sadeem-ai/khaliji-asr

Khaliji ASR

PublicLimited riskRev-ShareSpeechTransformersArabic2.2.0

Test in Sandbox

7.4K 210K/30d 460

WER (Gulf blind set)

11.2%

WER (Qatari)

12.4%

Dialect ID accuracy

0.93

RTF (A10G)

0.14

About this model

Conformer-based speech recogniser purpose-built for Gulf dialects, Qatari, Emirati, Kuwaiti, Bahraini, and eastern Saudi, with automatic dialect identification and MSA fallback. Winner of the Khaliji ASR Challenge 2025 on the blind Gulf-dialect test set.

Intended use

Transcription of dialectal audio in contact centres, media monitoring, and field research where Gulf-dialect accuracy matters more than broad multilingual coverage.

Training data lineage

Trained on 3,400 hours from the khaliji-speech-corpus including the challenge training split and licensed broadcast material.

Limitations & bias notes

Performance on North African and Levantine dialects is well below Gulf-dialect levels; route such audio to a general model. Very short utterances under two seconds show elevated deletion errors.

Evaluation metrics

WER (Gulf blind set)	11.2%
WER (Qatari)	12.4%
Dialect ID accuracy	0.93
RTF (A10G)	0.14

#asr#gulf-dialect#khaliji#speech#conformer#challenge-winner

Try it

Live sandbox

call_centre_qa_0912.wav

fixture

Feedback

Owning entity

saSadeem AI

Updated: 2026-07-17
Latest version: 2.2.0
License: Revenue Share
Access: Open

Trained on

khaliji-speech-corpus →