COMPARING UZBEK GRAPHEME-TO-PHONEME CONVERTERS: EPITRAN, ESPEAK NG, AND UZG2P ON SHARED EVALUATION CRITERIA
Keywords:
Keywords: Uzbek phoneme-based speech pipelines need consistent text-to-phoneme conversion. Epitran supplies a built-in map; eSpeak NG provides a native Uzbek voice; uzg2p adds normalization, foreign-token flagging, stress, and map repairs.Abstract
We compare stock Epitran uzb-Latn, eSpeak NG’s uz voice, and uzg2p, a patched Epitran wrapper, on robustness, phoneme inventory, digraphs, apostrophe canonicalization, and stress. All three complete a 356-item vocabulary without crashes or empty outputs, but their inventories differ substantially. uzg2p offers normalization, configurable oʻ→/ɵ/, and stress control; eSpeak supplies a synthesis baseline. Inventory harmonization is essential for phoneme error rate (PER) comparisons. Without independently annotated pronunciations, these findings establish implementation differences, not universal phonological superiority.
References
1.Mortensen, D. R., Dalmia, S., & Littell, P. (2018). Epitran: Precision G2P for Many Languages. LREC 2018.
2.eSpeak NG contributors. eSpeak NG supported languages. GitHub.
3.Azimov, U. (2026). uzg2p: Auditing Uzbek Grapheme-to-Phoneme Conversion and a Phoneme Error Rate Benchmark for Agglutinative-Language ASR. Preprint.
4.uzinfocom-edu-ai. uzbek-asr-curated-701h. HuggingFace Datasets.
5.Canopy Labs. Orpheus-TTS. GitHub: canopyai/Orpheus-TTS.