Seminar at Speech, Music and Hearing:
Non-invasive automatic detection of phonatory state transitions
AbstractThe vocal folds can oscillate in several different ways, manifest to practitioners and clinicians as ‘registers’ or ‘mechanisms’, of which the two most commonly considered are modal voice and falsetto voice. Here we will refer to them as different phonatory states. State transitions are common in nonlinear oscillators; the transitions are usually abrupt and impossible to predict exactly. Switching state is much like switching to a different voice. Therefore, phonatory states are a source of confounding variation, for instance, when acquiring a voice range profile (VRP). In the quest for a state-aware, non-invasive VRP analysis, a method based on the electroglottographic signal (EGG) was developed to identify automatically phonatory state transitions of various kinds, including the modal/falsetto switch.
Following Herbst et al. (Science, vol 337 no 6094, 2012), the state-space loop representation of the high-pass filtered EGG signal is constructed, cycle by cycle. Then, a combination of Fourier descriptor analysis and entropy estimation is applied to the loop, yielding a new period-synchronous signal, which peaks when the attractor shape changes abruptly, but remains close to zero when the phonatory state is stable. A corpus was compiled of 50 EGG recordings of 1-4 known state transitions (‘register breaks’), in sustained phonation tasks similar to those used in VRP protocols. After some initial tuning, the algorithm located 96% of these breaks on the time axis. It also indicated numerous places where the phonatory state appeared to be changing in other ways, such as into or out of diplophonia. The next step will be to recognise different states automatically.
15:15 - 17:00
Tuesday June 11, 2013
The seminar is held in Fantum.
| Show complete seminar list