=============================================================================
	README file for the example files bdg_TDNN.xxx
=============================================================================


Description:	Time-delay network (TDNN) for the recognition of the
============	phonemes 'b', 'd', 'g'

The file bdg_TDNN.net describes the time delay network discussed in 

	Connectionist Glue: Modular Design of Neural Speech Systems,
	Alex Waibel, CMU, in Proceedings of the 1988 Connectionist 
	Models Summer School, 1988

The net classifies the phonemes obtained by spelling the letters B, D,
and G from 16 melscale spectral coeficients. The input therefore does
not resemble the figures B, D, or G, but their accoustic
representations.

See the user manual for a detailed description of TDNNs and their usage.


Network-Files:	bdg_TDNN.net
==============

This network file contains a trained TDNN network for the phoneme
recognition task described above. The standard configuration file
for this network is bdg_TDNN.cfg (one 2D display only).


Pattern-Files:	bdg_TDNN.pat
==============	bdg_variable.pat

In the old SNNS pattern format exemplified by bdg_TDNN.pat, each
position shift of a fixed moving window over a time-varying input
signal resulted in a different imput pattern. This kind of pattern
file is given in bdg_TDNN.pat. For space reasons, only three sample
patterns are provided here.

In the new pattern file format with variable length input patterns one
can use the subpattern mechanism to extract subpatterns 'on the fly'
from one large pattern, e.g. from a time series pattern. This is
demonstrated with the file bdg_variable.pat
See the SNNS manual for details about using the subpattern mechanism.


=============================================================================
	End of README file
=============================================================================