Pronunciation Group Meeting 16.5.97 (Roni Rosenfeld)

PRESENT: Mike Riley, Bill Byrne, Sanjeev Khudanpur, Michael Finke,
Murat Saraclar, George Zavaliagkos, Takayuki Arai, Harriet Nock, Chuck
Wooters, Roni Rosenfeld.


BEFORE-WORKSHOP TASK LIST:
~~~~~~~~~~~~~~~~~~~~~~~~~~

1. TIMIT pronunciations (Chuck, DONE!)
2. ICSI normalization (ICSI, unknown date, optional)
3. Alignment of baseforms (from ICSI, later from TIMIT) (Murat, 2-3 days)
4. VTL (JHU, 1 week to code, starting now;  ~1 day to run)
5. Building trees (Murat)
6. Network compilation (Mike, 1 week to code, <1 day to run)
7. Build acoustic models (JHU, 2 weeks each run)
8. Build test lattices (Harriet)
9. Build pronunciation networks for acoustic training data
	 (Mike/JHU, ~1 day to run)


DEPENDENCIES: 
~~~~~~~~~~~~~
(See Mike's block diagram in the LVCSR proceedings)

For Pronunciation model construction: 1,2, then 3, then 5.
For Pronunciation network construction: 6.
For test-set recognition: 8.
For Acoustic model generation: 4,9, then 7.


OTHER PLANS/NEEDS:
~~~~~~~~~~~~~~~~~~

Roni:	Needs only #1,2,3.  Currently working on fast ME algorithm - plans
	to apply it to reducing perplexity of phone string.
Chuck:	Needs no new tools.  Plans to do error analysis using ICSI
	transcriptions.
Harriet:Work on test lattices.
George:	Plans to look at the data and induce word-specific trees, then
	try to generalize to the whole vocab.
Murat:	Works on ME/feature selection for phone string prediction.
	Currently using Ristad's toolkit, maybe develop interface to FSM.
Michael:Work on continuous pronunciation modeling (of common words?)
Sanjeev:Work on ME with Murat.  Look for linguistic cues within the tree,
	e.g. word freq.  Eventually convert to FST
Bill:   ??? [I was away. please fill in]
Mike:   ??? [I was away. please fill in]


NEXT MEETING:
~~~~~~~~~~~~~
There will be a June 23 (Monday) meeting at CLSP to start the
pre-workshop training.


DELIVERABLES:
~~~~~~~~~~~~~
Mike et. al. gave a FST course at Penn.  Mike will send around a
laundary list of course material, then collect specific requests.