void LDA<Scalar>::fit(const Eigen::MatrixXi &X) {
auto corpus = get_corpus(X);
for (size_t i=0; i<iterations_; i++) {
partial_fit(corpus);
}
}
int resynthesize(Options& opts) {
readCoefOptions(opts);
auto index = util::parse<unsigned>(opts.input);
auto corpus = get_corpus(opts);
assert(index < corpus.size());
std::vector<PhonemeInstance> input = corpus.input(index);
INFO("Input file: " << alphabet_test.file_data_of(input[0]).file);
INFO("Total duration: " << get_total_duration(input));
INFO("Original cost: " << concat_cost(input, crf, crf.lambda, input));
std::vector<int> path;
traverse_automaton<MinPathFindFunctions, CRF, 1>(input, crf, crf.lambda, &path);
std::vector<PhonemeInstance> output = crf.alphabet().to_phonemes(path);
SynthPrinter sp(crf.alphabet(), labels_all);
if(opts.has_opt("verbose"))
sp.print_synth(path, input);
sp.print_textgrid(path, input, labels_synth, opts.text_grid);
CRF::Stats stats;
INFO("Resynth. cost: " << concat_cost(output, crf, crf.lambda, input, &stats));
//INFO("Second best cost: " << costs[1]);
auto baselineCost = concat_cost(output, baseline_crf, baseline_crf.lambda, input);
INFO("Baseline cost: " << baselineCost);
outputStats(crf.lambda, stats, opts);
outputPath(opts, output, input);
auto sws = SpeechWaveSynthesis(output, input, crf.alphabet());
Wave outputSignal = sws.get_resynthesis(opts);
outputSignal.write(opts.get_opt<std::string>("output", "resynth.wav"));
auto sws2 = SpeechWaveSynthesis(input, input, alphabet_test);
auto concatenation = sws2.get_concatenation();
concatenation.write(opts.get_opt<std::string>("original", "original.wav"));
if(opts.has_opt("verbose")) {
Comparisons cmp;
cmp.fill(concatenation, outputSignal);
INFO("LogSpectrum = " << cmp.LogSpectrum);
INFO("LogSpectrumCritical = " << cmp.LogSpectrumCritical);
INFO("SegSNR = " << cmp.SegSNR);
INFO("MFCC = " << cmp.MFCC);
INFO("WSS = " << cmp.WSS);
outputComparisons(opts, cmp, baselineCost);
}
return 0;
}
int baseline(const Options& opts) {
readCoefOptions(opts);
auto index = opts.get_opt<unsigned>("input", 0);
auto corpus = get_corpus(opts);
std::vector<PhonemeInstance> input = corpus.input(index);
INFO("Input file: " << alphabet_test.file_data_of(input[0]).file);
INFO("Total duration: " << get_total_duration(input));
std::vector<int> path;
traverse_automaton<MinPathFindFunctions>(input, baseline_crf,
baseline_crf.lambda, &path);
std::vector<PhonemeInstance> output = baseline_crf.alphabet().to_phonemes(path);
Wave outputSignal = SpeechWaveSynthesis(output, input, baseline_crf.alphabet())
.get_resynthesis_td();
outputSignal.write(opts.get_opt<std::string>("output", "baseline.wav"));
auto sws2 = SpeechWaveSynthesis(input, input, alphabet_test);
auto concatenation = sws2.get_concatenation();
concatenation.write(opts.get_opt<std::string>("original", "original.wav"));
CRF::Stats stats;
INFO("Baseline cost in original: " << concat_cost(output, crf, crf.lambda, input, &stats));
auto baselineCost = concat_cost(output, baseline_crf, baseline_crf.lambda, input);
INFO("Baseline cost:" << baselineCost);
outputStats(crf.lambda, stats, opts);
outputPath(opts, output, input);
if(opts.has_opt("verbose")) {
auto sws = SpeechWaveSynthesis(input, input, alphabet_test);
auto concatenation = sws.get_concatenation();
Comparisons cmp;
cmp.fill(concatenation, outputSignal);
INFO("LogSpectrum = " << cmp.LogSpectrum);
INFO("LogSpectrumCritical = " << cmp.LogSpectrumCritical);
