Comparing and combining classifiers for self-taught vocal interfaces

摘要

An attractive approach to enable the use of vocal interfaces by impaired users with dysarthric speech is the use of a system which learns from the end-user. To enable such technology, it is imperative that the learning is fast to reduce the time spent training the interface. In this paper we investigate to what extend various machine learning techniques are able to learn from only a single or a few spoken training samples. Additionally, we explore whether these techniques can be combined through boosting to improve the performance. Our evaluations on a small, but highly realistic home automation database reveal that non-negative matrix factorization seems best suited for fast learning and that some of the boosting approaches can indeed improve performance, especially for small amounts of training data.