Semi-Supervised Transductive Speaker Identification

We present an application of transductive semi-supervised learning to the problem of speaker identification. Formulating this problem as one of transduction is the most natural choice in some scenarios, such as when annotating archived speech data. Experiments with the CHAINS corpus show that, using the basic MFCC-encoding of recorded utterances, a well known simple semi-supervised algorithm, label spread, can solve this problem well. With only a small number of labelled utterances, the semi-supervised algorithm drastically outperforms a state of the art supervised support vector machine algorithm. Although we restrict ourselves to the transductive setting in this paper, the results encourage future work on semi-supervised learning for inductive speaker identification.