RoboShaulTiny is a subset of RoboShaul

SASpeech Dataset - automatic data readme

This dataset contains approximately 30 hours of audio spoken by Shaul Amsterdamski in a recording studio with corresponding machine-generated transcriptions.

There is no overlap with the gold-standard subset.

The Data

Data Description

The audio is all sampled at 44100Hz and was divided by searching for silences in the source recordings. The shortest files are around 1 second long, and the longest is slightly over 3 minutes.

The transcriptions were creating by running OpenAI Whisper large-v2 model on every file after splitting files by silences. Automated diacritics (Nikud) were added using the open-source neural net Nakdimon (paper: Restoring Hebrew Diacritics Without a Dictionary).

In this case, each audio file may contain multiple utterances. The reason for this is that timing data generated by Whisper is not always accurate, so the user must decide how to separate each utterance from the audio file.

These recordings were done over a period of multiple years. This creates a few oddities, for example:

• Some source files were compressed as MP3 flies
• Most of the speech is reading from a script, some is spontaneous speech
• Recording environments may be different between files
• The speaker's voice changed over time

Extra metadata is provided to allow users to differentiate between different source recording sessions for utternaces in the dataset.

Data Organization

The dataset is organized as follows:

wavs/
metadata.csv
README.md

Where metadata.csv is a CSV with the separator |, with the following structure per row:

<utterance id>|<file id>|<start time (secs)>|<end time>|<transcription>|<source file id>

The file wavs/<file id>.wav is the corresponding audio. Source file ID is an identifier for the source file from which the utterances were extracted. This is useful because there is a 1-to-1 relationship between the source file ID and a recording session. Different recording sessions may have different properties as described in the previous chapter.

Notes

If you wish to remove diacritics, it is achievable by using the hebrew library on PyPI (pip install hebrew). Example usage:

>>> from hebrew import Hebrew
>>> s_nikud = 'אַתֶּם ×”Ö¶×Ö±×–Ö·× Ö°×ªÖ¼Ö¶× לִחְיוֹת כִּיס'
>>> Hebrew(s_nikud).no_niqqud().string
'אתם ×”××–× ×ª× לחיות כיס'

License

By using this Dataset, you accept the terms and conditions in the license agreement (Hebrew-only). The license is also provided with the dataset. In case of conflict between the attached license and the version available online, the online version takes precedence.

A summary of the terms in English:

Copyright for the recordings and corresponding transcriptions is owned solely by the Israeli Public Broadcast Corporation, the IPBC.

The dataset is free for use for non-commercial purposes, under the following limitations, whether by positive act or by omission:

• You may not present your use of the Dataset in a way that suggests that the IPBC supports or endorses you or your use of the Dataset
• You may not make use of the Dataset in a manner that brings harm to Shaul Amsterdamski and/or the IPBC, including defamation
• You may not make use of the Dataset for commercial or broadcast needs
• You may not make use of the Dataset for political needs
• You may not make use of the Dataset in a manner that breaches any applicable law

Disclaimer

The Dataset is licensed "AS IS", and the IPBC excludes all representations, warranties, obligations and liabilities, whether explicit or implied, in relation to the Dataset, including as regards the suitability of the Dataset for your use.

The IPBC is not liable for any error or omission in the Dataset and shall not be held liable for any loss, injury or damage of any kind, whether direct or indirect, caused by its use to you or to third parties.

The IPBC is not liable for changes made to the Dataset by you or by any third party.

Credits & Contact

The dataset contains recordings of Shaul Amsterdamski, who volunteered his voice for this dataset.

The Israeli Public Broadcasting Corporation generously allowed the release of the dataset under the attached terms and conditions.

Tamir Cohen of the IPBC provided the technical support to make the data available.

Uri Eliyabayev, the manager of MDLI, was instrumental in making this project come together.

Data was processed and prepared by Orian Sharoni and Roee Shenberg of Up·AI.

For any questions regarding the data, contact us at [email protected].