Speech Recognition Algorithm Based On Straight Spectrum Study

Language is the most important of human intercourse. With the advent of the information era, a set of techniques for speech information processing has been indispensable in Modern Society. Automated Speech Recognition (ASR for short) is one of the most important factors, which converts speech to text, and disengages people from dependence on keyboard. In a variety of parameters describing the characteristics of speech signals, spectrogram has many advantages over others, which absorbs their merit of analysis in both time and frequency domains. With the development of the computer technology, the disadvantage is no longer apparent.The LPCC and MFCC, which parameters are used most extensively in ASR, all decompose voice signal into a separate excitation source parameters and filter parameters, and get the cepstrum coefficients from the filter parameters. In the same way, The STRAIGHT algorithm is to resolve voice into the convolution of channel spectrum and a series of pulses, and displays the channel spectrum in the form of spectrogram, which is called STRAIGHT spectrum.The difference subspace is an effective training algorithm for models. Theoretical proof and experimental results show that the template trained by this method has no reference to selection of training samples, and it will be more representative as the sample grows larger. In the system designed in this paper, STRAIGHT spectrum is selected as the identifiable feature, and the difference subspace is for template training. In addition, the mapping with corresponding points is introduced, which is effective than DTW for solving the problem of alignment between different sub-word units, thereby solving the problem of speaker independent.The speech recognition system is built in MATLAB Environment. The Speech of Chinese digits 1-10 which are recorded by various persons at different times, are used in experiments. The result shows that this method reaches satisfied results.