Research & Implementation Of Morphology-Based Speech Recognition System

Speech recognition is a hotspot in the artificial intelligence. It is popularly used in many fields, which improves human-computer interaction and bring great benefits to us. The performance of most speech recognition systems degrades rapidly in adverse environments that are inevitable in real-word applications. It is therefore important to research the robustness of speech recognition systems. A very important class of nonlinear digital signal processing stems from the concepts of mathematical morphology and is known as morphological filters. Morphological filters have found wide applications in many research fields, such as image analysis and processing, computer vision, etc. In recent years morphological filtering has applied to one-dimensional signal processing tasks step by step and is given increasing attention in speech processing due to the good performance.This paper focuses on the research of mathematical morphology in one-dimensional speech signal processing. Based on the traits of digital speech signals, a new morphological filtering algorithm is proposed in speech enhancement. Experiments show that the new filter is quite effective to suppress noise and improve the quality of noisy speech. Further, we introduce morphological filtering into speech recognition and propose a new noisy speech recognition method based on morphology. To reduce the distortion effect causing by enhancement in speech recognition, the morphological filter is employed to preprocess the clean training speech examples which is the same as the enhancement algorithm for the testing speech data. The new method with robust performance in noisy condition can improve the match between training and testing.In this paper, an isolated-word, continuous density hidden Markov model (CHMM) recognition system based on the proposed morphological speech recognition method using MFCCs as feature vectors is constructed. The recognition results show that the new method is of great robustness and the performance of recognition system can be improved effectively under noisy environments.