Automatic speech recognition (ASR) has seen significant improvements with the advent of deep learning and end-to-end models based on Transformer architectures. Arabic ASR has remained a challenging task due to the language’s complexity, especially in terms of its dialectal variety. We explore the performance of several state-of-the-art ASR models on a large-scale Arabic speech dataset – the SADA (Saudi Audio Dataset for Arabic) comprising 668 h of high-quality audio. The dataset includes multiple dialects and environments, and specifically a noisy subset; all of these make it particularly challenging for ASR. We evaluate the performance of the models on the SADA test set, and we explore the impact of finetuning, language models, as well as noise and denoising on their performance. We find that the best performing model is the MMS 1B model finetuned on SADA with a 4-gram language model that achieves a WER of 40.9% and a CER of 17.6% on the SADA test clean set. We find that the best path towards improving the performance of the models in noise is finetuning them on the noisy data, with denoising adversely impacting performance overall and only leading to improvement the noisiest samples.