Caca

Abstract
This paper addresses the problem of voice activity detection
for distant-talking speech recognition in noisy and reverberant
environment. The proposed algorithm is based on the sameCross-power Spectrum Phase analysis that is used for talker location
and tracking purposes. A normalized feature is derived,
which is shown to be more effective than an energy-based one.
The algorithmexploits that feature by dynamically updating the
threshold as a non-linear average value computed during the
preceding pause. Given a real multichannel database, recorded
with the speaker at 2.5meter distance from the microphones,
experiments show that the proposed algorithm provides a relevant
relative error rate reduction.
1. Introduction
Robust multi-sensory based interfaces representa key issue for
the development of perceptual systems able to understand a
given audio-visual scenario. Future development of the related
technologies is envisaged for applications in contexts asmeetings,
video-conferencing, interviews, etc.
This study is being conducted under the project PEACH
(Personal Experience Active Cultural Heritage)1, whose objective
is to develop and experimentadvanced technologies that
can enhance cultural heritage fruition in a museum. Among
many other objectives, the project refers to the realization of
multisensory analysis for tracking a singlevisitor and recognizing
distant-talking speech utterances.
Distant-talking ASR in a noisy and reverberant environment
represents a difficult problem. In a realistic situation, one
has to take intoaccount several factors, generally neglected in
a close-talking interaction scenario as, for instance, the fact that
talker’s position and his/her head orientation may be unknown
and time-varying in anunpredictable fashion. The trend is that
of approaching the ”disappearing computing” concept, but this
may also cause a speaking style sometimes difficult to characterize,
rich of spontaneous...