Hybrid Architecture for Robust Speech Recognition System

The robustness and intelligibility of automatic speech recognition system (ASR) is a great challenge still today due to the variability involved with acoustic signals. Almost all the speech processing systems usually support noise free environments and are designed to recognize the words from predefined vocabulary. However, in the real world environment, out of vocabulary words (OOVs) as well as the background interference are the major sources of errors in recognizing the spontaneously spoken utterances. a novel approach by using speech enhancement technique such as RASTA at front-end and confidence measuring at back-end, to cope with noise and OOVs collectively is proposed in this paper. This new framework combines the long and short term features, and by taking the advantages of neural networks and statistical Gaussian mixtures by joining them in a single system. The experimental results, applied on Hindi language show a significant improvement over conventional system.