|
MayaFlux 1.0.0
Digital-First Multimedia Processing Framework
|
How speech runs are found and measured. More...
#include <SpeechSegmenter.hpp>
Public Attributes | |
| uint32_t | max_run { 0 } |
| Longest speech interval in samples, 0 for no cap. | |
| bool | measure { false } |
| Also measure the voiced fraction and pitch of each speech interval. | |
| uint32_t | min_silence { 1600 } |
| Shortest silence that ends a speech run, in samples. | |
| uint32_t | min_speech { 4000 } |
| Shortest speech run kept, in samples. | |
| uint32_t | padding { 480 } |
| Samples added on each side of a speech run. | |
| double | threshold { 0.5 } |
| Speech probability at or above which speech starts. | |
| double | voicing_threshold { 0.5 } |
| Periodicity at or above which a window is voiced. | |
How speech runs are found and measured.
The scan follows Silero's get_speech_timestamps: a run starts at threshold, ends after min_silence without speech, and is widened by padding on both sides. Runs shorter than min_speech are dropped. max_run caps one speech interval in samples, 0 for no cap. All lengths are in samples at 16 kHz.
With measure on, each speech interval also gets a voiced fraction and a pitch from the waveform windows whose start lies in it. A window is voiced when its periodicity, divided by the share of pairs its lag can pair, is at least voicing_threshold. The threshold is a starting choice, not a measured one.
Definition at line 37 of file SpeechSegmenter.hpp.