Example Based Large Vocabulary Speech Recognition

Опубликовано 17 августа 2016, 20:53

Research into Example Based Speech Recognition got revived over the past decade. Example based recognition is appealing for a number of reasons: there is considerable evidence from the linguistic literature that humans store actual traces of at least some sentences or phrases. Moreover, after 40 years of refining the HMM framework, we are still stuck with a model that is fundamentally flawed in a number of manners, most of the all the first order Markov assumption. Example based recognition avoids some of the traps of HMMs: (i) the data is not compacted into suboptimal models; (ii) all the data - with all its detail - is available at the moment of recognition. In this talk we will address besides the global framework some of the major challenges that we encountered. The example based approach is by and large the non-parametric statistical counterpart of the parametric HMMs. In this process we were confronted with some issues less prominent in HMMs distance metrics, outliers, merit of individual data points, distinguishing 'outright wrong' vs. 'unusual but correct', ... For some of these we came up with novel and interesting solutions, for others we surely don't have a definite answer. Also, for some of these, our interpretation had to be revised as we are moving to increasingly large databases.