A Comparative Analysis of Traditional and Deep Learning Approaches for Adressing Challenges in Speaker Diarization

Speaker diarization is the task of distinguishing and segmenting speech from multiple speakers in an audio recording, a crucial task for various applications such as meeting transcription, voice activated systems, and audio indexing. Traditional clustering-based methods have been widely used, but they struggle with challenges in real-world scenarios, including noisy environments, overlapping speech, speaker variability and variable recording conditions. This study addresses these limitations by focusing on deep learning-based approaches, which have demonstrated significant advancements in improving the accuracy of multi-speaker diarization. The aim of this study is to compare traditional clustering methods with new deep learning techniques, including Time Delay Neural Networks (TDNN), End-to-End Neural Diarization (EEND), and the Fully Supervised UIS-RNN, to solve the challenges of multi-speaker diarization. The results show that on the CallHome dataset, TDNN systems indicated slight improvements in non-overlapping speech, with a Diarization Error Rate (DER) of 12-14%, in comparison to 13-15% for traditional clustering methods. However, in overlapping speech, EEND outperformed traditional methods, achieving a DER of 12.6%, which was significantly lower than the 23.7% observed with traditonal clustering. The Fully Supervised UIS-RNN model delivered the best overall performance, achieving a DER of 7.6%. Future research should focus on integrating the strengths of traditional and deep learning techniques while reducing the computational and data requirements for more accessible, real-time speaker diarization systems. The findings indicated that deep learning will make a substantial contribution to the field of speaker diarisation.