Implementation of a D-vector based speaker diarization system using hybrid voice activity detection

In today's world with rapid technological developments, systems that use machines and software instead of humans are increasing day after day. These systems need to be developed in the area of Digital Speech Processing (DSP), as in many other fields. Speaker Diarization, one of the DSP applications, requires automatic extraction of "who spoke when" from an audio recording containing speech. Developing a speaker diarization system working with high performance is still one of the challenging issues for researchers studying in this area. In order to develop speaker diarization systems with lower error rates, sub-systems such as Speech Pre-processing, Voice Activity Detection, Speaker Segmentation and Speaker Clustering, which constitute a speaker diarization system, need to be improved. In this thesis, it is aimed to design a speaker diarization system with low error rate by developing a hybrid model that has not been proposed before for the voice activity detection system which is one of the stages of speaker dialization systems. In hybrid voice activity detection system where supervised and unsupervised learning is combined with logical operators, feature thresholding was used for unsupervised learning while long-short term memory (LSTM), a deep learning architecture, was utilized for supervised learning. In the continuation of the speaker dialization system, d-vectors were extracted from a pre-trained artificial neural network, and after Spectral Clustering was applied on these vectors, "who spoke when" was detected in the audio recording. At the evalution phase of the proposed speaker diarization system, Miss and False Alarm (FA) metrics, which can be occurred due to used Voice Activity Detectors in Speaker Diarization Systems, were interpreted in detail. It was observed that using an Hybrid VAD in diarization systems has achieved low Miss and False Alarm (FA) error rate.