Speech recognition based on unified model of acoustic and language aspects of speech

Automatic speech recognition has been attracting a lot of attention recently and is considered an important technique to achieve natural interaction between humans and machines. However, recognizing spontaneous speech is still considered to be difficult owing to the wide variety of patterns in spontaneous speech. We have been researching ways to overcome this problem and have developed a method to express both the acoustic and linguistic aspects of speech recognizers in a unified representation by integrating powerful frameworks of deep learning and a weighted finite-state transducer. We evaluated the proposed method ill an experiment to recognize a lecture speech dataset, which is coilsidered as a spontaneous speech dataset, and confirmed that the proposed method is promising for recognizing spontaneous speech.