Developing a method to build Japanese speech recognition system based on 3-gram language model expansion with Google database

Abstract

We have developed a method to build a Japanese automatic speech recognition (ASR) system based on 3-gram language model expansion with the Google database. Our aim is to enhance the recognition accuracy of ASR systems based on the 3-gram language model, even in cases where the language model is trained using short text segments. We investigate a practical approach to expanding language models by using 3-gram information from external web documents. In addition, we filter 3-gram entries on the basis of term frequency-inverse document frequency (TF-IDF) scores and the output of the Yahoo! web API to prevent the unnecessary addition of redundant or irrelevant 3-gram entries. In the experiments, we achieved an improvement of 0.71% in the word error rate and proved that the recognition accuracy can be improved by combining the proposed method and the traditional back-off smoothing technique without any costs being incurred in collecting additional text for training the model.