... + 'nstates', 4 ... % predefined number of states + ); + end + + % --- + % member functions + % --- + methods + + % --- + % constructor: pointer to feature in database + % --- + function feature = MTTAudioFeatureHMM(varargin) + + feature = feature@MTTAudioFeature(varargin{:}); + + end + % --- + % extract feature data from raw audio features + % --- + function data = extract(feature, clip) + % --- + % get Basic Summary audio features. this includes possible + % local normalisations + % --- + + global globalvars; + + % --- + % get casimir child clip if available + % --- + if isa(clip, 'CASIMIRClip') + baseclip = clip.child_clip(); + else + baseclip = clip; + end + if isa(baseclip, 'MTTClip') + rawf = baseclip.audio_features_raw(); + elseif isa(baseclip, 'MSDClip') + rawf = baseclip.features('MSDAudioFeatureRAW'); + end + + % --- + % now extract the features + % first step: chroma clustering + % --- + weights = [rawf.data.segments_duration]; + + % normalise weights + weights = weights / rawf.data.duration; + + % get the chroma features + chroma = [rawf.data.segments_pitches]'; + + % --- + % TODO: train hmm + % --- + + + % save hmm into data variable + data.mu = mu1 + data.transmat1 = mu1 + + + + + + + % prepare field for final features + data.final.vector = []; + data.final.vector_info = struct(); + data.final.dim = 0; + + % save info data + data.info.type = 'MTTAudioFeatureBasicSm'; + data.info.owner = clip; + data.info.owner_id = clip.id; + data.info.creatorrev = feature.my_revision; + + % save parameters + data.info.params = feature.my_params; + end + + function define_global_transform(features) + % calculate and set normalization factors from the group of + % input features. These features will be set for the full database + + + + end + + + function finalise(feature) + % applies a final transformation and + % collects the information of this feature within a single vector + % see info for types in specific dimensions + + for i = 1:numel(feature) + + % check for neccesary parameters + if isempty(feature(i).my_db.commondb) + + error('Define the global transformation first') + return; + end + + if feature(1).my_params.ntimbres > 0 + % --- + % normalise features + % --- + % norm timbre features if neccesary + timbren = []; + if feature(i).my_params.norm_timbres + for j = 1:numel(feature(i).data.timbre) + + timbren = cat(1, timbren, ... + MTTAudioFeatureBasicSm.norm_timbre... + (feature(i).data.timbre(j).means, feature(i).my_db.commondb.post_normf.timbre)); + end + else + + timbren = cat(1, timbren, feature(i).data.timbre(:).means); + end + end + + % --- + % construct resulting feature vector out of features + % --- + vec = []; + info = {}; + if feature(i).my_params.nchromas > 0 + + info{numel(vec)+ 1} = 'chroma'; + vec = cat(1, vec, feature(i).data.chroma(:).means); + + info{numel(vec)+ 1} = 'chroma weights'; + vec = cat(1, vec, [feature(i).data.chroma(:).means_weight]'); + + % --- + % NORMALISE Chroma variance + % --- + if feature(i).my_params.chroma_var >= 1 + + info{numel(vec)+ 1} = 'chroma variance'; + + % normalise this pack of variance vectors + tmp_var = mapminmax('apply', [feature(i).data.chroma(:).vars],... + feature(i).common.post_normf.chroma_var); + + % concatenate normalised data to vector + for vari = 1:size(tmp_var,2) + + vec = cat(1, vec, tmp_var(:, vari)); + end + end + end + + + if feature(i).my_params.ntimbres > 0 + + info{numel(vec)+ 1} = 'timbre'; + vec = cat(1, vec, timbren); + + info{numel(vec)+ 