A Hybrid System with Symbolic AI and Statistical Methods for Speech Recognition

In this paper we show how symbolic Artii-cial Intelligence (AI) techniques can be combined with traditional Digital Signal Processing (DSP) and statistical methods to enhance speech recognition. We implement a hybrid system which uses a rule-based expert system to create a Conceptual Dependency (CD) representation of the spoken input. Conceptual Dependencies are used for natural language understanding of written text, but until now have not been applied to speech recognition. Our hybrid systems uses a three-step process. First, we implement continuous speech recognition using a keyword spotting system, based on Hidden Markov Models (HMM). Then, the recognized keywords are used to create the CD representation. Finally, the gaps in the CD representation are used as contextual cues for reinterpret-ing the speech input and so increase the speech recognition accuracy. Using these three techniques together increases the accuracy of meaning representation from 20% to 70%, a signii-cant improvement.