% Copyright David Rowe 2016
% This program is distributed under the terms of the GNU General Public License
% Version 2
-%
-% Octave script to batch process model parameters using the new
-% amplitude model. Used for generating samples we can listen to.
-%
#{
+ Octave script to batch process model parameters using the new
+ amplitude model, version 1. Outputs anotehr set of model parameters
+ that can be fed to c2sim for listening tests. The companion
+ newamp1_fbf.m script is used to visualise the processing frame by frame
+
+ c2sim -> dump files -> newamp1_batch.m -> output model params -> c2sim -> play
+
+ The newamp1_xxx scripts have evolved to (i) resample {Am} using a
+ mel frequency axis, (ii) 2 stage VQ the mean removed vector. Seems to work
+ OK at 700 bit/s.
+
Usage:
~/codec2-dev/build_linux/src$ ./c2sim ../../raw/hts1a.raw --dump hts1a
octave:14> newamp1_batch("../build_linux/src/hts1a")
~/codec2-dev/build_linux/src$ ./c2sim ../../raw/hts1a.raw --amread hts1a_am.out -o - | play -t raw -r 8000 -s -2 -
- Or with a little more processing, first dump energy and voicing, the import Wo, voicing, phase spectra:
+ Or with a little more processing, first dump energy and voicing, the
+ import Wo, voicing, phase spectra which simulates all the decoder
+ DSP. We switch on lpc 10 just to dump voicing.
$ ./c2sim ../../raw/vk5qi.raw --phase0 --postfilter --dump vk5qi --lpc 10 --dump_pitch_e vk5qi_pitche.txt
octave:14> newamp1_batch("../build_linux/src/vk5qi", "../build_linux/src/vk5qi_am_Wo.out");
#}
-% process a whole file and write results
-% TODO:
-% [ ] refactor decimate-in-time to avoid extra to/from model conversions
-% [ ] switches to turn on/off quantisation
-% [ ] rename mask_sample_freqs, do we need "mask" any more
-
-#{
- [ ] refactor
-#}
-
% In general, this function processes a bunch of amplitudes, we then
-% use c2sim to hear the results
+% use c2sim to hear the results. Bunch of different experiments below
function surface = newamp1_batch(samname, optional_Am_out_name, optional_Aw_out_name)
newamp;
more off;
max_amp = 80;
- postfilter = 0;
+ postfilter = 0; % optional postfiler that runs on Am, not used atm
synth_phase = 1;
model_name = strcat(samname,"_model.txt");
%[model_ surface] = experiment_mel_freq(model, 1, 1, voicing);
%model_ = experiment_dec_abys(model, 8, 1, 1, 1, voicing);
- [amodel_ avoicing_ indexes] = experiment_rate_K_dec(model, voicing);
- [model_ voicing_] = model_from_indexes(indexes);
+
+ [amodel_ avoicing_ indexes] = experiment_rate_K_dec(model, voicing); % encoder, toss away results except for indexes
+ [model_ voicing_] = model_from_indexes(indexes); % decoder uses just indexes, outputs vecs for synthesis
%model_ = experiment_dec_linear(model_);
%model_ = experiment_energy_rate_linear(model, 1, 0);
endfunction
-% experiment to resample freq axis on mel scale, then optionally vq
-
-function [model_ rate_K_surface] = experiment_mel_freq(model, vq_en=0, plots=1, voicing)
- [frames nc] = size(model);
- K = 20;
- [rate_K_surface rate_K_sample_freqs_kHz] = resample_const_rate_f_mel(model, K);
-
- if plots
- figure(1); clf; mesh(rate_K_surface);
- figure(2); clf; plot_dft_surface(rate_K_surface)
- end
-
- for f=1:frames
- mean_f(f) = mean(rate_K_surface(f,:));
