In this paper, by using the three different lan-guage models, the contents containing more re-lated grammatical errors will help to improvethe speech recognition accuracy in the GEDsystem. We introduced the generation of textwith grammatical errors by using an NMTmodel in the GED system, which is helpful forlanguage modeling without taking much time tolabel the transcriptions from spoken utterances.Normally, the NMT only translates one sen-tence to one different sentence. In order to copewith this limitation in our GED system, thereare two methods: one is keeping more grammat-ical errors in one generated sentence; the otherone is splitting the long sentence into severalshort ones in order to add more errors. In thefuture, we plan to calculate more English speechdate from Japanese native students and changethe architecture in NMT model with the state-of-the-art techniques to automatically generatemeaningful sentences with errors for the GEDsystem.
展开▼
