Spontaneous Mandarin speech understanding using Utterance Classification: A case study

As speech recognition matures and becomes more practical in commercial English applications, localization has quickly become the bottleneck for having more speech features. Not only are some technologies highly language dependent, there are simply not enough speech experts in the large number of target languages to develop the data modules and investigate potential performance related issues. This paper shows how data driven methods like Utterance Classification (UC) successfully address these major issues. Our experiments demonstrate that UC performs as well as or better than hand crafted Context Free Grammars (CFGs) for spontaneous Mandarin speech understanding, even when applied without linguistic knowledge. We also discuss two pragmatic modifications of the UC algorithm adopted to handle multiple choice answers and to be more robust to feature selections.