Automatic allocation of training data for rapid prototyping of speech understanding based on multiple model combination

The optimal choice of speech understanding method depends on the amount of training data available in rapid prototyping. A statistical method is ultimately chosen, but it is not clear at which point in the increase in training data a statistical method become effective. Our framework combines multiple automatic speech recognition (ASR) and language understanding (LU) modules to provide a set of speech understanding results and selects the best result among them. The issue is how to allocate training data to statistical modules and the selection module in order to avoid overfitting in training and obtain better performance. This paper presents an automatic training data allocation method that is based on the change in the coefficients of the logistic regression functions used in the selection module. Experimental evaluation showed that our allocation method outperformed baseline methods that use a single ASR module and a single LU module at every point while training data increase.