A Passwords Recognition System Based On ZCPA Feature Parameter

Speech recognition technology after decades of development, great progress has been made。In particular, the successful application of HMM, making the recognition rate has been greatly improved。However, HMM-based speech recognition systems model for the training environment and test environment mismatch,Because of the affect of convolution noise and background noise, Test speech feature distribution is different from the distribution of characteristic parameters described by Model, Recognition results will be significantly decreased。Therefore , Improve the performance of speech recognition system In different environments,That is Improve the robustness of recognition system, It has become to the priorities and hot spots of the current speech recognition research.First, based on the principles of speech contamination and statistic models, the reason why mismatch will degrade the performance is introduced.Second, the paper introduced the common methods of improving the system robustness. Generally, because the training and testing conditions does not match the performance degradation caused by, through the specific environment re-training to be compensated. Nevertheless, the way is able to solution to environmental mismatch problems. In practice, it is difficult to collect large amounts of field training data. Therefore, this re-training method is very limited scope of application. So far, many researchers do a lot of work to improve the performance of speech recognition system in noise environment, they also propose a lot of effective methods. For example: Speech enhancement,Parameter compensation,Model compensation and so on. These methods can improve the system robustness, but they have some limitations in actual use.Third, based on the mechanism of human hearing perception, extract better robust characteristic parameters. If we can extract a characteristic parameters, which less affected by noise, it can also be effective in improving system performance. We know that the human ear can do good job in a noisy environment, this show that human speech perception and recognition capabilities and noise robustness better than all the automatic speech recognition system. So if human auditory perception mechanism of speech information can be used to speech recognition, there may improve speech recognition rate. ZCPA is a characteristic parameter which based on that, the parameter simulates the neural firing pattern occurring in human cochlea. The frequency information of speech signal is obtained by zero-crossing intervals, and intensity information is incorporated by a peak detector and a compressive nonlinearity. A passwords recognition system combines ZCPA with HMM .Experimental results of the passwords recognition system in noisy environments show that ZCPA has a better robustness than MFCC.