A Review of Recent Advances in Speaker Diarization with Bayesian Methods

Abstract

This chapter aims to present some of the recent Bayesian approaches to speaker diarization (SD). SD is the task of grouping an audio document into homogenous regions, where each region should ideally correspond to the complete set of utterances that belong to a single speaker. Rich transcription, speaker adaptation of speech recognition systems and speaker recognition are some of the applications that require such a clustering procedure. Broadcast News, meeting, and telephone conversations are the main domains that SD is applied to. SD is a fully unsupervised clustering task. Not only we are not allowed to use any target-speaker enrollment data to detect the target speakers through the acoustic stream, but the number of speakers should be considered as an unknown, too. Moreover, text-independence should also be assumed, meaning that no transcript is available, either. Despite the effectiveness of several approaches and frameworks that have been proposed and tested in literature, the most natural and systematic approach to SD is to treat it as a model’s order selection task. Once the order is estimated (i.e. the number of speakers) the task reduces to a familiar (but not trivial at all) machine learning task where the latent variables (i.e. the speaker indicators of each utterance) of given cardinality should be estimated from the observations. Therefore, a major issue we deal with is how to assess the number of speakers in a way that is simultaneously robust and efficient. Bayesian machine learning is a highly principled paradigm and can naturally tackle model selection problems. It does so by applying consistently the rules of probability in order to infer the desired quantities, including the order of the model. Its superiority over the frequentistic statistical framework (e.g. Maximum Likelihood estimates, Classical Hypothesis testing) or semi-Bayesian approaches (e.g. MAP estimation, penalized maximum likelihood criteria) in model selection, averaging and density estimation has been verified in most (if not all) of the speaker related tasks, including identification and verification. Several drawbacks however still exist, most of which stem from the intractability of the majority of the ideal Bayesian solutions. Many well known and effective machine learning tools cannot be applied or require severe adaptation that may drastically increase their computational complexity. Nevertheless, the introduction of powerful approximate inference method (e.g. Variational Bayes, Expectation Propagation), novel Markov-Chain Monte Carlo techniques, along with the rapid development of the Bayesian nonparametric models 11