Combination of deep speaker embeddings for diarisation

Significant progress has recently been made in speaker diarisation after the introduction of d-vectors as embeddings extracted from neural network (NN) classifiers for clustering speech segments. To extract better-performing and more robust embeddings, this paper proposes a c-vector method by combining multiple sets complementary derived systems with different NN components. Three structures are used to implement c-vectors, namely 2D self-attentive, gated additive, bilinear pooling structures, relying on attention mechanisms, gating mechanism, low-rank mechanism respectively. Furthermore, neural-based single-pass pipeline is also proposed paper, which uses NNs achieve voice activity detection, change point embedding extraction. Experiments detailed analyses conducted challenging AMI NIST RT05 datasets consist real meetings 4--10 speakers wide range acoustic conditions. For trained training set, relative error rate (SER) reductions 13% 29% obtained using c-vectors instead dev eval respectively, reduction 15% SER observed RT05, shows robustness methods. By incorporating VoxCeleb data into best system achieved 7%, 17% and16% compared d-vector dev, eval, respectively