AI
New system tackles Hebrew's phonetic ambiguity to improve speech synthesis
Researchers develop Phonikud, an open-source grapheme-to-phoneme converter that handles Hebrew's underspecified phonetics, enabling more natural text-to-speech systems.
Creating natural-sounding speech from Hebrew text has proven tricky for technologists, largely because the language's written form leaves out critical phonetic details. Researchers have now released Phonikud, a framework designed to address these gaps and bring Hebrew TTS closer to the quality of commercial systems.
The challenge of Hebrew orthography#
Hebrew writing omits many phonetic features that are essential for proper pronunciation, particularly stress patterns. This underspecification creates a bottleneck for speech synthesis systems: without knowing exactly where emphasis falls or which vowels are intended, automated systems struggle to generate intelligible speech. Existing TTS approaches for Hebrew have largely sidestepped this problem rather than solving it.
What Phonikud does#
The new framework centers on a grapheme-to-phoneme (G2P) converter that takes written Hebrew and outputs fully specified International Phonetic Alphabet (IPA) transcriptions. Built by extending a baseline diacritizer—a tool that adds diacritical marks to clarify pronunciation—Phonikud captures phonetic details previously overlooked by earlier methods.
New datasets and benchmarks#
To support the work, the researchers created two key resources:
- ILSpeech corpus: a collection of paired Hebrew audio recordings, corresponding text, and expert IPA annotations
- Hebrew G2P benchmark: the first standardized test for this previously unmeasured conversion task
They also developed audio-to-IPA models that extract detailed phonetic information automatically, enabling more nuanced evaluation of TTS output.
Performance gains#
Testing shows that Phonikud predicts Hebrew phonemes more accurately than prior approaches. When paired with smaller, locally-run TTS models, the system achieves results comparable to large proprietary services. The team has released their code, data, and trained models as open-source resources for further research and development.
Read the full paper on arXiv.
