|
Overview
The ability to operate computers using conversational speech
has been popularized in science fiction, and it offers to
provide powerful and economical tools for a number of applications,
including:
- Preventing repetitive use syndrome
- Air traffic control training
- Hands-busy environments (e.g., cockpits, factory floors,
laboratories, and operating rooms)
- Foreign language instruction
Although speech recognition research began nearly 50 years
ago at Bell
Labs, state-of-the-art systems have high computational
demands, low recognition accuracies under ideal conditions,
and very poor performance in noisy or stressful
real-world environments. Tanner's speech research has
targeted both computational efficiency and robustness with
respect to background noise and speaker variability.
The Tanner Approach
Tanner Labs has explored a variety of techniques for tackling
these problems. We have designed custom hardware, including
several analog ICs as well as a reconfigurable computer, to
provide compact systems that perform significantly faster
than a software implementation. (For example, as part of our
"AIM in hardware" project,
we developed a reconfigurable computer that included a custom
analog IC as a component.)
To achieve high performance in realistic environments, we
have studied and refined state-of-the-art algorithms, both
traditional techniques and biologically inspired neural networks
and cochlear models. Moreover, we have demonstrated that our
innovative architecture and
training system produce an isolated-word recognizer that is
more robust to noise and speaker variability than any commercially
available product. We are currently extending this technology
to recognize continuous speech. Another
approach that we are currently developing is an audiovisual
recognizer that uses lip reading
to recognize speech.
|
