Fusion of Multimodal Audio Data for Enhanced Speaker Identification Using Kolmogorov-Arnold Networks

Speaker identification using audio data is quite challenging because of inherent differences between people, ambient noise, and variable recording conditions. Although the classical deep learning methods are effective, they have rather high computational cost, which leads to usually cumbersome parameter tuning processes and hence reduce their applicability to real-world deployments. In this paper, we discuss an efficient version of Kolmogorov-Arnold Network for audio-based speaker identification using three different types of audio inputs: bone conduction, air conduction, and throat microphone recordings, alongside evaluation on the publicly available LibriSpeech dataset to enhance reproducibility. To optimize the performance of these modalities, we investigate data-level, feature-level, and decision-level fusion techniques. An exhaustive comparison is performed against some of the prominent deep learning architectures, namely VGG-19, ResNet-152, EfficientNetV2, and DenseNet-121, both as standalone architectures and within multimodal combinations for the purpose of evaluating the proposed framework. We demonstrate that our model achieves a significant increase in accuracy up to 99.375% through the use of MFCC and GFCC feature inputs with data-level fusion through MFCC features and achieves this level of accuracy using substantially fewer parameters than those of conventional deep learning models. This makes the KAN framework a highly efficient choice for audio-based speaker identification. One of the critical innovations within this design is a class of trainable activation functions that eliminate the need for traditional weight parameters, thereby providing scalable and computationally efficient solutions. This work contributes to the field of biometrics by providing a refined, resource-efficient approach to speaker identification based on multimodal audio data.