maxmelichov retweeted
Everything is open: paper, code, models, and ILSpeech, a studio-quality Hebrew TTS dataset with IPA annotations.
📄 Paper: arxiv.org/abs/2506.12311
🌐 Project: phonikud.github.io
🔊 Follow-up (ReNikud): renikud.github.io
🤗 Data: huggingface.co/Phonikud
Huge thanks to @maxmelichov, @oyd11, and @MorrisAlper 🙏 More soon :)
#interspeech2026
maxmelichov retweeted
This led to a follow-up project with a bigger idea: ReNikud. What if Hebrew pronunciation could be learned directly from audio, without manual nikud labels? We run Hebrew ASR + IPA ASR on 1,700 hours of speech, keep only transcripts that agree, and train on 1.5M pseudo-labeled sentences with everyday spoken Hebrew including slang, loanwords, and colloquial forms. 🧵
maxmelichov retweeted
Does it work? Listen for yourself.
הוא רצה את זה, "he wanted it" (ratsa) vs. היא רצה מהר, "she ran fast" (ratsa, different stress). Same letters.
Phonikud handles cases that trip up commercial systems, trained with as little as 2 hours of data. 🧵
maxmelichov retweeted
Phonikud extends a Hebrew diacritizer with lightweight adapters that predict new enhanced marks indicating lexical stress and vocal Shva. The enhanced Nikud then maps directly to full IPA phonemes.
Small enough for real-time TTS on edge devices like a Raspberry Pi. 🧵
maxmelichov retweeted
Hebrew is written without vowels. ספר can be "book" (sefer), "barber" (sapar), or "he counted" (safar). Deep learning diacritizers handle this by restoring vowel marks (nikud).
But nikud doesn't mark lexical stress (בוקר: "boker" morning vs. "boker" cowboy) or vocal Shva (בלונדון "belondon" vs. בלונדיני "blondini").
For TTS, that's not enough. 🧵
maxmelichov retweeted
Our Hebrew G2P, Phonikud, was accepted to Interspeech! 🎉
In Hebrew, בירה can be "beer" ('bira) or "capital" (bi'ra). Same vowel marks, different stress. Even fully vocalized Hebrew underspecifies pronunciation. 🧵
Thank you @izakid664 for sponsoring me on @github. You can join them at my sponsors profile: github.com/sponsors/maxmelic…
31
maxmelichov retweeted
Yakov Kolani, Maxim Melichov, Cobi Calev, Morris Alper, "Phonikud: Hebrew Grapheme-to-Phoneme Conversion for Real-Time Text-to-Speech," arxiv.org/abs/2506.12311
maxmelichov retweeted
I'm pretty sure the buildings I had previously modeled are all way out of scale. #gamedev
Don't worry about passwords ever again, sign up for LastPass now! lastpass.com/f?47403262
Just tried out this awesome plugin Social Traffic Pop - Its Amazing! MEGADOWN.CO.IL












