The BWE stimuli had noisy artifacts. Now the BWE signals are windowed…

… to exclude the noisy part at the end. There was also an issue with the reproducibility of the BWE, which was due to some random components in the algorithm. This was changed by explicitly seeding the random number generator. The seed was carefully selected to avoid any further artifacts. Under certain (as of yet unknown) circumstances, the BEW algorithm otherwise produces audible artifacts within the male /u/. The selected see minimizes this artifact so that it can be assumed to be no longer audible. (#3) Co-authored-by: Simon Stone <[email protected]>
TUD-STKS · Oct 21, 2021 · 0da559c · 0da559c
1 parent b105eed
commit 0da559c
Show file tree

Hide file tree

Showing 65 changed files with 21 additions and 9 deletions.
diff --git a/include/cleanupBweSignal.m b/include/cleanupBweSignal.m
@@ -4,21 +4,24 @@
 %   parts around the speech signal. This function gets rid of those by
 %   fading the actual speech signal in and out.
 %
-%   y: Bandwith-extended speech signal consisting of [silence, speech,
+%   y: Bandwidth-extended speech signal consisting of [silence, speech,
 %   silence]
 %   sil: Duration of the symmetric (!) silence before and after the
 %   speech
 
-% Number of non-silent samples (with a little bit of slack)
-nSpeech = length(y) - floor(1.8*sil);
+% Number of non-silent samples
+nSpeech = length(y) - floor(2.2*sil);
 
 % Fading part of the window
-win = tukeywin(nSpeech, 0.05);
+win = tukeywin(nSpeech, 0.1);
 
 % Extend by zeros
 win = padarray(win, floor((length(y) - length(win)) / 2), 'pre');
 win = padarray(win, length(y) - length(win), 'post');
 
+% Move window back a little bit to get rid of noise tail
+win = circshift(win, floor(-0.1*sil));
+
 y = y .* win;
 end
 
diff --git a/include/extend_to_8kHz.m b/include/extend_to_8kHz.m
@@ -13,7 +13,7 @@
 global path_to_GMM;
 path_to_GMM = '../include/bandwidth_extension/ABE_explicit_memory_ICASSP18/2_GMM_training/existing_models/';
 
-past_frames = 2; future_frames=2; inp_feature= 'LogMFE_zs_pca'; dimX=10; dimY=10; 
+past_frames = 1; future_frames=1; inp_feature= 'LogMFE_zs_pca'; dimX=10; dimY=10; 
 x_wb = logmfe_lpc_abe(x_nb, inp_feature, past_frames, future_frames, dimX, dimY);
 
 if size(x_wb,1)==1

diff --git a/stimuli/StimuliGeneration.m b/stimuli/StimuliGeneration.m
@@ -1,6 +1,10 @@
 %% Generation of the stimuli for the listening experiments
 
 clc; close all; clearvars;
+% Make BWE results reproducible
+rng(17); % Tiny artifact in male /u/
+
+
 %%
 addpath('../include')
 addpath('../include/bandwidth_extension')
@@ -9,7 +13,7 @@
 
 load('filters.mat');
 
-plot_blended_tf = true;
+plot_blended_tf = false;
 
 %% Transfer functions
 tf_mm_path = '../transfer-functions/multimodal';
@@ -78,9 +82,14 @@
 Ug.male = Ug.male .* p_male;
 Ug.female = Ug.female .* p_female;
 
-% Add initial and final silence
-Ug.male = padarray(Ug.male, floor(sil_s * Fs_out), 'both');
-Ug.female = padarray(Ug.female, floor(sil_s * Fs_out), 'both');
+% Add final silence 
+Ug.male = padarray(Ug.male, floor(sil_s * Fs_out) - length(Ug.male) + length(natural_ref.pressure_contour), 'post');
+Ug.female = padarray(Ug.female, floor(sil_s * Fs_out) - length(Ug.female) + length(natural_ref.pressure_contour), 'post');
+
+% Add initial
+Ug.male = padarray(Ug.male, floor(sil_s * Fs_out), 'pre');
+Ug.female = padarray(Ug.female, floor(sil_s * Fs_out), 'pre');
+
 %% Synthesize
 playlist = {};
 for file = tf_mm_files'

