Speech Stream Detection in Strong Noise based on Linear Prediction

The speech signal is usually mixed with a great deal of noise, and the noise weakens seriously the performance of the algorithms to detect speech signal. This paper presents a robust algorithm for speech signal detection in low SNR based on the linear prediction technology. The proposed approach firstly decreases noise by using linear predication residual, then decides whether speech signal is contained or not according to the coefficients statistics of LPC. The operation to speech component is only taken in prediction residual by enhancement, which produces little impairment to the formant of the speech. As most of the energy components of speech signals exist in the region between 300 Hz and 3000 Hz, only those LPC coefficients in this region are taken into account, which also reduces the influence from noise. Experiments show that it is particularly immunized for the proposed algorithm to the strong noise, especially in the white noise. At last, the performance of the algorithm is compared to the approach based on short-term energy in various noise condition, and quantified using the probability of correct classification. The results show that the proposed algorithm has an overall better performance than the referred approach, such as white noise and factory noise to low SNR