14th August
Perception of Lexical Pitch Accents (“limited tones”) by Serbian and French listeners: New Version of Free Classification Task
1. The Old Task
The free classification task has been used to describe explore how listeners might classify accents, dialects, and speech categories based on their similarity, without imposing any labels upon them. The task was first developed by Imai and Gardner (1956), and it was later re-imagined for the research on regional dialects (e.g., where the speaker comes from) by Clopper (2008) and Clopper and Pisoni (2007). The task requires the participants to listen to a number of different stimuli arbitrarily labelled, that is, labelled without any meaning (e.g., AA, BB, CC, etc.), and place the similar stimuli in the same group without having experimenter-imposed categories. The participants are given a power point presentation with the spreadsheet inside of it (Figure 1 below). Then, they asked to drag the recordings into the spreadsheet and, without having to worry about where in the spreadsheet they place the recordings, they should place similar stimuli into the same group. How do they decide which ones are similar? Well, by listening to the stimuli as many times as they want to.
Figure 1. The free classification task set-up taken from Atagi and Bent, 2016.
2. The New Task
Having to switch all our experiments online, we (my supervisor, Dr. Stephen Winters and I) could not send out the ppt to our participants and have them do the task. We wanted to make it easier and more interactive for the participants. Therefore, we found a wonderful resource, which we adapted for our purposes. The resource in question is here (Donhauser & Klein, 2023). We used this toolbox to set up our own free classification task, “the new task”. The task includes a big circle and eight small circles around it. Each of the small circles is a recording, which in our case represents a nonword produced by a male or a female talker with a specific Serbian lexical pitch accent. The participants are asked to drag each smaller circle into the big circle and position them based on similarity — the more similar they are, the closer they should be. The lines are used to help participants visualise where a circle being dragged is in comparison to the rest of the circles. (Figure 2).
Figure 2. The new version of the free classification task adapted from Donhauser & Klein, 2023.
3. Data Collection
From this free classification task, we receive “raw” coordinates of where in the circular space the circles were placed in a 0.5 by -0.5 grid, whereby the maximum distance between the two tokens would then be 1 (Figure 3 is merely a visual representation of the received coordinates in a non-circular environment).
Figure 3. Coordinates obtained from the task.
The following python code is applied:
python audio_tokens.py –infile example_results.csv –outfile converted/example_results
And the coordinates are converted into similarity matrices. 0 means most dissimilar, 1 means most similar. The matrices are saved in the .csv format and they look like this:
Item 1 Item 2 Item 3
Item 1 1 0.4 0.6
Item 2 1 0.7
Item 3 1
4. Data Analysis
There are different ways to analyze the data. We chose the Multidimensional Scaling Analysis (MDS) because the analysis reduces the number of dimensions that exist in reality, and help us visualise the distances between the tokens. That is, with MDS, we are able to re-create the so-called “participants’ perceptual space”, which otherwise, with different non-visualization techniques, we would not be able to. The MDS analysis first requires one to determine the number of dimensions most optimal for the model. To do so, we followed Levshina’s (2015) procedure on the “elbow” technique. For the majority of the data, the most optimal number of dimension was 3 – the point where the distribution is not drastically changed (Figure 4). The received MDS figures are represented in Figure 5.
Figure 4. Number of dimensions determined with the “elbow” technique.
Figure 5. An example of the output of the MDS analysis.
5. The Code
The entire experiment was run in Cognition.run platform. You can read more about this platform on their website, but you should know that it is a very neat and user-friendly platform for running experiments in jsPsych. Therefore, the entire code was written in jsPsych, except for the Donhauser and Klein’s toolbox. The Data Analysis was performed in RStudio. The entire jsPsych and RStudio code are available in the OSF platform, on my account, which you can access any time because we are all about open science!