Transformer parameters' effects on performance of automatic speech recognition

Rapid advancements in artificial intelligence and deep learning technologies have brought significant transformations, particularly in the fields of natural language processing and automatic speech recognition. The Transformer architecture offers superior accuracy and computational efficiency compared to traditional neural networks by effectively processing long-context data. Initially developed for text-based applications, this architecture is now widely adopted in speech recognition systems. In this thesis, key hyperparameters influencing the performance of Transformer-based automatic speech recognition models are systematically examined. The effects of parameters such as the number of neurons in feedforward networks, the depth of hidden layers, and the number of training epochs are experimentally evaluated. The model development process was conducted using the Python programming language and the Keras library, with training performed both in local environments and cloud-based platforms. Experiments utilized a large-scale English speech and text dataset sourced from a single speaker. Performance metrics obtained through various parameter configurations were thoroughly analyzed. The findings provide valuable insights into optimizing speech recognition systems developed on the Transformer architecture. In conclusion, this study reveals the impact of hyperparameter tuning on the success of Transformer-based speech recognition models and offers practical experimental data to guide researchers and practitioners in the field. Ultimately, it aims to contribute to the development of more effective and accurate speech recognition systems.