"Phoneme Segmentation for Unit Selection Synthesis"
John Kominek

As part of improved support for building unit selection voices, Festival
now includes two algorithms for automatic labeling of wavefile data. The
first method employs dynamic time warping to align a given wavefile
against a known reference. This usually requires having a synthesizer
already built for the target language -- a restriction to be averted if
possible. The second, more recent addition makes use of the HMM-based
acoustic modeling component from Sphinx-2. We have found that one
technique is not clearly superior to the other but that the error
characteristics are distinctly different. DTW is the more accurate
method in 60-70% of cases, but is also more prone to gross labeling
errors. Gross label errors are disastrous in a synthetic voice and need
to be corrected before an acceptable voice can be constructed. This talk
will illustrate these findings and indicate how a hybrid approach can
eliminate such outliers without compromising overall accuracy.