The Research And Realization Of Speech Feature Extraction Based On Non-parametric Spectral Estimation

The research of automatic speech recognition technology is an important topic in pattern recognition field, and also plays an important role in the actual application. At present, most of the current speech recognizers get more acceptable recognition accuracy for clean speech, while the performance degrades sharply when they are subjected to noise which often exists in practical environments. Therefore, it is necessary and of great importance for the research of speech recognition.This thesis mainly focused on the robustness of the front-end processing in speech recognition--speech features extraction, and analyzed the shortage of the traditional speech feature extraction based on spectrum estimation. In this paper, a speech feature extraction method was proposed which combined with the characteristics of the human voice and auditory properties with better robustness and recognition performance, and the speech feature extraction algorithm was transplanted in the embedded DSP platform.Firstly, this paper analyzed kinds of non-parametric spectrum estimation methods, and pointed out that the linear predictive coding (LPC) spectrum estimation method and the fast Fourier transform (FFT) spectrum estimation method applied to speech feature extraction, however, the spectrum resolution was poor, which leaded to its robustness of speech recognition greatly influence in the actual environment. As a result, Minimum Variance Distortionless Response (MVDR) was introduced to realize higher resolution and higher robustness of speech recognition feature extraction.Secondly, consideration of the characteristic of non-stationary random signal of the speech signal, this paper proposed a new robust speech feature extraction method--the MHCC (Hilbert-MFCC) feature extraction method for speech, which introduced the Hilbert transform to MFCC process. Using this method, effective instantaneous magnitude characteristics and instantaneous frequency characteristics were extracted, while improving the discriminative of the feature, so as to the purpose of improving the robustness and recognition rate of speech recognition. By training and recognition, the experimental results showed that compared with the classic MFCC feature extraction and the traditional MVDR-MFCC feature extraction, the proposed MHCC feature extraction methods achieved better performance, that is to say, lower algorithm complexity, higher anti-noise, better recognition, and more suitable for the realization of the embedded platform.At last, the programmes of MFCC and MHCC algorithm were compiled using the C language. The successful C procedure was transplanted to integrated development environment CCS3.1. In order to realize in the TMS320C6713DSK and real-time extraction parameter values using proposed two methods, we made full use of the advantages of processing signal of DSP and optimized algorithm and code.