Bachelor's Thesis

AsTRiQue is described in detail in my bachelor's thesis (Institute of Phonetics, Faculty of Arts, Charles University). You can download the thesis here.

MEDAL Poster Presentation

AsTRiQue was presented as a flash talk and poster at the MEDAL Summer School of Computational Modeling in Birmingham on June 24, 2025. You can download the poster here.

Reference

[1] Kocjančič, T., & Bořil, T. (2025). Voicing in Czech children’s sibilants: children’s productions and adult’s perception, International Clinical Phonetics and Linguistics Association (ICPLA), June 24-27, 2025, Patras.

AsTRiQue

ASk The RIght QUEstions: An active machine learning framework for perception experiments

AsTRiQue is a tool designed to streamline perception experiments that involve large amounts of stimuli. In traditional perception experiments, participants often have to classify a large number of items, which can be time-consuming and tiring. AsTRiQue minimizes this burden through active learning; a kind of machine learning where a model is continuously updated based on the participant’s input to decide what to ask next, following in the footsteps of Einfeldt et al. (2024).

The experiment starts by selecting a few stimuli for the participant ("oracle") to classify (using a grid-based sampling strategy). Based on the answers, a logistic regression model is trained to predict how the participant might respond to all the remaining stimuli. Then, the model picks the stimulus it is most uncertain about (uncertainty sampling). By targeting these hard-to-predict cases, the model learns much faster.

To keep things balanced, there is also an option to occasionally include a very easy (high-certainty) stimulus as a “cleanser” to reduce participant fatigue from repetitive difficult stimuli.

This process repeats: the participant answers, the model updates, and a new uncertain stimulus is selected. Once the model is confident enough in its predictions of all the stimuli the participant hasn't been presented with (based on a pre-determined confidence value), the experiment ends.

To test and evaluate this model, there is an option to use a lookup table as the oracle (as a so-called virtual agent) instead of a human participant.

ℹ️ Example Chart

In this example, a virtual agent simulated participant 3 from Kocjančič and Bořil (2025) with the following config in AsTRiQue:

python STRATIFIED_SAMPLING_RESOLUTION = 3 # resolution of the n by n grid from which samples are taken MIN_ITERATIONS = 30 # minimum number of iterations CLEANSER_FREQUENCY = 0 # insert a high-certainty sample every nth iteration to prevent participant fatigue (irrelevant for virtual agents); 0 to disable MODEL_CERTAINTY_CUTOFF = 0.95 # stopping certainty threshold PARTICIPANT_TO_MODEL = 'p03' # participant ID to simulate

The oracle (virtual agent) classified 57 stimuli out of 104. The remaining 47 were classified by the model with a 97.9% accuracy. This specific case would, with a human participant, translate to reducing the participant workload by about 45% while still producing data that is nearly as reliable as if the participant had answered every item themselves—with an effective overall accuracy of over 99% (including the participant's answers).

Of course, this is an exceptionally good scenario; the model's efficiency can be impacted by the source data and especially by the participant's predictability (the less consistent the participant is in in their answers, the more difficult it is to model them in the first place).

🧑🏾 Human Workflow Showcase

If you'd like to see AsTRiQue in action with yourself as the participant, you can run the participant.ipynb notebook locally or check out the showcase notebook online.

In the notebook, you'll be asked to classify all 104 stimuli. The model will stop training after the confidence threshold is reached, the rest of your classifications will be used to evaluate the model's performance.

🤖 Virtual Agent Showcase

Not feeling like doing the perception experiment yourself? You can still see AsTRiQue in action by running the virtual_agent.ipynb notebook locally or checking out the virtual agent showcase notebook online. Instead of querying you as the oracle, the model will query a lookup table of a real participant who took part in the original experiment.

📊 Dataset

The showcase makes use of data from Kocjančič and Bořil (2025), where they investigated categorization of Czech sibilants /s/ vs. /z/ and /ʃ/ vs. /ʒ/ as a function of two acoustic parameters: voicing (quantified as the percentage of the segment exhibiting periodic vocal fold vibration) and segmental duration (in ms). For the purposes of the showcase, /s/ and /ʃ/ were batched together, as were /z/ and /ʒ/.

🗂️ Data Structure

📁 data/data.csv

This spreadsheet contains the filenames of all recordings used in the experiment, as well as their parameters (voicing and duration).

| filename | voicing | duration | | -------------------- | ------- | -------- | | NJ8sC0File017.wav | 0 | 72 | | NJ1zV8File036.wav | 8 | 127 | | ... | ... | ... |

📁 data/participants/p01.csv

This spreadsheet contains participant 1's answers in the real experiment. | filename | answerbatch | | -------------------- | ------------ | | NJ8sC0File017.wav | s | | ... | ... |

⚙️ Config & Technical Stuff

The model takes two numbers (e.g. voicing and duration) as predictors for a binary classification (e.g. batched /s/ /ʃ/ vs. batched /z/ /ʒ/)

```python

data structure config

PREDICTOR1 = 'voicing' # first predictor column name PREDICTOR2 = 'duration' # second predictor column name TARGET = 'answer' # target column name FILENAMECOL = 'filename' # filename column name LABELMAPPING = {'s': 0, 'z': 1} # binary output label mapping DATAPATH = 'data/data.csv' # sound info data file path PARTICIPANTCSV_DIR = 'data/participants' # participant CSV directory

model parameters

STRATIFIEDSAMPLINGRESOLUTION = 3 # resolution of the n by n grid from which samples are taken MINITERATIONS = 20 # minimum number of iterations MODELCERTAINTYCUTOFF = 0.95 # stopping certainty threshold PARTICIPANTTO_MODEL = 'p01' # participant ID to simulate ```

🔄 Customization Tips

• See how MODEL_CERTAINTY_CUTOFF affects the number of samples collected and prediction quality
• Human participant: see how CLEANSER_FREQUENCY affects fatigue (by preventing long stretches of ambiguous stimuli)
• Virtual agent: simulate other participants by changing PARTICIPANT_TO_MODEL

References

Einfeldt, M., Sevastjanova, R., Zahner-Ritter, K., Kazak, E., & Braun, B. (2024). The use of Active Learning systems for stimulus selection and response modelling in perception experiments. Computer Speech & Language, 83, 101537. https://doi.org/10.1016/j.csl.2023.101537

Kocjančič, T., & Bořil, T. (2025). Voicing in Czech children’s sibilants: children’s productions and adult’s perception, International Clinical Phonetics and Linguistics Association (ICPLA), June 24-27, 2025, Patras.