A Method Of Keyword Recognition Based On Fuzzy Theory

Keyword recognition, which is used to identify one or more specific keywords from a continuous speech stream, is a branch of continuous speech recognition and is an important research direction in speech signal processing. Keyword recognition has some merits relatively to continuous speech recognition, such as short time-consuming and high accuracy. The technology is paid more and more attention with a wide range of applications in many areas.In this paper, the Chinese small vocabulary of keywords based on Hidden Markov Model technology is oriented. Due to the unrestrict of speaker and speaking method, a number of confusing similar to the non-keywords or keywords similarity between pronunciations often appear in a continuous stream of speech, the system identification rate has dropped rapidly. self-recorded voice is used in this paper. With regard to such problems, from the beginning of pre-treatment of voice signals, short time average energy method and short time average magnitude are adopted to estimate the starting point and the end point of the speech signal in order to remove the silent voice signal segment. Continuous hidden Markov model are used to establish models for keywords and non-keywords which is needed to extract 10-dimensional MFCC cepstral coefficients and its first-order differential parameters as feature vectors. In the phase of training, Baum-Welch algorithm is used for iterative revaluation. In the detection phase of keywords spotting, frame synchronous Viterbi search algorithm is adopted. In order to truly contain the keyword entered, a number of candidate keywords are often selected.In the stage of keywords confirmation, a fuzzy c-means clustering algorithm is introduced in order to avoid that keywords are replaced by the non-keywords which sound similar to the true ones and that the anti-keywords are not selected perfectly causing the recognition rate dropped. The membership principle is introduced. Taking Keyword models and anti-keyword models for cluster centers, the candidate keywords required to cluster. The candidate keywords are confirmed to its own category according to the result of clustering. The candidate ones which is clustered around the anti-keywods are refused first. At the confirmation stage, the application of the keywords likelihood rule are used for final confirmation. In addition, the research and experiments of keywords confusing network are studied on, and are compared with keyword recognition based on fuzzy theory.Experiments show that the recognition rate increased markedly in the entire system because of the introduction of fuzzy clustering, indicating the effectiveness of the algorithm. In the process, the clustering of the wild away from the cluster center point of the problem often generate, an improved fuzzy c-means clustering method is used to make better and more accurate clustering results.