Models / mcit

mcit/qatari-speech-to-text

Qatari Speech To Text

PublicLimited riskOpenSpeechTransformersArabic3.0.0
11K 148K/30d 640

WER (MSA)

8.4%

WER (Qatari dialect)

13.1%

WER (code-switched)

16.5%

RTF (A10G)

0.18

About this model

Whisper large-v3 fine-tune covering Modern Standard Arabic and the Qatari dialect, optimised for call-centre audio, public consultations, and broadcast material. Handles code-switching into English tokens common in service conversations.

Intended use

Transcription of citizen-facing calls, interviews, and public hearings by government entities and licensed contact-centre operators.

Training data lineage

Fine-tuned on 1,900 hours from the khaliji-speech-corpus with a Qatari-dialect weighted sampling schedule.

Limitations & bias notes

Word error rate rises noticeably on heavy Bedouin-register speech and on overlapping speakers in majlis-style recordings. Numerals spoken in mixed Arabic-English sequences are occasionally normalised inconsistently.

Evaluation metrics

WER (MSA)8.4%
WER (Qatari dialect)13.1%
WER (code-switched)16.5%
RTF (A10G)0.18
#asr#whisper#arabic#qatari-dialect#speech#call-centre

Try it

Live sandbox

call_centre_qa_0912.wav

fixture

Feedback

Owning entity

mcMCIT
Updated
2026-06-30
Latest version
3.0.0
License
Open
Access
Open