Dialect identification is exploited as a valuable tool in forensic investigations, aiding in geolocating suspects, identifying suspects from recordings, and authenticating recordings. Additionally, it assists investigators in narrowing down the search space and facilitating the recognition process. This paper aims to assess fine-grained Algerian dialect identification. Algerian speakers often engage in code-switching, seamlessly alternating between Modern Standard Arabic (MSA), Algerian Arabic, and even French within the same conversation. This linguistic fluidity adds further complexity to Algerian dialects. In this paper, we introduce Recurrent Neural Network with Long Short-Term Memory (RNN-LSTM) architecture to accurately and efficiently identify the four main Algerian accents. The classifier is trained using a fine-grained Algerian corpus, with a feature vector composed of 39 Mel-Frequency Cepstral Coefficients (MFCCs) extracted from audio signals. Experimental results demonstrate that the proposed RNN-LSTM classifier achieves interesting results with a precision of 0.79, 0.8, 0.79 and 0.79, recall of 0.83, 0.79, 0.82 and 0.69 and F1-score of 0.82, 0.79, 0.81 and 0.73 for the Central accent, Eastern accent, Western accent and Southern accent respectively. A global correct identification rate is about 79%.