Speech Sensory Mapping For Autonomous Mental Developing Robot

Most people communicate with each other by listening and speaking. Compare to other methods, such as writing, communication by speech does not require educational and professional background of users; it is the most popular communication method. So if speech can be used as the interface between intelligent robot and human users, robot will be much more easy to use and will bring much more convenience to people. At the same time, as a promising theory, Autonomous Mental Development is believed to be able solve complicated, muddy and task nonspecific problems. Therefore it is worth to try to combine the advanced Autonomous Mental Development paradigm with classical speech processing methods.In the past, all speech processing systems and methods were designed for specific tasks in limited condition. This paradigm is not suitable for a intelligent robot which will work in a natural environment. Focus on this problem, based on autonomous mental development paradigm and classical Hidden Markov Model, Double Hidden Layer Markov Model is introduced in this thesis. This model and its related algorithms meet the requirements of incremental computing, online computing which are demanded in Autonomous Mental Development paradigm. This model can automatically generate and tune the internal representation for speech. Thus a system based on this model is capable to interact with users without limit of domain and specific task, is expected in this article.For instance, acoustic model and language model must be trained and tuned by operator before a speech recognition engine can work, thus the performance will decreased dramatically when the environment and domain are changed. In contrast to this, a series of rules like incremental and online computing, in place computing, nonspecific task were introduced by autonomous mental development paradigm in recent years. Conforming these rules, algorithms systems will perform well in a changing environment.On the basis of above discussion, the author defines the problem of sensory mapping for autonomous mental development robot speech processing, and develops a solution for this problem. The key of the solution, Double Hidden Layer Markov Model and related searching and training algorithms are introduced in this article. Experiments show that this solution is better than traditional methods in ability of adaption to new environment and domain.