Phonikud: Overcoming Phonetic Underspecification for Hebrew Text-To-Speech

What it is about

Modern Hebrew is written without most vowels, and even pointed text leaves out stress and whether a shva is pronounced. Phonikud is a grapheme-to-phoneme pipeline: a frozen DictaBERT diacritiser gets a small trainable head that predicts three extra symbols (stress on a non-final syllable, a shva read as /e/, and a bar at the end of a cliticised prefix). A deterministic transducer with dictionary matching then produces IPA. Its training labels for stress, prefix boundaries and shva come from a morpho-phonological API, corrected by hand on the 1,000 most frequent word types of a 5-million-line corpus. On 100 items of a new expert-annotated set (ILSpeech), WER is 17.4 against 39.5 for DictaBERT and 40.5 for Nakdimon, both with the same IPA step. On 250 homograph sentences, the full system has a stress WER of 3.2 and exact-match 77.0; removing the stress symbols gives 8.4 and 46.4. Small TTS models built on the output approach large proprietary systems. The ILSpeech corpus is released under a non-commercial licence.

Ideas

• I1. Mark the end of a cliticised prefix (the one-letter particles ו, ה, ב, ל, כ, מ, ש and their like) so that a dictionary lookup can match irregular words after the prefix is removed. For tools/hbo (the prefix tables) and 2. Script and spelling, test how many irregular forms match with and without the boundary mark.
• I2. A deterministic finite-state transducer for the grapheme-to-sound step, with a dictionary for exceptions, that handles letters read out of linear order (their example: a vowel letter read after the consonant it follows). Write it as a rule table in 3. Phonology and test it on a word list with gold readings. This fits Mova's rules-first design.
• I3. Correct the labels on the most frequent word types first. The paper's pseudo-labels are checked by hand on the top 1,000 types, which cover most tokens. Use the same leverage for any Hebrew rule set, and score the result with exact match on the whole sentence, not only with a per-word error rate.

For Mova

I1 goes into tools/hbo (prefix segmentation) and 2. Script and spelling. I2 goes into 3. Phonology. I3 is a method for the Modern Hebrew plan in 12. Mova and Modern Hebrew: where it stands, what it would need, since Mova does not parse Modern Hebrew yet. The ILSpeech corpus and the TTS models are not used in Mova builds, because the corpus is non-commercial.

Cite

The paper itself, not this card. BibTeX from the library's record of the work. arXiv 2506.12311

@misc{kolani2025phonikud,
  title = {{Phonikud: Overcoming Phonetic Underspecification for Hebrew Text-To-Speech}},
  author = {Kolani, Yakov and Melichov, Maxim and Calev, Cobi and Alper, Morris},
  year = {2025},
  eprint = {2506.12311},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2506.12311}
}

