SASpeech

This dataset contains 13+ hours of speech in Hebrew of single speaker in `44.1khz

The metadata.csv contains file_id|text|phonemes

Where the file_id is the name of the file in ./wav folder

The dataset is cleaned from numbers, it contains only Hebrew words.

Additional the words have Hebrew diacritics + phonetics marks (non standard) you may remove the non standard if you have use of the text.

The last column is phonemes created with phonikud

Created from https://www.openslr.org/134

License

Non commercial use only. See license in OpenSLR: https://www.openslr.org/134

Contents

The folder saspeech_manual/ contains 3 hours (~7GB) with hand annotated transcripts

The folder saspeech_automatic/ contains 12 hours (1GB) with automatic transcripts with ivrit.ai Whisper turbo and aggressive cleans (from 30 hours)

LJSpeech format

To convert the data to LJSpeech format, use the following:

import pandas as pd
df = pd.read_csv('metadata.csv', sep='\t', names=['file_id', 'text', 'phonemes'])
df[['file_id', 'phonemes']].to_csv('subset.csv', sep='|', header=False, index=False)

Resample

The dataset sample rate is 44.1khz You can resample to 22.05khz with the following:

from pydub import AudioSegment
from pathlib import Path
from tqdm import tqdm

in_dir = Path("wav")
out_dir = Path("wav_22050")
out_dir.mkdir(exist_ok=True)

for f in tqdm(list(in_dir.glob("*.wav"))):
    audio = AudioSegment.from_wav(f)
    audio = audio.set_frame_rate(22050).set_channels(1)
    audio.export(out_dir / f.name, format="wav")

Setup

uv pip install huggingface_hub
sudo apt install p7zip-full
uv run huggingface-cli download --repo-type dataset thewh1teagle/saspeech ./manual/saspeech_manual_v1.7z --local-dir .
7z x saspeech_v1.7z

Changelog saspeech manual

• v1: prepare files from manual transcript
• v2: enhance with Adobe enhance speech v2 and normalize to 22.05khz