1} = 'timbre weights'; + vec = cat(1, vec, [feature(i).data.timbre(:).means_weight]'); + + % --- + % NORMALISE timbre variance + % --- + if feature(i).my_params.timbre_var >= 1 + + info{numel(vec)+ 1} = 'timbre variance'; + + % normalise this pack of variance vectors + tmp_var = mapminmax('apply', [feature(i).data.timbre(:).vars],... + feature(i).common.post_normf.timbre_var); + + % concatenate normalised data to vector + for vari = 1:size(tmp_var,2) + + vec = cat(1, vec, tmp_var(:, vari)); + end + end + end + + if feature(i).my_params.nrhythms > 0 + + info{numel(vec)+ 1} = 'rhythm 8'; + vec = cat(1, vec, feature(i).data.rhythm.acorr8); + + info{numel(vec)+ 1} = 'int 8'; + vec = cat(1, vec, feature(i).data.rhythm.interval8); + + if feature(i).my_params.nrhythms >= 2 + + info{numel(vec)+ 1} = 'rhythm 16'; + vec = cat(1, vec, feature(i).data.rhythm.acorr16); + + info{numel(vec)+ 1} = 'int 16'; + vec = cat(1, vec, feature(i).data.rhythm.interval16); + end + end + + feature(i).data.final.vector = vec; + feature(i).data.final.dim = numel(feature(i).data.final.vector); + + % fill up info struct and append to feature + + info(end+1: feature(i).data.final.dim) = ... + cell(feature(i).data.final.dim - numel(info),1); + + feature(i).data.final.vector_info.labels = info; + end + + % --- + % TODO: Maybe delete more basic features again at this point? + % --- + end + + % --- + % destructor: do we really want to remove this + % from the database? No, but + % TODO: create marker for unused objects in db, and a cleanup + % function + % --- + function delete(feature) + + end + + + function visualise(feature) + % --- + % plots the different data types collected in this feature + % --- + for i = 1:numel(feature) + clip = feature(i).data.info.owner; + + % display raw features + if isa(clip, 'CASIMIRClip') + baseclip = clip.child_clip(); + else + baseclip = clip; + end + if isa(baseclip, 'MTTClip') + rawf = baseclip.audio_features_raw(); + elseif isa(baseclip, 'MSDClip') + rawf = baseclip.features('MSDAudioFeatureRAW'); + end + + % --- + % @todo: implement MSD feature visualisation + % --- + [a1, a2, a3] = rawf.visualise(); + + % --- + % Display chroma features + % --- + if isfield(feature(i).data, 'chroma') + + chroma_labels = {'c', 'c#', 'd','d#', 'e', 'f','f#', 'g','g#', 'a', 'a#', 'h'}; + mode_labels = {'minor', 'major'}; + + % change labels to reflect detected mode + chroma_labels{rawf.data.key + 1} = ... + sprintf('(%s) %s',mode_labels{rawf.data.mode + 1}, chroma_labels{rawf.data.key + 1}); + + % transpose labels and data + chroma_labels = circshift(chroma_labels, [0, feature(i).data.chroma(1).shift]); + chromar = circshift([rawf.data.segments_pitches], [feature(i).data.chroma(1).shift, 0]); + + % image transposed chromas again + segments = [rawf.data.segments_start]; + segments(end) = rawf.data.duration; + + hold(a1); + uimagesc(segments, 0:11, chromar, 'Parent', a1); + set(a1,'YTick',[0:11], 'YTickLabel', chroma_labels); + + % enlarge plot and plot new data after the old ones + ax = axis(a1); + ax(2) = ax(2) + 2*feature(i).my_params.nchromas + 0.5; + axis(a1, 'xy'); + axis(a1, ax); + + imagesc(rawf.data.duration + (1:feature(i).my_params.nchromas), (-1:11), ... + [ feature(i).data.chroma(:).means_weight; feature(i).data.chroma(:).means],... + 'Parent', a1); + % variance calculated? + if isfield(feature(i).data.chroma, 'vars') + + imagesc(rawf.data.duration + feature(i).my_params.nchromas + (1:feature(i).my_params.nchromas), (-1:11), ... + [feature(i).data.chroma(:).vars],... + 'Parent', a1); + end + end + + % --- + % Display timbre features + % --- + if isfield(feature(i).data, 'timbre') + + % enlarge plot and plot new data after the old ones + hold(a2); + ax = axis(a2); + ax(2) = ax(2) + 2*feature(i).my_params.ntimbres + 0.5; + + axis(a2, ax); + imagesc(rawf.data.duration + (1:feature(i).my_params.ntimbres), (-1:11), ... + [ feature(i).data.timbre(:).means_weight; feature(i).data.timbre(:).means],... + 'Parent', a2); + if isfield(feature(i).data.timbre, 'vars') + + imagesc(rawf.data.duration + feature(i).my_params.ntimbres + (1:feature(i).my_params.ntimbres), (-1:11), ... + [feature(i).data.timbre(:).vars],... + 'Parent', a1); + end + end + + % --- + % Display rhythm features + % --- + if isfield(feature(i).data, 'rhythm') + % data.rhythm.interval + % get timecode + eightt = feature(i).data.rhythm.energy8_time; + sixt = feature(i).data.rhythm.energy16_time; + + hold(a3); + % plot sixteens acorr and energy + plot(sixt, feature(i).data.rhythm.energy16, 'bx') + + plot(sixt, feature(i).data.rhythm.acorr16, 'b') + + % plot eights acorr and energy + plot(eightt, feature(i).data.rhythm.energy8, 'rx') + + plot(eightt, feature(i).data.rhythm.acorr8, 'r') + + % broaden view by fixed 4 seconds + ax = axis(a3); + axis(a3, [max(0, eightt(1)-( eightt(end) - eightt(1) + 4 )) ... + min(rawf.data.duration, eightt(end) +4) ... + ax(3:4)]); + end + end + end + end + + + methods (Hidden = true) + + function [env, time] = energy_envelope(feature, clip) + % extracts the envelope of energy for the given clip + + % --- + % TODO: externalise envelope etc in external audio features + % --- + + [null, src] = evalc('miraudio(clip.mp3file_full())'); + [null, env] = evalc('mirenvelope(src, ''Sampling'', feature.my_params.energy_sr)'); + + time = get(env,'Time'); + time = time{1}{1}; + env = mirgetdata(env); + end + + function [acorr, base_sig, base_t] = beat_histogram(feature, startt, interval, signal, signal_t) + % acorr = beat_histogram(feature, startt, interval, signal, time) + % + % compute correlation for beats of specified length in energy curve + + % get corresponding energy values + dt = signal_t(2) - signal_t(1); + base_t = startt:interval:(startt + (feature.my_params.nints*2-1) * interval); + base_sig = signal( min( numel(signal), max(1,round((base_t - signal_t(1))/dt)))); + + % normalise energy + acbase_sig = base_sig./max(base_sig); + + % calculate their cyclic autocorrelation + acorr = circshift(xcorr(acbase_sig,acbase_sig(1:end/2)),... + [numel(acbase_sig) 0]); + + % cut acorr to relevant points, normalise and square + acorr = (acorr(1:feature.my_params.nints)./feature.my_params.nints).^2; + + % --- + % NOTE: we normalise the autocorrelation locally, to compare the + % (rhythmic) shape + % --- + if feature.my_params.norm_acorr; + + acorr = acorr - min(acorr); + acorr = acorr/max(acorr); + end + end + end + + methods(Static) + + function timbre = norm_timbre(in, normfs) + % returns normed timbre data + + % --- + % individually scale the data using + % the dimensions factors + % --- + timbre = zeros(size(in)); + for i = 1:size(in,2) + + timbre(:,i) = normfs .* in(:,i); + end + + % shift to positive values + timbre = (1 + timbre) /2; + + % clip features to [0,1] + timbre = min(1, max(timbre, 0)); + end + + % --- + % returns parameter md5 hash for comparison + % --- + end + +end \ No newline at end of file