diff --git a/stimuli/dev/f_a_MM_1d_modal.wav b/stimuli/dev/f_a_MM_1d_modal.wav
diff --git a/stimuli/dev/f_a_MM_1d_pressed.wav b/stimuli/dev/f_a_MM_1d_pressed.wav
diff --git a/stimuli/dev/f_a_MM_MM_modal.wav b/stimuli/dev/f_a_MM_MM_modal.wav
diff --git a/stimuli/dev/f_a_MM_MM_pressed.wav b/stimuli/dev/f_a_MM_MM_pressed.wav
diff --git a/stimuli/dev/f_a_MM_bwe_modal.wav b/stimuli/dev/f_a_MM_bwe_modal.wav
diff --git a/stimuli/dev/f_a_MM_bwe_pressed.wav b/stimuli/dev/f_a_MM_bwe_pressed.wav
diff --git a/stimuli/dev/f_e_MM_1d_modal.wav b/stimuli/dev/f_e_MM_1d_modal.wav
diff --git a/stimuli/dev/f_e_MM_1d_pressed.wav b/stimuli/dev/f_e_MM_1d_pressed.wav
diff --git a/stimuli/dev/f_e_MM_MM_modal.wav b/stimuli/dev/f_e_MM_MM_modal.wav
diff --git a/stimuli/dev/f_e_MM_MM_pressed.wav b/stimuli/dev/f_e_MM_MM_pressed.wav
diff --git a/stimuli/dev/f_e_MM_bwe_modal.wav b/stimuli/dev/f_e_MM_bwe_modal.wav
diff --git a/stimuli/dev/f_e_MM_bwe_pressed.wav b/stimuli/dev/f_e_MM_bwe_pressed.wav
diff --git a/stimuli/dev/f_i_MM_1d_modal.wav b/stimuli/dev/f_i_MM_1d_modal.wav
diff --git a/stimuli/dev/f_i_MM_1d_pressed.wav b/stimuli/dev/f_i_MM_1d_pressed.wav
diff --git a/stimuli/dev/f_i_MM_MM_modal.wav b/stimuli/dev/f_i_MM_MM_modal.wav
diff --git a/stimuli/dev/f_i_MM_MM_pressed.wav b/stimuli/dev/f_i_MM_MM_pressed.wav
diff --git a/stimuli/dev/f_i_MM_bwe_modal.wav b/stimuli/dev/f_i_MM_bwe_modal.wav
diff --git a/stimuli/dev/f_i_MM_bwe_pressed.wav b/stimuli/dev/f_i_MM_bwe_pressed.wav
diff --git a/stimuli/dev/f_o_MM_1d_modal.wav b/stimuli/dev/f_o_MM_1d_modal.wav
diff --git a/stimuli/dev/f_o_MM_1d_pressed.wav b/stimuli/dev/f_o_MM_1d_pressed.wav
diff --git a/stimuli/dev/f_o_MM_MM_modal.wav b/stimuli/dev/f_o_MM_MM_modal.wav
diff --git a/stimuli/dev/f_o_MM_MM_pressed.wav b/stimuli/dev/f_o_MM_MM_pressed.wav
diff --git a/stimuli/dev/f_o_MM_bwe_modal.wav b/stimuli/dev/f_o_MM_bwe_modal.wav
diff --git a/stimuli/dev/f_o_MM_bwe_pressed.wav b/stimuli/dev/f_o_MM_bwe_pressed.wav
diff --git a/stimuli/dev/f_u_MM_1d_modal.wav b/stimuli/dev/f_u_MM_1d_modal.wav
diff --git a/stimuli/dev/f_u_MM_1d_pressed.wav b/stimuli/dev/f_u_MM_1d_pressed.wav
diff --git a/stimuli/dev/f_u_MM_MM_modal.wav b/stimuli/dev/f_u_MM_MM_modal.wav
diff --git a/stimuli/dev/f_u_MM_MM_pressed.wav b/stimuli/dev/f_u_MM_MM_pressed.wav
diff --git a/stimuli/dev/f_u_MM_bwe_modal.wav b/stimuli/dev/f_u_MM_bwe_modal.wav
diff --git a/stimuli/dev/f_u_MM_bwe_pressed.wav b/stimuli/dev/f_u_MM_bwe_pressed.wav
diff --git a/stimuli/dev/m_a_MM_1d_modal.wav b/stimuli/dev/m_a_MM_1d_modal.wav