References

27 references, 7 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

1. M. P. Lewis, Ed., Ethnologue: Languages of the World, sixteenth ed. Dallas, TX, USA: SIL International, 2009. find
2. A. Roth, A. Turetzky, and Y. Adi, “A language modeling approach to diacritic-free hebrew tts,” in Proc. Interspeech 2024, 2024, pp. 2775–2779. find our card
3. E. Zeldes, O. Tal, and Y. Adi, “Enhancing tts stability in hebrew using discrete semantic units,” in ICASSP 2025-2025 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). IEEE, 2025, pp. 1–5. find our card
4. O. Sharoni, R. Shenberg, and E. Cooper, “Saspeech: A hebrew single speaker dataset for text to speech and voice conversion,” in Proc. Interspeech, 2023. find
5. V. Pratap, A. Tjandra, B. Shi, P. Tomasello, A. Babu, S. Kundu, A. Elkahky, Z. Ni, A. Vyas, M. Fazel-Zarandi et al., “Scaling speech technology to 1,000+ languages,” Journal of Machine Learning Research, vol. 25, no. 97, pp. 1–52, 2024. find
6. S. Shmidman, A. Shmidman, and M. Koppel, “Dictabert: A state-of-the-art bert suite for modern hebrew,” 2023. find our card
7. D. Graf and A. Ussishkin, “Emergent iambs: stress in modern hebrew,” Lingua, vol. 113, no. 3, pp. 239–270, 2003. find
8. W. Weinberg, “Spoken israeli hebrew: Trends in the departures from classical phonology,” Journal of Semitic Studies, vol. 11, no. 1, pp. 40–68, 1966. find
9. G. Mor-Lan, E. Levi, T. Sheafer, and S. R. Shenhav, “Israparltweet: The israeli parliamentary and twitter resource,” in Proceedings of the 2024 Joint International Conference on Computational Linguistics, Language Resources and Evaluation (LREC-COLING 2024), 2024, pp. 9372–9381. find our card
10. S. Izre’el, B. Hary, and G. Rahav, “Designing cosih: the corpus of spoken israeli hebrew,” International Journal of Corpus Linguistics, vol. 6, no. 2, pp. 171–197, 2001. find
11. J. Azogui, A. Lerner, and V. Silber-Varod, “The open university of israel map task corpus (matacop),” 2016. find
12. M. Marmorstein and N. Matalon, “The huji corpus of spoken hebrew: An interaction-oriented design of a corpus,” 2022. find
13. A. Conneau, M. Ma, S. Khanuja, Y. Zhang, V. Axelrod, S. Dalmia, J. Riesa, C. Rivera, and A. Bapna, “Fleurs: Few-shot learning evaluation of universal representations of speech,” in 2022 IEEE Spoken Language Technology Workshop (SLT). IEEE, 2023, pp. 798–805. find
14. Y. Marmor, K. Misgav, and Y. Lifshitz, “ivrit. ai: A comprehensive dataset of hebrew speech for ai research and development,” arXiv preprint arXiv:2307.08720, 2023. ↗ our card
15. A. Turetzky, O. Tal, Y. Segal, Y. Dissen, E. Zeldes, A. Roth, E. Cohen, Y. Shrem, B. R. Chernyak, O. Seleznova et al., “Hebdb: a weakly supervised dataset for hebrew speech processing,” in Proc. Interspeech 2024, 2024, pp. 1360–1364. find our card
16. Adobe, “Adobe Podcast Enhance Speech v2,” 2025. [Online]. Available: https://podcast.adobe.com/en/enhance ↗
17. E. Gershuni and Y. Pinter, “Restoring hebrew diacritics without a dictionary,” in Findings of the Association for Computational Linguistics: NAACL 2022, 2022, pp. 1010–1018. find our card
18. eSpeak NG Contributors, “eSpeak NG Text-to-Speech,” 2025. [Online]. Available: https://github.com/espeak-ng/espeak-ng ↗
19. J. Zhu, C. Zhang, and D. Jurgens, “Byt5 model for massively multilingual grapheme-to-phoneme conversion,” in Proc. Interspeech 2022, 2022, pp. 446–450. find
20. Anthropic, “Introducing Claude Opus 4.6,” 2026. [Online]. Available: https://www.anthropic.com/news/claude-opus-4-6 ↗
21. The Gemini Team (Google), “Gemini 3.1 Pro: A smarter model for your most complex tasks,” 2026. [Online]. Available: https://blog.google/innovation-and-ai/models-and-research/gemini-models/gemini-3-1-pro/ ↗
22. J. Kim, J. Kong, and J. Son, “Conditional variational autoencoder with adversarial learning for end-to-end text-to-speech,” in International Conference on Machine Learning. PMLR, 2021, pp. 5530–5540. find
23. Piper Contributors, “Piper: A Fast, Local Neural Text to Speech System,” 2025. [Online]. Available: https://github.com/rhasspy/piper ↗
24. Y. A. Li, C. Han, V. Raghavan, G. Mischler, and N. Mesgarani, “Styletts 2: Towards human-level text-to-speech through style diffusion and adversarial training with large speech language models,” Advances in Neural Information Processing Systems, vol. 36, pp. 19 594–19 621, 2023. find
25. M. Melichov and T. Hasson, “Robo-Shaul,” http://www.roboshaul.com/, 2023. [Online]. Available: https://github.com/maxmelichov/Text-To-speech ↗
26. Tulsee Doshi, “Gemini 2.5: Our most intelligent models are getting even better,” 2025. [Online]. Available: https://blog.google/innovation-and-ai/models-and-research/google-deepmind/google-gemini-updates-io-2025/ ↗
27. OpenAI, “GPT-4o mini TTS,” 2025. [Online]. Available: https://developers.openai.com/api/docs/models/gpt-4o-mini-tts ↗

This page in Markdown → https://md.mova.elamur.ai/free/en/docs/papers/langstream-2025-kolani-phonikud-overcoming-phonetic-underspecification-hebrew/