Datasets
DATASETInstitutional
5/5

MGB-2 — Arabic Multi-Genre Broadcast Speech Recognition Dataset

Arabic(ara)SpeechASRWest Africa
View Source

Dataset Info

Language
Arabic (ara)
Modality
Speech
Domain
News/Broadcast
Size
1200
Institution
Ahmed Ali; Peter Bell; James Glass; Yacine Messaoui; Hamdy Mubarak; Steve Renals; Yifan Zhang (QCRI/Aljazeera)
Last Updated
20 May 2026

License & Access

SPDX identifierLDC Custom

Restricted — review the source terms before use.

RedistributionRestricted

PII RISK

Low risk

Full Record

28 of 46 fields stated

Every column the registry tracks. Values marked derivedwere filled by a rule or read from the source's text rather than stated in its metadata; the rule or the quoted sentence is shown under the value.

Registry columns (23/27)

iso code
ara
Dataset name
MGB-2 — Arabic Multi-Genre Broadcast Speech Recognition Dataset
Modality
Speech
Type
dataset
Link
https://arabicspeech.org/resources/mgb2
NLP Task
ASR
Domain
News/Broadcast
Data Format
Not stated
Size Metric
Hours
Size Value
1200
License
LDC Custom
Open Access
Restricted
Source Repository
Institutional
Year Published
2016
Creator/Author(s)
Ahmed Ali; Peter Bell; James Glass; Yacine Messaoui; Hamdy Mubarak; Steve Renals; Yifan Zhang (QCRI/Aljazeera)
Funder/Partner
DARPA/IARPA
Dialect Coverage
Dialectal/MSA
Gender Balance
Not stated
Speaker Count
Not stated
Annotation Type
Human (light supervision)
Status
Active
Has Datasheet
Yes
PII Risk
Low
Consent Verified
Not stated
Verified By
Agboola Blessing
Date Indexed
2026-05-20
Notes / Flags
1,200 hours of Arabic broadcast speech from Aljazeera TV. Lightly supervised transcriptions. Five dialects: Egyptian, Gulf, Levantine, North African, MSA. LDC C…

Quality (DQI) (4/12)

Datasheet Fields Completed
Not stated
SPDX License ID
Not stated
Redistribution Allowed?
Restricted
Content Hash
Not stated
Submitted By
Not stated
Fetch Date
2026-05-20
Last Updated Date
Not stated
Annotation Guidelines Available?
Not stated
IAA Score
Not stated
Cleaning Notes
Not stated
Benchmark Exists?
Yes
Benchmark Citation
https://huggingface.co/atishay23/whisper-small-ar

Representativeness (DRI) (1/7)

Gender Distribution
Not stated
Region Coverage
Not stated
Region Distribution
Not stated
Age Range Coverage
Not stated
Register
Not stated
Marginalized Group Note
Not stated
Demographic Fields Reported Count
0

Readiness Score Breakdown

5/5
Data Coverage100%
Model Availability100%
Benchmarks100%
Documentation72%
Recency100%

Representativeness

Gender Balance
Regional Distribution
Dialect Coverage
Inclusivity

Provenance

Source URL
https://arabicspeech.org/resources/mgb2
Fetch Timestamp
20 May 2026
Content Hash
—

Audit Log

Record indexed from Institutional

20 May 2026

Source status: Active

20 May 2026

Related Datasets

dataset

Mozilla Common Voice 24.0 — Arabic (ar) Speech Dataset

Common Voice v24.0 Arabic: 136,040 clips, 157.28 hours total, 91.74 hours validated. 1,651 speakers. Text corpus: 1,801,544 sentences. CC0 licence. Includes …

9 languages2 copiesArabic(ara)SpeechASR
OTOther (Mozilla)
Open20 May 2026
dataset

Wojood — Arabic Named Entity Recognition Corpus (WojoodFine)

550K tokens, 75K named entity mentions. Nested and fine-grained NER with 21 entity types and 47+ subtypes (WojoodFine). Also includes WojoodGaza (60K tokens,…

Arabic(ara)TextNER
HUHugging Face
Open20 May 2026
dataset

Arabic Speech Corpus (ASC) — Nawar Halabi / University of Southampton

First open-source Arabic speech corpus for TTS synthesis. Recorded in studio (Damascene/South Levantine accent). 1,813 WAV files + TextGrid phoneme labels. A…

4 copiesArabic(ara)SpeechTTS/ASR
HUHugging Face
Open20 May 2026