HebDB

Paper: http://arxiv.org/abs/2407.07566

If you use our datasets, please use the following:

@article{turetzky2024hebdb,
  title={HebDB: a Weakly Supervised Dataset for Hebrew Speech Processing},
  author={Turetzky, Arnon and Tal, Or and Segal-Feldman, Yael and Dissen, Yehoshua and Zeldes, Ella and Roth, Amit and Cohen, Eyal and Shrem, Yosi and Chernyak, Bronya R and Seleznova, Olga and others},
  journal={arXiv preprint arXiv:2407.07566},
  year={2024}
}
  

Dataset Summary

A weakly supervised dataset for spoken language processing in the Hebrew language. HEBDB offers roughly 2500 hours of natural and spontaneous speech recordings in the Hebrew language, consisting of a large variety of speakers and topics. We provide raw recordings together with a pre-processed, weakly supervised, and filtered version. The goal of HEBDB is to further enhance research and development of spoken language processing tools for the Hebrew language.

Data variants are: pre, raw. Note variants share the same columns to ease the usage of dataset subsets but raw only use the columns: fname, audio and is_raw.

How do I download this?

Using 🤗 Datasets

from datasets import load_dataset

# pre only
hebdb_pre = load_dataset("SLPRL-HUJI/HebDB", "pre")

# raw only
hebdb_raw = load_dataset("SLPRL-HUJI/HebDB", "raw")

# One specific source(see code list below), both raw and pre
geekonomy = load_dataset("SLPRL-HUJI/HebDB", "GK")

# One specific source, both raw and pre
geekonomy_pre = load_dataset("SLPRL-HUJI/HebDB", "GK_pre")

To avoid downloading the entire dataset you can load it in streaming mode using streaming=True, for example:

hebdb_pre = load_dataset("SLPRL-HUJI/HebDB", "pre", streaming=True)

You can also load and mix:

from datasets import concatenate_datasets, load_dataset

geekonomy = load_dataset("SLPRL-HUJI/HebDB", "GK_pre")
osim_history = load_dataset("SLPRL-HUJI/HebDB", "OH_pre")

# Concatenate both datasets
concatenated = concatenate_datasets([geekonomy, osim_history])

Sources

The 6 available sources are reported in the table below.

code	name
GK	Geekonomy
OH	Osim History
DK	The Dor Kahn Experience
YO	Yo! The podcast
GQ	Good Question
YV	Yad vashem

Data Fields

The data have several fields:

• fname: file name
• audio:
  • array: array of audio samples
  • sample_rate: audio sampling rate
  • path: path to the audio file saved location
• is_raw: Flag for raw/preprocessed
• raw:
  • fname: origin raw file name
  • start_sec: start time mark in seconds
  • end_sec: end time mark in seconds
• source: Source name
• n_samples: Number of samples
• text: Transcription
• normalized_text: Normalized transcription (details in paper)
• score: Transcription quality score obtained by forced aligner (details in paper)

Licensing Information

Data is licensed under the terms of the Creative Commons Attribution 4.0 International License (CC BY 4.0), The full text of the CC-BY 4.0 license is available at https://creativecommons.org/licenses/by/4.0/.

Acknowledgements

This research work was supported by the Israel Innovation Authority, grant number 78563.