aboutsummaryrefslogtreecommitdiff
path: root/octave/vq_700c_eq.m
diff options
context:
space:
mode:
Diffstat (limited to 'octave/vq_700c_eq.m')
-rw-r--r--octave/vq_700c_eq.m366
1 files changed, 0 insertions, 366 deletions
diff --git a/octave/vq_700c_eq.m b/octave/vq_700c_eq.m
deleted file mode 100644
index 3e57b9a..0000000
--- a/octave/vq_700c_eq.m
+++ /dev/null
@@ -1,366 +0,0 @@
-% vq_700c.m
-% David Rowe May 2019
-%
-% Researching Codec 2 700C VQ equaliser ideas
-% See also scripts/train_700c_quant.sh, tnewamp1.m
-
-melvq;
-newamp_700c;
-
-% general purpose plot function for looking at averages of K-band
-% sequences in scripts dir and VQs:
-% vq_700c_plots({"hts2a.f32" "vk5qi.f32" "train_120_1.txt"})
-
-function vq_700c_plots(fn_array)
- K = 20; rate_K_sample_freqs_kHz = mel_sample_freqs_kHz(K);
- freq_Hz = rate_K_sample_freqs_kHz * 1000;
-
- figure(1); clf; hold on; axis([200 4000 40 90]); title('Max Hold');
- figure(2); clf; hold on; axis([200 4000 0 40]); title('Average');
-
- for i=1:length(fn_array)
- [dir name ext] = fileparts(fn_array{i});
- if strcmp(ext, ".f32")
- % f32 feature file
- fn = sprintf("../build_linux/%s%s", name, ext)
- bands = load_f32(fn , K);
- else
- % text file (e.g. existing VQ)
- bands = load(fn_array{i});
- end
- % for max hold: break into segments of Nsec, find max, average maximums
- % this avoids very rare global peaks setting the max
- Nsec = 10; Tframe = 0.01; frames_per_seg = Nsec/Tframe
- Nsegs = floor(length(bands)/frames_per_seg)
- max_holds = zeros(Nsegs, K);
- if Nsegs == 0
- max_holds = max(bands)
- else
- for s=1:Nsegs
- st = (s-1)*frames_per_seg+1; en = st + frames_per_seg - 1;
- max_holds(s,:) = max(bands(st:en,:));
- end
- max_holds = mean(max_holds);
- end
- figure(1); plot(freq_Hz, max_holds, '+-', 'linewidth', 2);
- figure(2); plot(freq_Hz, mean(bands), '+-', 'linewidth', 2);
- end
- figure(1); legend(fn_array); grid; xlabel('Freq (Hz)'); ylabel('Amp dB');
- figure(2); legend(fn_array); grid; xlabel('Freq (Hz)'); ylabel('Amp dB');
-endfunction
-
-
-% limit mean of each vector to between lower_lim and upper_lim
-function vout = limit_vec(vin, lower_lim, upper_lim)
- m = mean(vin');
- vout = zeros(size(vin));
- for i=1:length(vin)
- vec_no_mean = vin(i,:) - m(i);
- if m(i) < lower_lim
- m(i) = lower_lim;
- end
- if m(i) > upper_lim
- m(i) = upper_lim;
- end
- vout(i,:) = vec_no_mean + m(i);
- end
-endfunction
-
-
-% single stage vq a target matrix
-function errors = vq_targets(vq, targets)
- errors = [];
- for i=1:length(targets)
- [mse_list index_list] = search_vq(vq, targets(i,:), 1);
- error = targets(i,:) - vq(index_list(1),:);
- errors = [errors; error];
- end
-endfunction
-
-
-% single stage vq a target matrix with adaptive EQ, this didn't work
-
-function [errors eqs] = vq_targets_adap_eq(vq, targets, eqs)
- errors = []; gain=0.02;
- eq = eqs(end,:);
- for i=1:length(targets)
- t = targets(i,:) - eq;
- mean(t)
- %t -= mean(t);
- [mse_list index_list] = search_vq(vq, t, 1);
