You agree to the following license terms:
This material and 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/.

Notwithstanding the foregoing, this material and data may only be used, modified and distributed for the express purpose of training AI models, and subject to the foregoing restriction. In addition, this material and data may not be used in order to create audiovisual material that simulates the voice or likeness of the specific individuals appearing or speaking in such materials and data (a “deep-fake”). To the extent this paragraph is inconsistent with the CC-BY-4.0 license, the terms of this paragraph shall govern.

By downloading or using any of this material or data, you agree that the Project makes no representations or warranties in respect of the data, and shall have no liability in respect thereof. These disclaimers and limitations are in addition to any disclaimers and limitations set forth in the CC-BY-4.0 license itself. You understand that the project is only able to make available the materials and data pursuant to these disclaimers and limitations, and without such disclaimers and limitations the project would not be able to make available the materials and data for your use.

Overview

This dataset provides full, machine-generated transcriptions for the entire audio-v2 dataset, containing >20k hours of Hebrew audio, all licensed under the ivrit.ai v1 license. It was released on May 18th, 2025.

You can find the full list of sources in this dataset under the audio-v2 dataset's sources.txt. All files were transcribed using the process.py pipeline, performing:

Frame-level VAD
Machine transcription using ivrit.ai's whisper-large-v3-turbo engine with the 2025.04.03 label

The VAD operation is used to reduce hallucinations in long silence durations.

Fetching the data

Use git to fetch the data. You will be asked for your username/password multiple times; use yoru token for both.

Command:

git clone https://huggingface.co/datasets/ivrit-ai/audio-v2-transcripts

Repository layout

Each episode contains two files:

full_transcript.json.gz: full transcript of source audio, including segment information, timestamps, avg_logprob etc.
speech_probs.frame.gz: speech probability per time period.

and in graphical form:

repository/
└── top/
    ├── podcast1/
    │   ├── episode1/
    │   │   ├── full_transcript.json.gz
    │   │   └── speech_probs.frame.gz
    │   ├── episode2/
    │   │   ├── full_transcript.json.gz
    │   │   └── speech_probs.frame.gz
    │   └── ...
    ├── podcast2/
    │   ├── episode1/
    │   │   ├── full_transcript.json.gz
    │   │   └── speech_probs.frame.gz
    │   ├── episode2/
    │   │   ├── full_transcript.json.gz
    │   │   └── speech_probs.frame.gz
    │   └── ...
    └── ...

Source audio is available under the audio-v2 repository, under podcast/episode.mp3 or podcast/episode.m4a.

Data usage

The full_transcript.json.gz contains, among others:

Full text transcript ('text' key).
Text segments ('segments'). Each segment contains start/end times, text, and other data provided by Whisper per-segment.
Original data prior to stable-ts based alignment ('ori_dict').

You can access it as follows:

import json
import gzip

>>> transcript = json.load(gzip.GzipFile('גיקונומי/2024.12.18 פרק #982 - פרופ׳ אייל זיסר וסוריה/full_transcript.json.gz'))
))
>>>
>>> len(transcript['text'])
57372
>>> len(transcript['segments'])
2183
>>> print(transcript['segments'][0])
{'start': 0.26, 'end': 2.18, 'text': ' גיקונומי ה-16 בדצמבר,', 'seek': 2674, 'tokens': [], 'temperature': 0.0, 'avg_logprob': -0.09050864100139192, 'compression_ratio': 2.0642857142857145, 'no_speech_prob': 0.0, 'words': [{'word': ' גיקונומי', 'start': 0.26, 'end': 0.88, 'probability': 0.92158203125, 'tokens': None}, {'word': ' ה', 'start': 0.88, 'end': 1.02, 'probability': 0.5625, 'tokens': None}, {'word': '-16', 'start': 1.06, 'end': 1.36, 'probability': 0.997314453125, 'tokens': None}, {'word': ' בדצמבר,', 'start': 1.36, 'end': 2.18, 'probability': 0.99990234375, 'tokens': None}]}

Credits

Paper: https://arxiv.org/abs/2307.08720

If you use our datasets, the following quote is preferable:

@misc{marmor2023ivritai,
      title={ivrit.ai: A Comprehensive Dataset of Hebrew Speech for AI Research and Development}, 
      author={Yanir Marmor and Kinneret Misgav and Yair Lifshitz},
      year={2023},
      eprint={2307.08720},
      archivePrefix={arXiv},
      primaryClass={eess.AS}
}

Downloads last month: 42