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

The article introduces Phonikud, an open-source Hebrew grapheme-to-phoneme (G2P) system that enhances text-to-speech (TTS) accuracy by providing fully-specified IPA transcriptions, addressing phonetic underspecification issues in Hebrew. It includes the ILSpeech corpus for Hebrew audio-text pairs, a benchmark for Hebrew G2P conversion, and models that capture nuanced phonetic details for TTS evaluation. This framework allows smaller TTS models to achieve performance comparable to larger proprietary systems, offering significant improvements for practitioners in Hebrew TTS development.