SAcC_train_v1.0

   This package shows how to train SAcC on a new dataset. Using
   genPitchLabel, the pseudo ground truth f0 can be generated from
   the outputs of multiple pitch trackers. (e.g. YIN, Wu, and SWIPE')

   You can download a zip file containing demo_SAcC_train.m and the
   other files used in this demo from SAcC_train_v1.0.zip.

   Note that the dataset files are not included. Please change the
   scripts files to run this demo according to the current dataset.

OBJECTIVE:
PROCEDURE:
REFERENCE:
USAGE:
Software packages:
Training data:
Subband PCA file:
CONTACT:

----------------------------------------

OBJECTIVE:

         Procedure for training Subband Autocorrelation Classification
         (SAcC) pitch tracker on a new speech corpus [1].

PROCEDURE:

   a. generate the labels using genPitchLabel with YIN, Wu, and SWIPE'.
      (the audio files are read from list files.)
      Save the labels in pfile format.
   b. calculate subband PCA features and save in pfile for MLP training
   c. train MLP from pfile using QuickNet

REFERENCE:

       [1] Byung Suk Lee, Noise Robust Pitch Tracking by Subband
           Autocorrelation Classification (SAcC),
           PhD Thesis, Columbia University, 2012
----------------------------------------

USAGE:

This package includes an example script:

    demo_SAcC_train.m

This example trains a SAcC for a Babel dataset. Due to large volumn, the dataset is not included. To run this script, the following is required.

Software packages:

   audioread, YIN, Wu, SWIPEP, QuickNet, genPitchLabel
   (Please update the paths of the software.)

Training data:

   lists/Babel_tr.list
   (The files in the list are required or to be updated as needed.)

Subband PCA file:

   ../../dat/keele/rbf_pinknoise/out/PCA_sr8000_bpo6_nchs30.mat
   (Pleas update the path.)

To train on a new dataset, replace the list with the new data files.

conf_fn = 'conf/config.trBabel_sr8k_bpo6_sb24_k10';
P = load_config(conf_fn);

maxyth = 0.5900;
maxsth = 0.1200;
dispflag = 0;
out_base = 'logs/';
tst_base = 'logs/tst/';
if ~exist(out_base,'dir'), mkdir(out_base); end;
if ~exist(tst_base,'dir'), mkdir(tst_base); end;
reject_target = 1;
nMLP = P.npcf + 1; % +1 for no-pitch
cexp = P.qe;
allR = [];
tstR = [];
name_base = 'Babel_';

tr_list_file = 'lists/Babel_tr.list';
tst_list_file = 'lists/Babel_tst.list';
tr_dat_pfile = 'tmp/Babel_tr_sr8k_bpo6_sb24_k10.pfile';
tst_dat_pfile = 'tmp/Babel_tst_sr8k_bpo6_sb24_k10.pfile';

% a. Generate pseudo pitch label using genPitchLabel with Wu, YIN, SWIPE'
cal_agreed_label(tr_list_file,P.pcf,maxyth,maxsth,dispflag);
cal_agreed_label(tst_list_file,P.pcf,maxyth,maxsth,dispflag);

% b. Calculate Subband Autocorrelation PCA and save in pfiles
[tr_l_pfile] = preprocessing_sbac(tr_dat_pfile,tr_list_file,P.sr,P.mapping,P.b2,P.a2,P.t2);
[tst_l_pfile] = preprocessing_sbac(tst_dat_pfile,tst_list_file,P.sr,P.mapping,P.b2,P.a2,P.t2);

% c. Train MLP using QuickNet
[Rd,Ud,Fd,Ld,Td,hdr_sized] = pfinfo(tr_dat_pfile);
trRange = ['0:',num2str(round(P.trR*(Ud+1)))];
[d_path,d_name,d_ext] = fileparts(tr_dat_pfile);
n_fn = [d_path,'/',d_name,'.norms'];
% generate norm file
if ~exist(n_fn,'file')
qnnrm = 'qnnorm '; qo1 = [' norm_ftrfile=',tr_dat_pfile];
qo2 = [' output_normfile=',n_fn];
cmd = [qnnrm,qo1,qo2]; disp(cmd); system(cmd);
end

TRname = [name_base,'h',num2str(P.hids)];
w_fn = [out_base,'wgt/',TRname,'.wgt'];
if ~exist([out_base,'wgt/'],'dir'), mkdir([out_base,'wgt/']); end;
w_log = [out_base,'log/',TRname,'.log'];
if ~exist([out_base,'log/'],'dir'), mkdir([out_base,'log/']); end;
w_chk = [out_base,'chk/',TRname,'.chk'];
if ~exist([out_base,'chk/'],'dir'), mkdir([out_base,'chk/']); end;
w_chklog = [out_base,'chk/',TRname,'.log'];
if ~exist(w_fn,'file')
qnstrn_wrap(tr_dat_pfile,n_fn,tr_l_fn,w_fn,w_log,w_chklog,w_chk,Fd,Ud,allR,trRange,P.hids,P.wl,nMLP,cexp,[],[],reject_target);
end

alg=SAcC_fast
nchs=24
sr=8000
fmin=100
bpo=6
q=8
n=2
ftype=2
dataset=Babel
ntype=radio_channel
tst_dataset=Babel
tst_ntype=radio_channel
feature_names=sbpca_features
pca_file=../../dat/keele/rbf_pinknoise/out/PCA_sr8000_bpo6_nchs30.mat
kdim=10
QN_Learn_Rate_Param=0
trR=0.6
hids=100
n_s=10
train_src_list=lists/sub-qtr-rats-src.list
test_src_list=lists/tst-qtr-rats-src.list
train_file=tmp/tr_rats_sr8k_bpo6_sb24_k10.pfile
train_label=tmp/sub-qtr-rats-lab.pfile
test_file=tmp/tst_rats_sr8k_bpo6_sb24_k10.pfile
test_label=tmp/tst-sub-qtr-rats-lab.pfile
Using Slaney-Patterson filterbank, frq=100..1600, bpo=6

...