Transfer Learning for Speaker Diarization on Bangla Audio Dataset

Abstract

Sound classification is a very intriguing concept in the field of artificial intelligence. Speaker Diarization is a very interesting task in the domain of sound classification. It has recently expanded significantly with the introduction of deep learning technology, which has transformed research and practices throughout speech application fields. Speaker diarization is the process of assigning labels to audio data that match the speaker's identity. It is quite beneficial when it comes to identifying audio information. In Bangla, very little work has been done on speaker diarization. This project aimed to build a Bangla dataset for the diarization process in this research. We described our deep learning model for the speaker diarization problem and demonstrated how transfer learning can be utilized to swiftly learn a model with minimal performance loss when compared to a fully trained one. To increase the universal applicability of our model, we focused on transfer learning and tweaked it manually across AMI and Bangla dataset. Additionally, we've been focusing our efforts on improving the Diarization Error Rate (DER) and experimenting with other embedding generation networks. We obtained a DER score of 0.24 using our transfer learning variation trained on Bangla dataset.