Gefilte Hebrew Sentences

Gefilte Hebrew Sentences Modern, everyday Hebrew sentences with nikud and an English translation, graded 1 to 5 by how hard they are to read aloud for a non-native speaker, plus recordings of them. Why it exists. Most Hebrew speech data comes from native speakers. This set is for a personal model of one immigrant's Hebrew: accented, with imperfect pronunciation. That speech is part of everyday Hebrew too. The nikud is there so the reader can get the vowels right, and the grades… See the full description on the dataset page: https://huggingface.co/datasets/danielrosehill/Gefilte-Hebrew-Sentences.

About This Project

Gefilte Hebrew Sentences provides a collection of modern Hebrew text and corresponding audio recordings specifically designed to represent non-native speech patterns. Each entry includes nikud to assist with pronunciation and a difficulty rating from one to five, based on the complexity of reading the text aloud. By capturing accented speech and imperfect pronunciation, the dataset offers a more realistic representation of the linguistic diversity found in everyday Hebrew conversation.

The project is intended for developers and researchers working on automatic speech recognition and text-to-speech models who require training data beyond native-speaker benchmarks. It is available on Hugging Face in parquet format and supports integration with libraries such as Dask and Polars, facilitating its use in machine learning pipelines focused on accented speech and language acquisition.

Project Details

Tags

task_categories:automatic-speech-recognitiontask_categories:text-to-speechlanguage:helanguage:enlicense:cc-by-2.0size_categories:10k<n<100kformat:parquetmodality:audiomodality:tabularmodality:textlibrary:datasetslibrary:dasklibrary:polarslibrary:mlcroissantregion:ushebrewnikudnon-native-speechaccented-speech

Explore More Projects