Phonikud: Overcoming Phonetic Underspecification for Hebrew Text-To-Speech
Researchers have developed Phonikud, an open-source system designed to improve text-to-speech (TTS) for Modern Hebrew by addressing phonetic underspecification. This framework includes a grapheme-to-phoneme system that outputs detailed International Phonetic Alphabet (IPA) transcriptions, a new corpus named ILSpeech with annotated Hebrew audio and text, and models for automatic TTS evaluation. The system demonstrates improved phoneme prediction compared to previous methods, with small TTS models utilizing Phonikud's phonetic input achieving performance comparable to larger proprietary systems. AI
IMPACT Enhances TTS capabilities for under-resourced languages by providing a more accurate phonetic representation.