Classification of Room Impulse Responses and its application for channel verification and diarization

This paper describes experiments on a classification of Room Impulse Responses (RIRs) from recordings where acoustic channel is described by corresponding RIRs. The classifiers are trained on large sets of synthetic RIRs and then used as extractors of RIR embeddings bearing information about acoustic characteristic of a room as well as the locations of sound source and microphone. Experiments on different datasets demonstrate that RIR embeddings can be used for verification of acoustic channel and for the diarization in meeting-like scenarios where speakers' positions are fixed. The verification experiments on VoxCeleb1 and BUT Speech@FIT Reverb show reasonable performance of RIR embeddings. The diarization results with RIR embeddings on LibriCSS dataset are better than those with state-of-the-art speaker embeddings that shows the potential of the proposed approach.