Speaker Diarization Using X-vectors-DNN Framework
Sanka Sreelakshmi, Devika Jayaram, Varsha R Ajith, Joel Jacob Mathew, Rajeev Rajan · 2023
The speaker diarization algorithm analyses an audio conversation and labels it with different speakers present for speaker turn. In this paper, a speaker diarization system using the x-vector-DNN framework is proposed. Conventionally, the x-vector-based system uses clustering to segment the speech based on the speaker’s turns. But we replace the clustering phase with a DNN-based supervised scheme. X-vectors of the individual speakers are used to train the DNN model. Augmentation schemes are also utilised for the efficient training of the model. During testing, the trained model segments the speech by computing the x-vectors. The performance is evaluated using the AMI corpus. Three performance metrics, namely diarization error rate (DER), purity, and coverage, are used for the performance evaluation. The performance is also compared with that of an attention-based model. The model shows the promise of the supervisory model in the speaker diarization task.