BSc. Thesis Repository

---

⚥ Gender Differences in Vocal Markers of Autism

Investigating Autistic Speech through Machine-Learning Generalizability

The speech of autistic individuals tends to show atypical patterns of prosody and voice quality. Machine learning (ML) has increasingly been used to identify vocal markers of autism and support automatic screening, however, the relation between vocal markers and gender remains underexplored. This study systematically examined gender differences in vocal markers of autism by assessing within- and cross-gender generalizability of ML models.

Using a gender-balanced dataset from a standardized social interaction task, a robust Long Short-Term Memory (LSTM) pipeline was developed to model the broader temporal profile of the task. Three models trained on male-only, female-only, and mixed-gender data were evaluated on held-out male and female test sets to examine gender-specific generalization. Feature importance analysis using the TIME framework was conducted to scaffold model interpretation, and a control analysis using a Support Vector Machine (SVM) with identical data partitions and evaluation logic assessed the benefits and trade-offs of different architectures.

Results showed substantial variability in generalization across models: the female-only model exhibited above-chance performance on both within- and cross-gender test sets, the mixed-gender model showed the most robust performance on the male test set, and the male-only model failed to generalize. Feature importance analysis did not identify stable individual acoustic predictors, supporting prior work that ASD vocal markers are multivariate and relational. The control analysis showed that LSTMs exhibited better cross-gender generalization, highlighting the relevance of modelling temporal patterns. While constrained by limited model performance and interpretability, these findings align with existing research on female masking behaviours and the heterogeneous nature of autistic speech and contribute to a more inclusive investigation of autistic vocal profiles.

---

📁 Repository Structure

python_code/

• lstm_main.py
  main script for training LSTM models, performing hyperparameter tuning, testing on held-out data, evaluating performance, and feature importance analysis using TIME
• lstm_functions.py
  contains functions for model training and evaluation done in lstm_main.py
• requirements.txt
  all packages and their versions that were used in the project
• results/ contains output from the ML pipeline: selected hyperparameters for each model, model weights, and detailed train/test outputs

r_code/

• grand_new.Rmd
  main script that handles data partitioning and min-max scaling for each gender group
• grand_function_new.R
  pipeline used in grand_new.Rmd for preprocessing steps
• functions/
  contains several helper functions used in grand_function_new.R

results/

contains performance overview for LSTM and SVM models across training, validation and held-out test sets; also contains ranked feature importance lists for all models

• lstm_svm_performance.csv
• mean_female_importance_list.csv
• mean_male_importance_list.csv
• mean_mixed_importance_list.csv

svm_control_analysis/

• main.py main script for training SVM models, performing hyperparameter tuning, testing on held-out data, and evaluating performance
• functions.py contains functions for model training and evaluation done in main.py

utterance_prep/

• utterance_preprocessing.Rmd R script for aggregation of frame-level data into utterances based on transcript timestamps