ISCA Archive VOQUAL 2003
ISCA Archive VOQUAL 2003

Voice characteristics of MARSEC speakers

Eric Keller

An examination of Bark scale spectra of some 30'000 vowel nuclei from the Marsec corpus showed significant voice quality differences for sex and speech style. Of the two, speech style appeared to be the stronger predictor in multiple correlation analyses. Clustering further documented the effects of speech style, with hierarchical clustering grouping voice profiles of similar-sounding styles. Sports, news and market reports were at one end of the clustering tree, while fiction, poetry and dialogues occupied other distinctive areas on the tree. Further, a factor analysis of the Bark scale spectrum showed four areas of relatively independent variation for the common speech frequencies involved in transmitting voice quality, indicating that considerable parameter reduction is possible in the prediction of voice quality-related spectral variation. The results suggest that high-quality spectrallybased speech synthesis systems could profit from a systematic biasing of spectral profiles to convey vocal quality, particularly in the lower frequencies (up to 600 Hz) and in the higher frequencies of speech (above 1600 Hz).

Cite as: Keller, E. (2003) Voice characteristics of MARSEC speakers. Proc. Voice Quality: Functions, Analysis and Synthesis (VOQUAL 2003), 97-102

  author={Eric Keller},
  title={{Voice characteristics of MARSEC speakers}},
  booktitle={Proc. Voice Quality: Functions, Analysis and Synthesis (VOQUAL 2003)},