No abstract

According to Jensen's inequality, the Bayesian information criterion (BIC) based on the Gaussian mixture model (GMM) is applied to speaker indexing. It can utilise the advantages of BIC and GMM. Experimental results have demonstrated that it is superior to both single-Gaussian-based BIC and GMM for speaker indexing.Introduction: Speaker indexing sequentially detects points where a speaker identity changes in a multispeaker audio stream, and categorises each speaker segment, without any prior knowledge about the speaker. In general, at the beginning, there is no sufficient speaker data to accurately estimate a speaker's model. So incremental speaker model updating has been popular in creating a speaker model in speaker indexing. That is to say, speaker model construction and speaker indexing are performed simultaneously [1]: for each utterance, the input speech is identified whether it belongs to one of the previous speakers. If so, the current speaker is regarded as one of the previous speakers and the corresponding speaker model can be updated. Otherwise, a new speaker model is created using the current speech.The conventional method of speaker indexing based on the Bayesian information criterion (BIC) is formulated in [2], which assumes a single Gaussian model for each speech segment and performs speaker clustering based on the BIC result. It utilises only a single Gaussian model (SGM) in BIC; and is called SGM-BIC. Because speaker information may not be fully represented with an SGM, SGM-BIC cannot cope with too short or too long speech segments effectively. To capture more speaker information, in this Letter, a Gaussian mixture model (GMM) is used to model every speech segment. According to Jensen's inequality, BIC based on GMM (GMM-BIC) is applied to speaker indexing.

No abstract