Machine learning paradigms are generally separated into supervised learning and unsupervised learning. Both of these paradigms have their own advantages in practice. But existing algorithms of these two paradigms also expose some hard problems in many different applications. In this paper, we first analyze the general problems of these two paradigms, and some successful techniques for boosting their performance. Then we propose a novel algorithm that can overcome some of existing problems through a mixture of these two paradigms. The algorithm is tested with a robot language-learning task. Equipped with this algorithm, our robot is able to acquire short audio information online, and gradually understand the audio input through human's intensive teaching.