The viewer is disabled because this dataset repo requires arbitrary Python code execution. Please consider
removing the
loading script
and relying on
automated data support
(you can use
convert_to_parquet
from the datasets
library). If this is not possible, please
open a discussion
for direct help.
Dataset Card for Arabic Speech Corpus
Dataset Summary
This Speech corpus has been developed as part of PhD work carried out by Nawar Halabi at the University of Southampton. The corpus was recorded in south Levantine Arabic (Damascian accent) using a professional studio. Synthesized speech as an output using this corpus has produced a high quality, natural voice.
Supported Tasks and Leaderboards
[Needs More Information]
Languages
The audio is in Arabic.
Dataset Structure
Data Instances
A typical data point comprises the path to the audio file, usually called file
and its transcription, called text
.
An example from the dataset is:
{
'file': '/Users/username/.cache/huggingface/datasets/downloads/extracted/baebe85e2cb67579f6f88e7117a87888c1ace390f4f14cb6c3e585c517ad9db0/arabic-speech-corpus/wav/ARA NORM 0002.wav',
'audio': {'path': '/Users/username/.cache/huggingface/datasets/downloads/extracted/baebe85e2cb67579f6f88e7117a87888c1ace390f4f14cb6c3e585c517ad9db0/arabic-speech-corpus/wav/ARA NORM 0002.wav',
'array': array([-0.00048828, -0.00018311, -0.00137329, ..., 0.00079346, 0.00091553, 0.00085449], dtype=float32),
'sampling_rate': 48000},
'orthographic': 'waraj~aHa Alt~aqoriyru Al~a*iy >aEad~ahu maEohadu >aboHaA^i haDabapi Alt~ibiti fiy Alo>akaAdiymiy~api AlS~iyniy~api liloEuluwmi - >ano tasotamir~a darajaAtu AloHaraArapi wamusotawayaAtu Alr~uTuwbapi fiy Alo<irotifaAEi TawaAla ha*aA Aloqarono',
'phonetic': "sil w a r a' jj A H a tt A q r ii0' r u0 ll a * i0 < a E a' dd a h u0 m a' E h a d u0 < a b H aa' ^ i0 h A D A' b a t i0 tt i1' b t i0 f i0 l < a k aa d ii0 m ii0' y a t i0 SS II0 n ii0' y a t i0 l u0 l E u0 l uu0' m i0 sil < a' n t a s t a m i0' rr a d a r a j aa' t u0 l H a r aa' r a t i0 w a m u0 s t a w a y aa' t u0 rr U0 T UU0' b a t i0 f i0 l Ah i0 r t i0 f aa' E i0 T A' w A l a h aa' * a l q A' r n sil",
'text': '\ufeffwaraj~aHa Alt~aqoriyru Al~aTHiy >aEad~ahu maEohadu >aboHaA^i haDabapi Alt~ibiti fiy Alo>akaAdiymiy~api AlS~iyniy~api liloEuluwmi - >ano tasotamir~a darajaAtu AloHaraArapi wamusotawayaAtu Alr~uTuwbapi fiy Alo<irotifaAEi TawaAla haTHaA Aloqarono'
}
Data Fields
file: A path to the downloaded audio file in .wav format.
audio: A dictionary containing the path to the downloaded audio file, the decoded audio array, and the sampling rate. Note that when accessing the audio column:
dataset[0]["audio"]
the audio file is automatically decoded and resampled todataset.features["audio"].sampling_rate
. Decoding and resampling of a large number of audio files might take a significant amount of time. Thus it is important to first query the sample index before the"audio"
column, i.e.dataset[0]["audio"]
should always be preferred overdataset["audio"][0]
.text: the transcription of the audio file.
phonetic: the transcription in phonentics format.
orthographic: the transcriptions written in orthographic format.
Data Splits
Train | Test | |
---|---|---|
dataset | 1813 | 100 |
Dataset Creation
Curation Rationale
The corpus was created with Speech Synthesis as the main application in mind. Although it has been used as part of a larger corpus for speech recognition and speech denoising. Here are some explanations why the corpus was built the way it is:
- Corpus size: Budget limitations and the research goal resulted in the decision not to gather more data. The goal was to show that high quality speech synthesis is possible with smaller corpora.
- Phonetic diversity: Just like with many corpora, the phonetic diversity was acheived using greedy methods. Start with a core set of utterances and add more utterances which contribute to adding more phonetic diversity the most iterativly. The measure of diversity is based on the diphone frequency.
- Content: News, sports, economics, fully diacritised content from the internet was gathered. The choice of utterances was random to avoid copyright issues. Because of corpus size, acheiving diversity of content type was difficult and was not the goal.
- Non-sense utterances: The corpus contains a large set of utterances that are generated computationally to compensate for the diphones missing in the main part of the corpus. The usefullness of non-sense utterances was not proven in the PhD thesis.
- The talent: The voice talent had a Syrian dialect from Damascus and spoke in formal Arabic.
Please refer to PhD thesis for more detailed information.
Source Data
Initial Data Collection and Normalization
News, sports, economics, fully diacritised content from the internet was gathered. The choice of utterances was random to avoid copyright issues. Because of corpus size, acheiving diversity of content type was difficult and was not the goal. We were restricted to content which was fully diacritised to make the annotation process easier.
Just like with many corpora, the phonetic diversity was acheived using greedy methods. Start with a core set of utterances and add more utterances which contribute to adding more phonetic diversity the most iterativly. The measure of diversity is based on the diphone frequency.
Please refer to PhD thesis.
Who are the source language producers?
Please refer to PhD thesis.
Annotations
Annotation process
Three annotators aligned audio with phonemes with the help of HTK forced alignment. They worked on overlapping parts as well to assess annotator agreement and the quality of the annotations. The entire corpus was checked by human annotators.
Please refer to PhD thesis.
Who are the annotators?
Nawar Halabi and two anonymous Arabic language teachers.
Personal and Sensitive Information
The dataset consists of people who have donated their voice online. You agree to not attempt to determine the identity of speakers in this dataset. The voice talent agreed in writing for their voice to be used in speech technologies as long as they stay anonymous.
Considerations for Using the Data
Social Impact of Dataset
[More Information Needed]
Discussion of Biases
[More Information Needed]
Other Known Limitations
[Needs More Information]
Additional Information
Dataset Curators
The corpus was recorded in south Levantine Arabic (Damascian accent) using a professional studio by Nawar Halabi.
Licensing Information
Citation Information
@phdthesis{halabi2016modern,
title={Modern standard Arabic phonetics for speech synthesis},
author={Halabi, Nawar},
year={2016},
school={University of Southampton}
}
Contributions
This dataset was created by:
- Nawar Halabi @nawarhalabi main creator and annotator.
- Two anonymous Arabic langauge teachers as annotators.
- One anonymous voice talent.
- Thanks to @zaidalyafeai for adding this dataset.
- Downloads last month
- 281