[Colloquium] Reminder: Stoehr/MS Presentation/May 10, 2013

[Margaret Jaffey]

This is a reminder about Mark's MS Presentation tomorrow.

------------------------------------------------------------------------------
Date:  Friday, May 10, 2013

Time:  12:00 PM

Place:  Ryerson 277

M.S. Candidate:  Mark Stoehr

M.S. Paper Title: Binary-Feature Detection Cascades for Speech
Recognition

Abstract:
Speech recognition is generally performed with Hidden Markov models,
which require a language model and an enormous training set to obtain
good performance. Moreover, HMMs poorly model phonetic trajectories
and detail. Template-based approaches are a promising alternative to
HMMs that require less training data, but can potentially better model
phonetic dynamics. In this work, we investigate the performance of a
template-based system at detecting short sequences of phones embedded
in continuous speech.

Template-based systems compare reference templates of the phone
sequence directly against the speech signal. A detection is declared
whenever a template closely matches a segment of the signal. We show
how template-matching can be performed using a cascade of standard
classification algorithms. The cascade has three steps: a
likelihood-ratio test that identifies a set of potential matches, a
clustering algorithm to reduces this set of matches, then an SVM
classifier that discards low-quality matches. Training the likelihood
ratio detector involves estimating a mixture model over segmented
examples of the target sequence, while training the SVM detector
requires speech that is known to not contain the target sequence but
is otherwise unlabeled.

We test this architecture on a variety of different feature sets
including raw-spectral data, binary edge features, and data-adapted
patch-based features. The binary edge features are computed on the
speech spectrogram producing a binary time-frequency representation.
The patch-based features are constructed using mixture models over
patches of the edge-based binary time-frequency representation. Other
feature sets are considered and we compare the detector performance
obtained for each choice of features

Our results indicate that a cascaded detector combining a likelihood
ratio test and an SVM classifier achieves better performance than
either detector alone. We also find that binary edge features and
patch features improve robustness to speaker variation and noise.

Mark's advisor is Prof. Yali Amit

Login to the Computer Science Department website for details:
 https://www.cs.uchicago.edu/phd/ms_announcements#stoehr

=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=
Margaret P. Jaffey            margaret at cs.uchicago.edu
Department of Computer Science
Student Support Rep (Ry 156)               (773) 702-6011
The University of Chicago      http://www.cs.uchicago.edu
=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=