In this research, we propose the pre-trained intra- and inter-framework for the cocktail party effect. Initially, all intra- and inter-architectures are frozen. The framework undergoes retraining in three stages, during which the learning rates are adjusted using a scheduling mechanism, and the layers of the framework are gradually unfrozen according to the schedule. This process allows the framework to be re-educated, refined, and refreshed based on prior knowledge. It proves to be an efficient method for enhancing the performance of the advanced model while considerably decreasing both the time and cost required for learning. This approach proves highly beneficial for leveraging existing models to tackle similar tasks or augmenting the performance of the model. Within this approach, the pre-trained architecture surpasses the non-pre-trained architecture as subsequent phases of the model’s learning repurpose features gleaned from the earlier training phases, enhancing overall performance. The proposed approach undergoes testing and evaluation using a standard dataset. Experimental results indicate that the methodology exhibits superior performance compared to the baseline approach and surpasses existing frameworks for the cocktail party effect.