Speech refinement using Bi-LSTM and improved spectral clustering in speaker diarization

In this digitally-driven culture, the need and demand for diarizing online meetings, classes, conferences, and medical diagnoses have increased a lot. Speaker Diarization, a sub-domain of Speaker Recognition has grown with the advent of neural networks in the last decade. Diarize generally refers to obtaining the duration of individual speakers in any event. Researchers have suggested various approaches for multiple-speaker diarization. However, it still suffers from a problem of various environmental noises, and non-speech sounds like laughter, murmuring, clapping, etc. in the datasets. Hence, this paper proposes an improved speaker diarization pipeline to deal with the noise present in a dataset having multiple speakers. This improved diarization pipeline uses Bi-directional Long Short-Term Memory (Bi-LSTM), based speech refinement pre-processing module, and Modified Spectral Clustering with Symmetrized Singular Value Decomposition (MSC-SSVD). MSC-SSVD is used to cater to the problem of spectral clustering in large datasets. The proposed diarization pipeline is evaluated using the publicly available VoxConverse dataset. The Diarization Error Rate (DER) obtained after experimentation are 37.2%, 37.1%, and 43.3% respectively for three batches of dataset under study. The results are also compared with the baseline system and significant change in DER by 6.1%, 4.7%, and 7% respectively for three batches is observed.