Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

HebDB: a Weakly Supervised Dataset for Hebrew Speech Processing

Arnon Turetzky, Or Tal, Yael Segal-Feldman, Yehoshua Dissen, Ella Zeldes, Amit Roth, Eyal Cohen, Yosi Shrem, Bronya R. Chernyak, Olga Seleznova, Joseph Keshet, Yossi Adi

TL;DR

HebDB addresses the need for a large-scale Hebrew speech resource by presenting approximately $2\,500$ hours of spontaneous Hebrew data with raw and weakly transcribed versions, including a pre-processed subset of about $1\,690$ hours (filtered to ~ $1\,470$ hours). A two-stage pre-processing and filtering pipeline—16 kHz mono resampling, silero-VAD segmentation, Whisper-based transcription, UNIKUD diacritization, Uroman transliteration for forced alignment, and a $0.3$ confidence threshold—yields usable data for model training. The authors provide two baselines, HuBERT (SSL) and Conformer (supervised), trained on HebDB and evaluated on the Hebrew FLEURS benchmark; HuBERT generally outperforms Conformer and shows competitive performance relative to multi-lingual models like Whisper and MMS at similar model sizes. This work, released under CC BY $4.0$, aims to catalyze Hebrew speech processing research and suggests future work to add higher-quality annotations and generative-data subsets for tasks such as TTS and voice conversion, thereby advancing practical Hebrew NLP and speech technologies.

Abstract

We present HebDB, 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. Hence, we additionally provide two baseline systems for Automatic Speech Recognition (ASR): (i) a self-supervised model; and (ii) a fully supervised model. We present the performance of these two methods optimized on HebDB and compare them to current multi-lingual ASR alternatives. Results suggest the proposed method reaches better results than the evaluated baselines considering similar model sizes. Dataset, code, and models are publicly available under https://pages.cs.huji.ac.il/adiyoss-lab/HebDB/.

HebDB: a Weakly Supervised Dataset for Hebrew Speech Processing

TL;DR

HebDB addresses the need for a large-scale Hebrew speech resource by presenting approximately hours of spontaneous Hebrew data with raw and weakly transcribed versions, including a pre-processed subset of about hours (filtered to ~ hours). A two-stage pre-processing and filtering pipeline—16 kHz mono resampling, silero-VAD segmentation, Whisper-based transcription, UNIKUD diacritization, Uroman transliteration for forced alignment, and a confidence threshold—yields usable data for model training. The authors provide two baselines, HuBERT (SSL) and Conformer (supervised), trained on HebDB and evaluated on the Hebrew FLEURS benchmark; HuBERT generally outperforms Conformer and shows competitive performance relative to multi-lingual models like Whisper and MMS at similar model sizes. This work, released under CC BY , aims to catalyze Hebrew speech processing research and suggests future work to add higher-quality annotations and generative-data subsets for tasks such as TTS and voice conversion, thereby advancing practical Hebrew NLP and speech technologies.

Abstract

We present HebDB, 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. Hence, we additionally provide two baseline systems for Automatic Speech Recognition (ASR): (i) a self-supervised model; and (ii) a fully supervised model. We present the performance of these two methods optimized on HebDB and compare them to current multi-lingual ASR alternatives. Results suggest the proposed method reaches better results than the evaluated baselines considering similar model sizes. Dataset, code, and models are publicly available under https://pages.cs.huji.ac.il/adiyoss-lab/HebDB/.
Paper Structure (9 sections, 2 figures, 3 tables)

This paper contains 9 sections, 2 figures, 3 tables.

Table of Contents

  1. Introduction
  2. Related work
  3. HebDB dataset
  4. Pre-processing
  5. Data filtering
  6. Baseline system
  7. Implementation details
  8. Results
  9. Conclusion & future work

Figures (2)

  • Figure 1: Score level histogram of the data filtering process. We use a threshold of 0.3 which filters roughly 13% of the data.
  • Figure 2: A boxplot of the processed data, percentages denote the corresponding portion of the unfiltered non-outlier instances with respect to each source. We provide statistics for both audio recordings (in hours) and transcriptions (in words count).