Aperiodicity analysis of filled pause in the Corpus of Spontaneous Japanese database

We applied a novel aperiodicity analysis method (Kawahara et al. 2016, SSW9) to CSJ database (Maekawa, 2003, ISCA & IEEE Workshop). The applied method derives the amount of aperiodicity as a time-frequency map, using three staged procedures. The first stage derives a probability map of the fundamental component witout prior information. The second stage finds the most probable trajectory of the fundamental component based on a state transition model and calculates the initial estimate of FO. The final stage refines this initial estimat using FO adaptive time axis warping and instantaneous frequency of each harmonic component. Recursive application of this refinemet procedure provides simultaneous estimation of FO and aperiodicity map. Classification of filled pause using the refined FO trajectories and the aperiodicity maps will be discussed. [This work was supported by JSPS KAKENHI Grant Number JP26284062.]