Dsp-based Speech Spectrum Envelope Extraction Method And Realization

The pitch period and the spectral envelope of speech represent acoustical characteristics of the speech signal. The pitch period shows the periodicity of lung-airflow in the process of phonation. And the spectral envelope of speech reflects on the resonance-filter property of the vocal tract model, which carries the semantic and personalized information of human. The extraction of speech spectral envelope is an important task of speech signal analysis. This paper mainly researches on the methods of extracting speech spectral envelope.Firstly, this paper researches on the conventional methods, including the time-domain method and frequency-domain method. Meanwhile, the influence of different methods and noise on the extraction has been discussed. Then, aiming at the problems of the combination and fictitiousness of formant when extracting the spectral envelope, an algorithm using piecewise linear predictive coding (PLPC) is proposed in this paper. In PLPC, multi-channel filter banks are used to decompose the sprectral coverage of speech to different signal component, and the linear predictive coding is applied to the different signal component for extracting speech spectral envelope. A bank of filters with uneven bandwidth is designed in order to fit the frequency resolving capability of the hearing in PLPC. The conventional linear predictive coding (LPC) is ameliorated in PLPC. The spectrum of noise is divided by the filters at the same time, so PLPC is not sensitive to noise. The results show that the extraction of speech spectral envelope based on PLPC is more pricise than conventional methods. Besides extracting the spectral envelope, PLPC could be used to estimate the formant frequency and detect the pitch period. At last, the implementation of PLPC is developed in this paper, based on the hardware of DSP chip TMS320C5509A. The fixed point implementation of the program code is discussed. The computational complexity and the memory storage requirement are also analyzed. The method based on PLPC could, efficiently, resolve the problems of the combination and fictitiousness of formant when extracting the spectral envelope. It has low computational complexity, and could be practical for speech signal analysis.