Speaker: Paul Smith
Title: Analog Architectures for Auditory Feature Extraction and
Recognition
Date: January 30, 2004
Time: 3:00 pm
Location: GCATT Room 325
Abstract:
Speech recognition systems have been implemented using a wide range of
signal processing techniques including Neuromorphic/biologically
inspired and Digital Signal Processing techniques.
Neuromorphic/biologically inspired techniques such as silicon cochlea
models are based on highly parallel computation and computational units.
While DSP techniques are based on block transforms and statistical or
error minimization methods that
occur serially.  Over the past years DSP has grown into a very popular
topic.  Applications using DSPs have also grown fueled by the increase
in computing power and decrease in cost, which has been a very positive
step, particularly in the area of speech recognition.

This research combines the areas of classical Digital Signal Processing
and Neuromorphic/biologically inspired techniques applied to speech
recognition.  Our system implements a continuous-time cepstrum as the
feature extraction block.  These features are then processed by a VQ and
HMM. This approach is used because our architecture operates in the
continuous-time domain, lending itself well to Neuromorphic and parallel
architectures, however the feature vectors are discrete pre-determined
vectors, which simplifies the signal flow to later processing blocks.
These pre-determined vectors are used to train and build our pattern
recognition block, which is based on Hidden-Markov Model decoding.



Biography:

http://users.ece.gatech.edu/pds

Slides: