This paper considers the online semi-supervised learning (OSSL) problem in which the data are a mixture of both labeled and unlabeled samples and appear in a sequential (stream) manner. OSSL is very common in real applications and similar to the human-like learning process. Prototypebased classifiers, which represent the data of different classes by some prototypes, are natural in a streaming scenario by updating the prototypes with online (incremental) learning. However, most of previous prototype-based models are either designed for supervised or unsupervised learning separately. In this paper, we propose a novel model called online adaptive vector quantization (OAVQ) aiming at improving the classification performance in case of OSSL. Specially, we use the learning vector quantization (LVQ) criterion for updating the prototypes when the data point is labeled, and the frequency sensitive competitive learning (FSCL) criterion for adjusting the prototypes when the data point is unlabeled. The labeled and unlabeled data are coming randomly in a sequential manner, and these two criteria are used alternatively to learn the positions of prototypes. In this way, we can make full use of both supervised and unsupervised information to further boost the performance. Experiment results on several databases verify the effectiveness and applicability of the proposed method in improving the performance for OSSL.