Reproduce “PhiNet: Speaker Verification with Phonetic Interpretability” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Yi Ma, Shuai Wang, Tianchi Liu, Haizhou Li
PhiNet introduces the first self-interpretable speaker verification network by integrating phonetic evidence directly into its decision-making process, offering both local (phoneme-specific) and global (overall phoneme distinctiveness) explanations for verification outcomes. It achieves speaker verification performance comparable to black-box systems while providing human-understandable evidence, aligning with forensic speaker comparison principles.
Key insights
- The model's learned phonetic weights, indicating discriminative power, align with observations from human forensic speaker comparison (e.g., nasals and some vowels are highly important).
- There is an observed trade-off between achieving optimal speaker verification performance and maintaining the faithfulness (fidelity) of local interpretability, particularly with varying training sample durations.
- Data augmentation is crucial for improving both speaker verification accuracy and the faithfulness of the interpretability, while mixed-duration training can destabilize phonetic weight learning.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.