kadirnar/arabic-arena-ar

Arabic Arena: STT benchmark sets (ar)

These are the evaluation sets of Arabic Arena, a multi-dataset WER/CER benchmark for speech-to-text with simulated telephony and call-center conditions.

Each config is a deterministic subset of a public test set: at most 1,000 utterances of 1–30 s, chosen by hash rank. Degraded configs (-noisy, -telephony, -callcenter) are generated with seeded, recorded parameters; see meta.degradation in every row and METHODOLOGY.md.

Columns: id, audio, sampling_rate, text (raw reference), duration, meta (JSON).

from datasets import load_dataset
ds = load_dataset("kadirnar/arabic-arena-ar", "sada-callcenter", split="test")

Config	Dialect	Domain	Condition	Utterances	Hours	Licence	Source
fleurs_ar-clean	MSA	read	original audio (16 kHz)	428	1.30	CC-BY-4.0	google/fleurs
cv18_ar-clean	MSA	read (crowd-sourced)	original audio (16 kHz)	1000	1.21	CC0-1.0	MohamedRashad/common-voice-18-arabic
masc_clean-clean	mixed	youtube	original audio (16 kHz)	1000	1.18	CC-BY-4.0	MohamedRashad/MASC-Arabic
sada-clean	Gulf (Saudi)	tv shows	original audio (16 kHz)	1000	1.87	CC-BY-NC-SA-4.0	MohamedRashad/SADA22
mixat_cs-clean	Emirati + English	podcast	original audio (16 kHz)	1000	3.68	CC-BY-NC-SA-4.0	sqrk/mixat-tri
masc_noisy-clean	mixed	youtube (noisy)	original audio (16 kHz)	1000	0.96	CC-BY-4.0	MohamedRashad/MASC-Arabic
fleurs_ar-noisy	MSA	read	+ background noise 0–15 dB SNR (16 kHz)	300	0.91	CC-BY-4.0	google/fleurs
cv18_ar-noisy	MSA	read (crowd-sourced)	+ background noise 0–15 dB SNR (16 kHz)	300	0.36	CC0-1.0	MohamedRashad/common-voice-18-arabic
sada-noisy	Gulf (Saudi)	tv shows	+ background noise 0–15 dB SNR (16 kHz)	300	0.62	CC-BY-NC-SA-4.0	MohamedRashad/SADA22
fleurs_ar-telephony	MSA	read	8 kHz phone channel: band-pass, codec, packet loss	300	0.91	CC-BY-4.0	google/fleurs
cv18_ar-telephony	MSA	read (crowd-sourced)	8 kHz phone channel: band-pass, codec, packet loss	300	0.36	CC0-1.0	MohamedRashad/common-voice-18-arabic
sada-telephony	Gulf (Saudi)	tv shows	8 kHz phone channel: band-pass, codec, packet loss	300	0.62	CC-BY-NC-SA-4.0	MohamedRashad/SADA22
fleurs_ar-callcenter	MSA	read	noise + reverb at the caller, then the phone channel (8 kHz)	300	0.91	CC-BY-4.0	google/fleurs
cv18_ar-callcenter	MSA	read (crowd-sourced)	noise + reverb at the caller, then the phone channel (8 kHz)	300	0.36	CC0-1.0	MohamedRashad/common-voice-18-arabic
sada-callcenter	Gulf (Saudi)	tv shows	noise + reverb at the caller, then the phone channel (8 kHz)	300	0.62	CC-BY-NC-SA-4.0	MohamedRashad/SADA22

Licences. Every config keeps the licence of its source dataset. NC = non-commercial only.

Sets not re-hosted

The following sets are part of the benchmark, but their licences do not allow redistribution. Rebuild them locally with scripts/build_benchmark.py; the exact ids are listed in manifests/.

Set	Source	Licence
mgb2	MohamedRashad/mgb2-arabic	QCRI MGB-2 agreement
casablanca	UBC-NLP/Casablanca	CC-BY-NC-ND-4.0
mgb3_egy	ArabicSpeech/MGB-3	unspecified
mgb5_mor	ArabicSpeech/MGB-5	unspecified
algerian_cs	oist/algerian-code-switched-asr-eval	controlled-access terms
nadi_robust	UBC-NLP/NADI2026_subtask1.1_Robust_ASR	unspecified