Benchmarks

Word error rate (%), lower is better
Test set	Tasjeel	Audar-ASR-V1-Turbo	Cohere Transcribe	whisper-large-v3	gpt-transcribe	gpt-4o-transcribe	Gemini 3.5 Transcribe
Average	15.43 lowest	23.36	21.21	30.63	19.55	31.41	23.37
SADA	25.85 lowest	28.92	37.47	55.96	34.85	61.53	42.73
MGB-2	9.50 lowest	11.08	15.54	16.26	10.37	23.82	14.04
MASC Clean	14.48 lowest	16.73	19.60	24.66	16.40	18.93	18.78
MASC Noisy	23.74 lowest	27.19	27.07	34.63	25.66	35.50	29.22
Common Voice 18	5.00 lowest	8.09	5.82	17.83	7.55	14.15	16.66
Casablanca (Full)	41.28 lowest	47.02	49.71	71.81	47.71	63.03	55.96
Quran Recitation (Internal Dataset)	2.37	3.76	12.25	6.91	1.74 lowest	2.97	6.79
Arabic (Internal Dataset)	8.10 lowest	20.21	11.07	23.36	14.85	29.83	13.46
English (Internal Dataset)	6.75 lowest	27.33	9.26	12.53	9.92	24.43	10.17
Code Switched (Internal Dataset)	17.21 lowest	43.26	24.34	42.36	26.45	39.90	25.93

Word error rate (%), lower is better

Compare by test set