Customising Czech Phonetic Alignment using HuBERT and manual segmentation

This paper presents Prak, a forced alignment tool developed for Czech, with a focus on transparent modular design and phonetic accuracy. In addition to a rule-based pronunciation module and exception handling, Prak introduces a novel application of non-deterministic, backward-processing FSTs to model complex regressive assimilation processes in Czech consonant clusters. We further describe the integration of a HuBERT-based transformer model and training including extensive manually time-aligned data to enhance phone classification accuracy while maintaining ease of installation and use. Evaluation against a manually aligned test corpus demonstrates that the enhanced model significantly outperforms both our earlier Prak-CV model and the long-established previous forced alignment baseline. The new model reduces major boundary errors and mismatches, bringing alignment accuracy closer to manual phonetic segmentation standards for Czech. We emphasize both methodological transparency and practical usability, aiming to support phoneticians working with Czech as well as developers interested in extending the tool for other languages.