- rate_K_surface_no_mean(f,:) = rate_K_surface(f,:) - mean_f(f);
- end
-
- if vq_en
- melvq;
- load train_120_vq; m=5;
-
- [res rate_K_surface_ ind] = mbest(train_120_vq, rate_K_surface_no_mean, m);
-
- for f=1:frames
- rate_K_surface_(f,:) = post_filter(rate_K_surface_(f,:), 1.5, voicing(f));
- end
-
- for f=1:frames
- rate_K_surface_(f,:) += mean_f(f);
- end
-
- rate_K_surface = rate_K_surface_;
- end
-
- model_ = resample_rate_L(model, rate_K_surface, rate_K_sample_freqs_kHz);
-
- if plots
- figure(3); clf; mesh(rate_K_surface_no_mean);
- figure(4); clf; plot_dft_surface(rate_K_surface)
- end
-
- for f=1:frames
- rate_K_surface(f,:) -= mean(rate_K_surface(f,:));
- end
-endfunction
-
-
-% mel spaced sampling, differential in time VQ. Curiously, couldn't
-% get very good results out of this, I suspect a bug
-
-function [model_ rate_K_surface_diff] = experiment_mel_diff_freq(model, vq_en=0)
- [frames nc] = size(model);
- K = 20;
- [rate_K_surface rate_K_sample_freqs_kHz] = resample_const_rate_f_mel(model, K);
-
- if vq_en
- melvq;
- load surface_diff_vq; m=5;
- end
-
- for f=1:frames
- mean_f(f,:) = mean(rate_K_surface(f,:));
- rate_K_surface_no_mean(f,:) = rate_K_surface(f,:) - mean_f(f,:);
- end
-
- rate_K_surface_no_mean_ = zeros(frames, K);
- rate_K_surface_no_mean_diff = zeros(frames, K);
- rate_K_surface_(1,:) = rate_K_surface_diff(1,:) = zeros(1, K);
-
- for f=2:frames
- rate_K_surface_diff(f,:) = rate_K_surface_no_mean(f,:) - 0.8*rate_K_surface_no_mean_(f-1,:);
- if vq_en
- [res arate_K_surface_diff_ ind] = mbest(surface_diff_vq, rate_K_surface_diff(f,:), m);
- rate_K_surface_diff_(f,:) = arate_K_surface_diff_;
- else
- rate_K_surface_diff_(f,:) = rate_K_surface_diff(f,:);
- end
- rate_K_surface_no_mean_(f,:) = 0.8*rate_K_surface_no_mean_(f-1,:) + rate_K_surface_diff_(f,:);
- end
-
- for f=1:frames
- rate_K_surface_(f,:) = rate_K_surface_no_mean_(f,:) + mean_f(f,:);
- end
-
- model_ = resample_rate_L(model, rate_K_surface_, rate_K_sample_freqs_kHz);
-
-endfunction
-
-
-% try vq with open and closed loop mean removal, turns out they give
-% identical results lol
-
-function [model_ rate_K_surface] = experiment_closed_loop_mean(model)
- [frames nc] = size(model);
- K = 15;
- mel_start = ftomel(300); mel_end = ftomel(3000);
- step = (mel_end-mel_start)/(K-1);
- mel = mel_start:step:mel_end;
- rate_K_sample_freqs_Hz = 700*((10 .^ (mel/2595)) - 1);
- rate_K_sample_freqs_kHz = rate_K_sample_freqs_Hz/1000;
-
- rate_K_surface = resample_const_rate_f(model, rate_K_sample_freqs_kHz);
-
- load surface_vq; m=1;
-
- for f=1:frames
- amean = mean(rate_K_surface(f,:));
- rate_K_target_no_mean = rate_K_surface(f,:) - amean;
- mse_open_loop(f) = search_vq2(surface_vq(:,:,1), rate_K_target_no_mean, m, 0);
- mse_closed_loop(f) = search_vq2(surface_vq(:,:,1), rate_K_surface(f,:), m, 1);
- end
-
- printf("rms open loop..: %f\nrms closed loop: %f\n", sqrt(mean(mse_open_loop)), sqrt(mean(mse_closed_loop)));
-
- % just return model_ as we have to so nothing breaks, it's not actually useful
-
- model_ = resample_rate_L(model, rate_K_surface, rate_K_sample_freqs_kHz);
-
-endfunction
-
-
-% Experiment with 10ms update rate for energy but 40ms update for spectrum,
-% using linear interpolation for spectrum.