- error = t - vq(index_list(1),:);
- eq = (1-gain)*eq + gain*error;
- errors = [errors; error]; eqs = [eqs; eq];
- end
-endfunction
-
-
-% single stage vq a target matrix with block adaptive EQ, this works
-% well with nblock == 10
-
-function [errors eq] = vq_targets_block_eq(vq, targets, eq, nblock)
- errors = []; n = 0; [tmp K] = size(vq); error_eq = zeros(1,K); gain=0.20;
- for i=1:length(targets)
- t = targets(i,:) - eq;
- [mse_list index_list] = search_vq(vq, t, 1);
- error = t - vq(index_list(1),:);
- error_eq += error;
- errors = [errors; error];
- n++;
- if n == nblock
- eq = 0.99*eq + gain*error_eq/nblock;
- n = 0; error_eq = zeros(1,K);
- end
- end
-endfunction
-
-
-% two stage mbest VQ a target matrix
-
-function [errors targets_] = vq_targets2(vq1, vq2, targets)
- vqset(:,:,1)= vq1; vqset(:,:,2)=vq2; m=5;
- [errors targets_] = mbest(vqset, targets, m);
-endfunction
-
-
-% two stage mbest VQ a target matrix, with adap_eq
-
-function [errors targets_ eq] = vq_targets2_adap_eq(vq1, vq2, targets, eq)
- vqset(:,:,1)= vq1; vqset(:,:,2)=vq2; m=5; gain=0.02;
- errors = []; targets_ = [];
- for i=1:length(targets)
- t = targets(i,:)-eq;
- t -= mean(t')';
- [error target_ indexes] = mbest(vqset, t, m);
- % use first stage VQ as error driving adaptive EQ
- eq_error = t - vq1(indexes(1),:);
- eq = (1-gain)*eq + gain*eq_error;
- errors = [errors; error]; targets_ = [targets_; target_];
- end
-endfunction
-
-
-% Given target and vq matrices, estimate eq via two metrics. First
-% metric seems to work best. Both uses first stage VQ error for EQ
-
-function [eq1 eq2] = est_eq(vq, targets)
- [ntargets K] = size(targets);
- [nvq K] = size(vq);
-
- eq1 = zeros(1,K); eq2 = zeros(1,K);
- for i=1:length(targets)
- [mse_list index_list] = search_vq(vq, targets(i,:), 1);
-
- % eq metric 1: average of error for best VQ entry
- eq1 += targets(i,:) - vq(index_list(1),:);
-
- % eq metric 2: average of error across all VQ entries
- for j=1:nvq
- eq2 += targets(i,:) - vq(j,:);
- end
- end
-
- eq1 /= ntargets;
- eq2 /= (ntargets*nvq);
-endfunction
-
-function [targets e] = load_targets(fn_target_f32)
- nb_features = 41;
- K = 20;
-
- % .f32 files are in scripts directory, first K values rate_K_no_mean vectors
- [dir name ext] = fileparts(fn_target_f32);
- fn = sprintf("../script/%s_feat.f32", name);
- feat = load_f32(fn, nb_features);
- e = feat(:,1);
- targets = feat(:,2:K+1);
-endfunction
-
-% rather simple EQ in front of VQ
-
-function [eqs ideal] = est_eq_front(targets)
- [tmp K] = size(targets);
- ideal = [ 8 10 12 14 14*ones(1,K-1-4) -20];
- eq = zeros(1,K); gain = 0.02;
- eqs = [];
- for i=1:length(targets)
- update = targets(i,:) - ideal;
- eq = (1-gain)*eq + gain*update;
- eq(find(eq < 0)) = 0;
- eqs = [eqs; eq];
- end
-endfunction
-
-function table_across_samples
- K = 20;
-
- % VQ is in .txt file in this directory, we have two to choose from. train_120 is the Codec 2 700C VQ,
