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

2025-06-14 · Source: cs.CL updates on arXiv.org · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Natural Language Processing & Speech · Depth: Expert, quick

Summary

Phonikud, an open-source framework, addresses the challenge of phonetic underspecification, particularly stress, in Modern Hebrew Text-to-Speech (TTS). The framework introduces four key contributions: Phonikud itself, a Hebrew grapheme-to-phoneme (G2P) system that augments a diacritizer to produce fully-specified International Phonetic Alphabet (IPA) transcriptions; the ILSpeech corpus, a new dataset featuring paired Hebrew audio, text, and expert IPA annotations; a novel benchmark for Hebrew G2P conversion, a task previously unmeasured; and Hebrew audio-to-IPA models designed to capture overlooked phonetic details for automated TTS evaluation. Experimental results demonstrate that Phonikud achieves higher accuracy in predicting Hebrew phonemes compared to existing methods. Furthermore, small, local TTS models leveraging phonetic input from Phonikud can achieve performance comparable to large proprietary systems. The code, data, and models are publicly released.

Key takeaway

For NLP Engineers developing Hebrew Text-to-Speech systems, Phonikud offers a critical advancement. You should integrate this open-source grapheme-to-phoneme system to overcome phonetic underspecification, particularly regarding stress. By doing so, you can achieve more accurate phoneme predictions and enable smaller, local TTS models to rival the performance of larger proprietary solutions. Consider using the ILSpeech corpus and the new G2P benchmark for your development and evaluation efforts.

Key insights

Fully specifying phonetic features like stress significantly improves Hebrew TTS accuracy and enables smaller models.

Principles

• Augmenting diacritizers improves G2P accuracy.
• Expert IPA annotations enhance corpus quality.
• Benchmarking unmeasured tasks drives progress.

Method

Phonikud augments a base diacritizer to generate fully-specified IPA transcriptions for Hebrew text, then uses these as input for TTS models.

In practice

• Use Phonikud for accurate Hebrew G2P.
• Integrate ILSpeech corpus for Hebrew TTS training.
• Evaluate TTS models with new audio-to-IPA models.

Topics

• Hebrew TTS
• Grapheme-to-Phoneme
• Phonetic Underspecification
• IPA Transcription
• ILSpeech Corpus
• Speech Synthesis

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

Related on AIssential

• Hard to Be Heard: Phoneme-Level ASR Analysis of Phonologically Complex, Low-Resource Endangered Languages — Data scarcity, not phonological complexity, primarily drives ASR errors in low-resource, typologically complex languages.
• 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.
• Evaluating and Preserving Lexical Stress in English-to-Chinese Speech-to-Speech Translation — A novel S2ST system and evaluation metric significantly improve cross-lingual lexical stress transfer from English to Chinese.
• PolySpeech-100: A Large-Scale Benchmark for Speech Understanding Across 100+ Languages and Dialects — PolySpeech-100 reveals E2E Speech-LLMs excel on dialects, but open-source models struggle with low-resource languages, and CoT prompting can degrade performance.
• AP-GRPO: Anchor-Gated Phonetic Alignment with Policy Optimization for Pathological Speech Reconstruction — AP-GRPO reconstructs pathological speech by aligning SLMs using audible anchors and phonetic compatibility, adapting to disease severity.
• Lexical Tone is Hard to Quantize: Probing Discrete Speech Units in Mandarin and Yor\`ub\'a — Current DSU quantization methods struggle to reliably encode lexical tone, prioritizing phonetic structure instead.

Open in AIssential →

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