Loading presentation...

Present Remotely

Send the link below via email or IM


Present to your audience

Start remote presentation

  • Invited audience members will follow you as you navigate and present
  • People invited to a presentation do not need a Prezi account
  • This link expires 10 minutes after you close the presentation
  • A maximum of 30 users can follow your presentation
  • Learn more about this feature in our knowledge base article

Do you really want to delete this prezi?

Neither you, nor the coeditors you shared it with will be able to recover it again.


Stats Grp: Analysing Ultrasound Articulation Data

No description

Martin Corley

on 30 January 2014

Comments (0)

Please log in to add your comment.

Report abuse

Transcript of Stats Grp: Analysing Ultrasound Articulation Data

...and look at it a different way
Analysing Ultrasound Articulation Data
Articulation as a Group Measure
Dealing with Noisy Data
The Timecourse of Articulation
Prediction & Articulation
Speech Production in Psychology
measure time-to-respond, or transcribe production
typical research question
understand commonalities in production
typical research approach
(Ries et al., 2012)
(Lisker & Abramson, 1970)
articulation is an important source of evidence
psychologists' aim: to describe how articulation reflects differences between antecedent conditions for speech
Martin Corley
Elli Drake

Psychology, PPLS, University of Edinburgh
are people's articulations systematically affected by their predictions?
can't compare absolute articulations
but can compare within-ppt
difference between any two articulations
= mean euclidean distance between relevant
frames of ultrasound video
approx 270ms (8 frames) to vowel ending ("CV")
multidimensional scaling
analysing group behaviour
how different are
each individual's articulations
from their "control" articulations, in
each experimental condition
stronger effect of experimental manipulation should result in greater deviance from control articulations
saying "cap" when you predict "tap" might result in a "less /k/-like" onset
cf. McMillan & Corley (2010)
8 participants, approx 400ms (12 frames) to vowel onset ("C")
not very convincing
difference from subject-specific mean control /t/ or /k/
a potential problem
ultrasound is inherently noisy
but no easy way to determine how noisy
the MDS plots suggest that known differences are captured better for some participants
can we quantify that?
Quality Metric =
mean distance between different CV combinations
mean distance between matched CV combinations
metric is calculated from "raw" distances (MDS plots are for illustration)
exact calculation would depend on experiment ("some factor that can be measured independently of predictions")
could use quality metrics to exclude ppts,
but what would the criterion be?
instead, use metrics to
subsequent regression models
scale to geometric mean to avoid misrepresenting power
can we improve on "different"?
we can determine whether word onsets are affected by prediction
we can weight noisy evidence
but we can't say much about
obvious contribution to be made from detailed articulatory analyses
but multi-participant summaries tend to involve categorisation of responses
instead, take the delta technique...
subsequent frames of one participant's ultrasound
differences between frames represent movement over time
can create a "trajectory" showing the extent the articulators move in the process of articulation
difference from subject-specific mean control /t/ or /k/
weighted regression shows a reliable difference:
onsets differ 11 more units from the mean than do
these differences seem to be caused by an early tendency to move (more) in the
condition (as though an articulation is started and then abandoned?)
multi-participant ultrasound analyses are possible, and may provide insight into the relationship between what is said and how long it takes to start speaking

precision weights rather than sampling weights
(sometimes) an alternative to data transformation
biggest disadvantage: weights must be
here, we have to take the estimates on faith
weights effectively scale observations in the model fit
affects implied numbers of observations
(greater precision ~ reduced variance ~ larger n)
solution(?): keep th
implied n constant
Full transcript