Paul Carter (pgc104@york.ac.uk)
Wed, 26 May 1999 15:06:55 +0100 (BST)
Greetings.
Here are some draft minutes from yesterday. If you have any changes you
would like me to make, please let me know as soon as possible. I'll aim
to send out a decent copy on Friday.
Paul
======================================================================
Minutes of a ProSynth project meeting, UCL, 25.5.99
===================================================
Present: Jill, Mark, Jana, Sarah, Sebastian, John, Paul.
=======
Matters arising from 18/19.3.99 minutes
======================================
* JKL is working on rules for Procsy.
* Laureate: BT need to know details of which bits of Laureate we need, and
we need to sign the usual forms. Procsy is nearly capable of taking our
rules but we will still get Laureate. Laureate may be hard to tune to our
linguistic system. MH suggested that the best version to have might be
for SunOS (perhaps also Windows).
***ACTION***: JKL will explore further and obtain Laureate with the aim of
running it at each site.
* York to regroup test stimuli by listening. Ideas for new tests not yet
explored.
***ACTION***: York to continue this.
* Procsy has been sent to York
* Circulating non-ProSynth papers: some circulated, some not.
***ACTION***: Circulate papers if not already done so.
* MH now has an account on the Cambridge SGIs.
* CSL paper: submissions ongoing
* XML features issues: MH has corrected earlier files and added stop
releases
* Web sites: more updates needed
***ACTION***: all sites to send progress reports and links to papers to
PGC
* Collaboration with Keller & Zellner: will be on an informal basis during
their sabbatical in York.
Progress Reports
================
Cambridge
---------
(see document)
Feedback needed on high-level Procsy parameters: any improvements found to
be passed on to Cambridge.
***ACTION***: MH to set up Procsy at UCL
Procsy manual needs improving
***ACTION***: all to pass on suggestions for improving the manual to
Sebastian.
Cambridge are revising some of their labelling due to inconsistencies.
A warning: when using Procsy, be very careful about durations of
periodicity because of interactions between the parameters.
MH suggested that one way forward might be to import formant and other
control values into XML.
There was discussion about whether real values should be used or analysis
(eg x number of datapoints for vowel y).
***ACTION***: SH & SH to talk to JKL & RAO about locus equations and (less
urgently) about the possibility of getting YorkTalk values.
York
----
(see document)
Sarah suggested subdividing the data by context (lingual versus labial).
JKL suggested separating unstressed vowels in function/nonfunction words.
MH confirmed that the issues of multiple matching and empty structures
should be solved by the use of ProXML. The prx interpreter has a switch
'-n' which results in no XML output: using "output", durations can then be
printed directly to stdout.
**ACTION***: MH to give PGC examples of how to use ProXML in this way; PGC
to feed back any problems.
MH asked that ambisyllabic consonants should have identifiers to
distinguish the first of a pair from the second (other than being in coda,
onset positions)
An aside: JKL asked MH about the possibility of obtaining means of OQ
measurements from the database. This would also be very useful for
Cambridge. Cambridge would like the following things:
* mean OQ from the database, including all voiced portions (except inside
stops), or maybe just vowels. Possibly split by stress.
* ultimately, an average time-aligned LX waveform in a window of 8 periods
around obstruent boundaries, split by voicing of the obstruent.
Ideally, Cambridge would like a tool to enable them to specify this sort
of request.
***ACTION***: in the short term, MH to put LX waveforms on the ftp site.
Sarah Palmer's work on LX and boundaries (including ICSLP 92) is germane
to this investigation.
***ACTION***: JH to hunt out relevant papers.
UCL - Mark
----------
(see document)
point 2
***ACTION***: JH to tell MH how to mark AG boundaries etc.
point 3
newattribute() suppresses errors but doesn't update the DTD. This seems
to be a problem with the Edinburgh tools.
***ACTION***: MH to take this up with Edinburgh.
UCL - Jana
----------
(no document)
A finer classification of ONS than had been used in the ICPhS paper was
examined with increased success of the model. Using onset duration to
predict the peak was less successful. Relating peak onset time (PON) and
peak offset time (POF) to rhyme was not as successful as relating them to
the syllable. Vowel type seemed not to matter.
However, expressing PON and POF as a proportion of the foot eliminated the
rightward shift previously observed when the foot had a tail. There was
still some sensitivity to the identity of the onset.
NB So far, only short feet have been examined.
JKL suggested eventually using CART techniques.
Agenda Item 1.1
===============
We are aiming to have something running on a PC in time for San Francisco
(Cambridge will be taking a laptop).
One possibility is to have PSOLA-manipulated waveforms of Mark's speech
(especially using f0 and timing information) to be produced from text from
a section / extension of the database. This might demonstrate that f0 and
timing are not making naturalness worse (for example, timing could be
compared to Festival timing). "Text in, synthesis out" to demonstrate
that there is a model.
The test set would need to be small, eg 5 sentences each of 5 structures.
Device-independence could be demonstrated.
***ACTION***: York to produce a ProXML script for durations; UCL to
produce one for f0.
Also a demo of Procsy on a very small set of data. Deadline for this:
mid-June; others: end of June.
***ACTION***: York to provide Cambridge with existing database files with
York timings.
Agenda Item 1.2
===============
How & where to express PON, POF and LON in structure. PON is sensitive to
onset type; POF is also sensitive to coda. JKL suggested percolating
headed information. For the moment, FOOT seems to be the best level
because of least variation, though other levels may be needed for POF.
For now, will use the computationally easy solution. More is likely to
fall out with further analysis.
***ACTION***: Jana to send data for peak and minimum f0 to Sebastian who
will look at it with respect to peak/minimum subglottal pressure.
***ACTION***: Jana to send a copy of her document to Cambridge and York.
***ACTION***: JKL to contact York Computing Service to see if York has a
PC site-licence for non-supported PCs to run SPSS (if so, give Cambridge
access).
Agenda Item 1.3
===============
The 'meat' of the CSL paper should contain at least summaries of: the
results from the perceptual testing, the (preliminary) intonation model
(ICPhS paper plus results presented at this meeting), (preliminary) timing
model.
This is a "position paper with preliminary results".
Agenda Item 1.4
===============
No further perceptual testing on York material before San Francisco.
Study of resonance will be limited to a production study on Mark, probably
using centre of gravity as a measure.
***ACTION***: Cambridge to complete RT experiment for UCL data, getting up
to 30 subjects.
Agenda Item 1.5
===============
See York list previously circulated. Items may be added based on Ali's
findings.
***ACTION***: All sites to check this list and add in other items they
might want. Circulate by 4.6.99.
***ACTION***: UCL to check there is an appropriate representation of
segments (particularly in stressed syllables).
Agenda Item 1.6
===============
No objections in principle to RAs going part-time. JKL needs to discuss
the issue with RAO. UCL will be looking for a replacement for Jana,
checking regulations re the need to advertise. An approach would need to
be made to EPSRC. Sebastian needs to work out what level of
part-timedness would suit him.
Agenda Item 2
=============
Hyperspeech, extending range of prosodic structures, function words (as
previously discussed).
***ACTION***: all to flesh out minute 8 from 19.3.99 to reconstruct what
we want for early September.
Date of next meeting
====================
5th (or possibly 6th) July, 1999, by consultation with RAO. Sarah to
check these dates are OK for her.
ENDS
This archive was generated by hypermail 2.0b3 on Wed May 26 1999 - 15:08:13 BST