Voice Keywords Recognition Based On Fuzzy Theory

Keyword spotting is an important area in speech recognition. Its objective is to identify and verify a few specified key-words in continuous speech. Comparing with keyword spotting, continuous speech recognition need more resources and its process speed is lower, and it's more vulnerable to noise. So continuous speech recognition is not suitable in many applications and keyword spotting is preferred.This paper conducts an in-depth study of keyword spotting algorithm. For the overlapping between categories in clustering process, a fuzzy pattern recognition algorithm is proposed and successfully solves the problem, moreover, it makes some improvements in bionic pattern recognition algorithm. The main work of this paper focuses on the following aspects:Chinese syllable is recognition unit in this paper. In order to detect every syllable accurately, a twice detection method based on the energy-frequency-value is proposed, and it can detect single syllable.Features extracted is briefly formulated. For example, short-time energy and zero-cross rate are used for endpoint detection, Mel-frequency cepstrum coefficient and critical-band feature vector is useful in the modeling process. As features in frequency domain have nothing to do with time, so complex calculation of dynamic time warping is avoidable .Modeling method of multidimensional hyper-ellipsoid has been perfected in bionic pattern recognition, equation and discriminant function of multidimensional hyper-ellipsoid is educed. Considering the complexity of realization, small chain of hypersphere instead of hyper-ellipsoid can greatly simplify algorithm. In addition, this paper analyses the overlapping between categories and proposed solutions.(1)A concept of degree of membership with category is proposed. The attribution of samples falls into the overlapping region is determined by comparing the degree of membership.(2)Membership functions based on distance, volume and discriminant function are proposed in hypersphere and hyper-ellipsoid cases. After compared these membership functions from feasibility and complexity of algorithm, this paper comes to an optimum membership function finally and proposes a whole recognition algorithm. (3)The confirmation of the identification results is detailedly described. Because of the degree of membership, the confirmation is realized by comparing the degree of membership with threshold.The experiments show that the rejection rate reduces 11 percent by applying fuzzy pattern recognition algorithm.