UIS-RNN

UIS-RNN (google/uis-rnn) is the library behind 'Fully Supervised Speaker Diarization' (arXiv 1810.04719), the 2018 paper that argued diarization's clustering step shouldn't be unsupervised when labeled data exists. The model is an unbounded interleaved-state RNN: each speaker gets their own RNN instance sharing parameters, new instances are born via a Bayesian non-parametric process (so the speaker count is learned, not guessed), and the interleaved states carry speaker information through time. Google featured it on the AI Blog in November 2018 ('Accurate Online Speaker Diarization'), and it influenced a wave of supervised diarization work. The open-source reality needs stating plainly: this is NOT an official Google product (says so in the README), the released code is a simplified reimplementation of the internal system, and — critically — Google cannot share the d-vector speaker-embedding system the paper's results depend on, so you get the algorithm but must supply your own embeddings and training data (the repo ships only toy data for the demo). Dependencies are pinned to the PyTorch 0.4/1.3 era, and the repo has been effectively archival for years. Include it for what it is: a landmark idea with runnable reference code, not a production component.

• Paper reports strong results on internal Google data with d-vector embeddings (author-reported, 2018)\n- The open release ships toy demo data only — no reproducible benchmark possible from the repo as published