Speaker Change Detection using Derivatives of Speech Embeddings from Bidirectional LSTM

Speaker change detection (SCD) is the fundamental task for other speech-related tasks in processing voice recordings of meetings. In this research, the Bidirectional Long Short-Term Memory (Bi-LSTM) model is used to provide synthesized features to predict speaker change from feature representations such as Mel Frequency Cepstral Coefficients and Mel Spectrogram audio features. This research explores the impact of long speaker duration in SCD within a voice meeting recording. We proposed a method to improve an existing SCD methodology that uses local maxima as a threshold to detect speaker change segments by using derivatives of Bi-LSTM predictions. Our proposed model provides a robust pipeline that uses 2 seconds as the segmentation duration which was able to achieve at most 0.76 purity along with 0.71 coverage for the voice recordings of meetings, each with the duration of 20 minutes.