Automatic Compound Word Reconstruction for Speech Recognition of Compounding Languages

This paper compares two approaches to lexical compound word reconstruction from a speech recognizer output where compound words are decomposed. The first method has been proposed earlier and uses a dedicated language model that models compound tails in the context of the preceding words and compound heads only in the context of the tail. A novel approach models imaginable compound particle connectors as hidden events and predicts such events using a simple N -gram language model. Experiments on two Estonian speech recognition tasks show that the second approach performs consistently better and achieves high accuracy.