8th International Conference on Spoken Language Processing

Jeju Island, Korea
October 4-8, 2004

A Forensically-Motivated Tool for Selecting Cepstrally-Consistent Steady-States from Non-Contemporaneous Vowel Utterances

Michael Barlow (1), Mehrdad Khodai-Joopari (1), Frantz Clermont (2)

(1) University of New South Wales (ADFA Campus), Australia
(2) American University of Paris, France

We describe a forensically-motivated, semi-automatic tool, which yields steady-state locations and cepstral parameters for contemporaneous and non-contemporaneous recordings of the five vowels in spoken Japanese. Using the notion of spectral prototype obtained from the mean cepstrum of a vowel's high-energy interval, coupled with the peak-sensitivity property of the index-weighted cepstral distance, the tool is able to find steady-state intervals that are the least-phonetically deviant from the prototype. In addition to the consistency in steady-state location afforded by this approach, non-contemporaneity is taken into account by seeking the minimum deviation across all recordings. The overall design of the tool draws its efficiency from the interactive ability to quickly alter settings and visualize intermediate results in the time and frequency domains.

Full Paper

Bibliographic reference.  Barlow, Michael / Khodai-Joopari, Mehrdad / Clermont, Frantz (2004): "A forensically-motivated tool for selecting cepstrally-consistent steady-states from non-contemporaneous vowel utterances", In INTERSPEECH-2004, 2393-2396.