Encoder-Decoder Based Attractors for End-to-End Neural Diarization

This paper investigates an end-to-end neural diarization (EEND) method for unknown number of speakers. In contrast to the conventional cascaded approach speaker diarization, EEND methods are better in terms overlap handling. However, still has a disadvantage that it cannot deal with flexible To remedy this problem, we introduce encoder-decoder-based attractor calculation module (EDA) EEND. Once frame-wise embeddings obtained, EDA sequentially generates speaker-wise attractors on basis sequence-to-sequence using LSTM encoder-decoder. The generation continues until stopping condition is satisfied; thus, can be flexible. Diarization results then estimated as dot products and embeddings. from overlaps result larger product values multiple attractors; overlaps. Because maximum output speakers limited by training set, also propose iterative inference remove restriction. Further, aligns external speech activity detector, which enables fair comparison against approaches. Extensive evaluations simulated real datasets show EEND-EDA outperforms approach.