ReNikud: Audio-Supervised Hebrew Grapheme-to-Phoneme Conversion

2026-06-19 · Source: cs.CL updates on arXiv.org · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Software Development & Engineering · Depth: Expert, short

Summary

ReNikud is a novel method addressing the challenges of Grapheme-to-Phoneme (G2P) conversion for Modern Hebrew, crucial for text-to-speech (TTS) applications. Hebrew's abjad writing system, which largely omits vowels, creates significant ambiguity that traditional G2P approaches struggle with due to scarce vocalization data or limitations with direct sequence-to-sequence IPA prediction. ReNikud overcomes these by employing two key insights: weak audio supervision via a phoneme-based automatic speech recognition (ASR) pseudo-labeling pipeline on thousands of hours of unlabeled Hebrew audio, and a pseudo-vocalization architecture that predicts IPA phonemes at each character position, enforcing character-level alignment. This approach generates phonemic transcriptions reflecting natural spoken norms and has surpassed previous state-of-the-art methods on existing Hebrew G2P benchmarks and the new MILIM benchmark.

Key takeaway

For NLP Engineers developing Hebrew text-to-speech systems, ReNikud offers a robust solution to the complex G2P problem. You should consider integrating its audio-supervised pseudo-labeling and pseudo-vocalization architecture to generate more accurate, natural-sounding phonemic transcriptions. This method directly addresses the scarcity of vocalization data and the nuances of spoken Hebrew, potentially improving your TTS model's naturalness and performance significantly.

Key insights

ReNikud uses audio-supervised pseudo-labeling and a pseudo-vocalization architecture to improve Hebrew G2P conversion.

Principles

• Audio supervision can generate natural spoken phonemic data.
• Character-level alignment improves abjad G2P models.

Method

ReNikud employs a phoneme-based ASR pipeline to pseudo-label thousands of hours of unlabeled Hebrew audio, then uses a pseudo-vocalization architecture to predict IPA phonemes per character.

In practice

• Apply ASR pseudo-labeling for low-resource language phonemic data.
• Design G2P models with character-level alignment for abjad systems.

Topics

• Hebrew G2P
• Text-to-Speech
• Audio Supervision
• ASR Pseudo-labeling
• Phonemic Transcription
• Abjad Systems

Best for: Research Scientist, AI Scientist, NLP Engineer, Machine Learning Engineer

Related on AIssential

• ReNikud: Audio-Supervised Hebrew Grapheme-to-Phoneme Conversion — ReNikud leverages weak audio supervision and a character-aligned pseudo-vocalization architecture for accurate Hebrew Grapheme-to-Phoneme conversion.
• Phonikud: Overcoming Phonetic Underspecification for Hebrew Text-To-Speech — Fully specifying phonetic features like stress significantly improves Hebrew TTS accuracy and enables smaller models.
• From A to B to A: Palindromic Zero-Shot Voice Conversion with Non-Parallel Data — A voice conversion framework uses KNN over WavLM to create synthetic training pairs from non-parallel data for zero-shot, multilingual VC.
• Scaling Human and G2P Supervision for Robust Phonetic Transcription — G2P supervision for phonetic transcription has diminishing returns beyond 20-30 hours of human annotation.
• LoSATok: Low-dimensional Semantic-Acoustic Tokenizer for Cross-Domain Audio Understanding and Generation — LoSATok unifies audio understanding and generation via low-dimensional semantic-acoustic tokens, reducing Diffusion Transformer modeling complexity.
• Rhythm of the Deep: A Computational-Linguistic Test of Duality of Patterning in Sperm Whale Codas — The study reveals a two-tier, duality-of-patterning-like architecture in sperm whale codas, with rhythmic lower-tier click combinations and upper-tier sequential dependence.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by cs.CL updates on arXiv.org.