Interpretable by design: An interpretability-centred approach to automatic L1 identification from L2 English speech

As AI and machine learning methods are increasingly used for high-stakes decision-making, the challenges surrounding understanding how these decisions are made has emerged as a critical concern, prompting a push towards the development of more interpretable methods. The field of forensic speech science (FSS) presents a particularly demanding context for exploring interpretability not only due to legal and regulatory requirements, but also the responsibility to the public. Although this is a broader challenge facing the field, the research presented in this thesis focuses on the gap in systems for automatic L1 identification from L2 speech, specifically its evidential applications.

This research adopts an interpretability-centred approach exploring: 1) how interpretability can be integrated into the design of an automatic L1 identification system, and 2) whether such design choices can address the limitations of existing systems for forensic applications. The thesis investigated three methods for integrating interpretability and contributed practical strategies for evaluation within an FSS context. A novel phoneme modelling method was introduced, the Variational Autoencoder with Spectrogram Reconstruction (VAE+SR) method, which enables real-time exploration of which underlying speech features are captured in the encodings. Additionally, existing forensic methods, MVKD and Bayesian Networks, are newly applied to the L1 identification task and explored as a means of improving different aspects of interpretability. Results demonstrated that while the presented methods offered ways to address the limitations of existing systems, there is still room for development as new limitations emerge.

As the field continues to evolve, this work demonstrates how deliberate design choices that prioritise interpretability can help forensic phonetics practitioners present data-driven decisions that can be explained and supported by existing research in linguistics and phonetics and encourages a rethinking of how we may integrate interpretability into system design.