A study on two stage acoustic classification neural network training algorithm from pretrained models for small scale data environments

Training data directly impacts neural network performance during machine learning. Limited training data causes performance degradation in larger neural networks compared to simpler ones. We propose a two stage neural network method using feature extraction and classifier networks with pretrained models to address data scarcity. Performance evaluation on small scale datasets compared our method against conventional networks. Our approach achieved improved classification performance at similar complexity levels. The method demonstrated improved performance of the proposed method even with complex models where traditional training models of similar complexity typically degrade performance, showing effectiveness of the proposed method under data constraints.