Research Of Voice Recognition System Based On Optical Cable Environment

Optical cable, as one of the main communication mediums, plays an important role in the modern communication systems. The broader optical cable lays, the more difficult cable maintenance work becomes. At the same time, various behavior of the surrounding environment along the optical cable may bring huge obstacles to the maintenance work. Therefore, it is necessary to monitor the running state of cable environment in order to take timely remedial measures for various kinds of emergent events. Hence guarantee the quality of communications.Voice, as a kind of inherent signal when activities occur, inevitably carries some specific information of this kind of activities. Through the analysis of such signals, we can distinguish among different events. According to the different nature of events, we can judge the pros and cons of the event on the cable environment so that we could take corresponding measures. Different sound signals can identify different events which bring about the sound. To achieve this goal, it needs voice recognition system.Recognition system mainly includes the following several parts: preprocessing, feature extraction and classification. The preprocessing includes four parts:sample quantization, pre-emphasis, frame and window and endpoint detection process. Preprocessing deals with sound signals acquired by sensors. Its main purpose is to filter and reduce noise so to reduce the interference of invalid elements. The preprocessing will also split the sound signals into short pieces so that they can maintain the smooth performance of sound. With the preprocessing step, we can conduct next process. The feature extraction process is to analyze the short-term characteristic component of the voice samples, which can distinguish with the other voice samples. Classification matching includes two aspects:study of known voice samples and training of the unknown ones. Good classification and matching has very good robustness.Based on the above contents, a corresponding recognition system experiment is carried out. Voice samples of the experiment are collected from an optical cable system which is in use. And the main platform for the experiment is MATLAB, with many toolboxes related to recognition supporting the experiment. The experiment begins with the analysis of a voice sample, in both time zone and frequency zone. In order to improve the purity of sample, noise deduction has to be done. When dealing with preprocessing, there’s a comparision among voice sample pieces with different length in both time and frequency zone. In case of feature extraction, I choose mel-frequency cepstrum coefficients, MFCC for short, as the feature. Besides, a diagram, from which we kown the lower ingredients mainly contains the information, is shown with MFCC extracted from the voice sample. Based on the diagram, a test with different ingredients of MFCC is carried out to show the recognition results of two voice samples by linear classifier. By choosing the fittest combination of features, there’s the recognition results of eleven voice samples. At last, a three-layer neural network is found by the neural network toolbox. By adjusting the learning algorithm and training times to simulate the recognition process of three voice samples, a table is diplayed to show the total running time and the recognition results with different neural network models.