This dataset is derived from public audio for research. By requesting access you agree to use it for research purposes, to respect the rights of the original content owners, and to honour removal requests.

Kejue Arabic Speech

A large, multi-dialect Arabic speech corpus with transcriptions, dialect labels and word-level timing — organized by dialect so any variety is a one-line load. Gated: request access and agree to the research-only terms.

Overview

7,732,647 clips · ~11,920 hours of dialect-tagged Arabic speech
21 dialect configs live (country-level dialects + MSA ar + code-switching ar-x) — filter to any dialect without downloading the rest
Each clip ships raw 16 kHz audio + a denoised copy, transcript, dialect + sub-dialect, code-switch flags, per-word timestamps, quality signals and full provenance
7,642,929 clips pass the quality gates · 4,367,155 code-switch with another language

Usage

from datasets import load_dataset, concatenate_datasets, get_dataset_config_names

# One dialect (streams the ar-eg/ shards only)
ds = load_dataset("KejueAI/kejue-common-voice", "ar-eg")

# Modern-Standard / undetermined Arabic
ds_msa = load_dataset("KejueAI/kejue-common-voice", "ar")

# Code-switching (Arabic mixed with any other language)
ds_cs = load_dataset("KejueAI/kejue-common-voice", "ar-x")

# The FULL corpus, each clip exactly once. Every dialect config is a disjoint
# partition; the code-switch config `ar-x` is skipped because its
# clips already live in their own dialect config (loading it too would double-count them).
configs = [c for c in get_dataset_config_names("KejueAI/kejue-common-voice") if c != "ar-x"]
full = concatenate_datasets([load_dataset("KejueAI/kejue-common-voice", c, split="train") for c in configs])

# High-quality subset: keep only clips that PASSED the quality gates, dropping the
# low-quality and uncertain ones. `quality_ok` is True (passed), False (flagged), or
# null (unsure); `is True` keeps only the ones we're confident are clean.
good = full.filter(lambda r: r["quality_ok"] is True)

Each config name is a BCP-47 dialect code, so filtering by ar, ar-eg, ar-sa, … is just the second argument to load_dataset. For the whole corpus, concatenate every config except ar-x — the dialect folders already contain those clips, so including it duplicates them.

Dialect configs

Config	Dialect	Clips	Hours
`ar-x`	Arabic code-switching (Arabic mixed with any other language)	4,367,155	0.0
`ar-sa`	Saudi	1,544,138	2465.6
`ar-jo`	Jordanian	828,408	888.1
`ar-eg`	Egyptian	665,975	1252.8
`ar-kw`	Kuwaiti	593,377	742.5
`ar-bh`	Bahraini	590,110	823.0
`ar-iq`	Iraqi	408,741	823.5
`ar-lb`	Lebanese	390,787	652.1
`ar-tn`	Tunisian	382,604	602.6
`ar-ps`	Palestinian	343,861	524.6
`ar-ma`	Moroccan	335,259	661.8
`ar-sy`	Syrian	280,744	452.3
`ar-ae`	Emirati	265,853	367.8
`ar-sd`	Sudanese	239,635	305.1
`ar-ye`	Yemeni	201,193	275.0
`ar-dz`	Algerian	176,536	356.2
`ar`	Arabic (MSA / undetermined)	148,091	220.2
`ar-ly`	Libyan	134,315	223.9
`ar-om`	Omani	123,735	169.9
`ar-qa`	Qatari	76,440	108.4
`ar-td`	ar-td	1,243	2.0

ar-x re-lists every code-switching clip on top of its own dialect config, so it can be loaded as a standalone subset (Arabic mixed with any other language).

Schema

Column	Type	Description
`id`	string	unique clip id (`{video_id}_{idx}`)
`audio`	Audio	raw 16 kHz mono clip audio
`audio_denoised`	Audio	GTCRN noise-cancelled copy (may be empty)
`text`	string	transcript
`continuation`	string	LLM-generated continuation of `text` in the same dialect and register (speech-continuation training target; may be empty)
`dialect`	string	country-level BCP-47 dialect (`ar-EG`, `ar-SA`, …)
`sub_dialect`	string	finer variety within the country (Najdi / Cairene / …)
`code_switching`	bool	true if Arabic is mixed with another language
`code_switch_languages`	list	non-Arabic languages present
`language`	string	dominant clip language
`languages`	list	all languages detected in the clip
`source`	string	`youtube` or the origin Hugging Face dataset
`source_kind`	string	`youtube`
`hf_dataset`	string	origin HF dataset id (for HF-sourced clips)
`video_id`	string	source video / shard id
`idx`	int32
`url`	string	source URL
`channel`	string	source channel / uploader
`license`	string	source license (when known)
`audio_r2_key`	string	object key of the original clip in R2 (for authorized re-fetch)
`audio_denoised_r2_key`	string	object key of the denoised clip in R2
`speaker`	string	diarized speaker label (within the video)
`speaker_id`	string	globally-unique speaker id
`duration_ms`	int32	speech duration
`start_ms`	int32	speech start on the source timeline
`end_ms`	int32	speech end on the source timeline
`audio_start_ms`	int32	padded cut start
`audio_end_ms`	int32	padded cut end
`pad_ms`	int32	padding applied around the speech span
`avg_confidence`	float64	mean ASR confidence
`speech_ratio`	float64	VAD speech fraction
`quality_ok`	bool	passed the quality gates
`quality_flags`	list	reasons a clip was flagged
`denoised`	bool	a denoised copy accompanies this clip
`sample_rate`	int32	audio sample rate (16000)
`channels`	int32	audio channels (1)
`words`	list	per-word transcript with timestamps + confidence
`pipeline_version`	int32	producing pipeline version
`soniox_model`	string	ASR model
`gemini_model`	string	dialect-tagging model

Methodology

Transcription — Soniox ASR on the raw audio, chunked into word-aligned clips.
Denoising — a GTCRN copy of every clip is published alongside the raw audio (audio_denoised).
Dialect tagging — a flagship LLM labels each clip's country-level dialect (never a regional bucket) plus a finer sub-dialect and Arabic↔other-language code-switching.
Quality gates — confidence + VAD speech-ratio thresholds flag low-quality clips (quality_ok, quality_flags); flagged clips are only published after human approval.

Governance

Derived from public audio for research use only. Access is gated. To request removal of any content, contact the dataset owners via the Hugging Face community tab; takedown requests are honoured. Provenance (source, video_id, url, license) is retained per clip.

Removed: overlap with the Open Universal Arabic ASR Leaderboard test sets

On 2026-10-14 we removed 4,016 rows (9.5 h) from 651 of the 25210 parquet files. Each removed row overlaps the test material of the Open Universal Arabic ASR Leaderboard (test sets with hits: casablanca, casablanca_validation, common_voice, masc_clean, masc_noisy, mgb2, sada).

How rows were matched: source ids, normalised text, and audio fingerprints on 100% of the rows (contamination guard version 2, arabic-english-stt #8).
Row ids: the removed rows are listed in contamination/removed_ids.txt.
Counts per test set: in contamination/summary.json.
Unchanged files: all other files are unchanged.
Earlier versions: earlier revisions of this repo keep the original files.

Downloads last month: 545

Total file size:

3.57 TB

Collection including Arabic-STT/kejue-common-voice

Keju

Collection

7 items • Updated 11 days ago