Abstarct: Audio Event Detection and classification (AED) in the workplace and life has become a need in the modern 21st century, so that a branch of research in artificial intelligence and, consequently, in the Computational Auditory Scene Analysis (CASA) group has been dedicated to this work. This thesis offers two important approaches for audio event detection and classification in environmental audio signals. In the first approach, we proposed a model with several smaller independent components, each with a strong point for the entire model. Most existing models view all expectations of modeling in a coherent and statistical definition, but here the main model is made up of a multi-component parallel composition in order to utilize the maximum power of raw data and extract and manage their information desirable. In addition, such modeling replaces the negative side effect between the components, especially the features, with the positive side effect. The F1 measure obtained in this method was 80.6% and 72.8% for the frxame-baxsed and event-baxsed audio detection, respectively. The second approach is baxsed on the using non-negative matrix factorization (NMF) on the observation signal to capture the sample spectral tissue from sounds as atoms or patches in a dictionary, and then reconstruct the test signal by producing a sparse representation of the dictionary atoms in the representation of the segments of that signal. This method is baxsed on the separating a signal into the primary components, that is, voice events. The F1 measure obtained in this method was 56.5% and 49.2% for the frxame-baxsed and event-baxsed audio detection, respectively. These results show the superiority of the proposed methods compared to previous methods done on the DCASE2013 databaxse.