/************************************************************************
Function: main
Description: Main entry point
Inputs: int argc, char *argv[] - main input params
Output: int - always 0.
Comments: none.
***********************************************************************/
int main(int argc, char **argv)
{
// Parse command line
unsigned int frame_rate;
double min_phoneme_length, max_phoneme_length;
string silence_symbol;
bool remove_silence;
double C = 0.0;
double beta1, beta2, beta3;
string scores_filelist;
string dists_filelist;
string phonemes_filelist;
string start_times_filelist;
string phonemes_filename;
string phoneme_stats_filename;
string classifier_filename;
string output_align;
string output_confidence;
learning::cmd_line cmdline;
cmdline.info("Forced Alignment based on Passive-Aggressive");
cmdline.add("-frame_rate", "frame rate (shift) in msec [10]",&frame_rate,10);
cmdline.add("-min_phoneme_length", "min. phoneme duration in msec [20]", &min_phoneme_length, 20);
cmdline.add("-max_phoneme_length", "max. phoneme duration in msec [330]", &max_phoneme_length, 330);
cmdline.add("-silence_symbol", "silence symbol [sil]", &silence_symbol, "sil");
cmdline.add("-remove_silence", "remove pre/post silence from data", &remove_silence, false);
cmdline.add("-beta1", "weight of the distance feature", &beta1, 1.0);
cmdline.add("-beta2", "weight of the duration feature", &beta2, 1.0);
cmdline.add("-beta3", "weight of the speaking rate feature", &beta3, 1.0);
cmdline.add("-output_align", "file list where the forced alignemnt is written", &output_align, "");
cmdline.add("-output_confidence", "single file where the forced alignemnt confidence is written", &output_confidence, "");
cmdline.add_master_option("scores_filelist", &scores_filelist);
cmdline.add_master_option("dists_filelist", &dists_filelist);
cmdline.add_master_option("phonemes_filelist", &phonemes_filelist);
cmdline.add_master_option("start_times_filelist [null]", &start_times_filelist);
cmdline.add_master_option("phonemes", &phonemes_filename);
cmdline.add_master_option("phoneme-stats", &phoneme_stats_filename);
cmdline.add_master_option("classifier", &classifier_filename);
int rc = cmdline.parse(argc, argv);
if (rc < 7) {
cmdline.print_help();
return EXIT_FAILURE;
}
// phoneme symbol to number mapping (Lee and Hon, 89)
PhonemeSequence::load_phoneme_map(phonemes_filename, silence_symbol);
// Initiate classifier
Classifier classifier(frame_rate, min_phoneme_length, max_phoneme_length, C, beta1, beta2, beta3, 0.0);
classifier.load(classifier_filename);
classifier.load_phoneme_stats(phoneme_stats_filename);
// begining of the training set
Dataset test_dataset(scores_filelist, dists_filelist, phonemes_filelist, start_times_filelist);
int num_boundaries = 0;
int cummulative_loss = 0;
int cum_loss_less_than[NUM_CUM_LOSS_RESOLUTIONS+1];
for (uint t=1; t <= NUM_CUM_LOSS_RESOLUTIONS; t++)
cum_loss_less_than[t] = 0;
StringVector output_align_files;
if (output_align != "")
output_align_files.read(output_align);
ofstream output_confidence_ofs;
if (output_confidence != "") {
output_confidence_ofs.open(output_confidence.c_str());
if (!output_confidence_ofs.good()) {
cerr << "Error: unable to open " << output_confidence << "for writing." << endl;
}
}
// Run over all dataset
for (uint i=0; i < test_dataset.size(); i++) {
SpeechUtterance x;
StartTimeSequence y;
StartTimeSequence y_hat;
cout << "==================================================================================" << endl;
// read next example for dataset
test_dataset.read(x, y, remove_silence);
y_hat.resize(x.phonemes.size());
