A noise estimation method based on improved VAD used in noise spectral suppression under highly non-stationary noise environments

A noise spectral estimation method, which is used in spectral suppression noise cancellers, is proposed for highly non-stationary noise environments. Speech and non-speech frames are detected by using the entropy-based voice activity detector (VAD). An adaptive parameter and variable thresholds are newly introduced for the VAD. The former can stabilize the entropy used in the VAD. The latter is used to discriminate the noisy speech into three categories, a non-speech frame, a quasi-speech frame and a speech frame. The noise spectrum is estimated by using the noisy speech spectrum in the non-speech frames and the weighted noisy speech spectrum in the quasi-speech frame and the speech frame. The weighting function used in the quasi-speech frame is modified from the conventional approach to suppress over estimation. These proposed techniques are very useful for rapid change in the noise spectrum and power. Simulations are carried out by using many kinds of noises, including white, babble, car, pink, factory and tank. The segmental SNR can be improved by 1.7 ~ 2.8dB and 0.6 ~ 1.8dB for the input SNR=3dB and 9dB, respectively. The noise spectral estimation error can be improved by 1.42 ~ 2.4dB and 0.6 ~ 1.4dB for the input SNR=3dB and 9dB, respectively.