1. Intro

Phonetic labeling was done for internal purposes.  We are happy
to distribute the results of this effort, but this should be kept in mind.

The phonetic labeling procedure involved a first pass by a trained labeler;
generation of spectrograms with time-aligned phonetic labelers; inspection
of the spectrograms by an expert; and a final pass where boundaries were
moved or labels changed in line with the corrections.

Certain categories which we do not currently use in our classifier
(such [ao] as in caught) were nevertheless labeled for future use, but
not as consistently as possible.  For example, the [pv] or prevoicing
label, was often used before voiced stops when strong prevoicing
was present, but it was not always used.  Similarly, [epi], or epinthetic
closure, was sometimes but not always used.

The categories [ao], [aa] and [aw] will be re-checked for consistency
by R. Cole in the near future.  These and other corrections will be
announced when available.

We would appreciate feedback on the labeling.  Send email to
vincew@cse.ogi.edu.  

2. Symbols

The label set is very similar to that used by TIMIT (see the included
file timit.phonecode.doc from the TIMIT distribution CD).  The new
symbols are

-	A segment labeled "-" indicates that the previous segment was
	incomplete. This should only occur as the last segment in an
	utterance when was recording was terminated during speech.

pv	Prevoicing.

ao-r	Common in names, it is the sound in Portland and York.

bn	Background noise.

ns	Caller sounds which are not speech.

ls	Lip smack.

ln	Line noise.

br	Breath.

glot	Glottalization.

unk     Unknown.  This label was used when the speech could not be
	described using the label set; typically when a caller with a
	foreign accent produced sounds not used in English.

3. Procedure

Several different full and part time employees participated.  The
labeler could see the waveform and play selected regions. All
labels were scanned by a second person to look for obvious errors.
Even so, we are aware that inconsistencies remain.

Because of the time required, the redundant labeling used for word
transcriptions was not used for phonetic labeling.

4. Spelling

The phonetic labeling of spelled utterances was done some time prior
to the labeling of names and may not follow the conventions exactly. 

5. Resolution

For historical reasons, all labels were produced with a 3ms
resolution.  Since frame sizes used in practice are typically larger,
we feel this is acceptable, although using a finer resolution would
have been a better idea.