Using Genetic Algorithm to Improve the Performance of Speech Recognition
The development for speech recognition system has been for a while. The recognition platform can be divided into three types. Dynamic Time Warping (DTW) (Sakoe, 1978), the earliest platform, uses the variation in frame’s time for adjustment and further recognition. Later, Artificial Neural Network (ANN) replaced DTW. Finally, Hidden Markov Model was developed to adopt statistics for improved recognition performance. Besides the recognition platform, the process of speech recognition also includes: recording of voice signal, point detect, pre-emphasis, speech feature capture, etc. The final step is to transfer the input sampling feature to recognition platform for matching. In recent years, study on Genetic Algorithm can be found in many research papers (Chu, 2003a; Chen, 2003; Chu, 2003b). They demonstrated different characteristics in Genetic Algorithm than others. For example, parallel search based on random multi-points, instead of a single point, was adopted to avoid being limited to local optimum. In the operation of Genetic Algorithm, it only needs to establish the objective function without auxiliary operations, such as differential operation. Therefore, it can be used for the objective functions for all types of problems.
Because artificial neural network has better speech recognition speed and less calculation
load than others, it is suitable for chips with lower computing capability. Therefore, artificial
neural network was adopted in this study as speech recognition platform. Most artificial
neural networks for speech recognition are back-propagation neural networks. The local
optimum problem (Yeh, 1993) with Steepest Descent Method makes it fail to reach the
highest recognition rate. In this study, Genetic Algorithm was used to improve the
Consequently, the mission of this chapter is the experiment of speech recognition under the
recognition structure of Artificial Neural Network (ANN) which is trained by the Genetic
Algorithm (GA). This chapter adopted Artificial Neural Network (ANN) to recognize
Mandarin digit speech. Genetic algorithm (GA) was used to complement Steepest Descent
Method (SDM) and make a global search of optimal weight in neural network. Thus, the
performance of speech recognition was improved. The nonspecific speaker speech
recognition was the target of this chapter.