-
-function model_c = experiment_energy_rate_linear(model, vq_en, plot_en)
- max_amp = 80;
- [frames nc] = size(model);
-
- % 10ms mel freq modelling and VQ
-
- model_ = experiment_mel_freq(model, vq_en, plot_en);
-
- % Remove energy. Hmmmm, this is done on Ams rather than surface but that's
- % similar I guess
-
- e = zeros(1,frames);
- model_a = zeros(frames,max_amp+3);
- for f=1:frames
- L = min([model_(f,2) max_amp-1]);
- Am_ = model_(f,3:(L+2));
- AmdB_ = 20*log10(Am_);
- mean_f(f) = mean(AmdB_);
- AmdB_ -= mean_f(f);
- model_a(f,1) = model_(f,1); model_a(f,2) = L; model_a(f,3:(L+2)) = 10 .^ (AmdB_(1:L)/20);
- end
-
- % linear interp after removing energy (mean AmdB)
-
- model_b = experiment_dec_linear(model_a);
-
- % add back in energy
-
- model_c = zeros(frames,max_amp+3);
- for f=1:frames
- L = min([model_b(f,2) max_amp-1]);
- Am_ = model_b(f,3:(L+2));
- AmdB_ = 20*log10(Am_);
- AmdB_ += mean_f(f);
- model_c(f,1) = model_b(f,1); model_c(f,2) = L; model_c(f,3:(L+2)) = 10 .^ (AmdB_(1:L)/20);
- end
-
-endfunction
-
-
-% conventional decimation in time without any filtering, then linear
-% interpolation. Linear interpolation is a two-tap (weak) form of fir
-% filtering that may have problems with signals with high freq
-% components, for example after quantisation noise is added. Need to
-% look into this some more.
-
-function model_ = experiment_dec_linear(model)
- newamp;
- max_amp = 80;
-
- [frames nc] = size(model);
- model_ = zeros(frames, max_amp+3);
- decimate = 4;
- for f=1:frames
- AmdB_ = decimate_frame_rate(model, decimate, f, frames);
- L = length(AmdB_);
- model_(f,1) = model(f,1); model_(f,2) = L; model_(f,3:(L+2)) = 10 .^ (AmdB_(1:L)/20);
- end
-endfunction
-
-
+% -----------------------------------------------------------------------------------------
% Linear decimator/interpolator that operates at rate K, includes VQ, post filter, and Wo/E
-% quantisation. Evolved from abys decimator below.
+% quantisation. Evolved from abys decimator below. Simulates the entire encoder/decoder.