INFO("SegSNR = " << cmp.SegSNR);
INFO("MFCC = " << cmp.MFCC);
INFO("WSS = " << cmp.WSS);
outputComparisons(opts, cmp, baselineCost);
}
return 0;
}
int psola(const Options& opts) {
auto inputString = opts.get_opt<std::string>("input", "");
auto inputPhonemes = util::split_string(inputString, ',');
std::vector<int> input(inputPhonemes.size());
std::transform(inputPhonemes.begin(), inputPhonemes.end(), input.begin(), util::parse<int>);
std::vector<PhonemeInstance> phonemeInput;
std::vector<PhonemeInstance> phonemeOutput;
auto pitchScale = opts.get_opt<double>("pitch-scale", 1.0);
auto durationScale = opts.get_opt<double>("duration-scale", 1.0);
auto& alphabet = opts.has_opt("synth") ? alphabet_synth : alphabet_test;
if(input.size() > 1) {
phonemeInput = alphabet.to_phonemes(input);
} else {
Corpus& corpus = get_corpus(opts);
phonemeInput = corpus.input(input[0]);
INFO("Input file: " << alphabet.file_data_of(phonemeInput[0]).file);
}
for(auto p : phonemeInput) {
p.end += p.duration * std::abs(1 - durationScale);
p.duration += p.duration * std::abs(1 - durationScale);
p.pitch_contour[0] += std::log(pitchScale);
p.pitch_contour[1] += std::log(pitchScale);
phonemeOutput.push_back(p);
}
auto sws = SpeechWaveSynthesis(phonemeInput, phonemeOutput, alphabet);
auto outputSignal = sws.get_resynthesis(opts);
auto original = sws.get_concatenation();
original.write(opts.get_opt<std::string>("original", ""));
outputSignal.write(opts.get_opt<std::string>("output", ""));
return 0;
}
AlignmentCorpus::AlignmentCorpus(string file_name): number_source_words_(0), number_target_words_(0)
{
if(file_name.find("training_data_file_names_english_to_french",0) == 0){
translation_type_ = "english_to_french";
}
else if(file_name.find("training_data_file_names_french_to_english",0) == 0){
translation_type_ = "french_to_english";
}
else if(file_name.find("test_data_file_names_english_to_french.",0) == 0){
translation_type_ = "english_to_french";
}
else if(file_name.find("test_data_file_names_french_to_english.",0) == 0){
translation_type_ = "french_to_english";
}
else{
cout << "A type of error happened! "<< endl;
exit(1);
}
ifstream data_file;
data_file.open(file_name.c_str());
string data_line, source_file_name, target_file_name;
string::size_type split_position;
int i(0);
while(true){
getline(data_file,data_line);
split_position = data_line.find(' ', 0);
target_file_name = data_line.substr(split_position + 1);
source_file_name = data_line.substr(0,split_position);
if(source_file_name.size() == 0){
break;
}
get_corpus(source_file_name, target_file_name);
}
data_file.close();
}
typename LDA<Scalar>::MatrixX LDA<Scalar>::transform(const Eigen::MatrixXi& X) {
// cast the parameters to what is needed
auto model = std::static_pointer_cast<parameters::ModelParameters<Scalar> >(
model_parameters_
);
// make some room for the transformed data
MatrixX gammas(model->beta.rows(), X.cols());
// make a corpus to use
auto corpus = get_corpus(X);
// Queue all the documents
for (size_t i=0; i<corpus->size(); i++) {
queue_in_.emplace_back(corpus, i);
}
// create the thread pool
create_worker_pool();
// Extract variational parameters and calculate the doc_e_step
for (size_t i=0; i<corpus->size(); i++) {
std::shared_ptr<parameters::Parameters> vp;
size_t index;
std::tie(vp, index) = extract_vp_from_queue();
gammas.col(index) = std::static_pointer_cast<parameters::VariationalParameters<Scalar> >(vp)->gamma;
