Oral Business English Recognition Method Based on RankNet Model and Endpoint Detection Algorithm

Abstract

The prosodic feature of “stressed syllables” plays an important role in business English learning. For the majority of nonnative language learners learning English, the pronunciation of native language can easily affect the expression of spoken English. For those who cooperate and communicate in business, if they want to have more idiomatic and accurate spoken English, they need to control the necessary condition of stress first. This puts forward new requirements for the existing speech recognition technology: more accurate recognition of stressed syllables, reading complex and different emotional colors, and helping to correct oral expression. The experimental results show that (1) the higher the Fisher Ratio of the feature, the easier it is to distinguish the stressed syllables; adjusting the weight of features can effectively improve the recognition accuracy. (2) The recognition rate will decrease with the increase of noise. (3) InwMS method can distinguish features better than min-max method, but the recognition rate of stressed syllables is very low. Linear recognition based on single feature is not recommended. (4) The error rates of the two methods are 20.6% and 19.32%. If any feature of the fusion feature is removed, the error recognition rate of the model will increase by at least 3%. (5) For sentence stress recognition based on fusion features, the recognition error rate on RankNet model is as high as 42.51%. The final result of system operation is good, simple, and convenient.