MayaFlux 1.0.0
Digital-First Multimedia Processing Framework
Loading...
Searching...
No Matches

◆ bound_events()

template<ComputeData InputType = std::vector<Kakshya::DataVariant>, ComputeData OutputType = Eigen::MatrixXd>
void MayaFlux::Yantra::PhonemeAnalyzer< InputType, OutputType >::bound_events ( PhonemeAnalysis & analysis,
size_t samples )
inlinestaticprivate

Gives every event its sample range.

The first frame of the later of two neighbours is the one, after the earlier spike and up to the later one, that maximises the log probability of the earlier phoneme on the frames before it plus that of the later phoneme from it on, probabilities taken among the phonemes only.

Definition at line 467 of file PhonemeAnalyzer.hpp.

468 {
469 constexpr size_t classes = HubertPhonemeModel::classes;
470 constexpr size_t phonemes = HubertPhonemeModel::phonemes;
472 auto& events = analysis.events;
473 if (events.empty()) {
474 return;
475 }
476 const auto evidence = [&](size_t frame, uint32_t phoneme) {
477 const float* row = analysis.posteriors.data() + frame * classes;
478 double mass = 0.0;
479 for (size_t index = 0; index < phonemes; ++index) {
480 mass += row[index];
481 }
482 return mass > 0.0 ? std::log(row[phoneme] / mass + 1.0e-30) : 0.0;
483 };
484
485 events.front().start = 0;
486 for (size_t k = 0; k + 1 < events.size(); ++k) {
487 const size_t first = events[k].frame;
488 const size_t last = events[k + 1].frame;
489 const uint32_t earlier = events[k].phoneme;
490 const uint32_t later = events[k + 1].phoneme;
491
492 double score = 0.0;
493 for (size_t frame = first + 1; frame <= last; ++frame) {
495 }
496 double best = score;
497 size_t split = first + 1;
498 for (size_t candidate = first + 2; candidate <= last; ++candidate) {
500 if (score > best) {
501 best = score;
503 }
504 }
506 events[k].end = boundary;
507 events[k + 1].start = boundary;
508 }
509 events.back().end = samples;
510 }
Which phoneme is spoken near each frame of a mono 16 kHz utterance, from a learned network.

Referenced by MayaFlux::Yantra::PhonemeAnalyzer< std::vector< Kakshya::DataVariant >, Eigen::MatrixXd >::analyze_implementation().

Here is the caller graph for this function: