Hebrish
Technical

Accessory models

Not speech models, but the speech models' pipelines need them.

Classifiers

Text written in English with Hebrew words in Latin letters looks like plain English to every standard tool; the token classifier finds the Hebrew in it, so a pipeline can hand each Hebrew word to a Hebrew voice.

Hebrew-Latin Token Classifier small model

Meta · XLM-RoBERTa base

Marks each romanized Hebrew word in an English sentence, so text-to-speech can send it to a Hebrew voice. Span F1 0.941 on terms it never saw in training.

Current
v1 (2026-10-15)
Last updated
Formats
Transformers (safetensors)
Parameters
277M
Result
Span F1 0.941 on terms it never saw in training
DOIs
2

Formats, evals and DOIs Hugging Face collection

Other model groups: Speech recognition models · Speech synthesis models · All models. Use of every Hebrish model is governed by the terms of use.