function [model_ voicing_ indexes] = experiment_rate_K_dec(model, voicing)
max_amp = 80;
endfunction
-% Stand alone decoder that takes indexes and creates model_
+% ---------------------------------------------------------------------------------------
+% Stand alone decoder that takes indexes and creates model_, models
+% decoder and an important step in proving everything works
function [model_ voicing_] = model_from_indexes(indexes)
max_amp = 80;
endfunction
+% ---------------------------------------------------------------------------------------
+% Various experiments tried during development
+
+% experiment to resample freq axis on mel scale, then optionally vq
+
+function [model_ rate_K_surface] = experiment_mel_freq(model, vq_en=0, plots=1, voicing)
+ [frames nc] = size(model);
+ K = 20;
+ [rate_K_surface rate_K_sample_freqs_kHz] = resample_const_rate_f_mel(model, K);
+
+ if plots
+ figure(1); clf; mesh(rate_K_surface);
+ figure(2); clf; plot_dft_surface(rate_K_surface)
+ end
+
+ for f=1:frames
+ mean_f(f) = mean(rate_K_surface(f,:));
+ rate_K_surface_no_mean(f,:) = rate_K_surface(f,:) - mean_f(f);
+ end
+
+ if vq_en
+ melvq;
+ load train_120_vq; m=5;
+
+ [res rate_K_surface_ ind] = mbest(train_120_vq, rate_K_surface_no_mean, m);
+
+ for f=1:frames
+ rate_K_surface_(f,:) = post_filter(rate_K_surface_(f,:), 1.5, voicing(f));
+ end
+
+ for f=1:frames
+ rate_K_surface_(f,:) += mean_f(f);
+ end
+
+ rate_K_surface = rate_K_surface_;
+ end
+
+ model_ = resample_rate_L(model, rate_K_surface, rate_K_sample_freqs_kHz);
+
+ if plots
+ figure(3); clf; mesh(rate_K_surface_no_mean);
+ figure(4); clf; plot_dft_surface(rate_K_surface)
+ end
+
+ for f=1:frames
+ rate_K_surface(f,:) -= mean(rate_K_surface(f,:));
+ end
+endfunction
+
+
+% mel spaced sampling, differential in time VQ. Curiously, couldn't
+% get very good results out of this, I suspect a bug
+
+function [model_ rate_K_surface_diff] = experiment_mel_diff_freq(model, vq_en=0)
+ [frames nc] = size(model);
+ K = 20;
+ [rate_K_surface rate_K_sample_freqs_kHz] = resample_const_rate_f_mel(model, K);
+
+ if vq_en
+ melvq;
+ load surface_diff_vq; m=5;
+ end
+
+ for f=1:frames
+ mean_f(f,:) = mean(rate_K_surface(f,:));
+ rate_K_surface_no_mean(f,:) = rate_K_surface(f,:) - mean_f(f,:);
+ end
+
+ rate_K_surface_no_mean_ = zeros(frames, K);
+ rate_K_surface_no_mean_diff = zeros(frames, K);
+ rate_K_surface_(1,:) = rate_K_surface_diff(1,:) = zeros(1, K);
+
+ for f=2:frames
+ rate_K_surface_diff(f,:) = rate_K_surface_no_mean(f,:) - 0.8*rate_K_surface_no_mean_(f-1,:);
+ if vq_en
+ [res arate_K_surface_diff_ ind] = mbest(surface_diff_vq, rate_K_surface_diff(f,:), m);
+ rate_K_surface_diff_(f,:) = arate_K_surface_diff_;
+ else
+ rate_K_surface_diff_(f,:) = rate_K_surface_diff(f,:);
+ end
+ rate_K_surface_no_mean_(f,:) = 0.8*rate_K_surface_no_mean_(f-1,:) + rate_K_surface_diff_(f,:);
+ end
+
+ for f=1:frames
+ rate_K_surface_(f,:) = rate_K_surface_no_mean_(f,:) + mean_f(f,:);
+ end
+
+ model_ = resample_rate_L(model, rate_K_surface_, rate_K_sample_freqs_kHz);
+
+endfunction
+
+
+% try vq with open and closed loop mean removal, turns out they give
+% identical results lol
+
+function [model_ rate_K_surface] = experiment_closed_loop_mean(model)
+ [frames nc] = size(model);
+ K = 15;
+ mel_start = ftomel(300); mel_end = ftomel(3000);
+ step = (mel_end-mel_start)/(K-1);
+ mel = mel_start:step:mel_end;
+ rate_K_sample_freqs_Hz = 700*((10 .^ (mel/2595)) - 1);
+ rate_K_sample_freqs_kHz = rate_K_sample_freqs_Hz/1000;
+
+ rate_K_surface = resample_const_rate_f(model, rate_K_sample_freqs_kHz);
+
+ load surface_vq; m=1;
+
+ for f=1:frames
+ amean = mean(rate_K_surface(f,:));
+ rate_K_target_no_mean = rate_K_surface(f,:) - amean;
+ mse_open_loop(f) = search_vq2(surface_vq(:,:,1), rate_K_target_no_mean, m, 0);
+ mse_closed_loop(f) = search_vq2(surface_vq(:,:,1), rate_K_surface(f,:), m, 1);
+ end
+
+ printf("rms open loop..: %f\nrms closed loop: %f\n", sqrt(mean(mse_open_loop)), sqrt(mean(mse_closed_loop)));
+
+ % just return model_ as we have to so nothing breaks, it's not actually useful
+
+ model_ = resample_rate_L(model, rate_K_surface, rate_K_sample_freqs_kHz);
+
+endfunction
+
+
+% Experiment with 10ms update rate for energy but 40ms update for spectrum,
+% using linear interpolation for spectrum.
