CSR 1995 Hub-4 Marketplace Training Epoch Data
 
		       NIST Speech Disc 26-1.1

			      March, 1996

This disc contains training epoch data for the ARPA CSR 1995 Hub-4
tests.  This data consists of 10 complete Marketplace radio
broadcasts, each of which are approximately 29 minutes long.  The
broadcasts were chosen so as to evenly represent different days of the
week across a training epoch (931112 - 940331).  The epoch was chosen
so as to match the training epoch in the 1994 CSRNAB1 data.

The recorded waveforms are stored in SPHERE-headered, files.  The
directory and file structure on this CD-ROM is as follows:

	csr95/h4/train/<BROADCAST-DATE>/<FILES>

Each broadcast date directory contains a waveform file
"<BROADCAST-DATE>.wav" for the broadcast as well as a
transcription file "<BROADCAST-DATE>.txt" and a speaker
information file "<BROADCAST-DATE>.spk".

The following broadcasts are included in this training epoch set:

        Date    Day-of-week
	------	-----------
	931117  Wednesday 
	931210  Friday
	931230  Thursday
	940112  Wednesday
	940124  Monday
	940204  Friday	
	940215  Tuesday
	940303  Thursday
	940315  Tuesday	
	940328  Monday

The file, "csr95/h4/doc/transpec.doc", contains a description of the
transcription convention used for this data.

Note that the transcription (.txt) files and speaker info (.spk) files
in the Hub-4 training data have not received the scrutiny that the
Hub-4 Development Test and Evaluation Test data have and may contain
some transcription errors.  Note also that unlike the test data
transcriptions, the training data transcriptions have only been
time-marked at the story level.