![[2602.18899] [b]=[d]-[t]+[p]: Self-supervised Speech Models Discover Phonological Vector Arithmetic](https://arxiv.org/static/browse/0.3.4/images/arxiv-logo-fb.png){kind=logo}
[2602.18899] [b]=[d]-[t]+[p]: Self-supervised Speech Models Discover Phonological Vector Arithmetic
Summary
This article explores how self-supervised speech models (S3Ms) encode phonological information, revealing linear relationships in their representation space that correspond to phonological features across 96 languages.
Why It Matters
Understanding the structure of phonological information in S3Ms is crucial for advancing natural language processing and speech recognition technologies. This research provides insights into how these models can be improved to better understand and generate human speech, which has implications for various applications in AI and linguistics.
Key Takeaways
- S3Ms encode rich phonetic information that is structured in a meaningful way.
- Linear directions in the model's representation correspond to phonological features.
- The scale of phonological vectors correlates with the acoustic realization of features.
- The study demonstrates phonological vector arithmetic, enhancing our understanding of speech models.
- Findings could lead to improved applications in speech recognition and natural language processing.
Electrical Engineering and Systems Science > Audio and Speech Processing arXiv:2602.18899 (eess) [Submitted on 21 Feb 2026] Title:[b]=[d]-[t]+[p]: Self-supervised Speech Models Discover Phonological Vector Arithmetic Authors:Kwanghee Choi, Eunjung Yeo, Cheol Jun Cho, David Harwath, David R. Mortensen View a PDF of the paper titled [b]=[d]-[t]+[p]: Self-supervised Speech Models Discover Phonological Vector Arithmetic, by Kwanghee Choi and Eunjung Yeo and Cheol Jun Cho and David Harwath and David R. Mortensen View PDF Abstract:Self-supervised speech models (S3Ms) are known to encode rich phonetic information, yet how this information is structured remains underexplored. We conduct a comprehensive study across 96 languages to analyze the underlying structure of S3M representations, with particular attention to phonological vectors. We first show that there exist linear directions within the model's representation space that correspond to phonological features. We further demonstrate that the scale of these phonological vectors correlate to the degree of acoustic realization of their corresponding phonological features in a continuous manner. For example, the difference between [d] and [t] yields a voicing vector: adding this vector to [p] produces [b], while scaling it results in a continuum of voicing. Together, these findings indicate that S3Ms encode speech using phonologically interpretable and compositional vectors, demonstrating phonological vector arithmetic. All code ...