diff --git a/stimuli/dev/m_a_MM_1d_pressed.wav b/stimuli/dev/m_a_MM_1d_pressed.wav
diff --git a/stimuli/dev/m_a_MM_MM_modal.wav b/stimuli/dev/m_a_MM_MM_modal.wav
diff --git a/stimuli/dev/m_a_MM_MM_pressed.wav b/stimuli/dev/m_a_MM_MM_pressed.wav
diff --git a/stimuli/dev/m_a_MM_bwe_modal.wav b/stimuli/dev/m_a_MM_bwe_modal.wav
diff --git a/stimuli/dev/m_a_MM_bwe_pressed.wav b/stimuli/dev/m_a_MM_bwe_pressed.wav
diff --git a/stimuli/dev/m_e_MM_1d_modal.wav b/stimuli/dev/m_e_MM_1d_modal.wav
diff --git a/stimuli/dev/m_e_MM_1d_pressed.wav b/stimuli/dev/m_e_MM_1d_pressed.wav
diff --git a/stimuli/dev/m_e_MM_MM_modal.wav b/stimuli/dev/m_e_MM_MM_modal.wav
diff --git a/stimuli/dev/m_e_MM_MM_pressed.wav b/stimuli/dev/m_e_MM_MM_pressed.wav
diff --git a/stimuli/dev/m_e_MM_bwe_modal.wav b/stimuli/dev/m_e_MM_bwe_modal.wav
diff --git a/stimuli/dev/m_e_MM_bwe_pressed.wav b/stimuli/dev/m_e_MM_bwe_pressed.wav
diff --git a/stimuli/dev/m_i_MM_1d_modal.wav b/stimuli/dev/m_i_MM_1d_modal.wav
diff --git a/stimuli/dev/m_i_MM_1d_pressed.wav b/stimuli/dev/m_i_MM_1d_pressed.wav
diff --git a/stimuli/dev/m_i_MM_MM_modal.wav b/stimuli/dev/m_i_MM_MM_modal.wav
diff --git a/stimuli/dev/m_i_MM_MM_pressed.wav b/stimuli/dev/m_i_MM_MM_pressed.wav
diff --git a/stimuli/dev/m_i_MM_bwe_modal.wav b/stimuli/dev/m_i_MM_bwe_modal.wav
diff --git a/stimuli/dev/m_i_MM_bwe_pressed.wav b/stimuli/dev/m_i_MM_bwe_pressed.wav
diff --git a/stimuli/dev/m_o_MM_1d_modal.wav b/stimuli/dev/m_o_MM_1d_modal.wav
diff --git a/stimuli/dev/m_o_MM_1d_pressed.wav b/stimuli/dev/m_o_MM_1d_pressed.wav
diff --git a/stimuli/dev/m_o_MM_MM_modal.wav b/stimuli/dev/m_o_MM_MM_modal.wav
diff --git a/stimuli/dev/m_o_MM_MM_pressed.wav b/stimuli/dev/m_o_MM_MM_pressed.wav
diff --git a/stimuli/dev/m_o_MM_bwe_modal.wav b/stimuli/dev/m_o_MM_bwe_modal.wav
diff --git a/stimuli/dev/m_o_MM_bwe_pressed.wav b/stimuli/dev/m_o_MM_bwe_pressed.wav
diff --git a/stimuli/dev/m_u_MM_1d_modal.wav b/stimuli/dev/m_u_MM_1d_modal.wav
diff --git a/stimuli/dev/m_u_MM_1d_pressed.wav b/stimuli/dev/m_u_MM_1d_pressed.wav
diff --git a/stimuli/dev/m_u_MM_MM_modal.wav b/stimuli/dev/m_u_MM_MM_modal.wav
diff --git a/stimuli/dev/m_u_MM_MM_pressed.wav b/stimuli/dev/m_u_MM_MM_pressed.wav
diff --git a/stimuli/dev/m_u_MM_bwe_modal.wav b/stimuli/dev/m_u_MM_bwe_modal.wav
diff --git a/stimuli/dev/m_u_MM_bwe_pressed.wav b/stimuli/dev/m_u_MM_bwe_pressed.wav
diff --git a/stimuli/dev/zf_o_MM_Peter_MM_modal.wav b/stimuli/dev/zf_o_MM_Peter_MM_modal.wav
diff --git a/stimuli/dev/zf_o_MM_Peter_bwe_modal.wav b/stimuli/dev/zf_o_MM_Peter_bwe_modal.wav