Combining Outputs of Multiple LVCSR Models by Machine Learning

摘要

This paper proposes to apply machine learning techniques to the task of combining outputs of multiple LVCSR models, where, as features of machine learning, information such as the models which output the hypothesized word, its part-of-speech, and its syllable length are useful for improving the word recognition rate. Experimental results show that the combination result outperforms several baselines including model combination by voting such as ROVER in the word recognition rate. Furthermore, unlike model combination by voting, word recognition rate of model combination by machine learning is not damaged even in the case where only the minority of the participating models perform well in the word recognition task.