|
Articulatory gestures for speech recognition
Using information about the movements people make when producing speech,
this kind of approach to automatic speech recognition promises to provide
improvements in accuracy over conventional systems, particularly for fluent or
conversational speech or in the presence of noise.
|
|
Time-frequency analysis
|
It is typical to perform some sort of short-term frequency analysis of
time-varying signals, but there are high-resolution transformations and
adaptive filtering methods that can give better definition of certain
features in an audio signal.
One application of this is in
speech coding.
|
|
|
|
Audio classification
Classification is an important task when it comes to information retrieval,
or trying to index a lot of material, e.g., from TV broadcasts.
Research in the group has sought to identify certain characteristics or
"events" from the audio stream with a view to enhancing the
amount of meaningful data that can be used for tagging.
|
Periodic-aperiodic decomposition of speech signals
Many speech sounds contain contributions both from voicing in the larynx and from
air-turbulence noise.
We use a signal processing technique, called the pitch-scaled harmonic filter,
to try to separate these two components so that they can each be analysed
separately.
Such acoustic scene analysis has also been shown to help in other applications
of speech processing.
|
|
|
