A new method for sample selection in active learning

Abstract

Speech recognition systems are usually trained using tremendous transcribed samples, and training data preparation is intensively time-consuming and costly. Aiming at achieving better performance of acoustic model with less transcribed samples, active learning is adopted in acoustic model training to iteratively select the most informative samples corresponding to some sample selection method. And as the key part of active learning, sample selection method decides the performance. However, in active learning for acoustic speech recognition modeling, samples are always selected based on single predictor such as likelihood posterior probability and so on, which can not overall evaluate the samples. This paper proposes a sample selection method based on support vector machine using combination of several predictors in active learning for acoustic modeling. And our experiments show that active learning using our proposed sample selection method can achieve satisfying performance.