Having promised not to be able to do the minutes quickly, I did them on the
train home yesterday. Enclosed, in text format and w6/win95 formats. PLease
let me have any amendments.
Richard
Minutes of the ProSynth meeting,
UCL, 15/02/00
Present: Jill, Rachael (for the morning), Mark, Sarah, Sebastian, Richard, Paul, John, Jana (from lunchtime).
1. Progress.
1.1 Cambridge
See the written report.
The pretests will start later this week, starting wtih York's material.
Sebastian and Sarah talked about the ways that the resonance effects spread: they are extremely complex and there still isn't a clear story to tell.
ACTION: the words 'yes' and 'no' which are used for the experiments with higher cognitive load should be replaced with a tick and a cross.
1.2 UCL
See written report from Rachael.
There are two turning points in each foot: where they are depends on the number of syllables within the foot. There are some problems that are shared with the temporal model, mostly around how to join feet.
We discussed microprosodic effects: these might come from free in HLsyn, which does intrinsic f0 unless you turn that feature off.
It seems that none of us knows how much data is enough.
Mark reported orally on the work he has done. The association between foot and AG has been broken; ` is used to mark both a new foot and a new AG, while \ is used to mark just a new foot. The new rules for f0 are currently in the wrong domain: uttmbrol needs changing. Log and power have been added to XML. The unix version of ProSynth has not yet been updated; York needs this. The webpages have been cleaned up.
ACTION: Mark will update ProSynth into unix and sort out a problem with a header file.
1.3 York
See written report from Paul.
There are problems with the current timing model(s), and a number of solutions are possible. We know that the data sent to Cambridge for the pretests is problematic; this is because there is no proper model of consonants at the moment.
Data for the pre-tests will be changed so that we compare a model with no CV information against one based on z-scores for consonants.
We have demonstrated that it is possible to generate rhythm with a rather small database.
It might be possible to compare Laureate origninals and MBROLA synthesis by using dynamic time-warping.
JKL reported that Rob Fletcher has just developed a unix version of HLsyn; he and Rob are just checking the input and output to the program. It should be possible to pipe the output to PROCSY.
ACTION: York to send new material for pretests to Cambridge.
Sarah asked us to note that there is now a formal connection between Cambridge University and MIT. This means that Sarah and Ken Stevens are now officially colleagues. We should use this network. Eventually there will be money in the link. Information about the link ('CMI') is available on the web, from a link at www.cam.ac.uk.
2. CSL paper
The paper has been accepted and revisions made along the lines suggested by the reviewers. Richard thanked everyone for making the job of getting the final corrections so easy. There was just one figure left to do, which Sarah is supplying. Sarah will deliver a hard copy of the paper to the editor on Wednesday and send copies to the journal's offices.
CSL have also asked us to supply audio files for the website. Material that could be supplied immediately is the stuff we already have for last year's perceptual tests. We could also put up more up-to-date material.
ACTION: € Richard will check with Steve Young what the practicalities are for supplying audio files and let everyone know. € Mark will make an html page for CSL if necessary. € Jana and Paul will send data and descriptions to Sebastian and Mark. Mark will update our website with the material where appropriate.
3. Testing
3.1 Pretests
The pretests for UCL were the ones we discussed most. We decided the following:
€ use MBROLA + York's timings
€ increase the number of test sentences
€ ask for true/false judgements
€ use dummy sentences as well as ones we want, in order to compound predictability and focus subjects' attention on the end of the utterance. eg. snow is white/black/cold (and test just one of these)
€ a within-groups design might be possible this way (ie. each subject hears both 'right' and 'wrong' versions of sentences)
€ have the fillers spoken 2-3 times 0% false
€ about 50% of fillers should be true, 5
€ words in dummy sentences need to be phonetically similar to their congenors in about 50% of cases eg. we often see stars at night/nine; a dog is a kind of animal/animus
Within-group designs are not much liked by psychologists; we might prefer to do this experiment with a between-groups design.
We need to check the amount of repetition in the pre-test sentences and replace ones where 'unusual' words are repeated so as to eliminate priming effects.
ACTION: York to check their pre-test sentences for repetition of words.
3.2 Factorial
Sebastian has a method for getting York's timings into PROCSY.
ACTION: JKL and Sebastian to talk more about this so Rob can incorporate it into HLsyn.
We had a rather depressing discussion about the factorial experiments. We do not know whether EPSRC are willing to use some excess money from Cambridge's grant to continue the project beyhond March. The result is that we are not sure whether we will be able to carry out the factorial experiment. We are very clear that we do not want to run the tests unless we can do it well. We decided to leave any further decisions about this till Monday 21/2/00.
4. Webpages and final report.
Mark has produced a website which brings together our current material in a more systematic and accessible fashion. Updates are needed for the webpages from each site. The material for the webpages could form the bare bones of the final report.
ACTION:
€ Jill to write a paragraph about UCL's 'aims'.
€ All to sort out example files
€ Internal reports are needed on all aspects of ProSynth (with exceptions on a couple of things from Cambridge, which just need updating); York needs to produce a report on the timing model.
€ Everyone should consider what conclusions and implications there might be from the project as a whole.
This archive was generated by hypermail 2b29 : Wed Feb 16 2000 - 10:30:19 GMT