Summary: may have led to a poor allocation of network resources during
training.
5. CONCLUSIONS
We have presented preliminary results for two methods of adding
five distinctive speech features (sonorant, fricative, nasal,
vocalic, and voiced) to our system. We showed that locally inte­
grating knowledge about distinctive speech features into an MLP
by training specific hidden units to recognize specific features
can improve performance a small amount, but that allocating net­
work resources for the same purpose in a distributed fashion
using the architecture described above does not appreciably
improve performance. These results show a small performance
improvement; however, ``oracle'' experiments with improved
feature detectors have shown that further improvements in fea­
ture classifiers that can be incorporated into specific hidden units
are likely to lead to substantial performance improvements.
ACKNOWLEDGMENTS
Part of this work was supported by DARPA Contract MDA904­
90­C­5253 and part by a contract with NTT Data Communica­
tions Systems Corporation.

Source: Abrash, Victor - Speech Technology & Research Laboratory, SRI International

Collections: Computer Technologies and Information Sciences