You need to agree to share your contact information to access this model

This repository is publicly accessible, but you have to accept the conditions to access its files and content.

This model derives from the SASPEECH dataset (© Israeli Public Broadcasting Corporation). It is licensed for NON-COMMERCIAL use only. By accessing it you accept the SASPEECH terms reproduced in the model card below.

Kokoro-Hebrew (SASPEECH) — mlx

A Kokoro-82M text-to-speech model fine-tuned for Hebrew on the SASPEECH dataset (speaker: Shaul Amsterdamski), provided in mlx format for Apple-Silicon inference via mlx-audio.

Hebrew grapheme→phoneme uses phonikud (vocalized Hebrew → IPA, normalized to Kokoro's 178-token vocab). The G2P must be applied identically at training and inference time.

Model & training

This release is fine-tuned from Kokoro-82M with a two-stage recipe (StyleTTS2 trainer / kikiri) on both SASPEECH subsets, to trade breadth for polish:

Stage 1 — breadth (~20 h): the SASPEECH automatic subset (OpenSLR SLR134, ~26 h machine-transcribed with Whisper and machine-vocalized with Nakdimon; ~20 h after filtering). This gives the model wide phonetic and prosodic coverage, but noisier labels.
Stage 2 — polish (~4 h): the clean, human-vocalized gold subset, which sharpens pronunciation and prosody on top of the stage-1 foundation.

Diacritics for the automatic subset were restored where missing with phonikud's neural diacritizer; the exact same phonikud G2P is used at train and inference.

Each checkpoint was tracked on held-out sentences for intelligibility (Whisper round-trip WER) and speaker similarity (ECAPA-TDNN cosine to real Shaul references); this release was selected by listening, favouring natural prosody. It corresponds to the stage-2 epoch-3 checkpoint.

Files

kokoro-v1_0.safetensors — the model (bf16, mlx-audio layout)
voices/he_shaul.safetensors — the Shaul voicepack
config.json, vocab.json, kokoro_symbols.py, hebrew_g2p.py — inference support

⚠️ License — NON-COMMERCIAL (SASPEECH / IPBC)

This model is a derivative of the SASPEECH dataset, whose recordings and transcriptions are © the Israeli Public Broadcasting Corporation (IPBC) and released for non-commercial use only (OpenSLR SLR134). These terms govern this model. By using it you accept that you:

may use it for non-commercial purposes only — not for commercial or broadcast needs;
will not imply that the IPBC endorses you or your use;
will not use it in a way that harms or defames Shaul Amsterdamski or the IPBC, including defamation;
will not use it for political needs;
will not use it in a manner that breaches any applicable law.

A code/wrapper license (e.g. MIT) does not relax these terms — the voice itself is bound by the SASPEECH terms above, regardless of how it is packaged.

Intended use

Personal, research, and other non-commercial Hebrew TTS.

Out of scope

Commercial products or broadcast; political use; anything intended to harm, defame, or impersonate Shaul Amsterdamski or the IPBC; anything unlawful. For commercial use, source/train a model on a commercially-licensed dataset instead.

Disclaimer

Provided "AS IS". Neither the IPBC nor the uploader makes any warranty or is liable for any loss, injury, or damage arising from use, nor for changes made to the model by you or any third party.

Citation (dataset)

Please cite the SASPEECH dataset — see OpenSLR SLR134 (Sharon et al., SASPEECH: A Hebrew Single Speaker Dataset, INTERSPEECH 2023).

Downloads last month: 18

MLX

Hardware compatibility

Quantized