Proceedings

Comparing Phonological Feature Sets for Low-Resource ASR

---

Abstract

In this paper, we explore an alternative ASR framework in which phonological features are predicted as an explicit intermediate representation, rather than predicting phones directly. Because feature systems encode cross-linguistically meaningful structure, this intermediate representation can reduce sample complexity by constraining what must be learned from limited data, while also enabling rapid adaptation to new languages through changes to the phone-to-feature mapping rather than retraining the model. As a result, this approach is particularly well suited to low-resource settings. We retrained Phonet models on two different feature sets to see the extent to which specific theories of phonological features facilitate better phoneme recognition, using a low-resourced language (Yan-nhangu, Pama-Nyungan) as a testing ground for performance. We use a naïve greedy decoding strategy to isolate the effect of feature set choice, and find that IPA features lead to the best transcription accuracy, followed closely by a featureless baseline.

Keywords: phonet, ASR, low-resource languages, phonological features

How to Cite:

Daul, M., Tosolini, A., Karakas, A. & Bowern, C., (2026) “Comparing Phonological Feature Sets for Low-Resource ASR”, Proceedings of the Annual Meetings on Phonology 2(1). doi: https://doi.org/10.7275/amphonology.3704