Precision of Phoneme Boundaries Derived using Hidden Markov
Models Phonemes (to the extent that they are real)
don't have sharp boundaries. But some are better defined
than others. Humans and machines generall agree which pairs
of phonemes have a sharp contrast that makes it easy to
draw a line. Perhaps the sharper boundaries are more
important for human language than the fuzzy ones?
Same/Different
in Speech. We investigate variability of speech by
developing a novel measurement of the distance between two
utterances. The idea is to train the algorithm so that it
responds only to the differences that are important to
humans. We then apply this to measure the difference
between repeated speech and more normal speech (phrases
read from a list).
Measuring
the position of accents may seem easy: just look for
the peak of fundamental frequency. But there are previously
unsuspected problems, caused by segmental effects. Some
vowels and consonants have a higher fundamental frequency
than others, and this can move peaks substantially. It can
even cause a simple unimodal distribution of accent
position to appear as two phonologically distinct places
where a peak can appear.
Report on
the Impact of Information Technology on Humanities
Research. Commissioned by the Arts and Humanities
Research Council, UK. (What will change in Humanities
research when we can search for audio and video clips? What
are the prospects for good search technologies? What about
automatic transcription and annotation of audiovisual
materials? How does copyright law and DRM impact humanities
research? This is the result of an 18-month study involving
extensive research into developing technologies and
interviews with potential users of upcoming technologies.)
[Local copy.]
What
Marks the Beat of Speech? In repetitive speech, like
poetry or chants, what acoustic properties mark the beat?
It turns out to be the loudness, relative to the
neighbouring syllables.
Evidence
for Attractors in English Intonation (A draft version
can be found here,
and a poster
here. We show that there are special intonation
patterns in the mind, and when you mimic speech, you end up
compromising between these special patterns and what you
hear. The debate about categorical perception of intonation
seems to have an intermediate answer: perception is not
entirely categorical, yet there are some discrete
structures in the perceptual process.)
