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

A new open-source system called Phonikud addresses a long-standing gap in Hebrew text-to-speech by producing fully specified phonetic transcriptions that include features such as stress, which earlier tools overlooked [1][2]. The framework, detailed in a paper by Morris Alper and colleagues, was first submitted to arXiv in June 2025 and revised through June 2026 [1]. It centers on a grapheme-to-phoneme converter that outputs International Phonetic Alphabet (IPA) transcriptions, built by augmenting a base diacritizer [2]. The authors argue that Modern Hebrew’s orthographic complexity has historically forced TTS systems to ignore underspecified phonetic features, degrading naturalness [2]. Alongside Phonikud, the team released the ILSpeech corpus, which pairs Hebrew audio and text with expert IPA annotations [2]. They also established a benchmark for Hebrew G2P conversion, a task that had not been systematically measured before [2]. The paper reports that Phonikud predicts Hebrew phonemes more accurately than prior methods, and that small, locally run TTS models fed with Phonikud’s output can approach the quality of large proprietary systems [2]. The work appears at a time when open-source AI releases are drawing heightened attention. Platforms such as Hugging Face have built infrastructure to link papers directly to models, datasets, and interactive demos, making it easier for researchers to share artifacts and for users to test systems without writing code [4][5]. The Phonikud project follows this pattern, with code, data, and models publicly available [2]. Hebrew TTS has lagged behind languages with more straightforward orthography-to-sound mappings. By providing fully specified IPA transcriptions, Phonikud gives downstream TTS engines the phonetic detail they need to produce more accurate speech [2]. The ILSpeech corpus also supplies a foundation for training audio-to-IPA models that can capture previously disregarded details for automatic evaluation [2]. The final version of the paper, posted in June 2026, weighed in at 608 KB, down from earlier submissions of 1,380 KB and 1,433 KB [1]. The authors have not yet disclosed whether the system will be integrated into commercial TTS services, but the open-source release invites further development and testing by the wider research community [2].