linagora/linto-dataset-audio-ar-tn-augmented

LinTO DataSet Audio for Arabic Tunisian Augmented
A collection of Tunisian dialect audio and its annotations for STT task

This is the augmented datasets used to train the Linto Tunisian dialect with code-switching STT linagora/linto-asr-ar-tn.

• Dataset Summary
• Dataset composition
  • Sources
  • Content Types
  • Languages and Dialects
• Example use (python)
• License
• Citations

Dataset Summary

The LinTO DataSet Audio for Arabic Tunisian Augmented is a dataset that builds on LinTO DataSet Audio for Arabic Tunisian, using a subset of the original audio data. Augmentation techniques, including noise reduction and SoftVC VITS Singing Voice Conversion (SVC), have been applied to enhance the dataset for improved performance in Arabic Tunisian Automatic Speech Recognition (ASR) tasks.

Dataset Composition:

The LinTO DataSet Audio for Arabic Tunisian Augmented comprises a diverse range of augmented audio samples using different techniques. Below is a breakdown of the dataset’s composition:

Sources

subset	audio duration	labeled audio duration	# audios	# segments	# words	# characters
ApprendreLeTunisienVCA	2h 40m 6s	2h 40m 6s	6146	6146	8078	36687
MASC_NoiseLess	2h 49m 56s	1h 38m 17s	48	1742	11909	59876
MASC_NoiseLess_VCA	19h 49m 31s	11h 27m 59s	336	12194	83377	411999
OneStoryVCA	9h 16m 51s	9h 7m 32s	216	2964	73962	341670
TunSwitchCS_VCA	59h 39m 10s	59h 39m 10s	37639	37639	531727	2760268
TunSwitchTO_VCA	18h 57m 34s	18h 57m 34s	15365	15365	129304	659295
Youtube_AbdelAzizErwi_VCA	122h 51m 1s	109h 32m 39s	125	109700	657720	3117170
Youtube_BayariBilionaireVCA	4h 54m 8s	4h 35m 25s	30	5400	39065	199155
Youtube_DiwanFM_VCA	38h 10m 6s	28h 18m 58s	252	32690	212170	1066464
Youtube_HkeyetTounsiaMensia_VCA	12h 13m 29s	9h 53m 22s	35	10626	73696	360990
Youtube_LobnaMajjedi_VCA	6h 41m 38s	6h 12m 31s	14	6202	42938	211512
Youtube_MohamedKhammessi_VCA	12h 7m 7s	10h 58m 21s	14	12775	92512	448987
Youtube_Shorts_VCA	26h 26m 25s	23h 45m 25s	945	14154	201138	1021713
Youtube_TNScrapped_V1_NoiseLess	4h 2m 9s	2h 31m 05s	52	2538	18777	92530
Youtube_TNScrapped_V1_NoiseLess_VCA	28h 15m 1s	17h 37m 36s	364	17766	132587	642292
Youtube_TV_VCA	4h 16m 16s	3h 40m 56s	28	4676	33376	311500
TOTAL	373h 10m 28s	320h 36m 58s	61609	292257	2342336	11742108

Data Proccessing:

• Noise Reduction: Applying techniques to minimize background noise and enhance audio clarity for better model performance. For this, we used Deezer Spleeter, a library with pretrained models, to separate vocals from music.
• Voice Conversion: Modifying speaker characteristics (e.g., pitch) through voice conversion techniques to simulate diverse speaker profiles and enrich the dataset. For this, we chose SoftVC VITS Singing Voice Conversion (SVC) to alter the original voices using 7 different pretrained models.

The image below shows the difference between the original and the augmented audio:

• The first row shows the original waveform.
• The second row shows the audio after noise reduction.
• The last row shows the audio with voice conversion augmentation.

Content Types

• FootBall: Includes recordings of football news and reviews.
• Documentaries: Audio from documentaries about history and nature.
• Podcasts: Conversations and discussions from various podcast episodes.
• Authors: Audio recordings of authors reading or discussing different stories: horror, children's literature, life lessons, and others.
• Lessons: Learning resources for the Tunisian dialect.
• Others: Mixed recordings with various subjects.

Languages and Dialects

• Tunisian Arabic: The primary focus of the dataset, including Tunisian Arabic and some Modern Standard Arabic (MSA).
• French: Some instances of French code-switching.
• English: Some instances of English code-switching.

Characteristics

• Audio Duration: The dataset contains more than 317 hours of audio recordings.
• Segments Duration: This dataset contains segments, each with a duration of less than 30 seconds.
• Labeled Data: Includes annotations and transcriptions for a significant portion of the audio content.

Data Distribution

• Training Set: Includes a diverse range of augmented audio with 5 to 7 different voices, as well as noise reduction applied to two datasets.

Example use (python)

• Load the dataset in python:

from datasets import load_dataset

# dataset will be loaded as a DatasetDict of train and test
dataset = load_dataset("linagora/linto-dataset-audio-ar-tn-augmented")

Check the containt of dataset:

example = dataset['train'][0] 
audio_array = example['audio']["array"]
segments = example['segments']
transcription = example['transcript']

print(f"Audio array: {audio_array}")
print(f"Segments: {segments}")
print(f"Transcription: {transcription}")

Example

Audio array: [0. 0. 0. ... 0. 0. 0.]
Transcription: أسبقية قبل أنا ما وصلت خممت فيه كيما باش نحكيو من بعد إلا ما أنا كإنطريبرنور كباعث مشروع صارولي برشا مشاكل فالجستين و صارولي مشاكل مع لعباد لي كانت موفرتلي اللوجسيل ولا اللوجسيل أوف لنيه ولا لوجسيل بيراتي
segments: [{'end': 14.113, 'start': 0.0, 'transcript': 'أسبقية قبل أنا ما وصلت خممت فيه كيما باش نحكيو من بعد إلا ما أنا كإنطريبرنور كباعث مشروع صارولي برشا مشاكل فالجستين و صارولي مشاكل مع لعباد لي كانت موفرتلي اللوجسيل ولا اللوجسيل أوف لنيه ولا لوجسيل بيراتي'}]

License

Given that some of the corpora used for training and evaluation are available only under CC-BY-4.0 licenses, we have chosen to license the entire dataset under CC-BY-4.0.

Citations

When using the LinTO DataSet Audio for Arabic Tunisian corpus, please cite this page:

@misc{linagora2024Linto-tn,
  author = {Hedi Naouara and Jérôme Louradour and Jean-Pierre Lorré},
  title = {LinTO Audio and Textual Datasets to Train and Evaluate Automatic Speech Recognition in Tunisian Arabic Dialect},
  year = {2024},
  publisher = {HuggingFace},
  journal = {HuggingFace},
  howpublished = {\url{https://huggingface.co/datasets/linagora/linto-dataset-audio-ar-tn}},
}

@misc{abdallah2023leveraging,
      title={Leveraging Data Collection and Unsupervised Learning for Code-switched Tunisian Arabic Automatic Speech Recognition}, 
      author={Ahmed Amine Ben Abdallah and Ata Kabboudi and Amir Kanoun and Salah Zaiem},
      year={2023},
      eprint={2309.11327},
      archivePrefix={arXiv},
      primaryClass={eess.AS}
}

@data{e1qb-jv46-21,
doi = {10.21227/e1qb-jv46},
url = {https://dx.doi.org/10.21227/e1qb-jv46},
author = {Al-Fetyani, Mohammad and Al-Barham, Muhammad and Abandah, Gheith and Alsharkawi, Adham and Dawas, Maha},
publisher = {IEEE Dataport},
title = {MASC: Massive Arabic Speech Corpus},
year = {2021} }