sc_compute_suf_stats.rar_Factor Analysis

% make_suf_stat(scp_script,out_dir) % Makes sufficient statistic for FA % scp_script input files % out_dir output directory % load model params (as supervectors) m=load('models/ubm_means')'; v=load('models/ubm_variances')'; w=load('models/ubm_weights')'; n_mixtures = size(w, 1); dim = size(m, 1) / n_mixtures; % we load the model as superverctors, so we reshape it to have each gaussian in % one column m = reshape(m, dim, n_mixtures); v = reshape(v, dim, n_mixtures); % these are the sets that we want to extract the stats for data_sets{1,1} = 'enroll_stats'; data_sets{2,1} = 'fa_train_eigenchannels_stats'; data_sets{3,1} = 'fa_train_eigenvoices_stats'; data_sets{4,1} = 'test_stats'; % process each dataset for set_i = 1:size(data_sets,1) set_list_file = [ 'lists/' data_sets{set_i,1} '.lst' ]; disp(['Processing list ' set_list_file]); % process the file list (logical=physical) [spk_logical spk_physical] = parse_list(set_list_file); n_sessions = size(spk_logical, 1); % initialize the matrices for the stats % one row per session N = zeros(n_sessions, n_mixtures); F = zeros(n_sessions, n_mixtures * dim); % process each session for session_i = 1:n_sessions session_name = [ spk_physical{session_i,1} '.ascii' ]; disp(['Reading feature file ' session_name]); data = load(session_name, '-ascii')'; disp('Processing...'); % process the feature file tic [Ni Fi] = collect_suf_stats(data, m, v, w); toc N(session_i,:) = Ni; F(session_i,:) = Fi; end out_stats_file = [ '/data/stats/' data_sets{set_i,1} '.mat' ]; disp(['Saving stats to ' out_stats_file]); % save(out_stats_file, 'N', 'F', 'spk_logical'); end