Speaker Identification Using Small Artificial Neural Network on Small Dataset
Luka Loina · 2022 International Conference on Smart Systems and Technologies (SST) · 2022
Speaker recognition provides an answer to the question “Who is speaking?”. Most of the research in the field of speaker recognition focuses on models for recognizing thousands of speakers and uses large datasets for doing so. In this paper, we explore the possibility of using small neural networks that could quickly be trained on small datasets for doing speaker recognition for a small number of speakers with high accuracy. To investigate this matter experimental analysis was conducted by using hyperparameter optimization to find the optimal combination of the structure and parameters for multiple configurations of neural networks. Furthermore, the investigation was performed with multiple preprocessing methods to find the effect of preprocessing on the size and accuracy of the resulting models. As the result of the experiment, we found a residual neural network with 387,461 parameters that was able to classify speakers with high accuracy. In comparison state of the art speaker recognition uses 4.2 million parameters.