There is a broad range of languages around the globe, each of which has a variety of spoken dialects. Dialect refers to a variety in a language’s vocabulary, idioms, grammar, and pronunciation. Dialect Identification (DID) is an automatic process to detect to which dialect a speech utterance belongs. Moreover, Arabic Dialect Identification (ADI) is considered one of the most challenging DID tasks since some dialects share many vocabularies, and there is a massive overlap in the linguistic and acoustic features between the countries that belong to the same region. Furthermore, while the work is saturated on DID systems in other languages, the Arabic language still lags behind. Therefore, we present an approach that utilizes transfer learning techniques for classifying both regions and countries. We used a subset of the MGB-5 dataset, and employed Mel-Frequency Cepstral Coefficients (MFCC) as a feature extraction method, then fed the features into the ResNet-LSTM model. However, in this method, we proposed a multi-output approach that combines the classification of regions and countries into one single model, which proved to achieve better results than the single-output approach. After that, we utilized a pre-trained audio model that proved to be highly effective for language-speech tasks, called Wav2Vec2. We used the XLSR variant fine-tuned on the MGB-3 dataset, to train the model on the same small subset and it achieved an accuracy of 93.49% and 93.20% on regions and countries, respectively, which shows the potential of applying state-of-the-art pre-trained audio models on the ADI task.