- % train_all_speech was trained up from a different, longer database, as a later exercise
- vq_name = "train_120";
- #vq_name = "train_all_speech";
- vq1 = load(sprintf("%s_1.txt", vq_name));
- vq2 = load(sprintf("%s_2.txt", vq_name));
-
- printf("----------------------------------------------------------------------------------\n");
- printf("Sample Initial vq1 vq1_eq2 vq1_eq2 vq2 vq2_eq1 vq2_eq2 \n");
- printf("----------------------------------------------------------------------------------\n");
-
- fn_targets = { "cq_freedv_8k_lfboost" "cq_freedv_8k_hfcut" "cq_freedv_8k" "hts1a" "hts2a" "cq_ref" "ve9qrp_10s" "vk5qi" "c01_01_8k" "ma01_01"};
- #fn_targets = {"cq_freedv_8k_lfboost"};
- figs=1;
- for i=1:length(fn_targets)
-
- % load target and estimate eq
- [targets e] = load_targets(fn_targets{i});
- eq1 = est_eq(vq1, targets);
- eq2s = est_eq_front(targets);
- % for these simulation uses fixed EQ sample, rather than letting it vary frame by frame
- eq2 = eq2s(end,:);
-
- % first stage VQ -----------------
-
- errors1 = vq_targets(vq1, targets);
- errors1_eq1 = vq_targets(vq1, targets-eq1);
- errors1_eq2 = vq_targets(vq1, targets-eq2);
-
- % two stage mbest VQ --------------
-
- [errors2 targets_] = vq_targets2(vq1, vq2, targets);
- [errors2_eq1 targets_eq1_] = vq_targets2(vq1, vq2, targets-eq1);
- [errors2_eq2 targets_eq2_] = vq_targets2(vq1, vq2, targets-eq2);
-
- % save to .f32 files for listening tests
- if strcmp(vq_name,"train_120")
- save_f32(sprintf("../script/%s_vq2.f32", fn_targets{i}), targets_);
- save_f32(sprintf("../script/%s_vq2_eq1.f32", fn_targets{i}), targets_eq1_);
- save_f32(sprintf("../script/%s_vq2_eq2.f32", fn_targets{i}), targets_eq2_);
- else
- save_f32(sprintf("../script/%s_vq2_as.f32", fn_targets{i}), targets_);
- save_f32(sprintf("../script/%s_vq2_as_eq.f32", fn_targets{i}), targets_eq_);
- end
- printf("%-21s %6.2f %6.2f %6.2f %6.2f %6.2f %6.2f %6.2f\n", fn_targets{i},
- var(targets(:)), var(errors1(:)), var(errors1_eq1(:)), var(errors1_eq2(:)),
- var(errors2(:)), var(errors2_eq1(:)), var(errors2_eq2(:)));
-
- figure(figs++); clf;
- %plot(var(errors2'),'b;vq2;'); hold on; plot(var(errors2_eq1'),'g;vq2_eq1;'); plot(var(errors2_eq2'),'r;vq2_eq2;'); hold off;
- plot(mean(targets),'g;mean(targets);'); hold on; plot(mean(vq1),'g;mean(vq1);'); plot(eq2,'r;eq2;'); hold off;
- title(fn_targets{i}); axis([1 K -20 30]);
- end
-endfunction
-
-
-% interactve, menu driven frame by frame plots
-
-function interactive(fn_vq_txt, fn_target_f32)
- K = 20;
- vq = load("train_120_1.txt");
- [targets e] = load_targets(fn_target_f32);
- eq1 = est_eq(vq, targets);
-
- [errors1_eq2 eqs2] = vq_targets_adap_eq(vq, targets, zeros(1,K));
- [errors1_eq2 eqs2] = vq_targets_adap_eq(vq, targets, eqs2(end,:));
- eq2 = eqs2(end,:);
-
- figure(1); clf;
- mesh(e+targets)
- figure(2); clf;
- plot(eq1,'b;eq1;')
- hold on;
- plot(mean(targets),'c;mean(targets);'); plot(eq2,'g;eq2;');
- hold off;
- figure(3); clf; mesh(eqs2); title('eq2 evolving')
