|
MayaFlux 1.0.0
Digital-First Multimedia Processing Framework
|
One phoneme the network recognised. More...
#include <PhonemeAnalyzer.hpp>
Public Attributes | |
| size_t | end { 0 } |
| Estimated sample one past the phoneme's last. | |
| size_t | frame { 0 } |
| Frame it was recognised on. | |
| uint32_t | phoneme { 0 } |
| Class of the phoneme, 0 to HubertPhonemeModel::phonemes - 1. | |
| float | probability { 0.0F } |
| Probability the network gave the phoneme on that frame. | |
| size_t | sample { 0 } |
| First sample of that frame. | |
| size_t | start { 0 } |
| Estimated first sample of the phoneme. | |
One phoneme the network recognised.
start and end are estimates. The network marks a phoneme on one frame, and the boundary between two neighbours is put where the weak evidence on the frames between them changes sides. Measured on 60 LibriSpeech utterances against a forced alignment, the boundary between two correctly recognised neighbours fell a median of 12.5 ms off, with 94% within 40 ms. Pauses are not detected: a phoneme before a pause runs through it, the first starts at sample 0 and the last ends with the utterance.
Definition at line 70 of file PhonemeAnalyzer.hpp.