Speech Enhancement Algorithm Research Base On Short-Term Spectrum Estimation

Usually, the equipment in the mobile environment such as mobile phones, car-phones and so on code the voice signal based-on the linear predictive coding(LPC), but the LPC is very sensitive to the noise, therefore, in the noise environment, especially in the strong noise environment using the LPC to code the voice is unacceptable. Speech enhancement algorithm picks-up the pure voice from the noise speech. Therefore, speech enhancement algorithm research on mobile communication systems is great significance.Speech enhancement algorithm's early work is to estimate the noise, then the noise is applied in suitable denoising algorithm, the denoising algorithm is the key of the whole.Generally noise's mean is assumed zero, we only need to estimate the noise's variance, whether the noise's variance is accurate, will make a major impact to the follow-up denoising algorithm. In this paper, the classical noise estimation algorithm—the minimum tracking method was studied first. Using MATLAB simulation software simulation, the result shows that noise-estimation is very accurate in stationary noise environment, but non-stationary not. In non-stationary environment this paper researches another noise-estimation algorithm—non-stationary noise adaptive algorithm based on statistical information, the proposed algorithm estimates speech-presence probability exploiting of power spectral components in neighboring frames. The factor is computed based on the speech-presence probability, and the noise estimate is updated in each frame using a time-frequency dependent smoothing. Through simulation shows that the noise-estimation is very accurate in non-stationary noise environment, following-up the changing noise timely.About denoising algorithm this paper researches focused on Spectral Subtraction algorithm and improved form, STSA-MMSE algorithm, using MATLAB tool to simulate these algorithms. This paper improves the Classical STSA-MMSE algorithm, and applies non-stationary noise adaptive algorithm based on statistical information to the denoising algorithm. The results shows that the improved algorithm restrains the background noise and "music noise" effectively, which remnants in the original algorithm. This paper also takes the masking effect to the enhancement algorithm, improving voice's clarity and comfort. Comparing to the SNR, Itakura_Distance, PESQ and subjective feelings, whether from an objective evaluation criteria, or the subjective feelings, the enhanced voice by improved STSA-MMSE algorithm has been improved noticeably to the original algorithm.In this paper finally designs a really voice communication system, including three parts: Acoustics Echo Cancellation(AEC), Speech Enhancement and automatic gain control(AGC). AEC eliminates the echo of far-end speech in near-end speech; Speech Enhancement eliminates the noise in near-end speech; AGC solves the problem in near-end speech, which is the voice of near-end headset suddenly becoming big or small, because the far-end speech may be changed quickly, or the distance from the speaker in far-end to the microphone changes quickly, which makes the near-end speech changed quickly.