+
+function model_c = experiment_energy_rate_linear(model, vq_en, plot_en)
+ max_amp = 80;
+ [frames nc] = size(model);
+
+ % 10ms mel freq modelling and VQ
+
+ model_ = experiment_mel_freq(model, vq_en, plot_en);
+
+ % Remove energy. Hmmmm, this is done on Ams rather than surface but that's
+ % similar I guess
+
+ e = zeros(1,frames);
+ model_a = zeros(frames,max_amp+3);
+ for f=1:frames
+ L = min([model_(f,2) max_amp-1]);
+ Am_ = model_(f,3:(L+2));
+ AmdB_ = 20*log10(Am_);
+ mean_f(f) = mean(AmdB_);
+ AmdB_ -= mean_f(f);
+ model_a(f,1) = model_(f,1); model_a(f,2) = L; model_a(f,3:(L+2)) = 10 .^ (AmdB_(1:L)/20);
+ end
+
+ % linear interp after removing energy (mean AmdB)
+
+ model_b = experiment_dec_linear(model_a);
+
+ % add back in energy
+
+ model_c = zeros(frames,max_amp+3);
+ for f=1:frames
+ L = min([model_b(f,2) max_amp-1]);
+ Am_ = model_b(f,3:(L+2));
+ AmdB_ = 20*log10(Am_);
+ AmdB_ += mean_f(f);
+ model_c(f,1) = model_b(f,1); model_c(f,2) = L; model_c(f,3:(L+2)) = 10 .^ (AmdB_(1:L)/20);
+ end
+
+endfunction
+
+
+% conventional decimation in time without any filtering, then linear
+% interpolation. Linear interpolation is a two-tap (weak) form of fir
+% filtering that may have problems with signals with high freq
+% components, for example after quantisation noise is added. Need to
+% look into this some more.
+
+function model_ = experiment_dec_linear(model)
+ newamp;
+ max_amp = 80;
+
+ [frames nc] = size(model);
+ model_ = zeros(frames, max_amp+3);
+ decimate = 4;
+ for f=1:frames
+ AmdB_ = decimate_frame_rate(model, decimate, f, frames);
+ L = length(AmdB_);
+ model_(f,1) = model(f,1); model_(f,2) = L; model_(f,3:(L+2)) = 10 .^ (AmdB_(1:L)/20);
+ end
+endfunction
+
+
% Experimental AbyS decimator that chooses best frames to match
% surface based on AbyS approach. Can apply post filter at different
#{
-todo: get this working again
+
+ My original idea was to used a 3-4 "resonators" to construct a
+ piecewise model of the spectrum. Kind of got distracted by the
+ surface and mel sampling that ended up working OK. This method was
+ working OK, soem issues with backgorund noise but rather easy to
+ quantise.
+
+ todo: get this working again
function model_ = experiment_piecewise(model)
% encoder loop ------------------------------------------------------
#}
-
-
-
% early test, devised to test rate K<->L changes along frequency axis
function model_ = resample_half_frame_offset(model, rate_K_surface, rate_K_sample_freqs_kHz)
endfunction
+
+% helper function to plot surfaces
+
function plot_dft_surface(surface)
[frames K] = size(surface);
projects / freetel-svn-tracking.git / commitdiff