Automatic Speaker Clustering Using a Voice Characteristic Reference Space and Maximum Purity Estimation

Abstract

This paper investigates the problem of automatically grouping unknown speech utterances based on their associated speakers. In attempts to determine which utterances should be grouped together, it is necessary to measure the voice similarities between utterances. Since most existing methods measure the inter-utterance similarities based directly on the spectrum-based features, the resulting clusters may not be well-related to speakers, but to various acoustic classes instead.This study remedies this shortcoming by projecting utterances onto a reference space trained to cover the generic voice characteristics underlying the whole utterance collection. The resultant projection vectors naturally reflect the relationships of voice similarities among all the utterances, and hence are more robust against interference from non-speaker factors. Then, a clustering method based on maximum purity estimation is proposed, with the aim of maximizing the similarities between utterances within all the clusters. This method employs a genetic algorithm to determine the cluster to which each utterance should be assigned, which overcomes the limitation of conventional hierarchical clustering that the final result can only reach the local optimum. The proposed clustering method also incorporates a Bayesian information criterion to determine how many clusters should be created.

Statistical Utterance Comparison for Speaker Clustering Using Factor Analysis

“…where is the mixture occupation probability of Gaussian component at time : (21) Applying (15) to the first two Gaussian terms in (19), and , we have…”

“…We compared our model with two well-known similarity metrics: the Generalized Likelihood Ratio (GLR) and the Cross-Likelihood Ratio (CLR). The Generalized Likelihood Ratio (GLR) [21] between some utterance and some utterance is defined as (53) where is a single set of model parameters trained on both and merged together. The Cross-Likelihood Ratio [16] is (54) A previous study used either GMMs or a VQ model [16] for the parameters and , but here, we simply used the speaker factor vector obtained via ML-estimation [19] to build an adapted GMM for each utterance via (5).…”

“…Since the clustering task is done in the context of speaker diarization, however, we also use the well-known NIST diarization error rate (DER) with the standard collar of 0.25 s [22] to get a better feel for how each distance measure would fare in a standard diarization setting. Cluster purity is defined as [21] (55)…”

“…Note, however, that also when each cluster contains only one (speaker-pure) utterance, i.e., if we do not do any clustering at all. Therefore, the purity measure only makes sense when used in conjunction with a measure on how accurately the number of speakers was estimated, as was done in other experiments [16], [21]. We use the following measure for the overall speaker number accuracy [16]:…”

Speaker clustering using vector representation with long-term feature for lecture speech recognition

“…Recently, Tsai et al [6] presented an automatic speaker clustering algorithm using a voice characteristic reference space and maximum purity estimation, with the aim of maximizing the similarities between utterances within clusters. Tang et al [7] applied a complete treatment for a partially supervised speaker clustering to assist the unsupervised speaker clustering process.…”