// tell the thread safe event dispatcher to process the events from the
// workers
process_worker_events();
}
// destroy the thread pool
destroy_worker_pool();
return gammas;
}
void LDA<Scalar>::partial_fit(const Eigen::MatrixXi &X, const Eigen::VectorXi &y) {
partial_fit(get_corpus(X, y));
}
void parse(int fd, wchar_t* buf, char* dir)
{
/* word word */
CorpusInfo *corpus = NULL;
wchar_t *ptr;
int direction = 0;
nat_boolean_t exact_match = FALSE;
nat_boolean_t both = FALSE;
wchar_t words[50][150];
int i = 0;
wchar_t *token = NULL;
char tmp[150];
chomp(buf);
for (i=0;i<50;i++)
wcscpy(words[i], L"");
i = 0;
if (wcscmp(buf, L"") == 0) return;
#if DEBUG
LOG("Request was [%s]", buf);
#endif
token = wcstok(buf, L" ", &ptr);
while(token) {
wcscpy(words[i], token);
i++;
token = wcstok(NULL, L" ", &ptr);
}
if (wcsncmp(words[0], L"LIST", 4) == 0) {
dump_corpora_list(fd, LAST_CORPORA, CORPORA);
return;
} else if (wcsncmp(words[0], L"??", 2) == 0) {
dump_all_conf(get_corpus(atoi((char*)words[1])), fd);
return;
} else if (wcsncmp(words[0], L"?", 1) == 0) {
sprintf(tmp, "%ls", words[2]);
dump_conf(get_corpus(atoi((char*)words[1])), fd, tmp);
return;
} else if (wcsncmp(words[0], L"~>", 2) == 0) {
corpus = get_corpus(atoi((char*)words[1]));
dump_dict_w(fd, words[2], 1, corpus);
return;
} else if (wcsncmp(words[0], L"~#>", 3) == 0) {
corpus = get_corpus(atoi((char*)words[1]));
dump_dict_n(fd, words[2], 1, corpus);
return;
} else if (wcsncmp(words[0], L"<~", 2) == 0) {
corpus = get_corpus(atoi((char*)words[1]));
dump_dict_w(fd, words[2], -1, corpus);
return;
} else if (wcsncmp(words[0], L"<#~", 3) == 0) {
corpus = get_corpus(atoi((char*)words[1]));
dump_dict_n(fd, words[2], -1, corpus);
return;
} else if (wcsncmp(words[0], L"<->", 3) == 0) {
direction = 1;
both = TRUE;
exact_match = FALSE;
corpus = get_corpus(atoi((char*)words[1]));
dump_conc(fd, corpus, direction, both, exact_match, words, i);
} else if (wcsncmp(words[0], L"<=>", 3) == 0) {
direction = 1;
both = TRUE;
exact_match = TRUE;
corpus = get_corpus(atoi((char*)words[1]));
dump_conc(fd, corpus, direction, both, exact_match, words, i);
} else if (wcsncmp(words[0], L"<-", 2) == 0) {
direction = -1;
both = FALSE;
exact_match = FALSE;
corpus = get_corpus(atoi((char*)words[1]));
dump_conc(fd, corpus, direction, both, exact_match, words, i);
} else if (wcsncmp(words[0], L"->", 2) == 0) {
direction = 1;
both = FALSE;
exact_match = FALSE;
corpus = get_corpus(atoi((char*)words[1]));
dump_conc(fd, corpus, direction, both, exact_match, words, i);
} else if (wcsncmp(words[0], L"<=", 2) == 0) {
direction = -1;
both = FALSE;
exact_match = TRUE;
corpus = get_corpus(atoi((char*)words[1]));
dump_conc(fd, corpus, direction, both, exact_match, words, i);
} else if (wcsncmp(words[0], L"=>", 2) == 0) {
direction = 1;
both = FALSE;
exact_match = TRUE;
corpus = get_corpus(atoi((char*)words[1]));
dump_conc(fd, corpus, direction, both, exact_match, words, i);
} else if (wcsncmp(words[0], L":>", 2) == 0) {
direction = 1;
corpus = get_corpus(atoi((char*)words[1]));
dump_ngrams(fd, corpus, direction, words, i );
} else if (wcsncmp(words[0], L"<:", 2) == 0) {
direction = -1;
corpus = get_corpus(atoi((char*)words[1]));
dump_ngrams(fd, corpus, direction, words, i );
} else if (wcsncmp(words[0], L"GET", 3) == 0) {
LOG("Playing http server");
play(fd);
return;
} else {
ERROR(fd);
}
}