// predict label
double confidence = classifier.predict(x, y_hat);
cout << "phonemes=" << x.phonemes << endl;
if (test_dataset.labels_given())
cout << "alignment= " << y << endl;
cout << "predicted= " << y_hat << endl;
cout << "confidence= " << confidence << endl;
if (output_align != "") {
ofstream output_align_ofs(output_align_files[i].c_str());
if (output_align_ofs.good()) {
for (uint j=0; j < y_hat.size(); j++) {
output_align_ofs << y_hat[j] << endl;
}
output_align_ofs.close();
}
}
if (output_confidence != "" && output_confidence_ofs.good())
output_confidence_ofs << confidence << endl;
// calculate the error
if (test_dataset.labels_given()) {
int file_loss = 0;
int cur_loss;
for (unsigned int j=0; j < y.size(); ++j) {
if (y[j] > y_hat[j]) {
cur_loss = y[j] - y_hat[j];
} else {
cur_loss = y_hat[j] - y[j];
}
file_loss += cur_loss;
cummulative_loss += cur_loss;
for (int t=1; t <= NUM_CUM_LOSS_RESOLUTIONS; t++)
if ( cur_loss <= t ) cum_loss_less_than[t]++;
}
num_boundaries += y.size();
cout << "File loss = " << file_loss/double(y.size()) << endl;
cout << "Cum loss = " << cummulative_loss/double(num_boundaries) << endl;
for (uint t = NUM_CUM_LOSS_RESOLUTIONS; t >= 1; t--) {
cout << "% Boundaries (t <= " << t*frame_rate << "ms) = "
<< 100.0*cum_loss_less_than[t]/double(num_boundaries) << "\n";
}
cout << endl;
}
}
if (output_confidence != "" && output_confidence_ofs.good())
output_confidence_ofs.close();
cout << "Done." << endl;
return EXIT_SUCCESS;
}
/************************************************************************
Function: Dataset::read
Description: Read next instance and label
Inputs: SpeechUtterance&
StartTimeSequence&
Output: void.
Comments: none.
***********************************************************************/
uint Dataset::read(SpeechUtterance &x, StartTimeSequence &y, bool remove_silence, bool enable_printouts)
{
uint sil_offset = 0;
if (single_file_mode) {
std::cout << "current file=" << current_file << std::endl;
std::cout << scores_file_list[0] << std::endl;
std::cout << dists_file_list[0] << std::endl;
std::cout << phonemes_file_list[0] << std::endl;
x.read(scores_file_list[0], dists_file_list[0],
phonemes_file_list[current_file],single_file_mode);
}
else {
if (enable_printouts) {
std::cout << "current file=" << current_file << std::endl;
std::cout << scores_file_list[current_file] << std::endl;
std::cout << dists_file_list[current_file] << std::endl;
std::cout << phonemes_file_list[current_file] << std::endl;
}
x.read(scores_file_list[current_file], dists_file_list[current_file],
phonemes_file_list[current_file],single_file_mode);
}
if (read_labels) {
std::cout << start_times_file_list[current_file] << std::endl;
y.read(start_times_file_list[current_file]);
}
else {
y.resize(x.phonemes.size());
}
++current_file;
if (remove_silence) {
///////////////////////////
// the code below is used to remove the leading silence at the
// begining of each utterance and the following silnce at the end
x.silence_offset = y[1];
x.last_silence = y[y.size()-1];
infra::matrix tmp1(x.scores);
x.scores.resize(y[y.size()-1]-y[1],tmp1.width());
x.scores = tmp1.submatrix(y[1],0,x.scores.height(),x.scores.width());
// remove silences
for (uint i=0;i<x.phonemes.size()-2;++i)
x.phonemes[i] = x.phonemes[i+1];
x.phonemes.resize(x.phonemes.size()-2);
infra::matrix tmp2(x.distances);
x.distances.resize(y[y.size()-1]-y[1],tmp2.width());
x.distances = tmp2.submatrix(y[1],0,x.distances.height(),x.distances.width());
uint tmp3 = y[1];
for (uint i=0;i<y.size()-2;++i)
y[i] = y[i+1]-tmp3;
y.resize(y.size()-2);
///////////////////////////
}
return sil_offset;
}
