Unsupervised Speech Morphing between Utterances of any Speakers

• Main Author: Hartmut R. Pfitzinger
• Other Authors: The Pennsylvania State University CiteSeerX Archives
• Format: Text
• Language: English
• Published: 2004
• Subjects: Arctic
• Online Access: http://citeseerx.ist.psu.edu/viewdoc/summary?doi=10.1.1.61.8736; http://www.phonetik.uni-muenchen.de/~hpt/pub/Pfitzinger_SST04.pdf

A new approach to speech morphing is presented which avoids the extraction of fundamental and formant frequencies as well as the detection of phone or syllable boundaries. All prominent spectral and temporal features of the source and target utterances are automatically related and interpolated. The method consists of three main parts: LPC-based source-filter decomposition, separate interpolation, and composition of the morphed speech signal. The paper focuses on the alignment and interpolation problems on three speech signal layers: the timing structure on a phone- and syllable-level, the shape of the frequency spectrum including formants and other spectral properties, and the micro-timing of the source signal. Particularly, the source signal alignment and interpolation is described since it is most crucial for the resulting quality of the modified speech signal. The new morphing procedure was applied to utterances taken from the freely available CMU ARCTIC speech corpus and assessed by a perceptual MOS experiment. Preliminary