google/uis-rnn

# Technical Summary The library implements a clustering algorithm for sequential data using interleaved-state RNNs with PyTorch, accepting variable-length embedding sequences and outputting speaker cluster assignments. It features block-wise data augmentation during training to improve robustness with limited data, and supports both list-based and concatenated sequence input formats for flexibility across different dataset sizes. The model operates on pre-computed embeddings (speaker d-vectors) rather than raw audio, enabling integration into larger speaker diarization pipelines.

1,589 stars. No commits in the last 6 months.