Look who's not talking

TL;DR

This paper introduces a unified speaker diarisation approach that leverages speaker embedding norms for speech activity detection, simplifying the process and improving accuracy in real-world recordings.

Contribution

It presents a novel method that uses speaker embedding norms for speech activity detection, eliminating the need for separate models and enhancing diarisation performance.

Findings

Outperforms popular baselines on multiple datasets

Effective speech activity detection without separate models

Simplifies speaker diarisation pipeline

Abstract

The objective of this work is speaker diarisation of speech recordings 'in the wild'. The ability to determine speech segments is a crucial part of diarisation systems, accounting for a large proportion of errors. In this paper, we present a simple but effective solution for speech activity detection based on the speaker embeddings. In particular, we discover that the norm of the speaker embedding is an extremely effective indicator of speech activity. The method does not require an independent model for speech activity detection, therefore allows speaker diarisation to be performed using a unified representation for both speaker modelling and speech activity detection. We perform a number of experiments on in-house and public datasets, in which our method outperforms popular baselines.