-
- % enter single step loop
- f = 20; neq = 0; eq=zeros(1,K);
- do
- figure(4); clf;
- t = targets(f,:) - eq;
- [mse_list index_list] = search_vq(vq, t, 1);
- error = t - vq(index_list(1),:);
- plot(e(f)+t,'b;target;');
- hold on;
- plot(e(f)+vq(index_list,:),'g;vq;');
- plot(error,'r;error;');
- plot(eq,'c;eq;');
- plot([1 K],[e(f) e(f)],'--')
- hold off;
- axis([1 K -20 80])
- % interactive menu
-
- printf("\r f: %2d eq: %d ind: %3d var: %3.1f menu: n-next b-back e-eq q-quit", f, neq, index_list(1), var(error));
- fflush(stdout);
- k = kbhit();
-
- if k == 'n' f+=1; end
- if k == 'e'
- neq++;
- end
- if neq == 3 neq = 0; end
- if neq == 0 eq = zeros(1,K); end
- if neq == 1 eq = eq1; end
- if neq == 2 eq = eqs2(f,:); end
- if k == 'b' f-=1; end
- until (k == 'q')
- printf("\n");
-endfunction
-
-
-% Experiment to test iterative approach of block update and remove
-% mean (ie frame energy), shows some promise at reducing HF energy
-% over several iterations while not affecting already good samples
-
-function experiment_iterate_block(fn_vq_txt, fn_target_f32)
- K = 20;
- vq = load("train_120_1.txt");
- [targets e] = load_targets(fn_target_f32);
-
- figure(1); clf;
- plot(mean(targets),'b;mean(targets);');
- hold on;
- plot(mean(vq), 'g;mean(vq);');
- figure(2); clf; hold on;
- eq = zeros(1,K);
- for i=1:3
- [errors eq] = vq_targets_block_eq(vq, targets, eq, 10);
- figure(1); plot(mean(targets-eq));
- figure(2); plot(eq);
- printf("i: %d %6.2f\n", i, var(errors(:)))
- end
-endfunction
-
-% Experiment to test EQ of input (before) VQ. We set a threshold on
-% when to equalise, so we don't upset already flat-ish samples. This
-% is the algorithm used for C at the time of writing (newamp1.c, newamp_700c.m)
-
-function experiment_front_eq(fn_vq_txt, fn_target_f32)
- K = 20;
- vq = load("train_120_1.txt");
- [targets e] = load_targets(fn_target_f32);
-
- [eqs ideal] = est_eq_front(targets);
-
- figure(1); clf;
- plot(mean(targets),'b;mean(targets);');
- hold on;
- plot(ideal, 'g;ideal;');
- plot(eqs(end,:), 'r;eq;');
- plot(mean(targets)-eqs(end,:), 'c;equalised;');
- plot(mean(vq),'b--;mean(vq);');
- hold off;
- figure(2); clf; mesh(eqs(1:100,:)); title('EQ weights over time');
- ylabel('Time (frames'); xlabel('Freq (mel)');
-endfunction
-
-more off
-
-% choose one of these to run first
-% You'll need to run scripts/train_700C_quant.sh first to generate the .f32 files
-
-%interactive("train_120_1.txt", "cq_freedv_8k_lfboost.f32")
-%table_across_samples;
-%vq_700c_plots({"all_speech_8k.f32" "all_speech_8k_hp300.f32" "dev-clean-8k.f32" "train_8k.f32" } )
-%vq_700c_plots({"ve9qrp_10s.f32" "cq_freedv_8k_lfboost.f32" "cq_ref.f32" "hts1a.f32" "vk5qi.f32"})
-%experiment_iterate_block("train_120_1.txt", "ve9qrp_10s.f32")
-%experiment_iterate_block("train_120_1.txt", "cq_freedv_8k_lfboost.f32")
-%experiment_front_eq("train_120_1.txt", "cq_freedv_8k_lfboost.f32")
generated by cgit v1.2.3 (git 2.46.0) at 2026-01-26 22:15:29 +0000