int Classemes_init(const Program_Options* _options)
{
char buffer[MAX_CHARACTERS_LINE];
int res;
assert(_options);
options = _options;
/* load Phi */
Phi = (Matrix*) malloc(sizeof(Matrix));
sprintf(buffer, "%s%c"PHI_CLASSEMES_BF, options->parameters_dir, SYSTEM_PATH_SEPARATOR);
res = read_float_matrix(Phi, buffer);
if(res) {
return ERROR_PARAMETERS_LOADING;
}
/* load Tau */
Tau = (Matrix*) malloc(sizeof(Matrix));
sprintf(buffer, "%s%c"TAU_CLASSEMES_BF, options->parameters_dir, SYSTEM_PATH_SEPARATOR);
res = read_float_matrix(Tau, buffer);
if(res) {
return ERROR_PARAMETERS_LOADING;
}
return 0;
}
bool Lab2VitMain::init_utt()
{
if (m_audioStrm.peek() == EOF)
return false;
m_idStr = read_float_matrix(m_audioStrm, m_inAudio);
cout << "Processing utterance ID: " << m_idStr << endl;
m_frontEnd.get_feats(m_inAudio, m_feats);
if (m_feats.size2() != m_gmmSet.get_dim_count())
throw runtime_error("Mismatch in GMM and feat dim.");
if (m_doAlign)
{
if (m_graphStrm.peek() == EOF)
throw runtime_error("Mismatch in number of audio files "
"and FSM's.");
m_graph.read(m_graphStrm, m_idStr);
}
if (m_graph.get_gmm_count() > m_gmmSet.get_gmm_count())
throw runtime_error("Mismatch in number of GMM's between "
"FSM and GmmSet.");
m_gmmSet.calc_gmm_probs(m_feats, m_gmmProbs);
// Initialize dynamic programming chart.
m_chart.resize(m_feats.size1() + 1, m_graph.get_state_count());
m_chart.clear();
if (m_graph.get_start_state() < 0)
throw runtime_error("Graph has no start state.");
return true;
}
bool Lab2TrainMain::init_utt()
{
if (m_audioStrm.peek() == EOF)
return false;
m_idStr = read_float_matrix(m_audioStrm, m_inAudio);
m_frontEnd.get_feats(m_inAudio, m_feats);
if (m_feats.size2() != m_gmmSet.get_dim_count())
throw runtime_error("Mismatch in GMM and feat dim.");
if (m_alignStrm.peek() == EOF)
throw runtime_error("Mismatch in number of audio files "
"and alignments.");
read_int_vector(m_alignStrm, m_gmmList, m_idStr);
if (m_gmmList.size() != m_feats.size1())
throw runtime_error("Mismatch in alignment and feat lengths.");
int gmmCnt = m_gmmSet.get_gmm_count();
int frmCnt = m_gmmList.size();
m_gmmCountList.clear();
for (int frmIdx = 0; frmIdx < frmCnt; ++frmIdx)
{
int gmmIdx = m_gmmList[frmIdx];
if ((gmmIdx < 0) || (gmmIdx >= gmmCnt))
throw runtime_error("Out of range GMM index.");
m_gmmCountList.push_back(GmmCount(gmmIdx, frmIdx, 1.0));
}
return true;
}
void main_loop(const char** argv)
{
// Process command line arguments.
map<string, string> params;
process_cmd_line(argv, params);
bool verbose = get_bool_param(params, "verbose");
// Load feature files for templates.
// Get template label from matrix name.
vector<string> templateLabelList;
vector<matrix<double> > templateMatList;
ifstream templateStrm(
get_required_string_param(params, "template_file").c_str());
while (templateStrm.peek() != EOF)
{
templateMatList.push_back(matrix<double>());
string labelStr = read_float_matrix(templateStrm,
templateMatList.back());
templateLabelList.push_back(labelStr);
}
templateStrm.close();
if (templateMatList.empty())
throw runtime_error("No templates supplied.");
// Load correct label for each feature file, if present.
vector<string> featLabelList;
string labelFile = get_string_param(params, "feat_label_list");
if (!labelFile.empty())
read_string_list(labelFile, featLabelList);
// The main loop.
ifstream featStrm(
get_required_string_param(params, "feat_file").c_str());
matrix<double> feats;
unsigned templCnt = templateLabelList.size();
unsigned uttCnt = 0;
unsigned correctCnt = 0;
while (featStrm.peek() != EOF)
{
int uttIdx = uttCnt++;
string idStr = read_float_matrix(featStrm, feats);
// Find closest template.
int bestTempl = -1;
double bestScore = DBL_MAX;
for (unsigned templIdx = 0; templIdx < templCnt; ++templIdx)
{
if (feats.size2() != templateMatList[templIdx].size2())
throw runtime_error("Mismatch in test/template feature dim.");
double curScore = compute_distance(feats,
templateMatList[templIdx]);
if (verbose)
cout << format(" %s: %.3f") % templateLabelList[templIdx] %
curScore << endl;
if (curScore < bestScore)
bestScore = curScore, bestTempl = templIdx;
}
if (bestTempl < 0)
throw runtime_error("No alignments found in DTW.");
string hypLabel = (bestTempl >= 0) ? templateLabelList[bestTempl] : "";
if (!featLabelList.empty())
{
// If have reference labels, print ref and hyp classes.
if (uttIdx >= (int) featLabelList.size())
throw runtime_error("Mismatch in number of utterances "
"and labels.");
string refLabel = featLabelList[uttIdx];
cout << format("Reference: %s, Hyp: %s, Correct: %d") %
refLabel % hypLabel % (hypLabel == refLabel) << endl;
if (hypLabel == refLabel)
++correctCnt;
}
else
// If don't have reference labels, just print hyp class.
cout << hypLabel << " (" << idStr << ")" << endl;
}
featStrm.close();
if (!featLabelList.empty())
{
// If have reference labels, print accuracy.
unsigned errCnt = uttCnt - correctCnt;
cout << format("Accuracy: %.2f%% (%d/%d), Error rate: %.2f%% (%d/%d)")
% (100.0 * correctCnt / uttCnt) % correctCnt % uttCnt %
(100.0 * errCnt / uttCnt) % errCnt % uttCnt << endl;
}
}
